Title: MidSurfNet: Learnable Face Pairing and Interference Implicit Fields for Generalized Mid-surface Abstraction

Abstract:

Generating mid-surface abstractions is a critical step in the finite element analysis of thin-walled CAD models. While traditional methods depend on handcrafted geometric heuristics for face pairing, they often struggle with complex real-world industrial designs. Specifically, rule-based approaches frequently fail when encountering multi-wall-thickness regions, self-matching face configurations, or the need for non-center offset surfaces. To overcome these challenges, we introduce MidSurfNet, a learning-augmented framework featuring two innovative components. First, a neural face pairing module utilizes geometric and topological features to predict face pair confidence, effectively managing complex pairing scenarios that exceed the capabilities of rule-based systems. Second, we propose an interference implicit field that models mid-surfaces through the interference of two signed distance functions, allowing for generalized offset control and flexible positioning within downstream CAE/FEA workflows.

To support this research, we compiled a large-scale dataset comprising over 1,500 manually annotated CAD models. Our experimental results show that MidSurfNet reaches an 87.32% accuracy in face pairing. Notably, it successfully addresses multi-wall-thickness scenarios with a 61.90% completion rate and self-matching situations with a 52.94% completion rate—tasks that have previously confounded all existing methods. Ultimately, MidSurfNet offers a learning-driven solution for generalized mid-surface abstraction, providing arbitrary offset control tailored for CAE-oriented applications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC