Generalized multiscale finite element method. Symmetric interior penalty coupling

Yalchin Efendiev, Juan Galvis, Raytcho D. Lazarov, M. Moon, Marcus Sarkis

TL;DR

This paper introduces a novel Discontinuous Galerkin (DG) framework for multiscale finite element methods (MSFEMs) to efficiently simulate fluid flow in highly heterogeneous porous media. By constructing specialized coarse-grid basis functions using local spectral problems and a symmetric interior penalty, the method achieves accurate solutions with significantly reduced computational cost, especially for high-contrast materials.

What Problem Does This Paper Address?

Simulating fluid flow in porous media with highly varying properties (high contrast) is computationally demanding. Traditional finite element methods require extremely fine grids to capture all the small-scale details, leading to massive systems of equations that are expensive to solve. Existing multiscale methods aim to reduce this computational burden by using coarser grids and constructing specialized basis functions that implicitly capture the multiscale physics. However, effectively handling high-contrast materials and ensuring stability and accuracy within these reduced-order models remains a challenge. This paper addresses this by developing a robust MSFEM framework that leverages the strengths of Discontinuous Galerkin methods to handle the discontinuities arising from local basis function construction, particularly in the presence of high contrast.

Methods

The core of this work lies in developing a multiscale finite element method within the Discontinuous Galerkin (DG) framework. The approach involves a two-stage process:

  • Offline Stage: This is where the multiscale basis functions are constructed. The authors propose three methods for generating these local basis functions on a coarse grid:
  • Local Eigenvalue Problem (Method I): Based on a local eigenvalue problem using a weighted L2-norm for the mass matrix.
  • Amended Eigenvalue Problem (Method II): Modifies the first method by adding a penalty term to the eigenvalue problem, aiming for improved error estimates.
  • Snapshot Space (Method III): Generates a space of local solutions (snapshots) and then selects a reduced subspace using spectral decomposition.
  • Online Stage: The pre-computed multiscale basis functions are used to solve the problem on the coarse grid, leading to rapid computations for different parameter realizations.

A key innovation is the use of the Symmetric Interior Penalty Discontinuous Galerkin (SIPG) method. This method is particularly suited for coupling the locally constructed, potentially discontinuous, multiscale basis functions. The SIPG formulation includes a penalty term that is crucial for stability, and the paper investigates its scaling and impact on accuracy. The authors also derive theoretical error estimates and conduct extensive numerical experiments to validate their approach.

Datasets

Not discussed in the paper. The paper focuses on numerical experiments using a synthetic high-contrast coefficient field, illustrated in Figure 1, which represents channels and inclusions within a porous medium.

Key Insights and Ideas

  • DG for MSFEMs: The Discontinuous Galerkin method provides a natural and effective framework for coupling locally constructed multiscale basis functions, which are often discontinuous across coarse-grid interfaces.
  • Symmetric Interior Penalty (SIP): The SIP formulation, combined with a carefully chosen penalty parameter, is shown to be crucial for stabilizing the DG method and achieving good accuracy, especially in high-contrast scenarios.
  • Local Spectral Problems: Using local eigenvalue problems to generate basis functions that capture dominant local physics is a powerful technique for model reduction. The amended eigenvalue problem (Method II) shows slightly better performance in numerical tests.
  • Error Decay with Basis Functions: The numerical results demonstrate that the error decreases as more multiscale basis functions are added to the coarse-grid space, and this decay is often proportional to the smallest excluded eigenvalue.
  • Penalty Parameter Sensitivity: The boundary error is more sensitive to the penalty scaling than the interior error. Choosing an appropriate penalty parameter is important for balancing interior and interface errors.

Key Results

  • The proposed DG-MSFEM methods achieve significant error reduction as the number of multiscale basis functions increases.
  • For a contrast of 10^4 and 10^6, the methods show convergence to the fine-grid solution.
  • With a coarse mesh size H = 1/10 and fine mesh size h = 1/100, only a few additional basis functions (e.g., 3-4) are needed to achieve a substantial reduction in interior error (around 17%).
  • The numerical experiments suggest that the penalty parameter does not need to be excessively large; a small penalty can still yield good results, particularly for time-dependent problems.
  • Method II (amended eigenvalue problem) generally yields slightly better results than Method I.

Limitations and Future Work

  • Penalty Parameter Optimization: While the paper investigates the effect of penalty scaling, it does not delve into finding an optimal penalty parameter for all scenarios, especially for time-dependent problems. This is suggested as future work.
  • Other Coupling Strategies: The authors mention that other techniques for coupling discontinuous spectral basis functions, such as hybridized DG methods or mortar methods, could be explored to avoid artificial penalty terms.
  • Snapshot Space Selection: The paper focuses on demonstrating the DG framework's flexibility with different snapshot spaces but does not extensively explore the optimal selection of these spaces, which can depend on the input space.
  • Smoothness Dependence: Remark 3.6 points out that further bounds on the error might require studying the decay of coefficients related to eigenvalues, which can depend on the smoothness of the solution and warrants further research.

Summary updated

Abstract

Motivated by applications to numerical simulation of flows in highly heterogeneous porous media, we develop multiscale finite element methods for second order elliptic equations. We discuss a multiscale model reduction technique in the framework of the discontinuous Galerkin finite element method. We propose three different finite element spaces on the coarse mesh. The first space is based on a local eigenvalue problem that uses a weighted $L_2-$norm for computing the "mass" matrix. The second space is generated by amending the eigenvalue problem of the first case with a term related to the penalty. The third choice is based on generation of a large space of snapshots and subsequent selection of a subspace of a reduced dimension. The approximation with these spaces is based on the discontinuous Galerkin finite element method framework. We investigate the stability and derive error estimates for the methods and further experimentally study their performance on a representative number of numerical examples.
BibTeX
@article{efendiev2013generalized,
  author = {Yalchin Efendiev and Juan Galvis and Raytcho D. Lazarov and M. Moon and Marcus Sarkis},
  title = {{Generalized multiscale finite element method. Symmetric interior penalty coupling}},
  journal = {Journal of Computational Physics},
  year = {2013},
  doi = {10.1016/j.jcp.2013.07.028},
}
Figure