Reduced-order modelling of parametrized unsteady Navier-Stokes equations and application to flow around cylinders with periodic changing boundary conditions

This paper presents a reduced-order model combining Proper Orthogonal Decomposition (POD) and Radial Basis Functions (RBF) to efficiently and accurately predict unsteady flows with periodic boundary conditions, demonstrating a CPU time reduction of over 99% in a 3D flow around cylinders.

The Gist

Imagine you are trying to predict the exact movement of every single leaf in a massive, swirling storm. To do this perfectly, you would need a supercomputer and days of calculation. But what if you only needed to know the general shape of the storm to stay safe? You wouldn't want to wait days for the answer; you’d need a "shortcut."

This paper describes a mathematical "shortcut" for predicting how fluids (like air or water) move around complex objects, like cylinders, when the conditions are constantly changing.

Here is the breakdown of how they did it, using everyday analogies.

1. The Problem: The "High-Definition" Headache

In engineering, scientists use CFD (Computational Fluid Dynamics). Think of this as trying to simulate a river by tracking the position of every single microscopic drop of water. It is incredibly accurate, but it is "heavy." It requires massive supercomputers and a huge amount of time. If you are trying to design a car or a wind turbine, you might need to run this simulation thousands of times. Doing that with the "full-detail" method would take years.

2. The Solution: The "Sketch Artist" Approach

The researchers used two main tools to create a Reduced-Order Model (ROM)—which is essentially a "lite" version of the simulation.

Step A: POD (The "Highlight Reel")

Imagine you are watching a 2-hour long, high-definition movie of a dancer. Instead of saving every single frame, you decide to only save the "essential" movements—the big leaps, the spins, and the poses.

This is Proper Orthogonal Decomposition (POD). The researchers took a massive amount of data from a full simulation and stripped away the "noise." They identified the "main characters" of the flow (the big swirls and currents) and ignored the tiny, unimportant ripples. They turned a massive, complex movie into a few key "dance moves."

Step B: RBF (The "Connect-the-Dots" Master)

Now that they have the "dance moves," they need to know how those moves change when the wind speed changes.

Imagine you have a few dots on a piece of paper representing different wind speeds. Radial Basis Function (RBF) is like a master artist who can look at those dots and instantly draw a smooth, beautiful curve that connects them all. It allows the computer to "guess" (interpolate) what the flow will look like at a wind speed it has never actually seen before, just by looking at the patterns of the speeds it has seen.

3. The Result: Speed vs. Accuracy

The researchers tested this on water flowing around three cylinders where the water speed was pulsing up and down like a heartbeat (sinusoidal flow).

• The Speed Boost: The full simulation (the "heavy" way) took about 20,000 seconds. The new shortcut (the "lite" way) took only 62 seconds. That is a 99% reduction in time! It’s like the difference between reading an entire encyclopedia and just reading the summary on the back cover.
• The Accuracy: Even though they skipped most of the details, they were still incredibly accurate. Their "guesses" were off by only about 5%.

4. The "Goldilocks" Warning (Overfitting)

The researchers discovered something very important: More is not always better.

If you try to include too many tiny details (too many "dance moves") in your shortcut, the model becomes "overfit." This is like a student who memorizes the exact answers to a practice test instead of learning the actual math. When the real test comes with slightly different numbers, the student fails.

The researchers found that if they used too many modes, the prediction got worse. They had to find the "Goldilocks" number—not too many, not too few—to get the perfect prediction.

Summary

In short: This paper provides a way to turn a massive, slow, "high-definition" physics problem into a fast, "sketch-style" mathematical model. It allows engineers to predict complex, changing fluid flows in seconds rather than hours, without losing much accuracy.

Technical Summary

Technical Summary: Reduced-Order Modelling of Parametrized Unsteady Navier–Stokes Equations

1. Problem Statement

Computational Fluid Dynamics (CFD) simulations are essential for engineering but are computationally prohibitive when applied to problems requiring repeated calculations, such as optimization, uncertainty analysis, or real-time emergency management. While many Reduced-Order Models (ROMs) exist, most are designed for reconstructing flow fields within a fixed parameter space (interpolation) rather than predicting future states in unsteady flows. Predicting unsteady flows with periodic, time-varying boundary conditions remains a significant challenge due to the non-linear and dynamic nature of the Navier–Stokes equations.

2. Methodology

The authors propose a hybrid ROM framework combining Proper Orthogonal Decomposition (POD) and Radial Basis Function (RBF) interpolation. The approach follows an offline-online strategy:

• Offline Phase (Full-Order Model - FOM):

  • High-fidelity snapshots of the velocity and pressure fields are generated using the Finite Volume Method (FVM) to solve the unsteady incompressible Navier–Stokes equations.
  • POD Generation: The snapshots are decomposed using Singular Value Decomposition (SVD) (or an equivalent correlation matrix method) to extract orthogonal basis functions (modes). These modes capture the dominant coherent structures of the flow.
  • Coefficient Extraction: The high-dimensional snapshots are projected onto these low-dimensional POD modes to obtain temporal coefficients ($a_i$ for velocity and $b_i$ for pressure).

• Online Phase (ROM Prediction):

  • RBF Interpolation: Instead of using the computationally expensive POD-Galerkin projection, the authors use RBFs to establish a non-linear mapping between the current state (POD coefficients at time $t$) and the future state (coefficients at $t + \Delta t$).
  • Markov Assumption: The model assumes that for sufficiently small time steps, the unsteady flow can be treated as a Markov process, where the next state depends primarily on the current state.
  • Reconstruction: The predicted coefficients are then used to reconstruct the full-dimensional flow field via a linear combination of the pre-calculated POD modes.

3. Key Contributions

• Predictive Capability: Unlike standard POD methods that focus on interpolation within a known parameter range, this method specifically addresses the prediction of unsteady flows with periodic boundary conditions.
• RBF-based Non-linear Mapping: By using RBF interpolation for the coefficients rather than Galerkin projection, the model avoids the high computational cost of integral evaluations over the entire domain and handles non-linearity more efficiently.
• Identification of Overfitting: The study provides a critical insight into the relationship between the number of POD modes and prediction accuracy, identifying that increasing modes does not always improve prediction due to "overfitting" (where the model captures noise rather than physics).

4. Results and Validation

The method was validated using a 3D numerical case of unsteady flow around three cylinders with a sinusoidal inlet velocity.

• Accuracy (Reconstruction): For the reconstruction of known snapshots, the model was highly accurate, with a maximum $L_2$ relative error of less than 3.89%.
• Accuracy (Prediction): For predicting future states (the period $150\text{s} < t \le 230\text{s}$), the model achieved a maximum $L_2$ relative error of 5.13% using the optimal number of modes.
• Optimal Mode Selection: The authors found that while 20 modes were good for reconstruction, 14 modes were optimal for prediction. Using too many modes (e.g., 25) caused the error to double due to overfitting.
• Efficiency: The ROM demonstrated massive computational savings. Compared to the FOM (which required a 64-processor cluster), the ROM (run on a standard laptop) reduced CPU time by more than 99% (from $\approx 20,034$ seconds to $\approx 62$ seconds).

5. Significance

This research demonstrates that a POD-RBF framework can provide a highly efficient and stable alternative to full-scale CFD for unsteady, periodic flows. The ability to reduce computational costs by two orders of magnitude while maintaining an error rate below 6% makes this approach highly significant for real-time engineering applications, such as monitoring fluid-structure interactions or managing transient hydraulic systems where rapid decision-making is required.