Reduced order modelling of nonaffine problems on parameterized NURBS multipatch geometries

Basic Information

• Paper ID: 2211.07348
• Title: Reduced order modelling of nonaffine problems on parameterized NURBS multipatch geometries
• Authors: Margarita Chasapi, Pablo Antolin, Annalisa Buffa
• Classification: math.NA cs.NA
• Submission Date: November 14, 2022 (arXiv submission)
• Paper Link: https://arxiv.org/abs/2211.07348

Abstract

This paper explores the combined application of Reduced Basis Methods (RBM) and Isogeometric Analysis (IGA) in solving parameterized partial differential equations. The introduction of isogeometric analysis enables a unified simulation framework for design and analysis based on a single geometric representation. Particularly for complex geometries modeled by multiple patches, this work focuses on nonaffine problems characterized by high-dimensional parametric spaces. The Empirical Interpolation Method (EIM) is employed to recover affine parametric dependence, combined with domain decomposition for dimensionality reduction. Spline patches are coupled in the parametric setting using the Static Condensation Reduced Basis Element (SCRBE) method. Static condensation procedures are applied at common interfaces between adjacent patches, while reduced basis approximations are implemented in the interior to achieve efficient offline/online decomposition.

Research Background and Motivation

Problem Definition

1. Core Problem: When solving partial differential equations on complex parameterized geometries, traditional full-order models incur prohibitive computational costs, particularly in multi-query and real-time application scenarios.
2. Technical Challenges:
   • Nonaffine parametric dependence induced by geometric parameterization undermines the efficiency of reduced basis methods
   • High-dimensional parametric spaces degrade the performance of standard reduced basis approaches
   • The complexity of multipatch geometries increases modeling difficulty
3. Application Requirements: Complex geometries in industrial applications typically require multipatch representations and involve numerous geometric parameters, necessitating efficient parameterized solution methods.

Research Motivation

• Unified Framework: Leveraging IGA to achieve unified representation of geometric design and numerical analysis
• Computational Efficiency: Implementing fast solution of parameterized problems through reduced basis methods
• Industrial Relevance: Addressing complex multipatch geometric problems with industrial significance

Core Contributions

1. Method Integration: Systematically combining IGA with reduced basis methods for nonaffine multipatch geometric problems for the first time
2. EIM-IGA Integration: Reformulating the empirical interpolation method within the spline approximation framework, utilizing exact geometric mappings to recover affine decomposition
3. SCRBE Extension: Extending the static condensation reduced basis element method to NURBS multipatch geometries
4. Domain Decomposition Strategy: Proposing a patch-based domain decomposition approach for handling high-dimensional parametric spaces
5. Numerical Validation: Verifying method effectiveness on three-dimensional multipatch geometries

Methodology Details

Problem Formulation

Solving parameterized elliptic partial differential equations: $-\Delta u = f \text{ in } \Omega(\mu)$$u = 0 \text{ on } \Gamma_D(\mu)$$\frac{\partial u}{\partial n} = 0 \text{ on } \Gamma_N(\mu)$

where $\Omega(\mu) \subset \mathbb{R}^d$ is the parameterized domain and $\mu \in \mathcal{P} \subset \mathbb{R}^P$ represents geometric parameters.

Model Architecture

1. NURBS Multipatch Geometric Representation

• Parametric Domain: $\hat{\Omega} = [0,1]^{\hat{d}}$
• Geometric Mapping: $F(\xi;\mu) = \sum_i B_{i,p}(\xi)P_i(\mu)$
• Multipatch Decomposition: $\Omega(\mu) = \bigcup_{k=1}^{n_p} \Omega_k(\mu^{(k)})$

2. Empirical Interpolation Method (EIM)

Converting nonaffine parametric dependence into an interpolation problem on the reference domain:

Transformed Weak Form: $\alpha(u_h, v_h;\mu) = \int_{\hat{\Omega}} (\nabla \hat{u}_h DF^{-1}) \cdot (\nabla \hat{v}_h DF^{-1})|\det DF| d\hat{\Omega}$

EIM Approximation: $g_M^\alpha(\xi;\mu) = I_{\xi}^M g^\alpha(\xi;\mu) = \sum_{m=1}^{M^\alpha} \theta_m^\alpha(\mu)\phi_m^\alpha(\xi)$

3. SCRBE Method

Static Condensation Decomposition:

• Bubble Space: $V_{h,b}^k = \{v_h \in V_h^k : v_h|_{\gamma_{k,j}} = 0\}$
• Port Space: $V_{h,\gamma}^{k,j} = \{v_h \in V_h^k : v_h|_{\gamma_{k,j}} \neq 0\}$

Solution Representation: $u_h(\mu)|_{\Omega_k} = b_{h,f}^k(\mu^{(k)}) + \sum_{j=1}^{n_\gamma^k} \sum_{r=1}^{N_{\gamma}^{k,j}} \hat{u}_{G_k(j),r}(\mu) \phi_{h}^{k,j,r}(\mu^{(k)})$

Technical Innovations

1. Geometric Mapping Utilization: Directly leveraging NURBS exact geometric mappings for variable transformation, avoiding complex mapping constructions in traditional finite elements
2. Hierarchical Dimensionality Reduction:
   • Interior of patches: Reduced basis approximation of bubble functions
   • Patch interfaces: POD-constructed port modes
3. Affinity Recovery: Recovering affine decomposition of geometric parameterization through EIM interpolation on the parametric domain

Experimental Setup

Geometric Model

• Domain: Three-dimensional curved multipatch geometry
• Parameters: 8 geometric parameters $\mu = (\mu_1, \mu_2, ..., \mu_8)$
• Patch Distribution: 4 patches, 2 parameters per patch
• Mesh: 64 elements per patch, 1225 total degrees of freedom
• Polynomial Order: $p = 3$

Boundary Conditions

• Dirichlet Boundary: $u = 0$ on $\Gamma_D$
• Neumann Boundary: $\frac{\partial u}{\partial n} = 0$ on $\Gamma_N$
• Source Term: $\tilde{f} = 2xyz$

Training Parameters

• EIM Training Set: $|P_{train}^{EIM}| = 250$
• RB Training Set: $|P_{train}| = 250$
• POD Snapshots: $N_s = 25$ (Latin hypercube sampling)
• Tolerances: EIM $\varepsilon = 10^{-7}$, RB $\varepsilon = 10^{-5}$

Experimental Results

EIM Approximation Performance

• EIM basis functions for stiffness matrices and right-hand sides are relatively few ($M^\alpha \leq 33, M^f \leq 14$)
• Rapid convergence at $10^{-7}$ accuracy

Reduced Basis Performance

Port Mode Analysis

• Rapid Singular Value Decay: POD singular values decay rapidly at all interfaces
• Port Dimensions: Each interface reduced to ≤25 port modes
• Interface RB: Average of 16 basis functions per port mode

Computational Efficiency

• Average Online Evaluation: 42ms per patch
• Port Mode Online Time: 29.3ms
• Overall Acceleration: Significant improvement compared to full-order models

Related Work

IGA-RB Combined Research

1. Fluid Problems: Manzoni et al. 2015 application to NASA airfoil flow
2. Certification Methods: Devaud and Rozza 2017 affine parametric IGA-RB methods
3. Parabolic Problems: Zhu et al. 2017 combining POD for time-dependent problems
4. Non-intrusive Methods: Reduced order modeling based on interpolation and deep learning

Domain Decomposition Methods

1. SCRBE Foundation: Huynh et al. 2013 static condensation reduced basis element method
2. Port Dimension Reduction: Eftang and Patera 2013 port dimension reduction techniques
3. Component Methods: Component-oriented approaches for industrial-scale problems

EIM Development

1. Foundational Theory: Barrault et al. 2004 empirical interpolation method
2. Error Estimation: Development of a posteriori error bounds
3. Geometric Applications: Extension of applications in parameterized geometries

Conclusions and Discussion

Main Conclusions

1. Method Feasibility: Successfully combining IGA with RB methods for nonaffine multipatch geometric problems
2. Computational Efficiency: Achieving significant dimensionality reduction and acceleration
3. Accuracy Preservation: Maintaining good approximation accuracy with relatively small reduced basis dimensions
4. Extensibility: The method framework has potential for extension to more complex industrial problems

Limitations

1. Problem Complexity: Currently validated only on relatively simple model problems
2. Parameter Dimensionality: While handling 8-dimensional parameters, performance on higher-dimensional parametric spaces requires further investigation
3. Geometric Complexity: Applicability to more complex industrial geometries needs verification
4. Nonlinear Extension: Current framework primarily addresses linear problems

Future Directions

1. Industrial Applications: Extension to complex geometries with industrial relevance
2. Parameter Optimization: Integration into parameter optimization processes
3. Nonlinear Problems: Extension to nonlinear partial differential equations
4. Parallel Computing: Development of more efficient parallel implementations

In-Depth Evaluation

Strengths

1. Theoretical Completeness: Systematically establishing a theoretical framework for IGA-RB-EIM
2. Method Innovation: Cleverly utilizing NURBS geometric mappings to simplify EIM implementation
3. Practical Value: Providing an effective approach for parameterized solution of complex geometries
4. Sufficient Validation: Numerical experiments verify the effectiveness of each component

Weaknesses

1. Application Scope: Validation cases are relatively simple, lacking testing on complex industrial problems
2. Theoretical Analysis: Lacking rigorous theoretical analysis of convergence and error bounds
3. Comparative Study: Insufficient systematic comparison with other reduced order methods
4. Computational Complexity: Detailed computational complexity analysis not provided

Impact

1. Academic Contribution: Providing important theoretical foundation for combining IGA and RB methods
2. Practical Value: Significant implications for CAD/CAE integration
3. Reproducibility: Using open-source software packages with good reproducibility
4. Extension Potential: Providing solid methodological foundation for subsequent research

Applicable Scenarios

1. Geometric Design Optimization: CAD parameter-driven design optimization
2. Multi-query Simulation: Engineering problems requiring numerous parameter evaluations
3. Real-time Simulation: Applications with strict computational time requirements
4. Shape Optimization: Structure optimization based on geometric parameters

References

This paper cites important literature from reduced basis methods, isogeometric analysis, and domain decomposition, including:

• Hughes et al. 2005: Foundational IGA theory
• Quarteroni et al. 2016: Reduced basis methods survey
• Barrault et al. 2004: Empirical interpolation method
• Huynh et al. 2013: SCRBE method foundation

Overall Assessment: This is a paper of significant importance in computational mathematics, systematically addressing efficient solution of complex geometric parameterized problems. The method has solid theoretical foundations and excellent practical prospects, opening new research directions for combining IGA and RB methods.