Nonlinear energy-preserving model reduction with lifting transformations that quadratize the energy

Harsh Sharma, Juan Diego Draxl Giannoni, Boris Kramer · 2025

Read Nonlinear energy-preserving model reduction with lifting transformations that quadratize the energy by Harsh Sharma, Juan Diego Draxl Giannoni, Boris Kramer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-02273.txt.

Introduction

Simulating waves or fields on a fine grid can take hours, so engineers shrink the model, but a shrunk model that leaks fake energy gives wrong answers over time. Harsh Sharma, Juan Diego Draxl Giannoni, and Boris Kramer present a shrink method in a 2025 paper that keeps the books balanced. Their trick adds helper parts that turn messy energy terms into neat squared ones.

Once the energy is in squared form, the shrunk model can hold it exactly, even on the full grid. The team tried the method on four wave-style systems, from a one-line wave to coupled field pairs. Results matched the best rival method in speed and accuracy during runs, while the setup stage ran much faster. For design shops that rerun models often, the saving lands where it counts.

Catalog record

CREATORHarsh Sharma, Juan Diego Draxl Giannoni, Boris Kramer
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE114 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-02273

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Nonlinear energy-preserving model reduction with lifting transformations that quadratize the energy Harsh Sharma, Juan Diego Draxl Giannoni, Boris Kramer Abstract Existing model reduction techniques for high-dimensional models of conservative partial differential equations (PDEs) encounter computational bottlenecks when dealing with systems featuring non-polynomial nonlinearities. This work presents a nonlinear model reduction method that employs lifting variable transformations to derive structure-preserving quadratic reduced-order models for conservative PDEs with general nonlinearities. We present an energy-quadratization strategy that defines the auxiliary variable in terms of the nonlinear term in the energy expression to derive an equivalent quadratic lifted system with quadratic system energy. The proposed strategy combined with proper orthogonal decomposition model reduction yields quadratic reduced-order models that conserve the quadratized lifted energy exactly in high dimensions. We demonstrate the proposed model reduction approach on four nonlinear conservative PDEs: the one-dimensional wave equation with exponential nonlinearity, the two-dimensional sine-Gordon equation, the two-dimensional Klein-Gordon equation with parametric dependence, and the two-dimensional Klein-Gordon-Zakharov equations. The numerical results show that the proposed lifting approach is competitive with the state-of-the-art structure-preserving hyper-reduction method in terms of both accuracy and computational efficiency in the online stage while providing significant computational gains in the offline stage. 1 Introduction High-dimensional nonlinear full-order models (FOMs) of conservative PDEs appear in a wide variety of science and engineering areas ranging from climate modeling to plasma physics. These nonlinear FOMs are typically derived via structure-preserving methods [ 1 ] that discretize conservative PDEs such that the resulting space-discretized models satisfy nonlinear conservation laws. To meet the rapidly increasing demand for accurate numerical simulations of complex physical systems, the field of projection-based model reduction [ 2 ] has developed principled techniques to derive computationally efficient reduced-order models (ROMs) via the projection of FOM operators on low-dimensional subspaces. A major challenge in the context of model reduction of conservative nonlinear FOMs is preserving the qualitative properties of the dynamics in the reduced setting, especially the energy conservation law, as it plays a crucial role in characterizing the nonlinear dynamics. Structure-preserving model reduction was first explored in the context of Lagrangian mechanical systems [ 3 ] where the authors derived nonlinear Lagrangian ROMs that conserved the nonlinear FOM energy. For nonlinear FOMs of interconnected systems modeled using the port-Hamiltonian framework, structure-preserving model reduction via preservation of the underlying Dirac structure has been presented in, e.g., [ 4 , 5 , 6 , 7 ] . In the context of the Hamiltonian formulation of conservative FOMs, symplectic model reduction of Hamiltonian systems was introduced in [ 8 ] where the authors derived Hamiltonian ROMs by projecting the Hamiltonian FOM operators onto a symplectic subspace obtained via proper symplectic decomposition. Building on this work, reduced basis methods for structure-preserving model reduction of parametric Hamiltonian systems and Poisson systems have been presented in [ 9 ] and [ 10 ] , respectively. A more general symplectic model reduction approach with non-orthonormal bases was proposed in [ 11 ] . In a similar direction, the authors in [ 12 ] modified the standard POD-Galerkin approach such that the Hamiltonian structure is preserved after the Galerkin projection step. The above methods laid the foundation for recent developments in structure-preserving model reduction for Hamiltonian systems and more general gradient systems in both intrusive [ 13 , 14 , 15 , 16 , 17 , 18 ] and nonintrusive settings [ 19 , 20 , 21 , 22 , 23 , 24 ] . However, all of the aforementioned structure-preserving approaches suffer from computational efficiency issues as the evaluation of the nonlinear components of the structure-preserving ROM vector field still scales with the FOM dimension. This computational bottleneck is a well-known issue in the nonlinear model reduction community and has led to the development of hyper-reduction methods. These methods introduce a second level of approximation to reduce the computational cost involved in the evaluation of the nonlinear terms. In particular, the discrete empirical interpolation method (DEIM) [ 25 , 26 , 27 ] has been shown to be effective for nonlinear model reduction over a range of applications. However, the standard hyper-reduction approaches for general nonlinear systems do not preserve the underlying Hamiltonian structure. While the subfield of hyper-reduction methods has grown considerably in the last decade, progress on structure-preserving hyper-reduction has been less rapid. The authors in [ 28 ] presented a cubature approach, the energy-conserving sampling and weighting method, where the vector field obtained from the nonlinear conservative FOM is approximated with a weighted average of the nonlinear vector field components on a coarser mesh. However, this approach is limited to nonlinear conservative FOMs obtained via finite element discretization of conservative PDEs. In [ 29 ] , the authors presented a DEIM method for nonlinear mechanical oscillators in structural dynamics where the trajectory does not deviate drastically from equilibrium. The underlying assumption about trajectories being localized around an equilibrium point, however, is generally not met by nonlinear conservative FOMs considered herein. A modification of the DEIM method for nonlinear port-Hamiltonian systems has been presented in [ 30 ] where the nonlinear gradient of the Hamiltonian function is approximated in the space where the DEIM projection is orthogonal. Even though this method can be adapted to the Hamiltonian setting to derive nonlinear ROMs that preserve the underlying Hamiltonian structure, the paper does not provide state space error bounds for the resulting ROM. For dynamical systems with a first integral, an energy-conserving hyper-reduction approach via preservation of the skew-symmetric structure has been proposed in [ 31 ] but the cost of deriving the ROM operators for this method is computationally prohibitive as it scales with the FOM dimension. In [ 32 ] , the authors presented an energy-conserving hyper-reduction method for energy-and momentum-conserving ROMs of incompressible Navier-Stokes equations. However, the problem-specific formulation of this method does not generalize to many nonlinear conservative PDEs. The authors in [ 33 ] obtained a structure-preserving DEIM approximation for general nonlinear Hamiltonian systems by decomposing the Hamiltonian into the Euclidean product of the nonlinear vector field with a constant vector. Building on this idea, the authors in [ 34 ] developed a gradient-preserving hyper-reduction approach that guarantees the preservation of the FOM Hamiltonian asymptotically. Moreover, this work also provided state space error bounds by extending the a priori error estimate result for the standard POD-DEIM in [ 35 ] to the Hamiltonian setting. This structure-preserving DEIM (spDEIM) method, however, requires a computationally demanding offline phase for parametric problems. Projection-based nonlinear model reduction via lifting [ 36 , 37 , 38 , 39 , 40 ] has emerged as a promising research direction. These methods first introduce auxiliary variables to transform a general nonlinear system into an equivalent finite-dimensional lifted system with quadratic dynamics and then project the quadratic operators of the high-dimensional lifted system onto a reduced space to obtain a quadratic ROM. The key advantage of quadratizing (polynomializing) the dynamics before nonlinear model reduction is that it eliminates the need for additional hyper-reduction/interpolation techniques. This lifting approach has also been extended to the nonintrusive setting in [ 41 , 42 , 43 , 44 , 45 ] where lifting transformations are exploited to learn low-order polynomial ROMs of complex nonlinear FOMs directly from the lifted data. While

Read the complete text