To stall-cell or not to stall-cell: Variational data assimilation of 3D mean flow past a stalled airfoil

This study demonstrates that three-dimensional variational data assimilation, when combined with a Spalart--Allmaras turbulence model and sparse planar PIV measurements, can successfully reconstruct the full 3D mean flow field and essential features of stall cells on a post-stall NACA 0012 wing.

The Gist

The Big Picture: Reconstructing a 3D Puzzle from a Few Pieces

Imagine you are trying to understand the shape of a complex, invisible 3D sculpture floating in the air. You can't see the whole thing at once. All you have is a few flat, 2D photographs taken from specific angles. Your goal is to figure out what the entire 3D sculpture looks like based only on those few snapshots.

This is exactly what the researchers did, but instead of a sculpture, they were studying air flowing over an airplane wing. Specifically, they were looking at a chaotic phenomenon called "stall cells."

The Problem: The "Stall Cell" Mystery

When an airplane wing tilts too high (a high angle of attack), the smooth air flowing over it breaks apart and swirls wildly. This is called "stalling." Sometimes, this stalling doesn't happen evenly across the whole wing. Instead, it forms distinct, 3D bubbles of swirling air that look like mushrooms or "cells" moving along the wing.

• The Challenge: To see these cells, you usually need expensive, high-tech 3D cameras (like a CT scan for air). But these are hard to set up and very slow.
• The Reality: Most experiments only take 2D "slices" (like taking a single photo of a loaf of bread). The problem is that a single photo doesn't tell you how the air is moving sideways or how the "mushrooms" are arranged in 3D space.
• The Computer's Failure: The researchers tried using standard computer simulations (RANS) to predict these cells. It was like trying to guess the shape of a cloud by looking at a flat drawing; the computer predicted the air would separate, but it missed the complex 3D "mushroom" shapes entirely.

The Solution: The "Smart Guessing" Machine

The team used a technique called Variational Data Assimilation. Think of this as a super-smart detective who has two tools:

1. A Rulebook: The laws of physics (fluid dynamics) that say how air should behave.
2. A Clue: A few real-world photos (experimental data) showing what the air actually did in a few specific spots.

The detective's job is to tweak the "Rulebook" just enough so that the computer's prediction matches the real-world photos. But here is the magic: because the detective knows the laws of physics (specifically that air cannot just disappear or appear out of nowhere), the computer is forced to "fill in the blanks" for the parts of the wing where no photos were taken.

How They Did It

1. The Experiment: They put a model wing (NACA 0012) in a wind tunnel and took 2D photos of the air flow at four different spots along the wing's length.
2. The Data: These photos showed that the air separation was different at each spot (some spots had huge bubbles, others had small ones), proving that 3D "stall cells" were present.
3. The Reconstruction: They fed these photos into their computer model. The model adjusted its internal "knobs" (mathematical corrections to the turbulence) to match the photos.
4. The Result: Even though they only gave the computer data from one or two slices, the computer successfully reconstructed the entire 3D structure of the stall cells.

Key Findings (The "Aha!" Moments)

• One Slice is Enough (Sort of): Surprisingly, feeding the computer data from just one slice of the wing was enough to recover the essential features of the stall cells, including the swirling vortices.
• The Sweet Spot: The best results came when they used two slices that were close together but showed very different behavior (one with a huge separation bubble, one with a small one). This gave the computer a clear "before and after" picture of how the air was changing rapidly, allowing it to build a very sharp, detailed 3D model.
• The Anchor: The researchers found that the "anchor" of these 3D cells (where the swirling starts near the wing tip) was always in the same spot, regardless of which photos they used. This suggests that the physical boundary of the wing (the splitter plate) acts like a magnet, holding the cell in place, while the photos help define the rest of the shape.
• The Missing Piece: The computer figured out the "missing" sideways movement of the air (which wasn't in the photos) by strictly following the law of continuity (air must flow smoothly). This allowed the 2D photos to magically expand into a full 3D picture.

The Bottom Line

The paper proves that you don't need a massive, expensive 3D scan to understand complex 3D airflow. If you have a good physics model and just a few smartly placed 2D snapshots, you can mathematically "grow" the full 3D picture.

In the authors' own words, they successfully answered the question: "To stall-cell or not to stall-cell?" Yes, by using this method, they could reconstruct the stall cells from sparse data, revealing the hidden 3D "mushroom" structures that standard computer models missed.

Technical Summary

1. Problem Statement

The reconstruction of full-field three-dimensional (3D) turbulent flows from sparse experimental measurements remains a significant challenge, particularly for flows exhibiting complex 3D separation such as stall cells.

• Experimental Limitations: Planar Particle Image Velocimetry (PIV) provides only 2-component (2D) velocity data on single planes. It lacks the out-of-plane velocity component, making the data inherently non-divergence-free, and often captures only a small portion of the flow volume.
• Computational Limitations: Direct Numerical Simulation (DNS) and Large Eddy Simulation (LES) are too computationally expensive for high Reynolds number practical flows. Reynolds-Averaged Navier-Stokes (RANS) simulations are computationally tractable but often fail to accurately predict the onset, location, and 3D topology of stall cells (spanwise coherent structures) without data correction.
• The Gap: There is a lack of methods capable of reconstructing the full 3D physics of stall cells using only sparse, 2-component planar experimental data, validated on planes not used in the assimilation.

2. Methodology

The authors employ 3D Variational Data Assimilation (3DVar DA) within a Field Inversion framework to correct a baseline RANS model using sparse experimental data.

• Experimental Setup:

  • Geometry: NACA 0012 wing with a 0.75 m span and 0.3 m chord.
  • Conditions: Reynolds number $Re_c \approx 450,000$, Angle of Attack $\alpha = 14^\circ$.
  • Data Acquisition: Planar PIV was performed on four spanwise planes ($z/c = 1.1, 0.9, 0.71, 0.52$). The data provided 2-component mean velocity fields ($U_x, U_y$).
  • Validation: Statistical stationarity was confirmed via bootstrap resampling, and spanwise coherence was analyzed using Proper Orthogonal Decomposition (POD).

• Computational Framework:

  • Baseline Model: The Spalart–Allmaras (SA) RANS turbulence model.
  • Field Inversion: A spatially varying scalar multiplier, $\beta(x, y, z)$, is introduced to the production term of the SA turbulence transport equation:
    $$\frac{D\tilde{\nu}}{Dt} = \beta(x, y, z)P(\tilde{\nu}, w) + T(\nabla\tilde{\nu}, w) - D(\tilde{\nu}, w)$$
  • Optimization: The goal is to minimize the discrepancy between the RANS solution and experimental data by optimizing $\beta$.
    • Objective Function: $J = \frac{1}{2} \| Q(u, \beta) - \tilde{Q} \|^2$, where $\tilde{Q}$ is the experimental data and $Q$ is the projection operator.
    • Solver: The discrete adjoint method (using DAFoam) computes sensitivities, and Sequential Quadratic Programming (SQP) via SNOPT performs the optimization.
  • Key Mechanism: Although experimental data is only 2D and sparse, the continuity equation ($\nabla \cdot \mathbf{U} = 0$) in the 3D RANS solver forces the development of the missing spanwise velocity component ($U_z$) to satisfy mass conservation, thereby reconstructing the full 3D flow field.

3. Key Contributions

1. Full-Field 3D Reconstruction from Sparse 2D Data: The study demonstrates that a full 3D mean flow field, including complex stall cell structures, can be reconstructed using data from as few as one or two spanwise planes of 2-component PIV data.
2. Validation on Unseen Data: Unlike many data-driven studies, the reconstruction quality was assessed on spanwise planes that were not used in the assimilation process, proving the method's ability to generalize corrections across the domain.
3. Insight into Data Placement: The paper quantifies how the number and placement of reference planes affect the reconstruction. It identifies that two planes close together but with markedly different separation extents yield the most compact and accurate stall cell reconstruction.
4. Role of Boundary Conditions: The study highlights the complementary role of computational boundary conditions (specifically the symmetry plane/splitter plate) in anchoring the stall cell structure, working in tandem with sparse experimental data.

4. Key Results

• Baseline Failure: The baseline SA model failed to predict stall cells, showing a uniform separation line and lacking the characteristic counter-rotating vortices, despite showing weak 3D effects.
• Reconstruction of Stall Cells:
  • The assimilated flows successfully recovered the counter-rotating streamwise vortices and focal points on the suction surface, characteristic of stall cells.
  • Spanwise Wavelength: The reconstructed stall cells exhibited wavelengths ($\lambda_z$) between $1.3c$ and $1.6c$, consistent with literature values.
  • Vortex Topology: Volumetric streamlines revealed an "arch-like" or inverted U-shaped vortical structure connecting the surface focal points, consistent with previous experimental observations of stall cells.
• Performance of Assimilation Cases:
  • Single-Plane Cases: Using a single plane (e.g., $z/c=0.71$ or $0.52$) improved the flow field on non-assimilated planes. However, the outboard vortex core location varied depending on the data plane's proximity to the boundary.
  • Dual-Plane Cases: The case using two planes ($z/c=1.1$ and $0.9$) that were close in space but had different separation extents produced the most robust and compact stall cell structure.
  • Inboard Anchor: Across all cases, the inboard focal point of the stall cell consistently formed at $z/c \approx 0.5\text{--}0.6$, demonstrating that the symmetry boundary condition effectively anchors one end of the structure regardless of data placement.
• Error Reduction: Significant reductions in the $L_1$ norm error were observed on both reference and non-reference planes, particularly in the shear layers and recirculation regions.

5. Significance

This work bridges the gap between sparse experimental measurements and high-fidelity 3D flow physics. It proves that variational data assimilation can overcome the limitations of planar PIV (lack of 3D data and divergence-free constraints) by leveraging the governing equations of fluid dynamics.

The findings are significant for:

• Aerodynamic Design: Providing a cost-effective method to characterize complex 3D separation phenomena (stall cells) without requiring expensive volumetric measurements (like Tomographic PIV) or prohibitive DNS.
• Model Correction: Validating the Field Inversion and Machine Learning (FIML) pipeline for correcting turbulence models in 3D separated flows.
• Experimental Strategy: Offering guidelines for experimentalists, suggesting that strategic placement of a few measurement planes (specifically those capturing varying separation states) is sufficient to reconstruct the global 3D flow topology.

In conclusion, the paper demonstrates that "to stall-cell or not to stall-cell" is a question that can be answered by combining minimal experimental data with variational data assimilation, successfully recovering the full 3D physics of stall cells from sparse 2D inputs.