Physics-informed cluster analysis and a priori efficiency criterion for the construction of local reduced-order bases

Thomas Daniel, Ali Ketata, Fabien Casenave, David Ryckelynck

Abstract

Nonlinear model order reduction has opened the door to parameter optimization and uncertainty quantification in complex physics problems governed by nonlinear equations. In particular, the computational cost of solving these equations can be reduced by means of local reduced-order bases. This article examines the benefits of a physics-informed cluster analysis for the construction of cluster-specific reduced-order bases. We illustrate that the choice of the dissimilarity measure for clustering is fundamental and highly affects the performances of the local reduced-order bases. It is shown that clustering with an angle-based dissimilarity on simulation data efficiently decreases the intra-cluster Kolmogorov $N$-width. Additionally, an a priori efficiency criterion is introduced to assess the relevance of a ROM-net, a methodology for the reduction of nonlinear physics problems introduced in our previous work. This criterion also provides engineers with a very practical method for ROM-nets' hyperparameters calibration under constrained computational costs for the training phase. On five different physics problems, our physics-informed clustering strategy significantly outperforms classic strategies for the construction of local reduced-order bases in terms of projection errors.

Introduction

In computational physics, high-fidelity simulations for complex nonlinear problems—such as fluid dynamics or structural mechanics—often face prohibitive costs in "many-query" contexts like optimization and uncertainty quantification. While Model Order Reduction (MOR) typically addresses these challenges, many "non-reducible" problems involving wave propagation or advection exhibit a slow decay of the Kolmogorov $N$-width. This implies that a single global linear subspace cannot efficiently approximate the solution manifold without becoming excessively large.

To overcome this, local reduced-order bases are employed. By partitioning the solution space into clusters and building specialized bases for each, accuracy can be maintained with lower-dimensional subspaces. However, the effectiveness of this approach hinges on the clustering strategy. Traditional methods often rely on Euclidean distance, which may fail to capture the underlying physical similarities between solutions.

Methodology: Physics-Informed ROM-nets

The authors propose a "ROM-net" architecture, where a dictionary of local models is utilized to approximate solutions across the parameter space.

Workflow of the ROM-net methodology for local reduced-order modeling.

Physics-Informed Sine Dissimilarity

A key innovation is the shift from $L^2$ (Euclidean) distance to a sine dissimilarity measure for clustering. Euclidean clustering often groups solutions based on magnitude rather than "shape." In contrast, the sine dissimilarity focuses on the angle between subspaces spanned by the solutions. This is mathematically linked to the relative projection error, ensuring that solutions within a cluster are well-represented by the same local basis. The authors utilize k-medoids clustering to handle this non-Euclidean metric robustly.

A Priori Efficiency Criterion

Building ROM-nets is computationally expensive. To address this, the paper introduces a "Gain" metric ($G$) to evaluate if a local ROM-net is likely to outperform a single global ROM before committing to high-fidelity data generation. This criterion is calculated using low-fidelity snapshots (e.g., from coarse meshes), allowing for early calibration of hyperparameters such as the number of clusters ($K$) and basis size ($N$).

Experimental Evaluation

The approach was validated on five distinct physics benchmarks:

  • 1D Steady Heat Equation: A non-reducible heterogeneous medium problem.
  • 2D Advection Equation: Testing the limits of wave-like transport.
  • 2D Incompressible Navier-Stokes: Simulating air flow around a circular object.
  • 2D Transient Heat Equation: A localized volumetric heat source problem.
  • 3D Nonlinear Structural Mechanics: A quasistatic rotation of a viscoplastic object.

Key Results and Discussion

The study highlights that "magnitude is a distraction" in many physical systems. In the 2D advection case, Euclidean clustering incorrectly grouped low-magnitude snapshots regardless of their physical direction. The sine-based approach correctly grouped them by advection direction, leading to superior local bases.

Furthermore, the results demonstrate that an improper dissimilarity measure can make a local ROM perform worse than a global one. Locality only provides a benefit when it aligns with the physics of projection. In the 1D heat equation test, the sine-based ROM dictionary achieved a mean projection error of 0.31, significantly lower than the global ROM (0.56) and the Euclidean-based local ROM (0.71).

The "Gain" metric successfully identified hyperparameter "sweet spots," showing that increasing the number of clusters eventually yields diminishing returns due to the increased difficulty of the classification task at runtime.

Conclusion

Physics-informed clustering using sine dissimilarity significantly improves the construction of local reduced-order bases. While the ROM-net approach shows great promise, challenges remain regarding classifier accuracy as the number of clusters increases and scaling to massive 3D industrial datasets. Future work will explore data augmentation for classifiers and more sophisticated temporal handling for time-dependent problems.

Reference: Daniel, T., et al. (2021). Physics-informed cluster analysis and a priori efficiency criterion for the construction of local reduced-order bases. arXiv:2103.13683v2 [math.NA].

Summary updated

Abstract

Nonlinear model order reduction has opened the door to parameter optimization and uncertainty quantification in complex physics problems governed by nonlinear equations. In particular, the computational cost of solving these equations can be reduced by means of local reduced-order bases. This article examines the benefits of a physics-informed cluster analysis for the construction of cluster-specific reduced-order bases. We illustrate that the choice of the dissimilarity measure for clustering is fundamental and highly affects the performances of the local reduced-order bases. It is shown that clustering with an angle-based dissimilarity on simulation data efficiently decreases the intra-cluster Kolmogorov $N$-width. Additionally, an a priori efficiency criterion is introduced to assess the relevance of a ROM-net, a methodology for the reduction of nonlinear physics problems introduced in our previous work in [T. Daniel, F. Casenave, N. Akkari, D. Ryckelynck, Model order reduction assisted by deep neural networks (ROM-net), Advanced Modeling and Simulation in Engineering Sciences 7 (16), 2020]. This criterion also provides engineers with a very practical method for ROM-nets' hyperparameters calibration under constrained computational costs for the training phase. On five different physics problems, our physics-informed clustering strategy significantly outperforms classic strategies for the construction of local reduced-order bases in terms of projection errors.
BibTeX
@article{daniel2022physics,
  author = {Thomas Daniel and Ali Ketata and Fabien Casenave and David Ryckelynck},
  title = {{Physics-informed cluster analysis and a priori efficiency criterion for the construction of local reduced-order bases}},
  journal = {Journal of Computational Physics},
  year = {2022},
  url = {https://openreview.net/forum?id=Iw0aetqDjI},
}
Figure