Intelligent generation method for pure quadrilateral partition based on diffusion model

Abstract

The intelligent generation method for the pure quadrilateral partition based on the diffusion model comprises the following steps: S1, constructing a global integer linear programming model containing a plurality of objective functions, and generating a quadrilateral grid with symmetric and aligned singular points by minimizing the objective functions; s2, obtaining a two-dimensional profile diagram of the CAD model according to the open source three-dimensional CAD model data set ABC; generating a triangular mesh model according to the two-dimensional contour map; s3, boundary fairing processing is carried out on the triangular mesh model, and hole digging and random segmentation are carried out in the triangular mesh model; s4, generating a symbol distance field according to the triangular mesh model; extracting pure quadrilateral partitions according to the quadrilateral grids; obtaining a model outer contour image according to the pure quadrilateral partition; s5, constructing a pure quadrilateral partition to generate a diffusion model; and training a pure quadrilateral partition by using the symbol distance field and the model outer contour image to generate a diffusion model. According to the method, the quick generation of the pure quadrilateral partition is realized, and meanwhile, the training time of the diffusion network model is greatly shortened.

Description

Technical Field

[0001] The present invention relates to the fields of computer graphics and deep learning, and in particular to a structured partitioning generation method and system for complex geometric models, which are used to improve mesh quality and numerical simulation efficiency. Background Art

[0002] In the fields of computer graphics and engineering simulation, quadrilateral meshes are widely used in finite element analysis, geometric modeling, animation production and other scenarios due to their advantages in boundary alignment, surface direction control, geometric accuracy and numerical calculation performance.

[0003] In the process of quadrilateral mesh generation in the existing technology, partitioning is a key pre-step. For complex geometric models, it is very difficult to directly generate a global quadrilateral mesh because multiple conflicting optimization goals need to be met at the same time, such as uniformity of cell size, optimization of cell quality, rationality of topological structure, and boundary alignment. By dividing the model into multiple smaller partitions, each partition can generate a quadrilateral mesh independently, thereby decomposing the complex global problem into multiple relatively simple local problems. After partitioning, each partition can generate a mesh in parallel, thereby speeding up the mesh generation speed and significantly improving the efficiency of mesh generation. In addition, the quadrilateral mesh usually needs to be aligned with the boundaries and directions of the model to ensure the geometric accuracy and numerical stability of the mesh. Partitioning can better achieve this alignment and direction control, thereby generating high-quality quadrilateral meshes. Good partitioning should solve quadrilaterals as much as possible, but the number of partitions should be as small as possible.

[0004] However, generating high-quality pure quadrilateral partitions is a complex and challenging task. Traditional methods such as frame field-based methods have the following problems:

[0005] (1) High computational complexity: The calculation of the frame field depends on the background triangular mesh and requires iterative optimization of the energy function. The computational efficiency is low, especially when dealing with complex geometric models, the computational time increases significantly.

[0006] (2) Unstable partitioning results: The solution of the frame field is easily affected by the quality of the background triangular mesh and the initial conditions. The same geometric model may produce different results under different mesh partitioning, which reduces the reliability and predictability of the method. Summary of the Invention

[0007] This paper proposes for the first time a new technical route for using the diffusion model to generate pure quadrilateral partitions. By fine-tuning the Stable Diffusion network pre-trained on massive data and fully tapping its rich prior knowledge, it achieves the rapid generation of pure quadrilateral partitions of the model, while significantly shortening the training time of the diffusion network model and demonstrating a certain generalization ability.

[0008] This method extracts a large number of 3D models from the open-source dataset ABC and converts them into symmetric planar models, generating symmetric pure quadrilateral partitions. It then optimizes the quadrilateral meshing algorithm to generate quadrilateral meshes with symmetry and alignment around singular points. It then reverse-extracts motorcycle graphs as pure quadrilateral partitions. Experimental results demonstrate that the dataset generated by this method enables the network to rapidly generate pure quadrilateral partitions that outperform traditional frame field methods, addressing the limitations of existing techniques and improving the consistency and stability of network output.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0010] The method for intelligently generating pure quadrilateral partitions based on the diffusion model includes the following steps:

[0011] S1. Building a global integer linear programming model based on the existing partitions, wherein the global integer linear programming model includes multiple objective functions, and generating a quadrilateral mesh with symmetry and alignment of singular points by minimizing the objective functions;

[0012] S2. Obtain a two-dimensional contour image of the CAD model based on the open source three-dimensional CAD model dataset ABC; and generate a triangular mesh model based on the two-dimensional contour image;

[0013] S3, performing boundary smoothing processing on the triangular mesh model, and then performing hole digging and random segmentation inside the triangular mesh model;

[0014] S4, generating a signed distance field according to the triangular mesh model; extracting pure quadrilateral partitions according to the quadrilateral mesh; and obtaining a model outer contour image according to the pure quadrilateral partitions;

[0015] S5. Constructing a pure quadrilateral partition generation diffusion model; and training the pure quadrilateral partition generation diffusion model using the signed distance field and the model outer contour image.

[0016] Preferably, in step S3, 3-7 circular holes are dug relative to the geometric center of the triangular mesh model to simulate the typical features of cooling holes and weight reduction holes in the CAD model; according to the normalized coordinates of the geometric center of the model, the spacing of the circular holes is limited to meet the minimum spacing constraint d min ≥L, L is the minimum length and width of the triangular mesh model; the aperture of the circular hole is 0.08L-0.2L.

[0017] Preferably, the random segmentation in step S3 includes: calculating the symmetry axis of the triangular mesh model, taking a random offset in the normal direction of the triangular mesh model and adding it to the symmetry axis of the triangular mesh model to obtain the symmetry axis of the new triangular mesh model; after segmenting the left half of the triangular mesh model along the symmetry axis of the new triangular mesh model, the remaining right half is mirrored to form a complete symmetrical triangular mesh model.

[0018] Preferably, the resolution of the signed distance field and the model outline image are both 512×512.

[0019] Advantageously, the pure quadrilateral partition generative diffusion model comprises a variational autoencoder for encoding the pure quadrilateral partition and the signed distance field map into a latent space.

[0020] Preferably, the two single-channel images of the model outer contour image and the signed distance field image are spliced ​​along the channel dimension to form a dual-channel input image; the dual-channel input image is the input of the variational autoencoder.

[0021] Preferably, the objective function includes a parity constraint function, a consistency constraint function, a regularity constraint function, a singular point alignment constraint function and a balance objective function; the balance objective function is composed of an equality objective, a regularity objective and a singular point alignment objective, and is weighted and summed by parameters.

[0022] Preferably, there are multiple two-dimensional contour maps; a symmetry detection algorithm is used to screen the two-dimensional contour maps according to a preset symmetry threshold; the screened two-dimensional contour maps are cropped along the symmetry axis of the two-dimensional model, retaining only the right half of the two-dimensional contour maps; and the triangular mesh model based on image pixels is extracted from the cropped two-dimensional contour maps.

[0023] The beneficial effects of the present invention are:

[0024] 1. Unlike existing pure quadrilateral partition generation methods, this method utilizes an end-to-end mapping technique based on a deep learning model. This allows the partition generation process to avoid complex frame field calculations and reliance on a background triangular mesh, significantly reducing computational resource consumption. High-quality structured partitions can be generated in a very short time. Experiments show that the diffusion model-based pure quadrilateral partition generation method takes only about one-tenth the time of traditional frame field methods on average, significantly improving processing efficiency.

[0025] 2. Unlike existing methods that generate purely quadrilateral partitions, this method generates diverse training data through operations such as internal model hole digging and random segmentation. It improves the quadrilateral mesh generation algorithm to effectively generate quadrilateral meshes with symmetry and alignment at singular points. It innovatively uses a technical solution that extracts MotorcycleGraphs from quadrilateral meshes with symmetry and alignment at singular points as the model partitioning standard for network learning. This results in neatly shaped partitions with clear boundaries and a minimal number of partitions. This method significantly outperforms traditional frame field methods, demonstrates good generalization capabilities on both symmetric and asymmetric models, and can handle a variety of complex geometric features. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A network structure is generated for the pure quadrilateral partitioning based on Stable Diffusion in Example 1 of the present invention;

[0027] Figure 2 This is a flow chart of the technical solution of Example 1 of the present invention;

[0028] Figure 3 This is an example of singular point alignment constraint in embodiment 1 of the present invention;

[0029] Figure 4 This is the main process of obtaining training data in Example 1 of the present invention;

[0030] Figure 5 A schematic diagram of forming multiple similar models by digging holes inside the same model according to Example 1 of the present invention;

[0031] Figure 6 This is a schematic diagram of a model expanded by cutting according to Example 1 of the present invention;

[0032] Figure 7 Schematic diagram of the quadrilateral partitioning results of the model based on the frame field (left) and network intelligent generation (right) according to Example 1 of the present invention;

[0033] Figure 8 Schematic diagram of the generalization ability of the test network for CAD models in Example 1 of the present invention. DETAILED DESCRIPTION

[0034] In order to make the technical means, creative features, objectives and effects of the invention easier to understand, the present invention is further described with reference to specific figures. However, the present invention is not limited to the following implementation cases.

[0035] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention.

[0036] In recent years, deep learning technology has made significant progress in image processing and model processing. Diffusion models, through a progressive generative paradigm, have demonstrated powerful generative capabilities in tasks such as image generation, video generation, and 3D model generation. This paper innovatively proposes using diffusion models to intelligently partition geometric models. By taking the boundary information and internal features of the geometric model as input in the form of images, the diffusion model can learn the mapping from geometric features to partition results, thereby achieving efficient and stable partition generation, providing a new method for automated partition generation of geometric models.

[0037] Example 1:

[0038] like Figure 1 As shown in the figure, the present invention proposes a pure quadrilateral partition generation network structure based on Stable Diffusion. This structure fully utilizes the rich prior knowledge contained in the latent space of the Stable Diffusion model pre-trained on massive data to effectively recover clear partition information from noisy data.

[0039] like Figure 2 The method for intelligently generating pure quadrilateral partitions based on a stable diffusion model is shown, and includes the following steps:

[0040] Step S1, optimizing the quadrilateral mesh generation algorithm to generate a quadrilateral mesh with symmetry and alignment of singular points;

[0041] To generate a quadrilateral mesh with symmetry and alignment of singular points, this paper constructs a global integer linear programming (ILP) model based on the partitioning obtained by traditional methods to determine the number of discrete mesh edges for each partition line. The ILP model optimizes the topology of the quadrilateral mesh by defining and minimizing multiple objective functions to ensure symmetry and alignment of singular points. Specific objective functions include:

[0042] S1-1. Parity Constraint: During quadrilateral mesh generation, the sum of the discrete segments of a partition's edges must be an even number to ensure topological closure. Therefore, a parity constraint is imposed on each partition's boundary edge group to ensure the subsequent generation of a quadrilateral mesh.

[0043] S1-2, consistency constraint: quantify the deviation between the actual number of discrete segments of each partition edge and the preset ideal value through a weighted penalty term to optimize the regularity of the grid structure:

[0044]

[0045] Where ei represents the ideal number of discrete grid segments for the ith edge.

[0046] S1-3. Regularity constraint: The objective function for partitions with different numbers of edges is defined as the sum of the absolute values ​​of the differences between discrete segments of the edges.

[0047] Regularity objective function for three-sided partitioning:

[0048]

[0049] Regularity objective function for four-side partitioning:

[0050]

[0051] Regularity objective function for five-sided partitioning:

[0052]

[0053] In view of the particularity of hexagonal partitioning, the present invention proposes an additional constraint: the sum of the number of edges of the three interval sides of the partition must satisfy the parity matching. Specifically, if the sum of the number of edges of the three interval sides is odd, then the sum of the remaining three sides must also be odd, otherwise the sum must be even. To this end, a binary auxiliary variable is introduced 0,1}, whose value corresponds to the parity of the sum of the interval edges. By adding the parity constraint of this variable to the objective function, the regularity condition of the hexagon can be enforced. Finally, the regularity objective function of the hexagonal partition can be formalized as:

[0054]

[0055] in, ,h∈Z.

[0056] S1-4, Singular point alignment goal: This goal requires that the singular points in the generated quadrilateral mesh are aligned with each other as much as possible. The goal is to align the singular points in two adjacent non-quadrilateral partitions with each other so that symmetrical and small high-quality partitions can be efficiently extracted from the quadrilateral mesh.

[0057] Taking two opposing non-T-shaped partitions as an example (this objective function is not applied to partitions containing T-shaped nodes. A T-shaped partition is a special mesh unit in which one edge is shared by two other edges, forming a shape similar to the letter "T". This type of partition may introduce some discontinuities or irregularities in the mesh, thereby affecting the quality of the mesh and subsequent computational analysis.), the present invention introduces the following penalty terms on the i-th side and the j-th side of the partition respectively to ensure that the topological lines flowing out of the singular points of the two partitions can be aligned with each other, thereby generating a quadrilateral mesh with aligned singular points. Figure 3 An example of this situation is shown.

[0058]

[0059] in, represents the number of discrete grid segments on the sides of the first partition, Represents the number of discrete grid segments on the sides of the second partition.

[0060] In addition, considering the symmetric nature of the partition, the partitions on both sides of the symmetry axis are adjacent and have the same number of sides, so a weight is multiplied on the basis of the above formula :

[0061]

[0062] Where Ni represents the number of edges in one partition, and Nj represents the number of edges in its adjacent partition.

[0063] S1-5. Balance objective function: The final balance objective function consists of the equality objective, the regularity objective, and the singular point alignment objective, which are weighted summed by parameters:

[0064]

[0065] in In the present invention, and The ratio of the two parameters is about 1:50.

[0066] Step S2: Based on the open source 3D CAD model dataset ABC (A Big CAD Model Dataset), multiple 2D contour images are obtained, and symmetrical contours are screened based on the 2D contour images to generate a rough triangular mesh model;

[0067] The complex three-dimensional model in the open source dataset ABC is converted into a symmetrical two-dimensional model, preparing for the subsequent generation of symmetrical pure quadrilateral partitions, so as to facilitate the evaluation of the quality of the generated partitions and improve the consistency and stability of the network output results. Figure 4 As shown, the specific implementation steps of the present invention to obtain the triangular mesh model are as follows:

[0068] S2-1. 3D to 2D projection: Orthogonally project the 3D model obtained from the open source 3D CAD model dataset ABC onto three coordinate planes (XOY, XOZ, YOZ) to generate the corresponding 2D contour map.

[0069] S2-2, Symmetry detection: Use a symmetry detection algorithm to screen out two-dimensional images with high symmetry according to a preset symmetry threshold.

[0070] S2-3. Image cropping: crop the filtered two-dimensional image along the model symmetry axis, retaining only the right half for subsequent symmetry operations.

[0071] S2-4. Extracting a rough triangular mesh model with a jagged outline boundary based on image pixels from the cropped two-dimensional image.

[0072] Step S3: Smoothe the triangular mesh model's boundaries, dig holes within the triangular mesh, and perform random segmentation to further enhance data diversity and complexity. The triangular mesh model from step S2 is then remeshed and Laplacian smoothed to improve mesh quality. Some circular holes are dug within the triangular mesh model based on relative positions, and the triangular mesh is randomly segmented. The model is then symmetrically reshaped to form a triangular mesh with a new outer contour.

[0073] S3-1. Hole generation inside the model: dig N holes relative to the geometric center inside the triangular mesh model. [3,7] (one hole is dug out from the geometric center, and the remaining holes are symmetrically distributed relative to the geometric center) circular holes are used to simulate typical features such as cooling holes and weight reduction holes in most CAD models, such as Figure 5 As shown. According to the normalized coordinates of the model geometric center, the hole spacing is limited to meet the minimum spacing constraint d min ≥L (L is the minimum length and width of the triangular mesh model), to avoid hole overlap or edge collapse. Randomly set the aperture r [0.08L,0.2L].

[0074] S3-2. Symmetry axis cutting extension: Calculate the symmetry axis A of the triangular mesh model main , take a random offset ∆A in the normal direction to get the new symmetry axis A cut :

[0075]

[0076] Where ∆A (−0.3L, 0.3L), for the triangular mesh model along Acut After splitting the left half, the remaining right half is mirrored to form a complete symmetrical triangular mesh model, such as Figure 6 shown.

[0077] This method can generate a series of new triangular mesh models with similar but not identical geometric features, which not only significantly increases the size of the dataset, but also provides more diverse training samples for the neural network, effectively increasing the diversity of the dataset, helping to improve the model's generalization ability for different geometric characteristics, and thus improving the robustness of the intelligent generation of model partitions. It is worth noting that the singular structures around holes often show a certain regularity and repeatability. These features are valuable learning materials for deep learning because they provide clear spatial relationships and geometric constraints, which help the model learn more accurate and robust representations. This allows for a better understanding and characterization of complex singular structure distributions;

[0078] Step S4: constructing pure quadrilateral partitions from the quadrilateral mesh based on the signed distance field map of the triangular mesh generation model and the quadrilateral mesh with symmetry and alignment of the singular points;

[0079] Based on the triangular mesh model obtained in step S3, a signed distance field based on the triangular mesh is generated and an image of size 512*512 is used as the input of the subsequent network;

[0080] The steps of generating a signed distance field based on a triangular mesh model are as follows:

[0081] Input: Closed 2D triangular mesh model Output: Image resolution is The signed distance field

[0082] S4-1. The specific implementation process of the signed distance field generation algorithm based on the triangular mesh is as follows:

[0083] 1. Bounding box calculation based on triangle mesh vertices:

[0084] , , ,

[0085] Where V is the set of vertices of the triangle mesh

[0086] 2. Define pixel coordinates , scale the coordinates of the vertices in the bounding box to In-plane, where:

[0087] 3. Nearest triangle face search:

[0088] Traverse the triangle patch, if Inside the model (determined by the triangle normal direction: ) Calculate according to step 4 arrive Projection distance ;if Outside the model, directly set the distance to

[0089] 4. Triangle projection distance calculation (for triangle and point ), calculation steps:

[0090] Calculate the vectors formed by the three sides of a triangle:

[0091] Calculate the center of gravity coordinates :

[0092]

[0093] if :

[0094] Otherwise, proceed to the next step. The shortest distance to the three edges.

[0095] 5. Edge projection distance calculation (opposite side and point )point The projection is:

[0096]

[0097] in,

[0098]

[0099]

[0100] S4-2. Based on the quadrilateral mesh algorithm optimized in step S1, a quadrilateral mesh with symmetric and aligned singular points is generated. Motorcycle Graphs are extracted from this quadrilateral mesh as pure quadrilateral partitions of the model. Motorcycle Graphs are essentially the same as pure quadrilateral partitions obtained from traditional frame fields. Furthermore, partitions extracted from quadrilateral meshes have the advantage of being close to quadrilaterals, meeting the basic requirements for partitioning. Motorcycle Graphs are constructed based on quadrilateral meshes. The core concept is to use singular points and boundary concave points in the quadrilateral mesh as starting points. From these key locations, "motorcycles" are launched along the edges of the quadrilateral mesh. These "motorcycles" move along predefined paths within the mesh, the paths selected being based on the topological structure and geometric characteristics of the quadrilateral mesh. During their movement, the "motorcycles" continuously track the mesh boundaries and internal structure until they encounter the model boundary or collide with the paths of other "motorcycles." These closed paths divide the entire model into multiple partitions. The boundaries of each sub-region are defined by the paths of the "motorcycles", and since these paths are based on quad meshes, this ensures that the shape of the partitions is as close to a quad as possible.

[0101] The steps for constructing Motorcycle Graphs based on quadrilateral meshes are as follows:

[0102] Input: A quadrilateral mesh with symmetric and aligned singular points

[0103] Output: Image resolution is Model Motorcycle Graphs

[0104] 1. The system initializes and establishes a set of visited quadrilateral mesh vertices, Vvisited, initially adding all singular points (vertices with degree not equal to 4) and concave points (boundary points with degree greater than 2). It also establishes a set of visited quadrilateral edges, Evisited, initially adding all boundary edges. It also establishes a queue of currently processed extended edges, Qcurrent, with all edges emitted by singular points as initial elements. It also establishes a queue of next round of processing, Qnext, initialized to the empty set ∅.

[0105] 2. Loop through the extended edges to determine whether the current processing queue Qcurrent is empty: If Qcurrent is not empty, execute steps 3 to 11 in a loop.

[0106] 3. Edge element traversal: Take each edge e = (u, v) from Qcurrent in turn for processing;

[0107] 4. Find the forward vertex of the next extended edge and determine the next unvisited mesh vertex w along the direction of edge e, that is, w is not in Vvisited. If there is no vertex w that meets the conditions, skip the current edge processing and return to step 3 to process the next edge.

[0108] 5. Find the next extended edge based on the adjacent quadrilateral faces of the edge and find its adjacent faces based on edge e. Make , ; Determine the next edge e′=(w,x) based on the topological relationship of the surface;

[0109] 6. Queue update: add the successor edge e′ to the next round of processing queue Qnext;

[0110] 7. The visit status mark adds the vertex w to the visited vertex set Vvisited; adds the current edge e to the visited edge set Evisited;

[0111] 8. The Motorcycle edge record stores the ordered edge pair (e, e′) into the output edge set Emotor;

[0112] 9. The single-edge processing is completed. The current edge e is processed and the process returns to step S103 to continue processing the next edge in the queue.

[0113] 10. Process the round update and assign the Qnext queue to Qcurrent; reset Qnext to the empty set ∅;

[0114] 11. Algorithm Termination When Qcurrent is empty, the algorithm ends. The final Emotor is output as the edge set that constitutes the Motorcycle Graphs. The boundary edges of the Motorcycle Graphs and the model divide the entire model into multiple closed pure quadrilateral regions, which are the final model quadrilateral partitions. After obtaining the edge set of the Motorcycle Graphs, the Motorcycle Graphs are drawn with a resolution of image.

[0115] The Motorcycle Graphs obtained through the above steps can efficiently extract partition boundaries from the quadrilateral mesh. These partition boundaries not only have clear geometric meanings, but are also closely related to the singular points and boundary concave points of the quadrilateral mesh, thus ensuring the rationality and effectiveness of the partitioning.

[0116] Step S5: construct and train the pure quadrilateral partition generation diffusion model proposed in the present invention.

[0117] This invention innovatively converts discretized triangular mesh models into fixed 512×512 resolution signed distance field (SDF) images and contour images suitable for diffusion model network input. This fundamentally eliminates the uncertainty in representation of the same model due to variations in triangular mesh discretization. By encoding the entire topological and boundary information of the geometry in image form, it enables direct use of established, high-performance image diffusion networks for training and inference. This fully utilizes the parallel acceleration capabilities of GPUs, reducing the cost of data augmentation (rotation, scaling, and cropping) to near-zero.

[0118] The pre-trained text-to-image latent diffusion model (LDM) is used and fine-tuned to adapt it to the partition intelligent generation task. The architecture contains a frozen variational autoencoder (VAE) that is used to extract the model's outer contour (after extracting the model's boundary edges, the model is scaled down in the same way as the MotorcycleGraphs image is generated). The proposed method encodes the boundary map and the signed distance field map (SDF) into the latent space to train the conditional denoising diffusion model. Since the VAE encoder was originally designed for single-channel input, and the input data of the present invention includes two single-channel images (i.e., the boundary map and the signed distance field map), the present invention concatenates these two single-channel images along the channel dimension to form a dual-channel input.

[0119] In the denoising process, the present invention achieves conditional encoding of images by modifying the U-Net structure, thereby significantly improving the training efficiency and maintaining the ability to generate high-resolution images. Specifically, this improvement enables the model to better handle complex geometric information and accelerates the convergence speed while ensuring the quality of generation. Figure 7 As shown in the figure, the traditional frame field takes an average of 43 seconds (for a model with about 10,000 triangular mesh units) to generate quadrilateral partitions. The method of the present invention is tested on an Intel i9-12900K + RTX 4090 and takes only about 2.3 seconds on average. This method not only simplifies the processing flow of input data, but also enhances the adaptability and robustness of the model to different geometric structures, such as Figure 8 As shown, it provides strong support for quadrilateral partition generation.

Claims

1. A pure quadrilateral partition intelligent generation method based on a diffusion model, characterized by: The following steps are involved: S1. Building a global integer linear programming model based on the existing partitions, wherein the global integer linear programming model includes multiple objective functions, and generating a quadrilateral mesh with symmetry and alignment of singular points by minimizing the objective functions; S2. Obtain a two-dimensional contour image of the CAD model based on the open source three-dimensional CAD model dataset ABC; and generate a triangular mesh model based on the two-dimensional contour image; S3, performing boundary smoothing processing on the triangular mesh model, and then performing hole digging and random segmentation inside the triangular mesh model; S4, generating a signed distance field according to the triangular mesh model; extracting pure quadrilateral partitions according to the quadrilateral mesh; and obtaining a model outer contour image according to the pure quadrilateral partitions; S5. Constructing a pure quadrilateral partition generation diffusion model; and training the pure quadrilateral partition generation diffusion model using the signed distance field and the model outer contour image.

2. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 1, characterized in that: In step S3, 3-7 circular holes are dug relative to the geometric center of the triangular mesh model to simulate the typical features of cooling holes and weight reduction holes in the CAD model; the spacing of the circular holes is limited to meet the minimum spacing constraint d min Greater than or equal to the minimum value L of the length and width of the triangular mesh model; the aperture of the circular hole is 0.08L-0.2L.

3. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 2, characterized in that: The random segmentation in step S3 includes: calculating the symmetry axis of the triangular mesh model, taking a random offset in the normal direction of the triangular mesh model and adding it to the symmetry axis of the triangular mesh model to obtain the symmetry axis of the new triangular mesh model; after segmenting the left half of the triangular mesh model along the symmetry axis of the new triangular mesh model, the remaining right half is mirrored to form a complete symmetrical triangular mesh model.

4. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 1, characterized in that: The resolutions of the signed distance field and the model outline image are both 512×512.

5. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 4, characterized in that: The pure quadrilateral partition generative diffusion model includes a variational autoencoder for encoding the pure quadrilateral partition and the signed distance field map into a latent space.

6. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 5, characterized in that: The two single-channel images of the model outer contour image and the signed distance field image are spliced ​​along the channel dimension to form a dual-channel input image; the dual-channel input image is the input of the variational autoencoder.

7. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 1, characterized in that: The objective function includes a parity constraint function, a consistency constraint function, a regularity constraint function, a singular point alignment constraint function and a balance objective function; the balance objective function is composed of an equality objective, a regularity objective and a singular point alignment objective, and is weighted and summed by parameters.

8. The method for intelligently generating pure quadrilateral partitions based on a diffusion model according to claim 1, characterized in that: There are multiple two-dimensional contour images; using a symmetry detection algorithm, according to a preset symmetry threshold, the two-dimensional contour images are screened; the screened two-dimensional contour images are cropped along the symmetry axis of the two-dimensional model, and only the right half of the two-dimensional contour images is retained; The triangular mesh model based on image pixels is extracted from the cropped two-dimensional contour image.

Images