Multiscale Simulations for Coupled Flow and Transport Using the Generalized Multiscale Finite Element Method

Eric T. Chung, Yalchin Efendiev, Wing Tat Leung, Jun RenComput 2015 · DOI

TL;DR

This paper introduces a novel multiscale finite element method (GMsFEM) for simulating coupled flow and transport in heterogeneous porous media. By constructing specialized multiscale basis functions, the method achieves accurate results on a coarse grid, significantly reducing computational cost while maintaining mass conservation.

What Problem Does This Paper Address?

Simulating fluid flow and solute transport in porous media is crucial for many applications, such as groundwater contamination, oil recovery, and carbon sequestration. These problems often involve highly heterogeneous materials, meaning properties like permeability vary drastically over short distances. Traditional numerical methods struggle with this scale disparity: using a fine grid to capture all the small-scale details is computationally prohibitive, while using a coarse grid with standard methods leads to significant errors. Existing "upscaling" methods that average properties on a coarse grid can also lose critical information. This paper tackles this challenge by developing a multiscale method that intelligently constructs basis functions on a coarse grid to represent the complex fine-scale physics.

Methods

The core of this work is the Generalized Multiscale Finite Element Method (GMsFEM). The authors adapt GMsFEM for a coupled system of a steady-state elliptic flow equation (governed by Darcy's law) and a convection-dominated parabolic transport equation.

Key methodological contributions include:

  • Mixed Formulation: The authors employ a mixed formulation for both the flow and transport equations. This is crucial for ensuring mass conservation, a fundamental physical principle, and also helps stabilize the convection-dominated transport equation.
  • Multiscale Basis Function Construction:
    • Snapshot Spaces: For both the velocity/flux and concentration fields, they first define "snapshot spaces." These spaces are generated by solving local problems on each coarse grid block with various boundary conditions that represent all possible scenarios. This captures the essential multiscale information within each block.
  • Local Spectral Decomposition: To reduce the dimensionality of these snapshot spaces, they perform local spectral decompositions. This process identifies the dominant modes (eigenvectors corresponding to the smallest eigenvalues) that best represent the multiscale behavior.
  • Offline Construction: These dominant modes are then used to construct the multiscale basis functions for the coarse grid. This construction is done "offline," meaning it's a one-time cost before the actual simulation begins.
  • Coupled System Solver: The GMsFEM framework is applied to the coupled flow and transport system, allowing for a full coarse-grid approximation.

Datasets

Not discussed in the paper. The paper uses synthetic examples with defined permeability fields and source terms for numerical experiments.

Key Insights and Ideas

  • More Basis Functions, Better Accuracy: A key finding is that using just one multiscale basis function per coarse edge (analogous to traditional flow-based upscaling) is insufficient for capturing the fine-scale solution behavior, especially for the concentration field. Employing two basis functions significantly improves accuracy and captures the essential features of the solution.
  • Mass Conservation is Key: The use of a mixed formulation is highlighted as essential for guaranteeing mass conservation, which is critical for accurate transport simulations.
  • GMsFEM's Adaptability: The GMsFEM framework, with its offline construction of multiscale basis functions, offers a robust way to handle complex heterogeneities without requiring a fine-grid simulation. The method is also noted to handle arbitrary coarse-grid geometries.
  • Convection Dominance Challenge: The paper specifically addresses the challenge of convection-dominated transport, where standard finite element methods can suffer from spurious oscillations. The chosen mixed formulation and test spaces help mitigate this.

Key Results

The paper demonstrates the effectiveness of their GMsFEM approach through numerical experiments.

  • Reduced Errors with More Basis Functions: In the first experiment (Figure 4), using two multiscale basis functions per edge for both velocity and flux-concentration reduced L² errors in concentration from over 20% (with one basis function) to around 8-11% at various time points.
  • Improved Accuracy in Convection-Dominated Cases: In a second, more convection-dominated scenario (Figure 5), the improvement was even more pronounced. Using one basis function resulted in very high L² errors (over 70%), while two basis functions reduced these errors to around 15%.
  • Time-Dependent Velocity Case: The method also showed comparable accuracy for a time-dependent velocity case (Figure 6), with L² errors around 5-8%.

Limitations and Future Work

  • Piecewise Constant Concentration Basis: The current implementation uses piecewise constant basis functions for concentration. While this keeps computational costs low and avoids inf-sup condition issues, it limits the accuracy to O(H) (where H is the coarse grid size).
  • Future Work: The authors plan to:
  • Systematically design multiscale basis functions for the concentration that can provide stable discretizations.
  • Address time-dependent velocity fields more comprehensively, potentially involving offline computation of basis functions or online adaptation using residuals.
  • Explore joint multiscale basis construction for concentration without needing to solve the velocity field first.
  • Investigate more complex scenarios, such as those involving compressibility or time-varying concentration influencing velocity.

Summary updated

BibTeX
@article{chung2015multiscale,
  author = {Eric T. Chung and Yalchin Efendiev and Wing Tat Leung and Jun Ren},
  title = {{Multiscale Simulations for Coupled Flow and Transport Using the Generalized Multiscale Finite Element Method}},
  journal = {comput.},
  year = {2015},
  doi = {10.3390/computation3040670},
}
Figure