Projection Based Semi-Implicit Partitioned Reduced Basis Method for Fluid-Structure Interaction Problems

Monica Nonino, Francesco Ballarin, Gianluigi Rozza, Yvon Maday

TL;DR

This paper introduces a novel Reduced Order Model (ROM) for simulating Fluid-Structure Interaction (FSI) problems. By employing a partitioned, semi-implicit approach combined with Proper Orthogonal Decomposition (POD), the method significantly accelerates simulations of complex FSI scenarios, including those with geometric and physical parameter variations, while maintaining good accuracy.

What Problem Does This Paper Address?

Simulating Fluid-Structure Interaction (FSI) problems, where fluids and solid structures interact dynamically (e.g., blood flow in arteries, airflow over aircraft wings), is computationally intensive. Traditional methods often rely on either monolithic approaches (solving fluid and solid equations simultaneously) or partitioned approaches (solving them separately and coupling the results). Monolithic methods offer stability but can be inflexible and computationally expensive. Partitioned methods are more flexible, leveraging existing solvers, but can suffer from stability issues, particularly the "added mass effect," and require careful treatment of coupling conditions. This paper aims to develop an efficient and stable ROM for FSI problems that can handle unsteady dynamics and parameter variations, making complex simulations more accessible.

Methods

The core of this work is a partitioned, semi-implicit Reduced Order Model (ROM). Here's a breakdown of the key techniques:

  • Partitioned Approach: The fluid and solid problems are solved separately, allowing for the use of specialized solvers and easier integration of existing codes.
  • Semi-Implicit Coupling: To balance stability and computational cost, the coupling conditions between the fluid and solid are treated in a semi-implicit manner. This avoids the prohibitive cost of fully implicit coupling while improving stability over explicit methods.
  • Chorin-Temam Projection Scheme: This standard method is used for solving the incompressible Navier-Stokes equations for the fluid.
  • Arbitrary Lagrangian-Eulerian (ALE) Formulation: This formalism is employed to handle the moving fluid domain, which deforms with the solid structure.
  • Proper Orthogonal Decomposition (POD): This technique is used to generate reduced-order basis functions from snapshots of the high-fidelity solution. This drastically reduces the dimensionality of the problem for online simulations.
  • Change of Variable for Fluid Velocity: A clever change of variable is introduced for the fluid velocity to transform a non-homogeneous boundary condition at the FSI interface into a homogeneous one, simplifying the ROM formulation and avoiding Lagrange multipliers.
  • Harmonic Extension for Fluid Displacement: Instead of a direct POD on fluid displacement snapshots, an harmonic extension of the reduced solid displacement modes is used. This ensures the continuity of displacements at the interface and simplifies the online computation.
  • POD-Greedy Strategy (for parametrized cases): For problems with parameters, a multi-stage POD approach is used to efficiently generate basis functions from a large number of snapshots across parameter space.

Datasets

The paper does not introduce new benchmark datasets. Instead, it uses standard FSI test cases, including:

  • An unsteady, non-parametrized FSI problem.
  • A problem with geometric parametrization (leaflets' length).
  • A problem with both geometric and physical parametrization (leaflets' length and shear modulus).

The "datasets" in this context are the high-fidelity simulation results (snapshots) generated from these test cases, which are then used to train the ROMs.

Key Insights and Ideas

  • Balancing Act: The semi-implicit coupling is crucial for achieving a good trade-off between simulation stability and computational cost, a common challenge in FSI.
  • Simplifying the Interface: The change of variable for fluid velocity and the harmonic extension for fluid displacement are elegant ways to handle interface conditions, reducing the complexity of the ROM and improving efficiency.
  • Parameterization Power: The method effectively extends to parametrized problems, demonstrating that ROMs can capture the behavior of systems with varying geometric or material properties.
  • POD-Greedy for High Dimensions: The multi-stage POD approach is essential for handling the high dimensionality introduced by parameter spaces, making ROM generation feasible where a naive POD would fail.
  • Pressure Stability: The choice of a Chorin-Temam scheme with a pressure Poisson formulation, even without supremizer enrichment, proves effective for stable pressure approximation in the online phase.

Key Results

The paper demonstrates the effectiveness of the proposed ROM through several test cases:

  • Accuracy: The ROM solutions show good agreement with the full-order model (FOM) solutions, with relative errors typically in the range of 10⁻¹ to 10⁻³. For instance, in the non-parametric case, relative errors are generally below 10⁻² for velocity and displacement, and around 10⁻⁴ for pressure.
  • Speedup: While not explicitly quantified in terms of a single speedup factor, the ROM significantly reduces computational cost. The paper highlights that the online phase involves solving much smaller systems, and the number of internal iterations for convergence is often lower or comparable to the FOM, especially with basis refinement. For example, CPU time comparisons show substantial reductions (e.g., FOM CPU time of 24763.47 seconds vs. ROM CPU times in the range of 10,000-16,000 seconds for a given tolerance).
  • Parametrization: The method successfully captures the behavior of FSI systems under geometric variations (leaflets' length) and physical property variations (shear modulus). For example, changes in leaflet length lead to predictable changes in displacement, well-reproduced by the ROM.
  • Basis Refinement: Results show that increasing the number of basis functions (modes) generally leads to better accuracy, though diminishing returns are observed, and too many modes can sometimes introduce noise.

Limitations and Future Work

  • Computational Cost of Offline Phase: Generating snapshots for training the ROM, especially for parametrized problems with many parameter samples, remains computationally expensive.
  • Error Accumulation: The paper notes an oscillating behavior in pressure error towards the end of simulations, potentially due to error accumulation. This is linked to the lack of supremizer enrichment for pressure approximation.
  • Parameter Sampling: The accuracy of the ROM is sensitive to the number and distribution of parameter samples used in the offline phase. Refining this sampling can improve accuracy but significantly increases offline cost.
  • Future Work:
    • Investigating the impact of supremizer enrichment on pressure approximation stability.
  • Developing error estimators to guide parameter sampling and basis selection more efficiently.
  • Exploring hyper-reduction techniques (like ECSW) for further computational savings.
  • Extending the method to handle different time-stepping schemes for fluid and solid domains.
  • Considering approaches for efficient online-offline decoupling, such as the Empirical Interpolation Method (EIM).
  • Investigating the use of vibrational modes instead of POD for solid basis functions.

Summary updated

Abstract

We present a partitioned Model Order Reduction method for multiphysics problems, that is based on a semi-implicit treatment of the coupling conditions, and on a projection scheme. The proposed Reduced Order Method is based on the Proper Orthogonal Decomposition and on a Galerkin projection onto the reduced basis spaces; we aim of addressing both time-dependent and time-dependent, parametrized Fluid-Structure Interaction problems, where the fluid is incompressible and the structure is linear, elastic and two dimensional.
BibTeX
@article{nonino2023projection,
  author = {Monica Nonino and Francesco Ballarin and Gianluigi Rozza and Yvon Maday},
  title = {{Projection Based Semi-Implicit Partitioned Reduced Basis Method for Fluid-Structure Interaction Problems}},
  journal = {J. Sci. Comput},
  year = {2023},
  doi = {10.1007/s10915-022-02049-6},
}
Figure