TL;DR
This paper introduces a novel data-driven approach to improve the accuracy of reduced order models (ROMs) for fluid flows, specifically focusing on enhancing pressure prediction. By incorporating data-driven correction terms into both the momentum and continuity equations, the proposed method significantly boosts the accuracy of both velocity and pressure approximations, outperforming standard ROMs and previous data-driven techniques.
What Problem Does This Paper Address?
Reduced order models (ROMs) are powerful tools for accelerating complex simulations, particularly in fluid dynamics. However, standard ROMs often struggle with accuracy, especially in "marginally-resolved" or "under-resolved" regimes where the number of modes used is insufficient to fully capture the underlying physics. This is particularly true for pressure prediction, which is crucial for many engineering applications (e.g., calculating lift and drag) and can be challenging to reconstruct accurately, especially when the ROM basis is not perfectly divergence-free. Existing data-driven ROMs have primarily focused on improving velocity accuracy, leaving a gap in addressing the critical need for accurate pressure prediction.
Methods
The authors build upon the data-driven variational multiscale (DD-VMS) ROM framework. The core idea is to introduce data-driven closure/correction terms into the ROM equations to model the effects of neglected modes. This paper extends previous work by proposing and implementing correction terms not only for the velocity (momentum equation) but also for the pressure (continuity equation).
Two main ROM formulations are explored:
- SUP-ROM (Supremizer ROM): This approach enriches the velocity space with additional "supremizer" modes to satisfy the inf-sup condition, which is important for stable pressure-velocity coupling.
- PPE-ROM (Pressure Poisson Equation ROM): This formulation replaces the continuity equation with a Poisson equation for pressure, providing a dedicated equation for pressure approximation.
For both formulations, the data-driven correction terms are constructed by:
- Offline Stage: Postulating an ansatz (model form) for the correction term and then solving a least-squares problem to find the optimal parameters that best fit the exact correction term computed from full-order model (FOM) data.
- Online Stage: Integrating the ROM with the learned correction terms.
The paper also investigates different approaches for constructing these correction terms, including:
- Unconstrained Data-Driven Correction: A standard least-squares fitting approach.
- Constrained Data-Driven Correction: Incorporating physical constraints (e.g., semi-definiteness, skew-symmetry) into the optimization problem to potentially improve accuracy.
- Combined Data-Driven Corrections: Merging multiple correction terms into a single optimization problem for efficiency.
Datasets
The study uses numerical simulations of a 2D flow past a circular cylinder at a Reynolds number of Re = 50,000. The full-order model (FOM) solutions are generated using the open-source software OpenFOAM with the SST κ-ω turbulence model. The FOM snapshots are collected at a high frequency (0.004 s) to serve as the "ground truth" for training and evaluating the ROMs. No new benchmark datasets are introduced; rather, the standard flow problem is used to demonstrate the effectiveness of the proposed methods.
Key Insights and Ideas
- Pressure Matters: The paper strongly emphasizes the critical importance of accurate pressure prediction in ROMs, especially for engineering applications and when dealing with non-divergence-free bases.
- Data-Driven Pressure Corrections are Novel: This work presents the first pressure-based data-driven variational multiscale ROM, a significant advancement beyond previous velocity-focused approaches.
- PPE-ROM is Key for Pressure: The PPE-ROM formulation, which explicitly solves for pressure via a Poisson equation, proves to be much more effective for incorporating pressure correction terms than the SUP-ROM. This is likely because the SUP-ROM lacks a dedicated pressure equation, limiting the direct impact of pressure-specific corrections.
- Velocity Corrections Help Pressure (Indirectly): Even in the SUP-ROM, adding velocity correction terms indirectly improves pressure accuracy, though not as significantly as dedicated pressure corrections in the PPE-ROM.
- Time Window Selection is Crucial: The choice of time intervals for collecting snapshots (for POD basis construction) and for building correction terms significantly impacts the accuracy, especially for predictive simulations. For periodic flows, at least one flow period is recommended for building robust data-driven corrections.
Key Results
- The proposed pressure DD-VMS-ROM significantly improves both velocity and pressure approximations compared to standard ROMs and previous DD-VMS-ROMs (which only corrected velocity).
- In the SUP-ROM, velocity corrections improved both velocity and pressure, but pressure corrections alone had a negligible effect.
- In the PPE-ROM, adding pressure correction terms led to a substantial improvement in pressure accuracy, with combined velocity and pressure corrections yielding the best results.
- The constrained velocity correction in the SUP-ROM outperformed the unconstrained version, particularly in predictive scenarios.
- The PPE-ROM with combined velocity and pressure corrections achieved results very close to the reconstruction error (the best possible outcome given a fixed number of modes).
- Qualitative inspection of velocity and pressure fields (Figures 13 and 14) showed that the data-driven corrected fields were closer to the FOM solutions, especially near the cylinder, leading to better reconstruction of forces.
Limitations and Future Work
- SUP-ROM Pressure Correction Ineffectiveness: The paper notes that pressure corrections in the SUP-ROM formulation did not yield significant improvements. This is attributed to the lack of a dedicated pressure equation in this formulation, limiting the direct influence of pressure-specific corrections.
- Turbulence Modeling: The current work does not explicitly include turbulence modeling within the ROM framework.
- Reynolds Number Dependence: While the study was conducted at Re = 50,000, the authors suggest that higher Reynolds number test cases need to be carefully analyzed, as the effectiveness of certain terms might change.
Future work suggested includes:
- Exploring further data-driven terms for the SUP-ROM formulation that might better capture pressure dynamics.
- Extending the data-driven techniques to include turbulence modeling.
- Investigating the combined effects of data-driven corrections with other reduced-order modeling techniques.
- Applying and analyzing the proposed methods for higher Reynolds number flows.
Summary updated