Variationally Consistent Hamiltonian Model Reduction

Anthony Gruber, Irina TezaurSIAM J. Appl. Dyn. Syst 2025 · DOI

TL;DR

This paper introduces a new method for creating reduced-order models (ROMs) of Hamiltonian systems that preserves their fundamental physical properties, like energy conservation and symplecticity. This "variationally consistent" approach leads to more accurate and stable simulations, especially for complex systems where traditional methods fail.

What Problem Does This Paper Address?

Hamiltonian systems are fundamental for describing conservative physical phenomena, from classical mechanics to fluid dynamics. However, reducing the complexity of these systems for computational efficiency (model reduction) is challenging. Standard projection-based ROMs often break the inherent Hamiltonian structure, leading to unphysical behavior like artificial energy growth or instabilities. Existing methods to preserve this structure either require specialized, difficult-to-construct bases or can still suffer from significant errors. This paper aims to bridge this gap by developing a method that is both variationally consistent and flexible in its choice of reduced basis, offering improved accuracy and stability.

Methods

The core of the paper is a novel approach to constructing variationally consistent Hamiltonian ROMs. Instead of relying on standard Galerkin projection, which can distort the Hamiltonian structure, the authors propose using a Petrov-Galerkin projection. This technique, combined with a specific choice of test functions, ensures that the reduced system remains Hamiltonian.

Key methodological contributions include:

  • Variationally Consistent Petrov-Galerkin Projection: This is the central innovation, allowing for the projection of Hamiltonian systems onto reduced spaces while preserving their structure, regardless of the chosen basis.
  • Intrusive and Nonintrusive ROMs: The framework is developed for both intrusive (requiring access to FOM operators) and nonintrusive (data-driven, OpInf) settings.
  • Error Decomposition: The authors provide interpretable error estimates that decompose the ROM error into two parts: the error from basis projection and the error from deviation from canonicity.
  • Centering Procedure: An existing technique for preserving linear invariants is adapted and extended to the nonintrusive OpInf context.
  • Re-projection: A technique to improve the accuracy of nonintrusive ROMs by ensuring they pre-asymptotically recover intrusive ROM performance.

Datasets

The paper evaluates its methods on:

  • 1D Linear Wave Equation: A standard benchmark problem used to demonstrate the basic behavior of the proposed ROMs and compare them against existing methods.
  • 3D Linear Elasticity Examples:
    • Cantilever Plate: A classical solid mechanics benchmark involving a vibrating rectangular plate.
  • Flexible Bracket: A more complex geometry featuring a circular hole, designed to challenge ROMs with wave interaction and propagation in multiple spatial directions.

These examples are simulated using finite element methods, and snapshot data are collected to build POD bases and train/evaluate the ROMs.

Key Insights and Ideas

  • Variational Consistency is Crucial for Stability and Accuracy: The paper strongly demonstrates that preserving the Hamiltonian structure through variational consistency is not just about theoretical elegance but is essential for obtaining accurate and stable long-term simulations, especially for complex dynamics.
  • Petrov-Galerkin Offers Flexibility: The use of Petrov-Galerkin projection, rather than standard Galerkin, allows for a more flexible choice of reduced basis without sacrificing Hamiltonian structure. This is a significant advantage over methods requiring specially constructed bases.
  • Error Decomposition Provides Insight: The ability to decompose ROM error into basis projection error and deviation-from-canonicity error offers a valuable diagnostic tool for understanding ROM performance and identifying sources of inaccuracy.
  • Nonintrusive Methods Can Match Intrusive Performance: Through techniques like re-projection and centering, the proposed nonintrusive OpInf ROMs can achieve performance comparable to their intrusive counterparts, making them highly practical.
  • Inconsistent Methods Lead to Divergence: The paper clearly shows that even slight deviations from variational consistency (as seen with the "inconsistent H-ROM") can lead to significant errors and divergence, particularly in predictive simulations.

Key Results

  • Improved Accuracy and Stability: The proposed variationally consistent Hamiltonian ROMs consistently outperform traditional methods, especially in time-predictive simulations and for complex geometries. For instance, in the flexible bracket example, the consistent H-ROM solution closely matches the FOM solution, while the inconsistent H-ROM diverges significantly.
  • Energy Conservation: All proposed Hamiltonian ROMs (consistent and inconsistent) conserve energy to a high degree (order 10⁻¹³), a property not shared by the Galerkin ROMs.
  • Basis Flexibility: The method works well with various basis constructions, including standard POD, but shows particular promise when combined with more expressive bases.
  • Nonintrusive ROMs Achieve Comparable Performance: The developed nonintrusive OpInf versions of the variationally consistent ROM achieve accuracy and stability comparable to their intrusive counterparts, especially when using re-projection.

Limitations and Future Work

  • Focus on Linear Systems: While the paper demonstrates excellent results for linear systems, the authors note that extending the framework to nonlinear Hamiltonian systems, quadratic or nonlinear bases, and noncanonical Hamiltonian systems is a crucial area for future work.
  • Lagrangian Counterparts: The paper mentions that canonical Hamiltonian systems have natural Lagrangian counterparts. Future work will involve comparing and contrasting the proposed Hamiltonian ROMs with Lagrangian ROMs derived from second-order equations of motion.
  • Complex Nonlinearities: Extending the centering procedure to nonintrusive ROMs with more complex nonlinearities is identified as an interesting avenue for future research.

Summary updated

BibTeX
@article{gruber2025variationally,
  author = {Anthony Gruber and Irina Tezaur},
  title = {{Variationally Consistent Hamiltonian Model Reduction}},
  journal = {SIAM J. Appl. Dyn. Syst},
  year = {2025},
  doi = {10.1137/24m1652490},
}
Figure