MASKED TOPOLOGY MODELING FOR SELF-SUPERVISED LEARNING ON PARAMETRIC CAD

How can we improve the data efficiency of AI systems for Computer-Aided Design (CAD) given the scarcity of labeled, proprietary B-rep data? The authors investigate whether the unique topological structure of B-rep models—specifically the face-adjacency graph—can be leveraged as a self-supervised signal to learn better geometric representations without manual labels.

The authors introduce Masked Topology Modeling (MTM), a novel self-supervised task. In this framework, the B-rep model is treated as a graph where faces are nodes and shared edges are graph edges. During pretraining, a fraction of these edges are masked (hidden from the encoder). A lightweight head is then trained to reconstruct the convexity (e.g., concave, convex) and curve type (e.g., line, circle, B-spline) of these edges using only the features of the adjacent faces. This objective is combined with MoCo-style contrastive learning and a region-based masked reconstruction task, using both the ABC dataset and a new procedurally generated synthetic dataset.

MTM forces the encoder to learn geometric relationships that standard face-level objectives ignore. Theoretical analysis demonstrates that MTM is strictly more discriminative than face-level methods, as it can distinguish between solids that share identical face geometries but differ in their crease topology. Empirically, the authors show that their method achieves state-of-the-art or competitive performance across several downstream CAD benchmarks, particularly in low-label settings.

This work provides a way to extract high-quality training signals from the intrinsic structure of CAD models without requiring human annotation. By effectively utilizing the "free" topological information inherent in B-rep files, the method addresses the critical bottleneck of data scarcity in industrial engineering and manufacturing AI applications.