TL;DR
This paper introduces a novel "Embedded Boundary Integral" (EBI) method for efficiently solving the Stokes equations in complex geometries. By combining boundary integral equations with regular grid solvers, the EBI method achieves second-order accuracy and optimal algorithmic complexity, making it suitable for problems with moving boundaries where traditional mesh generation is prohibitive.
What Problem Does This Paper Address?
The Stokes equations are fundamental for modeling low Reynolds number flows and are crucial building blocks for more complex fluid dynamics simulations, such as the Navier-Stokes equations. Solving these equations in complex or moving geometries presents a significant challenge, primarily due to the difficulty and computational cost associated with generating and adapting high-quality meshes. Traditional methods like finite elements require extensive mesh generation, which is a bottleneck for time-dependent problems. Furthermore, existing fast solvers for integral equations often rely on analytic kernel expansions that can be computationally expensive and difficult to implement, especially for problems with singularities or near-field interactions. This paper aims to provide a robust, efficient, and implementable solver that bypasses these limitations.
Methods
The core of the proposed method is the "Embedded Boundary Integral" (EBI) approach, which extends prior work on Poisson and biharmonic equations. The strategy involves:
- Domain Embedding: The complex geometry domain ($\omega$) is embedded within a simpler, rectangular domain ($\Omega$) for which fast solvers are readily available.
- Problem Decomposition: The original Stokes problem is split into two parts:
- A problem on the larger domain $\Omega$ with an inhomogeneous body force and zero boundary conditions.
- A problem on the original domain $\omega$ with no body force but nontrivial boundary conditions.
- Boundary Integral Formulation: An indirect double-layer potential formulation is used to solve the boundary integral equation for the second problem. This approach leads to a Fredholm equation of the second kind, which is then discretized using Nyström's method.
- Fast Integral Equation Solver: To accelerate the solution of the boundary integral equation, a fast matrix-vector multiplication algorithm is employed. This algorithm leverages the fact that the interaction matrix is often numerically low-rank, especially for well-separated boundary segments. It uses Singular Value Decomposition (SVD) to sparsify large low-rank blocks of the discretized double-layer operator. This approach is kernel-independent and easier to implement than Fast Multipole Methods (FMM).
- Regular Grid Solver: For the problems on the rectangular domain $\Omega$, a finite element discretization (Q1-Q1 bilinear elements with stabilization) is used. This formulation is shown to be equivalent to a finite difference discretization for certain stencils.
- Taylor Expansion Stencil Corrections (TESC): To ensure second-order accuracy for the regular grid solver, especially near the embedded boundary, Taylor expansion stencil corrections are computed. These corrections account for the jumps in the solution across the interface, derived from the boundary integral equation solution.
Datasets
Not discussed in the paper. The paper focuses on developing a numerical method and tests it on various analytical solutions and geometries.
Key Insights and Ideas
- Bridging Integral and Grid Methods: The EBI method elegantly combines the strengths of boundary integral equations (handling complex geometries efficiently) and regular grid solvers (fast, efficient for bulk computations). This hybrid approach circumvents the need for complex mesh generation.
- Embedded Boundary as Interface: Treating the embedded boundary as an interface where jump conditions are imposed is a key conceptual simplification. This allows for decoupling the problem into a regular grid part and a boundary integral part.
- SVD for Matrix Compression: The application of SVD to sparsify the dense interaction matrices arising from boundary integral discretizations is a powerful technique. It provides a practical way to achieve near-linear complexity for matrix-vector products, which is crucial for scalability.
- TESC for Accuracy: The Taylor Expansion Stencil Corrections are vital for achieving second-order accuracy in the regular grid solver, particularly when dealing with discontinuities introduced by the embedded boundary.
Key Results
- Second-Order Accuracy: The method consistently achieves second-order accuracy for both velocities and pressures, as demonstrated in numerical experiments.
- Optimal Algorithmic Complexity: The EBI method achieves an optimal O(N log N) complexity, where N is the number of discretization points. This is a significant improvement over methods that rely on dense matrix inversions or expensive mesh generation.
- Scalability: Numerical results show that the SVD-based approach scales almost linearly with the problem size, outperforming dense matrix methods which scale quadratically. For instance, for a problem with 1024² grid points, the 2L-ASM preconditioner solves the system in 19 seconds.
- Robustness: The method is demonstrated to work for various geometries, including those with holes and complex shapes, and for different flow types (Poiseuille, cubic flow, body force flow, Stokeslet).
Limitations and Future Work
- Geometry Regularity: The current implementation has stringent requirements on the regularity of the boundary geometry. While singularities at corners can be handled analytically, this becomes more complex in 3D.
- Piecewise Constant Coefficients: The method is presented for problems with piecewise constant coefficients. Extending it to variable coefficients would require further development.
- Higher-Order Extensions: The authors suggest that higher-order accuracy can be achieved by differentiating the hydrodynamic density further and using high-order or spectral regular grid discretizations.
- Neumann and Mixed Boundary Value Problems: The paper notes that the method can be extended to Neumann and mixed boundary value problems, which would be a valuable future direction.
- 3D Extension: While the paper focuses on 2D, the underlying principles are generalizable to 3D, which is a natural area for future research.
Summary updated