In computer vision, 3D shapes are often treated as "fuzzy" collections of points or triangles. But in engineering, objects are defined by Boundary Representations (B-Reps)—precise mathematical descriptions of surfaces (faces) and the curves (edges) where they meet. Recent research has shifted from processing raw spatial coordinates to learning directly on the topological "logic" of these B-Reps by representing them as graphs.
The B-Rep as a Face-Adjacency Graph
The core intuition behind this research direction is that the identity of a mechanical part lies in its connectivity, not just its coordinates. While a point cloud might tell you where a surface is, it doesn't explicitly state that a specific cylinder is "joined" to a specific plane.
To capture this, researchers like Pradeep Kumar Jayaraman and Karl D. D. Willis introduced UV-Net, which models a 3D solid as a Face-Adjacency Graph (FAG). In this structure:
- Nodes represent individual surfaces (like a plane, cylinder, or sphere).
- Edges represent the shared boundaries where these surfaces intersect.
By using Graph Neural Networks (GNNs) on this representation, models can reason about the global structure of an assembly. As shown in the Fusion 360 Gallery, this structural encoding allows networks to understand how individual features relate to one another, making it far more effective than standard meshes for engineering tasks.
Preserving Engineering Intent
Why use a graph instead of a high-resolution mesh? The answer lies in "design intent." A mesh loses the distinction between a single hole and the surrounding plate; it’s just a sea of triangles. In contrast, a B-Rep graph preserves the exact geometric primitives.
This precision is critical for mechanical assembly. In JoinABLe, Karl D. D. Willis and colleagues demonstrated that operating on B-Rep graphs allows a model to predict how parts should fit together with near-human accuracy. Because the graph preserves the "analytic" nature of the surfaces (knowing a hole is a perfect cylinder), the model can calculate precise alignment axes that point clouds simply cannot resolve.
Enforcing Geometric Logic
One of the hardest problems in 3D reconstruction is "leakiness"—when a predicted shape has tiny gaps between surfaces that should be closed. Traditional methods struggle with this because they don't understand the underlying topology.
Researchers have begun using the B-Rep graph as a form of topological supervision. For example, SpelsNet, developed by Kseniya Cherenkova and others, uses the graph structure to ensure that predicted surfaces and curves meet at shared boundaries. By forcing the network to respect the "logic" of the B-Rep, the model prevents disconnected or physically impossible reconstructions. Similarly, EquiCAD by Yonghao Su combines this topological reasoning with geometric equivariance, ensuring the model's understanding of the part remains consistent even if the object is rotated or reflected.
Reverse Engineering the Construction Sequence
Beyond just identifying shapes, this graph-based approach enables "reverse engineering"—turning a static 3D model back into an editable CAD program.
Models like CADParser by Shengdi Zhou treat the B-Rep graph as a starting point to infer a sequence of modeling operations (like extrusions or fillets). Because the graph captures the spatial relationships between features, the model can "read" the history of the part's construction. This direction is furthered by RLCAD, which uses reinforcement learning to "play" with CAD operations in a simulated environment, using the B-Rep graph to provide robust shape embeddings that guide the agent toward a perfect reconstruction.
By moving from "where the points are" to "how the surfaces connect," this research direction provides a bridge between the unstructured world of deep learning and the precise, programmatic world of computer-aided design.