Model reduction method avoids snapping and enforces boundary conditions exactly
The Snapping Phenomenon and Boundary Enforcement in Model Reduction: Mechanisms and Mitigation
Computational Engineering, Finance, and Science
Summary
Complex computer simulations that depend on parameters can be sped up by creating simpler models, but these models often get stuck on certain training examples, making them less flexible. The authors identify different reasons why this 'snapping' happens and develop a new technique that includes a flexible spatial shift to prevent sticking and accurately apply boundary rules based on parameters. This approach keeps the model’s predictions smooth and accurate across different conditions without extra computational cost in use. They test their method on problems in fluid flow and structural engineering to show it works well.
What this means in practice
- •For aerospace engineers: Speed up aerodynamic simulations with more accurate reduced models that handle boundary conditions correctly and avoid sticking to specific scenarios.
- •For structural analysis teams: Improve parametric structural simulations by eliminating errors from boundary condition enforcement and snapshot latching in reduced models.
Authors
Charbel Farhat, Yihong Zhu, Faisal As'ad, Tommaso Taddei
Abstract
Projection-based model order reduction mitigates the computational cost of high-fidelity parametric simulations but can suffer from the "snapping phenomenon", whereby a projection-based reduced-order model (PROM) rigidly latches onto a training snapshot, impairing parametric interpolation. Standard methods also struggle to enforce parameter-dependent boundary conditions without costly constraints or penalty schemes. This work proposes a tripartite taxonomy of snapping, distinguishing intrinsic, extrinsic, and numerical mechanisms. Focusing on intrinsic and numerical snapping, we introduce a unified framework based on a parameter-dependent shift vector. By embedding spatial transport into the state ansatz, the method enforces boundary conditions exactly, restores continuous parametric sensitivity, and eliminates snapshot latching in affine and regression-augmented PROMs without increasing online complexity. We establish the framework's mathematical foundations in the linear regime, showing how it overcomes test-space orthogonality blind spots. In nonlinear computational fluid dynamics, we identify the roles of boundary conditions, mesh grading, and discrete equation scaling in numerical snapping. We also show that formulation-toggling is an unreliable heuristic that can compromise near-field accuracy. Finally, we develop a field-theoretic realization using a spatially varying shift vector governed by a scaled Laplace equation. The approach is validated on structural and aerodynamic benchmarks, including viscous flow around a NACA airfoil and a three-dimensional NASA Common Research Model.