[Paper Review] AutoMate: A Dataset and Learning Approach for Automatic Mating of CAD Assemblies

B.E. Jones, Dalton Hildreth+4 more|arXiv (Cornell University)|May 25, 2021

3D Shape Modeling and Analysis35 references9 citations

TL;DR

This paper introduces SB-GCN, a heterogeneous graph neural network for learning topological embeddings from BREP-based CAD models, enabling accurate prediction of parametric mates in CAD assemblies. The authors release the first large-scale BREP assembly dataset and demonstrate a 72.2% accuracy in mate suggestion within a commercial CAD workflow using their model.

ABSTRACT

Assembly modeling is a core task of computer aided design (CAD), comprising around one third of the work in a CAD workflow. Optimizing this process therefore represents a huge opportunity in the design of a CAD system, but current research of assembly based modeling is not directly applicable to modern CAD systems because it eschews the dominant data structure of modern CAD: parametric boundary representations (BREPs). CAD assembly modeling defines assemblies as a system of pairwise constraints, called mates, between parts, which are defined relative to BREP topology rather than in world coordinates common to existing work. We propose SB-GCN, a representation learning scheme on BREPs that retains the topological structure of parts, and use these learned representations to predict CAD type mates. To train our system, we compiled the first large scale dataset of BREP CAD assemblies, which we are releasing along with benchmark mate prediction tasks. Finally, we demonstrate the compatibility of our model with an existing commercial CAD system by building a tool that assists users in mate creation by suggesting mate completions, with 72.2% accuracy.

Motivation & Objective

• To address the lack of machine learning models that operate directly on BREP geometry, the native data structure in commercial CAD systems.
• To overcome the absence of large-scale, BREP-based datasets with explicit mate annotations for training assembly prediction models.
• To develop a representation learning framework that captures heterogeneous topological entities (faces, edges, vertices) in BREPs for accurate mate prediction.
• To integrate a learnable mate prediction system into a real commercial CAD environment, improving design efficiency.
• To release a new benchmark dataset and model for future research in BREP-based assembly modeling.

Proposed method

• Propose SB-GCN, a structured graph convolutional network that models BREP geometry as a heterogeneous graph with nodes for topological entities and edges for boundary/component relations.
• Design a message-passing mechanism that propagates embeddings across different entity types (vertex → edge → loop → face) while preserving topological semantics.
• Use mating coordinate frames (MCFs) defined relative to topological entities as input to the model, enabling analytic precision in mate prediction.
• Train the model using a novel dataset of 1,000+ BREP assemblies with annotated mates, compiled from public CAD repositories and cleaned via geometric and topological fingerprinting.
• Implement a user-in-the-loop system that suggests mate completions based on user-selected MCFs, with predictions scored via a classifier on SB-GCN embeddings.
• Integrate the model into a commercial CAD system using the Parasolid kernel, validating compatibility and usability in real design workflows.

Experimental results

Research questions

• RQ1 Can a graph neural network effectively learn heterogeneous topological embeddings from BREP geometry to support mate prediction?
• RQ2 Is it possible to construct a large-scale, real-world dataset of BREP assemblies with explicit mate annotations for training such models?
• RQ3 Can a learned mate prediction system achieve high accuracy while remaining compatible with commercial CAD systems and their parametric modeling workflows?
• RQ4 How does the model’s performance scale with increasing complexity in part geometry and mate types?
• RQ5 What are the key limitations in generalization due to symmetry, part orientation, or geometric ambiguity in mate selection?

Key findings

• SB-GCN successfully learns joint embeddings across heterogeneous topological entities (faces, edges, vertices, loops) in BREP geometry, enabling effective representation learning.
• The authors compiled and released the first large-scale dataset of BREP-based CAD assemblies with annotated mates, containing over 1,000 assemblies and more than 10,000 unique mate pairs.
• The system achieves 72.2% accuracy in suggesting correct mates when integrated into a commercial CAD environment, significantly reducing manual effort in assembly creation.
• The model generalizes across different part geometries and mate types, demonstrating robustness to variations in part orientation and topology.
• The approach is compatible with the Parasolid kernel and can be adapted to other BREP modeling kernels with minimal modification, enabling broad applicability.
• Limitations include restricted MCF search space due to memory constraints and the inability to handle symmetric or non-canonical part orientations without further extension.