Masked Topology Modeling boosts CAD AI with self-supervised learning
New method learns from face-adjacency graphs to predict edge properties, no labels needed.
Heinrich Jiang and Jennifer Jang have introduced Masked Topology Modeling (MTM), a self-supervised learning method for parametric CAD models that exploits the unique structure of boundary representations (B-Rep). Instead of relying on expensive labeled data, MTM randomly masks a fraction of edges in the face-adjacency graph and trains a lightweight head to predict each edge’s convexity and curve type from the encoder’s post-message-passing face features. This approach combines with MoCo-style momentum-queue contrastive learning over B-Rep-aware augmentations and a BFS-connected face-region masked-reconstruction objective.
The team pretrained their model on the ABC dataset and a new procedurally generated dataset, then evaluated on several downstream benchmarks. Results show significant improvement over prior methods, demonstrating that MTM effectively captures geometric and topological patterns critical for CAD tasks like sketch editing, feature recognition, and shape completion. This work addresses the scarcity of large annotated CAD datasets by enabling data-efficient learning directly from unlabeled parametric designs.
- MTM masks edges in a face-adjacency graph and predicts convexity and curve type, a novel self-supervised task for B-Reps.
- Combined with MoCo contrastive learning and BFS-connected region reconstruction, outperforming prior methods on multiple CAD benchmarks.
- Pretrained on the ABC dataset and a new procedurally generated dataset, reducing reliance on labeled CAD data.
Why It Matters
Unsupervised pretraining for CAD can dramatically cut the cost of building AI tools for design automation and engineering.