2026 Liu DualBrep (1)

In the realm of Computer-Aided Design (CAD), Boundary Representation (B-rep) is widely utilized for its analytical precision; however, its complex structure presents challenges for neural network models. This study introduces DualBrep, a novel continuous representation designed to bridge the gap between discrete and continuous domains in B-rep learning. We reformulate traditional B-rep data by encoding geometry via a Signed Distance Function (SDF) and topology through a Generalized Voronoi Diagram (GVD) field, consolidating both into a shared latent space. This integrated approach facilitates extensive optimization, enabling joint sampling of geometry and topology, while supporting advanced features such as deterministic reverse engineering and flexible segmentation using a UDF-guided method. We demonstrate that DualBrep overcomes the limitations of conventional methods by allowing models to operate in a fully structured Euclidean domain, effectively addressing the combinatorial complexity of CAD models. Extensive qualitative comparisons reveal that DualBrep not only enables efficient reconstruction of watertight models but also enhances the overall capabilities of generative design processes. The findings suggest a significant advancement in efficient B-rep modeling and its applications in computational design, paving the way for future explorations in the field.