Rapid prediction method applied to digital twinning of surface flow field of aircraft

Abstract

The invention discloses a rapid prediction method applied to digital twinning of a flow field on the surface of an aircraft, and relates to the technical field of aerodynamic design of aircrafts. The problems of low calculation efficiency, poor geometric mutation region prediction precision and high high-fidelity data acquisition cost in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of aircraft aerodynamic design technology, and more specifically to a rapid prediction method for digital twins of flow fields on aircraft surfaces. Background Technology

[0002] Accurate prediction of the flow field on the surface of supersonic aircraft (involving complex physical phenomena such as shock wave interference, boundary layer separation, and aerodynamic heating) is a core foundation in aircraft aerodynamic layout optimization, thermal protection design, and flight safety assessment.

[0003] The existing technology has the following main problems: 1. Low computational efficiency: Traditional computational fluid dynamics (CFD) methods require the construction of millions to tens of millions of mesh elements, and single-condition calculations can take hours to days, which cannot meet the needs of rapid iteration in flight control, online trajectory planning, and conceptual design stages.

[0004] 2. Insufficient prediction accuracy for geometrically abrupt regions: Existing point cloud-based deep learning methods (such as PointNet++) only use basic features like coordinates and normal vectors, lacking explicit representation of key geometric features such as surface curvature. This results in a failure to accurately identify shock wave interference when dealing with geometrically abrupt regions such as wing leading edges, wing-body junctions, and rudder-wing gaps, leading to significant prediction errors.

[0005] 3. Dataset construction is difficult and costly: Three-dimensional flow fields involve a high-dimensional parameter space (Mach number, angle of attack, geometric parameters, etc.). Existing methods lack scientific guidance on parameter sensitivity analysis and often rely on uniform sampling and high-precision CFD data, resulting in huge computational resource consumption for dataset construction and weak model generalization ability across working conditions. Summary of the Invention

[0006] To overcome the shortcomings of the above technologies, this invention provides a rapid prediction method for digital twins of flow fields on aircraft surfaces, which utilizes curvature-enhanced point cloud neural networks, active subspace sampling, and multi-fidelity transfer learning strategies.

[0007] The technical solution adopted by this invention to overcome its technical problems is: A rapid prediction method for digital twins of flow fields on aircraft surfaces includes: S1. Generate an STL-formatted geometric model of the canard missile vehicle in SpaceClaim software, and obtain the set of face normal vectors for the triangular facets of the vehicle from the geometric model. and the cloud of points on the surface of the aircraft , , , For the first The face normal vectors of a triangular facet. , The number of triangular pieces, For the first A point cloud, This represents the number of point clouds; S2. Based on point cloud set Calculate neighborhood radius ; S3. Utilizing neighborhood radius Computing local voxel mesh ; S4. Based on the local voxel mesh The curvature was calculated ; S5. Calculate the first... using the face normal vectors of the triangular facets. Point cloud vertex normal vector ; S6. Use ICEM CFD software to construct a mesh model based on the geometry of the canard missile vehicle; S7. Determine the range of variation of the incoming flow parameters; S8. Input the mesh model into Ansys Fluent simulation software to obtain the lift coefficient of the canard missile vehicle through simulation. and drag coefficient ; S9. Lift coefficient of canard missile vehicles using the active subspace method and drag coefficient The analysis is performed to obtain the sensitivity, and the sampling point results are obtained based on the sensitivity. The target domain dataset is then generated using the sampling point results. S10. Obtain physical quantities using the target domain dataset. ; S11. For physical quantities Normalization is performed to obtain the physical quantity. Preserve normalized parameters ; S12. Calculate the rotated coordinates using the incoming flow parameters. and normal vector ; S13. Construct the PointCurvNeXt neural network model, and rotate the coordinates... Normal vector curvature The input is fed into the PointCurvNeXt neural network model, and the output yields the physical quantities of the flow field on the aircraft surface. ; S14. The physical quantities of the flow field on the aircraft surface will be obtained. Use normalization parameters Perform inverse normalization to restore the physical quantity.

[0008] Furthermore, step S2 includes the following steps: S2-1. Constructing a point cloud The bounding box, where the minimum X-axis coordinate of all point clouds within the bounding box is... The maximum value of the X-axis coordinate of all point clouds in the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the X direction. The minimum value of the Y-axis coordinate of all point clouds in the bounding box is The maximum value of the Y-axis coordinate of all point clouds in the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Y direction. The minimum Z-axis coordinate of all point clouds in the bounding box is The maximum Z-axis coordinate of all point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Z direction. ; S2-2. Through formula The bounding box volume of the point cloud was calculated. ; S2-3. Using the formula The equivalent total surface area was calculated. ; S2-4. Using the formula Calculate the neighborhood radius In the formula The number of target points in the neighborhood. .

[0009] Furthermore, step S3 includes the following steps: S3-1. Regarding the first Point cloud Perform neighborhood radius based The k-nearest neighbor index is obtained. A neighborhood point cloud; S3-2. Using the formula Calculate the first Point cloud Local density ; S3-3. Through formula The adaptive neighborhood radius is calculated. In the formula, For minimum density, Maximum density; S3-4. Regarding the first Point cloud Perform adaptive neighborhood radius The k-nearest neighbor index is obtained. A neighborhood point cloud; S3-5. Construction The bounding box of a neighborhood point cloud, where the minimum X-axis coordinate of all neighborhood point clouds within the bounding box is... The maximum value of the X-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the X direction. The minimum value of the Y-axis coordinate of all neighboring point clouds within the bounding box is The maximum value of the Y-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Y direction. The minimum Z-axis coordinate of all neighboring point clouds within the bounding box is The maximum Z-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Z direction. , size vector Dimension vector Dimension vector Each part is divided into three equal parts to construct a 3×3×3 local voxel mesh. .

[0010] Furthermore, step S4 includes the following steps: S4-1. Perform local voxel meshes using the breadth-first search algorithm. Access was performed and the results were obtained. The final sampling point set consists of 10 final sampling points. , ,in For the first The coordinates of the final sampling points ; S4-2. Through formula Calculate the first final sampling points Neighbor mean points ; S4-3. Through formula Calculate the offset vector ; S4-4. Through formula The covariance matrix with three rows and three columns was calculated. In the formula, For transpose; S4-5. On the covariance matrix Eigenvalue decomposition yields three non-negative eigenvalues. , , , Through formula The curvature was calculated .

[0011] Furthermore, in step S5, the formula is used... Calculate the first Point cloud vertex normal vector In the formula, This is a modulo operation.

[0012] Preferably, the incoming flow parameters in step S7 include: Mach number. Angle of attack and sideslip angle , The value ranges from 1.6 to 12.0. The value ranges from -20.0 to 20.0. The value ranges from 0 to 15.0.

[0013] Furthermore, step S9 includes the following steps: S9-1. Obtaining Mach numbers using the active subspace method. For lift coefficient and drag coefficient sensitivity Angle of attack For lift coefficient and drag coefficient sensitivity Sideslip angle For lift coefficient and drag coefficient sensitivity , ; S9-2. When Mach number When the value is between 1.6 and 2.0, it depends on the sensitivity. Mach number Using dense sampling with a step size of 0.2, when the Mach number... When the value is between 2.0 and 12.0, it depends on the sensitivity. Mach number A segmented, equal-step dense sampling method with a sampling step size of 0.5 and an endpoint not exceeding 12.0 was used to obtain a Mach number sampling point set. Based on sensitivity... angle of attack By employing dense sampling with a step size of 0.5, a set of angle-of-attack sampling points is obtained. Based on sensitivity... Side slip angle A sparse sampling step of 3.0 is used to obtain the sideslip angle sampling point set; S9-3. Based on the Mach number sampling point set, angle of attack sampling point set, and sideslip angle sampling point set, use Flunet software to solve the CFD and generate high-precision and low-precision flow field data. Use the high-precision and low-precision flow field data as the target domain dataset.

[0014] Furthermore, step S10 includes the following steps: S10-1. Generate surface unstructured meshes for the target domain dataset using pointwise software; S10-2. Using three-point interpolation, the surface flow field is mapped onto the unstructured surface mesh to obtain physical quantities. ; Furthermore, step S12 includes the following steps: S12-1. Through formula Calculate the first Point cloud Angle of attack of rotation around the Y-axis Rotated coordinates In the formula, For the first Point cloud Coordinates in the flow field body coordinate system; S12-2. Through formula Calculate the first Point cloud Sideslip angle of rotation about the Y-axis The rotated coordinates ; S12-3. Through formula Calculate the vertex normal vector Angle of attack rotating around the Y-axis Rotated coordinates ; S12-4. Through formula Calculate the vertex normal vector Sideslip angle around Z-axis The rotated normal vector .

[0015] Furthermore, step S13 includes the following steps: The S13-1.PointCurvNeXt neural network model consists of a PointNeXt network, an inflow parameter encoding module env_encoder, and a regression output module output_head; S13-2. Rotate the coordinates... The rotated normal vector curvature The input features are fed into the PointNeXt network's input feature projection module of the PointCurvNeXt neural network model, and the output features are obtained. ; The incoming parameter encoding module env_encoder of the S13-3.PointCurvNeXt neural network model consists of a first linear mapping layer, a ReLU activation function, and a second linear mapping layer, which encodes the Mach number. The input is given to the incoming stream parameter encoding module env_encoder, and the output is the feature. ; S13-4. Features With features Perform a concatenation operation to obtain the fused feature representation. ; S13-5. Representing the fused features The input is fed into the encoder-decoder of the PointNeXt neural network model, and the output is the feature. ; The S13-6.PointCurvNeXt neural network model's regression output module, output_head, consists of a first ReLU activation function, a BN layer, a non-linear activation layer, and a second ReLU activation function, sequentially processing the features. The input is fed into the regression output module output_head, and the output is the physical quantities of the flow field on the aircraft surface. .

[0016] The beneficial effects of this invention are: (1) Significantly improved computational efficiency: Compared with the problem of long time consumption of traditional CFD methods, the method of this invention can achieve rapid prediction after the neural network training is completed, which meets the real-time requirements of digital twin systems and supports online flow field prediction and real-time aerodynamic performance evaluation during flight.

[0017] (2) Improved accuracy and robustness of curvature calculation: The adaptive neighborhood radius method dynamically adjusts the neighborhood size according to the local point cloud density, effectively solving the technical problem of traditional neighbor methods being sensitive to point cloud density and losing accuracy in areas with varying density. The voxel effectively avoids the problem of the neighborhood crossing geometric boundaries, solving the key technical difficulty of the traditional method where the neighborhood index fails when the point clouds of two adjacent boundary surfaces are close, leading to errors in curvature calculation. This improves the accuracy and robustness of local geometric feature extraction.

[0018] (3) Reduced training sample requirements: This invention introduces active subspace sensitivity analysis and a non-uniform sampling strategy. By identifying key parameters and sampling them densely, and sampling minor parameters sparsely, the number of high-fidelity CFD calculations is reduced, the data acquisition cycle is shortened, and the data acquisition cost is lowered.

[0019] (4) High accuracy in flow field prediction: Based on the curvature-enhanced PointCurvNeXt network combined with encoder-decoder architecture and skip connection design, it can accurately capture complex flow phenomena such as shock waves, boundary layer separation, reattachment, and expansion waves. It achieves high accuracy in predicting key flow field features such as surface pressure, temperature, shock wave location, stagnation point pressure, and separation point location. Detailed Implementation

[0020] The present invention will be further described below.

[0021] A rapid prediction method for digital twins of flow fields on aircraft surfaces includes: S1. Generate an STL-formatted geometric model of the canard missile vehicle in SpaceClaim software, and obtain the set of face normal vectors for the triangular facets of the vehicle from the geometric model. and the cloud of points on the surface of the aircraft , , , For the first The face normal vectors of a triangular facet. , The number of triangular pieces, For the first A point cloud, This represents the number of point clouds.

[0022] S2. Based on point cloud set Calculate neighborhood radius .

[0023] S3. Utilizing neighborhood radius Computing local voxel mesh .

[0024] S4. Based on the local voxel mesh The curvature was calculated .

[0025] S5. Calculate the first... using the face normal vectors of the triangular facets. Point cloud vertex normal vector .

[0026] S6. Use ICEM CFD software to construct a mesh model based on the geometry of the canard missile aircraft.

[0027] S7. Determine the range of variation of incoming flow parameters.

[0028] S8. Input the mesh model into Ansys Fluent simulation software to obtain the lift coefficient of the canard missile vehicle through simulation. and drag coefficient .

[0029] S9. Lift coefficient of canard missile vehicles using the active subspace method and drag coefficient The analysis is performed to obtain the sensitivity, and the sampling point results are obtained based on the sensitivity. The target domain dataset is then generated using the sampling point results.

[0030] S10. Obtain physical quantities using the target domain dataset. .

[0031] S11. For physical quantities Normalization is performed to obtain the physical quantity. Preserve normalized parameters .

[0032] S12. Calculate the rotated coordinates using the incoming flow parameters. and normal vector .

[0033] S13. Construct the PointCurvNeXt neural network model, and rotate the coordinates... Normal vector curvature The input is fed into the PointCurvNeXt neural network model, and the output yields the physical quantities of the flow field on the aircraft surface. .

[0034] S14. The physical quantities of the flow field on the aircraft surface will be obtained. Use normalization parameters Inverse normalization is performed to restore the physical quantities. The restored physical quantities are then used to display the entire flow field contour map.

[0035] By employing curvature-enhanced point cloud neural networks, active subspace sampling, and multi-fidelity transfer learning strategies, this approach addresses the problems of low computational efficiency, poor prediction accuracy for geometrically abrupt regions, and high cost of acquiring high-fidelity data in existing technologies.

[0036] In one embodiment of the present invention, step S2 includes the following steps: S2-1. Constructing a point cloud The bounding box, where the minimum X-axis coordinate of all point clouds within the bounding box is... The maximum value of the X-axis coordinate of all point clouds in the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the X direction. The minimum value of the Y-axis coordinate of all point clouds in the bounding box is The maximum value of the Y-axis coordinate of all point clouds in the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Y direction. The minimum Z-axis coordinate of all point clouds in the bounding box is The maximum Z-axis coordinate of all point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Z direction. .

[0037] S2-2. Through formula The bounding box volume of the point cloud was calculated. .

[0038] S2-3. Using the formula The equivalent total surface area was calculated. .

[0039] S2-4. Using the formula Calculate the neighborhood radius In the formula The number of target points in the neighborhood. .

[0040] In one embodiment of the present invention, step S3 includes the following steps: S3-1. Regarding the first Point cloud Perform neighborhood radius based The k-nearest neighbor index is obtained. A neighborhood point cloud.

[0041] S3-2. Using the formula Calculate the first Point cloud Local density .

[0042] S3-3. Through formula The adaptive neighborhood radius is calculated. In the formula, For minimum density, This represents the maximum density.

[0043] S3-4. Regarding the first Point cloud Perform adaptive neighborhood radius The k-nearest neighbor index is obtained. A neighborhood point cloud.

[0044] S3-5. Construction The bounding box of a neighborhood point cloud, where the minimum X-axis coordinate of all neighborhood point clouds within the bounding box is... The maximum value of the X-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the X direction. The minimum value of the Y-axis coordinate of all neighboring point clouds within the bounding box is The maximum value of the Y-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Y direction. The minimum Z-axis coordinate of all neighboring point clouds within the bounding box is The maximum Z-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Z direction. , size vector Dimension vector Dimension vector Each part is divided into three equal parts to construct a 3×3×3 local voxel mesh. .

[0045] In one embodiment of the present invention, step S4 includes the following steps: S4-1. Perform local voxel mesh search using the breadth-first search (BFS) algorithm. Access was performed and the results were obtained. The final sampling point set consists of 10 final sampling points. , ,in For the first The coordinates of the final sampling points Breadth-first search (BFS) guarantees that the voxel closest to the starting point is visited first. This method uses face connectivity markers (based on 26 connectivity parameters, including face, edge, and corner connectivity) to indicate voxels adjacent to the center voxel, and limits the search depth to two layers to control the size of the sampling region. Starting from the voxel containing the center point, the algorithm searches layer by layer along the connectivity direction: when an empty voxel (i.e., a voxel without a point) is detected, further expansion in that direction is terminated, ensuring search efficiency while maintaining the validity of the sampling points.

[0046] S4-2. Through formula Calculate the first final sampling points Neighbor mean points .

[0047] S4-3. Through formula Calculate the offset vector .

[0048] S4-4. Through formula The covariance matrix with three rows and three columns was calculated. In the formula, This is a transpose.

[0049] S4-5. On the covariance matrix Eigenvalue decomposition yields three non-negative eigenvalues. , , , Through formula The curvature was calculated The obtained curvature can reflect the bending characteristics of the local surface.

[0050] In step S5, the formula is used. Calculate the first Point cloud vertex normal vector In the formula, This is a modulo operation.

[0051] In one embodiment of the present invention, the incoming flow parameters in step S7 include: Mach number. Angle of attack and sideslip angle , The value ranges from 1.6 to 12.0. The value ranges from -20.0 to 20.0. The value ranges from 0 to 15.0.

[0052] In one embodiment of the present invention, step S9 includes the following steps: S9-1. Obtaining Mach numbers using the active subspace method. For lift coefficient and drag coefficient sensitivity Angle of attack For lift coefficient and drag coefficient sensitivity Sideslip angle For lift coefficient and drag coefficient sensitivity , .

[0053] S9-2. When Mach number When the value is between 1.6 and 2.0, it depends on the sensitivity. Mach number Using dense sampling with a step size of 0.2, when the Mach number... When the value is between 2.0 and 12.0, it depends on the sensitivity. Mach number A segmented, equal-step dense sampling method with a sampling step size of 0.5 and an endpoint not exceeding 12.0 was used to obtain a Mach number sampling point set. Based on sensitivity... angle of attack By employing dense sampling with a step size of 0.5, a set of angle-of-attack sampling points is obtained. Based on sensitivity... Side slip angle Sparse sampling with a step size of 3.0 is used to obtain the sideslip angle sampling point set, reducing redundant calculations.

[0054] S9-3. Based on the Mach number sampling point set, angle of attack sampling point set, and sideslip angle sampling point set, use Flunet software to solve the CFD and generate high-precision (NS equation) and low-precision (inviscid Euler equation) flow field data. Use the high-precision and low-precision flow field data as the target domain dataset.

[0055] In one embodiment of the present invention, step S10 includes the following steps: S10-1. Generate surface unstructured meshes from the target domain dataset using pointwise software.

[0056] S10-2. Using three-point interpolation, the surface flow field is mapped onto the unstructured surface mesh to obtain physical quantities. .

[0057] In one embodiment of the present invention, step S12 includes the following steps: S12-1. Through formula Calculate the first Point cloud Angle of attack of rotation around the Y-axis Rotated coordinates In the formula, For the first Point cloud Coordinates in the flow field body coordinate system.

[0058] S12-2. Through formula Calculate the first Point cloud Sideslip angle of rotation about the Y-axis Rotated coordinates .

[0059] S12-3. Through formula Calculate the vertex normal vector Angle of attack rotating around the Y-axis Rotated coordinates .

[0060] S12-4. Through formula Calculate the vertex normal vector Sideslip angle around Z-axis The rotated normal vector By employing the aforementioned dual-rotation structure, rapid simulation of changes in angle of attack and sideslip tilt angle can be achieved without modifying the original point cloud topology. The complete neural network training dataset is then established through the steps outlined above.

[0061] In one embodiment of the present invention, step S13 includes the following steps: The S13-1.PointCurvNeXt neural network model consists of the PointNeXt network, the incoming flow parameter encoding module env_encoder, and the regression output module output_head.

[0062] S13-2. Rotate the coordinates... The rotated normal vector curvature The input features are fed into the PointNeXt network's input feature projection module of the PointCurvNeXt neural network model, and the output features are obtained. .

[0063] The incoming parameter encoding module env_encoder of the S13-3.PointCurvNeXt neural network model consists of a first linear mapping layer, a ReLU activation function, and a second linear mapping layer, which encodes the Mach number. The input is given to the incoming stream parameter encoding module env_encoder, and the output is the feature. .

[0064] S13-4. Features With features Perform a concatenation operation to obtain the fused feature representation. .

[0065] S13-5. Representing the fused features The input is fed into the encoder-decoder of the PointNeXt neural network model, and the output is the feature. .

[0066] The S13-6.PointCurvNeXt neural network model's regression output module, output_head, consists of a first ReLU activation function, a BN layer, a non-linear activation layer, and a second ReLU activation function, sequentially processing the features. The input is fed into the regression output module output_head, and the output is the physical quantities of the flow field on the aircraft surface. .

[0067] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rapid prediction method for digital twins of flow fields on aircraft surfaces, characterized in that, include: S1. Generate an STL-formatted geometric model of the canard missile vehicle in SpaceClaim software, and obtain the set of face normal vectors for the triangular facets of the vehicle from the geometric model. and the cloud of points on the surface of the aircraft , , , For the first The face normal vectors of a triangular facet. , The number of triangular pieces, For the first A point cloud, This represents the number of point clouds; S2. Based on point cloud set Calculate neighborhood radius ; S3. Utilizing neighborhood radius Computing local voxel mesh ; S4. Based on the local voxel mesh The curvature was calculated ; S5. Calculate the first... using the face normal vectors of the triangular facets. Point cloud vertex normal vector ; S6. Use ICEM CFD software to construct a mesh model based on the geometry of the canard missile vehicle; S7. Determine the range of variation of the incoming flow parameters; S8. Input the mesh model into Ansys Fluent simulation software to obtain the lift coefficient of the canard missile vehicle through simulation. and drag coefficient ; S9. Lift coefficient of canard missile vehicles using the active subspace method and drag coefficient The analysis is performed to obtain the sensitivity, and the sampling point results are obtained based on the sensitivity. The target domain dataset is then generated using the sampling point results. S10. Obtain physical quantities using the target domain dataset. ; S11. For physical quantities Normalization is performed to obtain the physical quantity. Preserve normalized parameters ; S12. Calculate the rotated coordinates using the incoming flow parameters. and normal vector ; S13. Construct the PointCurvNeXt neural network model, and rotate the coordinates... Normal vector curvature The input is fed into the PointCurvNeXt neural network model, and the output yields the physical quantities of the flow field on the aircraft surface. ; S14. The physical quantities of the flow field on the aircraft surface will be obtained. Use normalization parameters Perform inverse normalization to restore the physical quantity.

2. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 1, characterized in that, Step S2 includes the following steps: S2-1. Constructing a point cloud The bounding box, where the minimum X-axis coordinate of all point clouds within the bounding box is... The maximum value of the X-axis coordinate of all point clouds in the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the X direction. The minimum value of the Y-axis coordinate of all point clouds in the bounding box is The maximum value of the Y-axis coordinate of all point clouds in the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Y direction. The minimum Z-axis coordinate of all point clouds in the bounding box is The maximum Z-axis coordinate of all point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Z direction. ; S2-2. Through formula The bounding box volume of the point cloud was calculated. ; S2-3. Using the formula The equivalent total surface area was calculated. ; S2-4. Using the formula Calculate the neighborhood radius In the formula The number of target points in the neighborhood. .

3. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 1, characterized in that, Step S3 includes the following steps: S3-1. Regarding the first Point cloud Perform neighborhood radius based The k-nearest neighbor index is obtained. A neighborhood point cloud; S3-2. Using the formula Calculate the first Point cloud Local density ; S3-3. Through formula The adaptive neighborhood radius is calculated. In the formula, For minimum density, Maximum density; S3-4. Regarding the first Point cloud Perform adaptive neighborhood radius The k-nearest neighbor index is obtained. A neighborhood point cloud; S3-5. Construction The bounding box of a neighborhood point cloud, where the minimum X-axis coordinate of all neighborhood point clouds within the bounding box is... The maximum value of the X-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the X direction. The minimum value of the Y-axis coordinate of all neighboring point clouds within the bounding box is The maximum value of the Y-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Y direction. The minimum Z-axis coordinate of all neighboring point clouds within the bounding box is The maximum Z-axis coordinate of all neighboring point clouds within the bounding box is ,Will and The difference is denoted as the size vector of the bounding box in the Z direction. , size vector Dimension vector Dimension vector Each part is divided into three equal parts to construct a 3×3×3 local voxel mesh. .

4. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 1, characterized in that, Step S4 includes the following steps: S4-1. Perform local voxel meshes using the breadth-first search algorithm. Access was performed and the results were obtained. The final sampling point set consists of 10 final sampling points. , ,in For the first The coordinates of the final sampling points ; S4-2. Through formula Calculate the first final sampling points Neighbor mean points ; S4-3. Through formula Calculate the offset vector ; S4-4. Through formula The covariance matrix with three rows and three columns was calculated. In the formula, For transpose; S4-5. On the covariance matrix Eigenvalue decomposition yields three non-negative eigenvalues. , , , Through formula The curvature was calculated .

5. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 1, characterized in that: In step S5, the formula is used. Calculate the first Point cloud vertex normal vector In the formula, This is a modulo operation.

6. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 1, characterized in that: The incoming flow parameters in step S7 include: Mach number. Angle of attack and sideslip angle , The value ranges from 1.6 to 12.

0. The value ranges from -20.0 to 20.

0. The value ranges from 0 to 15.

0.

7. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 5, characterized in that, Step S9 includes the following steps: S9-1. Obtaining Mach numbers using the active subspace method. For lift coefficient and drag coefficient sensitivity Angle of attack For lift coefficient and drag coefficient sensitivity Sideslip angle For lift coefficient and drag coefficient sensitivity , ; S9-2. When Mach number When the value is between 1.6 and 2.0, it depends on the sensitivity. Mach number Using dense sampling with a step size of 0.2, when the Mach number... When the value is between 2.0 and 12.0, it depends on the sensitivity. Mach number A segmented, equal-step dense sampling method with a sampling step size of 0.5 and an endpoint not exceeding 12.0 was used to obtain a Mach number sampling point set. Based on sensitivity... angle of attack By employing dense sampling with a step size of 0.5, a set of angle-of-attack sampling points is obtained. Based on sensitivity... Side slip angle A sparse sampling step of 3.0 is used to obtain the sideslip angle sampling point set; S9-3. Based on the Mach number sampling point set, angle of attack sampling point set, and sideslip angle sampling point set, use Flunet software to solve the CFD and generate high-precision and low-precision flow field data. Use the high-precision and low-precision flow field data as the target domain dataset.

8. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 5, characterized in that, Step S10 includes the following steps: S10-1. Generate surface unstructured meshes for the target domain dataset using pointwise software; S10-2. Using three-point interpolation, the surface flow field is mapped onto the unstructured surface mesh to obtain physical quantities. .

9. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 6, characterized in that, Step S12 includes the following steps: S12-1. Through formula Calculate the first Point cloud Angle of attack of rotation around the Y-axis Rotated coordinates In the formula, For the first Point cloud Coordinates in the flow field body coordinate system; S12-2. Through formula Calculate the first Point cloud Sideslip angle of rotation about the Y-axis Rotated coordinates ; S12-3. Through formula Calculate the vertex normal vector Angle of attack rotating around the Y-axis Rotated coordinates ; S12-4. Through formula Calculate the vertex normal vector Sideslip angle around Z-axis The rotated normal vector .

10. The rapid prediction method for digital twins of flow fields on aircraft surfaces according to claim 8, characterized in that, Step S13 includes the following steps: The S13-1.PointCurvNeXt neural network model consists of a PointNeXt network, an inflow parameter encoding module env_encoder, and a regression output module output_head; S13-2. Rotate the coordinates... The rotated normal vector curvature The input features are fed into the PointNeXt network's input feature projection module of the PointCurvNeXt neural network model, and the output features are obtained. ; The incoming parameter encoding module env_encoder of the S13-3.PointCurvNeXt neural network model consists of a first linear mapping layer, a ReLU activation function, and a second linear mapping layer, which encodes the Mach number. The input is given to the incoming stream parameter encoding module env_encoder, and the output is the feature. ; S13-4. Features With features Perform a concatenation operation to obtain the fused feature representation. ; S13-5. Representing the fused features The input is fed into the encoder-decoder of the PointNeXt neural network model, and the output is the feature. ; The S13-6.PointCurvNeXt neural network model's regression output module, output_head, consists of a first ReLU activation function, a BN layer, a non-linear activation layer, and a second ReLU activation function, sequentially processing the features. The input is fed into the regression output module output_head, and the output is the physical quantities of the flow field on the aircraft surface. .

Images