Automobile aerodynamic flow field prediction method, device and system

Abstract

The application discloses a kind of automobile aerodynamic flow field prediction method, device and system, it is related to automobile aerodynamics technical field.The method includes: obtaining automobile geometric model;Extract global macroscopic point cloud and local high-curvature microscopic point cloud;Based on the sign distance function of query point and surface curvature generates physical bias mask;Double-scale point cloud features are input into pre-trained double-scale transformer model, the model includes macroscopic encoder, microscopic encoder and cross-scale cross attention layer, and the physical priori constraint is injected by superimposing the physical bias mask in attention calculation, and the velocity field, pressure field and aerodynamic force coefficient are output by decoder.The application can quickly output high-precision flow field prediction results in line with physical laws only by geometric data.

Description

Technical Field

[0001] This invention relates to the field of automotive aerodynamics, and in particular to a method, apparatus and system for predicting automotive aerodynamic flow fields. Background Technology

[0002] In the automotive R&D process, optimizing aerodynamic characteristics is crucial for reducing energy consumption and improving driving stability. Traditional computational fluid dynamics (CFD) simulations typically involve geometry cleaning, mesh generation, and numerical solution of the Navier-Stokes equations. For each new automotive styling design, this process usually requires more than 72 hours of computing cluster resources, severely limiting the efficiency of R&D iteration.

[0003] Data-driven surrogate models, represented by deep learning, have been widely used in flow field prediction. However, existing methods based on convolutional neural networks (CNNs) suffer from high curvature detail loss due to voxelization; methods based on graph neural networks (GNNs) struggle to efficiently establish long-range flow relationships; and methods based on standard transformers suffer from issues such as attention weights lacking fluid physics information, uneven resource allocation across multi-scale regions, and a lack of fluid conservation constraints, resulting in insufficient analytical accuracy of the near-wall boundary layer and a tendency to produce non-physical prediction results.

[0004] Therefore, there is an urgent need for a method to predict aerodynamic characteristics that can both respond quickly and reproduce the physical laws of the flow field with high fidelity. Summary of the Invention

[0005] The purpose of this invention is to provide a method, device and system for predicting automotive aerodynamic flow fields. By extracting macroscopic and microscopic point clouds and using a dual-scale transformer model with a physical bias mask generated based on the sign distance function and surface curvature, flow field prediction is performed. This allows the model to output high-precision velocity field, pressure field and aerodynamic coefficients that conform to the laws of fluid physics in a very short time, requiring only geometric data as input.

[0006] In a first aspect, the present invention provides a method for predicting the aerodynamic flow field of an automobile, comprising: acquiring geometric model data of the automobile to be predicted; extracting automobile point cloud data from the geometric model data; wherein the automobile point cloud data includes a macroscopic point cloud representing the global contour of the automobile and a microscopic point cloud representing a local region, wherein the surface curvature of the local region is higher than a preset discrete average curvature threshold; calculating the signed distance function value from each query point to the automobile surface and the corresponding surface curvature based on query points in a preset flow field space, and generating a physical bias mask based on the signed distance function value and the surface curvature; inputting the features of the automobile point cloud data into a pre-trained dual-scale transformer model, and outputting the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point based on the target query point coordinates through the decoder of the dual-scale transformer model; wherein the dual-scale transformer model includes a macroscopic encoder, a microscopic encoder, and a cross-scale attention layer connecting the macroscopic encoder and the microscopic encoder, and superimposing a physical bias mask on the attention weight matrix during the attention calculation process to inject physical prior constraints.

[0007] In some preferred embodiments of the present invention, extracting vehicle point cloud data from geometric model data includes: performing Poisson disk sampling on the geometric model data to obtain a macro point cloud with a first preset number of nodes, and calculating the surface normal vector of each macro point in the macro point cloud; calculating the discrete average curvature of the geometric model surface, and retaining nodes in regions where the discrete average curvature is greater than a preset discrete average curvature threshold as micro point cloud, the micro point cloud including multiple patches, each patch containing nodes within a second preset number of nodes.

[0008] In some preferred embodiments of the present invention, generating a physical bias mask based on the symbolic distance function value and surface curvature includes: performing a nonlinear mapping on the symbolic distance function value based on a preset learnable multilayer perceptron to obtain a first bias component; wherein the learnable multilayer perceptron includes a preset number of hidden layers, each hidden layer has a preset dimension, and the initialized weights satisfy the zero mean of the output; determining the product of the surface curvature of the corresponding query point and a preset learnable scalar parameter as a second bias component; wherein the learnable scalar parameter has a preset initial value and is adaptively adjusted through backpropagation; and determining the sum of the first bias component and the second bias component as the physical bias mask.

[0009] In some preferred embodiments of the present invention, the dual-scale transformer model includes a macro encoder and a micro encoder connected by a cross-scale cross-attention layer; the macro encoder consists of multiple layers of standard self-attention transformer blocks for global self-attention encoding of macro point cloud features; the micro encoder consists of multiple layers of transformer blocks with introduced physical bias masks; the method further includes: in the micro encoder, injecting the physical bias mask into the attention weight matrix of the micro point cloud features; in the cross-scale cross-attention layer, performing a cross-attention operation using the micro features output by the micro encoder as a query vector and the macro features output by the macro encoder as a key vector and a value vector, so that the local features of the local region obtain background information from the macro global features.

[0010] In some preferred embodiments of the present invention, the decoder of the dual-scale transformer model outputs the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point based on the target query point coordinates, including: applying position encoding to the target query point coordinates, using the encoded position vector as the query vector, using the multi-scale latent features output by the dual-scale transformer model as the key vector and value vector, and outputting the three-dimensional velocity components and pressure values ​​at the target query point through a multi-layer cross-attention decoding network; performing area-weighted integration on the pressure values ​​predicted at each query point on the vehicle surface, and calculating the global aerodynamic coefficient in combination with the incoming flow direction; or, performing global average pooling on the multi-scale latent features and inputting them into a fully connected network to directly output the global aerodynamic coefficient.

[0011] In some preferred embodiments of the present invention, the method further includes: acquiring multiple vehicle geometric samples and corresponding computational fluid dynamics simulation ground truth data; wherein, the computational fluid dynamics simulation ground truth data includes volumetric velocity field, surface pressure field, and aerodynamic coefficients; extracting macroscopic point cloud and microscopic point cloud for each vehicle geometric sample, and calculating the signed distance function and surface curvature to generate a physical bias mask; inputting the features of the macroscopic point cloud, the features of the microscopic point cloud, and the physical bias mask into the dual-scale transformer model to be trained to obtain the predicted velocity field, pressure field, and aerodynamic coefficients; constructing a joint loss function, and updating the model parameters of the dual-scale transformer model to be trained according to the joint loss function until the joint loss function converges; wherein, the joint loss function includes a volumetric velocity regression loss term, a surface pressure regression loss term, an aerodynamic coefficient regression loss term, and a physical conservation penalty term based on the predicted velocity field divergence, wherein the preset weight coefficient of the surface pressure regression loss term is greater than the preset weight coefficient of the volumetric velocity regression loss term, and the preset weight coefficient of the aerodynamic coefficient regression loss term is greater than the preset weight coefficient of the surface pressure regression loss term.

[0012] In some preferred embodiments of the present invention, the physical conservation penalty term is calculated based on the divergence of the incompressible continuity equation and is expressed as the square mean of the sum of the partial derivatives of the predicted velocity components with respect to spatial coordinates; the physical conservation penalty term has a dynamic weight coefficient in the joint loss function, which takes a preset initial weight value in the initial stage of training, gradually increases with the training process, and maintains a preset maximum weight value in the later stage of training.

[0013] In some preferred embodiments of the present invention, the dynamic weight coefficient maintains a preset initial weight value in the first preset iteration ratio stage of the training process, linearly increases from the preset initial weight value to a preset maximum weight value in the subsequent second preset iteration ratio stage, and maintains the preset maximum weight value in the remaining iteration ratio stage.

[0014] Secondly, the present invention provides an automotive aerodynamic flow field prediction device, comprising: a data acquisition module for acquiring geometric model data of the vehicle to be predicted; a point cloud acquisition module for extracting vehicle point cloud data from the geometric model data; wherein the vehicle point cloud data includes a macroscopic point cloud representing the global contour of the vehicle and a microscopic point cloud representing a local region, the surface curvature of the local region being higher than a preset discrete average curvature threshold; a data processing module for calculating the symbolic distance function value from each query point to the vehicle surface and the corresponding surface curvature based on query points within a preset flow field space, and generating a physical bias mask based on the symbolic distance function value and the surface curvature; and a flow field prediction module for inputting the features of the vehicle point cloud data into a pre-trained dual-scale transformer model, and outputting the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point based on the target query point coordinates through the decoder of the dual-scale transformer model; wherein the dual-scale transformer model includes a macroscopic encoder, a microscopic encoder, and a cross-scale attention layer connecting the macroscopic encoder and the microscopic encoder, and a physical bias mask is superimposed on the attention weight matrix during the attention calculation process to inject physical prior constraints.

[0015] Thirdly, the present invention provides an automotive aerodynamic flow field prediction system, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the method provided in the first aspect above.

[0016] This invention brings the following beneficial effects: This invention provides a method, apparatus, and system for predicting automotive aerodynamic flow fields. The method includes: acquiring geometric model data of the vehicle to be predicted; extracting vehicle point cloud data from the geometric model data; wherein the vehicle point cloud data includes macroscopic point clouds representing the global contour of the vehicle and microscopic point clouds representing local regions, the surface curvature of the local regions being higher than a preset discrete average curvature threshold; calculating the signed distance function value from each query point to the vehicle surface and the corresponding surface curvature based on query points within a preset flow field space, and generating a physical bias mask based on the signed distance function value and surface curvature; inputting the features of the vehicle point cloud data into a pre-trained dual-scale transformer model, and then processing the data through the decoder of the dual-scale transformer model. Based on the coordinates of the target query point, the system outputs the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficients at the target query point. The dual-scale converter model includes a macro encoder, a micro encoder, and a cross-scale attention layer connecting the macro encoder and the micro encoder. During the attention calculation process, a physical bias mask is superimposed on the attention weight matrix to inject physical prior constraints. By extracting macro and micro dual-scale point clouds and using the dual-scale converter model superimposed with a physical bias mask generated based on the signed distance function and surface curvature for flow field prediction, the model can output high-precision velocity field, pressure field, and aerodynamic coefficients that conform to the laws of fluid physics in a very short time, requiring only geometric data as input. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A system overall architecture data flow diagram provided in this embodiment of the invention; Figure 2 A flowchart of a method for predicting automotive aerodynamic flow field provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the computation of a physical information-guided attention mechanism provided in an embodiment of the present invention; Figure 4 A diagram of a dual-channel feature fusion converter encoder architecture is provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the structure of an automotive aerodynamic flow field prediction device provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an automotive aerodynamic flow field prediction system provided in an embodiment of the present invention.

[0019] Icons: 310 - Data acquisition module; 320 - Point cloud acquisition module; 330 - Data processing module; 340 - Flow field prediction module; 400 - Memory; 401 - Processor; 402 - Bus; 403 - Communication interface. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] Furthermore, terms such as "horizontal," "vertical," and "sag" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0025] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] This invention provides a method for predicting automotive aerodynamic flow fields, which is applied to automotive aerodynamic flow field prediction systems.

[0028] For details, see Figure 1 The illustrated embodiment of the present invention provides a system architecture data flow diagram, which shows the complete data flow path from the input of the vehicle geometry model to the final output of the flow field and aerodynamic coefficients. The entire system includes core processing stages such as data acquisition, dual-scale point cloud extraction, physical bias mask calculation, and dual-scale transformer model encoding and decoding. This method can be executed on electronic devices such as high-performance personal computers, graphics workstations, cloud servers, or server clusters. The following detailed description is provided in conjunction with specific steps.

[0029] See Figure 2 The flowchart shown in this embodiment of the invention provides a method for predicting automotive aerodynamic flow fields, the method comprising: Step S102: Obtain the geometric model data of the car to be predicted.

[0030] Specifically, the geometric model data refers to three-dimensional surface geometry data describing the external contour surface of a car body. This data can be in STL, OBJ, STEP, or other common CAD file formats. These formats can accurately express the car's external geometry in the form of triangular patches or parametric surfaces. After acquiring the geometric model data, optional steps include preliminary geometric cleanup, such as removing duplicate faces, repairing non-manifold edges, and filling small holes, to ensure the stability and accuracy of subsequent processing. Alternatively, the geometric model data can also be directly derived from the three-dimensional surface description parameters generated in real-time by parametric modeling software, without the need for a file export and import process.

[0031] Step S104: Extract vehicle point cloud data from geometric model data; wherein the vehicle point cloud data includes macro point cloud representing the global contour of the vehicle and micro point cloud representing local regions, and the surface curvature of the local regions is higher than a preset discrete average curvature threshold.

[0032] Specifically, the contribution of most flat areas of the vehicle body to the overall aerodynamic force is relatively uniform and gradual, and can be summarized using a sparse macroscopic skeleton point cloud. However, areas with abrupt curvature changes (such as the base of the rearview mirror, the junction of the A-pillar and fender, and the edge of the air intake grille) are key areas for flow separation and vortex generation, requiring the retention of a high-density microscopic point cloud to capture subtle flow field changes. The division between the "macroscopic point cloud" and the "microscopic point cloud" is based on whether the surface curvature of the local area is higher than a preset discrete average curvature threshold. The selection of this threshold directly determines the boundary between the two scales, and its specific value can be preset according to the vehicle type and geometric complexity or determined through parameter optimization.

[0033] The specific extraction process includes: First, sampling the geometric model data to obtain a macroscopic point cloud with a first preset number of nodes, and calculating the surface normal vector of each macroscopic point in the macroscopic point cloud. The sampling method should ensure that the macroscopic nodes are evenly distributed on the vehicle surface to fully capture the contour features of the car. Second, calculating the discrete mean curvature of the geometric model surface, and retaining the nodes in the region where the discrete mean curvature is greater than the preset discrete mean curvature threshold as the microscopic point cloud. The microscopic point cloud includes multiple patches, and each patch contains a certain number of nodes within a certain range. The curvature calculation can be implemented using discrete differential geometry operators. Finally, after obtaining the coordinates and geometric attributes of the macroscopic point cloud and the microscopic point cloud, position encoding is applied to the coordinate information to adapt to the input of the transformer model. The role of position encoding is to map low-dimensional coordinates to high-dimensional space to facilitate the transformer network's perception of spatial position. By performing dual-scale sampling and separation on the original geometric data, two sets of point cloud representations that take into account both global topology and local details are formed.

[0034] Step S106: Based on the query points in the preset flow field space, calculate the symbolic distance function value from each query point to the vehicle surface and the corresponding surface curvature, and generate a physical offset mask based on the symbolic distance function value and the surface curvature.

[0035] Specifically, the core step of generating physical prior constraints provided in this embodiment aims to explicitly encode the decisive influence of wall distance on boundary layer development in fluid mechanics into the model's attention mechanism. The generated physical bias mask will be used for the "additive injection" of attention weights in the subsequent transformer model.

[0036] See Figure 3The diagram illustrates a calculation of a physically-informed attention mechanism according to an embodiment of the present invention. The diagram shows the symbolic distance function value. Input-learnable multilayer perceptron Generate the first bias component, surface curvature With learnable scalar parameters Multiplying them generates a second bias component, and summing the two yields the physical bias mask. The complete process involves injecting the attention weight matrix in an additive manner.

[0037] First, a series of query points are generated within a pre-defined flow field space surrounding the vehicle. The size of this flow field space is typically several times the vehicle's dimensions to ensure the wake region is completely contained. The query point sampling strategy can combine uniform sampling with near-wall layered sampling to achieve higher resolution in critical areas. For each query point, the signed shortest distance to the vehicle's surface geometry is precisely calculated, yielding the signed distance function value (denoted as ). Simultaneously, for each query point, it is also necessary to determine its nearest projection point on the vehicle surface and obtain the surface curvature value at that projection point (denoted as ). ).

[0038] Secondly, a first bias component is obtained by nonlinearly mapping the symbolic distance function value based on a preset learnable multilayer perceptron; wherein, the learnable multilayer perceptron contains a preset number of hidden layers, each hidden layer has a preset dimension, and the initialized weights satisfy the zero-mean output. This learnable multilayer perceptron (denoted as...) This maps the scalar distance value to a scalar bias value that is compatible with the model's attention mechanism.

[0039] Next, the product of the surface curvature corresponding to the query point and a preset learnable scalar parameter is determined as the second bias component; wherein, the learnable scalar parameter has a preset initial value and is adaptively adjusted through backpropagation. This learnable scalar parameter (denoted as...) This allows the network to adaptively learn the importance weights of curvature information in flow field prediction.

[0040] Finally, the sum of the first bias component and the second bias component is determined as the physical bias mask, denoted as . Its mathematical expression is: ; The physical significance of this design lies in the fact that when the query point is far from the wall... Larger, through The first bias component after mapping approaches a negative constant, thus weakening the attention weights between far-field nodes after the subsequent Softmax operation, reflecting the physical fact that far-field flow is less directly affected by the wall; when the query point is located in a high-curvature near-wall region... , Larger, at this time Approaching a negative constant allows the attention mechanism to be fully activated in such critical regions, enhancing information interaction between near-wall nodes to resolve the strong nonlinear flow characteristics in the boundary layer and separation region.

[0041] Step S108: Input the features of the car point cloud data into the pre-trained dual-scale transformer model. Through the decoder of the dual-scale transformer model, output the three-dimensional velocity vector, pressure scalar and global aerodynamic coefficient at the target query point according to the coordinates of the target query point. The dual-scale transformer model includes a macro encoder, a micro encoder and a cross-scale attention layer connecting the macro encoder and the micro encoder. During the attention calculation process, a physical bias mask is superimposed on the attention weight matrix to inject physical prior constraints.

[0042] For details, see Figure 4 The diagram shown illustrates a dual-channel feature fusion transformer encoder architecture according to an embodiment of the present invention. The macro encoder consists of multiple layers of standard self-attention transformer blocks, used for global self-attention encoding of the features of the macro point cloud. Each transformer block includes a multi-head self-attention layer and a feedforward neural network layer, supplemented by residual connections and layer normalization. The macro encoder receives the macro point cloud feature sequence as input, establishes long-range dependencies between global macro nodes through a self-attention mechanism, enabling pressure changes at the front of the vehicle body to be correlated with the tail wake structure, and finally outputs the encoded macro feature sequence.

[0043] The microencoder consists of multiple layers of transformer blocks that incorporate the physical bias mask. In the self-attention calculation of each layer, the formula for calculating the attention weights is modified as follows: ; Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k Let be the dimension of the key vector. This is a physical bias mask. Through this additive injection method, wall effects and curvature effects in fluid dynamics are explicitly encoded into the receptive fields between microscopic features.

[0044] To achieve the fusion of macroscopic global information and microscopic local information, the dual-scale transformer model also includes a cross-scale attention layer. This layer employs an asymmetric cross-attention mechanism, using the microscopic features output by the micro-encoder as the query vector and the macroscopic features output by the macro-encoder as the key and value vectors to perform the cross-attention operation. This means that each microscopic node representing a high-curvature local region has the opportunity to retrieve and aggregate information from all macroscopic nodes representing the global skeleton, thereby obtaining the macroscopic background information of its local region within the entire flow field. After processing by the macro-encoder, micro-encoder, and the cross-scale attention layer, a multi-scale latent feature representation that integrates multi-scale information is finally output.

[0045] After obtaining the multi-scale latent features, the decoder of the dual-scale transformer model outputs the prediction result based on the coordinates of the target query point. The coordinates of the target query point can be any spatial point specified by the user. The specific prediction process includes: applying position encoding to the coordinates of the target query point, using the encoded position vector as the query vector, and using the multi-scale latent features as the key and value vectors, outputting the three-dimensional velocity components and pressure values ​​at the target query point through a multi-layer cross-attention decoding network. Global aerodynamic coefficients can be obtained through physical integration or aggregation multilayer perceptron methods.

[0046] Furthermore, in some preferred embodiments of the present invention, extracting vehicle point cloud data from geometric model data includes: performing Poisson disk sampling on the geometric model data to obtain a macro point cloud with a first preset number of nodes, and calculating the surface normal vector of each macro point in the macro point cloud; calculating the discrete average curvature of the geometric model surface, and retaining nodes in regions where the discrete average curvature is greater than a preset discrete average curvature threshold as micro point cloud, the micro point cloud including multiple patches, each patch containing nodes within a second preset number of nodes.

[0047] Specifically, Poisson disk sampling ensures a uniform distance between sampling points, preventing local over-density or under-density of the point cloud. As a specific configuration, the first preset number of nodes can be set to approximately 5000 nodes; for commercial trucks with relatively simple geometry, this number can be reduced to 3000 nodes; for supercars with complex surface transitions, this number can be increased to 8000 nodes. The surface normal vector can be obtained by averaging the area of ​​the normal vectors of the triangular facet containing the node, or estimated using principal component analysis based on the local neighborhood point cloud.

[0048] Discrete mean curvature is calculated on the original surface mesh or high-density sampled point cloud using the discretized Laplace-Beltrami operator, yielding the discrete mean curvature at each vertex location. The preset discrete mean curvature threshold Values: For sedans and sport utility vehicles (SUVs). 0.15m can be taken. -1 For commercial trucks, 0.08m can be taken. -1 For small aircraft such as drones, 0.20m can be taken. -1 Will satisfy The nodes are divided into multiple patches based on spatial connectivity, and each patch contains between 1,000 and 3,000 of the second preset nodes. In general, the total number of micro-nodes in a passenger car is typically between 10,000 and 20,000.

[0049] This invention enables efficient separation of the vehicle body surface geometry into a macroscopic skeleton and microscopic high-curvature patches. Poisson disk sampling ensures the uniformity of macroscopic points and contour capture capability; patch segmentation based on discrete average curvature thresholds ensures high-density preservation of key flow separation areas such as rearview mirrors and A-pillar corners. This dual-scale representation allows the model to simultaneously consider global aerodynamic topology and local fine flow characteristics, avoiding the loss of high-curvature details caused by uniform sampling, thereby significantly improving the fidelity of flow field prediction.

[0050] Furthermore, in some preferred embodiments of the present invention, generating a physical bias mask based on the symbolic distance function value and surface curvature includes: performing a nonlinear mapping on the symbolic distance function value based on a preset learnable multilayer perceptron to obtain a first bias component; wherein the learnable multilayer perceptron includes a preset number of hidden layers, each hidden layer has a preset dimension, and the initialized weights satisfy the zero-mean output; determining the product of the surface curvature of the corresponding query point and a preset learnable scalar parameter as a second bias component; wherein the learnable scalar parameter has a preset initial value and is adaptively adjusted through backpropagation; and determining the sum of the first bias component and the second bias component as the physical bias mask.

[0051] Specifically, learnable multilayer perceptrons The preset number of hidden layers is 2, with each hidden layer having a dimension of 64. The activation function can be ReLU or GELU. Weight initialization uses Xavier or Kaiming initialization, and the bias term is adjusted to ensure that the initial output expectation is zero, guaranteeing that the bias components have zero mean in the early stages of training and do not interfere with model convergence. The learnable scalar parameters... The preset initial value is taken During training, no additional regularization constraints are imposed; the gradient descent algorithm is used to automatically find the curvature penalty strength suitable for the current dataset.

[0052] This embodiment employs a learnable multilayer perceptron with two hidden layers and a dimension of 64 to map the symbolic distance function value. Combined with a learnable scalar curvature penalty term with an initial value of 0.01, it can inject the physical relationship between wall distance and local curvature into the attention mechanism in a concise, parameterized form while maintaining the model's expressive power. This structure enables the model to adaptively learn the correlation between boundary layer thickness and wall distance based on data, significantly improving the prediction accuracy of near-wall flow characteristics.

[0053] Furthermore, in some preferred embodiments of the present invention, the dual-scale transformer model includes a macro encoder and a micro encoder connected through a cross-scale cross-attention layer; the macro encoder consists of multiple layers of standard self-attention transformer blocks, used for global self-attention encoding of features of the macro point cloud; the micro encoder consists of multiple layers of transformer blocks with introduced physical bias masks; the method further includes: in the micro encoder, injecting the physical bias mask into the attention weight matrix of the micro point cloud features; in the cross-scale cross-attention layer, performing a cross-attention operation using the micro features output by the micro encoder as a query vector and the macro features output by the macro encoder as a key vector and a value vector, so that the local features of the local region obtain background information from the macro global features.

[0054] Specifically, both the macro-encoder and micro-encoder can be preset to 4 layers, but the number of layers does not need to be exactly the same and can be adjusted according to the task complexity. The number of attention heads is preset to 8, and the dimension of each head is the model input dimension divided by the number of heads. The hidden layer dimension of the feedforward neural network is preset to 1024. The injection of the physical bias mask occurs in the self-attention calculation of each micro-encoder layer to ensure that physical prior constraints can be carried out throughout the entire micro-feature extraction process. The cross-scale cross-attention operation can be performed once after a certain number of encoding layers, for example, cross-attention fusion is performed after every 2 encoding layers, so as to achieve multi-scale information interaction while maintaining computational efficiency.

[0055] Microscopic features (microscopic cloud feature sequence) ) as a Query Macroeconomic characteristics as Key and Value This allows locally high-curvature nodes to preferentially extract macroscopic background information about large-scale flows from the global topology; simultaneously, it utilizes macroscopic features (macroscopic point cloud feature sequences). ) as a Query Microscopic features serve as Key / Value pairs, enabling global skeleton nodes to perceive fine separation flow cues in high-curvature regions; the fused features are compressed through a linear projection layer. Multi-scale latent feature representation.

[0056] This invention, through the configuration of a 4-layer macroscopic encoder and a 4-layer microscopic encoder with physical bias masks, and employing an 8-head attention mechanism and a 1024-dimensional feedforward network, provides the model with sufficient capacity to learn the nonlinear mapping relationships of complex aerodynamic flow fields. In particular, the introduction of asymmetric cross-scale attention layers enables microscopic nodes in high-curvature local regions to explicitly extract large-scale flow background information from the global macroscopic skeleton. This allows for the acquisition of contextual constraints on the global flow topology while maintaining the ability to resolve local geometric details, effectively improving the prediction accuracy of separated and re-attached flow regions.

[0057] Furthermore, in some preferred embodiments of the present invention, the decoder of the dual-scale transformer model outputs the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point based on the target query point coordinates, including: applying position encoding to the target query point coordinates, using the encoded position vector as the query vector, using the multi-scale latent features output by the dual-scale transformer model as the key vector and value vector, and outputting the three-dimensional velocity components and pressure values ​​at the target query point through a multi-layer cross-attention decoding network; performing area-weighted integration on the pressure values ​​predicted at each query point on the vehicle surface, and calculating the global aerodynamic coefficient in combination with the incoming flow direction; or, performing global average pooling on the multi-scale latent features and inputting them into a fully connected network to directly output the global aerodynamic coefficient.

[0058] Specifically, the location encoding can employ Fourier feature location encoding, with a frequency count of L=8, resulting in a dimensionality of 48 for the encoded location features. In the joint task decoder and multi-constraint output, the inputs are multi-scale latent feature representations and the coordinates of the target query point to be predicted. (Can be any point in space, not limited to grid nodes during training). Output: A 3D velocity vector for each query point. With pressure scalar Global drag coefficient With lift coefficient .

[0059] The decoder employs a query-based cross-attention decoding mechanism. Microscopic field prediction task: The location is encoded as a query, represented as a key / value pair using latent features, and after decoding through two layers of cross-attention, it is output as each query point via an MLP (dimensions 512→256→4). . Global coefficient prediction task: Two strategies are available: Physical integration method (recommended): Predicted pressure values ​​for all surface query points. Perform area-weighted integration and calculate based on the direction of incoming flow. Aligned with the true integral result; Aggregated MLP approach: The global average pooling result of latent features is used as input, and the output is directly processed by an MLP (dimensions 512→256→2). .

[0060] This invention provides two complementary approaches to predicting global aerodynamic coefficients: the physical integration approach directly corresponds to the standard post-processing workflow of computational fluid dynamics, offering clear physical interpretability and facilitating engineer understanding and verification; the aggregated multilayer perceptron approach achieves complete end-to-end learning, avoiding the computational overhead of explicit integration and is suitable for applications with extremely high real-time requirements. The two approaches can be flexibly switched, balancing physical consistency of predictions with computational efficiency.

[0061] Furthermore, in some preferred embodiments of the present invention, the method further includes: acquiring multiple vehicle geometric samples and corresponding computational fluid dynamics simulation ground truth data; wherein, the computational fluid dynamics simulation ground truth data includes volumetric velocity field, surface pressure field, and aerodynamic coefficients; extracting macroscopic point cloud and microscopic point cloud for each vehicle geometric sample, and calculating the signed distance function and surface curvature to generate a physical bias mask; inputting the features of the macroscopic point cloud, the features of the microscopic point cloud, and the physical bias mask into the dual-scale transformer model to be trained to obtain the predicted velocity field, pressure field, and aerodynamic coefficients; constructing a joint loss function, and updating the model parameters of the dual-scale transformer model to be trained according to the joint loss function until the joint loss function converges; wherein, the joint loss function includes a volumetric velocity regression loss term, a surface pressure regression loss term, an aerodynamic coefficient regression loss term, and a physical conservation penalty term based on the predicted velocity field divergence, wherein the preset weight coefficient of the surface pressure regression loss term is greater than the preset weight coefficient of the volumetric velocity regression loss term, and the preset weight coefficient of the aerodynamic coefficient regression loss term is greater than the preset weight coefficient of the surface pressure regression loss term.

[0062] Specifically, a large-scale aerodynamic database was constructed. As a concrete example, 732 sets of car geometry samples were collected, covering multi-dimensional geometric variations such as different wheelbases, body lengths, roof heights, and front fascia chamfers. Computational fluid dynamics simulations were performed on each set of samples using a Reynolds-averaged Navier-Stokes (RANS) steady-state solver, and the turbulence model adopted... SST model, incoming flow velocity The corresponding Reynolds number The simulation output includes the true velocity field. Surface pressure field drag coefficient Lift coefficient Each CFD simulation (including mesh generation) took approximately 12 hours, and a total of 64-core computing clusters were used for parallel processing. The total data generation time for 500 sets was approximately 4 days (with 64 nodes operating at full capacity in parallel). The dataset was divided proportionally into 586 training sets, 73 validation sets, and 73 test sets.

[0063] Then, the same feature extraction steps as in the inference phase are performed on each car geometry sample to obtain the predicted velocity field. and pressure field and aerodynamic coefficient , .

[0064] The joint loss function is constructed, and its mathematical expression is as follows: ; in, This is the volumetric velocity regression loss term; This is the surface pressure regression loss term; This is the regression loss term for aerodynamic coefficients; This is a physical conservation penalty term based on the predicted velocity field divergence. The weights are dynamic and follow a phased adjustment strategy: the first 20% of training steps are set to... Prioritize reducing regression loss in the data; linearly increase weights during 20%–80% of training steps. Maintain 20% of your pace after training. .

[0065] Recommended hyperparameter configuration: , (Surface pressure has a higher weight). (Directly monitor aerodynamic coefficients) , .

[0066] Each forward propagation randomly samples the spatial flow field. These points are used as the query coordinates for the current batch. The AdamW optimizer is employed. ), initial learning rate Using a cosine annealing strategy to decay to The batch size was set to 4 groups of car samples, and the training was conducted for a total of 200 rounds.

[0067] See Table 1 for a summary table of model hyperparameter configurations provided by an embodiment of the present invention: Table 1

[0068] This training method jointly optimizes volumetric velocity regression, surface pressure regression, aerodynamic coefficient regression, and a divergence-based physical conservation penalty term, and employs... The hierarchical weighting configuration allows the model to prioritize the aerodynamic coefficient, the most critical macroscopic indicator, during training, while also considering the detailed reproduction of surface pressure distribution and spatial flow field. A physical conservation penalty term is incorporated as a soft constraint, guiding the model output to satisfy the incompressible continuity equation and effectively suppressing non-physical predictions near the separation zone. Experiments on 732 sets of passenger car data show that this joint training strategy reduces the relative error of the drag coefficient from 8.1% in the baseline model to 2.3%.

[0069] Furthermore, in some preferred embodiments of the present invention, the physical conservation penalty term is calculated based on the divergence of the incompressible continuity equation and is expressed as the square mean of the sum of the partial derivatives of the predicted velocity components with respect to spatial coordinates; the physical conservation penalty term has a dynamic weight coefficient in the joint loss function, which takes a preset initial weight value in the initial stage of training, gradually increases with the training process, and maintains a preset maximum weight value in the later stage of training.

[0070] Specifically, since the model's predictions are not accurate in the early stages of training, imposing excessively strong physical constraints at this time can lead to the optimization process getting stuck in local minima or slow convergence. Therefore, a dynamic weighting strategy is adopted: in the initial stage of training, a small preset initial weight value is used to allow the model to preferentially fit the true data values; as training progresses, the physical constraint weights are gradually increased to guide the model output to satisfy the incompressible continuity equation. Partial derivatives are accurately obtained from the network output with respect to the coordinates of the input query point using the automatic differentiation function, without the need for finite difference approximation. The dynamic weighting strategy effectively balances the contributions of data regression loss and physical regularization loss at different training stages. In the early stages of training, data-driven approaches are the primary focus to ensure that the model converges quickly to near the true value; in the later stages of training, physical constraints are gradually increased to guide the model to satisfy the fluid conservation law while maintaining prediction accuracy. This strategy avoids the convergence difficulties or underfitting of data caused by excessively strong physical constraints under fixed weights.

[0071] Furthermore, in some preferred embodiments of the present invention, the dynamic weight coefficient maintains a preset initial weight value in the first preset iteration ratio stage of the training process, linearly increases from the preset initial weight value to a preset maximum weight value in the subsequent second preset iteration ratio stage, and maintains the preset maximum weight value in the remaining iteration ratio stage.

[0072] Specifically, this embodiment provides a specific numerical adjustment scheme for the dynamic weight coefficient. As a preferred configuration, the first preset iteration ratio is 20%, the second preset iteration ratio is 60% (i.e., the stage from 20% to 80%), and the remaining iteration ratio is 20%. The preset initial weight value... Preset maximum weight value That is, in the first 20% of training rounds, In 20% to 80% of the rounds, linearly increase to In the last 20% of rounds, It remains unchanged.

[0073] This invention provides a simple and robust dynamic weighting implementation. The first 20% of rounds, with low weights, allow ample room for the data regression terms to decrease; the linear growth period from 20% to 80% smoothly introduces physical constraints, avoiding abrupt changes in the optimization trajectory; the final 20% of rounds, with high weights, ensures the model eventually converges to a solution that is both accurate and satisfies the incompressible continuity equation. Ablation experiments show that this strategy reduces the mean divergence in the wake region from... Reduce to This represents an increase of approximately two orders of magnitude.

[0074] After training, the model with fixed weights can be used for rapid prediction of new vehicle geometry. For a completely new CAD model, the system automatically performs the above steps. The pre-calculation of the signed distance function takes approximately 10 seconds, and the model forward inference takes approximately 8 seconds, for a total of approximately 18 seconds to output the complete 3D flow field, surface pressure contour map, and aerodynamic coefficients. Traditional high-fidelity CFD simulation (including mesh generation of approximately 20 hours and solution of approximately 52 hours) takes approximately 72 hours; this method speeds up the inference stage by approximately 32,400 times.

[0075] To verify the technical effectiveness of this invention, a quantitative evaluation was conducted on 73 independent test sets. Table 2 below shows a comparison of ablation experiment results under different model configurations. In the table, "baseline transformer" refers to a pure transformer model that does not employ dual-scale encoding, does not use a physical information-guided attention mechanism, and does not impose a physical conservation penalty term.

[0076] Table 2

[0077] The experimental results show that the complete solution of this invention reduces the drag coefficient. The mean absolute percentage error (MAPE) decreased from 8.1% at the baseline to 2.3%, and the root mean square error (RMSE / range) of the surface pressure field was 4.1%; mean divergence in the wake region decreased by approximately two orders of magnitude (from...). Down to This fully demonstrates the significant superiority of the method provided in the embodiments of the present invention in terms of prediction accuracy and physical consistency. The synergistic effect of the three technologies far exceeds the simple summation of any single effect.

[0078] This invention provides a method for predicting automotive aerodynamic flow fields, comprising: acquiring geometric model data of the vehicle to be predicted; extracting macroscopic point clouds representing the global contour of the vehicle and microscopic point clouds representing local regions from the geometric model data; wherein the surface curvature of the local regions is higher than a preset discrete average curvature threshold; calculating the signed distance function value from each query point to the vehicle surface and the corresponding surface curvature based on query points in a preset flow field space, and generating a physical bias mask based on the signed distance function value and the surface curvature; inputting the features of the macroscopic point cloud and the features of the microscopic point cloud into a pre-trained dual-scale transformer model, and using the decoder of the dual-scale transformer model to predict the target query point... The system outputs the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficients at the target query point. The dual-scale converter model includes a macro encoder, a micro encoder, and a cross-scale attention layer connecting the macro encoder and the micro encoder. During the attention calculation process, a physical bias mask is superimposed on the attention weight matrix to inject physical prior constraints. By extracting macro and micro dual-scale point clouds and using the dual-scale converter model superimposed with a physical bias mask generated based on the symbolic distance function and surface curvature for flow field prediction, the model can output high-precision velocity field, pressure field, and aerodynamic coefficients that conform to the laws of fluid physics in a very short time, requiring only geometric data as input.

[0079] Based on the above embodiments, this invention provides an automotive aerodynamic flow field prediction device, see [link to relevant documentation]. Figure 5 The diagram shown is a structural schematic of an automotive aerodynamic flow field prediction device according to an embodiment of the present invention. The device includes: The data acquisition module 310 is used to acquire the geometric model data of the car to be predicted; The point cloud acquisition module 320 is used to extract vehicle point cloud data from geometric model data; wherein, the vehicle point cloud data includes macro point cloud representing the global contour of the vehicle and micro point cloud representing local regions, and the surface curvature of the local regions is higher than a preset discrete average curvature threshold. The data processing module 330 is used to calculate the symbolic distance function value from each query point to the vehicle surface and the corresponding surface curvature based on the query points in the preset flow field space, and to generate a physical offset mask based on the symbolic distance function value and the surface curvature. The flow field prediction module 340 is used to input the features of the vehicle point cloud data into a pre-trained dual-scale transformer model. Through the decoder of the dual-scale transformer model, the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point are output according to the coordinates of the target query point. The dual-scale transformer model includes a macro encoder, a micro encoder, and a cross-scale attention layer connecting the macro encoder and the micro encoder. During the attention calculation process, a physical bias mask is superimposed on the attention weight matrix to inject physical prior constraints.

[0080] Furthermore, in some preferred embodiments of the present invention, the point cloud acquisition module 320 is used to perform Poisson disk sampling on the geometric model data to obtain a macro point cloud with a first preset number of nodes, and calculate the surface normal vector of each macro point in the macro point cloud; calculate the discrete average curvature of the surface of the geometric model, and retain the nodes in the region where the discrete average curvature is greater than a preset discrete average curvature threshold as micro point cloud, the micro point cloud including multiple patches, each patch containing nodes within a second preset number of nodes.

[0081] Furthermore, in some preferred embodiments of the present invention, the data processing module 330 is used to perform nonlinear mapping on the symbolic distance function value based on a preset learnable multilayer perceptron to obtain a first bias component; wherein the learnable multilayer perceptron includes a preset number of hidden layers, each hidden layer has a preset dimension, and the initialized weights satisfy the zero mean of the output; the product of the surface curvature of the corresponding query point and a preset learnable scalar parameter is determined as a second bias component; wherein the learnable scalar parameter has a preset initial value and is adaptively adjusted through backpropagation; the sum of the first bias component and the second bias component is determined as a physical bias mask.

[0082] Furthermore, in some preferred embodiments of the present invention, the dual-scale transformer model includes a macro encoder and a micro encoder connected through a cross-scale cross-attention layer; the macro encoder consists of multiple layers of standard self-attention transformer blocks, used for global self-attention encoding of features of the macro point cloud; the micro encoder consists of multiple layers of transformer blocks with introduced physical bias masks; the method further includes: in the micro encoder, injecting the physical bias mask into the attention weight matrix of the micro point cloud features; in the cross-scale cross-attention layer, performing a cross-attention operation using the micro features output by the micro encoder as a query vector and the macro features output by the macro encoder as a key vector and a value vector, so that the local features of the local region obtain background information from the macro global features.

[0083] Furthermore, in some preferred embodiments of the present invention, the flow field prediction module 340 is used to apply position encoding to the coordinates of the target query point, using the encoded position vector as the query vector, and using the multi-scale latent features output by the dual-scale transformer model as the key vector and value vector. The three-dimensional velocity components and pressure values ​​at the target query point are output through a multi-layer cross-attention decoding network. The area-weighted integral of the pressure values ​​predicted at each query point on the vehicle surface is performed, and the global aerodynamic coefficient is calculated in combination with the incoming flow direction. Alternatively, the multi-scale latent features are globally averaged and then input into a fully connected network to directly output the global aerodynamic coefficient.

[0084] Furthermore, in some preferred embodiments of the present invention, the apparatus further includes: a model training module, used to acquire multiple vehicle geometric samples and corresponding computational fluid dynamics simulation ground truth data; wherein, the computational fluid dynamics simulation ground truth data includes volumetric velocity field, surface pressure field, and aerodynamic coefficients; extracting macroscopic point cloud and microscopic point cloud for each vehicle geometric sample, and calculating the signed distance function and surface curvature to generate a physical bias mask; inputting the features of the macroscopic point cloud, the features of the microscopic point cloud, and the physical bias mask into the dual-scale transformer model to be trained to obtain the predicted velocity field, pressure field, and aerodynamic coefficients; constructing a joint loss function, and updating the model parameters of the dual-scale transformer model to be trained according to the joint loss function until the joint loss function converges; wherein, the joint loss function includes a volumetric velocity regression loss term, a surface pressure regression loss term, an aerodynamic coefficient regression loss term, and a physical conservation penalty term based on the predicted velocity field divergence, wherein the preset weight coefficient of the surface pressure regression loss term is greater than the preset weight coefficient of the volumetric velocity regression loss term, and the preset weight coefficient of the aerodynamic coefficient regression loss term is greater than the preset weight coefficient of the surface pressure regression loss term.

[0085] Furthermore, in some preferred embodiments of the present invention, the physical conservation penalty term is calculated based on the divergence of the incompressible continuity equation and is expressed as the square mean of the sum of the partial derivatives of the predicted velocity components with respect to spatial coordinates; the physical conservation penalty term has a dynamic weight coefficient in the joint loss function, which takes a preset initial weight value in the initial stage of training, gradually increases with the training process, and maintains a preset maximum weight value in the later stage of training.

[0086] Furthermore, in some preferred embodiments of the present invention, the dynamic weight coefficient maintains a preset initial weight value in the first preset iteration ratio stage of the training process, linearly increases from the preset initial weight value to a preset maximum weight value in the subsequent second preset iteration ratio stage, and maintains the preset maximum weight value in the remaining iteration ratio stage.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the automotive aerodynamic flow field prediction device described above can be referred to the corresponding process in the aforementioned embodiments of the automotive aerodynamic flow field prediction method, and will not be repeated here.

[0088] This invention also provides an automotive aerodynamic flow field prediction system for running an automotive aerodynamic flow field prediction method; see [link to related documentation]. Figure 6 The diagram shown is a structural schematic of an automotive aerodynamic flow field prediction system provided by an embodiment of the present invention. The automotive aerodynamic flow field prediction system includes a memory 400 and a processor 401. The memory 400 is used to store one or more computer instructions, and the one or more computer instructions are executed by the processor 401 to implement the above-mentioned automotive aerodynamic flow field prediction method.

[0089] Furthermore, Figure 6 The automotive aerodynamic flow prediction system shown also includes a bus 402 and a communication interface 403. The processor 401, the communication interface 403, and the memory 400 are connected via the bus 402.

[0090] The memory 400 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 403 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 402 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0091] Processor 401 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 401 or by instructions in software form. Processor 401 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 400, and processor 401 reads information from memory 400 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0092] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the above-described automotive aerodynamic flow field prediction method. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0093] The computer program products of the automotive aerodynamic flow field prediction method, device, and system provided in the embodiments of the present invention include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and / or device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0095] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0096] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting automotive aerodynamic flow fields, characterized in that, include: Obtain the geometric model data of the car to be predicted; Vehicle point cloud data is extracted from the geometric model data; wherein, the vehicle point cloud data includes macro point clouds representing the global outline of the vehicle and micro point clouds representing local regions; Based on the query points in the preset flow field space, calculate the symbolic distance function value from each query point to the vehicle surface and the corresponding surface curvature, and generate a physical offset mask based on the symbolic distance function value and the surface curvature. The features of the vehicle point cloud data are input into a pre-trained dual-scale transformer model, and the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point are output according to the coordinates of the target query point. In the attention calculation process, the physical bias mask is superimposed on the attention weight matrix to inject physical prior constraints.

2. The method according to claim 1, characterized in that, Extracting vehicle point cloud data from the geometric model data includes: The geometric model data is sampled using a Poisson disk to obtain the macroscopic point cloud with a first preset number of nodes, and the surface normal vector of each macroscopic point in the macroscopic point cloud is calculated. Calculate the discrete average curvature of the surface of the geometric model, and retain the nodes in the region where the discrete average curvature is greater than the preset discrete average curvature threshold as the micro-view cloud. The micro-view cloud includes multiple patches, and each patch contains nodes within a second preset number of nodes.

3. The method according to claim 1, characterized in that, Generating a physical bias mask based on the symbolic distance function value and the surface curvature includes: The first bias component is obtained by nonlinearly mapping the symbol distance function value based on a preset learnable multilayer perceptron; wherein the learnable multilayer perceptron contains a preset number of hidden layers, each of which has a preset dimension and the weights are initialized to satisfy the zero mean of the output. The product of the surface curvature corresponding to the query point and a preset learnable scalar parameter is determined as the second bias component; wherein the learnable scalar parameter has a preset initial value and is adaptively adjusted through backpropagation; The sum of the first bias component and the second bias component is determined as the physical bias mask.

4. The method according to claim 1, characterized in that, The dual-scale transformer model includes a macro encoder and a micro encoder connected through a cross-scale attention layer; the macro encoder consists of multiple layers of standard self-attention transformer blocks, used for global self-attention encoding of the features of the macro point cloud; the micro encoder consists of multiple layers of transformer blocks that introduce the physical bias mask; the method further includes: In the micro encoder, the physical bias mask is injected into the attention weight matrix of the micro view cloud feature; In the cross-scale attention layer, the micro-features output by the micro encoder are used as query vectors, and the macro-features output by the macro encoder are used as key vectors and value vectors to perform cross-attention operations, so that the local features of the local region can obtain background information from the macro-global features.

5. The method according to claim 1, characterized in that, Based on the coordinates of the target query point, the output includes the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficients at the target query point, including: Position encoding is applied to the coordinates of the target query point. The encoded position vector is used as the query vector. The multi-scale latent features output by the dual-scale transformer model are used as the key vector and value vector. The three-dimensional velocity component and pressure value at the target query point are output through a multi-layer cross-attention decoding network. The area-weighted integral is performed on the predicted pressure values ​​at each query point on the vehicle surface, and the global aerodynamic coefficient is calculated in combination with the incoming flow direction. Alternatively, the multi-scale latent features can be globally averaged and pooled before being input into a fully connected network to directly output the global aerodynamic coefficients.

6. The method according to claim 1, characterized in that, The method further includes: Multiple automotive geometry samples and corresponding computational fluid dynamics (CFD) simulation ground truth data are obtained; wherein, the CFD simulation ground truth data includes volumetric velocity field, surface pressure field, and aerodynamic coefficients; For each of the vehicle geometry samples, the macroscopic point cloud and the microscopic point cloud are extracted, and the symbolic distance function and the surface curvature are calculated to generate the physical bias mask; The features of the macroscopic point cloud, the features of the microscopic point cloud, and the physical bias mask are input into the dual-scale transformer model to be trained to obtain the predicted velocity field, pressure field, and aerodynamic coefficients. A joint loss function is constructed, and the model parameters of the dual-scale transformer model to be trained are updated according to the joint loss function until the joint loss function converges; wherein, the joint loss function includes a volume velocity regression loss term, a surface pressure regression loss term, an aerodynamic coefficient regression loss term, and a physical conservation penalty term based on the predicted velocity field divergence, wherein the preset weight coefficient of the surface pressure regression loss term is greater than the preset weight coefficient of the volume velocity regression loss term, and the preset weight coefficient of the aerodynamic coefficient regression loss term is greater than the preset weight coefficient of the surface pressure regression loss term.

7. The method according to claim 6, characterized in that, The physical conservation penalty term is calculated based on the divergence of the incompressible continuity equation and is expressed as the square mean of the sum of the partial derivatives of the predicted velocity components with respect to spatial coordinates. The physical conservation penalty term has a dynamic weight coefficient in the joint loss function. The dynamic weight coefficient takes a preset initial weight value in the initial stage of training, gradually increases as the training progresses, and maintains a preset maximum weight value in the later stage of training.

8. The method according to claim 7, characterized in that, The dynamic weight coefficient maintains the preset initial weight value in the first preset iteration ratio stage of the training process, linearly increases from the preset initial weight value to the preset maximum weight value in the subsequent second preset iteration ratio stage, and maintains the preset maximum weight value in the remaining iteration ratio stage.

9. A device for predicting automotive aerodynamic flow fields, characterized in that, include: The data acquisition module is used to acquire the geometric model data of the car to be predicted; The point cloud acquisition module is used to extract vehicle point cloud data from the geometric model data; wherein, the vehicle point cloud data includes macro point clouds representing the global outline of the vehicle and micro point clouds representing local areas. The data processing module is used to calculate the symbolic distance function value from each query point to the vehicle surface and the corresponding surface curvature based on the query points in the preset flow field space, and to generate a physical offset mask based on the symbolic distance function value and the surface curvature. The flow field prediction module is used to input the features of the vehicle point cloud data into a pre-trained dual-scale transformer model, and output the three-dimensional velocity vector, pressure scalar, and global aerodynamic coefficient at the target query point according to the target query point coordinates; wherein, during the attention calculation process, the physical bias mask is superimposed on the attention weight matrix to inject physical prior constraints.

10. A vehicle aerodynamic flow field prediction system, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method according to any one of claims 1 to 8.

Images