A hybrid shifted Laplacian multigrid and domain decomposition preconditioner for the elastic Helmholtz equations

Eran Treister, Rachel Yovel

TL;DR

This paper introduces a novel hybrid preconditioner that combines a shifted Laplacian multigrid method with domain decomposition for efficiently solving the elastic Helmholtz equation. This approach significantly improves the scalability and parallelizability of solving these complex equations, which are crucial for seismic imaging and other geophysical applications.

What Problem Does This Paper Address?

The elastic Helmholtz equation is a fundamental tool for modeling wave propagation in the Earth's subsurface, particularly for seismic imaging and full waveform inversion. However, solving this equation is computationally challenging due to its complexity, which is roughly three times that of the acoustic Helmholtz equation. This increased difficulty arises from the inclusion of shear wave physics and a "nullity-rich" grad-div term, making it harder for standard numerical methods like multigrid solvers to converge efficiently. Existing methods often struggle with large problem sizes, memory consumption, and parallelization, especially for 3D scenarios.

Methods

The authors propose a hybrid approach that leverages the strengths of two powerful techniques:

  • Shifted Laplacian Multigrid (SLMG): This method is an extension of a successful technique for the acoustic Helmholtz equation. It involves reformulating the elastic Helmholtz equation into a mixed formulation and then applying a multigrid solver to an "attenuated" or "shifted" version of the system. This shift helps to make the problem more amenable to multigrid solvers. A key innovation here is the use of a "Vanka" cell-wise relaxation scheme, which is adapted for the elastic case and shows good performance across a range of material properties (Poisson's ratio).
  • Domain Decomposition (DD): To address the memory and parallelization challenges of large-scale problems, the authors integrate DD techniques. DD breaks down the large domain into smaller, more manageable subdomains. These subdomains can then be solved in parallel. The paper explores different ways to combine DD with SLMG:
  • Using DD to solve the coarse grid problem within the multigrid hierarchy.
  • Using DD as a top-level preconditioner, with SLMG solving the subproblems.
  • A combined DD-MG-DD approach.

The paper also details the specific discretization (second-order staggered grid) and the mathematical formulation (mixed formulation) used to enable these methods.

Datasets

Not discussed in the paper. The paper focuses on numerical methods and uses synthetic models (e.g., constant media, linear models, Marmousi2) for validation.

Key Insights and Ideas

  • Elasticity is not that much harder: The paper's core message is that the elastic Helmholtz equation, while more complex than the acoustic one, can be solved with comparable efficiency using the right techniques. The "added difficulty" is manageable.
  • Mixed formulation is key: Reformulating the elastic equation into a mixed system (introducing a pressure variable) is crucial for enabling efficient multigrid solvers, particularly for handling the "nearly incompressible" material cases.
  • Cell-wise Vanka smoother is effective: The adapted Vanka smoother shows good performance and scalability with respect to material properties like Poisson's ratio, which is a significant advantage over some other methods.
  • DD and SLMG are synergistic: The local nature of both DD and SLMG allows them to be combined effectively. DD provides parallelism and helps manage large problem sizes, while SLMG efficiently solves the subproblems. This combination avoids significant loss in convergence rate.
  • Local nature of attenuation: The attenuation introduced by the shifted Laplacian has a local support, meaning it doesn't require global information to be captured. This property is essential for the success of domain decomposition.

Key Results

  • Scalability with Poisson's Ratio: The proposed hybrid method shows performance that is largely independent of the Poisson's ratio, a significant improvement over some existing methods that struggle with nearly incompressible materials.
  • Comparable to Acoustic Solvers: The elastic solver achieves performance comparable to well-established shifted Laplacian solvers for the acoustic Helmholtz equation, with the primary difference being the relevant wavenumber (shear vs. pressure).
  • Efficient Parallelization: The hybrid DD-SLMG approach enables efficient parallelization, allowing for the solution of large 2D and 3D problems that were previously intractable.
  • Reduced Memory Footprint: By using DD, the memory requirements are managed, especially for 3D problems, by breaking them into smaller subdomains.
  • Numerical Examples: The paper presents numerical results on various models (linear, Marmousi2) demonstrating the method's efficiency and robustness in terms of convergence cycles and scalability. For instance, in 2D experiments, the number of preconditioning cycles remained relatively stable as the Lamé parameter (related to Poisson's ratio) increased.

Limitations and Future Work

  • Coarse Grid Problem: For very large 3D problems, the coarsest grid in the multigrid hierarchy can still become quite large, posing memory challenges for direct solvers (LU decomposition). The paper addresses this by using DD for the coarse grid problem.
  • "Ideal" Coarse Grid Correction: The theoretical analysis relies on the assumption of an "ideal" coarse grid correction, which is an approximation. The paper notes that the two-grid factor can provide a more accurate prediction when this assumption doesn't hold perfectly.
  • Future Work:
    • Investigating more sophisticated DD methods and sweeping orders.
  • Exploring adaptive discretization to better handle varying shear wavenumbers.
  • Further research into dealing with the indefiniteness of Helmholtz problems.
  • Leveraging the relationship between acoustic and elastic solvers to develop new methods.

Summary updated

Abstract

Highlights•A scalable Shifted Laplacian multigrid method for the Elastic Helmholtz equation.•The elastic Helmholtz equation is solved at the same efficiency as the acoustic.•We provide detailed local Fourier analysis to validate our numerical results.•Domain decomposition is combined to enhance parallelism and reduce memory costs.•The two parts of the hybrid method work in synergy due to their local nature.
BibTeX
@article{treister2024hybrid,
  author = {Eran Treister and Rachel Yovel},
  title = {{A hybrid shifted Laplacian multigrid and domain decomposition preconditioner for the elastic Helmholtz equations}},
  journal = {Journal of Computational Physics},
  year = {2024},
  doi = {10.1016/j.jcp.2023.112622},
}
Figure