Nonlinear model reduction on metric spaces. Application to one-dimensional conservative PDEs in Wasserstein spaces

Virginie Ehrlacher, Damiano Lombardi, Olga Mula, François-Xavier VialardarXiv 2019 · OpenReview

Introduction

Model Order Reduction (MOR) is a critical tool in computational science, designed to approximate the solutions of complex, parameterized Partial Differential Equations (PDEs) with minimal computational overhead. By reducing the dimensionality of the problem, MOR enables real-time predictions and efficient uncertainty quantification in engineering and control systems. Traditionally, these methods rely on identifying an optimal linear subspace—often through Principal Component Analysis (PCA) or Reduced Basis Methods—to represent the solution manifold.

However, linear MOR techniques encounter a fundamental bottleneck when applied to hyperbolic or transport-dominated problems, such as wave propagation or fluid dynamics with moving shocks. In these scenarios, the "Kolmogorov width" of the solution set—a measure of how well the set can be approximated by a linear space of a certain dimension—decays very slowly. Because the location of a shock or wave front changes based on system parameters, a linear combination of static basis vectors fails to capture the dynamics efficiently, leading to high-dimensional requirements or poor accuracy. This paper proposes a paradigm shift: performing model reduction within nonlinear metric spaces rather than standard vector spaces.

Methodology: MOR in the Wasserstein Space

For one-dimensional conservative PDEs where mass is preserved, the authors leverage the $L^2$-Wasserstein space. This metric space is uniquely suited for transport problems because it treats the solution as a probability measure, where the distance between two states is defined by the optimal cost of moving "mass" from one configuration to another. The researchers introduce two distinct, data-driven nonlinear approaches:

1. Tangent PCA (tPCA)

This method adapts classical PCA to the geometry of Riemannian manifolds. The process involves:

  • Logarithmic Mapping: Projecting the snapshot data from the nonlinear manifold into a flat "tangent space."
  • Dimensionality Reduction: Performing standard PCA within this tangent space to identify the principal components of the transformed data.
  • Exponential Mapping: Mapping the reduced representations back to the original manifold for reconstruction.

To ensure efficiency during the online phase, the PCA coefficients are interpolated based on the system parameters, allowing for rapid emulation without re-solving the underlying physics.

2. Barycentric Greedy Algorithm (gBar)

Recognizing that tPCA can occasionally suffer from numerical instability—where projections in the tangent space fall outside the domain of the exponential map—the authors developed the Barycentric Greedy algorithm. This approach does not require a Riemannian structure. Instead, it uses a greedy selection process to identify a small set of representative "basis" snapshots. New solutions are then approximated as a Wasserstein barycenter—a generalized weighted average in the metric space—of these snapshots. This method provides inherent stability, ensuring the reconstructed solution remains a valid physical measure.

Experimental Results and Insights

The framework was tested against four benchmark 1D conservative PDEs: the Inviscid and Viscous Burgers' equations, the Camassa-Holm equation, and the Korteweg-de Vries (KdV) equation. For each, the models were trained on 5,000 snapshots and validated against a 500-snapshot test set.

Accuracy and Visual Fidelity

Both tPCA and gBar demonstrated significant improvements over standard linear PCA. In the case of the Burgers' and Camassa-Holm equations, the metric space methods achieved average errors several orders of magnitude lower than linear approaches when measured in the transport-appropriate $H^{-1}$ norm. Visually, while standard PCA produced highly oscillatory and non-physical artifacts near sharp wave fronts, the Wasserstein-based methods preserved the sharp edges and physical integrity of the waves.

Computational Efficiency

The proposed methods act as fast emulators, bypassing the need to solve reduced versions of the physics equations online. For the Viscous Burgers' equation, the reduced models achieved a median online speedup of approximately 500x compared to high-fidelity PDE solvers. This speedup is primarily due to the existence of explicit, computationally cheap formulas for the exponential map in 1D Wasserstein space.

Limitations and Future Directions

Despite the success in transport-dominated scenarios, the study identified specific challenges. The methods struggled with the KdV equation during events where waves (peakons) merge or separate. The Wasserstein transformation is optimized for translation but is less effective at handling changes in the topology of the mass distribution.

Furthermore, while the 1D implementation is highly efficient, extending these methods to 2D or 3D domains presents a computational challenge, as the exponential map and barycenter calculations become significantly more complex. Future research will explore the use of the Hellinger-Kantorovich distance to handle mass creation and destruction, as well as the integration of machine learning to learn optimal metric mappings directly from high-dimensional data.

Summary updated

BibTeX
@misc{ehrlacher2019nonlinear,
  author = {Virginie Ehrlacher and Damiano Lombardi and Olga Mula and François-Xavier Vialard},
  title = {{Nonlinear model reduction on metric spaces. Application to one-dimensional conservative PDEs in Wasserstein spaces}},
  year = {2019},
  url = {https://openreview.net/forum?id=70ESvZY3Hm},
}
Figure