High-fidelity pneumatic agent modeling and optimization method and device, equipment, storage medium

Abstract

The embodiment of the application relates to the field of simulation models, in particular to a high-fidelity aerodynamic proxy modeling and optimization method and device, equipment and a storage medium, the method takes a triangular facet as a basic modeling unit, and extracts space information and topological information of each facet of a target model. Based on the space information and the topological information, spatial channel coding and structural feature coding are carried out to obtain a distribution feature vector and a structural feature vector of each facet. The distribution feature vector and the structural feature vector are aggregated to obtain a geometric feature vector of the target model. The geometric feature vector is input into a trained multi-task collaborative processing model to obtain a corresponding resistance coefficient and a volume flow field of the target model. Through the facet-level feature extraction mechanism, the application accurately captures the boundary details and topological structure of complex geometry. At the same time, through multi-task collaborative prediction based on the extracted accurate features of the same model, the efficiency of the parameter determination process is improved, and it is ensured that the prediction result conforms to the physical law.

Description

Technical Field

[0001] This application relates to the field of simulation models, and includes, but is not limited to, a high-fidelity pneumatic proxy modeling and optimization method, apparatus, device, and storage medium. Background Technology

[0002] In aerodynamic modeling for fields such as automotive and aerospace, computational fluid dynamics (CFD) simulation is a core tool for evaluating aerodynamic performance. It can obtain key physical quantities such as drag coefficient, surface pressure, and volumetric flow field, providing a basis for product optimization. Existing methods are complex in predicting model physical parameters, resulting in low efficiency and accuracy. Furthermore, existing methods often focus on predicting single physical quantities (such as drag coefficient), neglecting the co-modeling of surface fields (pressure, shear stress) and volumetric flow fields (velocity vectors). This leads to physical distortions in prediction results in high-gradient regions such as the wake and boundary layer, such as the velocity field not satisfying divergence constraints and pressure gradients not matching actual flow patterns. Summary of the Invention

[0003] In view of this, the high-fidelity pneumatic proxy modeling and optimization method, apparatus, equipment, and storage medium provided in the embodiments of this application improve the efficiency and accuracy of determining physical parameters in the process of fluid dynamics simulation, while ensuring that the obtained physical quantities conform to actual laws.

[0004] The high-fidelity pneumatic proxy modeling and optimization method, apparatus, device, and storage medium provided in this application embodiment are implemented as follows:

[0005] One aspect of this application provides a high-fidelity pneumatic agent modeling and optimization method, the method comprising:

[0006] Using triangular facets as the basic modeling unit, the spatial and topological information of each facet in the target model is extracted. The spatial information includes the center coordinates, displacement vector, and unit normal vector, while the topological information includes the adjacent facet index.

[0007] Spatial channel encoding is performed based on the spatial information of each patch to obtain the corresponding distribution feature vector;

[0008] Structural feature encoding is performed based on the spatial and topological information of each facet to obtain the corresponding structural feature vector;

[0009] Vector aggregation is performed on the distribution feature vectors and structural feature vectors of each patch to obtain the geometric feature vectors of the target model;

[0010] By inputting the geometric feature vectors into the trained multi-task collaborative processing model, the drag coefficient and volumetric flow field corresponding to the target model are obtained.

[0011] In one possible implementation, spatial channel encoding is performed based on the spatial information of each patch to obtain the corresponding distribution feature vector, including:

[0012] The center coordinates of each facet are encoded by a multilayer perceptron to obtain the corresponding distribution feature vector.

[0013] In one possible implementation, structural feature encoding is performed based on the spatial and topological information of each facet to obtain the corresponding structural feature vector, including:

[0014] Extract the geometric shape features of each facet based on its displacement vector;

[0015] The topological relationship features of each facet are extracted based on its unit normal vector and topological information.

[0016] Based on the geometric shape and topological characteristics of each facet, the corresponding structural feature vector is determined.

[0017] In one possible implementation, the geometric shape features of the facets are extracted based on the displacement vectors of each facet, including:

[0018] Based on the three-dimensional displacement vector of each patch, determine the corresponding three two-dimensional sub-vectors;

[0019] The local features corresponding to the three two-dimensional sub-vectors of each facet are extracted respectively;

[0020] By using a multilayer perceptron to fuse the three local features corresponding to each facet, geometric shape features are obtained.

[0021] In one possible implementation, the topological relationship features of the facets are extracted based on the unit normal vectors and topological information of each facet, including:

[0022] The corresponding adjacent facets are determined based on the topological information of each facet;

[0023] The set of normal vectors is determined based on the unit normal vector of each facet and the unit normal vector of the corresponding adjacent facets;

[0024] For each facet, the spatial distribution similarity between the unit normal vectors of the facet is calculated based on the set of normal vectors and the preset set of sphere kernel vectors to obtain the topological relationship features.

[0025] In one possible implementation, the distribution feature vectors and structural feature vectors of each facet are aggregated to obtain the geometric feature vectors of the target model, including:

[0026] For each patch, the distribution feature vector and structural feature vector are processed by multiple convolutions and then the features are aggregated to obtain the patch convolution features;

[0027] The geometric feature vector of the target model is determined based on the convolution features of each facet.

[0028] In one possible implementation, the geometric feature vector is input into the trained multi-task collaborative processing model to obtain the drag coefficient and volumetric flow field corresponding to the target model, including:

[0029] Input the geometric feature vector into the feature transformation module to determine the physical feature vector corresponding to the target model;

[0030] Input the physical feature vector into the drag prediction module to obtain the drag coefficient corresponding to the target model;

[0031] Input the physical feature vector into the volume flow field prediction module to obtain the volume flow field corresponding to the target model.

[0032] Another aspect of this application embodiment provides a high-fidelity pneumatic proxy modeling and optimization device, the device comprising:

[0033] The feature extraction module is used to extract the spatial and topological information of each facet in the target model, using triangular facets as the basic modeling unit. The spatial information includes the center coordinates, displacement vector, and unit normal vector, and the topological information includes the adjacent facet index.

[0034] The first encoding module is used to perform spatial channel encoding based on the spatial information of each patch to obtain the corresponding distribution feature vector;

[0035] The second encoding module is used to encode structural features based on the spatial and topological information of each facet to obtain the corresponding structural feature vector.

[0036] The feature aggregation module is used to aggregate the distribution feature vectors and structural feature vectors of each facet to obtain the geometric feature vectors of the target model.

[0037] The task processing module is used to input geometric feature vectors into the multi-task collaborative processing model obtained through training, and to obtain the drag coefficient and volumetric flow field corresponding to the target model.

[0038] In one possible implementation, the first encoding module is further used for:

[0039] The center coordinates of each facet are encoded by a multilayer perceptron to obtain the corresponding distribution feature vector.

[0040] In one possible implementation, the second encoding module is further used for:

[0041] Extract the geometric shape features of each facet based on its displacement vector;

[0042] The topological relationship features of each facet are extracted based on its unit normal vector and topological information.

[0043] Based on the geometric shape and topological characteristics of each facet, the corresponding structural feature vector is determined.

[0044] In one possible implementation, the second encoding module is further used for:

[0045] Based on the three-dimensional displacement vector of each patch, determine the corresponding three two-dimensional sub-vectors;

[0046] The local features corresponding to the three two-dimensional sub-vectors of each facet are extracted respectively;

[0047] By using a multilayer perceptron to fuse the three local features corresponding to each facet, geometric shape features are obtained.

[0048] In one possible implementation, the second encoding module is further used for:

[0049] The corresponding adjacent facets are determined based on the topological information of each facet;

[0050] The set of normal vectors is determined based on the unit normal vector of each facet and the unit normal vector of the corresponding adjacent facets;

[0051] For each facet, the spatial distribution similarity between the unit normal vectors of the facet is calculated based on the set of normal vectors and the preset set of sphere kernel vectors to obtain the topological relationship features.

[0052] In one possible implementation, the feature aggregation module is further used for:

[0053] For each patch, the distribution feature vector and structural feature vector are processed by multiple convolutions and then the features are aggregated to obtain the patch convolution features;

[0054] The geometric feature vector of the target model is determined based on the convolution features of each facet.

[0055] In one possible implementation, the task processing module is further used for:

[0056] Input the geometric feature vector into the feature transformation module to determine the physical feature vector corresponding to the target model;

[0057] Input the physical feature vector into the drag prediction module to obtain the drag coefficient corresponding to the target model;

[0058] Input the physical feature vector into the volume flow field prediction module to obtain the volume flow field corresponding to the target model.

[0059] The electronic device provided in this application includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the method described in this application.

[0060] The computer-readable storage medium provided in this application embodiment stores a computer program thereon, which, when executed by a processor, implements the method provided in this application embodiment.

[0061] In some embodiments, this method uses triangular facets as basic modeling units to extract spatial and topological information for each facet of the target model. Based on the spatial and topological information, spatial channel encoding and structural feature encoding are performed to obtain the distribution feature vector and structural feature vector for each facet. Vector aggregation is then performed on the distribution feature vector and structural feature vector to obtain the geometric feature vector of the target model. The geometric feature vector is input into a trained multi-task collaborative processing model to obtain the drag coefficient and volumetric flow field corresponding to the target model. This application accurately captures the boundary details and topological structure of complex geometries through a facet-level feature extraction mechanism. Simultaneously, by using the same model to perform multi-task collaborative prediction based on the extracted precise features, the efficiency of the parameter determination process is improved, and the prediction results are ensured to conform to physical laws. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart illustrating a high-fidelity pneumatic agent modeling and optimization method according to an embodiment of this application is shown.

[0064] Figure 2 A schematic diagram illustrating a process for extracting model information according to an embodiment of this application is shown;

[0065] Figure 3 A schematic diagram of a feature conversion module according to an embodiment of this application is shown;

[0066] Figure 4 A schematic diagram illustrating a high-fidelity pneumatic proxy modeling and optimization process according to an embodiment of this application is shown;

[0067] Figure 5 This diagram illustrates a high-fidelity pneumatic proxy modeling and optimization device according to an embodiment of this application;

[0068] Figure 6A schematic diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0070] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0071] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0072] It should be noted that the terms "first, second, third" used in the embodiments of this application are used to distinguish similar or different objects and do not represent a specific order of objects. It can be understood that "first, second, third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0073] The high-fidelity pneumatic proxy modeling and optimization method of this application embodiment can be executed by any electronic device, including but not limited to mobile phones, wearable devices (such as smartwatches, smart bracelets, smart glasses, etc.), tablet computers, laptops, vehicle terminals, PCs (Personal Computers), etc. The functions implemented by this method can be achieved by the processor in the electronic device calling program code. Of course, the program code can be stored in the computer storage medium. It can be seen that the electronic device includes at least a processor and a storage medium.

[0074] The high-fidelity aerodynamic proxy modeling and optimization method of this application can be used for any application scenario that requires constructing an aerodynamic model by determining physical parameters. For example, this application can be applied to aircraft design in the aerospace field. Alternatively, it can be applied to automobile design in the automotive industry.

[0075] With the intelligent transformation of the automotive, aerospace, and high-end equipment manufacturing industries, product development processes are undergoing a profound shift from "experience-driven" to "data-driven." Aerodynamic performance, as a core support for multiple indicators such as vehicle energy consumption optimization, driving stability, and wind noise control, plays a crucial role in the evaluation and optimization process of modern design. In the early stages of vehicle design, styling design must not only meet aesthetic standards and brand language consistency, but also balance multiple objectives such as aerodynamic performance, cooling requirements, and manufacturing processes within limited resources and tight timeframes. This multi-objective coupled task poses unprecedented challenges to the design process's response speed, modeling accuracy, and feedback loop.

[0076] In the process of constructing aerodynamic models using relevant technologies, engineers rely on Computational Fluid Dynamics (CFD) simulations or wind tunnel experiments to evaluate and iterate on design schemes. Although CFD offers extremely high simulation accuracy and controllability, its high computational resource consumption and complex mesh construction and convergence parameter tuning processes result in lengthy remodeling and simulation cycles for each geometric update. In highly complex scenarios at the vehicle level, a single CFD simulation can take hours or even tens of hours, severely limiting the depth of shape space exploration and optimization efficiency.

[0077] The breakthroughs in deep learning for image, text, and multimodal modeling in recent years have spurred extensive exploration of data-driven methods in the field of CFD. Early studies often used multilayer perceptrons (MLPs) or convolutional neural networks (CNNs) to fit the relationship between parametric geometry and flow field variables. Some studies also attempted to achieve structure-performance mapping using point clouds, voxel meshes, or implicit representations. However, these methods still face significant challenges in handling complex boundary geometry and local physical discontinuities. Point cloud methods lack structural connectivity information, voxel methods face a trade-off between resolution and memory consumption, and implicit methods, while expressive, lack analytical boundary control.

[0078] Meanwhile, related technologies often focus on predicting single physical quantities (such as drag coefficients) while neglecting the collaborative modeling of surface fields (pressure, shear stress) and volumetric flow fields (velocity vectors). This leads to physical distortions in the prediction results in high gradient regions such as wakes and boundary layers, such as the velocity field not satisfying divergence constraints and the pressure gradient not matching the actual flow patterns.

[0079] Therefore, the technical problem solved by the embodiments of this application is how to improve the efficiency and accuracy of determining physical parameters in the process of fluid mechanics simulation, and how to ensure that each physical quantity in the simulation results conforms to the actual law.

[0080] The high-fidelity pneumatic agent modeling and optimization scheme of this application embodiment will be described in detail below with reference to the accompanying drawings.

[0081] Figure 1 A flowchart illustrating a high-fidelity pneumatic agent modeling and optimization method according to an embodiment of this application is shown. Figure 1 As shown, the high-fidelity pneumatic agent modeling and optimization method of this application embodiment may include the following steps S10-S50.

[0082] For ease of description, the high-fidelity pneumatic proxy modeling and optimization method of this application embodiment is described using an electronic device as the execution subject. It should be understood that the execution subject of this application embodiment can also be a processor or chip in an electronic device, and this application embodiment does not impose any limitations.

[0083] Step S10: Using triangular facets as the basic modeling unit, extract the spatial and topological information of each facet in the target model.

[0084] In one possible implementation, after acquiring the three-dimensional target model for which an aerodynamic model needs to be constructed, the electronic device in this embodiment can extract the spatial and topological information of each facet of the target model, using triangular facets as the basic modeling unit. The type of target model varies depending on the application scenario. For example, in an application scenario requiring the construction of a vehicle aerodynamic model, the target model can be determined as a three-dimensional model of the target vehicle; in an application scenario requiring the construction of an aircraft aerodynamic model, the target model can be determined as a three-dimensional model of the target aircraft.

[0085] In some embodiments, during the aerodynamic modeling of a target object, to achieve geometric representation and physical consistency modeling in complex model structure modeling tasks, and to effectively characterize the geometric structure of the target object's shape and capture its detailed features, this application embodiment can use triangular facets as basic modeling units to extract the spatial and topological information of each facet in the target model. This, combined with a facet unit modeling strategy, achieves high-fidelity geometric feature extraction. The extracted spatial information includes the center coordinates, displacement vector, and unit normal vector of each facet, while the topological information includes the adjacent facet indices of each facet.

[0086] Optionally, in the preprocessing stage, the electronic device can first convert the target model into a set of faces. If the face set includes non-triangular faces, the electronic device can automatically convert them into triangular faces using a triangulation algorithm. For each triangular face, the electronic device can define its spatial geometric properties to obtain spatial information and define its adjacent faces to obtain topological information. The spatial information includes center coordinates, displacement vector, and unit normal vector. The center coordinates can be determined by calculating the average of the face's three vertices, the displacement vector is the vector from the center point to each vertex of the face, and the unit normal vector can be used to characterize the orientation of the face's surface. The topological information includes adjacent face indices, used to record the three faces adjacent to each face. When determining the adjacent face indices of a face, if the number of adjacent faces corresponding to a face is less than three, the adjacent face indices of that face can be determined by self-indexing. In cases where the unit normal vector of a facet is calculated abnormally (e.g., the normal vector magnitude exceeds the range of 0.9-1.1), the electronic device can correct the unit normal vector of the facet by weighted interpolation of the unit normal vectors of adjacent faces. The weight coefficients used in the calculation process can be set according to the proportion of the adjacent edge lengths.

[0087] Figure 2 This diagram illustrates a process for extracting model information according to an embodiment of this application. Figure 2 As shown, after acquiring the target model, the electronic device first converts the target model into a set of facets. Further, the electronic device determines the center coordinates, displacement vector, and unit normal vector of each facet to determine the spatial information used to describe the spatial characteristics of each facet. Simultaneously, the topological information of each facet is determined by recording adjacent facet indices.

[0088] Step S20: Perform spatial channel encoding based on the spatial information of each patch to obtain the corresponding distribution feature vector.

[0089] In one possible implementation, after determining the spatial information of each panel, the electronic device can perform spatial channel encoding in the spatial dimension based on the center coordinates of the panels in the spatial information to obtain the distribution feature vector corresponding to each panel. This spatial channel encoding process can be implemented using a multilayer perceptron; that is, the electronic device can use a multilayer perceptron to encode the center coordinates of each panel to capture the absolute position and spatial distribution characteristics of the panels, thereby obtaining the distribution feature vector corresponding to each panel.

[0090] Alternatively, the electronic device can encode different facets based on a multilayer perceptron with the same parameters. That is, it can encode the centers of all facets using a multilayer perceptron (MLP) with shared parameters to extract their spatial location information. Compared with point cloud feature extraction methods, this approach can achieve unified encoding of the geometric positions of facets without introducing explicit topological structures.

[0091] Step S30: Based on the spatial and topological information of each facet, perform structural feature encoding to obtain the corresponding structural feature vector.

[0092] In one possible implementation, the electronic device in this embodiment performs spatial channel encoding based on the spatial information of the facets, and also performs structural feature encoding based on the spatial and topological information of each facet to obtain the corresponding structural feature vector. That is, this embodiment uses dual-channel encoding to extract facet features, thereby obtaining accurate facet features.

[0093] In some embodiments, when encoding structural features, the electronic device extracts features from two dimensions: geometric shape and topological relationship. Specifically, the electronic device first extracts the geometric shape features of each facet based on its displacement vector, and then extracts the topological relationship features based on the unit normal vector and topological information of each facet. Finally, based on the geometric shape and topological relationship features of each facet, the corresponding structural feature vector is determined.

[0094] Optionally, when extracting the geometric shape features of the facets, the electronic device can determine three corresponding two-dimensional sub-vectors based on the three-dimensional displacement vector of each facet. Then, it can extract the corresponding local features from the three two-dimensional sub-vectors of each facet. Finally, it can fuse the three local features corresponding to each facet using a multilayer perceptron to obtain the geometric shape features.

[0095] For example, considering that each triangular facet consists of three corner vectors, the electronic device can define the displacement vector of the facet as a three-dimensional displacement vector {v1, v2, v3} pointing from the center point to the three vertices, where each element is a two-dimensional vector pointing from the center point to a corner point. The electronic device can decompose the three-dimensional displacement vector to obtain the corresponding three two-dimensional sub-vectors (v1, v2), (v2, v3), and (v3, v1), and then further apply a shared function f(·) to the above three sub-vectors for local feature fusion, and aggregate them into a structural representation of the facet as a set shape feature through average pooling and a multilayer perceptron MLPg(·) with shared parameters. That is, the electronic device can use the formula Determine the geometric shape features. Among them, For local feature extraction operators, As a multilayer perceptron with shared parameters, this structure is equivalent to a local patch convolution mechanism with corner pairs as the kernel. By rotating and traversing the corner pairs, it models the internal morphological changes of the patches, thereby enhancing the perception of structural information such as boundary curvature and contour abrupt changes.

[0096] Optionally, when extracting the topological relationship features of facets, the electronic device can determine the corresponding adjacent facets based on the topological information of each facet. It can also determine the set of normal vectors based on the unit normal vectors of each facet and the unit normal vectors of its corresponding adjacent facets. For each facet, the spatial distribution similarity between the unit normal vectors of the facets is calculated based on the set of normal vectors and a preset set of sphere kernel vectors to obtain the topological relationship features.

[0097] For example, an electronic device can propose a kernel correlation mechanism that encodes based on the spatial distribution similarity between the normal vectors of facets. The encoding process of this correlation mechanism is as follows: Let N be the set of normal vectors of the i-th facet. i (Including its own and the normal vectors of all adjacent faces), construct a set of learnable spherical kernels, where each kernel k is represented by a set of normal vectors M. k Its direction is determined by the spherical coordinate parameters (θ, Control, mapped to Euclidean space as Where θ ∈ [0, π], ∈ [0, 2π].

[0098] Furthermore, kernel similarity is defined as the kernel function K. σ The weighted average of (n, m), where n ∈ N i , m ∈ M k The kernel function is in the form of a Gaussian kernel. Where ∥ · ∥ represents the vector length in Euclidean space, and σ is a hyperparameter controlling the resolving power of the convolution kernel or its tolerance to source variations. The kernel correlation degree between patch i and kernel k is calculated as follows: Where N i M is the set of normals of the i-th face and its neighbors. k It is the set of normals in the k-th convolutional kernel. Using the above definition, we can obtain the similarity vector between each patch and all kernels, forming the structural feature representation `houter ∈ R`. M , which serves as the topological relationship feature of the facets. Here, M represents the number of kernels. This mechanism can adaptively learn the common normal arrangement distributions between facets, enhancing the model's ability to represent complex structural regions such as boundary layer perturbations and tail vortex regions.

[0099] Step S40: Perform vector aggregation on the distribution feature vector and the structural feature vector of each of the facets to obtain the geometric feature vector of the target model.

[0100] In one possible implementation, after determining the distribution feature vector and structural feature vector of each facet through dual-channel encoding, the electronic device aggregates these vectors to obtain the geometric feature vector of the target model. Specifically, the electronic device can first perform multi-layer convolution processing on the distribution feature vector and structural feature vector of each facet before feature aggregation to obtain facet convolution features. Then, the geometric feature vector of the target model is determined based on the facet convolution features corresponding to each facet.

[0101] In some embodiments, to achieve effective propagation of structural context, based on the structural features of each facet and its adjacency index, this application embodiment can aggregate the structural vectors of the current facet and all its neighboring faces to achieve local structural information fusion. This aggregation method can include three typical aggregation methods: average pooling, max pooling, and concatenation. This method concatenates the feature pairs of each facet and its neighboring faces, and then performs max pooling after sharing an MLP mapping to obtain the facet convolutional features of each facet. This feature concatenation method can preserve the response of highly activated regions while enhancing the network's ability to express different adjacency patterns.

[0102] Furthermore, after determining the convolutional features of each facet, the electronic device can directly use the convolutional features of each facet as a dimension of the vector to obtain the geometric feature vector of the target model.

[0103] Step S50: Input the geometric feature vector into the trained multi-task collaborative processing model to obtain the drag coefficient and volumetric flow field corresponding to the target model.

[0104] In one possible implementation, the electronic device of this application embodiment predetermines a multi-task collaborative model for aerodynamic modeling based on the geometric feature vectors of the target model, thereby obtaining the drag coefficient and volumetric flow field corresponding to the target model. Optionally, in different application scenarios, the electronic device can further determine other physical parameters of the target model, such as surface physical quantities, based on the multi-task collaborative processing model.

[0105] Optionally, the multi-task collaborative processing model in this embodiment may include a feature transformation model for performing feature transformation, and a multi-task processing module for performing task processing based on the transformed features. In different application scenarios, the multi-task processing module may be further divided into multiple sub-modules based on the physical parameters that need to be determined. For example, an electronic device can input a geometric feature vector into the feature transformation module to determine the physical feature vector corresponding to the target model. The physical feature vector is then input into the drag prediction module to obtain the drag coefficient corresponding to the target model. Finally, the physical feature vector is input into the volumetric flow field prediction module to obtain the volumetric flow field corresponding to the target model.

[0106] In some embodiments, the feature transformation module is used to convert the input geometric features into physical features, and can employ the U-Shape Mamba (USM) semantic modeling architecture. This architecture adopts a symmetric Encoder-Decoder structure, with the Mamba (state-space architecture) main module at its core to construct deep semantic channels. The input to this module is the geometric feature vector of the target model, and the output is the structural latent variables. It is used to drive the downstream multi-task prediction branch.

[0107] Figure 3 A schematic diagram of a feature conversion module according to an embodiment of this application is shown. Figure 3 As shown, the overall structure of the feature transformation module consists of input feature compression, deep encoding path, semantic bottleneck connection, symmetric decoding path, and state modeling mechanism. First, the geometric feature vector of the target model is input. By integrating spatial and structural information, and compressing it through convolutional downsampling operations, a dimensionality-reduced sequence representation is obtained. To enhance the model's ability to model long-range dependencies, the network employs 12 Mamba-based main modules in the encoder section, gradually extracting high-dimensional features with global semantic representation capabilities through inter-layer downsampling operations. The bottleneck layer, located in the middle of the network, acts as a bridge connecting the encoder and decoder, effectively fusing multi-scale contextual information while preserving deep geometric and structural semantics, providing a strong representational foundation for physical field reconstruction in the decoding stage. Situated at the center of the U-shaped structure, the bottleneck layer connects deep structural semantics with shallow geometric details, serving as a crucial bridge for cross-scale and long-range modeling capabilities.

[0108] The decoder structure is designed to be symmetrical with the encoder. It recovers the feature space dimension and reconstructs boundary structures and local geometric details through layer-by-layer upsampling operations and skip connections, thereby improving the continuity and spatial fidelity of the prediction results. Each decoding stage preserves shallow features while guiding the model to gradually reconstruct the physical semantics consistent with the input 3D mesh structure. To enhance sequence modeling capabilities and the model's adaptability to different physical control parameters, each main module of USM integrates a time-step normalization mechanism (AdaLN) and a Mamba block for state-space modeling, enabling the network to effectively capture feature dependencies at both spatial and temporal scales. The network also supports the introduction of external control variables and the injection of multimodal information through conditional attention mechanisms, further enhancing the model's flexibility and generalization ability in multi-task scenarios.

[0109] Ultimately, USM outputs a dimension R. 256 Unified latent variable representation This structure serves as the physical feature vector of the target module and as a shared input for subsequent tasks such as drag coefficient prediction, surface physics field reconstruction, and 3D volumetric flow field estimation, constructing a unified representation framework from structural space to multiphysics response. This structure not only demonstrates excellent accuracy and robustness in the physical consistency modeling of complex 3D structures but also provides a scalable, controllable, and computationally efficient modeling path for high-resolution, high-complexity CFD surrogate prediction.

[0110] In some embodiments, the multi-task processing module can collaboratively predict drag coefficients and volumetric flow fields based on physical feature vectors. The drag coefficient can be predicted by inputting the physical feature vectors into the drag prediction module to obtain the drag coefficient corresponding to the target model. The volumetric flow field can be predicted by inputting the physical feature vectors into the volumetric flow field prediction module to obtain the volumetric flow field corresponding to the target model.

[0111] For example, regarding the drag coefficient C in aerodynamic performance indicators... d This application embodiment constructs a lightweight regression network branch, utilizing the physical feature vectors derived from USM for learning. Specifically, the electronic device can learn from the physical feature vectors... The input module consists of five convolutional layers, each followed by a batch normalization layer and a max pooling operation, to extract hierarchical features layer by layer and effectively compress spatial dimensions. After feature extraction, the output tensor is flattened and passed to a regressor consisting of three fully connected layers to map the compressed high-dimensional features into accurate drag coefficient predictions. This module focuses on compressing high-dimensional structural representations into the performance prediction space, avoiding overfitting issues introduced by redundant geometric details. Due to the nonlinear and nonlocal nature of the structure-flow response relationship, a physical prior regularization term is introduced during network training to enhance the learning of regions of interest in boundary layer morphology. This module achieves a good balance between inference efficiency and physical consistency, and can quickly provide performance references for downstream optimization.

[0112] Furthermore, in order to fully reconstruct the three-dimensional velocity vector distribution (Ux, Uy, Uz) within the volumetric flow field, the volumetric flow field prediction module of this embodiment first... Surface physical quantities are calculated, and based on the predicted surface physical quantities, a decoder module driven by U-Shape Mamba is further extended to perform local query modeling by combining structure-aware features and spatial coordinates. The input physical feature vector, C = 4, corresponds to one pressure channel and three shear stress channels, respectively. The network preserves low-level geometric features through a skip connection mechanism and introduces local geometric guidance weights during the prediction process to adapt to the vortex structure in the boundary region.

[0113] Specifically, after calculating the surface physical quantities, the electronic device will transfer the physical feature vector. The decoder restores the volumetric features to a spatially consistent three-dimensional representation, which is then jointly encoded with the three-dimensional coordinates of the target location to form the query input for velocity prediction. This process can be represented as follows: ,in This represents the feature vector obtained based on structural decoding features and local neighborhood interpolation, where (x, y, z) are the position coordinates of the target point in three-dimensional space. By fusing and nonlinearly mapping the above information through a multilayer perceptron (MLP), the network can output a three-dimensional velocity estimate for the corresponding location.

[0114] To enhance the physical plausibility and numerical stability of this module, a smoothness loss and a divergence-free penalty are introduced during the training phase to constrain the continuity and physical consistency of the predicted flow field. This branch enables the model to not only possess strong structure awareness but also to perform velocity tensor predictions at arbitrary spatial resolutions, meeting the detailed modeling needs of industrial CFD for local vortex structure reconstruction, boundary layer analysis, and flow mechanism visualization.

[0115] Figure 4 This diagram illustrates a high-fidelity pneumatic proxy modeling and optimization process according to an embodiment of this application. Figure 4As shown, in this embodiment, after acquiring the target model, the electronic device first uses triangular facets as the basic modeling unit to extract the spatial and topological information of each facet in the target model. The spatial information includes the center coordinates, displacement vector, and unit normal vector, while the topological information includes the adjacent facet index. Then, through a dual-channel encoding method of spatial channel encoding and structural feature encoding, the geometric feature vector of the target model is obtained. The obtained geometric feature vector is then input into the feature conversion module to determine the physical feature vector corresponding to the target model. The physical feature vector is input into the drag prediction module to obtain the drag coefficient corresponding to the target model. Finally, the physical feature vector is input into the volumetric flow field prediction module to obtain the volumetric flow field corresponding to the target model, thus achieving the prediction of the physical parameters of the target model.

[0116] Based on the aforementioned technical features, the accuracy of the physical parameter prediction results in this application embodiment is significantly improved. Furthermore, through divergence constraints and gradient preservation mechanisms, the predicted velocity field satisfies the incompressible fluid continuity condition, the surface pressure gradient matches the actual flow pattern, and the deviation between the wake vortex structure morphology and the true value is reduced, enhancing physical consistency. Simultaneously, this application embodiment exhibits short inference time, high efficiency, and strong industrial generalization capability.

[0117] It should be understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0118] Based on the foregoing embodiments, this application provides a high-fidelity pneumatic agent modeling and optimization device. The device includes various modules and units included in each module, which can be implemented by a processor; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), microprocessor (MPU), digital signal processor (DSP), or field programmable gate array (FPGA), etc.

[0119] Figure 5 A schematic diagram of a high-fidelity pneumatic proxy modeling and optimization device according to an embodiment of this application is shown. Figure 5 As shown, the high-fidelity pneumatic proxy modeling and optimization device in this application embodiment includes:

[0120] The feature extraction module 50 is used to extract the spatial and topological information of each facet in the target model, using triangular facets as the basic modeling unit. The spatial information includes the center coordinates, displacement vector, and unit normal vector, and the topological information includes the adjacent facet index.

[0121] The first encoding module 51 is used to perform spatial channel encoding based on the spatial information of each patch to obtain the corresponding distribution feature vector;

[0122] The second encoding module 52 is used to encode structural features based on the spatial and topological information of each facet to obtain the corresponding structural feature vector.

[0123] The feature aggregation module 53 is used to aggregate the distribution feature vectors and structural feature vectors of each facet to obtain the geometric feature vectors of the target model.

[0124] The task processing module 54 is used to input the geometric feature vector into the trained multi-task collaborative processing model to obtain the drag coefficient and volumetric flow field corresponding to the target model.

[0125] In one possible implementation, the first encoding module 51 is further configured to:

[0126] The center coordinates of each facet are encoded by a multilayer perceptron to obtain the corresponding distribution feature vector.

[0127] In one possible implementation, the second encoding module 52 is further used for:

[0128] Extract the geometric shape features of each facet based on its displacement vector;

[0129] The topological relationship features of each facet are extracted based on its unit normal vector and topological information.

[0130] Based on the geometric shape and topological characteristics of each facet, the corresponding structural feature vector is determined.

[0131] In one possible implementation, the second encoding module 52 is further used for:

[0132] Based on the three-dimensional displacement vector of each patch, determine the corresponding three two-dimensional sub-vectors;

[0133] The local features corresponding to the three two-dimensional sub-vectors of each facet are extracted respectively;

[0134] By using a multilayer perceptron to fuse the three local features corresponding to each facet, geometric shape features are obtained.

[0135] In one possible implementation, the second encoding module 52 is further used for:

[0136] The corresponding adjacent facets are determined based on the topological information of each facet;

[0137] The set of normal vectors is determined based on the unit normal vector of each facet and the unit normal vector of the corresponding adjacent facets;

[0138] For each facet, the spatial distribution similarity between the unit normal vectors of the facet is calculated based on the set of normal vectors and the preset set of sphere kernel vectors to obtain the topological relationship features.

[0139] In one possible implementation, the feature aggregation module 53 is further used for:

[0140] For each patch, the distribution feature vector and structural feature vector are processed by multiple convolutions and then the features are aggregated to obtain the patch convolution features;

[0141] The geometric feature vector of the target model is determined based on the convolution features of each facet.

[0142] In one possible implementation, the task processing module 54 is further configured to:

[0143] Input the geometric feature vector into the feature transformation module to determine the physical feature vector corresponding to the target model;

[0144] Input the physical feature vector into the drag prediction module to obtain the drag coefficient corresponding to the target model;

[0145] Input the physical feature vector into the volume flow field prediction module to obtain the volume flow field corresponding to the target model.

[0146] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0147] It should be noted that, in the embodiments of this application... Figure 5 The module division of the high-fidelity pneumatic proxy modeling and optimization device shown is illustrative and represents only one logical functional division; in actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, exist as separate physical units, or have two or more units integrated into one unit. The integrated units can be implemented in hardware, as software functional units, or a combination of both.

[0148] It should be noted that, in the embodiments of this application, if the above-described methods are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, 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 an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0149] Figure 6 A schematic diagram of an electronic device according to an embodiment of this application is shown. For example... Figure 6 As shown in the figure, this application provides an electronic device, which can be a server, and its internal structure diagram can be as follows. Figure 6 As shown, the electronic device includes a processor 620, a memory, and a transceiver 640 connected via a system bus 610. The processor 620 provides computing and control capabilities. The memory includes a non-volatile storage medium 631 and internal memory 632. The non-volatile storage medium 631 stores an operating system, computer programs, and a database. The internal memory 632 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium 631. The database stores data. The transceiver 640 communicates with external terminals via a network connection. The computer program, when executed by the processor 620, implements the methods described above.

[0150] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor 620, implements the steps of the method provided in the above embodiments.

[0151] This application provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps in the method provided in the above-described method embodiments.

[0152] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0153] In one possible implementation, the shooting prompting device provided in this application can be implemented as a computer program, which can be configured as follows: Figure 6 The device operates on the electronic device shown. The memory of the electronic device can store the various program modules that make up the above-described apparatus. The computer program composed of the various program modules causes the processor 620 to execute the steps of the methods in the various embodiments of this application described in this specification.

[0154] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium, storage medium, and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0155] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" mentioned throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, phrases such as "in one possible implementation," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, they will not be repeated here.

[0156] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three kinds of relationships. For example, object A and / or object B can represent three situations: object A exists alone, object A and object B exist simultaneously, and object B exists alone.

[0157] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0158] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or modules can be electrical, mechanical, or other forms.

[0159] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0160] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0161] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0162] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, 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 an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0163] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0164] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0165] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0166] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A high-fidelity pneumatic proxy modeling and optimization method, characterized in that, The method includes: Using triangular facets as the basic modeling unit, the spatial and topological information of each facet in the target model is extracted. The spatial information includes the center coordinates, displacement vector, and unit normal vector, and the topological information includes the adjacent facet index. Spatial channel encoding is performed based on the spatial information of each patch to obtain the corresponding distribution feature vector; Structural feature encoding is performed based on the spatial and topological information of each patch to obtain the corresponding structural feature vector; Vector aggregation is performed on the distribution feature vector and the structural feature vector of each of the aforementioned facets to obtain the geometric feature vector of the target model; The geometric feature vector is input into the multi-task collaborative processing model obtained through training to obtain the drag coefficient and volumetric flow field corresponding to the target model.

2. The method according to claim 1, characterized in that, The spatial channel encoding based on the spatial information of each of the patches to obtain the corresponding distribution feature vector includes: The center coordinates of each patch are encoded using a multilayer perceptron to obtain the corresponding distribution feature vector.

3. The method according to claim 1, characterized in that, The structural feature encoding based on the spatial and topological information of each of the facets to obtain the corresponding structural feature vector includes: The geometric shape features of each facet are extracted based on the displacement vector of each facet. The topological relationship features of each facet are extracted based on its unit normal vector and topological information. Based on the geometric shape and topological characteristics of each facet, the corresponding structural feature vector is determined.

4. The method according to claim 3, characterized in that, The step of extracting the geometric shape features of the facets based on the displacement vectors of each facet includes: Based on the three-dimensional displacement vector of each of the aforementioned facets, determine the corresponding three two-dimensional sub-vectors; The local features corresponding to the three two-dimensional sub-vectors of each of the aforementioned facets are extracted respectively; The geometric shape features are obtained by fusing the three local features corresponding to each surface patch using a multilayer perceptron.

5. The method according to claim 3, characterized in that, The step of extracting the topological relationship features of the facets based on their unit normal vectors and topological information includes: The corresponding adjacent facets are determined based on the topological information of each facet; The set of normal vectors is determined based on the unit normal vector of each facet and the unit normal vector of the corresponding adjacent facet. For each facet, the spatial distribution similarity between the unit normal vectors of the facet is calculated based on the set of normal vectors and the preset set of sphere kernel vectors to obtain the topological relationship features.

6. The method according to claim 1, characterized in that, The step of aggregating the distribution feature vectors and structural feature vectors of each of the aforementioned facets to obtain the geometric feature vectors of the target model includes: For each of the patches, the distribution feature vector and the structural feature vector are subjected to multi-layer convolution processing and then feature aggregation is performed to obtain the patch convolution features; The geometric feature vector of the target model is determined based on the convolution features of each of the aforementioned facets.

7. The method according to claim 1, characterized in that, The step of inputting the geometric feature vector into the trained multi-task collaborative processing model to obtain the drag coefficient and volumetric flow field corresponding to the target model includes: The geometric feature vector is input into the feature conversion module to determine the physical feature vector corresponding to the target model; The physical feature vector is input into the resistance prediction module to obtain the resistance coefficient corresponding to the target model; The physical feature vector is input into the volume flow field prediction module to obtain the volume flow field corresponding to the target model.

8. A high-fidelity pneumatic proxy modeling and optimization device, characterized in that, The device includes: The feature extraction module is used to extract the spatial and topological information of each facet in the target model, using triangular facets as the basic modeling unit. The spatial information includes the center coordinates, displacement vector, and unit normal vector, and the topological information includes the adjacent facet index. The first encoding module is used to perform spatial channel encoding based on the spatial information of each of the facets to obtain the corresponding distribution feature vector; The second encoding module is used to encode structural features based on the spatial and topological information of each facet to obtain the corresponding structural feature vector. The feature aggregation module is used to perform vector aggregation on the distribution feature vector and the structural feature vector of each of the facets to obtain the geometric feature vector of the target model; The task processing module is used to input the geometric feature vector into the multi-task collaborative processing model trained to obtain the drag coefficient and volumetric flow field corresponding to the target model.

9. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images