TL;DR
This paper introduces a novel data-driven surrogate modeling approach for complex fluid dynamics problems governed by the Navier-Stokes equations. By combining Proper Orthogonal Decomposition (POD), Dynamic Mode Decomposition (DMD), and manifold interpolation, the method efficiently captures time-dependent solutions, even for challenging scenarios near turbulent behavior, and can recover frequencies not present in the training data.
What Problem Does This Paper Address?
Simulating complex fluid dynamics, particularly time-dependent flows described by the Navier-Stokes equations, is computationally expensive. This is especially true when studying how these flows change with varying parameters (e.g., Grashof number in buoyancy-driven flows). Traditional reduced-order modeling (ROM) techniques often struggle with capturing the intricate dynamics, such as time-periodic solutions with changing frequencies, that arise in these systems, particularly near bifurcations where the flow behavior dramatically shifts. The challenge is to develop a ROM that is both accurate and efficient, capable of predicting these complex behaviors across a range of parameters without requiring expensive full-order simulations for each new parameter value.
Methods
The proposed approach is a three-stage, data-driven ROM:
- Proper Orthogonal Decomposition (POD): This initial step reduces the dimensionality of the high-fidelity simulation data. It identifies dominant spatial modes that capture the most variance in the flow field. A two-tier POD approach is used to manage storage requirements for long time trajectories.
- Dynamic Mode Decomposition (DMD): After projecting the data onto the POD space, DMD is applied to capture the temporal dynamics. DMD decomposes the flow into modes that evolve linearly in time, effectively identifying dominant frequencies and growth/decay rates. This is particularly useful for time-periodic phenomena.
- Manifold Interpolation: This is the core novelty for handling parameter variations. Instead of interpolating the DMD modes directly, the method interpolates the reduced DMD operators (Koopman operator) on a manifold. This structure-preserving interpolation allows for accurate prediction of dynamics at new, unseen parameter values, crucially enabling the recovery of frequencies that differ from those in the training data.
Datasets
The method is tested and validated on the Rayleigh-Bénard cavity flow problem. This is a well-studied benchmark problem in fluid dynamics that exhibits rich bifurcating behavior, including steady-state solutions, time-periodic solutions, and chaotic behavior as the Grashof number (Gr) increases. The authors consider two parameter domains: a "medium Gr range" ([100e3, 150e3]) and a "high Gr range" ([650e3, 700e3]), which is closer to turbulent regimes.
Key Insights and Ideas
- Interpolating Operators, Not Modes: The key insight is that interpolating the reduced DMD operators on a manifold, rather than the DMD modes themselves, is crucial for accurately capturing parameter-dependent dynamics, especially when frequencies change.
- Recovering Unseen Frequencies: The method's ability to recover frequencies that are not present in the sampled data is a significant advantage, particularly for time-periodic solutions where these frequencies can shift with parameters.
- Manifold Interpolation for Dynamics: Applying manifold interpolation to the DMD operators provides a robust way to handle the non-linear nature of dynamical systems and their parameter dependencies.
- Base Point Sensitivity: The accuracy of manifold interpolation is sensitive to the choice of the base point, highlighting the importance of appropriate sampling density and selection of nearby sample points.
Key Results
- The proposed ROM approach achieves good accuracy for the Rayleigh-Bénard cavity flow in both medium and high Grashof number regimes.
- In the medium Gr range, mean relative errors for velocity are generally below 1%, with maximum errors up to 8% in less optimal cases.
- In the high Gr range, mean relative errors are higher, up to 15% on average, with maximum errors reaching 25%. This is still considered engineering accuracy, especially when compared to the full-order simulation errors.
- The method successfully captures important flow features and time-periodic behavior, even in challenging regimes close to chaotic transitions.
- Qualitative comparisons show that the ROM captures most important features, though some phase differences can occur in more complex scenarios.
Limitations and Future Work
- DMD Stability: The authors note that the stability of the DMD algorithm could be further improved using existing techniques.
- Non-Flat Metrics: For manifold interpolation, using non-flat metrics could potentially enhance accuracy.
- Complex DMD: Exploring complex DMD could also be a avenue for improvement.
- Broader Applications: The authors suggest applying the proposed approach to other engineering problems, higher-dimensional parameter domains, and quasi-periodic systems.
- Accuracy in High Gr Regime: While achieving engineering accuracy, the mean relative error in the high Grashof regime remains above 10% in some instances, suggesting room for improvement.
Summary updated