Graph Neural Networks for Mesh and Boundary Representations

299 concepts212 papers1 proposal

Traditional deep learning architectures excel at processing grid-like, Euclidean data such as 2D images. However, when extended to 3D geometric data, standard representations introduce severe trade-offs. Volumetric voxels suffer from cubic scaling costs ($O(N^3)$), which limits resolution and results in a blocky appearance. Unstructured point clouds avoid this spatial inflation but discard explicit connectivity, leaving the network to guess how points relate across a continuous surface.

To resolve these limitations, a prominent research direction treats 3D shapes—such as polygonal meshes and Computer-Aided Design (CAD) Boundary Representations (B-Reps)—as non-Euclidean graphs. By mapping geometric entities to nodes and their structural connectivity to edges, graph neural networks (GNNs) can perform message passing directly along the surface manifold. This approach builds a powerful inductive bias into the network, preserving topological integrity and enforcing geometric constraints without wasting computation on empty space.

Mapping Geometry to Graph Topologies

To process a 3D surface as a discrete graph, researchers structure the underlying geometry using two primary paradigms: primal graphs and dual (or face-adjacency) graphs.

In a primal graph representation, the physical vertices of a 3D mesh serve as graph nodes, and the mesh edges define the connectivity. As highlighted by Saleh et al., this approach preserves the original topological integrity of an object during complex deformation tasks. By maintaining explicit vertex-to-vertex connections, the model avoids the information loss and surface disruption that occurs when subsampling continuous geometry into unstructured point clouds.

Primal Graph (Mesh Topology):
   Vertex (Node) ---> Edge (Graph Edge) ---> Vertex (Node)

Dual Graph (Face Adjacency):
   Surface Face (Node) ---> Shared Boundary (Graph Edge) ---> Surface Face (Node)

Conversely, for solid modeling and CAD applications, researchers frequently employ dual graphs or face-adjacency graphs. In this setup, individual geometric faces are mapped to nodes, while the shared boundaries or curves form the graph edges. For example, Jayaraman et al. introduced this structure in UV-Net to integrate local geometric surface embeddings into a global connectivity graph. This allows message-passing networks to learn how distinct geometric features relate topologically across a model.

Shifting the inductive bias from raw spatial coordinates to explicit connectivity yields significant advantages. Prezzavento utilized a dual graph structure to adapt Transformer architectures to the irregular, non-sequential world of 3D meshes, demonstrating that prioritizing the "graph-ness" of a mesh provides a more robust foundation for semantic segmentation than relying on spatial coordinates alone.

Convolutions on Non-Euclidean Manifolds

Once geometry is structured as a graph, standard convolutional operators must be adapted to handle irregular neighborhoods. The literature splits these operations into spatial and spectral domains.

Spatial Message Passing and Edge Convolutions

Spatial methods operate directly on local neighborhoods by aggregating features from adjacent nodes or edges. Edge-based convolutions on triangular meshes extract local features like curvature and surface normals by processing explicit connectivity data. As shown by Cameron et al., leveraging this explicit connectivity allows neural operators to satisfy boundary conditions more accurately in physics-informed learning. Because the network understands exactly how points are linked, it can compute continuous surface gradients that point-cloud methods struggle to resolve.

Spectral Convolutions and Graph Laplacians

Spectral methods approach the problem by transforming graph signals into the frequency domain using the graph Laplacian. Wang et al. demonstrated that incorporating spectral convolutions allows neural networks to learn the underlying geometric structure of an organ rather than just memorizing absolute spatial coordinates. This mechanism improves the model's ability to generalize across anatomical variations and deformations. This mathematical formulation enables the extraction of hierarchical features from highly irregular geometric structures, a property utilized by Jiang et al. to successfully separate identity from expression in 3D face shapes.

Why It Works: Inductive Biases and Topological Supervision

The core reason graph-based mesh representations outperform unstructured alternatives lies in their strong topological inductive bias. In a point cloud, a model must infer which points belong to the same surface. In a graph-based representation, surface manifoldness is explicitly baked into the data structure and remains invariant to rigid transformations like rotation and translation.

This explicit structure enables models to enforce geometric logic directly during training. For instance, Cherenkova demonstrated that integrating topological graph learning as a direct supervision signal forces predicted surfaces and curves to meet precisely at shared boundaries. This structural constraint prevents the formation of disjointed or "leaky" 3D reconstructions that frequently plague unconstrained coordinate-based networks.

Representation Memory Scaling Surface Continuity Topological Awareness
Volumetric Voxels Cubic ($O(N^3)$) Discontinuous (Blocky) Implicit
Point Clouds Linear ($O(N)$) None (Isolated Points) None
Graph-Based Meshes Sparse/Linear ($O(V+E)$) Continuous Explicit

Furthermore, graph-based representations are highly compact. Pontes et al. showed that storing relationships between discrete models via graphs provides a memory-efficient alternative to volumetric grids. Systems can achieve high geometric fidelity without the cubic scaling costs that typically restrict the resolution of spatial grids.

Applications Across Domains

By bridging the gap between deep learning and non-Euclidean geometry, this paradigm has unlocked breakthroughs across several distinct fields:

  • Computer-Aided Design (CAD): In engineering and manufacturing, Willis et al. leveraged face-adjacency graphs to build the Fusion 360 Gallery, enabling machine learning agents to parse programmatic CAD construction sequences by aggregating local geometric information across a model's topology.
  • Biomedical Shape Analysis: In medical imaging, Besson et al. applied graph convolutional neural networks to 3D surface meshes of the human brain. The authors proved that treating the cortex as a continuous, folded non-Euclidean sheet captures complex anatomical biomarkers far more effectively than traditional voxel-based analysis, which often ignores the structural relationships of the tissue ribbon.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graph_based_mesh_representation_geometri_83c42627,
  title = {graph-based mesh representation (geometric deep learning)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-mesh-representation-geometric-deep-learning/rdr_83c426272caf4b5c9e2707bc2f608e6b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_based_mesh_representation_geometri_83c42627_v1,
  title = {graph-based mesh representation (geometric deep learning)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-mesh-representation-geometric-deep-learning/rdr_83c426272caf4b5c9e2707bc2f608e6b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake