Non-intrusive Balancing Transformation of Highly Stiff Systems with Lightly-damped Impulse Response

Elnaz Rezaian, Cheng Huang, Karthik DuraisamyarXiv 2021 · DOI · OpenReview

TL;DR

This paper introduces a novel non-intrusive method for creating reduced-order models (ROMs) of highly stiff and lightly-damped dynamical systems. By leveraging the Eigensystem Realization Algorithm (ERA), the authors overcome limitations of traditional methods, enabling accurate predictions for complex systems like reactive flows, even with challenging stiffness and full-state outputs.

What Problem Does This Paper Address?

High-fidelity simulations are crucial for understanding complex phenomena in fields like turbulence and chemical reactions. However, these simulations are computationally expensive, limiting their use in applications requiring rapid design iterations, control, or uncertainty quantification. Model reduction techniques aim to create simpler, faster models (ROMs) that capture the essential dynamics of these complex systems.

Traditional model reduction methods, like Balanced Truncation (BT), rely on transforming the system into a coordinate frame where its most important dynamics are both "observable" and "reachable." This is typically done by solving computationally intensive Lyapunov equations to obtain system Gramians. This approach faces significant challenges when dealing with:

  • Highly stiff systems: Extreme stiffness, often introduced by phenomena like chemical reactions, can lead to numerical instabilities and errors in the analytical computation of Gramians.
  • Lightly-damped systems: These systems exhibit oscillations that persist for a long time after initial disturbances, making it difficult to capture their full dynamics with limited data.
  • Full-state outputs: When all states of the system are observable, traditional BT methods become computationally prohibitive due to the large size of the required matrices.

The paper addresses the need for a scalable, accurate, and non-intrusive model reduction technique that can handle these challenging system characteristics, particularly for systems where direct access to the full-order model (FOM) operators is not feasible.

Methods

The core of this work is the development of a non-intrusive, data-driven Balanced Truncation (BT) method using the Eigensystem Realization Algorithm (ERA). Here's a breakdown of the key techniques:

  • Eigensystem Realization Algorithm (ERA): Traditionally used in system identification, ERA is adapted here for model reduction. It works by analyzing the system's impulse response (Markov parameters) to construct a Hankel matrix. The singular value decomposition (SVD) of this Hankel matrix allows for the identification of balanced system matrices without requiring direct access to the FOM's Jacobian or performing adjoint simulations. This is a significant advantage for systems with full-state outputs or experimental data.
  • Non-intrusive Approach: Unlike traditional methods that require manipulation of the FOM's governing equations or operators, ERA operates solely on input-output data (impulse response). This makes it suitable for "black-box" systems or experimental data where internal model details are unknown or inaccessible.
  • Handling Stiffness: ERA's data-driven nature bypasses the direct computation of Gramians via Lyapunov equations, which are prone to instability and errors in stiff systems.
  • Output Domain Decomposition: To address the computational cost associated with high sampling frequencies needed for sharp gradients, the authors propose dividing the output space into subdomains. This allows for higher sampling rates only in regions with sharp gradients, reducing overall computational burden.
  • Tangential Interpolation: This technique is used to further reduce the dimensionality of the problem by projecting the impulse response onto dominant tangential directions. This helps in managing the size of the Hankel matrix, especially for multi-input multi-output (MIMO) systems.

The paper also discusses standard methods for comparison:

  • Proper Orthogonal Decomposition (POD): A common dimensionality reduction technique.
  • POD-Galerkin Projection: A projection-based ROM method.
  • Least-Squares Petrov-Galerkin (LSPG) Projection: Another projection-based method, often used for stabilization.

Datasets

The primary "dataset" used in this paper is the impulse response of a simulated one-dimensional reactive flow system. This impulse response is generated by the full-order model (FOM) of the Navier-Stokes equations with species transport and reaction. The system is impulsively excited by a pressure perturbation.

The authors collect a significant number of snapshots (up to 1000 or more) of this impulse response at various sampling frequencies to train the ROMs. The performance of the developed ROMs is then evaluated using unseen sinusoidal pressure forcing inputs at different frequencies.

Key Insights and Ideas

  • ERA as a Game-Changer for Stiff Systems: The paper convincingly demonstrates that ERA can effectively perform balancing transformations for highly stiff systems, a task that is problematic for analytical BT methods. This opens doors for ROMs in previously intractable applications.
  • Non-intrusiveness is Key for Scalability and Accessibility: By avoiding direct access to FOM operators and adjoint simulations, ERA-based BT becomes significantly more scalable and applicable to experimental data or complex simulations where internal model access is limited.
  • Sampling Properties are Paramount: The performance of data-driven ROMs, especially ERA-based ones, is highly sensitive to the frequency and duration of impulse response sampling. Insufficient sampling leads to inaccurate or unstable ROMs, while overly dense sampling can be computationally prohibitive.
  • Lightly-damped Tails Require Long Sampling: The presence of lightly-damped oscillations in the impulse response tail necessitates sampling for extended periods to accurately capture these persistent dynamics, which is crucial for predictive accuracy.
  • Output Domain Decomposition and Tangential Interpolation Offer a Practical Solution: These techniques are crucial for managing the computational cost associated with high-fidelity data (e.g., sharp gradients, full-state outputs) while maintaining accuracy. They allow for targeted high-resolution sampling and data reduction.
  • Predictive Power is the Ultimate Goal: The paper emphasizes that the true value of ROMs lies in their ability to predict system behavior under unseen conditions, a capability that traditional POD-Galerkin and LSPG methods struggle with in this specific application, while ERA-based BT excels.

Key Results

  • Accurate Predictive ROMs: The developed non-intrusive balanced ROMs (NI-BROMs) using ERA demonstrate excellent predictive performance for unseen sinusoidal forcing inputs, accurately capturing the dynamics of the reactive flow system.
  • Failure of Standard Methods: In contrast, standard POD-Galerkin and LSPG ROMs fail to accurately represent the true dynamics under similar predictive scenarios, often exhibiting unbounded numerical errors.
  • Significant Speedup: The online stage of the ERA-based ROMs achieves a substantial speedup factor (e.g., 138.12) compared to the FOM, making them suitable for real-time applications.
  • Handling Extreme Stiffness: The method successfully models a system with a Jacobian condition number of approximately $4.82 \times 10^{13}$, a level of stiffness that would typically prevent analytical BT.
  • Improved Accuracy with Domain Decomposition and Interpolation: The combination of output domain decomposition and tangential interpolation significantly improves the accuracy of ROMs, particularly in resolving sharp gradients, while also reducing offline computation costs.
  • Stability of Balanced ROMs: The eigenvalues of the constructed balanced ROMs lie within the unit circle, demonstrating discrete-time stability, even when using a much larger time step than the FOM.

Limitations and Future Work

  • Sampling Sensitivity: While the paper highlights the importance of sampling, achieving optimal sampling strategies for complex systems can still be challenging and may require significant computational effort for analysis. The authors note that sampling periods of $1 \mu$sec and $0.5 \mu$sec were insufficient, while $0.1 \mu$sec provided good agreement.
  • Accuracy vs. Cost Trade-off: The output domain decomposition and tangential interpolation methods, while effective, involve a trade-off between computational cost and accuracy. More aggressive truncation can reduce costs but may impact accuracy, especially in regions with sharp gradients.
  • Future Work: The paper suggests that further research could explore more advanced methods for automatically determining optimal sampling strategies and subdomain decomposition. The application of this non-intrusive BT approach to other complex, stiff, and lightly-damped systems (e.g., turbulence, combustion, fluid-structure interaction) is also a promising avenue. The authors also mention the potential for extending this to non-linear systems, although the current work focuses on linearized dynamics.

Summary updated

Abstract

Balanced truncation (BT) is a model reduction method that utilizes a coordinate transformation to retain eigen-directions that are highly observable and reachable. To address realizability and scalability of BT applied to highly stiff and lightly-damped systems, a non-intrusive data-driven method is developed for balancing discrete-time systems via the eigensystem realization algorithm (ERA). The advantage of ERA for balancing transformation makes full-state outputs tractable. Further, ERA enables balancing despite stiffness, by eliminating computation of balancing modes and adjoint simulations. As a demonstrative example, we create balanced ROMs for a one-dimensional reactive flow with pressure forcing, where the stiffness introduced by the chemical source term is extreme (condition number $10^{13}$), preventing analytical implementation of BT. We investigate the performance of ROMs in prediction of dynamics with unseen forcing inputs and demonstrate stability and accuracy of balanced ROMs in truly predictive scenarios whereas without ERA, POD-Galerkin and Least-squares Petrov-Galerkin projections fail to represent the true dynamics. We show that after the initial transients under unit impulse forcing, the system undergoes lightly-damped oscillations, which magnifies the influence of sampling properties on predictive performance of the balanced ROMs. We propose an output domain decomposition approach and couple it with tangential interpolation to resolve sharp gradients at reduced computational costs.
BibTeX
@misc{rezaian2021non,
  author = {Elnaz Rezaian and Cheng Huang and Karthik Duraisamy},
  title = {{Non-intrusive Balancing Transformation of Highly Stiff Systems with Lightly-damped Impulse Response}},
  year = {2021},
  eprint = {2109.10408},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2109.10408},
}
Figure