A localized reduced basis approach for unfitted domain methods on parameterized geometries

Basic Information

• Paper ID: 2212.11934
• Title: A localized reduced basis approach for unfitted domain methods on parameterized geometries
• Authors: Margarita Chasapi, Pablo Antolin, Annalisa Buffa (École Polytechnique Fédérale de Lausanne & CNR)
• Classification: math.NA cs.NA
• Published Journal: Computer Methods in Applied Mechanics and Engineering (Accepted)
• Paper Link: https://arxiv.org/abs/2212.11934

Abstract

This paper proposes a reduced-order modeling (ROM) framework for solving parameterized second-order linear elliptic partial differential equations defined on unfitted geometries. The method aims to construct efficient projection-based ROMs relying on reduced basis methods and discrete empirical interpolation techniques. Since geometric parameters in unfitted domain discretization present challenges to standard ROM applications, the paper proposes a methodology based on (i) extension of snapshots on the background mesh and (ii) localization strategies to reduce the number of reduced basis functions. The resulting method is computationally efficient and accurate while remaining agnostic to the underlying discretization choices.

Research Background and Motivation

Core Problems

1. Challenges of unfitted domain methods: In computer-aided design (CAD) and isogeometric analysis (IGA), complex geometries are typically handled through trimming operations, resulting in unfitted meshes where geometry is decoupled from solution discretization.
2. Difficulties in reduced-order modeling of parameterized geometries: When geometric parameters vary, the set of active basis functions changes, causing snapshot solution vectors to potentially have different lengths, which obstructs direct application of standard reduced basis methods.
3. Non-affine parameter dependence: Geometric parameters induce non-affine parameter dependence of differential operators, violating fundamental assumptions of reduced basis methods.

Research Significance

• Industrial relevance: Trimming is a mainstream technique for representing complex shapes in CAD, crucial for achieving unified design-to-analysis frameworks
• Multi-query scenarios: Multiple solutions of parameterized PDEs are required in design optimization, uncertainty quantification, and control applications
• Computational efficiency requirements: Efficient offline/online procedures are needed to achieve computational acceleration

Limitations of Existing Methods

• Standard ROM requires reference domain mappings, highly problem-dependent
• Existing unfitted domain ROM work does not address non-affine dependence, lacking hyper-reduction techniques
• Global reduced basis provides poor approximation of nonlinear manifolds of extended solutions, requiring numerous basis functions

Core Contributions

1. Proposes a complete reduced-order framework for parameterized unfitted geometries, combining hyper-reduction techniques to achieve efficient offline/online decomposition
2. Develops parameter-driven localization strategies, constructing multiple small-dimensional local subspaces via k-means clustering, significantly reducing the number of reduced basis functions
3. Designs a discretization-agnostic universal framework applicable to both finite element and isogeometric methods
4. Implements non-intrusive online coefficient computation based on radial basis function (RBF) interpolation, avoiding intrusive modifications to high-fidelity assembly routines
5. Demonstrates significant computational acceleration in numerical experiments: achieving 17.6× to 130× online computational speedup compared to standard ROM at equivalent accuracy levels

Methodology Details

Problem Formulation

Solving linear elliptic PDEs defined on parameterized unfitted domains Ω(μ):

Lu = f in Ω(μ)

where μ ∈ P ⊂ R^M is the geometric parameter vector, and the domain is defined through cutting operations:

Ω(μ) = Ω̂₀(μ) \ ⋃ᵢ₌₁ᴷ Ω̂ᵢ(μ)

Core Method Architecture

1. Unfitted Domain Discretization

• Background domain setup: Embedding parameterized domain Ω(μ) into a parameter-independent background domain Ω₀
• Discrete space construction: Defining discrete function space V_{h,0} on the background domain, with active space V_h(μ) spanned by basis functions whose support intersects Ω(μ)
• Extension strategy: Extending solution vectors at different parameters to the background domain, setting values to zero in inactive regions

2. Discrete Empirical Interpolation Method (DEIM)

• Affine decomposition: Approximating stiffness matrix and right-hand side vector in affine form:

  Â(μ) ≈ ∑_{q=1}^{Q_a} θ_q^a(μ)Âq
  f̂(μ) ≈ ∑_{q=1}^{Q_f} θ_q^f(μ)f̂q
  

• RBF interpolation: Using radial basis functions to rapidly compute parameter-dependent coefficients online:

  θ_q^a(μ) ≈ ∑_{j=1}^{N_s^d} ω_{q,j}^a φ_{q,j}(‖μ-μⱼ‖₂)
  

3. Localization Strategy

• Parameter space clustering: Partitioning parameter space into N_c subregions using k-means algorithm
• Local basis construction: Constructing DEIM approximations and reduced bases separately for each cluster
• Online switching: Selecting appropriate local basis according to distance from parameter to cluster centers

Technical Innovations

1. Geometry-driven clustering strategy: Parameter clustering based on boundary distance, ensuring similar geometric configurations are grouped
2. Multi-level localization: Simultaneous localization of both DEIM approximations and reduced bases
3. Non-intrusive implementation: Avoiding modifications to underlying assembly routines through RBF interpolation
4. Universal framework design: Decoupled from specific discretization methods, supporting multiple unfitted domain techniques

Experimental Setup

Test Problems

1. Poisson equation:
   • 1D parameterization: Square domain with circular hole, hole center moving along diagonal
   • 2D parameterization: Simultaneous variation of hole center location and radius
2. Linear elasticity problems:
   • Porous quarter cylinder: Multiple circular hole radii parameterization
   • 3D cubic sphere inclusion: Sphere radius parameterization

Discretization Settings

• Basis functions: Quadratic and cubic C¹-C² continuous B-splines
• Meshes: 32×32 Cartesian mesh (2D), 8×8×8 mesh (3D)
• Degrees of freedom: 1225-3000 DOF

Evaluation Metrics

• Accuracy: Relative errors in L² and H¹ norms
• Efficiency: Online CPU time and speedup ratios
• Dimensionality reduction: Singular value decay and number of basis functions

Parameter Settings

• POD tolerance: ε_ = 10⁻⁵, ε_^d = 10⁻⁷
• Training samples: N_s = 100-1000, N_s^d = 250-1000
• Number of clusters: N_c = 4-16, N_c^d = 8-16
• RBF type: Cubic radial basis functions

Experimental Results

Main Results

1. DEIM Approximation Performance

• Global method: Requires 349 matrix terms and 124 vector terms (1D case)
• Local method: Maximum 31-38 matrix terms and 11-15 vector terms (16 clusters)
• Accuracy: Achieves 10⁻⁵ level error

2. Reduced Basis Dimension Comparison

3. Computational Speedup Ratios

Ablation Studies

• Cluster number impact: Increasing cluster numbers significantly reduces basis function count and improves accuracy
• Training sample size: Sufficiently rich training sets (N_s^d ≥ 100) are necessary to ensure DEIM accuracy
• RBF interpolation accuracy: Cubic RBF provides good smoothness without spurious oscillations

3D Verification

In the 3D cubic sphere inclusion problem:

• DEIM basis functions reduced from 180 to 35
• Reduced basis dimension reduced from 35 to 14
• Validates method effectiveness in high-dimensional geometries

Related Work

Isogeometric Reduced-Order Modeling

• Existing work primarily focuses on combining IGA-ROM for shape-preserving geometries
• Successful applications in fluid mechanics, parabolic problems, and cardiac electrophysiology
• Lacks systematic investigation of unfitted domains

Unfitted Domain Reduced-Order Modeling

• Early work: Virtual domain method + PGD, low-rank approximation for time-varying interfaces
• Recent progress: CutFEM-ROM, Shifted Boundary Method-ROM
• Limitations: Lacks hyper-reduction techniques, insufficient treatment of non-affine dependence

Local Reduced Basis Methods

• Clustering methods: Local ROM based on solution snapshot clustering
• hp-reduced basis: Adaptive parameter space partitioning
• This paper's contribution: First parameter-driven localization for unfitted geometries

Conclusions and Discussion

Main Conclusions

1. Effectiveness validation: The proposed method significantly reduces online computational cost while maintaining accuracy
2. Universality: Framework is discretization-agnostic, applicable to multiple unfitted domain techniques
3. Scalability: Successfully extends to 3D complex geometries, demonstrating good scalability

Limitations

1. Cluster number selection: Requires prior selection of cluster numbers; while variance criteria are provided, empirical judgment remains necessary
2. Parameter dimensionality: High-dimensional parameter spaces may lead to ill-conditioning of RBF interpolation
3. Geometric complexity: Extremely complex geometric variations may require additional clusters

Future Directions

1. Greedy algorithms: Applying greedy algorithms for local reduced basis construction and error certification
2. Higher-order PDEs: Extension to more complex PDE problems
3. Optimization applications: Applications in design, shape, and topology optimization

In-Depth Evaluation

Strengths

1. Strong methodological innovation: First systematic application of localization strategies to unfitted domain ROM, addressing important technical challenges
2. Complete theoretical framework: Forms comprehensive methodology from snapshot extension to localization strategies
3. Comprehensive experimental validation: Covers 2D/3D, different PDE types, and multiple parameterization scenarios
4. High practical value: Significant computational acceleration and discretization-agnostic characteristics provide broad application prospects

Weaknesses

1. Limited theoretical analysis: Lacks rigorous theoretical analysis of convergence and error estimates
2. Parameter dependency: Method effectiveness highly depends on clustering strategy and parameter choices
3. Complexity trade-offs: While online efficiency is high, offline computational and storage costs increase significantly

Impact

1. Academic contribution: Opens new research directions for unfitted domain ROM, expected to inspire subsequent research
2. Engineering value: Important for CAD-analysis integration and industrial geometry optimization
3. Reproducibility: Provides detailed algorithm descriptions and parameter settings, facilitating reproduction and application

Applicable Scenarios

• Parameterized design optimization involving complex geometric variations
• Multi-query analysis of trimmed geometries
• Real-time simulation and control applications
• Geometric parameter studies in uncertainty quantification

References

The paper cites abundant relevant literature, including:

• Classical work on unfitted domain methods (Peskin, Hughes, etc.)
• Foundational literature on isogeometric analysis
• Theoretical foundations of reduced basis methods and empirical interpolation
• Cutting-edge research on local ROM

This paper provides important methodological contributions to reduced-order modeling of parameterized PDEs on unfitted domains, achieving good balance between theoretical innovation and practical application.