High-efficiency and high-fidelity rendering method for large-scale CAE (Computer Aided Engineering) simulation data

Abstract

The invention relates to a high-efficiency high-fidelity rendering method for large-scale CAE simulation data, and the method achieves the high-efficiency watertight rendering through two core stages: stage one, watertight texture reconstruction: parameterizing each base grid piece to a square texture domain with an aligned axis, and converting a texture coordinate into an integer to avoid the precision error of a floating-point number; the boundary displacement values of the adjacent grid pieces are ensured to be consistent in combination with manual bilinear interpolation; in the second stage, self-adaptive subdivision factor calculation is carried out, by mapping a screen space pixel error into a world space geometric error, combining the characteristics of unique identification codes of the patches, utilizing a three-layer full-connection neural network to predict an optimal subdivision factor, averaging the subdivision factors of shared edges of adjacent patches and then rounding upwards. According to the method, real-time crack-free rendering of a large-scale model is realized, visual fidelity and calculation efficiency are considered, the method is highly compatible with a standard hardware subdivision pipeline, and the method can be widely applied to the fields of computer graphics, virtual reality, digital content creation and the like.

Description

Technical Field

[0001] This invention relates to an efficient and high-fidelity rendering method for large-scale CAE (large-scale discrete model) simulation data, which is applicable to scenarios requiring high-quality 3D model visualization, such as real-time rendering, virtual reality, and digital content creation, and belongs to the field of computer graphics and computational geometry technology. Background Technology

[0002] Discrete meshes, due to their regular structure and strong compatibility, have become the mainstream standard for 3D geometric representation and rendering, and are widely used in fields such as computer graphics, virtual reality (VR), augmented reality (AR), and digital content creation (DCC). In 3D model rendering, to enhance visual realism, it is usually necessary to enhance geometric details through high-resolution meshes or auxiliary texture techniques (such as displacement mapping). However, such methods always face a triple trade-off between accuracy, efficiency, and visual consistency. Especially in real-time rendering scenarios, the constraints of limited hardware resources make the "balance between fidelity and performance" a long-standing and unresolved core challenge.

[0003] Existing discrete mesh-based rendering techniques suffer from two major, unavoidable, core flaws:

[0004] 1. UV coordinate seam and texture discontinuity issues

[0005] Discrete meshes require UV parametric mapping to project 2D textures onto 3D surfaces. However, UV segmentation of non-zero genus or complex topological models inevitably introduces seams. These seams not only cause visual breaks in material properties such as color, normals, and roughness, but also lead to serious problems in displacement mapping or surface subdivision techniques: the lack of geometric consistency between adjacent UV islands causes vertex misalignment, resulting in visible cracks or geometric tears. The root causes are twofold: first, traditional parametric algorithms cannot guarantee that adjacent mesh patches have consistent sampling topology on the shared 3D boundary UV boundaries, leading to differences in sampling displacement values; second, normalized floating-point texture coordinates have precision errors, and small calculation deviations during boundary sampling can cause texture interpolation to select different pixel neighborhoods, ultimately resulting in inconsistent displacement values ​​and compromising surface watertightness.

[0006] 2. The problem of difficulty in accurately determining subdivision factors

[0007] Tessellation technology optimizes visual effects and computational performance by dynamically adjusting geometric details, but accurately determining the tessellation factor (i.e., the number of sub-triangles to be split in each facet) has always been a technical challenge. Traditional methods rely on approximate estimations using "empirical formulas" (such as those based on the straight-line distance between the object and the camera, and changes in viewpoint). These methods do not fully consider the actual pixel coverage of the target object on the screen and the dynamic changes in viewpoint, ignoring actual geometric errors in screen space. This results in poor tessellation control: either insufficient tessellation leads to loss of geometric details and visual distortion, or excessive tessellation wastes hardware resources and reduces rendering frame rate, making it impossible to achieve a balance between "pixel-level precision" and "efficient computation" in real-time rendering.

[0008] In existing technologies, there is a lack of unified and efficient solutions for texture reconstruction and subdivision factor selection. These solutions cannot avoid the watertightness degradation caused by UV seams, nor can they achieve adaptive and precise subdivision control, severely limiting the application of discrete meshes in high-quality real-time visualization scenarios. Therefore, there is an urgent need for a rendering framework that can adaptively perceive screen space errors, solve the UV seam problem, and be compatible with general-purpose graphics hardware, thus overcoming the bottlenecks of existing technologies. Summary of the Invention

[0009] To overcome the shortcomings of existing discrete mesh rendering techniques, this invention provides an efficient and high-fidelity rendering method for large-scale CAE simulation data. Without significantly increasing the complexity of existing rendering pipelines, it achieves real-time, crack-free rendering of large-scale models, balancing visual fidelity and computational efficiency. It is highly compatible with standard hardware subdivision pipelines and can be widely applied in fields such as computer graphics, virtual reality, and digital content creation.

[0010] An efficient and high-fidelity rendering method for large-scale CAE simulation data includes a watertight texture reconstruction stage and an adaptive subdivision factor calculation stage:

[0011] The watertight texture reconstruction stage simplifies the original high-poly mesh into a low-poly mesh, assigns an independent axis-aligned square texture domain to each quadrilateral face, converts texture coordinates into integer pixel coordinates, and generates displacement maps and normal maps by manually bilinear interpolation sampling and baking to ensure that the displacement values ​​of adjacent face boundaries are consistent.

[0012] The adaptive subdivision factor calculation stage maps screen space pixel error to world space geometric error. Combining the features of the unique identifier encoding of the face, the initial subdivision factor is predicted through a three-layer fully connected neural network. The subdivision factor of the shared edge of adjacent face is averaged and rounded up to obtain the final subdivision factor.

[0013] Based on the displacement map, normal map, and final subdivision factor, the discrete mesh model is rendered to achieve watertight and efficient subdivision rendering of large-scale discrete models.

[0014] The watertight texture reconstruction stage specifically includes:

[0015] Preprocess the high-poly mesh to reduce the number of faces, repair non-manifold structures, generate quadrilateral facets and assign independent square texture domains;

[0016] Convert texture coordinates to integers and manually implement sampling and interpolation based on integer pixel indices;

[0017] Baking sampling is performed to calculate the grid position and normal of the corresponding pixel, and displacement map and normal map are generated by intersecting rays.

[0018] The core objective is to continuously and stably map a simple base mesh plane onto a complex target surface using optimized displacement encoding technology, ensuring consistent boundary displacement values ​​between adjacent mesh patches and achieving a watertight connection.

[0019] By employing a specially optimized displacement mapping technique, one or more simple base mesh planes Σ are continuously and stably mapped (permuted) to a complex target surface Γ.

[0020] Discrete meshes are chosen as the representation method for the target surface. Discrete meshes possess a well-defined spatial discretization structure, with their geometry defined by the connections between a set of vertices, edges, and faces in three-dimensional space. This characteristic allows for control over the level of detail in the model by adjusting the mesh resolution. Even with lower vertex densities, discrete meshes can still maintain the overall topology and geometric contours of the original model macroscopically. This multi-resolution representation capability makes it ideal as a carrier for displacement mapping: a coarse, low-polygon mesh can be used to represent its basic shape, and then displacement mapping can be used to reconstruct the high-frequency geometric details of the surface during the shading stage, thus achieving a good balance between rendering visual quality and runtime performance.

[0021] However, while displacement mapping can conveniently recover high-precision details from coarse, low-precision meshes, it is highly susceptible to generating cracks at the UV boundaries of different base meshes (patches), thereby compromising the watertightness of the surface and causing noticeable visual defects. This problem mainly stems from two aspects. First, there is inconsistency in sampling rules. Traditional parametric algorithms primarily aim to optimize the area or angular distortion of UV mapping, but they usually cannot guarantee that different mesh patches have consistent sampling topologies for their corresponding UV boundaries on shared 3D boundaries. When a continuous edge in 3D space is mapped to two separate boundaries in UV space with different discretization methods, the sampling process will read different displacement values, causing the reconstructed surface to fail to close tightly at that point. Second, floating-point precision representation itself has errors. Even with consistent sampling rules, traditional normalized floating-point texture coordinates (usually within the [0, 1] range) themselves have precision issues. Because the binary representation of floating-point numbers cannot accurately represent all decimal fractions, small calculation errors during boundary sampling can cause the coordinates of a sampling point on one surface to be slightly larger than the boundary value, while the coordinates on adjacent surfaces may be slightly smaller than the boundary value. This results in different pixel neighborhoods being selected during texture interpolation, ultimately leading to inconsistent displacement values. To address these issues, two key strategies were designed to ensure the continuity of the displacement field and the watertightness of the reconstructed surface.

[0022] Inconsistent sampling rules: All base meshes must be parameterized to an axis-aligned square texture domain. Specifically, each base mesh is mapped to an independent square region [umin, vmin]∗ [umax, vmax], and the boundaries of the base mesh are strictly aligned with the four edges of this square texture domain. The advantage of this design is that the boundaries of the square parameter domain can be perfectly aligned with the pixel mesh after discretization. When two base meshes are adjacent in 3D space, their shared boundaries may correspond to different edges in their respective square texture domains (e.g., the right boundary of one corresponds to the left boundary of the other). However, since the sampling points always fall on the precise location of the pixel mesh, their sampling rules and patterns are completely equivalent. This orientation-independent and boundary-alignable characteristic provides a fundamental guarantee for achieving displacement field continuity across meshes.

[0023] Floating-point precision error: Texture coordinates are converted from normalized floating-point numbers to integers. Computers can accurately represent integers, thus avoiding the errors that occur when representing decimals with floating-point numbers. Instead of relying on the built-in bilinear interpolation based on floating-point coordinates in the graphics hardware, sampling and interpolation based on integer pixel indices are manually implemented in the program. This strategy ensures that the pixel indices of adjacent surfaces on shared boundaries are completely consistent. During interpolation calculations, they strictly reference the same set of pixels (Texels), ensuring that the displacement value calculation process for every point on the boundary is exactly the same, and the results are also strictly consistent. This fundamentally eliminates gaps caused by numerical uncertainty, resulting in a seamless connection of the displacement field at the boundaries.

[0024] In summary, by combining the strategies of square parameter domains and integer texture coordinates, it is ensured that any adjacent base meshes can obtain completely consistent displacement values ​​on the shared boundary. Based on this, as long as the same subdivision factor is used for all meshes in the subsequent tessellation stage, a globally continuous, smooth, and completely watertight final surface can be generated.

[0025] By employing a dual design of "square texture domain alignment" and "integer texture coordinates," the sampling rules of adjacent mesh patches on the shared boundary are consistent, and the pixel indices are perfectly matched, ultimately obtaining consistent displacement values. This eliminates cracks at the source and ensures the watertightness of the curved surface.

[0026] The adaptive subdivision factor calculation stage specifically includes:

[0027] Calculate the maximum geometric error in world space corresponding to the pixel error in screen space;

[0028] Extract the feature vectors of the facets and input them into a three-layer fully connected neural network model to obtain the initial subdivision factor;

[0029] Determine the final subdivision factor for shared edges and internal edges.

[0030] The core objective is to accurately predict the optimal subdivision factor based on screen space geometric errors and patch features, while balancing visual fidelity and computational efficiency.

[0031] The lower bound of this adaptive subdivision factor is determined by geometric error constraints. To ensure that the projection deviation of the subdivided mesh on the screen does not exceed a preset pixel threshold p, the pixel error in screen space needs to be mapped to a distance metric in the world coordinate system. Let the number of pixels in the vertical direction of the screen be... The camera's vertical field of view is FOV, and the depth of the nearest point of the current patch's limiting mesh is z. To ensure that the geometric deviation of the subdivided mesh projected onto the screen does not exceed one pixel, the world space length corresponding to one pixel at depth z can be denoted as... The calculation formula is as follows:

[0032]

[0033] Considering the worst-case diagonal direction deviation, the maximum geometric error in world space (i.e., the radius of the bounding sphere) corresponding to a p-pixel error can be conservatively estimated as:

[0034]

[0035] Where p is the pixel threshold and FOV is the camera's vertical field of view. z is the number of pixels in the vertical direction of the screen, and z is the spatial depth of the patch.

[0036] Set the pixel threshold p to 1 pixel, therefore This represents the maximum permissible geometric deviation in world space corresponding to one pixel at depth z. Any subdivision level below this error tolerance will result in geometric distortion.

[0037] Based on the aforementioned error constraints, and combined with the geometric and parametric information contained in the unique identifier (ID) assigned to each facet, a lightweight neural network model is constructed as features. These features are input into a three-layer fully connected network (MLP), with the structure: input layer > hidden layer > output layer, to obtain the subdivision factor t. During the subdivision process, the same subdivision factor is uniformly assigned to the boundary edges and internal edges of each facet. When adjacent faces specify inconsistent subdivision factors for shared edges, to ensure mesh continuity and crack-free rendering, the subdivision factor for the shared edge is taken as the arithmetic mean of the subdivision factors of the two facets, rounded up to the nearest integer, as the final subdivision level applied to the shared edge.

[0038] This strategy effectively avoids geometric gaps caused by inconsistent subdivision factors, while ensuring a good balance between visual fidelity and computational overhead. It predicts the optimal subdivision factor that meets visual fidelity requirements, thereby achieving adaptive control of the mesh subdivision level while satisfying pixel-level geometric accuracy requirements.

[0039] In the watertight texture reconstruction stage, the high-poly mesh is preprocessed, including simplifying the number of mesh faces, repairing non-manifold structures, intelligently pairing triangular faces into quadrilaterals, discarding the original UVs, and assigning non-overlapping rectangular tiles to each quadrilateral. The boundary of each quadrilateral face is strictly aligned with the four sides of the square texture domain. The shared 3D boundary of adjacent faces corresponds to different sides in their respective texture domains, and the sampling rules are completely equivalent.

[0040] In the watertight texture reconstruction stage, during the baking sampling process, the axis-aligned pixel bounding box surrounding the quadrilateral patch is calculated. Each pixel within the bounding box is traversed, and the corresponding grid position and normal are calculated using bilinear interpolation. The data is then written into the baking data.

[0041] In the adaptive subdivision factor calculation stage, the preset pixel threshold p is 1 pixel, which is used to determine the maximum allowable geometric deviation in world space.

[0042] The structure of the three-layer fully connected neural network model consists of an input layer, a hidden layer, and an output layer. The input is a feature vector, and the output is a continuous subdivision factor.

[0043] The input features of the three-layer fully connected neural network include the unique identifier (ID) of the face assignment, the maximum pixel geometric error in the screen space, and the MLP model.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] Achieve native watertight rendering and completely eliminate geometric cracks: Through the innovative design of "square texture domain parameterization" and "integer texture coordinates", the problem of inconsistent displacement values ​​caused by UV seams is solved from the aspects of sampling rules and numerical accuracy. Globally continuous and watertight rendering surfaces can be generated without any post-processing repair, and visual defects are completely eliminated.

[0046] Adaptive and precise subdivision, balancing fidelity and efficiency: Based on the quantization constraints of screen space geometric errors and combined with the MLP prediction model of the inherent features of the patches, it achieves "pixel-level precision" adaptive control of the subdivision factor—avoiding geometric distortion caused by insufficient subdivision and preventing resource waste caused by excessive subdivision. In real-time rendering scenarios (such as VR / AR), it can achieve the optimal balance between frame rate and visual quality.

[0047] High compatibility and scalability: The framework is highly compatible with standard hardware tessellation pipelines and can be directly integrated into existing rendering engines (such as Unity and Unreal Engine); it also supports integration with modern graphics technologies such as Compute Shader and Mesh Shader, enabling more advanced software-level surface tessellation optimization and adapting to different hardware platforms and application scenarios.

[0048] Highly efficient and easy to implement, suitable for large-scale models: The mesh simplification and texture baking process in the preprocessing stage is efficient and controllable, and the core algorithms in the rendering stage (manual interpolation, MLP prediction) have low computational overhead. It can support real-time rendering of large-scale discrete models with millions of faces, and does not rely on dedicated hardware, making it highly versatile. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This diagram illustrates how cracks can occur when using traditional UV mapping techniques for texture mapping of 3D models, due to the inability to guarantee consistent sampling topology between adjacent UV boundaries and the precision errors in floating-point texture coordinates.

[0051] Figure 2 This is a schematic diagram illustrating the principle of a high-efficiency, high-fidelity rendering method for large-scale CAE simulation data, which uses square parameter domain mapping to achieve watertight connections.

[0052] Figure 3 This is a schematic diagram of the mapping model from screen space pixel error to world space geometric error in an efficient and high-fidelity rendering method for large-scale CAE simulation data according to the present invention.

[0053] Figure 4 This invention presents a high-efficiency, high-fidelity rendering method for large-scale CAE simulation data. The watertight subdivision rendering framework yields a square parameter domain mapping effect diagram.

[0054] Figure 5 This is a flowchart illustrating the calculation of the adaptive subdivision factor in an efficient and high-fidelity rendering method for large-scale CAE simulation data according to the present invention.

[0055] Figure 6 The image shows a comparison of the rendering effects of the efficient and high-fidelity rendering method for large-scale CAE simulation data proposed in this invention (the left side shows the original surface, and the right side shows the actual rendering effect using the framework of this invention). Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Reference Figure 2 An efficient and high-fidelity rendering method for large-scale CAE simulation data is proposed, comprising a watertight texture reconstruction stage and an adaptive subdivision factor calculation stage, which solves the problem through square parameter domain mapping. Figure 1To address the cracking issue, ensure that adjacent UVs have a consistent sampling topology, thereby avoiding crack formation:

[0058] The watertight texture reconstruction stage simplifies the original high-poly mesh into a low-poly mesh, assigns an independent axis-aligned square texture domain to each quadrilateral face, converts texture coordinates into integer pixel coordinates, and generates displacement maps and normal maps by manual bilinear interpolation sampling and baking to ensure that the displacement values ​​of adjacent face boundaries are consistent. The core objective is to continuously and stably map simple base mesh planes to complex target surfaces through optimized displacement encoding technology, ensuring that the displacement values ​​of adjacent mesh boundaries are consistent and achieving watertight connection.

[0059]

[0060] By employing a specially optimized displacement mapping technique, one or more simple base mesh planes Σ are continuously and stably mapped (permuted) to a complex target surface Γ.

[0061] Discrete meshes are chosen as the representation method for the target surface. Discrete meshes have a well-defined spatial discretization structure, and their geometry is defined by a set of connections between vertices, edges, and faces in three-dimensional space. This characteristic allows for control over the level of detail of the model by adjusting the mesh resolution. Even with a low vertex density, discrete meshes can still maintain the overall topology and geometric contours of the original model macroscopically. This multi-resolution representation capability makes it very suitable as a carrier for displacement mapping: a coarse low-polygon mesh can be used to represent its basic shape, and then the high-frequency geometric details of the surface can be reconstructed during the shading stage using displacement mapping, thus achieving a good balance between rendering visual quality and runtime performance. However, although displacement mapping can easily recover high-precision details from coarse low-precision meshes, it is prone to generating cracks at the UV boundaries of different base meshes (patches), thereby compromising the watertightness of the surface and causing strong visual defects. This problem mainly stems from two aspects. First, there is inconsistency in sampling rules. Traditional parametric algorithms mainly aim to optimize the area or angular distortion of UV mapping, but they usually cannot guarantee that different mesh patches have consistent sampling topologies for their corresponding UV boundaries on shared 3D boundaries. When a continuous edge in 3D space is mapped to two separate boundaries in UV space with different discretization methods, the sampling process will read different displacement values, causing the reconstructed surface to fail to close tightly at that point. Secondly, floating-point precision representation itself has errors. Even with consistent sampling rules, traditional normalized floating-point texture coordinates (usually within the [0, 1] interval) also have precision issues. Because the binary representation of floating-point numbers cannot accurately represent all decimal fractions, small calculation errors during boundary sampling may cause the coordinates of a sampling point on one surface to be slightly larger than the boundary value, while the coordinates on adjacent surfaces may be slightly smaller than the boundary value. This results in different pixel neighborhoods being selected during texture interpolation, ultimately leading to differing displacement values. To address these issues, two key strategies were designed to ensure the continuity of the displacement field and the watertightness of the reconstructed surface.

[0062] Inconsistent sampling rules: All base meshes must be parameterized to an axis-aligned square texture domain. Specifically, each base mesh is mapped to an independent square region [umin, vmin]∗ [umax, vmax], and the boundaries of the base mesh are strictly aligned with the four edges of this square texture domain. The advantage of this design is that the boundaries of the square parameter domain can be perfectly aligned with the pixel mesh after discretization. When two base meshes are adjacent in 3D space, their shared boundaries may correspond to different edges in their respective square texture domains (e.g., the right boundary of one corresponds to the left boundary of the other). However, since the sampling points always fall on the precise location of the pixel mesh, their sampling rules and patterns are completely equivalent. This orientation-independent and boundary-alignable characteristic provides a fundamental guarantee for achieving displacement field continuity across meshes.

[0063] Floating-point precision error: Texture coordinates are converted from normalized floating-point numbers to integers. Computers can accurately represent integers, thus avoiding the errors that occur when representing decimals with floating-point numbers. Instead of relying on the built-in bilinear interpolation based on floating-point coordinates in the graphics hardware, sampling and interpolation based on integer pixel indices are manually implemented in the program. This strategy ensures that the pixel indices of adjacent surfaces on shared boundaries are completely consistent. During interpolation calculations, they strictly reference the same set of pixels (Texels), ensuring that the displacement value calculation process for every point on the boundary is exactly the same, and the results are also strictly consistent. This fundamentally eliminates gaps caused by numerical uncertainty, resulting in a seamless connection of the displacement field at the boundaries.

[0064] In summary, by combining the strategies of square parameter domains and integer texture coordinates, it is ensured that any adjacent base meshes can obtain completely consistent displacement values ​​on the shared boundary. Based on this, as long as the same subdivision factor is used for all meshes in the subsequent tessellation stage, a globally continuous, smooth, and completely watertight final surface can be generated.

[0065] By employing a dual design of "square texture domain alignment" and "integer texture coordinates," the sampling rules of adjacent mesh patches on the shared boundary are consistent, and the pixel indices are perfectly matched, ultimately obtaining consistent displacement values. This eliminates cracks at the source and ensures the watertightness of the curved surface.

[0066] The adaptive subdivision factor calculation stage maps screen-space pixel errors to world-space geometric errors. Combining this with the unique identifier encoding of each facet, an initial subdivision factor is predicted using a three-layer fully connected neural network. The final subdivision factor is obtained by averaging the subdivision factors of shared edges among adjacent faces and rounding up. The core objective is to accurately predict the optimal subdivision factor based on screen-space geometric errors and facet features, balancing visual fidelity and computational efficiency.

[0067]

[0068] The lower bound of this adaptive subdivision factor is determined by geometric error constraints. To ensure that the projection deviation of the subdivided mesh on the screen does not exceed a preset pixel threshold p, the pixel error in screen space needs to be mapped to a distance metric in the world coordinate system. Let the number of pixels in the vertical direction of the screen be... The camera's vertical field of view is FOV, and the depth of the nearest point of the current patch's limiting mesh is z. To ensure that the geometric deviation of the subdivided mesh projected onto the screen does not exceed one pixel, the world space length corresponding to one pixel at depth z can be denoted as... The calculation formula is as follows:

[0069]

[0070] Considering the worst-case diagonal deviation, the maximum geometric error in world space (i.e., the radius of the bounding sphere) corresponding to a p-pixel error can be conservatively estimated as:

[0071]

[0072] Set the pixel threshold p to 1 pixel, therefore This represents the maximum permissible geometric deviation in world space corresponding to one pixel at depth z. Any subdivision level below this error tolerance will result in geometric distortion.

[0073] Based on the aforementioned error constraints, and combined with the geometric and parametric information contained in the unique identifier (ID) assigned to each facet, a lightweight neural network model is constructed as features. These features are input into a three-layer fully connected network (MLP), with the structure: input layer > hidden layer > output layer, to obtain the subdivision factor t. During the subdivision process, the same subdivision factor is uniformly assigned to the boundary edges and internal edges of each facet. When adjacent faces specify inconsistent subdivision factors for shared edges, to ensure mesh continuity and crack-free rendering, the subdivision factor for the shared edge is taken as the arithmetic mean of the subdivision factors of the two facets, rounded up to the nearest integer, as the final subdivision level applied to the shared edge.

[0074] This strategy effectively avoids geometric gaps caused by inconsistent subdivision factors, while ensuring a good balance between visual fidelity and computational cost. It predicts the optimal subdivision factor that meets visual fidelity requirements, thus achieving adaptive control of the mesh subdivision level while satisfying pixel-level geometric accuracy requirements.

[0075] Figure 3 To calculate the conversion of screen space error pixels to world space geometric error, ensure that the projection deviation of the subdivided grid on the screen does not exceed a preset one pixel; Figure 4 The watertight subdivision yields a square parameter domain mapping effect diagram; Figure 5 The internal subdivision factor is obtained by using the unique identifier (ID) assigned to the input patch by the three-segment neural network and the maximum pixel geometric error in the screen space. Then, the subdivision factor of the shared edge of the adjacent patches is averaged and rounded up to obtain the edge subdivision factor.

[0076] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A high-efficiency, high-fidelity rendering method for large-scale CAE simulation data, characterized in that: This includes the watertight texture reconstruction stage and the adaptive subdivision factor calculation stage: The watertight texture reconstruction stage simplifies the original high-poly mesh into a low-poly mesh, assigns an independent axis-aligned square texture domain to each quadrilateral face, converts texture coordinates into integer pixel coordinates, and generates displacement maps and normal maps through linear interpolation sampling and baking to ensure that the displacement values ​​of adjacent face boundaries are consistent. The adaptive subdivision factor calculation stage maps screen space pixel error to world space geometric error. Combined with the features of the unique identifier encoding of the face, the internal subdivision factor is predicted by a three-layer fully connected neural network. The subdivision factor of the shared edge of adjacent face is averaged and rounded up to obtain the edge subdivision factor. Based on the displacement map, normal map, and final subdivision factor, the discrete mesh model is rendered to achieve watertight and efficient subdivision rendering of large-scale discrete models.

2. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 1, characterized in that: The watertight texture reconstruction stage specifically includes: Preprocess the high-poly mesh to reduce the number of faces, repair non-manifold structures, generate quadrilateral facets and assign independent square texture domains; Convert texture coordinates to integers and manually implement sampling and interpolation based on integer pixel indices; Baking sampling is performed to calculate the grid position and normal of the corresponding pixel, and displacement map and normal map are generated by intersecting rays.

3. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 1, characterized in that: The adaptive subdivision factor calculation stage specifically includes: Calculate the maximum geometric error in world space corresponding to the pixel error in screen space; Extract the feature vectors of the facets and input them into a three-layer fully connected neural network model to obtain the initial subdivision factor; Determine the final subdivision factor for shared edges and internal edges.

4. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 2, characterized in that: In the watertight texture reconstruction stage, the high-poly mesh is preprocessed, including simplifying the number of mesh faces, repairing non-manifold structures, intelligently pairing triangular faces into quadrilaterals, discarding the original UVs, and assigning non-overlapping rectangular tiles to each quadrilateral. The boundary of each quadrilateral face is strictly aligned with the four sides of the square texture domain. The shared 3D boundary of adjacent faces corresponds to different sides in their respective texture domains, and the sampling rules are completely equivalent.

5. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 2, characterized in that: In the watertight texture reconstruction stage, during the baking sampling process, the axis-aligned pixel bounding box surrounding the quadrilateral patch is calculated. Each pixel within the bounding box is traversed, and the grid position and normal corresponding to the pixel are calculated through bilinear interpolation. The data is then written into the baking data.

6. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 3, characterized in that: In the adaptive subdivision factor calculation stage, the preset pixel threshold p is 1 pixel, which is used to determine the maximum allowable geometric deviation in world space.

7. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 3, characterized in that, The structure of the three-layer fully connected neural network model consists of an input layer, a hidden layer, and an output layer. The input is a feature vector, and the output is a continuous subdivision factor.

8. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 1, characterized in that: The method for converting texture coordinates to integer pixel coordinates is as follows: ； Where p is the pixel threshold, FOV is the camera's vertical field of view, and H... px z is the number of pixels in the vertical direction of the screen, and z is the spatial depth of the facet.

9. The efficient and high-fidelity rendering method for large-scale CAE simulation data according to claim 7, characterized in that: The input feature vector of the three-layer fully connected neural network includes a unique identifier for the face assignment, the maximum pixel geometric error in the screen space, and the MLP model.

Images