Bayesian-Enhanced Galerkin-Based Reduced Order Modelling for Unsteady Compressible Flows

This paper proposes a Bayesian-enhanced Galerkin-POD framework that reformulates model coefficient correction as a statistical inverse problem to systematically address uncertainties from mode truncation and data noise, thereby significantly improving the stability, robustness, and predictive accuracy of reduced-order models for unsteady compressible flows.

The Gist

The Big Picture: Predicting the Unpredictable

Imagine you are trying to predict the weather. The atmosphere is a chaotic, swirling mess of wind, heat, and pressure. To understand it, scientists use supercomputers to run massive simulations (like a digital twin of the atmosphere). These simulations are incredibly accurate, but they are also slow and expensive. Running one simulation might take days on a supercomputer.

Now, imagine you want to predict the weather in real-time, or run thousands of scenarios to design a better airplane wing. You can't wait days for the computer to finish. You need a "shortcut."

This is where Reduced Order Modelling (ROM) comes in. It's like taking a high-definition 4K movie and compressing it into a low-resolution GIF. You lose some detail, but you get the main action (the storm, the wind) much faster.

The Problem: The "Compressed" Movie Glitches

The authors of this paper are experts in a specific type of shortcut called Galerkin-POD. Think of this method as a way to break a complex fluid flow (like air rushing over a car) into a few "building blocks" or "modes."

• Mode 1: The big, slow swirls.
• Mode 2: The medium wiggles.
• Mode 3: The tiny jitters.

Usually, you only keep the first few big modes to save time. But here's the catch: The shortcut breaks.

If you try to run the simulation for too long, the math gets unstable. The "movie" starts to glitch, the numbers go crazy, and the prediction diverges from reality. It's like a compressed video file that starts to pixelate and freeze after a few minutes.

Why does this happen?

1. Missing Pieces: By throwing away the tiny "jitter" modes (to save space), you lose the natural "damping" or friction that keeps the system stable.
2. Noisy Data: The data used to build the model often has tiny errors or "static" (like a bad phone signal), which gets amplified when you try to calculate how the fluid moves.

The Solution: The Bayesian "Editor"

The authors propose a new method: Bayesian-Enhanced Galerkin-POD.

Think of the standard model as a rough draft written by a student. It has the right general idea, but the details are shaky.
Bayesian Inference is like a smart editor who knows two things:

1. What the student usually writes (The Prior): The student's general style and logic.
2. What the actual facts are (The Likelihood): The real data from the wind tunnel or supercomputer.

The editor doesn't just throw away the student's draft. Instead, they calibrate it. They look at the errors and say, "Okay, the student got the big picture right, but they missed a tiny bit of friction here and added a little too much noise there. Let's tweak the numbers slightly to match reality."

In technical terms, they treat the missing physics and the noisy data as uncertainties. They use statistics to find the "best guess" for the missing numbers, effectively "fixing" the shortcut so it doesn't crash.

The Two Test Cases

The team tested this "smart editor" on two very different challenges:

1. The Dimpled Surface (The Practice Run)

• The Scenario: Air flowing over a surface with a single dimple (like a golf ball). It creates a rhythmic, self-sustaining oscillation (a "hum").
• The Result: The standard shortcut worked okay for a second, then spiraled out of control. The Bayesian version, however, kept the rhythm perfectly stable for a long time. It corrected the tiny math errors that were causing the glitch.
• Analogy: It's like a drummer who keeps perfect time, whereas the uncorrected model is a drummer who starts speeding up and eventually crashes into the cymbals.

2. The Centrifugal Compressor (The Real Boss Fight)

• The Scenario: A high-speed jet engine compressor. This is a nightmare of complexity: spinning blades, leaking air at the tips, and air slamming into stationary vanes. It's chaotic, turbulent, and happens at very high speeds.
• The Challenge: In this case, you have to throw away so many details (modes) to make the model fast enough that the "missing pieces" problem is huge.
• The Result: The standard model failed almost immediately. The Bayesian model, however, successfully predicted the complex interactions between the spinning blades and the stationary parts. It captured the "vortex breakdown" (where the air swirls violently and collapses) and the "blade beating" (the noise created when blades pass vanes).
• Analogy: Imagine trying to predict the movement of a swarm of bees in a hurricane. The standard model says, "I'll just guess the general direction," and fails. The Bayesian model says, "I know the general direction, but I'll also adjust for the wind gusts and the bees' erratic buzzing," and it actually works.

Why This Matters

This paper is a big deal because it combines Physics (how fluids actually move) with Statistics (how to handle uncertainty).

• Before: You had to choose between a slow, perfect simulation or a fast, broken shortcut.
• Now: You have a fast shortcut that is stable, accurate, and aware of its own mistakes.

It's like upgrading from a paper map that gets you lost in traffic to a GPS that not only knows the route but also adjusts in real-time for road closures and accidents. This allows engineers to design better engines, planes, and cars much faster and more reliably, without needing a supercomputer for every single test.

Technical Summary

1. Problem Statement

Conventional Galerkin–Proper Orthogonal Decomposition (Galerkin–POD) is a popular reduced-order modelling (ROM) technique that projects the Navier–Stokes equations onto a low-dimensional basis to create a system of Ordinary Differential Equations (ODEs). However, this approach suffers from two critical limitations, particularly in compressible flows:

• Instability and Divergence: The truncated ODE systems are highly sensitive to parameters. Over long integration times, they often diverge or converge to erroneous states because the truncation of POD modes removes the dissipative effects of small-scale vortices that naturally stabilize the system.
• Limited Predictive Capability: The models often fail to capture complex unsteady dynamics in high-Reynolds-number turbulent flows or intricate geometries.
• Sources of Error: The discrepancy between the ROM and high-fidelity data arises from:
  • Model Uncertainty: Neglecting low-energy modes (POD truncation).
  • Data Uncertainty: Noise in experimental/CFD data and amplification of errors during numerical post-processing (e.g., calculating high-order derivatives for dissipation terms).

Existing stabilization methods (e.g., eddy-viscosity, spectral vanishing viscosity) are often physics-based heuristics, while data-driven calibration methods often lack rigorous uncertainty quantification.

2. Methodology

The authors propose a Bayesian-Enhanced Galerkin–POD framework that treats the calibration of the reduced ODE coefficients as a statistical inverse problem.

• Galerkin Projection Formulation:
  • The compressible Navier–Stokes equations are reformulated using primitive variables where density is inverted ($\zeta = 1/\rho$) to reduce cubic nonlinearities to quadratic terms.
  • The flow field is projected onto a truncated set of POD modes, yielding a low-dimensional ODE system: $\dot{a}_k = \sum Q a_i a_j + \sum L a_i + C$.
• Bayesian Inference Framework:
  • The problem is formulated as $y = Ax + \epsilon$, where $y$ represents the time derivatives of POD coefficients (from high-fidelity data), $A$ contains the basis terms, and $x$ represents the uncertain ODE coefficients to be inferred.
  • Prior Distribution: The prior for the coefficients ($x$) is derived from the standard Galerkin projection results. An Inverse-Gamma prior is assigned to the error variance ($\sigma^2$) to handle unknown noise levels.
  • Likelihood: A Gaussian likelihood is assumed for the observation noise.
  • Analytical Solution: Instead of using computationally expensive sampling methods (like MCMC), which are impractical for high-dimensional turbulent flows, the authors derive an analytical, sampling-free solution. The posterior distribution of the coefficients follows a multivariate Student's t-distribution, allowing for efficient computation of the posterior mean.
  • Correction Mechanism: The final coefficients are a weighted combination of the prior knowledge (standard Galerkin) and the Ordinary Least Squares (OLS) estimate derived from the data, effectively correcting the ODE system to account for both model and data uncertainties.

3. Key Contributions

• Unified Uncertainty Quantification: The framework systematically addresses both model uncertainty (truncation error) and data uncertainty (noise/numerical error) within a single probabilistic framework.
• Sampling-Free Efficiency: By utilizing conjugate priors (Gaussian likelihood + Inverse-Gamma prior), the method provides an analytical solution. This avoids the prohibitive computational cost of sampling-based Bayesian methods, making it applicable to high-dimensional turbulent systems ($D \gg 10$).
• Extension to Compressible Flows: The methodology is specifically adapted for compressible flows, handling thermodynamic coupling and the specific nonlinearities inherent in the compressible Navier–Stokes equations.
• Robustness in High-Re Flows: Demonstrates that Bayesian correction can stabilize ROMs even when retaining a very small fraction of the total flow energy (e.g., capturing only 40% of unsteady energy in a turbulent compressor).

4. Results

The framework was validated on two distinct cases:

Case 1: Self-Sustained Oscillating Flow over a Dimpled Surface ($Re \approx 3000$)

• Context: Moderate Reynolds number with a single dominant shear layer instability.
• Findings:
  • Standard Galerkin–POD diverged rapidly (within 2–4 oscillation periods).
  • The Bayesian-enhanced model perfectly reproduced the temporal dynamics, oscillation amplitudes, and phase-space trajectories of the Direct Numerical Simulation (DNS) data over long durations.
  • The correction was minimal for dominant modes (low noise sensitivity) but significant for higher-order modes where numerical noise from derivative calculations was amplified.

Case 2: Centrifugal Compressor ($Re \approx 1.8 \times 10^5$)

• Context: High-Reynolds-number turbulent flow with complex interactions (tip-leakage vortex breakdown, impeller–diffuser interaction).
• Challenge: The energy spectrum decays slowly, requiring ~46 modes to capture 80% of energy. The ROM was restricted to only 10 modes (capturing ~40% of energy), introducing massive model uncertainty.
• Findings:
  • Standard Galerkin–POD failed almost immediately, diverging within a fraction of a rotation period.
  • The Bayesian-enhanced model successfully captured the dominant unsteady structures (tip-leakage vortex breakdown and rotor-stator interactions) and frequency characteristics.
  • The predictive time window was extended by nearly an order of magnitude compared to the standard ROM.
  • The reconstructed 3D flow fields matched Large Eddy Simulation (LES) benchmarks qualitatively and quantitatively with negligible computational cost.

5. Significance

This work bridges the gap between physics-based modeling (Galerkin projection) and statistical rigor (Bayesian inference). It offers a general, computationally efficient, and uncertainty-aware framework for reduced-order modeling.

• Reliability: It transforms unstable ROMs into robust predictive tools capable of handling complex, high-Reynolds-number compressible flows.
• Efficiency: The analytical solution ensures the method remains scalable for industrial applications where sampling methods are too slow.
• Future Impact: The authors highlight its potential for integration into digital twin architectures for real-time flow prediction and control in aerospace and automotive engineering.