Mesh: Discrete Geometric Structures

• Mesh is a discrete geometric representation composed of vertices, edges, and faces that approximate surfaces and volumes across various scientific and graphics applications.
• Meshes enable efficient rendering, finite element analysis, and neural simulation by leveraging explicit connectivity and metric details.
• Recent advances integrate neural surrogates and continuous field representations to improve mesh quality, generalization, and adaptability in computational tasks.

to=arxiv_search.search qq的天天中彩票 天天爱彩票怎么json {"query":"(Kunz et al., 6 Jun 2026) Mesh Graph Neural Network Framework for Accelerating Finite Element Simulation for Arbitrary Geometries", "max_results": 5} to=arxiv_search.search 大发快三的 彩票招商json {"query":"Mesh Graph Neural Network Framework for Accelerating Finite Element Simulation for Arbitrary Geometries (Kunz et al., 6 Jun 2026) MeshGraphNet structural mechanics", "max_results": 10} to=arxiv_search.search ingerlaner 大发快三有json {"query":"MeshNet (Feng et al., 2018) DG-Mesh (Liu et al., 2024) SpaceMesh (Shen et al., 2024) Mesh BDF (Song et al., 30 Jun 2026) MeshFT (Noguchi et al., 1 May 2026)", "max_results": 10} A mesh is a discrete geometric representation of a continuous domain. In graphics and vision, a mesh is a set of vertices, edges, and faces, usually triangles, approximating a surface; in numerical modeling, meshes also include volume meshes, in practice tetrahedral meshes, representing a volumetric domain. Across computer graphics, computational mechanics, biophysics, robotics, and geometric deep learning, meshes function as a memory-efficient representation that supports efficient rendering, geometry processing, texture editing, animation, and the discretization and solution of partial differential equations. Several recent systems make the mesh the central object rather than a by-product: the same finite element mesh can be used both by the numerical solver and by a neural surrogate, and the same triangle mesh can serve simultaneously as a representation for reconstruction, mapping, and generative modeling (Lee et al., 2019, Kunz et al., 6 Jun 2026, Liu et al., 2024).

1. Geometric and algebraic structure

A surface mesh is a collection of 2D simplices embedded in 3D, and a volume mesh is a 3D simplicial mesh. In the triangle-mesh setting, the fundamental entities are vertices, edges, and faces; in tetrahedral settings, tetrahedra are added as volumetric cells. Mesh-based learning papers often choose one element type as the computational unit. MeshNet treats faces as the fundamental processing unit, defines connectivity by shared edges, and notes that each triangular face has degree at most three in the face adjacency graph, making the neighborhood structure simple and bounded (Feng et al., 2018). Mesh Field Theory instead formulates meshes as oriented cell complexes with cochain spaces and signed incidence operators $D_k$, explicitly separating the topological structure encoded by incidence matrices from metric structure encoded by positive-definite constitutive operators (Noguchi et al., 1 May 2026).

This distinction between topology and metric is foundational. Topology specifies which cells are adjacent and with what orientation; metric specifies lengths, areas, volumes, angles, and material-dependent operators. In MeshFT, the conservative interconnection is fixed uniquely by mesh topology, whereas metric effects enter only through constitutive relations and dissipation (Noguchi et al., 1 May 2026). SpaceMesh makes a related point from the perspective of geometry processing and generation: local cyclic neighbor relationships in a halfedge mesh representation are sufficient to guarantee edge-manifoldness and to represent general polygonal meshes of complex connectivity (Shen et al., 2024).

A recurring misconception is that a mesh is merely a visualization artifact. In biophysical modeling, this is explicitly rejected: a mesh is a mathematically critical discretization whose quality directly impacts the numerical solution of PDEs, the conditioning of the resulting algebraic system, and the robustness of iterative solvers (Lee et al., 2019). This suggests that the same data structure is simultaneously geometric, topological, and numerical.

2. Meshes in numerical simulation and adaptation

Finite element analysis solves PDEs by discretizing the geometry into a mesh of elements, assembling a large sparse linear system from element stiffness matrices, and solving for fields such as displacement and stress. In the structural-mechanics setting of plates with arbitrary holes, the mesh is generated with triangular elements, refined near holes, and used to compute displacements and von Mises stress; each simulation can take minutes, and design workflows may require hundreds or thousands of runs (Kunz et al., 6 Jun 2026). In structural biology, GAMer 2 converts electron tomography, contour stacks, and PDB/PQR data into surface and tetrahedral meshes suitable for FEM, finite volume, finite difference, MCell, FFEA, and FEniCS workflows (Lee et al., 2019).

Mesh quality is not secondary. Poor-quality meshes may be acceptable for visualization yet can destroy the quality of the discretization and produce algebraic systems that are badly conditioned and difficult to solve efficiently or accurately with iterative solvers (Lee et al., 2019). The moving mesh PDE literature makes this precise: if the meshing functional satisfies a coercivity condition, then the mesh of the semi-discrete MMPDE is nonsingular for all time if it is nonsingular initially, and the altitudes and volumes of its elements are bounded below by positive numbers depending only on the number of elements, the metric tensor, and the initial mesh (Huang et al., 2015).

Learning-based mesh adaptation replaces repeated solves of auxiliary equations or hand-crafted error heuristics with learned prediction of mesh density or motion. MeshingNet learns a mapping $F:(\Gamma,B,M,X)\to A(X)$, where $A(X)$ is the desired local element area upper bound, and then delegates actual mesh generation to Triangle (Zhang et al., 2020). M2N learns end-to-end mesh movement for PDE solvers, keeps connectivity fixed, and uses either neural splines or graph attention to move nodes while alleviating mesh tangling and maintaining boundary consistency (Song et al., 2022). AMBER predicts a sizing field, defined as a function mapping from the geometry to the local element size of the target mesh, and iteratively produces a new intermediate mesh using an out-of-the-box mesh generator; it uses a hierarchical graph neural network and automatically projects expert labels onto AMBER-generated data during training (Freymuth et al., 29 May 2025). Across these systems, the mesh is the object being optimized, not merely the container for a later solution.

3. Meshes as graphs for machine learning surrogates

When the mesh is treated as a graph, nodes correspond to mesh vertices and edges correspond to mesh adjacency. In the mesh graph network for finite element surrogacy, the finite element mesh is turned into a graph $G=(V,E,\mathbf{u})$: each FE mesh vertex becomes a graph node, an undirected edge is created when two mesh nodes share an element edge, and the global attribute encodes problem-level data such as applied load magnitude (Kunz et al., 6 Jun 2026). Node types encode fixed, applied load, hole, free, and interior roles; edge features are relative geometric features $[\Delta x_{ij},\Delta y_{ij},\ell_{ij}]$; and the resulting architecture is translation-invariant and robust to rotations because it does not use absolute coordinates as node features (Kunz et al., 6 Jun 2026).

This mesh-based representation is the stated reason the model generalizes across arbitrary geometries. The MGN was trained on 11 plate geometries under 20 load conditions and evaluated on 7 unseen geometries and 3 unseen loads. In the most favorable case, it achieves $R^2 \ge 0.97$ on an unseen geometry and unseen load, compared to $R^2 \approx 0.01$–$0.86$ for Random Forest, Gradient Boosting, and K-Nearest Neighbors trained on identical data (Kunz et al., 6 Jun 2026). The paper interprets this as learning local rules on the mesh—how stress depends on local geometry and node roles—and propagating information by message passing, rather than memorizing absolute coordinate-to-stress mappings.

Other mesh-learning architectures choose different units and different inductive biases. MeshNet is face-centric, splits per-face features into spatial and structural components, and uses face rotate convolution for inner triangle shape and face kernel correlation for the distribution of normals around a face (Feng et al., 2018). Mesh-MLP removes mesh convolutions altogether: it takes mesh vertices equipped with the heat kernel signature and dihedral angles as input, replaces convolution with MLP blocks, and uses Layer Normalization; the argument is that for classification and semantic part segmentation, large-scale intrinsic descriptors such as HKS already encode much of the relevant geometry (Dong et al., 2023). These designs differ, but all of them treat the mesh as an information-rich structure rather than reducing it to an unstructured point set.

4. Reconstruction, tracking, and mapping on meshes

In 3D reconstruction, meshes are often the desired end state because modern 3D engines and graphics pipelines require mesh as a memory-efficient representation and because meshes enable geometry processing, texture editing, and animation. DG-Mesh reconstructs a high-fidelity and time-consistent mesh from dynamic monocular observations by building on deformable 3D Gaussians, recovering meshes from Gaussian points, and tracking mesh vertices over time (Liu et al., 2024). Its Gaussian-Mesh Anchoring encourages evenly distributed Gaussians, enforces a near one-to-one relationship between mesh faces and Gaussians, and supports cycle-consistent deformation between canonical and deformed space. The mesh is not treated as a by-product: Gaussians and meshes co-evolve and remain consistent (Liu et al., 2024).

Single-image object reconstruction takes a different route. Mesh R-CNN augments Mask R-CNN with a 3D shape branch that predicts a coarse voxel representation, converts it to a triangle mesh via cubify, and refines that mesh with a graph convolution network operating over mesh vertices and edges (Gkioxari et al., 2019). The voxel stage provides arbitrary topology, while the mesh refinement stage improves geometric fidelity without changing connectivity. This design was motivated by a limitation of earlier template-deformation methods: meshes that start from an ellipsoid or sphere have fixed topology and cannot represent holes or disconnected components (Gkioxari et al., 2019).

In robotics, Mesh-LOAM makes the mesh itself the reference map. It incrementally reconstructs a triangular mesh from LiDAR scans via an implicit signed distance representation and marching cubes, then performs point-to-mesh odometry against the current global mesh (Zhu et al., 2023). On KITTI, the point-to-mesh variant achieves average translation error $0.51\%$ and rotation error $0.14$ deg/100m, outperforming earlier mesh-based baselines and improving over the paper’s own point-to-plane variant (Zhu et al., 2023). Here again, the mesh is not only for rendering or storage; it is the active object of localization.

5. Mesh generation and native mesh modeling

Direct mesh generation has historically been difficult because mesh topology is discrete and variable-length, while many generative models operate on fixed-size continuous tensors. MeshDiffusion addresses this by representing meshes with deformable tetrahedral grids, storing at each tetrahedral vertex a deformation vector and an SDF value, and training a diffusion model on this direct parametrization (Liu et al., 2023). The surface is then the zero-level set of the piecewise-linear SDF field, extracted by marching tetrahedra. The paper argues that meshes are more desirable in practice than voxels or point clouds because they enable arbitrary manipulation for relighting and simulation and fully leverage graphics pipelines (Liu et al., 2023).

Two later works attempt to represent mesh connectivity itself as a continuous object. SpaceMesh defines a continuous latent connectivity space at each mesh vertex; adjacency embeddings determine edges, and permutation embeddings generate cyclic neighbor relationships in a halfedge mesh representation, giving a guarantee of edge-manifoldness and the ability to represent general polygonal meshes (Shen et al., 2024). Mesh BDF introduces the Barycentric Dominance Field, a continuous scalar field on triangular mesh surfaces defined by $F:(\Gamma,B,M,X)\to A(X)$0 in barycentric coordinates, so that vertices become local maxima and topology can be reconstructed from the field by non-maximum suppression and flood fill (Song et al., 30 Jun 2026). Because BDF is an intrinsic mesh property and behaves like a texture map, it integrates into existing 3D diffusion pipelines without architectural modification (Song et al., 30 Jun 2026).

Autoregressive mesh generation remains important, but it exhibits a stated quality-speed trade-off because complex meshes require long token sequences. Mesh RAG addresses this at inference time rather than by changing the base generator: it segments an input point cloud into parts, generates each part with an existing autoregressive mesh model, retrieves a spatial transformation for each part by axis-aligned bounding box alignment and point-to-plane ICP, and assembles the final mesh (Sun et al., 20 Nov 2025). On averaged benchmarks across ShapeNet, Objaverse, Toys4K, and Thingi10K, Mesh RAG improves Chamfer distance, Hausdorff distance, normal consistency, and F-score over several autoregressive baselines (Sun et al., 20 Nov 2025). This suggests that compositionality and retrieval are becoming as central to mesh generation as the topology representation itself.

6. Quality, invariance, and persistent limitations

Mesh research repeatedly returns to three criteria: geometric fidelity, topological validity, and computational usability. For FEM and PDE solvers, this appears as element quality, conditioning, and nonsingularity (Huang et al., 2015, Lee et al., 2019). For learned surrogates, it appears as invariance, generalization across unseen geometries, and robustness to different triangulations (Kunz et al., 6 Jun 2026, Feng et al., 2018). For reconstruction and generation, it appears as manifoldness, temporal consistency, and the absence of self-intersections or degenerate faces (Liu et al., 2024, Shen et al., 2024).

Current systems still have clear limitations. The structural MGN for FEA is restricted to 2D plane-stress problems, linear elastic isotropic steel, and a target mesh size of 25 mm; it does not generalize across significantly different mesh resolutions without retraining (Kunz et al., 6 Jun 2026). DG-Mesh requires foreground segmentation and accurate tracking, and large or abrupt topology changes remain fundamentally hard (Liu et al., 2024). SpaceMesh guarantees edge-manifoldness in connectivity, but geometric self-intersections are not prevented (Shen et al., 2024). Mesh BDF is limited by the resolution of the Dual Contouring grid: very small faces that do not intersect any grid edge are not captured (Song et al., 30 Jun 2026). Mesh-LOAM assumes mostly static scenes and consumes substantial GPU memory (Zhu et al., 2023).

A broader implication is that “mesh” is no longer a single disciplinary object. In one setting it is the discretization on which PDEs are solved; in another it is the graph over which message passing occurs; in another it is the surface on which topology is encoded as a continuous field; in another it is the map against which a robot localizes. What unifies these uses is the combination of explicit connectivity, geometric detail, and compatibility with downstream algorithms. The recent literature therefore treats the mesh not as an inconvenient irregular structure to be bypassed, but as the native representation in which locality, topology, metric structure, and physical or visual constraints can be expressed directly (Noguchi et al., 1 May 2026, Shen et al., 2024, Kunz et al., 6 Jun 2026).