Answers are grounded in this paper and cite their sources.
Designing a neural network for Boundary Representation (B-Rep) models impacts fundamental tasks in Computer-Aided Design (CAD), Engineering (CAE), and Manufacturing (CAM). Existing methods primarily leverage face adjacency for B-Rep feature encoding, overlooking crucial boundary features from the B-Rep vertices. To address this, we introduce a novel dual-stream neural network, called BRepDSN, that explicitly leverages complementary features from both a Face-adjacency Graph (FG) and a Vertex-adjacency Graph (VG). Our method employs alternating message passing between these two graph streams with bidirectional information exchange. Specifically, faces gather boundary-aware cues from incident vertices through a dual mapping, while vertices are refined by face-level context, ensuring boundary-aware and semantically aligned feature encoding. To further boost the performance, we propose a face-vertex contrastive objective. This objective constructs contrastive pairs based on the duality between a face’s embedding and the aggregated embedding of its boundary vertices. Additionally, for segmentation tasks, we use an auxiliary classification loss to determine boundary edges, which acts as a dual task to the main face segmentation. BRepDSN can be used for multiple B-Rep processing tasks. We evaluate the following example tasks: 1) B-Rep Modeling Operation Segmentation, 2) Operation Step Segmentation, 3) B-Rep Classification, 4) Machining Feature Segmentation, and 5) Assembly Mating. Extensive experiments show that our method sets a new state of the art on all five tasks. For example, our method outperforms competing methods on the B-Rep modeling operation segmentation of CC3D (+14.65% mIoU), B-Rep classification of TMCAD (+4.9% oAcc), and the mating task of AutoMate [18] (+23.6% location acc, +21.0% type acc). The code for this paper are available on: https://github.com/shinodashx/BrepDSN.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: