TL;DR
This paper introduces a novel "Mixed-ROM" approach that combines traditional projection-based reduced order modeling (ROM) with data-driven techniques to efficiently simulate turbulent fluid flows. This hybrid method significantly improves the accuracy of ROMs for turbulent flows, particularly in capturing pressure fields and fluid dynamics forces, while maintaining computational efficiency.
What Problem Does This Paper Address?
Simulating turbulent fluid flows, especially at high Reynolds numbers, is computationally very expensive. Traditional numerical methods (Full Order Models or FOMs) require immense computational resources, making real-time control, optimization, or multi-query scenarios infeasible. Reduced Order Models (ROMs) offer a solution by creating simplified mathematical models that capture the essential dynamics of the flow with significantly reduced computational cost. However, existing ROMs for turbulent flows often struggle with energy stability issues and accurately representing the complex interactions of turbulent scales. Furthermore, accurately modeling the eddy viscosity, a key component in turbulence models, within a ROM framework has been a persistent challenge. This paper aims to bridge this gap by developing a ROM that is both accurate for turbulent flows and "independent" of the specific turbulence model used in the FOM.
Methods
The authors propose a "Mixed-ROM" approach that combines projection-based ROMs with data-driven techniques.
- Projection-Based ROMs: The core of the method uses Proper Orthogonal Decomposition (POD) to extract dominant spatial modes for velocity and pressure fields from FOM snapshots. These modes form reduced basis spaces. A Galerkin projection is then applied to the governing Navier-Stokes equations onto these POD spaces to derive a reduced-order dynamical system. To address stability issues in turbulent flows, a "supremizer enrichment" technique is employed for the velocity POD space, which helps in satisfying the continuity equation.
- Data-Driven ROM for Eddy Viscosity: The key innovation is the treatment of the eddy viscosity field. Instead of projecting the eddy viscosity equations (which can be complex and model-dependent), the authors use a data-driven approach, specifically Radial Basis Function (RBF) interpolation, to approximate the eddy viscosity coefficients. This allows the ROM to be more independent of the specific turbulence model used in the FOM. The RBF interpolation can be performed using either the time-parameter vector or, more innovatively, the reduced velocity coefficients and their time derivatives as input.
Datasets
The paper uses data generated from Full Order Models (FOMs) for two benchmark turbulent flow cases:
- Steady Backstep Flow: FOM simulations were run with varying inflow velocities (corresponding to Reynolds numbers from 1x10³ to 2.5x10⁴) to generate snapshots for training the ROM.
- Unsteady Flow Around a Circular Cylinder: FOM simulations were performed for varying inflow velocities (corresponding to Reynolds numbers from 7.5x10⁴ to 1.2x10⁵) and also for a fixed Reynolds number (10⁵) with different turbulence models (k-ε and SST k-ω). Snapshots were acquired over multiple vortex shedding cycles.
Key Insights and Ideas
- Hybrid Approach for Turbulence: The most significant insight is the successful integration of a data-driven method (RBF interpolation) for the eddy viscosity field within a projection-based ROM framework. This hybrid approach allows for accurate modeling of turbulent effects without the complexity of projecting turbulence transport equations.
- Eddy Viscosity Independence: The Mixed-ROM demonstrates a degree of independence from the specific turbulence model (k-ε vs. SST k-ω) used in the FOM, a crucial step towards more generalizable ROMs.
- Improved Pressure Accuracy: The Mixed-ROM shows a marked improvement in accurately capturing the pressure field compared to a purely projection-based ROM (P-ROM), which is often a challenge in ROMs for turbulent flows.
- Data-Driven Extrapolation in Time: By using reduced velocity coefficients as input for RBF interpolation, the method can potentially handle time extrapolation, a limitation of traditional ROMs.
Key Results
- Accuracy: The Mixed-ROM achieved significantly lower L² relative errors for velocity and pressure compared to the P-ROM. For the steady backstep case, errors were around 0.44% for velocity and 0.37% for pressure. For the unsteady cylinder case, errors were around 0.81% for velocity and 0.73% for pressure (with SST k-ω), while the P-ROM showed much higher errors (e.g., ~22% for pressure).
- Fluid Forces: The Mixed-ROM accurately predicted lift coefficients, with relative peak errors below 3.5% and L² relative errors below 5% for the unsteady cylinder case, outperforming the P-ROM which showed errors as high as 10-20%.
- Speedup: The paper reports significant speedups, around SU=10 for unsteady cases and SU=1000 for steady cases, demonstrating the computational efficiency of the method.
- Turbulence Model Robustness: The Mixed-ROM showed good agreement with FOM results across different turbulence models (k-ε and SST k-ω), indicating its robustness.
Limitations and Future Work
- Extrapolation in Time: While the RBF interpolation using velocity coefficients offers potential for time extrapolation, the paper notes that currently, it is not possible to extrapolate in time beyond the snapshot window. The accuracy of this extrapolation depends on how close the current ROM solution is to the offline snapshot data.
- Parameter Space Dependence: The accuracy of the data-driven interpolation for eddy viscosity is still dependent on the parameter values used during the offline training phase.
- Future Work: The authors suggest exploring Artificial Neural Networks (ANNs) to potentially improve the accuracy of eddy viscosity coefficient approximation. They also propose using Dynamic Mode Decomposition (DMD) for extrapolation problems and developing stabilization techniques for long-time integration and multi-physics problems.
Summary updated