An adaptive GMsFEM for high-contrast flow problems

1 Introduction

• Prior GMsFEM work derived a-priori bounds, but rigorous a-posteriori indicators are needed to drive adaptive enrichment.The proposed estimators use a special local eigenvalue problem and the eigenstructure of the offline space.Source: §1 · p. 2
• The paper derives local residual-based error indicators for adaptive enrichment of GMsFEM spaces in linear elliptic problems with multiscale high-contrast coefficients.The indicators use either the L2 norm or a coefficient-weighted H−1 norm of the local residual.Source: §1 · pp. 1–2
• The convergence analysis concerns enrichment on a fixed coarse mesh and uses variational indicators that avoid coefficient gradients, without requiring bounded coefficient gradients.The analysis assumes coarse-grid error is the main contributor and excludes fine-grid discretization error from local problems.Source: §1 · p. 2
• The method uses snapshot spaces to approximate the residual error for the weighted H−1 indicator, with an irreducible error relative to the fine-grid solution.The paper also compares fine-grid and snapshot solutions against the offline solution.Source: §1 · p. 2
• Weighted H−1 residual indicators are reported as more robust than L2 indicators for high-contrast media, and both are used to guide adaptive enrichment.Numerical tests on two high-contrast permeability fields also find results from the weighted indicator nearly similar to those from an exact energy-norm error indicator.Source: §1 · p. 2

2 Preliminaries

• The coefficient κ(x) is heterogeneous, multiscale, and high contrast; each coarse element is partitioned into fine-grid blocks.The computational domain has homogeneous Dirichlet boundary conditions, and the fine grid refines the coarse grid.Source: §2 · p. 3
• The notation ω_i denotes a coarse neighborhood, while K denotes a coarse element in the mesh illustration..Source: §2 · p. 3, Figure 1 · p. 3
• The continuous Galerkin solution is sought using multiple basis functions supported in each coarse neighborhood, with the basis functions spanning V_off.A discontinuous Galerkin coupling is also noted as an option.Source: §2 · p. 3

3 CG-based GMsFEM for high-contrast flow problems

• The outlined workflow generates a coarse grid, constructs snapshot and reduced offline spaces, then forms parameter-specific multiscale basis functions for coarse-grid solves.Iterative solvers may be used if needed.Source: §3 · p. 4

3.1 Local basis functions

• Local snapshots may comprise fine-scale basis functions or local problem solutions, including κ-harmonic extensions; a spectral problem reduces the snapshot space.The offline space retains eigenvectors associated with the smallest selected eigenvalues.Source: §3.1 · pp. 4–5

3.2 Global coupling

• Global coupling multiplies offline eigenfunctions by partition-of-unity functions to form continuous basis functions spanning the CG multiscale space.The resulting coarse system is formed by projecting the fine-scale stiffness matrix and forcing vector.Source: §3.2 · pp. 5–6

• The weighted H−1 residual indicator is more robust for high-contrast media, while the L2 residual works well when coefficients lack high-contrast regions.The section contrasts their high-contrast convergence behavior.Source: §4 · pp. 6–7
• Theorem 4.1 concerns u and ums, defined as the respective solutions of (11) and (12), and identifies Cerr as a uniform constant in its stated result.Source: §4 · p. 7
• The adaptive procedure solves in the current space, computes local residuals, selects regions for enrichment, and adds eigenfunction basis functions.The algorithm repeats this process at successive enrichment levels.Source: §4 · p. 7
• The enrichment algorithm has a contracting property, and its added basis functions are bounded by the snapshot-space dimension.The choice of added eigenfunctions is stated to ensure convergence.Source: §4 · pp. 7–8

5 Numerical Results

• Across the tested cases, adaptive GMsFEM converges; the weighted H−1 indicator estimates exact error well and outperforms the L2 indicator for high-contrast media.For equal-dimensional offline spaces, θ = 0.2 gives more accurate results than θ = 0.7.Source: §5 · pp. 9–10

5.1 Numerical results with harmonic basis

• With harmonic bases, the proposed and exact indicators yield similar convergence, while adaptive enrichment is more accurate with a smaller offline space than uniform enrichment.The θ = 0.2 case achieves a similar error level with a smaller offline space than θ = 0.7.Source: Table 1 · p. 10, §5.1 · p. 11, Table 3 · p. 11

5.2 Numerical results with spectral basis

• With spectral bases, adaptive enrichment converges and uses a smaller offline space with better accuracy than uniform enrichment; the exact indicator performs better in this case.The proposed and exact indicators have similar convergence behavior and offline-space dimensions.Source: §5.2 · pp. 13–14, Table 4 · p. 13, Table 6 · p. 13

5.3 Numerical results with the L2 indicator

• For harmonic bases, the L2 indicator requires more iterations and a larger offline space than the proposed indicator for similar accuracy, and its energy error decays slowly.The indicator’s basis distribution is similar, but it selects more basis functions than the proposed indicator.Source: §5.3 · p. 15, Table 7 · p. 15

5.4 Numerical results using snapshot solutions for the proposed indicator

• Computing the proposed indicator in snapshot space gives similar offline-space size and accuracy to fine-scale computation, but requires more iterations.The snapshot-space computation solves local problems in the snapshot space to reduce computational costs.Source: §5.4 · pp. 15–16
• 22 iterations: the algorithm converges for the harmonic basis using snapshot space to compute the proposed indicator, with θ = 0.7.Source: Table 8 · p. 16

6 Convergence analysis

• The analysis proves convergence of adaptive enrichment as the approximation space is enriched on a fixed coarse mesh.The proof uses stability of the coarse-grid projection and spectral estimates.Source: §6 · p. 16

6.1 Proof of Theorem 4.1

• The proof of Theorem 4.1 derives a-posteriori error estimates by decomposing the error through coarse-grid projections and local-region contributions.The inequalities bound the energy error using the local residuals.Source: §6.1 · pp. 16–17

6.2 Some auxiliary lemmas

• Auxiliary lemmas establish upper bounds for energy error from residuals and recursive properties of the error quantities across enrichment levels.The section treats both H−1-based and L2-based residuals.Source: §6.2 · pp. 17, 20, 21

6.3 Proof of Theorem 4.2

• The convergence proof unifies the two residual types and selects coarse regions for enrichment using a threshold parameter.The proof chooses parameters so that the resulting contraction factor satisfies 0 < ρ < 1.Source: §6.3 · pp. 22–24

7 Conclusions

• The weighted H−1 residual indicator is reported as more appropriate than the L2-based indicator for high-contrast multiscale problems.The paper also reports convergence of the indicators and similarities to results using the exact solution in the indicator.Source: §7 · p. 24
• The analysis excludes fine-grid discretization error, assumes coarse-grid error is the main contributor, and treats the fine-grid solution as the desired quantity.Adding discretization errors from basis computations is left for future research.Source: §7 · p. 24