Boundary Representations (B-Reps)

• Boundary Representations (B-Reps) are models that explicitly encode the boundaries of 3D solids using interconnected NURBS-based surfaces, edges, and vertices.
• They are fundamental in CAD and CAM systems, ensuring precise Boolean operations, watertightness, and robust analytic modeling for manufacturing and simulation.
• Recent advances like NeuroNURBS and decoupled diffusion pipelines enhance parametric fidelity, enabling efficient neural approximations and generative modeling.

A boundary representation (B-Rep) defines the geometry and topology of 3D solids by explicitly encoding their boundaries as a collection of connected parametric surfaces, edges, and vertices. B-Reps are the de facto representation for solid modeling in nearly all computer-aided design (CAD), computer-aided manufacturing (CAM), and engineering analysis systems, encompassing both classical industrial applications and contemporary data-driven pipelines. The B-Rep formalism combines the exactness and editability of parametric surfaces—almost universally Non-Uniform Rational B-Splines (NURBS)—with a topologically rigorous data structure that supports both analytic modeling, manufacturing, and machine learning tasks (Fan et al., 2024).

1. Formal Foundations and Topological Structure

A B-Rep solid is structured as a 2-manifold cell complex embedded in ℝ³, where the boundary is decomposed into faces (2D parameterized surface patches), edges (1D parameterized curves), and vertices (0D points) (Lambourne et al., 2021). Each face is bounded by one or more oriented loops of edges. The data structure encodes:

• Faces: Trimmed parametric surfaces, often NURBS patches, with boundaries defined by trimming curves in the (u,v) parameter space.
• Edges: Oriented curve segments (e.g., lines, circles, B-spline curves) connecting vertices, each with a pair of directed coedges ("half-edges") referencing their parent faces.
• Vertices: Points of intersection of incident edges/curves, typically stored as 3D coordinates.

Incidence and adjacency relations between these entities are represented through graphs or explicit combinatorial structures (e.g., half-edge, winged-edge, or Weiler–coedge data structures). The topology ensures watertightness and orientability, with the boundary operator formalism mapping closed cells to their incident lower-dimensional features (Lambourne et al., 2021).

2. Parametric Geometry and NURBS Surfaces

In industrial CAD, B-Rep faces are nearly always represented as NURBS surfaces, which generalize B-splines to allow local refinement, arbitrary continuity, and exact representation of conics and freeform shapes. A NURBS surface

$$S(u,v) = \frac{\sum_{i=0}^n\sum_{j=0}^m N_{i,p}(u) N_{j,q}(v) w_{ij} P_{ij}} {\sum_{i=0}^n\sum_{j=0}^m N_{i,p}(u) N_{j,q}(v) w_{ij}}$$

is defined by a grid of control points $P_{ij} \in \mathbb{R}^3$, associated weights $w_{ij} > 0$, degree parameters $(p,q)$, and non-uniform knot vectors $U$, $V$. The flexibility of the knot vectors enables modeling of smooth surfaces with varying continuity and sharp features. Trimming curves in (u,v)-space restrict the surface to its final boundary (Fan et al., 2024). The parametric nature allows exact analytic operations (boolean, projection, filleting) crucial for robust CAD workflows.

3. Data Structures and Industry Standards

B-Reps underpin major standardized CAD formats such as STEP (ISO 10303-21), Parasolid, and IGES, which use explicit records to represent vertices, edges, faces, and high-level connectivity. STEP encodes NURBS parameters, topology, and trimming in a schema-based ASCII format, though direct programmatic access can be challenging due to undocumented variation and incompatibilities between kernels (Izadyar et al., 4 Jun 2025). Modern systems increasingly migrate to more accessible HDF5-based or open data structures exposing geometry, topology, and mesh representations for large-scale ML consumption (Izadyar et al., 4 Jun 2025).

Typical components of B-Rep data structures: | Entity | Parametric Representation | Topological References | |----------|---------------------------------------------------|-----------------------------------------------| | Vertex | 3D point $(x, y, z)$ | Set of incident edges (start, end) | | Edge | Parametric curve $C(t)$ (e.g., NURBS, circle) | End-point vertices, incident faces/coedges | | Face | NURBS surface $S(u, v)$, trims as curve loops | Bounding edges/coedges (loops), shell membership |

The data structure supports efficient local updates (e.g., pointer-based graphs), enforcement of topological validity via the Euler–Poincaré formula, and exact Boolean and intersection queries (Bank et al., 3 Mar 2026).

4. B-Rep Learning and Neural Approximations

Learning methods for B-Reps must simultaneously model both geometry and topology. Early methods (UV-Net, BRepNet) represent geometry using UV-grids: sampling each face's parameter domain into a regular grid of 3D points and normals, and modeling topology as a face adjacency graph (Jayaraman et al., 2020, Lambourne et al., 2021). However, UV-grid discretization suffers from inefficiency and is lossy for subtle features; upsampling gives higher fidelity at the cost of increased data and compute (Fan et al., 2024).

Recent approaches avoid explicit grid sampling by directly encoding NURBS parameters (control points, weights, knots), achieving substantial efficiency gains and ensuring exact parametric reconstruction without grid-induced oscillations. Notably, NeuroNURBS employs a transformer-based variational autoencoder to jointly encode control nets, weights, and knot vectors, achieving 80–90% reductions in memory/GPU use and significantly faster geometry reconstruction compared to UV-grid pipelines; moreover, it exactly matches the degree and regularity of the original B-Rep surfaces, preventing undulations and over-elevated degrees that arise from re-fitting NURBS to sampled points (Fan et al., 2024).

Representation	Storage per Face	Data Size Reduction (vs. UV)	Geometric Fidelity
UV-Grid (32×32)	3072 floats	–	High at dense grid, lossy
NeuroNURBS	parametric (≪1k)	80–96% (surface/solid/model)	Exact parametric

Hybrid implicit/explicit strategies have also been explored: representing each parametric patch via a learned MLP neural field outputting both the 3D surface and the parametric signed distance to the face boundary, improving few-shot learning and facilitating CAD-native tasks (segmentation, classification) without reliance on point clouds or meshes (Jones et al., 2022).

5. Generative Modeling and Topology-Geometry Coupling

Generative B-Rep models face challenges stemming from the discrete-continuous coupling of geometry (parametric surfaces/curves) and topology (incidence graphs, face loops). Early works factorize the problem autorregressively by generating vertices, then edges, then faces (e.g., SolidGen, BrepGen), using tokenized representations and pointer networks to ensure hierarchical recoverability. Later models such as BrepGPT introduce unified local representations (Voronoi half-patches) that combine geometric and topological features at the half-edge level, compactly encoding all information needed for sequential autoregressive generation (Li et al., 27 Nov 2025).

Recent frameworks employ decoupled diffusion pipelines (DTGBrepGen): first generating face–edge and edge–vertex adjacency relations to ensure exact topological validity, then generating geometry (e.g., B-spline control grids) by topology-aware sequential denoising (Li et al., 17 Mar 2025). Holistic latent representations (HoLa) go further by eliminating explicit edges and vertices in the latent, instead inferring all topological connections via surface-to-surface neural intersection networks; this reframes topology learning as a geometric reconstruction problem and achieves state-of-the-art validity and complex geometry rates (Liu et al., 19 Apr 2025).

6. Practical Applications and Computational Considerations

B-Rep models are fundamental in simulation, machining, and collaborative CAD workflows. In manufacturing simulation, such as virtual machining and cutter–workpiece engagement (CWE), B-Rep exactness enables precise Boolean operations, avoidance of meshing artifacts, and robust engagement profile extraction. FusionCut achieves sub-degree angular error and full reproducibility by exploiting in-kernel B-Rep APIs, demonstrating feasibility for cloud-scale, open, and parallelized simulation (Bank et al., 3 Mar 2026).

B-Rep data also enables efficient ML pipelines for normal estimation, denoising, segmentation, and property prediction directly on analytic surfaces, as demonstrated by the Better STEP HDF5-based dataset and its associated query library (Izadyar et al., 4 Jun 2025). On benchmarks, B-Rep-aware ML models markedly outperform mesh or point-cloud–based approaches in both accuracy and computational efficiency (Jayaraman et al., 2020, Lambourne et al., 2021).

7. Current Trends and Future Directions

Research in B-Rep modeling is rapidly advancing toward:

• Direct parametric representations: Avoidance of intermediate sampling by learning on analytic parameters (e.g., NURBS, B-splines), enabling exact editing and inference.
• Unified geometry–topology latent spaces: Joint encoding of shape and structure, reducing cascading errors and enabling robust interactive editing, local inpainting, and scalability to complex assemblies (Lu et al., 25 Jan 2026, Liu et al., 19 Apr 2025).
• Hybrid and volumetric methods: Encoding entire solids with volumetric distance functions or hybrid neural fields, ensuring watertightness and robustness at the cost of grid resolution (Zhang et al., 18 Nov 2025).
• Rotation invariance and local/global feature fusion: Techniques such as FoV-Net employ local reference frames and field-of-view ray-casting to achieve rotation invariance and improved generalization in learning tasks, outperforming conventional UV-grid-based networks (Ballegeer et al., 27 Feb 2026).
• Standardization and accessibility: Migration to open data exchange and sampling libraries that decouple pipelines from proprietary kernels, facilitating large-scale, reproducible geometric learning (Izadyar et al., 4 Jun 2025).

Boundary representations have thus evolved from static exact modeling structures into a central substrate for differentiable, generative, and data-augmented design, with ongoing research addressing compactness, reliability, and integration with automated and user-guided design workflows (Fan et al., 2024, Xu et al., 2 Dec 2025).