Fast and robust solvers for pressure-correction in bubbly flow problems

Scott P. MacLachlan, Jok Man Tang, Cornelis VuikJournal of Computational Physics 2008 · DOI

TL;DR

This paper tackles the computational bottleneck in simulating bubbly flows: solving the pressure-correction equation. The authors compare advanced multilevel solvers, specifically deflation-based methods (DIC-CG) and various multigrid approaches (GMG, CCMG, BoxMG, AMG), demonstrating that structured multigrid methods like BoxMG-CG offer superior speed and robustness for these complex fluid dynamics problems.

What Problem Does This Paper Address?

Simulating two-phase flows, particularly those with bubbles or droplets (bubbly flows), is crucial for many industrial applications. These simulations are computationally demanding, largely due to the need to solve a pressure-correction equation at each time step. This equation often takes the form of a Poisson equation with discontinuous coefficients, arising from significant differences in fluid properties like density. Traditional methods struggle with these discontinuities, leading to slow convergence. The paper aims to find faster and more robust numerical methods to solve this pressure-correction system, enabling more efficient and accurate simulations of bubbly flows.

Methods

The paper investigates several advanced numerical techniques for solving the pressure-correction equation, which is formulated as a symmetric and positive semi-definite linear system:

  • Preconditioned Conjugate Gradient (PCG): A standard iterative method for solving linear systems.
  • Deflated Conjugate Gradient (DIC-CG): An extension of PCG that incorporates deflation techniques. Deflation aims to remove components associated with small eigenvalues that slow down convergence, often by projecting out errors related to known low-frequency modes. The authors use subdomain-based deflation vectors.
  • Multigrid Methods: These methods use a hierarchy of grids to accelerate convergence.
  • Geometric Multigrid (GMG): Uses a fixed coarsening strategy based on grid geometry. The paper explores GMG-CG, where GMG acts as a preconditioner for CG.
  • Galerkin Coarsening (CCMG): A multigrid variant that uses a coarse-grid matrix derived from the fine-grid operator, aiming to better handle discontinuous coefficients.
  • Black Box Multigrid (BoxMG): An operator-dependent multigrid method where interpolation weights are determined by the matrix coefficients, making it more robust for problems with varying coefficients. The paper uses BoxMG-CG, where BoxMG is a preconditioner.
  • Algebraic Multigrid (AMG): A more general multigrid approach that does not rely on a regular grid structure, determining coarsening and interpolation algebraically.

The authors compare these methods both as standalone solvers and as preconditioners for CG.

Datasets

Not discussed in the paper. The paper focuses on numerical experiments using simulated bubbly flow scenarios.

Key Insights and Ideas

  • Discontinuous Coefficients are the Bottleneck: The significant density contrast in bubbly flows creates discontinuous coefficients in the pressure-correction equation, severely degrading the performance of standard solvers and even some advanced multigrid methods.
  • Structured Multigrid Shines: For the specific problem of bubbly flows on structured grids, specialized multigrid methods like BoxMG (and its preconditioned variant BoxMG-CG) demonstrate superior scalability and robustness compared to other methods.
  • Deflation Offers a Good Balance: DIC-CG provides a good balance between performance and ease of implementation, especially when starting with an existing CG solver. It shows good scalability, though generally not as fast as BoxMG-CG.
  • AMG's Robustness Comes at a Cost: While AMG is a powerful and general method, its unstructured nature leads to higher computational costs per iteration, making it less competitive for this specific problem where structured methods are applicable.
  • Preconditioning is Key: For many methods, using them as preconditioners for the Conjugate Gradient (CG) algorithm significantly improves performance and robustness, effectively resolving the remaining error modes.

Key Results

  • BoxMG-CG is the Winner: In most stationary and time-dependent simulations, BoxMG-CG consistently outperforms other methods in terms of both computational time and the number of iterations required. It shows excellent scalability with grid size and is relatively insensitive to variations in bubble size, contrast, and number.
  • DIC-CG is a Strong Contender: DIC-CG also demonstrates good performance and scalability, often being the second-best method. Its convergence is generally robust, though its accuracy can be slightly lower due to the use of deflated residuals.
  • GMG and CCMG Struggle with High Contrasts: Geometric multigrid methods (GMG-CG and CCMG-CG) are highly sensitive to density contrasts, with their performance degrading significantly as the contrast increases.
  • Scalability: BoxMG-CG and DIC-CG exhibit near-optimal scalability with respect to grid size (N), with CPU time increasing by a factor of approximately 8 when doubling the grid points in each direction.
  • Difficult Test Case: For a challenging test problem with high density contrast and small bubbles, BoxMG-CG and DIC-CG were the only methods that could converge within reasonable limits, while others failed or were uncompetitive.

Limitations and Future Work

  • DIC-CG Coarse Grid Solvers: The paper suggests that improving the solvers for the coarse linear systems within DIC-CG could further enhance its efficiency and scalability.
  • AMG Implementation: The AMG implementation used was not optimized for structured grids, leading to higher costs. Future work could explore more efficient AMG implementations for such problems.
  • DIC-CG vs. BoxMG-CG: While BoxMG-CG generally performs better, further research could aim to improve DIC-CG to match BoxMG-CG's performance.
  • Parallelization: While the algorithms are amenable to parallelization, the presented experiments were conducted in a serial environment for convenience. The paper notes that parallel implementations of these methods exist and are an active area of research.

Summary updated

BibTeX
@article{maclachlan2008fast,
  author = {Scott P. MacLachlan and Jok Man Tang and Cornelis Vuik},
  title = {{Fast and robust solvers for pressure-correction in bubbly flow problems}},
  journal = {Journal of Computational Physics},
  year = {2008},
  doi = {10.1016/j.jcp.2008.07.022},
}
Figure