An algebraic preconditioner for the exactly divergence-free discontinuous Galerkin method for Stokes

Sander Rhebergen, Ben S. SouthwortharXiv 2022 · OpenReview

TL;DR

This paper introduces a new, "black-box" friendly algebraic preconditioner for solving the Stokes equations using a specific, high-accuracy numerical method (exactly divergence-free discontinuous Galerkin). By "hybridizing" the numerical method, they make it compatible with standard solvers, achieving optimal performance that scales well with mesh size and polynomial degree, and importantly, paving the way for similar methods for the more complex Navier-Stokes equations.

What Problem Does This Paper Address?

Solving the Stokes equations, which describe fluid flow at low Reynolds numbers, is a fundamental problem in computational fluid dynamics. The paper focuses on a particular numerical method called the "exactly divergence-free discontinuous Galerkin" (DG) method. While this method offers desirable properties like local conservation and energy stability, it uses a special type of basis function (H(div)-conforming) that makes it difficult for standard iterative solvers (like those using algebraic multigrid) to efficiently find solutions. Existing methods to overcome this often require custom, specialized solvers or complex mathematical machinery, limiting their applicability and generalization, especially to the more challenging Navier-Stokes equations which include advection. The authors aim to bridge this gap by developing a preconditioner that is effective, generalizable, and relies on readily available solver components.

Methods

The core innovation lies in "hybridizing" the H(div)-conforming finite element space for the velocity. Instead of directly using the H(div) space, they introduce a Lagrange multiplier on the element faces. This effectively transforms the velocity space into a more standard L²-DG space, which is much more amenable to standard solvers. The original H(div)-conformity constraint is then enforced through the Lagrange multiplier.

This hybridization leads to a saddle-point system. The authors then propose a preconditioner for this system. It consists of:

  1. An approximate inverse for the velocity block, achieved using standard Algebraic Multigrid (AMG) methods.
  2. A spectrally equivalent preconditioner for the pressure and Lagrange multiplier Schur complement. This is constructed using weighted mass matrices, where the weighting is carefully chosen based on theoretical inf-sup conditions of the discrete problem.

The resulting preconditioner, denoted by $P$, is designed to be "black-box" friendly, meaning it relies on standard linear algebra components and does not require custom solvers for the underlying finite element spaces.

Datasets

Not discussed in the paper. The paper focuses on numerical methods and their performance on representative test cases, rather than on specific datasets.

Key Insights and Ideas

  • Hybridization as a "Black-Box" Enabler: The key insight is that by hybridizing the H(div)-conforming space, the problem becomes compatible with standard, widely available solvers like AMG. This bypasses the need for specialized solvers tailored to H(div) spaces.
  • Weighted Mass Matrices for Schur Complement: The effectiveness of the preconditioner for the pressure/Lagrange multiplier block hinges on carefully weighting the mass matrices. This weighting, derived from theoretical stability conditions, significantly improves performance compared to unweighted approaches.
  • Super Robustness in Element Order: The proposed preconditioner demonstrates remarkable robustness with respect to the polynomial degree ($k$) of the finite elements. Iteration counts remain low and stable even for high-order elements, a significant advantage for achieving high accuracy.
  • Path to Navier-Stokes: The hybridization strategy naturally extends to the Navier-Stokes equations, offering a promising avenue for developing effective preconditioners for these more complex, advection-dominated flows.

Key Results

The paper presents numerical results demonstrating the effectiveness of their proposed preconditioner.

  • Optimal Convergence: The preconditioner is shown to be "optimal in $h$", meaning the number of iterations required to solve the system does not grow with decreasing mesh size ($h$).
  • Super Robustness in Element Order: For 2D and 3D test cases, the number of MINRES iterations required to reach a target residual tolerance (10⁻⁸) remains remarkably low and stable across polynomial degrees from 2 to 5. For example, in 2D, iteration counts range from 21 to 39 for $k=5$ with optimal weighting, compared to over 100 for the unweighted case.
  • Significant Speedup: The weighted mass matrix approach provides a 2-3x reduction in iteration counts compared to unweighted mass matrices.
  • 3D Effectiveness: The method extends effectively to 3D, with iteration counts often being lower than in 2D.
  • AMG Compatibility: The velocity block can be efficiently preconditioned using standard AMG, although higher-order discretizations may require more inner AMG iterations.
  • Contrast with Non-Hybridized DG: Numerical results highlight the poor performance of standard AMG on the non-hybridized H(div)-DG discretization of the vector Laplacian, underscoring the benefit of the hybridization approach.

Limitations and Future Work

  • Parameter Tuning: While the method is robust, the optimal weighting parameters for the mass matrices are derived from theoretical inf-sup constants, which are not always known in practice. The paper suggests practical strategies for choosing these parameters and shows that they are moderately independent of element order.
  • AMG Iterations for High Order: For higher-order DG discretizations of diffusion, more inner AMG iterations are needed for the velocity block. The authors suggest that alternative preconditioning frameworks might reduce this requirement.
  • Future Work: The authors explicitly state that their hybridization approach naturally extends to Navier-Stokes problems. Future work will focus on developing block preconditioning techniques for the extended hybridized system of Navier-Stokes, aiming for efficient solvers for high-Reynolds number flows.

Summary updated

Abstract

We present an optimal preconditioner for the exactly divergence-free discontinuous Galerkin (DG) discretization of Cockburn, Kanschat, and Sch\"otzau [J. Sci. Comput., 31 (2007), pp. 61--73] and Wang and Ye [SIAM J. Numer. Anal., 45 (2007), pp. 1269--1286] for the Stokes problem. This DG method uses finite elements that use an $H({\rm div})$-conforming basis, thereby significantly complicating its solution by iterative methods. Several preconditioners for this Stokes discretization have been developed, but each is based on specialized solvers or decompositions, and do not offer a clear framework to generalize to Navier--Stokes. To avoid requiring custom solvers, we hybridize the $H({\rm div})$-conforming finite element so that the velocity lives in a standard $L^2$-DG space, and present a simple algebraic preconditioner for the extended hybridized system. The proposed preconditioner is optimal in $h$, super robust in element order (demonstrated up to 5th order), effective in 2d and 3d, and only relies on standard relaxation and algebraic multigrid methods available in many packages. The hybridization also naturally extends to Navier--Stokes, providing a potential pathway to effective black-box preconditioners for exactly divergence-free DG discretizations of Navier--Stokes.
BibTeX
@misc{rhebergen2022algebraic,
  author = {Sander Rhebergen and Ben S. Southworth},
  title = {{An algebraic preconditioner for the exactly divergence-free discontinuous Galerkin method for Stokes}},
  year = {2022},
  eprint = {2203.14881},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.14881},
}
Figure