Three-dimensional variational data assimilation of separated flows using time-averaged experimental data

This paper presents a novel three-dimensional variational data assimilation framework that integrates planar PIV experimental data with the Spalart-Allmaras RANS model to effectively separate measurement errors from turbulence model deficiencies, thereby significantly improving flow predictions for separated flows over a NACA0012 airfoil compared to traditional two-dimensional methods.

The Gist

Imagine you are trying to bake the perfect cake, but your recipe (the computer model) keeps coming out slightly wrong. You also have a photo of a real cake (experimental data) that you want to match. The problem is, the photo is a bit blurry, missing some ingredients, and taken from a weird angle.

This paper is about a new way to fix the recipe so it matches the real cake better, even when the photo isn't perfect.

The Problem: The "Flat" Photo vs. The "Round" Reality

Scientists use computer models to predict how air moves over airplane wings. These models are like recipes. Sometimes, the recipe is a bit off, especially when the wing is in a "deep stall" (a situation where the wing stops working well, like a plane stalling in the sky).

To fix the recipe, scientists use a technique called Data Assimilation. They take real-world measurements (like a photo of the airflow) and force the computer model to match it.

However, there's a catch. The real-world measurements come from a technique called PIV (Particle Image Velocimetry), which takes a 2D "slice" or a flat photo of the air. But the air moving around a wing is actually 3D (it moves up, down, left, right, and also in and out of the photo).

The paper argues that previous methods tried to force this flat, 2D photo to fit a 3D reality by pretending the air only moves in two directions. This is like trying to fit a round orange into a square hole; you have to squeeze and distort it to make it fit.

The Old Way: The "Squeezed Orange" (2D Assimilation)

In the old method (called 2DVar), scientists took the flat photo and forced the computer model to obey the rules of a 2D world.

• The Analogy: Imagine the computer model is a student trying to solve a math problem. The teacher (the real data) gives them a blurry, slightly wrong answer. The student tries to change their own answer to match the teacher's.
• The Mistake: Because the teacher's answer was taken from a flat photo of a 3D world, it has "errors" (it doesn't balance perfectly). The student, trying to match this, starts changing their math in weird ways. They blame their own bad math for the teacher's blurry photo.
• The Result: The "correction" the student makes is huge and messy. It fixes the math errors and tries to fix the fact that the photo was flat. You can't tell what part of the correction was fixing the model and what part was just trying to fix the bad photo.

The New Way: The "3D Glasses" (3D Assimilation)

The authors of this paper invented a new method (called 3DVar). Instead of forcing the air to stay flat, they let the computer model breathe in the third dimension (the depth), even though the photo only shows a flat slice.

• The Analogy: Now, the student is wearing 3D glasses. They know the teacher's photo is just a slice of a 3D object. When the photo looks "unbalanced" (divergent), the student realizes, "Ah, the air must be moving into or out of the photo to make this balance!"
• The Fix: The computer model allows the air to move in that third direction. This naturally fixes the "unbalanced" parts of the photo without needing to force the math to break.
• The Result: The "correction" the student makes is now much smaller and cleaner. It only fixes the actual mistakes in the recipe (the turbulence model), not the flaws in the photo.

What They Found

They tested this on a NACA0012 airfoil (a specific wing shape) at a high speed where the air separates and swirls chaotically.

1. The Old Way (2D): The computer had to make massive, confusing changes to the physics equations to match the flat photo. It couldn't tell if it was fixing the model or just compensating for the missing 3D data.
2. The New Way (3D): The computer made smaller, smarter adjustments. It let the air flow naturally in 3D to balance the equations.
3. The Outcome: The new method predicted the lift (how much the wing pushes up) and the pressure on the wing much more accurately. It also gave a better picture of the "turbulence" (the swirling chaos) because the model wasn't being forced to do impossible things just to match a flat photo.

The Bottom Line

Think of it like this: If you try to describe a 3D sculpture using only a 2D shadow, you'll get confused. If you force a 2D drawing to look like that shadow, you'll have to distort the drawing until it looks nothing like the real sculpture.

This paper shows that if you let your drawing have depth (3D), even if you only have a 2D shadow to look at, you can reconstruct the real sculpture much more accurately. The computer model stops fighting the data and starts actually understanding the physics of the flow.

Technical Summary

Technical Summary: Three-dimensional variational data assimilation of separated flows using time-averaged experimental data

Problem Statement
The paper addresses the limitations of applying variational data assimilation (DA) to complex, separated flows using planar Particle Image Velocimetry (PIV) data. While PIV provides valuable two-component, time-averaged velocity fields, it inherently lacks the out-of-plane velocity component. In separated flows, such as deep stall over an airfoil, the flow is inherently three-dimensional (3D). Traditional DA approaches often enforce two-dimensional (2D) continuity constraints on this 2D data. The authors argue that this conflation leads to a misinterpretation of the corrective forcing term used in the momentum equations. Specifically, when 2D constraints are applied to non-divergence-free experimental data, the corrective forcing term compensates not only for deficiencies in the Reynolds-Averaged Navier–Stokes (RANS) turbulence model but also for the lack of the third velocity component and experimental noise. This makes it difficult to isolate physical modeling errors from measurement inconsistencies and hinders the accurate reconstruction of second-order statistics like Reynolds shear stress.

Methodology
The study employs a variational data assimilation framework using a discrete adjoint method implemented via the DAFoam solver. The core of the methodology involves optimizing a corrective forcing term ($f_{ci}$) added to the momentum equations to minimize the discrepancy between a RANS simulation (using the Spalart–Allmaras turbulence model) and reference data.

The paper compares two distinct setups:

1. 2DVar (Two-Dimensional Assimilation): The governing equations and continuity constraints are enforced strictly in two dimensions (streamwise and wall-normal). The spanwise domain is treated as a single cell layer ($N_z=1$), effectively forcing the flow to be 2D.
2. 3DVar (Three-Dimensional Assimilation): The authors propose a novel approach where the governing equations and constraints incorporate all three spatial dimensions. The computational domain is extended in the spanwise direction with sufficient resolution ($N_z > 1$) to allow the development of an out-of-plane velocity component. Crucially, the reference data remains 2D2C (two components, two dimensions), projected onto the 3D grid.

The objective function minimizes the $L_1$ norm of the velocity discrepancy between the simulation and the reference data (both synthetic and experimental). The optimization is performed using Sequential Quadratic Programming (SQP) with gradients computed via the discrete adjoint approach.

Key Contributions

• Novel 3D Assimilation Framework: The paper introduces a 3D variational DA method that enforces 3D continuity constraints while assimilating 2D2C experimental data.
• Decoupling of Errors: The study demonstrates that 3D constraints allow the corrective forcing term to primarily address experimental setup errors (e.g., divergence due to missing 3D data or noise), while the turbulence model is free to capture the underlying flow physics more accurately.
• Reconstruction of Unmeasured Quantities: The method is shown to successfully reconstruct physical quantities not directly measured, such as pressure fields and lift forces, as well as modeling quantities like Reynolds shear stress and eddy viscosity.

Results
The methodology was tested on a NACA0012 airfoil in deep stall ($\alpha = 15^\circ$, $Re_c \approx 7.5 \times 10^4$) using both synthetic data (satisfying 2D continuity) and experimental PIV data (violating 2D continuity).

• Synthetic Data: In the 2DVar case with synthetic, divergence-free data, the corrective forcing was small compared to the Reynolds forcing, indicating the baseline model could capture the physics with minor adjustments.
• Experimental Data (2DVar): When applied to experimental data, the 2DVar approach required a large corrective forcing term. The magnitude of this forcing was comparable to the Reynolds forcing, suggesting it was compensating for the 3D nature of the flow and divergence errors rather than just turbulence modeling deficiencies. This led to an inaccurate representation of the flow physics.
• Experimental Data (3DVar): The 3DVar approach successfully developed a spanwise velocity component to satisfy 3D continuity. Consequently, the magnitude of the corrective forcing term decreased significantly, becoming more comparable to the Reynolds forcing. This indicates that the 3D constraints allowed the turbulence model to function correctly, with the forcing term handling experimental artifacts.
• Reconstructed Quantities: The 3DVar approach yielded superior results for reconstructed variables. The pressure coefficient ($C_p$) distribution and lift coefficient ($C_L$) showed better agreement with experimental measurements compared to 2DVar. Furthermore, the reconstructed Reynolds shear stress and eddy viscosity fields in 3DVar aligned more closely with experimental observations, showing enhanced momentum transport and turbulence effects in the wake.

Significance and Claims
The paper claims that enforcing 2D continuity on inherently 3D separated flows during data assimilation distorts the balance of the momentum equations, forcing the corrective term to absorb errors that should be attributed to the flow's 3D nature. By implementing 3D constraints, the authors demonstrate that the baseline RANS model (Spalart–Allmaras) is capable of capturing complex separated flow physics more accurately than previously assumed when the data assimilation framework is physically consistent.

The study concludes that 3DVar is superior to 2DVar for assimilating planar PIV data of separated flows. It resolves the ambiguity in interpreting the corrective forcing term, allowing researchers to distinguish between turbulence model deficiencies and experimental inconsistencies. This approach improves the accuracy of both directly assimilated mean velocities and reconstructed quantities, such as lift and Reynolds stresses, without modifying the functional form of the turbulence model. The authors note that while their implementation used 2D2C data, the framework is designed to handle the complexities of practical, high-Reynolds-number separated flows.