A Reduced Order Cut Finite Element method for geometrically parametrized steady and unsteady Navier-Stokes problems

Efthymios N. Karatzas, Monica Nonino, Francesco Ballarin, Gianluigi Rozza

TL;DR

This paper introduces a novel approach for efficiently simulating fluid flow governed by the Navier-Stokes equations, particularly when the geometry of the flow domain changes. By combining an unfitted mesh Finite Element Method (CutFEM) with a Reduced Basis Method (RBM) powered by Proper Orthogonal Decomposition (POD), the authors achieve significant computational speedups while maintaining accuracy for both steady and unsteady flow problems with varying geometries.

What Problem Does This Paper Address?

Simulating fluid dynamics, especially for complex geometries or scenarios where the geometry itself changes (e.g., due to large deformations, fractures, or moving boundaries), is computationally expensive. Traditional Finite Element Methods (FEM) struggle with such problems because they often require remeshing the computational domain whenever the geometry changes, which is a time-consuming and complex process. While unfitted mesh methods like CutFEM can handle changing geometries without remeshing by using a fixed background mesh, solving the resulting system repeatedly for different geometric parameters can still be computationally prohibitive. This paper addresses this challenge by developing a Reduced Order Model (ROM) that leverages the strengths of CutFEM for geometric flexibility and RBM for computational efficiency.

Methods

The core of this work lies in the integration of two key methodologies:

  • Cut Finite Element Method (CutFEM): This approach uses a fixed background mesh that does not conform to the physical domain's boundaries. The physical domain is defined by a level-set function, and the mesh elements are "cut" by the domain's boundary. This allows for handling complex and changing geometries without the need for remeshing. Stabilization techniques like the Ghost Penalty are employed to manage stability issues arising from the unfitted nature of the mesh.
  • Reduced Basis Method (RBM) with Proper Orthogonal Decomposition (POD): RBM aims to create a computationally inexpensive surrogate model of a complex system. POD is used in the "offline" phase to extract dominant modes (reduced basis functions) from a set of "snapshots" (solutions computed for various geometric parameters). These basis functions then form a low-dimensional space onto which the original high-dimensional problem is projected (Galerkin projection). This results in a much smaller system that can be solved rapidly in the "online" phase. The authors also incorporate a "lifting function" to handle non-homogeneous Dirichlet boundary conditions and a "natural smooth extension" to ensure parameter-independent basis functions, crucial for the POD procedure. For improved pressure approximation stability, a supremizer enrichment technique is also employed.

Datasets

The paper does not introduce new benchmark datasets. Instead, it uses synthetic data generated by solving the Navier-Stokes equations for various geometric configurations defined by a parameter $\theta$. The training set for the POD (Ptrain) and the test set (Ptest) are generated by randomly sampling values of this parameter $\theta$.

Key Insights and Ideas

  • Geometry-Independent Reduced Basis: A significant contribution is the development of a geometry-independent reduced basis. This is achieved through the natural smooth extension of snapshots to the entire background mesh, allowing for the creation of a reduced-order model that works across a range of geometric configurations without needing to re-generate the basis for each new geometry.
  • Synergy of CutFEM and RBM: The paper demonstrates a powerful synergy between CutFEM's ability to handle complex and evolving geometries and RBM's efficiency for parameterized problems. This combination overcomes limitations of both methods when used in isolation for such problems.
  • Supremizer Enrichment for Pressure Stability: The inclusion of a supremizer enrichment technique is highlighted as crucial for achieving stable approximations of the fluid pressure, a known challenge in reduced-order modeling of incompressible flows.
  • Handling of Boundary Conditions: The paper effectively addresses the treatment of non-homogeneous Dirichlet boundary conditions (specifically at the inlet) using a lifting function, and the weak imposition of other boundary conditions via Nitsche's method, which is well-suited for unfitted methods.

Key Results

The paper presents numerical results for both steady and unsteady Navier-Stokes problems with varying geometries.

  • Accuracy: The authors demonstrate that the reduced order models achieve good accuracy, with relative errors for velocity and pressure that decrease as the number of basis functions increases. Figures 7 and 12 show that the relative errors are generally low, especially for a sufficient number of basis functions.
  • Efficiency: While not explicitly quantified with speedup factors, the methodology is designed to provide significant computational gains. The reduced order models, once generated, allow for very fast online simulations compared to solving the full-order problem repeatedly.
  • Geometric Flexibility: The results showcase the ability of the method to accurately capture flow behavior across a range of geometric parameters, as illustrated by the visualizations of velocity and pressure fields for different obstacle positions or wall shapes (Figures 8, 9, 13, 14, 19-24). The approximation error is shown to be concentrated near the cut elements, a phenomenon consistent with previous studies.

Limitations and Future Work

  • Limitations:
    • The authors acknowledge that the approximation error tends to concentrate near the cut between the physical domain and the background mesh. This phenomenon warrants further investigation.
  • The current implementation does not explicitly incorporate snapshot transportation techniques, which could potentially further improve accuracy, especially for time-dependent problems with moving boundaries.
  • Future Work:
    • Extend the procedure to nonlinear, time-dependent problems where the level-set geometry is itself time-dependent (i.e., known a priori).
  • Investigate the more general case where the motion of the levelset geometry is an unknown of the problem.
  • Apply the approach to time-dependent Fluid-Structure Interaction problems, geometrically shallow water flows, and phase flow Navier-Stokes systems.
  • Further develop and apply hyper-reduction techniques for unfitted discretizations to achieve even greater computational efficiency.
  • A more in-depth analysis of the error concentration near cut elements.

Summary updated

Abstract

This work focuses on steady and unsteady Navier-Stokes equations in a reduced order modeling framework. The methodology proposed is based on a Proper Orthogonal Decomposition within a levelset geometry description and the problems of interest are discretized with an unfitted mesh Finite Element Method. We construct and investigate a unified and geometry independent reduced basis which overcomes many barriers and complications of the past, that may occur whenever geometrical morphings are taking place. By employing a geometry independent reduced basis, we are able to avoid remeshing and transformation to reference configurations, and we are able to handle complex geometries. This combination of a fixed background mesh in a fixed extended background geometry with reduced order techniques appears beneficial and advantageous in many industrial and engineering applications, which could not be resolved efficiently in the past.
BibTeX
@article{karatzas2022reduced,
  author = {Efthymios N. Karatzas and Monica Nonino and Francesco Ballarin and Gianluigi Rozza},
  title = {{A Reduced Order Cut Finite Element method for geometrically parametrized steady and unsteady Navier-Stokes problems}},
  journal = {Comput. Math. Appl},
  year = {2022},
  doi = {10.1016/j.camwa.2021.07.016},
}
Figure