TopoGNN: Heterogeneous B-Rep graph learning with NURBS-based geometric embeddings for industrial machining feature recognition

Machining Feature Recognition (MFR) remains challenging on industrial CAD models because complex feature interactions, irregular topology, and free-form geometry are only partially captured by discretized 3D representations and face-only CAD graphs. We propose TopoGNN, a CAD-native framework that operates directly on STEP boundary representations and models CAD parts as heterogeneous face–edge–vertex graphs with orientation-aware incidence, adjacency, and loop-order relations. To encode geometry, the method combines NURBS- and UV-based descriptors with type-specific self-supervised variational autoencoders for faces, edges, and vertices, whose embeddings are processed by a relation-aware heterogeneous graph neural network for per-face prediction. We evaluate the approach on the Fusion 360 Gallery benchmark, the TMCAD classification benchmark, and a proprietary industrial dataset of annotated machining operations. TopoGNN achieves 97.21% accuracy and 0.8757 mIoU on Fusion 360 and 88.50% accuracy on TMCAD. On the industrial dataset, evaluated with multi-seed variance reporting and model-level cross-validation, TopoGNN reaches 83.0±0.4% accuracy and 0.507 ± 0.010 mIoU on the fixed split, outperforming five retrained CAD-native baselines in accuracy under every evaluation protocol. Ablation studies show that the gains arise from two components whose relative importance depends on the data regime: on the large-scale public benchmark, the heterogeneous face–edge topology and the learned geometric embeddings contribute jointly, whereas in the small-data industrial regime the discriminative signal is carried primarily by compact handcrafted descriptors together with self-supervised face geometry. These results indicate that explicit B-Rep topology together with learned parametric geometry is an effective basis for robust CAD-native recognition in both public and industrial settings.     «

Machining Feature Recognition (MFR) remains challenging on industrial CAD models because complex feature interactions, irregular topology, and free-form geometry are only partially captured by discretized 3D representations and face-only CAD graphs. We propose TopoGNN, a CAD-native framework that operates directly on STEP boundary representations and models CAD parts as heterogeneous face–edge–vertex graphs with orientation-aware incidence, adjacency, and loop-order relations. To encode geometry...     »