A fully automatic method for generating unstructured grids

Abstract

The application discloses a kind of non-structure grid full automatic generation method, it is related to computational fluid dynamics pre-processing technical field.The application includes: the component segmentation is carried out to aircraft CAD model, obtains segmentation result and key geometric feature line;Geometric defect identification and repair are carried out to each subcomponent, generate component-level surface grid;Fusion geometric data, component semantic label and calculation working condition, utilize neural network to predict the grid area type and control parameter of boundary line, drive generation optimized surface grid;Based on optimized surface grid, generate boundary layer and three-dimensional body grid, assemble and output complete simulation model.Intelligent segmentation is realized by the combination of multi-modal AI and geometric rules, defect adaptive repair and grid strategy intelligent decision are realized by using graph neural network, and hybrid body grid containing pyramid transition unit is automatically generated, end-to-end full automatic generation from CAD model to simulation grid is realized, and the grid generation efficiency, quality and adaptability to flow field characteristics are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of computational fluid dynamics, and more specifically, to a fully automated method for generating unstructured meshes. Background Technology

[0002] Currently, aerodynamic simulation analysis of aircraft (such as airplanes, missiles, and drones) heavily relies on computational fluid dynamics (CFD) technology, and high-quality unstructured mesh generation is both a prerequisite and a bottleneck for CFD simulation. Traditional unstructured mesh generation processes typically include: CAD model cleaning, component segmentation, surface mesh generation, boundary layer generation, volume mesh filling, and assembly.

[0003] Existing technologies suffer from several drawbacks and pain points: High reliance on manual labor and low efficiency: Aircraft CAD models typically consist of thousands of curved surfaces and contain numerous subtle features (such as rivets, gaps, and hatches). Traditional methods heavily rely on engineers' experience for component identification, segmentation, and geometric repair, resulting in extremely long processing times and difficulty adapting to the demands of rapid design iterations. Difficulty in handling geometric defects: Original CAD models often contain topological and geometric defects such as gaps, overlaps, and fragmented surfaces. Existing automated tools have low recognition rates for complex defects, often requiring interactive manual repair, and lack automated repair strategies tailored to specific aircraft domain knowledge. Lack of intelligent mesh strategy formulation: The size distribution and anisotropic parameters (such as boundary layer densification direction) of surface meshes are typically set by engineers based on experience. The lack of an intelligent decision-making mechanism that integrates geometric features, component semantics (such as whether it's a wing or fuselage), and computational conditions (such as Mach number and angle of attack) leads to unstable mesh quality—either too dense, wasting computational power, or too sparse, affecting accuracy. Full-process automated breakpoints: Existing tools are often isolated (segmentation tools, repair tools, and mesh drawing tools are separate), data flow is not smooth, and it is difficult to achieve end-to-end fully automated generation from raw CAD to solveable mesh. Especially when dealing with multi-part assembly and transition areas (such as the junction of boundary layer and core flow field), manual intervention is often required. Summary of the Invention

[0004] The present invention aims to provide a fully automated method for generating unstructured meshes, in order to solve the problems of excessive manual intervention, low efficiency, and insufficient intelligence in the existing technology.

[0005] To achieve the above-mentioned objectives, this invention provides a fully automated method for generating unstructured meshes, the method comprising:

[0006] Step 1: Segment the input aircraft CAD model to be processed into components, obtaining the component segmentation results and corresponding key geometric feature lines; Step 2: Based on the component segmentation results, identify and repair geometric defects in each sub-component, generating a component-level surface mesh; Step 3: Based on the key geometric feature lines, determine the boundary lines of the sub-component surface; based on the component-level surface mesh, extract the geometric data of each boundary line; fuse the geometric data, the semantic labels of the sub-components, and the computational conditions, and use a trained neural network model to classify the boundary lines into mesh regions. The model classifies the grid regions to which each boundary line belongs, including isotropic and anisotropic regions. Based on the grid region type classification results, a differential parameter prediction subnetwork corresponding to each grid region type in the neural network model is used to predict the corresponding set of grid control parameters for each boundary line. The surface mesh is generated based on the set of grid control parameters to obtain an optimized surface mesh. Step 4: Based on the optimized surface mesh, a boundary layer mesh and a three-dimensional volume mesh are generated. The volume meshes of each sub-component are assembled and output to obtain a complete unstructured volume mesh model for simulation.

[0007] The specific objectives of this invention include: Achieving intelligent component segmentation: Utilizing a multimodal AI large-scale model to automatically identify the semantics of aircraft components, and combining geometric rules to complete minute features, replacing manual segmentation. Achieving automatic geometric defect repair: Based on deep learning, automatically identifying and classifying geometric defects, and using a divide-and-conquer strategy (topology first, geometry second) for automatic repair, generating high-quality surface meshes. Achieving intelligent mesh parameter prediction: Integrating geometric, semantic, and operational information, utilizing neural networks to adaptively predict the mesh type (isotropic / anisotropic) and control parameters for each region, optimizing mesh quality. Achieving full-process closed-loop automation: Establishing a complete link from CAD input to final volume mesh output, automatically handling boundary layer generation, transition mesh construction, and multi-component assembly, outputting simulation models compatible with mainstream solvers.

[0008] Existing mesh generation processes are fragmented, lacking a fully automated end-to-end solution from CAD model to final simulation mesh, resulting in high manual costs and long cycles. To address this issue, this method constructs a four-step automated process: automatically splitting the CAD model and extracting feature lines for intelligent segmentation; automatically identifying and repairing defects and generating component-level surface meshes for intelligent repair; utilizing AI to fuse multi-dimensional information (geometric + semantic + working condition) to predict mesh parameters and generate optimized surface meshes for intelligent decision-making; and automatically generating boundary layer and volume meshes and assembling them for automated volume mesh generation and assembly. This achieves a fully automated unstructured mesh generation process for aircraft without manual intervention, significantly shortening preprocessing time, ensuring consistency and standardization of mesh generation, and improving simulation preparation efficiency.

[0009] Preferably, the boundary layer mesh generation step in step 4 specifically includes: determining the surface region for which boundary layer generation is required based on the optimized surface mesh; matching or specifying boundary layer parameters according to preset simulation conditions, wherein the boundary layer parameters include at least the first layer mesh height, mesh growth rate, and number of growth layers; generating multi-layer prismatic meshes by stretching them layer by layer along the surface normal direction at the surface mesh nodes in the surface region to form the boundary layer mesh; performing quality checks and optimizations on the generated boundary layer mesh, wherein the quality checks include at least mesh distortion, aspect ratio, and orthogonality indices; and using the outer surface of the optimized boundary layer mesh as the input boundary for subsequent volume mesh generation.

[0010] In viscous flow simulation, the boundary layer mesh is crucial, but traditional methods suffer from cumbersome parameter settings and lack automated quality control after generation, easily leading to negative volume or high distortion. To address this issue, this method explicitly identifies the object surface region, employs parametric control (first-layer height, growth rate, number of layers), generates prismatic meshes using normal stretching, introduces quality checks and optimization steps (distortion, aspect ratio, orthogonality), and explicitly defines the output outer surface as the volume mesh boundary. This ensures high orthogonality and low distortion of the boundary layer mesh, which is key to accurately simulating wall friction resistance and heat exchange. Automated quality checks prevent invalid meshes from flowing into subsequent steps, improving the convergence of the final simulation.

[0011] Preferably, the method for generating the three-dimensional volume mesh in step 4 specifically includes: using the non-object surface regions in the optimized surface mesh and the outer surface of the boundary layer mesh to jointly form a closed solution domain boundary; within the closed solution domain boundary, calling the spatial mesh filling algorithm to generate a tetrahedral mesh to fill the core flow field region; at the interface between the boundary layer mesh and the tetrahedral mesh, generating a pyramidal transition mesh to achieve a smooth connection of mesh types; wherein, the tetrahedral mesh and the pyramidal transition mesh together constitute the three-dimensional volume mesh of each sub-component.

[0012] In hybrid mesh generation, mesh mismatch or abrupt quality changes can easily occur at the junction of the boundary layer (prism) and the core domain (tetrahedron), leading to solver divergence. To address this issue, this method clarifies the composition of the closed boundary (non-object surface + boundary layer outer surface), fills the core domain with a tetrahedral mesh, automatically generates a pyramidal transition mesh at the junction, and uses tetrahedrons and pyramids together to form the volume mesh. This achieves a smooth topological transition between different mesh types. The pyramidal mesh solves the node connection problem between the quadrilateral face (prism top face) and the triangular face (tetrahedron lateral face), eliminates dangling nodes, ensures mesh continuity and conservation, and improves the stability and accuracy of complex flow field calculations.

[0013] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0014] Full-process automation and high efficiency: It completely changes the traditional mesh generation mode that relies on manual experience, and realizes one-click fully automatic generation from dirty CAD model to net simulation mesh. It shortens the pre-processing work that originally required several days or even weeks to hours or even minutes, greatly improving the efficiency of aircraft R&D iteration.

[0015] Highly intelligent and adaptive: It innovatively integrates multimodal large models (vision + language), graph neural networks, and expert rule bases. The system can not only understand geometric shapes, but also comprehend component semantics and flight conditions, thereby adaptively generating optimal mesh topology and size distribution, solving the problem of accuracy loss or resource waste caused by the one-size-fits-all approach of traditional algorithms.

[0016] High robustness and fault tolerance: Through a dual segmentation mechanism of AI recognition and rule completion, and a hierarchical repair strategy of prioritizing topology repair and following up with geometric repair, the system effectively overcomes common quality problems in industrial-grade CAD models, such as missing data, gaps, and overlaps, and significantly improves the system's adaptability to complex and non-standard models.

[0017] Improved simulation accuracy and solution stability: The specially designed pyramid transition mesh generation mechanism and automatic boundary layer quality optimization process ensure that the generated unstructured mesh has extremely high quality indicators (orthogonality, smoothness) in key areas (such as boundary layers and component junctions), which directly improves the convergence speed of subsequent CFD solutions and the physical reliability of the calculation results.

[0018] Domain knowledge accumulation and reuse: This invention solidifies expert knowledge in the field of aircraft design (such as component definition rules, defect repair logic, and mesh strategy experience) into the algorithm model and rule base, so that the mesh generation process no longer depends on the individual level of a specific expert, realizing the standardized accumulation and large-scale reuse of enterprise-level simulation capabilities. Attached Figure Description

[0019] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0020] Figure 1 This is a flowchart illustrating a fully automated method for generating unstructured meshes. Detailed Implementation

[0021] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0023] Example 1;

[0024] Currently, there is a lack of an intelligent, integrated solution that can mimic the comprehensive decision-making capabilities of experts, connect the entire process of geometric repair, surface meshing, and volume meshing, and continuously accumulate and reuse organizational knowledge. This invention addresses this long-standing technical pain point by proposing a fully automated unstructured mesh generation method.

[0025] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a fully automated method for generating unstructured meshes. The method includes:

[0026] Step 1 - Intelligent Component Segmentation and Feature Extraction of CAD Model: Segment the input aircraft CAD model to be processed into components to obtain the component segmentation results and the corresponding key geometric feature lines;

[0027] Step 2 - Component-level geometric defect identification and adaptive repair: Based on the component segmentation results, geometric defects are identified and repaired for each sub-component in the component segmentation results to generate a component-level surface mesh;

[0028] Step 3 - Multi-source fusion mesh strategy prediction and surface mesh optimization: Based on the key geometric feature lines, determine the boundary lines of the sub-component surface; based on the component-level surface mesh, extract the geometric data of each boundary line; fuse the geometric data, the semantic labels of the sub-component, and the computational conditions, and use a trained neural network model to classify the mesh region types of each boundary line, obtaining the mesh region type classification results of each boundary line, including isotropic regions and anisotropic regions; based on the mesh region type classification results, use the differential parameter prediction sub-network in the neural network model corresponding to each mesh region type to predict the corresponding set of mesh control parameters for each boundary line; drive surface mesh generation based on the set of mesh control parameters to obtain the optimized surface mesh;

[0029] Step 4 - Automatic generation and global assembly output of hybrid body mesh: Based on the optimized surface mesh, generate boundary layer mesh and three-dimensional volume mesh, assemble and output the volume mesh of each sub-component to obtain a complete unstructured body mesh model for simulation.

[0030] This invention provides a fully automated method for generating unstructured meshes in the aerospace field. Its core principle lies in constructing an end-to-end integrated workflow that connects multiple intelligent modules and mature algorithms, thereby disrupting the traditional serial manual process. First, AI-based component segmentation technology is used to intelligently decompose complex assemblies, laying the foundation for parallel processing. Second, intelligent identification and repair of geometric defects are seamlessly integrated to automatically generate watertight surface meshes. Furthermore, intelligent prediction of the mesh size field is introduced to drive the optimized regeneration of the surface mesh, ensuring both geometric fit and physical adaptability. Finally, based on the optimized surface mesh, boundary layer insertion and spatial volume mesh generation are automatically performed. This invention deeply couples multiple AI decision points, such as segmentation, repair, and size prediction, with traditional mesh generation algorithms, achieving full-process automation from defective original CAD to high-quality simulated volume meshes. Expert experience is solidified into the system, significantly improving preprocessing efficiency and quality consistency.

[0031] (1) The complete process of the method includes: This method, from importing the original model to outputting the final volume mesh, includes the following steps in sequence:

[0032] 1. Intelligent Component Segmentation: The input original CAD assembly model from the aerospace field is automatically segmented using a trimodal fusion-based intelligent component identification and segmentation method for aircraft CAD models, resulting in a series of structurally and functionally independent sub-components. 2. Intelligent Geometric Repair and Surface Mesh Generation: For each sub-component, an AI-based large-model geometric defect identification and processing method is invoked to automatically clean the geometry and generate an initial, watertight component-level surface triangular mesh. 3. Intelligent Mesh Size Prediction and Regeneration: Based on the clean surface model (or initial mesh) obtained in step 2, a method for predicting the mesh node distribution and anisotropic parameters of aircraft component boundary lines is used, fusing geometric features, component semantics, and physical conditions to intelligently predict the global non-uniform size field. Using this size field as the control target, the component surface mesh is adaptively regenerated to obtain a high-quality surface mesh. 4. Automatic Boundary Layer Mesh Generation: On the regenerated component surface mesh, a high-quality multi-layer prismatic boundary layer mesh is automatically generated along the surface normal based on preset or automatically calculated viscous simulation parameters (such as the first layer mesh height, growth rate, and number of layers). 5. Automatic Generation and Output of Volume Mesh: Using the optimized surface mesh obtained in step 3 and the outer surface of the boundary layer mesh generated in step 4 as boundaries, a three-dimensional tetrahedral / pyramid volume mesh is automatically generated. Finally, the volume meshes of each component are combined and output as a complete unstructured volume mesh model that can be used for CAE analysis.

[0033] (2) Detailed introduction of each step: Step 1: Intelligent component segmentation: Receive the original aerospace CAD model (such as the whole aircraft model) that may contain defects. The system automatically identifies the geometric features in the model (such as surface continuity, assembly relationships) and understands its engineering semantics (such as fuselage, wings, horizontal stabilizer). According to predefined or learned rules, the complex assembly is decomposed into multiple sub-components. This step reduces the dimensionality of the subsequent processing from the entire complex model to multiple more regular components, which is a prerequisite for realizing parallel and intelligent processing. This step is the first link in starting the fully automatic process. Its output is a series of sub-component models, which will flow into the subsequent parallel processing pipeline as independent units.

[0034] Step 2: Geometric Intelligent Repair and Surface Mesh Generation: For each incoming sub-component, the engine first uses an AI model to automatically identify and precisely locate its internal geometric defects (cracks, holes, overlaps, etc.). Then, based on the defect type, it automatically calls the corresponding repair algorithm (such as crack fusion and hole filling) to complete phased repair at the B-Rep level and discrete level. After repair, a watertight surface triangular mesh for the component is automatically generated. This step is the core link in ensuring geometric health. It receives the defective component from Step 1 and outputs a clean surface mesh, providing a qualified input basis for all subsequent mesh-based operations. Its connection with Step 1 realizes the transition from divide-and-conquer to comprehensive management.

[0035] Step 3: Intelligent Mesh Size Prediction and Regeneration: For the clean part surface obtained in the previous step (or using its initial mesh as a medium), the engine extracts its geometric features (such as curvature, feature edge distance) and combines them with user-specified operating parameters (such as Mach number, angle of attack) and part semantic labels inherited from Step 1 (such as leading edge, control surface). These multimodal features are input into the trained AI prediction model to obtain a non-uniform target size field covering the entire part surface. Subsequently, the surface mesh generator is driven to re-divide or optimize the mesh using this size field as a constraint, ultimately generating a high-quality surface mesh with a density distribution that highly matches physical requirements and geometric features. This step is crucial for improving mesh intelligence. It upgrades mesh generation from passive geometric discretization to an active, physically driven adaptive process. It receives the geometric output from Step 2 and injects a size strategy into it; the resulting optimized surface mesh is the direct boundary for generating the boundary layer and volume mesh.

[0036] Step 4: Automatic Boundary Layer Mesh Generation: Based on the optimized surface mesh generated in Step 3, the module automatically recommends boundary layer parameters according to the simulation type (e.g., steady / unsteady, turbulent model) or allows the user to specify them. These boundary layer parameters are input into the GridStar mesh generation software, which automatically stretches multiple high-quality prismatic meshes along the normal direction at the boundaries of the object's surface mesh, ensuring that the near-wall resolution meets the requirements for viscous flow calculations. This step is a dedicated processing step for high-precision CFD simulations. It uses the high-quality surface mesh from Step 3 as a foundation to construct a mesh structure for accurately capturing boundary layer physical phenomena. Step 4 utilizes existing functions built into the GridStar mesh generation software, which will not be elaborated upon in this embodiment.

[0037] Step 5: Automatic Volume Mesh Generation and Output: The optimized surface mesh (non-object surface region) generated in Step 3 and the outer surface of the boundary layer mesh generated in Step 4 together form the closed solution domain boundary. The GridStar mesh generation software will call mature spatial mesh filling algorithms (such as the Delaunay advancing wavefront method) to automatically generate tetrahedral meshes within the boundary, and may generate pyramidal transition meshes at the junction of the boundary layer and the core flow field mesh. Finally, the system automatically assembles the volume meshes generated by each component and outputs a complete unstructured volume mesh file compatible with mainstream solver formats. This step is the final integration and output stage of the entire process. It brings together the results of all previous steps (clean surface, optimized dimensions, boundary layer), completes the filling from two-dimensional surface to three-dimensional space, and produces a final product that can be directly used for CAE simulation. Step 5 uses the existing functions built into the GridStar mesh generation software, and the corresponding details are not elaborated in this embodiment of the invention.

[0038] Based on the end-to-end fully automated integrated generation technology proposed in this invention, by organically integrating multiple advanced technologies such as intelligent component segmentation, geometric intelligent repair, and intelligent size prediction, and deeply coupling them with mature mesh generation algorithms, the following significant beneficial effects are achieved:

[0039] (1) Achieving complete automation and a leap in efficiency of the unstructured mesh preprocessing workflow: By adopting a complete workflow of sequential intelligent segmentation, intelligent repair, intelligent size prediction, and automatic volume mesh generation, the system can compress the traditional serial process that requires several days or even weeks and is highly dependent on manual intervention into an automated pipeline operation triggered by one click. This fundamentally eliminates the manual switching and waiting between each step in the process (geometric cleaning, size setting, volume mesh generation), completely liberating engineers from tedious and repetitive labor, reducing preprocessing time by more than an order of magnitude, and significantly compressing the overall simulation cycle.

[0040] (2) Systematically integrating and solidifying the experience of experts from multiple fields into reusable digital assets: By employing three intelligent modules—AI for component segmentation, AI for geometric defect identification and repair, and AI for mesh size prediction—the system can encode, integrate, and solidify the tacit knowledge about model decomposition, geometric problem diagnosis, and mesh strategy formulation that was originally scattered among different experts. This allows the organization's best practices to be transformed into a stable and iteratively upgradeable digital expert system, ensuring that regardless of who operates it, it can output meshes that meet the highest standards and have a consistent style. This greatly reduces the reliance of CAE applications on specific senior experts and achieves effective knowledge transfer and rapid replication of team capabilities.

[0041] (3) Achieving intelligent mesh generation from geometry-driven to physics-semantic joint-driven: By adopting the technical means of introducing AI size prediction that integrates component semantics and operating parameters to drive surface mesh regeneration on the automatically repaired geometry, the density distribution of the mesh generated by the system no longer passively responds to the geometry, but can actively adapt to the physical requirements of the simulation. For example, it automatically densifies key physical parts such as the leading edge of the wing and the expected shock wave area, and reasonably coarsens in non-interest areas, thereby intelligently balancing computational accuracy and efficiency in the automated process and generating a highly adaptable mesh that is truly oriented towards the simulation target.

[0042] (4) Ensuring the consistency and reliability of mesh quality throughout the entire process, bridging the last mile of CAE simulation: Due to the adoption of a rigorous cascaded process design that first ensures geometric watertightness, then optimizes the size strategy, and finally generates the volume mesh, each downstream step is built on a reliable foundation that has been rigorously processed by the upstream intelligent module. This fundamentally eliminates the risk of downstream volume mesh failure or poor quality due to hidden upstream geometric problems or improper size settings in traditional processes, forming a robust closed loop of self-verification and step-by-step optimization, ultimately outputting a high-quality volume mesh that can be directly used for high-confidence CAE analysis, significantly improving the success rate of simulation work and the credibility of results.

[0043] Example 2;

[0044] Based on Embodiment 1, Embodiment 2 of the present invention provides a detailed description of the component identification and segmentation process:

[0045] Example 2 can solve the problems of low efficiency, poor consistency and automation breakpoints caused by relying on manual component identification in the aircraft mesh generation process. Example 2 provides a method that can automatically and accurately identify aircraft components and directly output semantic grouping and feature geometry that meet the requirements of mesh generation engineering, so as to realize full-process automation from geometric model to computational mesh.

[0046] Example 2 provides a method for intelligent component recognition and segmentation of aircraft CAD models based on three-modal fusion. The method includes: Step S1: Obtaining an aircraft CAD model to be processed, the aircraft CAD model having a boundary representation structure; Step S2: Based on the engineering requirements of mesh generation, performing engineering semantic annotation on the faces in the boundary representation structure of the aircraft CAD model to obtain component semantic labels; extracting the geometric parameters of each face from the boundary representation structure; generating a multi-view two-dimensional rendering image for each annotated face; associating and storing the multi-view two-dimensional rendering image, the geometric parameters, and the component semantic labels to form training samples; Step S3: Training a multimodal recognition model using the training samples, the multimodal recognition model being configured to fuse input image features and language semantic information, and outputting the faces in the aircraft CAD model. Component semantic labeling; Step S4: Input the aircraft CAD model to be identified into the trained multimodal recognition model, identify the faces in the aircraft CAD model, and obtain a preliminary face-component semantic label mapping relationship; Step S5: Based on the preliminary face-component semantic label mapping relationship, for faces whose recognition confidence output by the multimodal recognition model is lower than a preset threshold or whose geometric area ratio is less than a set threshold, activate the geometric rule auxiliary module to perform semantic completion and obtain the final component segmentation result; Step S6: According to the final component segmentation result, extract the key geometric feature lines corresponding to each component from the boundary representation structure of the aircraft CAD model; Step S7: Convert the final component segmentation result and the extracted key geometric feature lines into a structured engineering semantic format for target mesh generation software to read and utilize, and output it.

[0047] In Example 2, step S1 limits the processing object to a CAD model with a boundary representation structure, ensuring the feasibility of all subsequent geometric operations (topology analysis, parameter extraction). Steps S2-S3 utilize steps S2-S3 to construct a multimodal dataset oriented towards mesh engineering semantics and train a specialized model integrating visual, geometric, and linguistic information, enabling the method to understand engineering semantics and replacing manual identification by engineers. Steps S4-S5 employ a collaborative mechanism of AI recognition of main structural components and small feature geometric rule completion, utilizing AI to handle macroscopic patterns while using geometric rules to ensure detailed integrity, overcoming the limitations of single methods on complex engineering geometries. Steps S6-S7 automatically connect the semantic recognition results to the geometric kernel, extracting key geometric feature lines and converting them into a format directly readable by mesh software, thus completing the final step from recognition to application and eliminating manual conversion. This solves the end-to-end automation bottleneck problem in the aircraft CAD mesh generation process, where component recognition relies entirely on manual labor, resulting in extremely low efficiency, inconsistent results, and the inability to directly use the recognition results for subsequent engineering. It achieves a fully automated, high-precision processing flow from the original CAD model to mesh-generated ready engineering data, reducing the time required for manual work, which can take several weeks, to automated processing.

[0048] In step S2, the surfaces in the boundary representation structure of the aircraft CAD model are annotated with engineering semantics. Specifically, in the mesh preprocessing software environment, semantic labels are manually assigned to the surfaces contained in different components based on aerodynamic characteristics or structural functional differences. The geometric rule auxiliary module, based on the geometric topological relationships and parametric feature determination rules of the aircraft CAD model, determines and corrects the component semantic labels of the surfaces, forming the final component segmentation result.

[0049] Purely data-driven AI models are prone to low confidence or misjudgment when faced with small-sized, weakly textured, feature-fuzzy, or geometrically rare surfaces in the training data (such as slender wing trailing edges, tiny connecting protrusions, and hatches that smoothly transition into the main body). The above approach addresses this by introducing rule modules based on deterministic geometric knowledge to compensate for AI's shortcomings in such cases. General image segmentation models output pixel-level labels, which are difficult to automatically and error-free associate with the precise boundary representation (B-Rep) structure (faces, edges, vertices, and their topology) in CAD models. The above approach directly limits operations to the geometric and topological levels of the original CAD model, ensuring that any judgments and corrections are precisely applied to specific B-Rep surfaces, guaranteeing absolute geometric accuracy. By filling in gaps and correcting errors in the AI ​​recognition results, the final component segmentation results ensure that all geometric details affecting subsequent mesh generation and simulation analysis are covered, reducing the degradation of mesh quality or simulation accuracy caused by missing key small features. This approach eliminates reliance on a single AI model, creating a hybrid intelligence model that combines data-driven (AI) and knowledge-driven (rule-based) approaches. When faced with new aircraft models or those with unique geometries, even if the AI ​​model performs poorly initially, the method can still ensure the correct identification of basic and critical components through geometric rules, thus enhancing the method's generalization ability and practical stability.

[0050] The multi-view 2D rendered images are generated as follows: for each surface to be processed, multi-angle rendering is performed along its normal vector, and rendering is performed at each viewpoint to generate a series of 2D view images. Specifically, for each surface, multi-angle rendering is performed along its normal vector. This simulates the process of observing a geometric surface from various directions, ensuring that the generated images fully reflect the shape characteristics of the surface. This solves the problem of how to generate richly informative 2D image data from a single, planar aircraft CAD model surface, which can be learned by the visual model. It provides the visual model with multi-angle, comprehensive observation data, enhancing the model's robustness in recognizing component surfaces under different viewpoints and lighting conditions.

[0051] Preferably, the training method for the multimodal recognition model is as follows: A neural network architecture comprising a visual encoder, a language understanding module, and a cross-modal attention fusion module is constructed. First, the visual encoding module, based on Vision Transformer (ViT), is responsible for extracting features from the input image and outputting a visual token. Second, the language encoding module, based on a pre-trained large model, is responsible for encoding text prompts into text tokens. Finally, the cross-modal fusion module achieves bidirectional deep interaction between the visual and language tokens through a cross-attention mechanism. After multiple rounds of interaction, the features of the two modalities are aligned and fused in a unified semantic space to form a joint feature representation. The neural network architecture is trained under supervised conditions using multi-view 2D rendered images and corresponding component semantic label text from the training samples as input. A combined loss function is used to optimize the multimodal recognition model, including a classification cross-entropy loss function, a segmentation intersection-union loss function, and a boundary-weighted loss function. The architecture of visual encoder + language module + cross-modal attention fusion module clearly provides a technical framework for fusing multimodal information. Supervised learning is performed using a pre-constructed dataset, enabling the model to learn the mapping from images to semantic labels. By employing a combined loss function, classification accuracy, segmentation region consistency, and boundary precision are simultaneously optimized, ensuring model output quality from multiple dimensions. This addresses the core technical path of model implementation: how to specifically construct and train a dedicated recognition model capable of understanding image-semantic associations. An optimized and trained intelligent model capable of accurately associating surface images with component semantic labels has been implemented.

[0052] Specifically, step S4 involves identifying the main structural component surfaces in the aircraft CAD model; step S5 involves semantic completion of local feature surfaces, connecting area surfaces, or low-confidence identification surfaces in the multimodal recognition model. Steps S4 and S5 are defined as a hierarchical collaborative process: the first layer (AI) is responsible for the main structural components (efficiently handling regularities); the second layer (rules) is responsible for complex situations such as local features and connecting surfaces (accurately handling special cases). This design concept is an efficient way to solve engineering complexity problems. It achieves intelligent division of labor and collaboration in the recognition process, significantly improving the completeness of identifying small and special features while ensuring the efficiency of main structural component recognition.

[0053] The rules for determining the geometric topological relationships and parametric features of the aircraft CAD model include at least one of the following rules: Rule 1: If a surface is located at the geometric boundary of two identified main structural component surfaces, and the area of ​​this surface relative to the area of ​​the adjacent main structural component surface is less than a set threshold, then the semantic label of this surface is determined as a connecting feature surface or an edge surface; Rule 2: If a surface is directly adjacent to the main wing, and the angle between its normal vector and the normal vector of the main wing reference plane is greater than or equal to 45°, and this surface is located at the outermost spanwise direction of the wing and its area is between 0.5% and 5% of the main wing area, then it is classified as a winglet or wingtip. The main structural components are then identified. The area ratio rule can identify small surfaces (such as edges) at the boundary. The direction relationship rule can identify small surfaces (such as doors) that are not parallel to the main flow direction. This provides quantifiable and judicious rule examples, enabling the system to automatically identify specific types of geometric features such as the wing trailing edge and doors.

[0054] The extraction of key geometric feature lines corresponding to each component includes: for components identified by the multimodal recognition model, key geometric feature lines, including boundary lines and feature contour lines, are extracted using geometric algorithms by analyzing the curvature distribution and boundary topology of their constituent surfaces; for feature surfaces identified by the geometric rule-assisted module, corresponding key geometric feature lines are extracted based on their geometric definitions and relationships with adjacent surfaces. Specifically, for the main structural components identified by AI, geometric algorithms such as curvature and topology analysis are used to automatically extract their boundaries and contours. Small features identified by rules are extracted directly based on their geometric definitions (such as the intersection of two surfaces). Both methods ensure that the output feature lines are accurately associated with the semantic labels of the components. This achieves automated connection from semantic segmentation to geometric feature extraction, outputting geometric data directly dependent on operations such as mesh encryption and block division.

[0055] In step S5, before or after forming the final component segmentation result, a post-processing step is also included: smoothing the semantic labels of the components based on the adjacency relationship of the surfaces to integrate the recognition results of the multimodal recognition model with the supplementary results of the geometric rule auxiliary module, and eliminating misjudgments of isolated surfaces; and visualizing and exporting the final component segmentation result.

[0056] The label smoothing process is based on the adjacency relationship of faces, performing consistency checks and adjustments on the labels of adjacent faces to eliminate unreasonable abrupt changes or isolated points. The AI ​​results are fused with rule-based results, and visualization and export are supported, forming a complete closed loop from processing to delivery. Post-processing improves the topological consistency and visual smoothness of the segmentation results, and visualization and export functions make the results easy for engineers to verify and use.

[0057] The input to the multimodal recognition model includes natural language instructions; the method allows users to specify the type of part to be recognized or the segmentation requirements by inputting natural language instructions; the visual encoder adopts a Transformer-based architecture, the language understanding module adopts a pre-trained large language model, and the cross-modal attention fusion module adopts a cross-attention mechanism.

[0058] The model's input can include natural language commands, which are then translated into internal tasks that drive the model and geometry engine, achieving intelligent conversion from natural language to concrete operations. This enhances the system's usability and flexibility, allowing users to drive the system to perform specific tasks using natural language (such as segmenting all wing surfaces). The key modules in the architecture are concretized into corresponding selections, ensuring the feasibility of the solution: a visual encoder (ViT or other Transformer architectures), a language understanding module (QwenLM or other large models), and a fusion module (CrossAttn), ensuring the powerful capabilities of each module.

[0059] In step S7, the structured engineering semantic format includes: a set of component geometric groups for the mesh generation software to recognize and call, a set of feature lines describing the key geometric features of the components, and optional mesh generation strategy parameters associated with preset components. By explicitly defining the output format as the structured engineering semantic format required by industrial software, including the component group set, feature line set, and mesh strategy parameters, the output of this method is no longer merely a display result, but rather instructions or data that can be directly read and applied by CAE software, thus achieving a closed loop of engineering value.

[0060] The technical solution in Example 1 achieves full automation of component identification, greatly improving the efficiency of mesh preprocessing. Because this invention employs a visual-geometric-linguistic three-modal fusion recognition and hierarchical collaborative processing mechanism, it can automatically and accurately complete the identification and semantic segmentation of various aircraft components. The effect is that it shortens the process of identifying medium-sized passenger aircraft components, which originally relied entirely on manual identification and grouping by engineers and took 1-2 weeks, to automated processing, solving the primary efficiency bottleneck in the mesh generation process.

[0061] The technical solution of Embodiment 1 can output engineering semantics and geometric features that can directly drive mesh generation. Since the output of this invention is not a general image segmentation mask, but rather a face-part label mapping table and key geometric feature lines precisely associated with the original B-Rep model, the effect is that the recognition results can be directly converted into components, sets, or layers that mesh generation software can recognize, and key feature lines (such as wing leading / trailing edge lines) used for mesh partitioning and densification control are automatically extracted, achieving recognition-readiness and eliminating the large amount of manual conversion and interpretation work required in traditional methods.

[0062] The technical solution in Example 1 improves the completeness of identification and engineering practicality for small-sized and weakly featured components. Because this invention employs a layered strategy of AI recognition of main structural components + geometric completion of small features, it intelligently completes small or indistinctly featured components such as wing trailing edges and access hatches using geometric rules. The effect is a significant improvement in the completeness of the overall segmentation, ensuring that subsequent mesh generation covers all necessary geometric details, making the generated mesh more consistent with the geometric integrity requirements of high-fidelity simulation.

[0063] The technical solution in Embodiment 1 can support intelligent mesh strategy pre-setting through semantic understanding, thereby improving mesh quality. Because the model training of this invention integrates mesh engineering knowledge, it possesses engineering semantic understanding capabilities. The effect is that the system can automatically recommend or apply pre-set mesh generation strategies based on component semantics (such as the leading edge of an airfoil), thereby generating higher-quality meshes that are more suitable for computation in key aerodynamic regions.

[0064] The technical solution in Example 1 possesses strong domain generalization capabilities and knowledge accumulation value. Because the core framework of this invention does not depend on a specific aircraft model and supports incremental learning, the effect is that it can quickly adapt to new aircraft models by only requiring annotation and fine-tuning of a small amount of data from new aircraft models, greatly reducing technical maintenance and expansion costs. Simultaneously, the system can accumulate the mesh partitioning experience of senior engineers into reusable models and rules, forming core knowledge assets.

[0065] This invention provides a method for intelligent component recognition and segmentation of aircraft CAD models based on three-modal fusion, the method comprising:

[0066] Step S1 - Model acquisition and format confirmation; Step S2 - Construction of multimodal dataset for engineering semantics; Step S3 - Training of multimodal fusion recognition model; Step S4 - Initial identification of macroscopic components based on AI; Step S5 - Collaborative completion of small features based on geometric rules; Step S6 - Automatic extraction of key geometric feature lines; Step S7 - Structured engineering semantic output.

[0067] The technical solution of Example 2 provides fully automated and high-precision component identification and geometric feature extraction capabilities for mesh generation processes in computational fluid dynamics and finite element analysis, and is applicable to the digital simulation and manufacturing of complex equipment such as aviation, aerospace, and ships.

[0068] The technical solution of Embodiment 2 provides a method for intelligent recognition, semantic segmentation, and geometric feature extraction of aircraft components based on a three-modal fusion of vision, geometry, and language. The technical solution of Embodiment 2 addresses the problems of low efficiency, poor consistency, and automation breakpoints caused by reliance on manual component recognition in the aircraft mesh generation process. Specifically, it includes:

[0069] (1) Eliminate the manual dependence of component identification and realize automatic and accurate identification and semantic grouping of each component in the aircraft CAD model; (2) Provide engineering semantic output that can be directly used for mesh generation, including component face sets and key feature lines (such as boundary lines and densification lines); (3) Support the preset of differentiated mesh strategies for different components and realize intelligent mesh planning that is configured upon identification; (4) Improve the automation level and processing efficiency of the entire mesh generation process and provide reliable pre-processing tools for digital simulation.

[0070] The technical solution in Example 2 provides a joint modeling and processing method for the visual-geometric-linguistic three modalities. Its overall architecture includes four parts: data construction, model training, inference and recognition, geometric processing, and mesh strategy generation. The specific steps are as follows:

[0071] (1) Data Construction: Semantic annotation system for grid requirements: 1) B-Rep surface-level semantic annotation: In grid generation software such as NNW-GridStar, combined with the experience of aerodynamic grid and structural grid partitioning, surface-level annotation is performed on the aircraft CAD model represented in B-Rep format, including but not limited to: nose, forward fuselage, mid-fuselage, rear fuselage, upper wing surface, lower wing surface, vertical tail, horizontal tail, engine nacelle, etc. After annotation, each component is visualized with different colors for easy verification. 2) Multi-view image generation: For each B-Rep surface, 360-degree multi-view rendering is performed around its normal vector to generate a series of main view images. Scaling, rotation, and lighting adjustment are supported to enhance view coverage and lighting robustness. 3) Dataset enhancement and construction: Geometric transformation (rotation, translation, scaling), color perturbation, noise injection, etc. are used to enhance data diversity. A structured multimodal dataset is constructed, and each data includes: B-Rep surface ID, multi-view image set, semantic label, and original geometric parameters.

[0072] (2) Model Training: 1) Model Architecture Design: ViT (Vision Transformer) is used as the visual encoder to extract image features. CrossAttn (cross-attention mechanism) is used to align visual features with language features (from QwenLM). QwenLM, as the language understanding and generation module, is responsible for semantic mapping and prompt response. 2) Supervised Fine-tuning Strategy: The input is the main view image of the face, the corresponding semantic label, and the structured prompt words (e.g., the face belongs to [category]). 3) The loss function adopts a ternary combination: semantic cross-entropy loss (optimizes classification accuracy) + IoU loss (improves the overlap of segmented regions) + boundary weighted loss (strengthens the accuracy of boundary pixel recognition); the loss function is:

[0073] ;

[0074] in, For semantic cross-entropy loss, For IoU loss, For boundary-weighted loss, the weighting coefficients are initially set. , , Adjustments can be made based on subsequent experiments.

[0075] Semantic cross-entropy loss:

[0076] ;

[0077] In the formula, N is the total number of pixels, i.e., the total number of faces in the B-Rep model, C is the number of semantic categories of the component, and y i,c Let p be the true label of the i-th pixel belonging to category c. i,c Let be the probability predicted by the model that the i-th pixel belongs to category c.

[0078] The IoU loss is calculated by averaging the IoU for each part category:

[0079] ;

[0080] Boundary weighted loss ( The initial setting is 0.3 (which can be adjusted based on experiments), where, L cls For boundary binary classification loss, L dist Loss for distance transformation:

[0081] ;

[0082] In the formula, This represents the true boundary label of the i-th pixel. This indicates that the pixel is located inside the component. This indicates that the pixel is located on the geometric boundary of two different components; This represents the probability that a pixel predicted by the model is a boundary. Based on the topological relationships of the B-Rep model, when two adjacent faces belong to different components, their shared edge is marked as a boundary.

[0083] Loss for distance transformation:

[0084] ;

[0085] In the formula, The total number of boundary pixels, This represents the distance from pixel i to the nearest real geometric boundary. This represents the distance from pixel i predicted by the model to the boundary. This forces the predicted boundary lines to be closer to the actual geometric boundaries, improving the geometric usability of the segmentation results in CAD models.

[0086] 4) The training employs progressive fine-tuning, first pre-training on large-scale general image-text data, and then fine-tuning on domain image-text pairs labeled with semantic tags for each component of the aircraft. This approach can utilize prior knowledge of general visual language while enabling the model to accurately grasp the fine-grained semantics of the aircraft's unique components, thus improving domain adaptability and recognition accuracy.

[0087] (3) Reasoning and Recognition: 1) Input: The aircraft CAD model to be processed (B-Rep) and natural language prompts (e.g., please recognize all fuselage surfaces). 2) Layered recognition processing flow:

[0088] The first layer is the overall component recognition: the model extracts visual features face by face, combines them with the semantics of prompt words to make inferences, and prioritizes the identification of the main structural components, i.e., large components. The main structural components refer to the components with independent functions on the aircraft, such as the entire wing, the entire fuselage, and the tail. For simplicity, they will be referred to as large components in the following descriptions. The primary semantic labels of each B-Rep face are output to form a face-component mapping table. The second layer, geometric completion of small parts / features: For small surfaces (such as surfaces belonging to the wing trailing edge or local protrusions) whose confidence level was lower than the set threshold or ignored by the model in the previous round of recognition, the AI ​​model is not directly relied on for secondary recognition. Instead, the geometric rule auxiliary module is activated. Small parts / features mainly include the following types: connecting feature surfaces or edge surfaces: surfaces located at the junction of two large parts (such as the upper and lower surfaces of the wing), such as the wing trailing edge and leading edge; feature surfaces with opening, closing or maintenance functions: surfaces surrounded by large parts and whose orientation is significantly different from the main body, such as hatches and maintenance access covers; local protrusions or depressions: such as rivet heads, sensor mounting bases, refueling ports and other small geometric features; low-confidence recognition surfaces: surfaces that the AI ​​model cannot recognize with high confidence due to reasons such as blurred geometric features or insufficient training data coverage. 3) Working principle of the geometric rule auxiliary module: Input: the set of large parts recognized by the AI ​​(such as the upper surface of the left wing and the lower surface of the left wing) and all the B-Rep surfaces contained therein.

[0089] Processing: Based on the geometric topology and parametric information of the CAD model, preset geometric rules are executed. For example, Rule 1: If a surface belongs to the upper or lower surface of the wing, or is located at the boundary between the two, at the leading edge of the wing, has high curvature, and its spanwise length is close to the entire wing span, it is classified as the leading edge of the wing. Rule 2: If a surface is located at the boundary between the upper and lower surfaces of the wing, and its area is less than 1% of the overall wing area, it is classified as the trailing edge of the wing. Rule 3: If a surface is directly adjacent to the main wing but its normal is significantly deviated from the main wing plane, and it is located at the outermost point of the wing spanwise, with an area between 0.5% and 5% of the main wing area, it is classified as a winglet or wingtip. Rule 4: If a surface is surrounded by a large fuselage component, and its normal is approximately perpendicular to the main fuselage direction, and its area is small, it may be an access hatch or door, and further determination is made based on the geometric characteristics of its adjacent surfaces. Rule 5: If a face belongs to a large engine nacelle component, is located at the foremost end of the nacelle, and has a closed annular boundary and a smooth Gaussian positive curvature transition, it is classified as an air intake lip. Output: Supplement the annotations of smaller faces that conform to the geometric rules to the corresponding component categories, or create new subcategory labels for them. 4) Post-processing: Smooth the labels based on face adjacency relationships, integrate AI recognition results with geometric rule supplementation results, and avoid misjudgment of isolated faces. Supports component-level result visualization and export.

[0090] (4) Geometric Processing: 1) Key Geometric Feature Line Extraction: Based on the complete identification of all components (including large components identified by AI and small features supplemented by geometric rules), the boundary lines, curvature extremum lines, and feature contour lines of each component are extracted based on B-Rep topology and geometric information. For large components identified by AI, geometric algorithms (such as edge detection based on curvature and topological connection analysis) are used to achieve automated extraction. For small features identified by the geometric rule auxiliary module, their geometric definitions (such as the trailing edge line, i.e., the intersection line of the upper and lower wing surfaces) are directly used to extract feature lines, resulting in higher accuracy. 2) Output Interface: Feature lines are output in a format supported by NNW-GridStar. Seamless integration with the simulation process is supported, improving CAE analysis efficiency. The visual encoder (ViT) can be replaced by other mainstream networks (such as ConvNeXt, ResNet), and this embodiment of the invention does not impose any restrictions on it. The language model (QwenLM) can be replaced by other large language models (such as GLM, LLaMA), and this embodiment of the invention does not impose any restrictions on it. The specific judgment conditions and parameters of the geometric rules can be adjusted according to the characteristics of different components. The output interface can be adapted to different target mesh generation software (such as ANSA, HyperMesh).

[0091] Example 3;

[0092] Regarding the geometric defect identification and processing in Embodiment 1, Embodiment 3 of the present invention will be specifically described: Embodiment 3 of the present invention provides a geometric defect identification and processing method that can intelligently, accurately and efficiently identify and process geometric defects, so as to break through the limitations of traditional methods, significantly improve the level of pre-processing automation, and shorten the research and development cycle.

[0093] After acquiring the CAD model of the aircraft to be processed and dividing the CAD model into multiple parts, Embodiment 3 of the present invention provides a method for geometric defect identification and processing based on a large AI model, the method comprising:

[0094] Step a1: Obtain multiple labeled sample components, and convert the geometric data of each sample component into a structured feature representation for processing by a deep learning model to obtain a training set. The structured feature representation includes at least one of the graph structure, point cloud, and feature vector of the sample component. The labeling includes the type and spatial location information of the geometric defects present in the sample component. Step a2: Train a multi-task hierarchical graph neural network model using the training set to obtain the trained multi-task hierarchical graph neural network model. Step a3: For any segmented component, convert the geometric data of the component into the same structured feature representation as in step a1, and input the converted data into the trained multi-task hierarchical graph neural network model. The trained multi-task hierarchical graph neural network model outputs the type and spatial location information of the geometric defects contained in the component. Step a4: Based on the type and spatial location information of the geometric defects output in step a3, firstly, execute a topology repair algorithm for defects identified as gaps to restore the topological connectivity at that location; then, execute the corresponding geometric repair algorithm for defects identified as geometric anomalies, and perform mesh stitching on the repaired results of all components to generate the final surface mesh.

[0095] While artificial intelligence technologies, particularly large language models (LLM) and visual large models (VLM), have made progress in multimodal understanding, their direct application in 3D geometry processing still faces challenges, including: a lack of domain-specific large-scale labeled datasets; insufficient support for engineering-level accuracy in existing models; difficulty in effectively integrating prior knowledge from fields such as aerodynamics and manufacturing; and poor interpretability of model decisions leading to low engineer trust. This method first intelligently segments complex CAD assembly models into more regular sub-components; then, it trains a multi-task hierarchical graph neural network (MT-HGNN) that incorporates domain knowledge to automatically identify and evaluate geometric defects in each sub-component; finally, based on the identification results, it drives a phased repair algorithm process following the principle of topology first, then geometry, ultimately generating a watertight overall surface mesh. This solves the problems of low efficiency and insufficient accuracy in geometric defect identification caused by the reliance on human experience in traditional methods, achieving automated, intelligent identification and precise location of various defects in complex CAD models. Overcoming the limitations of traditional algorithms that are based on simple rules and lack semantic understanding, this approach enhances the engineering relevance and accuracy of the identification process by encoding aerospace expertise (such as aerodynamic sensitivity and manufacturing constraints) into the AI ​​model. A closed-loop automated process from intelligent defect identification to targeted repair has been established. Through a phased repair strategy that prioritizes topology over geometry, the identification results are efficiently and reliably transformed into high-quality simulation mesh models, freeing engineers from tedious manual labor and improving the overall efficiency and quality of digital simulation preprocessing.

[0096] This invention, in its third embodiment, pioneers a closed-loop paradigm of divide-and-conquer, identification, and repair. Its principles are: 1) Divide and Conquer: Reducing the dimensionality of complex problems. By segmenting components, the disordered and diverse defects at the assembly level are transformed into relatively regular and finite defect patterns within the component level, creating conditions for AI identification. 2) Intelligent Identification: Datafiing expert experience. By training a dedicated AI model (MT-HGNN), the model learns the mapping from component geometry to defects, mimicking and surpassing the identification and judgment capabilities of the human eye and brain. 3) Tiered Repair: Operationalizing the identification results. Based on the structured defect report output by the AI, a repair algorithm is driven to execute in a strict order (topology first, then geometry), transforming intelligent judgment into deterministic geometric operations, ultimately outputting a high-quality mesh that can be directly used for simulation.

[0097] In step a1, the structured features are represented as a graph structure; where the faces of the component are treated as nodes of the graph, and the shared edges between faces are treated as edges of the graph; node features include the geometric attributes of the faces, and edge features include the geometric attributes of the shared edges. This scheme effectively transforms unstructured CAD boundary representation (B-Rep) data into structured input suitable for deep learning model processing, providing a high-fidelity, information-rich geometric data representation method. It completely maps the faces, edges, and their relationships and attributes of a CAD model into an attribute graph, enabling graph neural networks (GNNs) to directly reason on the model's topology and geometry.

[0098] In step a2, the multi-task hierarchical graph neural network model includes: a multi-scale feature extraction layer for encoding geometric features at the node, edge, and subgraph levels; a hierarchical graph pooling and attention layer for hierarchical clustering of faces in different regions and assigning attention weights; and a multi-task output head, which includes parallel gap recognition, fragmented surface recognition, hole boundary recognition, and defect severity assessment heads. This solution addresses the problem that existing general-purpose neural network models struggle to simultaneously and accurately identify multiple types of geometric defects and assess their engineering impact. A single model cannot handle tasks of different granularities, such as gaps (edge ​​level), fragmented surfaces (face level), and holes (ring level). A multi-functional, high-precision AI recognition core specifically for geometric defect analysis has been constructed. This model can output the recognition results and severity assessments of multiple defects in parallel, with recognition accuracy and engineering practicality far exceeding algorithms based on simple rules. The multi-scale feature extraction layer extracts features from three levels: nodes (faces), edges (connections), and subgraphs (local regions), ensuring the model can perceive geometric information from micro to macro levels. Hierarchical graph pooling and attention: Through clustering and attention mechanisms, the model can adaptively focus on feature regions at different scales, such as simultaneously focusing on micrometer-level slits and defect clusters in high-curvature regions. Multi-task output heads: Dedicated output heads are designed for different defect types, allowing the network to learn optimal discriminative features for different tasks while sharing the underlying feature extractor. This design enables different defect type recognition tasks to mutually reinforce each other, sharing learned geometric context knowledge, thereby improving overall performance.

[0099] In the hierarchical graph pooling and attention layer, a domain-knowledge-based attention weight guidance mechanism is introduced. Different prior attention weight coefficients are assigned to the facets based on their location within aerodynamically critical regions or manufacturing process regions. During the computation of the neural network attention mechanism, attention weights are not entirely data-driven, but rather a prior attention bias is applied to facets with different functional attributes based on a predefined domain knowledge rule base (e.g., if the facet belongs to the wing leading edge, the weight coefficient increases by 20%). This is equivalent to injecting engineer experience into the model from the early stages of training, guiding the model to learn patterns in key regions more quickly, thereby achieving higher recognition sensitivity and accuracy in these regions. This solves the problem that AI models treat all defects equally when identifying them, failing to prioritize key aerodynamic or process-constrained regions that have a greater impact on the final simulation results; that is, the model lacks an engineering perspective for priority judgment. It gives the AI ​​model an expert's perspective, enabling it to identify defects in key regions more quickly and accurately. This optimizes the allocation of computational resources and ensures that defects with the greatest impact on simulation quality are discovered and addressed first.

[0100] In training the multi-task hierarchical graph neural network model, a weighted multi-task loss function is used. In this function, the weight of the loss term for each defect category is inversely proportional to the number of samples of that defect category in the training set. During backpropagation optimization, a larger weight (inversely proportional to the number of samples) is assigned to the loss term of defect categories with fewer samples, amplifying the penalty for incorrect classification of these categories on the model parameters. This forces the model to invest more attention in learning how to correctly identify these rare defect samples during training, thus correcting its natural bias towards the majority class and achieving a balanced improvement in the performance of various defect identification types. This solves the model training bias problem caused by the extreme imbalance in the number of defect samples in the training data (e.g., more normal edges than gap edges), where the model easily ignores minority classes (important defects), resulting in high overall accuracy but serious missed detections. It significantly improves the recall rate of minority class defects (e.g., gaps, fragmented surfaces), preventing the model from ignoring these key but rare defect types due to differences in sample numbers, making the model's identification results more comprehensive and practical.

[0101] In step a2, a domain-knowledge-enhanced data augmentation strategy is employed during the training phase of the multi-task hierarchical graph neural network model. This strategy includes simulating defect generation patterns in the aerospace field, specifically: randomly introducing micrometer-level gaps at surface joints and / or generating irregular holes in planar areas to simulate process damage. Unlike general image rotation and cropping, this method specifically simulates the physical processes of defects that may actually occur in aerospace manufacturing and assembly processes. For example, introducing tiny gaps at surface joints (simulating assembly gaps) and generating irregular holes in planar areas (simulating impacts or corrosion). Adding these simulated, domain-specific defect data to the training set allows the model to learn in virtual engineering practice, enabling it to better understand and identify various real-world defects it may encounter in the future. This solves the problem that models trained solely on limited historical labeled data lack generalization ability and struggle to cope with the complex and varied defect morphologies in real engineering. Purely data-driven approaches may overfit the training set and have weak ability to identify new defect morphologies. This significantly enhances the model's robustness in identifying defects that may appear in real-world engineering scenarios but are not present in the original dataset. By increasing the diversity and realism of the training data, the model's generalization ability and engineering applicability were improved.

[0102] Specifically, step a4 includes: Step a4.1: Based on the spatial location information of the gaps, the gap repair algorithm is invoked at the boundary representation level to restore the topological connection of adjacent surfaces, obtaining an intermediate model after B-Rep level topological repair; Step a4.2: The intermediate model is triangulated to obtain a surface mesh; based on the spatial location information of holes, overlaps, and fragments, the hole-filling algorithm, the overlapping surface deletion or merging algorithm, and the fragment fusion or removal algorithm are sequentially executed on the surface mesh; Step a4.3: For the surface meshes corresponding to all components that have completed the operation in step a4.2, a mesh stitching operation is performed at adjacent boundaries to generate the final surface mesh. The repair follows two sequences: 1) Representational hierarchy order (B-Rep before discrete mesh): First, gaps are repaired at the precise boundary representation (B-Rep) level to restore the mathematical integrity of the model; then, other operations are performed on the discretized surface mesh, utilizing the flexibility of the mesh algorithm. 2) Defect Nature Order (Topology First, Geometry Second): Topological defects (gap) that affect the continuity of the model are addressed first, providing a closed and connected geometric framework for subsequent processing; then geometric anomalies (holes, overlaps, etc.) that do not affect the topology but affect the shape integrity are addressed. This order conforms to the data structure of the CAD model and the inherent logic of geometric problems, and is a core engineering principle to ensure the success of automated repair. It solves the problem of blind and disordered repair processes, which can easily introduce new defects or lead to repair failure due to improper repair order. For example, filling a fake hole composed of gaps first may fail due to boundary discontinuities. A stable, reliable, and predictable automated repair process is established. This ensures a high success rate and high-quality output for repair work, avoiding repeated trial and error and secondary damage to the model common in manual repair. Specifically, in step a1, when converting the geometric data of the sample parts into a structured feature representation, the extracted geometric features include aerodynamic sensitivity features and / or process constraint features.

[0103] In this process, when converting geometric data into model input (feature representation), not only are general area and curvature calculated, but additional domain-specific features are also calculated and added. For example, aerodynamic sensitivity features might include local angle of attack and estimated pressure gradient; process constraint features might include manufacturing tolerances and material type for that area. By using these features as part of the input, the model can establish a correlation between specific geometric shapes, specific engineering attributes, and defect importance during the learning process. This allows it to learn to assess the severity of a defect from an engineering value perspective, much like a domain expert. This solves the problem that AI models, by only learning general geometric features, cannot understand the engineering significance of defects, potentially leading to a defect list that does not match actual engineering priorities. For example, it might identify a process defect with no aerodynamic impact but miss a critical micro-gap that causes airflow separation. This ensures that the AI ​​model's identification results are directly aligned with the final engineering goals (such as aerodynamic performance and manufacturability). The defect severity assessment output by the model has greater engineering reference value and can truly guide the optimal allocation of repair resources.

[0104] In step a3, the spatial location information output by the multi-task hierarchical graph neural network model identifies edges for gap defects, boundary loops for hole defects, and faces for overlapping or fragmented defects. Based on the geometric nature of the defects, the output of the identification results must be mapped back to the original data structure of the CAD model: gaps are necessarily associated with one or a set of edges; holes are necessarily defined by loops formed by their boundary edges; and overlapping or fragmented defects are necessarily associated with one or a set of faces. This mapping ensures that the spatial location information output by the AI ​​is not a vague coordinate range, but rather a specific geometric entity ID or index that downstream B-Rep or mesh repair algorithms can directly identify and manipulate, thus achieving precise repair. This solves the problem that the defect information output by the AI ​​model is inconsistent in format and too abstract, making it unable to directly and unambiguously drive downstream repair algorithms. Simply outputting "hole" or "gap" is insufficient to guide specific geometric operations. Seamless instruction transmission from intelligent identification to automated repair is achieved. The AI's identification results are transformed into precise geometric element identifiers that the repair algorithm can directly execute, bridging the final link between intelligent decision-making and deterministic operation.

[0105] Based on the integrated model segmentation, intelligent identification, and phased repair technology proposed in Example 3, compared with the traditional completely manual processing mode, the following significant beneficial effects have been achieved: (1) Significantly reducing the dependence of geometric repair on the experience of senior experts: Due to the adoption of AI-based component-level geometric defect automatic identification and marking technology, the system can imitate the expert's perspective and automatically and accurately locate various complex defects, so that engineers do not need to rely on the naked eye and experience to conduct time-consuming and easily overlooked global searches, thereby enabling junior engineers to start the repair process efficiently and accurately, greatly reducing the technical threshold for high-quality mesh preprocessing. (2) Achieving an order-of-magnitude improvement in geometric repair efficiency: Due to the adoption of component segmentation + AI batch identification to replace the overall model + manual inspection technology, the open and serial defect search process is transformed into a structured and parallel intelligent detection process, which can shorten the model cleaning work that may take several days or even weeks to several hours, greatly compressing the overall cycle of simulation analysis. (3) Systematically accumulate and inherit repair strategies and experience: By using the technical means of encoding defect identification experience into a trainable and iterative AI model, the originally discrete, implicit and highly personalized expert experience (such as the skill of identifying specific types of gaps) is transformed into standardized digital assets that can be permanently saved, continuously optimized and shared within the team, ensuring the consistency and traceability of repair quality and realizing the efficient reuse of organizational knowledge. (4) Realize a closed loop from random repair to reliable and standardized repair process: By using the technical means of tightly coupling AI precise positioning and phased targeted repair algorithms, the system can automatically call the optimal repair strategy (such as topological fusion followed by geometric cladding) according to the identified defect type, avoiding the secondary damage to the model caused by improper methods in traditional manual repair, and finally forming a stable, reliable and predictable automated processing flow, fundamentally improving the robustness and success rate of the mesh generation stage.

[0106] Embodiment 3 of this invention provides a method for geometric defect identification and processing based on an AI large-scale model. The method includes: step a1 - training data preparation; step a2 - AI model training; step a3 - defect identification; and step a4 - phased repair and mesh generation. Embodiment 3 of this invention provides an efficient method for processing geometric defects in complex CAD models based on artificial intelligence. Its core principle is to address the pain points of traditional manual processing, such as lengthy processes, diverse problems, and lack of unified rules, by proposing a new paradigm of model segmentation, intelligent identification, and phased repair. First, the complex assembly model is intelligently decomposed into more regular components according to engineering semantics or geometric features. The purpose of splitting into components is that because there are many defect categories in a single model and it is impossible to classify them well, dividing it into individual components, such as the nose of an aircraft, makes the defects in a single component relatively fixed. Therefore, an AI large-scale model can be trained for identification, and these trained AI large-scale models have better accuracy and generalization. Subsequently, a deep learning model is constructed to learn the mapping relationship between component geometric features and defect categories and spatial locations, enabling automated identification and labeling of geometric defects (such as gaps, overlaps, holes, and fragments) within each component. Finally, the AI ​​identification results are seamlessly integrated with a carefully designed phased geometric repair algorithm flow, first repairing the topology (gaps), then processing the geometry (holes, overlaps, etc.), and finally obtaining a watertight overall surface mesh through mesh stitching, thereby freeing engineers from heavy and repetitive manual inspection and repair work.

[0107] (1) Complete Method Process: This method includes the following steps from model preprocessing to final mesh generation: 1. Model preparation and component segmentation: The original CAD assembly model is automatically or semi-automatically segmented into multiple sub-components with stronger regularity based on geometric continuity, engineering functional semantics, or user-defined rules. 2. Data preparation and feature representation: Geometric samples of components with defect annotations are collected and converted into multimodal feature representations suitable for deep learning model processing (such as figure structure, point cloud, feature vector, etc.), and a training dataset is constructed. 3. AI geometric defect recognition model training: Using the constructed training dataset, a deep learning model (such as a neural network) is trained to accurately identify and locate various geometric defects and their boundaries in the input components. 4. Intelligent identification and localization of defects in new components: For the segmented new components, their features are extracted and input into the trained AI model. The model outputs the type, confidence level, and precise spatial location markers of all defects on the component. 5. Staged geometric repair and global mesh stitching: Based on the defect markers provided by AI, the gap repair algorithm is first called at the B-Rep level to restore the topological connectivity; then the repaired surface is discretized, and discrete repair and surface wrapping are performed for problems such as holes and overlaps to obtain a component-level watertight mesh; finally, mesh stitching operation is performed at the mesh boundaries of different components to generate the final global watertight surface mesh.

[0108] (2) Detailed introduction of each step: Model preparation and component segmentation; receiving the original CAD assembly model, which may contain various geometric defects. Based on geometric features (such as surface continuity, curvature abrupt changes), a predefined engineering semantic rule base (such as wings, fuselage), or user interactive input, the system decomposes the model into multiple relatively independent sub-components with more regular internal geometric features and potential defect types. The processing object is the entire CAD assembly. This step is the foundation of the model segmentation strategy, and its decomposition quality directly affects the accuracy of AI recognition and the convenience of repair. Component segmentation simplifies the geometric context within each sub-component, making it easier for the AI ​​model to learn and generalize.

[0109] Data preparation and feature representation: For each component sample (from historical projects or artificial constructions), experts annotate its geometric defects (gap edges, overlapping surfaces, hole rings, fragmented surfaces, etc.) and their type labels. Subsequently, the B-Rep model of the component is converted into one or more deep learning-friendly representations.

[0110] 1. Graph Representation: Using faces as nodes and shared edges between faces as edges, an attribute graph is constructed. Node features can include geometric attributes of the face (area, curvature, normal), and edge features can include the included angle and length of the edges. 2. Hybrid Representation: Combining graph structure with locally densely sampled point clouds. Based on the graph structure, point cloud sampling is performed on each face or its boundary region, adding local geometric details. This step is the sample preparation stage for AI model training. Transforming unstructured geometric data into structured, semantically rich feature representations is crucial for the model to effectively learn defect patterns. The pairing of labeled data (defect location and type) with feature representations constitutes the training samples.

[0111] AI Geometric Defect Recognition Model Training: This invention designs a multi-task hierarchical graph neural network (MT-HGNN) for aerospace geometric defects, instead of directly using a general GNN model. The customized network structure is as follows: a) Multi-scale feature extraction layer: Node-level encoder: A multilayer perceptron (MLP) is used to encode the basic geometric features (such as area, mean curvature, Gaussian curvature, and normal vector) of each patch node. Edge-level encoder: Features of shared edges (such as the included angle, edge length, and relative curvature change) are encoded. Subgraph-level encoder: Node and edge information is aggregated through graph convolutional layers (such as GATv2) to generate an embedded representation of a local subgraph (such as the neighborhood around a vertex or edge) to capture local context. b) Hierarchical Graph Pooling and Attention Mechanism: An adaptive topological pooling layer is designed to perform hierarchical clustering of graph nodes based on the curvature characteristics and functional semantics (e.g., distinguishing between high-curvature leading edges and smooth skin of an airfoil), generating graph representations at different levels of abstraction. This simultaneously captures microscopic defects (e.g., tiny gaps) and macroscopic defect distribution patterns. A domain-knowledge-guided attention mechanism is introduced. For example, higher attention weights are assigned to the edges of key aerodynamic regions such as the wing leading edge and wing-body junction, making the model focus more on these regions that significantly impact CFD results. Domain knowledge refers to prior rules and experience in aerospace engineering, such as: Aerodynamic knowledge: the wing leading edge, engine intake / exhaust ports, and wing-body junctions are highly aerodynamically sensitive, and geometric defects in these areas have a significant impact on CFD results. Manufacturing and process knowledge: different manufacturing processes (e.g., sheet metal forming, composite material laying) have their typical defect types and tolerance ranges. Functional semantic knowledge: certain regions are load-bearing structures, sealing surfaces, or moving surfaces, requiring extremely high geometric integrity.

[0112] Specific Implementation of the Domain Knowledge-Based Attention Weight Guidance Mechanism: To achieve model focus on key regions, this invention introduces a domain knowledge-based attention weight guidance mechanism in the hierarchical graph pooling and attention layers of the Multi-Task Hierarchical Graph Neural Network (MT-HGNN). The core of this mechanism lies in superimposing a prior weight bias determined by domain knowledge on top of the model's self-attention calculation, thereby guiding the model to pay more attention to engineering-significant key regions. Specifically, this includes: 1. Construction of a Domain Knowledge Rule Base: First, based on aerodynamic principles, structural design specifications, and manufacturing process manuals, a computable rule base is constructed. Each rule maps the geometric / topological conditions in the CAD model to an attention adjustment coefficient. For example: Rule A (Aerodynamically Critical Region): IF patch is located within the wing leading edge or engine lip region (determined by the patch's parameter coordinates or labels) THEN basic attention weight coefficient β = 1.5. Rule B (Secondary Non-Structural Region): IF patch belongs to a non-load-bearing fairing or interior panel THEN β = 0.8. Rule C (High-risk areas for connections and gaps): If a facet is adjacent to more than two other components (at the boundary of a complex assembly), THENβ = 1.3. 2. Integration of guidance mechanisms in attention calculation: Let a be the attention score of the i-th facet to the j-th facet obtained by standard computation in a graph neural network. ij This mechanism is adjusted in the following ways: ;in, ij The adjusted attention score is represented by λ, a trainable hyperparameter used to control the strength of domain knowledge guidance. This is introduced as an additive bias term. When both patches belong to critical regions (both with β values ​​greater than 1), this term is positive, significantly increasing their attention scores; conversely, it may decrease them if neither patch belongs to a critical region. β i and β j These are the patch-level prior attention modulation coefficients. They are two independent scalar values, representing the importance assessments of the i-th patch (node) and j-th patch (node) in the graph neural network based on domain knowledge, respectively.

[0113] This mechanism achieves the following: Accelerated convergence: The model is given a near-expert perspective from the early stages of training, eliminating the need for extensive, inefficient trial-and-error to discover key regions, thus accelerating training. Improved accuracy in key region identification: The model allocates more computational resources to aerodynamically sensitive areas and connection zones—regions with zero tolerance for defects—significantly improving the accuracy and recall of identifying minute gaps and surface fragments in these areas. Enhanced decision interpretability: The model's focus aligns with engineering consensus, making its identification results easier for domain experts to understand and trust, overcoming the obstacles of traditional AI black-box problems in engineering applications. To achieve targeted intelligent identification, this invention performs engineering semantic-based region division on the aircraft CAD model, mainly including two categories:

[0114] (1) Aerodynamic critical areas: These are areas whose geometric integrity directly determines the accuracy of aerodynamic simulation. For example, the leading and trailing edges of the wing and tail, the engine inlet lip, and the wing-body fairing. These areas have low tolerance for defects such as gaps and step differences. (2) Manufacturing process areas: These are areas divided according to the actual production method of the parts. Different areas have unique defect patterns and quality standards. For example, the skin of [part A] is marked as a composite material layup area, and its focus is on delamination and resin unevenness; the bracket of [part B] is marked as a precision casting area, and its focus is on shrinkage cavities and deformation.

[0115] 2. Application of Knowledge in the Model: The aforementioned region segmentation information is injected into the AI ​​model in two ways: as input for attention guidance: In the attention mechanism, patches belonging to the aerodynamic critical region are assigned a higher prior weight coefficient β to guide the model to focus. As input for feature engineering: Manufacturing process type feature labels are added to each patch, enabling the model to understand defects in conjunction with the process background.

[0116] c) Multi-task learning output heads: The network end is not a single classifier, but rather multiple specific task output heads connected in parallel: Gap identification head: performs binary classification on each edge in the graph (gap / non-gap). Fragment identification head: classifies each facet node (fragment / normal face). Hole boundary identification head: identifies and serializes the edges that constitute the hole boundary loop. Defect severity assessment head: regresses and predicts the impact coefficient of each identified defect on the mesh quality (e.g., a score between 0 and 1), providing a basis for subsequent repair priority ranking. This multi-task design enables the model to collaboratively learn the correlation between different defect types, improving overall recognition accuracy. Among them, the gap identification head is an edge-level binary classifier. It receives features extracted by the graph neural network, judges each edge in the model, and outputs whether the edge is a gap. It accurately locates the position of topological discontinuity. Gap is the primary cause of the model's lack of watertightness, and this output is a direct instruction to initiate topology repair. The fragment identification head is a face-level classifier. It receives features from each facet node and determines whether the facet is a fragmented facet (facet with too small an area, abnormal aspect ratio, or no substantial contribution to the geometry or fluid simulation). It identifies geometric noise that needs to be cleaned up or merged. The repair algorithm will delete or merge these faces based on this result to simplify the geometry and improve mesh quality. The hole boundary recognition head is a boundary loop recognition and serialization module. By analyzing the features and connectivity of edges, it identifies edge sequences that form closed loops but have no faces within the loops and marks them as holes. It accurately delineates the boundaries of missing surface regions. This output provides the geometry repair algorithm with clear, closed filling targets.

[0117] The defect severity assessment head is a regression or grading evaluator. For each identified defect (such as a crack or a hole), it outputs a numerical value (such as a score of 0-1) or a grade (such as high, medium, or low) representing its severity. This provides engineering prioritization for repairs. Severity assessment is based on predictions of downstream mesh generation success rates and quality impacts, enabling the system to prioritize the most critical defects and optimize the repair process.

[0118] Training Strategy: Loss Function: This invention employs a weighted multi-task loss function for end-to-end training of a multi-task hierarchical graph neural network. The total loss function is defined as follows:

[0119] The definitions of each sub-loss function are as follows:

[0120] For the loss function of the gap detection head, a weighted binary cross-entropy loss is used. Since the number of normal edges in the training set is much greater than that of gap edges, positive and negative samples are balanced by class weights, and the calculation formula is as follows:

[0121] ;

[0122] in, The true label for edge e (1 for gap, 0 for no gap). To predict probabilities for the model, is the weight of the positive samples, set as the ratio of the number of negative samples to the number of positive samples, and E is the set of all edges involved in the computation in the graph neural network. For the loss of the facet recognition head, the weighted binary cross-entropy loss is also used, and the definition method is the same as... Similarly, as shown in the following formula, the weights The ratio of broken surface samples to normal surface samples is set.

[0123] ;

[0124] For the hole boundary identification head, this task is a sequence labeling task, and either Connection Temporal Classification Loss (CTC Loss) or edge-level binary cross-entropy loss is used. This invention preferably uses edge-level binary cross-entropy loss, which performs binary classification on the edges constituting the hole boundary, with the loss form being the same as... As shown in the formula below, the weights The ratio of the hole boundary edge to the normal edge is set.

[0125] ;

[0126] For the loss of the defect severity assessment head, this task is a regression task, and the mean squared error loss is used:

[0127] ;

[0128] in, The severity score (0~1) of the defect as indicated by the experts. The severity score predicted by the model. Total number of identified defects. Weighting coefficient. The setup method is as follows: To address the issues of inconsistent convergence speeds and dimensional differences among tasks in multi-task learning, this invention employs an adaptive loss weighting method based on task uncertainty or an inverse proportional weighting method based on sample size. In a preferred embodiment, the inverse proportional weighting method is used: ;

[0129] in, This represents the number of valid samples for the corresponding task in the training set (the number of positive samples for classification tasks and the number of defect instances for regression tasks). This setting gives higher weights to tasks with smaller sample sizes, effectively alleviating class imbalance and improving the recall rate for identifying minority class defects. Here, z is the index variable for all tasks. This represents the total number of samples in the training set for the corresponding task. In typical implementation scenarios, the values ​​of each weight can range as follows: , , , And satisfy The specific values ​​can be dynamically calculated and determined based on the statistical distribution of the training set.

[0130] Domain-specific data augmentation: During training, in addition to conventional rotation and scaling, data augmentation is performed by simulating defect generation patterns unique to the aerospace field. For example, micron-level gaps are randomly introduced at surface joints, and irregular holes simulating process damage are generated in planar areas to improve the model's robustness to real engineering scenarios.

[0131] Intelligent identification and localization of defects in new components: For a new component to be processed after segmentation, the system automatically converts it into a feature representation consistent with the training data (as shown in the figure). This representation is then input into the trained AI model. The model performs forward inference and outputs a structured identification result. Execution of objects and logical relationships: The output is a machine-readable defect report that clearly indicates which edges in the component are identified as gaps, which surfaces are identified as overlaps or fragments, and the suspected boundary loop locations of holes. This report accurately marks the spatial location and type of defects, providing direct and clear input guidance for targeted repair in subsequent steps, replacing the traditional fully manual, experience-based visual search process.

[0132] Phased Geometric Repair and Global Mesh Stitching: This step is the core repair process, strictly following the principle of topology first, then geometry, and is executed in three stages: 1. B-Rep Level Gap Repair: Based on the gap markings output by AI, at the boundary representation (B-Rep) level of the original CAD, gap repair algorithms (such as boundary fusion, surface extension / trimming) are called to restore the topological connection of the marked adjacent surfaces, generating an intermediate B-Rep model that is topologically complete but may still have other defects. 2. Discretization and Surface Mesh Repair: The B-Rep model obtained in the previous step is triangulated and discretized to generate a surface mesh. Based on the AI's markings of holes, overlaps, and fragmented surfaces, corresponding algorithms are executed on this mesh: closed loops are constructed in the hole marking area using a loop-finding algorithm and holes are filled; marked overlapping surfaces are deleted or merged; marked fragmented surfaces are merged or removed. Subsequently, a surface wrapping algorithm (such as Alpha Wrap) is applied to generate a clean, watertight temporary closed surface mesh for the entire part, and then the part outside the original boundary is trimmed to obtain a watertight mesh of the part that preserves the boundary. 3. Global Mesh Stitching: For all processed component-level watertight meshes, mesh stitching techniques such as node merging and edge swapping are used at their adjacent boundaries to seamlessly connect the component meshes, ultimately generating a complete, watertight, and globally suitable surface mesh model for subsequent simulations. This step is crucial for transforming AI intelligent recognition results into high-quality geometric output. Through a rigorous algorithm, it assigns different types of defects to the most suitable repair stage, ensuring repair efficiency and result reliability, ultimately achieving fully (semi-)automated conversion from defective CAD to a clean computational mesh.

[0133] The key innovation of Embodiment 3 of this invention lies in explicitly encoding implicit expert knowledge in the aerospace field into the learning and reasoning process of the AI ​​model. This is achieved through the following three mechanisms: Mechanism 1: Knowledge Embedding in Feature Engineering: In data preparation and feature representation, the geometric features we extract are not only general attributes but also engineering features strongly correlated with aerodynamic performance and manufacturing processes. For example: Aerodynamic Sensitivity Features: Calculating the local angle of attack and the estimated flow separation risk coefficient at the location of the surface patch, and inputting these parameters as node features into the model, enabling the model to prioritize defects in aerodynamically sensitive areas. Process Constraint Features: Labeling the surface patch with its corresponding manufacturing process (such as sheet metal forming, composite material layup), and using the allowable tolerance range of the process as features, enabling the model to distinguish between acceptable process defects and fatal defects that must be repaired. Mechanism 2: Knowledge Guidance in the Attention Mechanism: As described in the network structure section, in the attention layer of the GNN, prior attention weights are assigned to different functional regions using a predefined rule base. The rule base is derived from aerodynamic principles and design specifications. For example, if the IF region belongs to the wing leading edge, the THEN attention weighting coefficient is 1.2; if the IF region belongs to the non-load-bearing fairing, the THEN attention weighting coefficient is 0.8.

[0134] This gives the model an expert perspective from the early stages of training, accelerating convergence and improving the recognition accuracy of key regions.

[0135] Mechanism 3: Knowledge integration of loss function and evaluation criteria;

[0136] The optimization objective (loss function) of the model not only pursues classification accuracy, but also links it to the success rate of downstream CFD grid partitioning.

[0137] When labeling training data, not only is the existence of defects indicated, but experts also indicate the urgency level of their repair based on historical experience (e.g., causing mesh generation failure, causing poor mesh quality, or having negligible impact).

[0138] Through learning, the model can output an engineering severity score for defects, thereby enabling intelligent prioritization based on engineering impact and guiding the repair process to address the defects that have the greatest impact on simulation results first.

[0139] This invention, in its third embodiment, proposes a new paradigm for automated processing of geometric defects in complex CAD models based on artificial intelligence and phased repair. The difference from traditional methods lies in: 1. A pioneering integrated architecture of component-level AI recognition and phased repair: Traditional methods involve manual global inspection of the entire complex model followed by individual repair of identified problems. This invention, in its third embodiment, employs intelligent component segmentation, parallel AI batch recognition of each component, and phased system repair based on defect type. This represents a fundamental change in workflow, transforming a sequential, manual-dependent model into a parallel, intelligently driven one. 2. A multi-task hierarchical graph neural network (MT-HGNN) architecture for aerospace CAD geometric defect recognition. This architecture protects the specific network components, including: a multi-scale feature extraction layer, an attention mechanism incorporating domain knowledge weights, an adaptive topological pooling layer, and a multi-task output head for collaboratively identifying gaps, fragments, holes, and assessing defect severity. 3. An AI-driven intelligent identification and localization method for geometric defects: This transforms the domain knowledge of CAD geometric defect recognition, which highly relies on expert experience, into a trainable and reusable AI model. Specific methods for encoding aerospace domain knowledge into the AI ​​model are also provided. Specifically, this includes (a) embedding engineering features such as aerodynamic sensitivity and manufacturing tolerances into geometric features; (b) utilizing a mechanism of guided graph attention weights based on a predefined rule base; and (c) incorporating expert experience, such as the urgency of defect repair, into the model loss function and output criteria. 4. A phased engineering process for repair: A repair order prioritizing topology over geometry: Prioritizing gaps (restoring B-Rep topology) before addressing geometric issues (holes, overlaps, etc.). Combining B-Rep layer repair with discrete layer repair: fully utilizing the advantages of both representation forms. A strategy of independent component repair + overall stitching: reducing processing complexity and improving fault tolerance.

[0140] Example 4;

[0141] This invention provides a detailed description of step 3 in embodiment 1 in embodiment 4: This invention provides a method for predicting the distribution of grid nodes and anisotropic parameters of aircraft component boundary lines, realizing full-process intelligentization from high-level design intent and physical conditions to low-level grid generation instructions. Specifically, it includes: realizing intelligent automated decision-making of grid generation strategies, realizing intelligent prediction of differentiated parameters in coordination with strategies, and constructing an end-to-end hierarchical intelligent decision-making framework and system.

[0142] Embodiment 4 of the present invention provides a method for predicting the distribution of mesh nodes and anisotropic parameters of aircraft component boundary lines. The method includes: Step b1: Determining the geometric data, component semantic labels, computational conditions, and boundary lines on the surface of the aircraft CAD model; wherein, the component semantic labels are used to identify the functional category associated with the mesh division strategy to which each component in the aircraft CAD model belongs; Step b2: Based on the geometric data, component semantic labels, and computational conditions, using a trained neural network model, classifying the mesh region types of each boundary line on the surface of the aircraft CAD model to obtain the mesh region type classification results of each boundary line; the mesh region types include isotropic regions and anisotropic regions; Step b3: According to the mesh region type classification results, using the differential parameter prediction subnetwork in the neural network model corresponding to each mesh region type, predicting the corresponding set of mesh control parameters for each boundary line; Step b4: Based on the set of mesh control parameters, driving a mesh generator to generate the surface mesh of the aircraft CAD model.

[0143] This method constructs a complete logical closed loop through four steps: Step b1 clarifies the basis for decision-making (geometry, semantics, working conditions, boundary lines); Step b2 uses an AI model to replace human judgment in determining the grid type; Step b3 intelligently predicts specific parameters based on the strategic decision results; Step b4 generates the grid based on intelligent instructions. This achieves fully automated mapping from semantics to classification to prediction and finally to the grid. It solves the fundamental efficiency and quality problems in traditional CFD preprocessing, where the two core decisions—grid region type judgment and specific parameter setting—rely entirely on manual intervention and are disconnected from each other. It establishes a new automated process framework of semantic-driven, intelligent decision-making, and automatic execution. It compresses days or even weeks of experience-based manual work into an automated intelligent computing process.

[0144] The neural network model is a multi-task graph neural network model. In step b2, the grid region type classification is performed on each boundary line on the surface of the aircraft CAD model. Specifically, a boundary line topology graph is constructed using each boundary line as a node and the topological connection relationship between the boundary lines as edges. The boundary line topology graph and the feature vectors of each node are input into the multi-task graph neural network model. The feature vectors of each node are formed based on the geometric features extracted from the geometric data of the corresponding boundary line, the semantic label of the component, and the calculation conditions. The grid region type classification head in the multi-task graph neural network model is used to perform binary classification on each node in the boundary line topology graph, and the probability that each boundary line belongs to an isotropic region or an anisotropic region is output as the grid region type classification result.

[0145] This approach involves constructing the entire boundary line system of the model as a graph, with boundary lines as nodes and connections as edges. When input into a graph neural network, the model, in judging a particular boundary line, can simultaneously perceive information from other connected boundary lines, thus making a more co-categorized classification that aligns with the overall geometry and flow logic. This clarifies how intelligent classification works. It solves the problems of single, isolated boundary line judgments lacking context and failing to guarantee the overall rationality and coherence of region division. It endows classification decisions with a global perspective and contextual awareness, resulting in more accurate and consistent classification results.

[0146] Specifically, step b3 involves the multi-task graph neural network model routing the boundary line node features belonging to isotropic regions to the first parameter prediction sub-network based on the grid region type classification result output by the grid region type classification head, in order to predict a first set of grid control parameters including the total number of nodes and the node size distribution function parameter; and routing the boundary line node features belonging to anisotropic regions to the second parameter prediction sub-network, in order to predict a second set of grid control parameters including the total number of nodes, the node size distribution function parameter, the first layer grid height, the grid growth rate, and the number of growth layers.

[0147] The model incorporates two independent and dedicated prediction sub-networks. The classification results act as routing signals, automatically guiding boundary features to the corresponding sub-networks: those heading towards isotropic networks only predict basic parameters; those heading towards anisotropic networks must predict a complete parameter set including first-layer height, growth rate, and layer number. This mechanism eliminates the possibility of missing key parameters. It also specifies how to differentiate predictions. This addresses the problem of a one-size-fits-all approach or parameter mismatch when predicting parameters for different strategy regions, ensuring that anisotropic regions always obtain the key parameters necessary for advancing the boundary layer. It achieves strong coupling between classification results and parameter prediction, guaranteeing the physical correctness and completeness of the parameter set.

[0148] The method for determining the boundary lines on the surface of the aircraft CAD model is as follows: The input aircraft CAD model is processed to obtain component segmentation results and key geometric feature lines. The component segmentation results define the correspondence between each surface of the aircraft CAD model and the semantic labels of the components. The key geometric feature lines include at least all boundary lines between components. All boundary lines between components are selected from the key geometric feature lines as the boundary lines on the surface of the aircraft CAD model. This method indicates that the boundary lines originate from the key geometric feature lines obtained after intelligent component segmentation of the CAD model, and the boundary lines between components are selected from them. This clarifies the automated acquisition source of the key input object, the boundary lines, avoiding the bottleneck of still needing manual extraction. It achieves automated pre-processing from the CAD model to the boundary line list, making the entry point of the entire process fully automated. The method for determining the component semantic labels is as follows: Based on the component segmentation results, the component semantic labels to which each surface of the aircraft CAD model belongs are determined. The component semantic labels include one or more of the following: wing leading edge, wing upper surface, wing lower surface, fuselage nose, and engine nacelle.

[0149] The multi-task graph neural network model is obtained through the following offline training method: A training dataset is constructed, where each sample data point corresponds to a boundary line in a historical successful computation example. The sample data includes input features and output labels. The input features include the geometric features, component semantic labels, and computational conditions of the boundary line. The output labels include the optimal grid region type label and the corresponding optimal grid control parameter label for the boundary line. The optimal grid region type label is used to supervise the classification of the grid region type, and the optimal grid control parameter label is used to supervise the set of grid control parameters. Based on the training dataset, the constructed multi-task graph neural network model is trained end-to-end. The loss function of the multi-task graph neural network model is a joint loss function, which includes a classification loss for optimizing grid region type classification, a first regression loss for optimizing isotropic parameter prediction, and a second regression loss for optimizing anisotropic parameter prediction. The geometric features include at least one of the length and curvature information of the boundary line; the computational conditions include at least one of flight Mach number and flight angle of attack. The node size distribution function parameter is used to control the density distribution of grid nodes along the length of the boundary line.

[0150] In step b4, based on the set of mesh control parameters, the mesh generator is driven to generate the surface mesh of the aircraft CAD model. Specifically, the set of mesh control parameters for each predicted boundary line is converted into a control script or API instruction corresponding to the target mesh generation software. The control script or API instruction is executed to drive the target mesh generation software to complete the discretization of all boundary lines, anisotropic mesh advancement, and surface triangulation, and output the final surface mesh file.

[0151] Because the technical solution of Embodiment 4 of this invention creatively adopts a hierarchical integrated technical solution that first intelligently classifies grid region types and then predicts differentiated parameters in a targeted manner, the following outstanding beneficial effects are achieved: It achieves a breakthrough in intelligent grid generation strategies: Through AI-based automatic classification of region types, this invention, for the first time, systematically automates strategic decisions (which grid to use) that rely on high-level experience, solving core pain points that existing methods cannot address and achieving true strategy-level intelligence. It ensures a high degree of self-consistency between grid parameters and physical requirements: By adopting a differentiated prediction architecture driven by classification results, it ensures that the parameters predicted for anisotropic regions necessarily include the key parameters necessary for boundary layer advancement. The generated grid's density distribution and anisotropic characteristics naturally match the flow physical field, resulting in high physical fidelity. It brings an order-of-magnitude improvement in preprocessing efficiency and quality standardization: It compresses the traditional manual decision-making process, which requires several days and relies on repeated trial and error, into automated calculations at the minute level. Simultaneously, through best practices embedded in the AI ​​model, it ensures the consistency, repeatability, and high level of grid generation strategies and quality across different projects and personnel. A transparent and optimizable standard intelligent process has been constructed: the fuzzy expert decision-making process is decomposed into two clear, learnable, and verifiable steps: classification and differentiated prediction, forming a transparent, explainable, and continuously optimizable standardized intelligent pre-processing process.

[0152] 1) Offline Training Phase: Step A: Construct a high-quality training dataset. Extract the input features (geometric, semantic, and operational conditions) and corresponding optimal output labels (number of nodes, distribution parameters, and anisotropic parameters) for each boundary line from historical successful examples. Step B: Construct and train a deep learning model. Design a multi-task learning model to learn the complex mapping relationship from input features to multiple output targets. 2) Online Application Phase: Step C: Perform intelligent prediction on the new model. Input the component semantics, boundary lines, and operational conditions of the new aircraft model, and the trained model automatically outputs all control parameters for each boundary line. Step D: Drive automated surface mesh generation. Directly input the output parameters from Step C into the mesh generator to generate a high-quality initial surface mesh with one click.

[0153] (1) Detailed introduction of each step: 1) Offline training stage; construct a high-quality training dataset with region type labels; systematically collect historical successful CFD cases, associate CAD models, working conditions and final mesh files, and clearly label the region type label (isotropic / anisotropic) for each boundary line, which is the basis for the model to learn the classification task. Training sample feature and label extraction; extract input feature vectors for each boundary line, including geometric features (such as length, curvature, etc.), component semantic labels (such as wing leading edge) and global calculation working conditions (such as Mach number, angle of attack, etc.); extract output labels, including region type labels and mesh parameter labels (number of nodes, distribution function parameters). If it is an anisotropic region, add anisotropic parameter labels (first layer height, growth rate, number of layers).

[0154] Construction and training of graph neural network model; (1) Graph structure input: Construct a graph with all boundary lines of the example as nodes and connection relationships as edges, input the graph into the model, so that the model has context awareness. (2) Model architecture: Design a multi-task graph neural network with an explicit hierarchical structure. First level: Region type classification head. This is a classifier that performs binary classification on each node (boundary line) in the graph and outputs the probability that it is an isotropic or anisotropic region. This step realizes the intelligent decision of grid strategy. Second level: Differential parameter prediction head. It contains two structurally independent and task-different prediction sub-networks. Sub-network A is an isotropic parameter prediction head, which is specifically used to process the node features classified as isotropic and outputs the node number and size distribution function parameters; Sub-network B is an anisotropic parameter prediction head, which is specifically used to process the node features classified as anisotropic and outputs the node number, size distribution function parameters and complete anisotropic advancement parameters. (3) Training mechanism: Through gating or attention mechanism, the classification results of the first level are used as routing information to dynamically guide the feature flow to the corresponding prediction sub-network of the next level. End-to-end training is performed using a joint loss function (classification loss + isotropic regression loss + anisotropic regression loss) to enable the model to learn to collaboratively optimize classification and prediction tasks.

[0155] The loss function is as follows:

[0156] ;

[0157] In the formula, , and The weight parameters are initially set to... Adjustments can be made based on subsequent experiments.

[0158] Classification loss The first-level classification head outputs a probability for each node i. This represents the probability that the boundary line belongs to an anisotropic region. (True label) (0 represents isotropic, 1 represents anisotropic). The classification loss uses binary cross-entropy:

[0159] ;

[0160] In the formula, N represents the total number of nodes in the current batch. For the true label... The number of nodes in the second-level subnetwork A is predicted to be [number]. The distribution function parameter vector is The actual number of nodes and the distribution parameter vector are respectively and Then the isotropic regression loss function is:

[0161] ;

[0162] In the formula, The total number of isotropic nodes. For the regression loss function, mean squared error (MSE) or smoothed L1 loss (Huber) can be chosen to enhance robustness to outliers. The weight parameters are initially set to... Adjustments can be made based on subsequent experiments.

[0163] For real labels The number of nodes in the second-level subnetwork B is predicted to be [number]. The distribution function parameter vector is The anisotropy parameters are: first floor height ,growth rate Number of floors The actual number of nodes and the distribution parameter vector are respectively and The anisotropy parameters are as follows: , , The isotropic regression loss function is then:

[0164]

[0165] In the formula, The initial weight parameters are set to represent the total number of isotropic nodes. Further adjustments can be made based on experiments. Model validation and deployment preparation: Evaluate the model's region classification accuracy and parameter prediction precision on the test set, and conduct final validation through parameter-driven mesh reconstruction. After successful validation, package the model as a standardized service. The trained AI model and test set.

[0166] 2) Online intelligent application stage; input and preprocessing; receive the new aircraft CAD model and working conditions, automatically perform component segmentation and semantic annotation, and extract all component boundary lines. Boundary line feature extraction and graph construction; calculate the geometric features for each boundary line, combine its semantic labels and global working conditions to form a feature vector. Based on the topological connection relationship of the boundary lines, construct a boundary line graph. Hierarchical intelligent collaborative decision-making; (1) Intelligent division of region types. The model first performs the first-level classification task and outputs a clear region type division map, clearly identifying the grid strategy type of all boundary lines. (2) Intelligent prediction of differentiated parameters. Based on the division results, the model automatically routes internally, activates the isotropic parameter prediction head and the anisotropic parameter prediction head in parallel, and outputs two sets of structured parameter sets respectively. Finally, a structured list containing the region type of each boundary line and its corresponding complete parameter set is output.

[0167] Automated mesh generation; The above structured parameter list is automatically converted into control scripts or API instructions that can be recognized by the target mesh generation software (such as NNW-GridStar), driving NNW-GridStar to automatically complete the discretization of all boundary lines, anisotropic mesh advancement and surface triangulation, and finally output a high-quality surface mesh file. The technical solution of Example 4 solves the two core bottlenecks in the prior art: the complete reliance on human experience for mesh region type judgment and the disconnect between parameter prediction and high-level mesh strategy, realizing full-process intelligentization from high-level design intent to low-level mesh instructions. Specifically, it realizes: (1) Automated mesh strategy decision-making: Based on component semantics and computing conditions, it intelligently and automatically determines the type of mesh region (isotropic or anisotropic) to which each boundary line should belong, replacing human experience judgment. (2) Differentiated parameter intelligent prediction: Based on the region type judgment, it differentiates and coordinates the prediction of a complete set of mesh control parameters with different structures for different types of regions, ensuring that the parameters are highly matched with the physical strategy. (3) Construct an end-to-end intelligent system: Establish a hierarchical decision-making system that integrates intelligent classification and differentiated prediction to achieve full-process automation from CAD model semantic annotation and working condition input to one-click generation of high-quality surface mesh.

[0168] The method in Embodiment 4 of this invention proposes an automatic method for classifying surface mesh regions into isotropic or anisotropic types based on component semantics and computational conditions during CFD preprocessing using an artificial intelligence model. AI is utilized to achieve automatic classification decisions for mesh region types. Embodiment 4 also proposes an intelligent system integrating region type discrimination and parameter prediction. Based on the region type discrimination results, it automatically selects and executes different parameter prediction logic branches, outputting parameter sets with different structures for isotropic and anisotropic regions. The method in Embodiment 4 constructs a complete, closed-loop framework comprising three layers: intelligent region partitioning, differentiated parameter prediction, and automatic mesh generation. This achieves intelligent mapping and execution across the entire chain from high-level semantic information to low-level mesh generation instructions. The method in Embodiment 4 is not only applicable to CFD mesh generation for aircraft, but its core idea—automatically determining the discretization and refinement settings based on component functional semantics and operating conditions—can be extended to any simulation preprocessing workflow involving complex geometric and physical field coupling, such as automobiles, ships, and energy equipment, and has broad industry application prospects. The hierarchical decision-making framework proposed in Embodiment 4 of this invention, which first performs intelligent classification based on object attributes and then differentiates processing based on the category, has wide applicability. It can be extended to other CAE simulation fields, such as: intelligently allocating different element types and mesh sizes based on material properties and stress states; and dividing contact mesh densities according to the relative motion relationship and importance of contact pairs.

[0169] In the method of Embodiment 4 of this invention, the specific architecture of the AI ​​model (such as a GNN classifier head or a CNN classifier) ​​for region classification and the regression network for parameter prediction can be replaced with other similar or related models or networks, and this invention does not impose any limitations on them. In the method of Embodiment 4 of this invention, the definition of region types is not limited to binary (isotropic / anisotropic) and can be extended to multi-class classification (such as adding shock wave capture regions, wake regions, etc.), and this invention does not impose any limitations on them. In the method of Embodiment 4 of this invention, the classification model and the prediction model can be tightly coupled multi-task models or loosely coupled but collaborative independent model combinations. In the method of Embodiment 4 of this invention, the third-party software that ultimately drives the generation of the mesh can be selected according to actual needs, and this invention does not impose any limitations on it.

Claims

1. A fully automated method for generating unstructured meshes, characterized in that, The method includes: Step 1: Perform component segmentation on the input aircraft CAD model to be processed, and obtain the component segmentation results and corresponding key geometric feature lines; Step 2: Based on the component segmentation results, identify and repair geometric defects in each sub-component of the component segmentation results to generate a component-level surface mesh; Step 3: Based on the key geometric feature lines, determine the boundary lines of the sub-component surface; based on the component-level surface mesh, extract the geometric data of each boundary line; fuse the geometric data, the semantic labels of the sub-component, and the computational conditions, and use a trained neural network model to classify the mesh region types of each boundary line, obtaining the mesh region type classification results of each boundary line, where the mesh region types include isotropic regions and anisotropic regions; based on the mesh region type classification results, use the differential parameter prediction sub-network in the neural network model corresponding to each mesh region type to predict the corresponding set of mesh control parameters for each boundary line; drive surface mesh generation based on the set of mesh control parameters to obtain the optimized surface mesh; Step 4: Based on the optimized surface mesh, generate the boundary layer mesh and the three-dimensional volume mesh, assemble and output the volume mesh of each sub-component to obtain a complete unstructured volume mesh model for simulation; Step 1 specifically includes: Step 1.1: Obtain the CAD model of the aircraft to be processed, wherein the CAD model of the aircraft has a boundary representation structure; Step 1.2: Based on the engineering requirements of mesh generation, perform engineering semantic annotation on the faces in the boundary representation structure of the aircraft CAD model to obtain component semantic labels; extract the geometric parameters of each face from the boundary representation structure; generate a multi-view 2D rendering image for each annotated face; associate and store the multi-view 2D rendering image, the geometric parameters, and the component semantic labels to form training samples; Step 1.3: Train a multimodal recognition model using the training samples. The multimodal recognition model is configured to fuse input image features and language semantic information to output semantic labels for components on the surface of the aircraft CAD model. Step 1.4: Input the CAD model of the aircraft to be identified into the trained multimodal recognition model, identify the surfaces in the aircraft CAD model, and obtain the preliminary surface-component semantic label mapping relationship; Step 1.5: Based on the preliminary face-component semantic label mapping relationship, for faces whose recognition confidence output by the multimodal recognition model is lower than a preset threshold or whose geometric area ratio is less than a set threshold, the geometric rule auxiliary module is activated to perform semantic completion to obtain the final component segmentation result. Step 1.6: Based on the final component segmentation results, extract the key geometric feature lines corresponding to each component from the boundary representation structure of the aircraft CAD model; Step 1.7: Convert the final component segmentation results and the extracted key geometric feature lines into a structured engineering semantic format for the target mesh generation software to read and utilize, and then output it.

2. The fully automatic unstructured mesh generation method according to claim 1, characterized in that, The geometric rule auxiliary module determines and corrects the semantic labels of the surface components based on the geometric topological relationships and parametric features of the aircraft CAD model, thus forming the final component segmentation result.

3. The fully automatic unstructured mesh generation method according to claim 1, characterized in that, Step 1.4 specifically involves identifying the main structural component surfaces in the aircraft CAD model; Step 1.5 specifically involves semantic completion of local feature surfaces, connecting area surfaces, or low-confidence identification surfaces in the aircraft CAD model.

4. The fully automatic unstructured mesh generation method according to claim 2, characterized in that, The rules for determining the geometric topological relationships and parametric features of the aircraft CAD model include at least one of the following rules: Rule 1: If a face is located at the geometric intersection of two identified main structural component faces, and the area of ​​the face is less than a set threshold relative to the area of ​​the adjacent main structural component face, then the component semantic label of the face is determined as a connecting feature face or an edge face. Rule 2: If a face is surrounded by an identified main structural component face, and the spatial normal of the face and the main direction of the main structural component face meet the preset conditions, then based on the geometry of the face and the relationship with adjacent faces, the semantic label of the face is determined to be a hatch, cover, or other feature face with opening, closing, or maintenance functions.

5. The fully automatic unstructured mesh generation method according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Obtain multiple labeled sample parts, and convert the geometric data of each sample part into a structured feature representation for processing by a deep learning model to obtain a training set. The structured feature representation includes at least one of the graph structure, point cloud, and feature vector of the sample part; the labeling includes the type and spatial location information of the geometric defects present in the sample part. Step 2.2: Use the training set to train a multi-task hierarchical graph neural network model to obtain the trained multi-task hierarchical graph neural network model; Step 2.3: For any component obtained from the segmentation in Step 1, convert the geometric data of the component into the same structured feature representation as in Step 2.1, and input the converted data into the trained multi-task hierarchical graph neural network model; the trained multi-task hierarchical graph neural network model outputs the type and spatial location information of the geometric defects contained in the component; Step 2.4: Based on the geometric defect type and spatial location information output in Step 2.3, firstly, execute the topology repair algorithm for defects identified as gaps to restore the topological connectivity at that location; then, execute the corresponding geometric repair algorithm for defects identified as geometric anomalies, and perform mesh stitching on the repaired results of all components to generate a component-level surface mesh.

6. The fully automatic unstructured mesh generation method according to claim 5, characterized in that, Step 2.4 specifically includes: Step 2.4.1: Based on the spatial location information of the gap, call the gap repair algorithm at the boundary representation level to restore the topological connection of adjacent surfaces and obtain the intermediate model after B-Rep level topological repair; Step 2.4.2: The intermediate model is triangulated to obtain a surface mesh; based on the spatial location information of holes, overlaps and fragments, the hole-filling algorithm, the algorithm for deleting or merging overlapping patches, and the algorithm for fusing or removing fragments are executed sequentially on the surface mesh. Step 2.4.3: Perform a mesh stitching operation at adjacent boundaries on the surface meshes corresponding to all components that have completed the operation in Step 2.4.2 to generate component-level surface meshes.

7. The fully automatic unstructured mesh generation method according to claim 1, characterized in that, The neural network model in step 3 is a multi-task graph neural network model; The grid region types of the aforementioned boundary lines are classified as follows: Using each boundary line as a node and the topological connections between the boundary lines as edges, construct a boundary line topology graph; The boundary line topology map and the feature vectors of each node are input into the multi-task graph neural network model; wherein, the feature vectors of each node are formed based on the geometric features extracted from the geometric data of the corresponding boundary line, the semantic labels of the component in which they are located, and the computing conditions; Using the grid region type classification head in the multi-task graph neural network model, each node in the boundary line topology graph is classified into two categories, and the probability that each boundary line belongs to an isotropic region or an anisotropic region is output as the grid region type classification result.

8. The fully automatic method for generating unstructured meshes according to claim 1, characterized in that, The generation steps of the boundary layer mesh in step 4 specifically include: Based on the optimized surface mesh, determine the surface region where boundary layer generation is required; Boundary layer parameters are matched or specified according to preset simulation conditions, and the boundary layer parameters include at least the first layer mesh height, mesh growth rate and number of growth layers; On the surface grid nodes of the object surface region, multiple prismatic grids are generated by stretching layer by layer along the normal direction of the object surface to form a boundary layer grid. The generated boundary layer mesh is subjected to quality checks and optimizations. The quality checks include at least the mesh distortion, aspect ratio, and orthogonality index. The outer surface of the optimized boundary layer mesh is used as the input boundary for subsequent volume mesh generation.

9. The fully automatic method for generating unstructured meshes according to claim 1, characterized in that, The specific methods for generating the three-dimensional volume mesh in step 4 include: The non-object surface regions in the optimized surface mesh and the outer surface of the boundary layer mesh together constitute a closed solution domain boundary; Within the closed solution domain boundary, a spatial mesh filling algorithm is invoked to generate a tetrahedral mesh to fill the core flow field region; At the interface between the boundary layer mesh and the tetrahedral mesh, a pyramidal transition mesh is generated to achieve a smooth connection of mesh types; The tetrahedral mesh and the pyramid transition mesh together constitute the three-dimensional volume mesh of each sub-component.

Images