BRep-GD: A Graph Diffusion Model for CAD Boundary Representation Generation.

Area of Science:

• Computer-aided design (CAD)
• Geometric modeling
• Graph-based machine learning

Background:

• Traditional B-rep generation methods use tree hierarchies, limiting efficiency and quality.
• These methods do not leverage the inherent graph structure of B-reps effectively.

Purpose of the Study:

• To propose BRep-GD, a graph diffusion model for improved B-rep generation.
• To address the limitations of existing tree-based B-rep generation techniques.

Main Methods:

• BRep-GD models B-reps as graphs with faces as nodes and boundary/vertex elements as edges.
• Utilizes a continuous topological graph data structure for B-rep generation.
• Employs sequential face and edge generation with continuous topology decoupling to reduce computational complexity.

Main Results:

• BRep-GD demonstrates superior performance in both unconditional and class-conditional B-rep generation tasks.
• Achieves higher quality in generating watertight solids and handling complex geometries.
• Significantly reduces geometric inconsistencies and improves generation efficiency compared to state-of-the-art methods.

Conclusions:

• BRep-GD offers a more efficient and effective approach to B-rep generation by exploiting graph structures.
• The model shows promise for advancing CAD applications requiring high-quality geometric modeling.
• Code, models, and datasets are publicly available for further research and development.