Abstract
A variety of reduced order modeling (ROM) methods for geometrically nonlinear structures have been developed over recent decades, each of which takes a distinct approach, and may have different advantages and disadvantages for a given application. This research challenge is motivated by the need for a consistent, reliable, and ongoing process for ROM comparison. In this chapter, seven state-of-the-art ROM methods are evaluated and compared in terms of accuracy and efficiency in capturing the nonlinear characteristics of a benchmark structure: a curved, perforated plate that is part of the exhaust system of a large diesel engine. Preliminary results comparing the full-order and ROM simulations are discussed. The predictions obtained by the various methods are compared to provide an understanding of the performance differences between the ROM methods participating in the challenge. Where possible, comments are provided on insight gained into how geometric nonlinearity contributes to the nonlinear behavior of the benchmark system.
Introduction: The Computational Bottleneck in Nonlinear Dynamics
In modern structural engineering, the ability to accurately predict the vibration and deformation of components is vital for ensuring long-term durability and preventing fatigue failure. However, when dealing with "geometrically nonlinear" structures—such as thin, curved plates—traditional simulation techniques like the Finite Element Method (FEM) become computationally prohibitive. These models often require thousands or even millions of degrees of freedom (DOF) to capture complex behaviors, making them unsuitable for rapid design iterations or real-time structural health monitoring.
Reduced Order Models (ROMs) offer a solution by mathematically "compressing" these high-dimensional systems into a handful of essential variables. Despite their potential, the ROM landscape is fragmented, with many methods tested only on simplified "toy" problems. The 2023 ROM Research Challenge was established to provide a rigorous, head-to-head comparison of state-of-the-art ROM techniques against a complex, industrially relevant benchmark.
The Benchmark: A Perforated Exhaust Cover Plate
The focal point of this challenge is a curved, perforated stainless steel plate, 317.5 mm in diameter, which serves as an exhaust cover for a large diesel engine. This component is known to experience significant geometric nonlinearity and has a history of fatigue failures in the field.
The "ground truth" for the challenge was established using a high-fidelity FEM consisting of 1,440 shell elements (8,566 DOF). This full-order model was solved using the Multi-Harmonic Balance (MHB) method to provide a baseline for comparing the accuracy and efficiency of the various ROMs.
Evaluated ROM Methodologies
Seven distinct ROM methods, representing different mathematical philosophies, participated in the challenge:
- Implicit Condensation (IC): Represented by ICE-GPR (Gaussian Process Regression) and ICE-IC (Inertial Compensation). These methods simplify dynamics by assuming high-frequency modes respond quasi-statically to low-frequency excitations.
- Modal Derivatives (MD): This approach augments linear vibration modes with "derivatives" that describe how mode shapes evolve as the structure deforms.
- Invariant Manifolds: Techniques such as DNF (Direct Normal Form), DPIM (Direct Parameterization of Invariant Manifolds), and SSM (Spectral Submanifolds) attempt to identify the low-dimensional "surface" in state-space that governs the system's primary dynamics.
- Data-Driven Learning: SSMLearn identifies these invariant manifolds directly from simulated trajectory data, bypassing the need for explicit underlying equations.
Comparative Performance and Backbone Curves
The primary metric for evaluation was the Nonlinear Normal Mode (NNM) backbone curve, which tracks how the vibration frequency of the system changes as a function of energy or displacement. This provides a much deeper insight into the nonlinear physics than simple time-history comparisons.
The results revealed several key insights:
- Softening-Hardening Transition: The plate initially exhibits "softening" behavior (frequency decreases as amplitude increases) before transitioning to "hardening." Most ROMs successfully captured this complex transition.
- The Convergence Limit: Nearly all methods, including the FEM solver, encountered a "convergence wall" at high energy levels (approximately 0.3 J). This is attributed to "snap-through" dynamics or internal resonances where the plate's behavior becomes highly coupled and chaotic.
- Efficiency vs. Accuracy: All ROMs achieved computational speedups of several orders of magnitude. The ICE methods showed remarkable accuracy with as few as 4 to 6 modes, while the MD ROM maintained accuracy up to 0.4 J before struggling with high-amplitude convergence.
- Truncation Effects: For manifold-based methods, the mathematical order of truncation proved critical. A 3rd-order DNF model failed at high amplitudes, whereas a 7th-order DPIM model remained accurate for a significantly larger range.
Conclusion and Future Directions
The 2023 ROM Research Challenge demonstrates that while modern ROM techniques are highly efficient, their performance is sensitive to the complexity of the underlying dynamics. The challenge highlighted the difficulty of modeling "snap-through" events and the importance of choosing the right truncation order and mode selection.
Future work will focus on expanding the challenge to include different types of nonlinear responses, investigating the impact of damping on ROM performance, and refining models to better handle the sophisticated dynamics encountered during structural buckling and state transitions.
Summary updated