Reproduce “Shape: A Self-Supervised 3D Geometry Foundation Model for Industrial CAD Analysis” on OpenResearch

About this paper

Bayangmbe Mounmo, Sam Chien, Mile Mitrovic

Shape learns representations from industrial CAD meshes by predicting masked local geometry tokens and balancing their per-dimension scales. In a controlled 20-epoch comparison, target normalization raises positive-pair retrieval from 87.6% to 97.1% among 2,048 candidates. The released model achieves 0.729 reconstruction R-squared and 98.1% retrieval. These are pretraining results, not demonstrated symmetry detection or simulation savings.

Key insights

• Calibrated per-dimension target normalization is the decisive intervention for stable pretraining on industrial CAD data, which has heterogeneous geometric signals and heavy-tailed curvature distributions.
• The model's pretraining objective inherently provides a per-token attribution map by mapping reconstruction errors directly back to the input mesh surface, offering built-in explainability without post-hoc methods.
• The architecture enables the learning of dense, transferable per-token representations from raw surface meshes, effectively compressing complex 3D structure and generalizing well to unseen CAD parts.

Summary by alphaXiv.