Structure-preserving model reduction on manifolds of port-Hamiltonian systems

This paper proposes a non-intrusive, structure-preserving model order reduction method based on generalized manifold Galerkin reduction that generates reduced-order port-Hamiltonian models for both linear and nonlinear systems while maintaining stability and passivity, demonstrating superior accuracy compared to existing techniques.

The Gist

Imagine you have a massive, incredibly detailed blueprint of a complex machine, like a high-speed train or a power grid. This blueprint (the Full-Order Model) is perfect; it captures every tiny gear, spring, and electrical circuit. However, if you try to run a simulation of this machine on a computer, it takes forever. It's like trying to simulate the movement of every single grain of sand on a beach just to see how a wave crashes. You need a way to make the blueprint smaller and faster to work with, without losing the essential "soul" of the machine.

This is the problem of Model Order Reduction (MOR).

The Challenge: Keeping the "Soul" Intact

The machine in this paper is a Port-Hamiltonian (pH) system. Think of this as a machine built on the laws of energy. It has a special "energy bank" (the Hamiltonian) that tracks how much energy is stored, how much is lost to friction (dissipation), and how much comes in or goes out through its ports (inputs and outputs).

The most important rule of these machines is Conservation of Energy. If you build a smaller, simplified version of the machine (a Reduced-Order Model or ROM), it must still obey the laws of energy. If your simplified model creates energy out of thin air or loses it magically, it will become unstable and give you wrong answers. It's like shrinking a car engine: if you shrink the fuel tank but keep the pistons the same size, the engine will explode.

The Old Way vs. The New Way

The Old Way:
Previous methods for shrinking these models were like taking a photo of the machine and then squishing the photo to make it smaller. While the picture looks similar, the physics inside might be broken. The "energy rules" often get lost in the squishing process, leading to unstable simulations.

The New Way (This Paper):
The authors, Silke Glas and Hongliang Mu, propose a smarter way to shrink the machine. They use a technique called Generalized Manifold Galerkin (GMG) reduction.

Here is the analogy:
Imagine the machine's behavior isn't just a flat sheet of paper, but a curved, flexible surface (a manifold) floating in a huge room. The machine's state (where it is and how fast it's moving) always stays on this specific curved surface.

• The Map (Embedding): Instead of just squishing the paper, they create a special "map" that projects the huge, complex 3D room down onto a smaller, simpler 2D surface that perfectly matches the curve of the machine's behavior.
• The Projection: They don't just guess the new path; they use a mathematical "laser" (the GMG reduction) that projects the machine's forces directly onto this new, smaller surface.
• The Result: Because they projected the forces along the curve of the energy surface, the new, smaller machine automatically obeys the laws of energy. It's like shrinking the engine but keeping the fuel tank and pistons perfectly synchronized because you shrank them together along the same curve.

Two Types of Maps

The paper tests two ways to draw this "map":

1. The Straight Map (Linear): This assumes the machine's behavior is like a flat sheet. It's good for simple machines (like a standard spring and damper).
2. The Curved Map (Quadratic): This assumes the machine's behavior is curved, like a bowl or a saddle. This is crucial for complex, non-linear machines (like a spring that gets stiffer the more you stretch it). The authors found that using this "curved map" gave much more accurate results for complex systems.

The "DEIM" Trick

There's a catch: even with a smaller model, calculating the energy of a complex, non-linear spring can still be slow because it depends on the size of the original giant blueprint.

To fix this, they use a trick called DEIM (Discrete Empirical Interpolation Method).

• Analogy: Imagine you want to know the average temperature of a whole city. Instead of checking every single house (which takes forever), you pick a few "smart" houses that represent the whole city perfectly. You only check those few spots to get a very accurate estimate of the whole city's temperature. DEIM does this for the math, picking the most important parts of the calculation to keep the simulation fast.

The Results

The authors tested their new method on two types of machines:

1. A Linear Machine: A row of masses connected by simple springs and dampers (like a train of toy cars).
2. A Non-Linear Machine: The same setup, but the springs get stiffer the more you stretch them (like a real rubber band).

The Outcome:
Their new method (GMG-POD and GMG-QM) produced smaller models that were:

• Faster: They ran simulations much quicker.
• More Accurate: They made fewer mistakes in predicting the machine's movement compared to older methods.
• Stable: They perfectly preserved the energy rules, meaning the models didn't "blow up" or behave strangely.

In Summary

This paper introduces a new, smarter way to shrink complex energy-based machines. Instead of just cutting corners, they use advanced geometry to fold the machine down into a smaller size while keeping its "energy soul" intact. Whether the machine is simple or wildly complex, this method ensures the simplified version behaves just like the real thing, making simulations faster and more reliable for engineers and scientists.

Technical Summary

Here is a detailed technical summary of the paper "Structure-Preserving Model Reduction on Manifolds of Port-Hamiltonian Systems" by Silke Glas and Hongliang Mu.

1. Problem Statement

Port-Hamiltonian (pH) systems are widely used to model multi-physics systems (e.g., electrical circuits, mechanical systems) because they inherently satisfy energy balance equations, ensuring properties like stability and passivity. However, high-dimensional pH systems are computationally expensive to simulate.

While Model Order Reduction (MOR) techniques exist to approximate these systems with Reduced-Order Models (ROMs), standard methods (like POD or IRKA) often fail to preserve the intrinsic pH structure in the reduced model. Furthermore, existing structure-preserving methods for nonlinear pH systems are limited:

• Many rely on specific linear approximations.
• Some are non-intrusive (data-driven) but lack general nonlinear approximation capabilities.
• There is a lack of intrusive structure-preserving MOR methods based on general nonlinear approximation maps for nonlinear pH systems.

The paper aims to close this gap by proposing a MOR framework that preserves the pH structure for both linear and nonlinear systems using general nonlinear embedding maps.

2. Methodology

The proposed methodology is built upon the Generalized Manifold Galerkin (GMG) reduction framework, adapted specifically for pH systems.

A. Theoretical Foundation: GMG for pH Systems

The authors define a reduced-order state $\check{x} \in \mathbb{R}^r$ and a reconstruction map $\phi: \mathbb{R}^r \to \mathbb{R}^N$ such that the full state $x(t) \approx \phi(\check{x}(t))$.
To preserve the pH structure, they derive a specific reduction map $W(\check{x})$ based on the Jacobian of the embedding map ($D_{\check{x}}\phi$) and the system's interconnection/dissipation matrices ($J$ and $R$).

Key Theorem (3.1):
The reduced model remains a valid pH system if:

1. The span of the input matrix $B$ is contained within the span of the Jacobian $D_{\check{x}}\phi(\check{x})$.
2. The Jacobian satisfies a non-degeneracy condition with respect to $(J-R)^{-1}$.

Under these conditions, the reduced system matrices are defined as:

• $\check{J}(\check{x}) = W(\check{x})^\top J W(\check{x})$ (Skew-symmetric)
• $\check{R}(\check{x}) = W(\check{x})^\top R W(\check{x})$ (Positive semi-definite)
• $\check{H}(\check{x}) = H(\phi(\check{x}))$
• $\check{B}(\check{x}) = W(\check{x})^\top B$

B. Specific Realizations

The paper details two specific constructions for the approximation map $\phi$:

1. Linear Embedding (GMG-POD-ROM):

   • $\phi(\check{x}) = B\check{x}_1 + V\check{x}_2$, where $V$ is derived via Proper Orthogonal Decomposition (POD) on snapshots orthogonal to the input matrix $B$.
   • This ensures the span condition is met by construction.

2. Quadratic Embedding (GMG-QM-ROM):

   • $\phi(\check{x}) = B\check{x}_1 + V_1\check{x}_2 + V_2 M(\check{x}_2 \otimes \check{x}_2)$.
   • This utilizes a quadratic manifold to capture nonlinear dynamics more effectively.
   • Matrices $V_1, V_2$ are derived via POD, and $M$ is computed via a least-squares optimization with regularization.

C. Handling Nonlinearity (DEIM)

For systems with nonlinear Hamiltonians, evaluating the gradient $\nabla_x H$ in the ROM can still depend on the full dimension $N$. To address this, the authors integrate the Structure-Preserving Discrete Empirical Interpolation Method (SP-DEIM). This approximates the nonlinear terms using a low-dimensional basis while strictly maintaining the pH structure.

3. Key Contributions

• General Nonlinear Framework: The paper proposes the first intrusive, structure-preserving MOR method for nonlinear pH systems based on general nonlinear approximation maps, extending beyond the linear or specific-structure methods found in literature.
• Theoretical Guarantee: It provides a rigorous proof (Theorem 3.1) that the resulting ROM retains the pH structure (skew-symmetry of $\check{J}$, symmetry of $\check{R}$, and correct input-output port definition) under specific geometric conditions.
• Dual Implementation: It offers concrete algorithms for both linear (GMG-POD-ROM) and quadratic (GMG-QM-ROM) embeddings, making the method applicable to a wide range of physical systems.
• Integration with DEIM: It successfully combines GMG reduction with structure-preserving DEIM to handle nonlinear Hamiltonians efficiently.

4. Numerical Results

The authors validated their methods on two mass-spring-damper systems:

1. Linear Mass-Spring-Damper System ($N=100$):

   • Comparison: GMG-POD-ROM and GMG-QM-ROM were compared against SP1-POD-ROM (a state-of-the-art linear method).
   • Outcome: The GMG-based methods achieved significantly lower relative reduction and output errors. The quadratic version (GMG-QM-ROM) outperformed the linear version, demonstrating the benefit of nonlinear embeddings even for linear systems (due to better manifold approximation).

2. Nonlinear Mass-Spring-Damper System ($N=1000$):

   • Comparison: Compared against SP2-POD-ROM (a method using gradient-based POD).
   • Outcome:
     • State Error: GMG-POD-ROM performed similarly to SP2-POD-ROM.
     • Output Error: GMG-POD-ROM outperformed SP2-POD-ROM.
     • Quadratic Improvement: The GMG-QM-ROM showed the best performance, significantly reducing both state and output errors compared to linear methods.
   • Energy Balance: All proposed ROMs preserved the energy balance equation with errors comparable to the Full-Order Model (FOM), confirming the structure-preserving nature of the reduction.

5. Significance

This work is significant because it bridges the gap between geometric model reduction and port-Hamiltonian systems.

• Robustness: By ensuring the reduced model is mathematically guaranteed to be a pH system, the ROM inherits stability and passivity properties without needing post-hoc correction.
• Accuracy: The ability to use general nonlinear maps (specifically quadratic embeddings) allows for high-fidelity approximations of systems where linear subspaces fail (e.g., systems with slowly decaying Kolmogorov $n$-widths).
• Applicability: The framework is versatile, applicable to both linear and nonlinear systems, and provides a clear path for future extensions, such as using neural networks for the embedding map or handling state-dependent system matrices.

In summary, the paper presents a mathematically rigorous and numerically superior approach to reducing complex physical systems while strictly maintaining their fundamental energy-based physical properties.