Optimal piecewise linear data compression for solutions of parametrized partial differential equations

Thomas Daniel, Fabien Casenave, Nissrine Akkari, David RyckelynckarXiv 2021 · arXiv

Abstract

Standard Model Order Reduction (MOR) techniques often struggle with non-reducible manifolds in parametrized partial differential equations (pPDEs). This paper establishes a mathematically rigorous framework for partitioning complex physics simulations into local neighborhoods that are more amenable to linear compression. By introducing a "sine dissimilarity" measure, the authors prove that optimal local Reduced-Order Models (ROMs) can be constructed using k-medoids clustering, providing a theoretical bridge between geometric data analysis and high-fidelity engineering simulations.

Introduction

In modern engineering, simulating complex phenomena—such as fluid dynamics or structural mechanics—requires solving High-Dimensional Models (HDMs). While these models offer high accuracy, their computational cost is often prohibitive for real-time applications or large-scale optimization. Model Order Reduction (MOR) aims to alleviate this by projecting the problem onto a low-dimensional subspace.

However, many real-world physical systems are "non-reducible," meaning their solution sets form curved, complex manifolds that cannot be accurately captured by a single, flat linear subspace. To address this, researchers have turned to piecewise linear approximations, which break the manifold into smaller, more manageable regions. This paper addresses the fundamental question: How can we partition these manifolds in a way that is mathematically optimal for accuracy?

Methodology

Normalized Kolmogorov Width

To evaluate the quality of a local subspace, the authors utilize a "normalized" version of the Kolmogorov width. Unlike standard error metrics that focus on absolute differences, this normalized approach measures how well a subspace captures the relative features of a solution. This ensures the compression is invariant to the scale or intensity of the solution, focusing instead on the underlying physical patterns.

Sine Dissimilarity

A core contribution of this work is the introduction of a distance metric based on the sine of the angle between solution vectors. Standard Euclidean distances often fail in high-dimensional physics because they may group two solutions simply because they are both low in magnitude, even if their physical patterns are entirely different. Sine dissimilarity ignores magnitude and groups solutions based on their "shape" or direction in the high-dimensional space.

Optimal Partitioning and Clustering

The authors provide a rigorous proof that for a one-dimensional approximation (N=1), the problem of finding the optimal partition of data is equivalent to performing k-medoids clustering using the sine dissimilarity measure. This finding transforms a complex optimization problem into a tractable algorithmic task.

For higher-dimensional local models (N > 1), the paper proposes a practical two-step heuristic:

  1. Cluster the data using the sine-based metric.
  2. Construct local bases for each cluster using standard methods like Proper Orthogonal Decomposition (POD).

Theoretical Insights

The Geometry of Subspaces

The research treats the search for an ideal ROM as a geometric problem on a Grassmannian—the manifold of all possible linear subspaces. By viewing the problem through this lens, the authors demonstrate that magnitude is often a "distraction" in physics problems. For instance, in thermal simulations, the intensity of heat may change, but the spatial pattern of the heat distribution often remains consistent. Sine dissimilarity captures this consistency where traditional L2 metrics would fail.

Symmetry and Robustness

The authors demonstrate that sine dissimilarity is a symmetric pseudometric. This property makes it significantly more robust than standard distances when comparing disparate physical states, ensuring that the resulting local models are stable and reliable across different regions of the parameter space.

Applications: High-Pressure Turbine Blades

While primarily theoretical, the framework is validated against the mechanical behavior of high-pressure turbine blades in aircraft engines. These components are subjected to extreme and uncertain thermal loads. The paper provides the theoretical foundation for why local ROM dictionaries—built using the proposed clustering method—outperform global models in capturing the stress and deformation of these critical parts.

Conclusion and Future Work

This paper successfully links optimal data partitioning to representative-based clustering, providing a rigorous justification for local MOR techniques. While the proof of optimality is currently limited to 1D local models, the heuristic for higher dimensions shows significant promise for industrial applications.

Future research will focus on developing a posteriori error bounds. Such bounds would allow the algorithm to refine partitions dynamically during a simulation, moving away from the current reliance on pre-computed "snapshots" and toward a more adaptive, autonomous reduction framework.

Summary updated

Abstract

Model order reduction has been extensively studied over the last two decades. Projection-based methods such as the Proper Orthogonal Decomposition and the Reduced Basis Method enjoy the important advantages of Galerkin methods in the derivation of the reduced problem, but are limited to linear data compression for which the reduced solution is sought as a linear combination of spatial modes. Nonlinear data compression must be used when the solution manifold is not embedded in a low-dimensional subspace. Early methods involve piecewise linear data compression, by constructing a dictionary of reduced-order models tailored to a partition of the solution manifold. In this work, we introduce the concept of optimal partition of the solution manifold in terms of normalized Kolmogorov widths, and prove that the optimal partitions can be found by means of a representative-based clustering algorithm using the sine dissimilarity measure on the solution manifold.
BibTeX
@misc{daniel2021optimal,
  author = {Thomas Daniel and Fabien Casenave and Nissrine Akkari and David Ryckelynck},
  title = {{Optimal piecewise linear data compression for solutions of parametrized partial differential equations}},
  year = {2021},
  eprint = {2108.12291},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.12291},
}
Figure