Cross-graph learning based mating interface identification for CAD-assisted assembly design

During CAD model exchange and reuse, explicit assembly metadata may be incomplete even when the geometries of candidate parts remain available. For a predefined pair of B-rep parts, an upstream task is to rank the face pairs most likely to constitute a mating interface before constraint type inference or pose estimation. However, existing methods pay insufficient attention to the interaction between faces across different parts. To address these issues, this paper proposes MFPNet, a lightweight cross-graph learning method for mating interface identification between B-rep part pairs. The proposed method constructs face representations by integrating geometric and shape information, while a topology-aware graph encoder captures intra-part contextual dependencies among B-rep faces. A matcher integrating cross-graph attention and bilinear feature interaction is subsequently employed to rank the candidate face pairs. Experiments on the Fusion 360 Gallery Assembly Dataset demonstrate that the proposed network achieves competitive recognition accuracy with reduced parameter overhead. Moreover, zero-shot tests are performed on the self-built aerospace part mating surface dataset, named Aero-MF. A prototype system further demonstrates how the ranked mating interface candidates can be presented to engineers as upstream decision support for CAD-assisted assembly design.