Source-linked AI summary
Generalized Multiscale Finite Element Methods (GMsFEM)
Yalchin Efendiev, Juan Galvis, Thomas Y. Hou
TL;DR
Multiscale PDE simulations with complex, high-dimensional inputs are expensive, while existing reduced spaces may not remain effective across changing forcings or boundary conditions. GMsFEM builds local snapshot and offline spaces, then efficiently selects online coarse bases through spectral decomposition for each input. The reported results show decreasing errors as local spectral spaces are enlarged, with robust convergence and reuse across forward problems.
Problem
Multiscale PDE simulations with uncertainties, many source terms, and high-dimensional parameter spaces are computationally expensive, while existing reduced spaces may not remain small or reusable when forcings or boundary conditions change.
Method
GMsFEM constructs local snapshot spaces, reduces them spectrally into offline spaces, and selects largest-eigenvalue modes online for each input parameter.
Results
GMsFEM errors decrease as local spectral-space dimensions increase, with three iterations across tested coarse-space dimensions; convergence is robust to contrast and error correlates with the omitted largest eigenvalue.
Takeaways & Limitations
The local offline-online construction enables efficient coarse-grid solves for varying parameters, source terms, and boundary conditions while eliminating unnecessary coarse-level degrees of freedom.
Takeaways & Limitations
The framework assumes κ(·,·;µ) is bilinear, coercive, continuous for each µ, and sufficiently smooth with respect to µ; δE must also be selected because it affects performance.
Abstract
from arXiv · showhide
In this paper, we propose a general approach called Generalized Multiscale Finite Element Method (GMsFEM) for performing multiscale simulations for problems without scale separation over a complex input space. As in multiscale finite element methods (MsFEMs), the main idea of the proposed approach is to construct a small dimensional local solution space that can be used to generate an efficient and accurate approximation to the multiscale solution with a potentially high dimensional input parameter space. In the proposed approach, we present a general procedure to construct the offline space that is used for a systematic enrichment of the coarse solution space in the online stage. The enrichment in the online stage is performed based on a spectral decomposition of the offline space. In the online stage, for any input parameter, a multiscale space is constructed to solve the global problem on a coarse grid. The online space is constructed via a spectral decomposition of the offline space and by choosing the eigenvectors corresponding to the largest eigenvalues. The computational saving is due to the fact that the construction of the online multiscale space for any input parameter is fast and this space can be re-used for solving the forward problem with any forcing and boundary condition. Compared with the other approaches where global snapshots are used, the local approach that we present in this paper allows us to eliminate unnecessary degrees of freedom on a coarse-grid level. We present various examples in the paper and some numerical results to demonstrate the effectiveness of our method.
1 Introduction
The paper introduces GMsFEM, a local multiscale model-reduction framework for expensive PDE simulations with multiscale structure and high-dimensional inputs. It constructs offline and online coarse spaces to reduce unnecessary coarse-level degrees of freedom while accommodating varying parameters, source terms, and boundary conditions.
- Motivation: Multiscale simulations are costly because small scales, uncertainties, and many source terms require repeated forward solves over high-dimensional input spaces.The paper frames these problems through an input-output relation whose output is often lower-dimensional than the input.
- Limitations of existing approaches: Existing reduced-order methods may fail to produce small solution spaces for multiscale structures and may require regeneration when forcing or boundary conditions change.Global snapshot approaches specifically require recomputation after such changes.
- Local model reduction: Local approaches approximate solutions in coarse-grid regions across input parameters without computing global solution snapshots.They first construct an offline space, then derive an online subspace for each input parameter.
- Local model reduction: Local model reduction eliminates unnecessary parameter-space degrees of freedom at the coarse-grid level and can use local spaces largely independent of source terms or boundary conditions.This avoids expensive global snapshots that may not contribute to the online solution.
- GMsFEM framework: GMsFEM constructs a small offline space and uses it to generate reusable multiscale basis functions for coarse-grid problems with different input parameters.The framework divides computation into offline construction and online basis-function computation.
- GMsFEM framework: The offline space is obtained by spectral decomposition of local snapshot spaces, while the online coarse space selects eigenvectors with the largest eigenvalues for each input parameter.The resulting online space is used in a finite element solve on the coarse grid.
- Methodological distinction: GMsFEM emphasizes systematic online enrichment and careful construction of initial partition-of-unity functions to complement local spaces and converge toward the fine-grid solution.The paper also discusses iterative coarse-space corrections for fine-scale convergence.
2 A Generalized Multiscale Finite Element Method.
GMsFEM constructs local snapshot and offline spaces, then adaptively forms a small online coarse space for each input parameter through spectral decomposition. The framework supports multiscale operators, complex parameter spaces, and several snapshot constructions while reducing online computational cost.
- General formulation: The framework accommodates heterogeneous operators with nonlinear dependence on spatial fields and linear dependence on source terms and boundary conditions.An affine operator representation enables coarse-scale projections to be precomputed offline and reused online, reducing computational cost.
- Generating snapshots: The method constructs an offline space by spectrally reducing the snapshot space and retaining eigenvectors associated with the largest eigenvalues.The offline space is built from weighted snapshot matrices and corresponding local eigenvalue problems.
- Computing online multiscale basis functions and their coupling: For each input parameter, GMsFEM computes an online space inside the offline space by selecting eigenvectors corresponding to the largest eigenvalues.The resulting local spaces are coupled globally, for example through a Galerkin formulation, to solve the original problem on a coarse grid.
- Generating snapshots: GMsFEM represents multiscale problems through local snapshot spaces that capture fine-grid solution behavior within coarse regions.Snapshots may use local solutions with selected boundary data, fine-grid basis functions, or local spectral problems.
- Generating snapshots: The source term can be excluded from local solution spaces when it is smooth, with its influence captured through global coarse-grid coupling.The paper also permits modified operators, alternative boundary conditions, and oversampled regions when constructing snapshots.
- Numerical illustration: In numerical tests, eigenvalues decay rapidly, although the decay depends on the choices of a_off, s_off, and the partition-of-unity function χ_0.The example uses unit force terms at specified locations and a high-contrast coefficient in a central subregion.
3 Case studies and relation to existing methods. Discussions and applications
The case studies place existing multiscale methods within the GMsFEM framework and evaluate spectral coarse spaces for high-contrast and parameter-dependent problems. Results show convergence, contrast-robust preconditioning, and the need for enough local or reduced-basis eigenvectors to capture conductivity features.
- Relation to existing methods: Existing multiscale methods can be formulated as special cases of the general GMsFEM framework.The section illustrates this relationship through specific examples and numerical results.
- Offline and online construction: The offline space is constructed from local snapshot spaces through spectral decomposition, and dominant eigenvectors are multiplied by partition of unity functions to form basis functions.In the parameter-free example, the offline space is used directly online, with stiffness matrices precomputed offline.
- Coarse-space design: The choice of partition of unity functions controls the number of large eigenvalues and therefore the dimension of the coarse space.Multiscale initial basis functions reduce errors relative to MsFEM with piecewise linear initial conditions.
- Case with no parameter: For contrast η = 10^6, the error decreases as eigenvectors are added and correlates with the largest omitted eigenvalue, with correlation coefficient 0.99.The reported convergence is robust with respect to contrast, and errors are smaller than those obtained with MsFEM using piecewise linear initial conditions.
- Two-level preconditioning: Spectral coarse spaces yield contrast-independent PCG iterations and estimated condition numbers, while energy-minimizing initial basis functions achieve the smallest coarse-space dimension.The two-level preconditioner uses several coarse-space constructions, including multiscale, energy-minimizing, and spectral spaces.
- Parameter-dependent examples: Parameter-dependent experiments require enough local or reduced-basis eigenvectors to represent conductivity features: Nrb ≥ 4 in one example and at least three realizations in another.Errors converge as the number of local eigenvectors or the reduced-basis size increases, although insufficient offline spaces can slow error decay.
- Applications: The proposed procedure extends to general linear problems such as parabolic and wave equations, with performance depending on the local model reduction encoded in a_off.
- Some generalizations: In a generalized example, increasing the dimension of local spectral spaces reduces errors while the iteration count remains 3, whereas MsFEM with one basis per node has a 47% energy-norm error.The computations use 10 averaged-solution values per block, 8 dominant local eigenvectors, and 80 snapshots per coarse region.
4 Conclusions
GMsFEM provides a general multiscale framework for complex input spaces, constructing offline and online spaces through local snapshots and spectral problems. It supports efficient online basis construction, covers existing multiscale methods, and extends to nonlinear problems.
- GMsFEM extends multiscale finite element methods to PDE problems with parameters, high contrast, and varying right-hand sides or boundary conditions.
- The method constructs local snapshot spaces, then builds an offline space by solving a spectral problem that prioritizes snapshots across the input space.
- For each new parameter and source term, GMsFEM efficiently constructs online multiscale basis functions from the offline space.
- When solutions depend nonlinearly on parameters, online coarse spaces are constructed by solving a spectral problem over the offline space.
- The framework discusses online correction with two-level domain decomposition, demonstrates preconditioner optimality through examples, covers existing multiscale methods, and considers nonlinear extensions.