High-order finite element-integral equation coupling on embedded meshes

Natalie N. Beams, Andreas Klöckner, Luke N. Olson

Abstract

Highlights: This work presents a coupled integral equation–finite element method for solving interface problems. Our integral equation (IE) representations lead to well-conditioned second-kind integral equations, ensuring stability during refinement. Theoretical analysis supports the existence, uniqueness, and convergence of solutions, while numerical experiments demonstrate high-order accuracy throughout the solution domain.

Introduction

Solving partial differential equations (PDEs) like the Poisson equation is fundamental to computational physics, yet it becomes significantly more complex when the domain involves intricate boundaries or internal interfaces where material properties exhibit "jumps." Traditional Finite Element Methods (FEM) typically require "conforming meshes"—grids that are meticulously aligned with the geometry. Generating such meshes for complex shapes is often computationally expensive and labor-intensive.

Standard interface problem geometry with interior domain $\Omega^i$, exterior domain $\Omega^e$, and interface $\Gamma$.

Immersed boundary methods offer an alternative by using simple, structured background grids (embedded meshes). While flexible, these methods often struggle to maintain high-order accuracy near the interface or necessitate complex modifications like custom basis functions. This research proposes a "best of both worlds" approach: combining the flexibility of embedded meshes with the mathematical rigor and high-order accuracy of integral equations.

Methodology: FE-IE Coupling

The core innovation is a coupling strategy based on a "notional splitting" of the solution into two parts: a volume-based Finite Element part and a boundary-based Integral Equation part.

Embedded mesh setup showing the physical domain $\Omega$ within a structured background grid $\hat{\Omega}$.
  1. Finite Element Component: A standard FEM solver handles the inhomogeneous "forcing function" on a simple, rectangular "fictitious domain" ($\hat{\Omega}$) that ignores the complex interface. This decoupling allows the geometry to change without requiring volume mesh regeneration.
  2. Integral Equation Component: A boundary integral solver manages the boundary and interface conditions. It utilizes layer potentials to "correct" the FE solution, ensuring it satisfies the actual physical boundary conditions.

Key Technical Innovations

  • Second-Kind Formulations: The authors developed new formulations for exterior and interface problems that result in "second-kind" integral equations. These are mathematically well-behaved, ensuring the system remains easy to solve as the grid is refined.
  • Quadrature by Expansion (QBX): To accurately evaluate integral equations near the interface—where singularities typically cause issues—the method employs QBX via the pytential library.
  • Efficient Solvers: The system is solved using an iterative GMRES approach, preconditioned with Algebraic Multigrid (AMG) for the FE component, ensuring scalability for large-scale problems.

Numerical Results and Insights

The method was evaluated using "manufactured solutions" across various scenarios, including interior, exterior, and interface problems with complex shapes like a "starfish."

  • High-Order Convergence: The method consistently demonstrated high-order convergence (up to 4th order) even in the immediate vicinity of embedded boundaries. In the starfish test case, $L^2$ errors as low as $10^{-9}$ were achieved on relatively coarse meshes.
  • Robustness to Discontinuities: The coupling effectively handles both homogeneous and inhomogeneous "jump conditions" (discontinuities in the solution or its derivative) without requiring specialized modifications to the FE stencils.
  • Efficiency: Due to the second-kind formulation, the number of GMRES iterations remained nearly constant during mesh refinement, proving the method's stability and efficiency.

A critical insight from the study is the importance of smooth extensions. High-order accuracy depends on the ability to smoothly extend forcing data across the interface; "rough" data leads to a drop in convergence order, a phenomenon the authors quantified.

Conclusion and Future Work

By treating the solution as an additive split of FE and IE parts, this method avoids the common pitfalls of "cut-cell" integration or modified FE stencils. It remains compatible with off-the-shelf, highly-optimized libraries like deal.II and pyamg.

Future research will focus on extending the implementation from 2D to 3D and generalizing the theoretical error analysis for sharper, non-convex, or more irregular geometries. Additionally, the current reliance on smooth extendability of the forcing function $f$ remains a theoretical constraint to be addressed in future iterations of the framework.

Summary updated

Abstract

Highlights•Coupled integral equation–finite element methods for interface problems.•Our IE representations lead to well-conditioned second-kind integral equations.•Analysis supports existence, uniqueness, and convergence of solutions.•The methods demonstrate high-order accuracy throughout the solution domain.
BibTeX
@article{beams2018high,
  author = {Natalie N. Beams and Andreas Klöckner and Luke N. Olson},
  title = {{High-order finite element-integral equation coupling on embedded meshes}},
  journal = {Journal of Computational Physics},
  year = {2018},
  url = {https://openreview.net/forum?id=8UmTVwC5kR},
}
Figure