Effect of Turbulence-Closure Consistency on Airfoil Identification

This paper demonstrates that reliable airfoil shape identification from wake signatures requires turbulence-closure consistency, as single-condition inversions are ill-posed and model inconsistencies can cause divergent geometric predictions and up to 250% variation in sensitivities, necessitating that turbulence models prioritize physically consistent sensitivities alongside accuracy.

The Gist

The Big Idea: Guessing a Shape from Its Shadow

Imagine you are in a dark room. You can't see a sculpture, but you can see the shadow it casts on the wall when a light shines on it. Your job is to figure out exactly what the sculpture looks like just by studying that shadow.

In the world of engineering, this is called Inverse Design. Instead of designing a wing and then testing how the wind blows around it, engineers try to work backward: "We know how the wind is blowing behind the wing (the wake); what shape must the wing be to create that pattern?"

This paper, written by researchers at Brown University, tackles two major problems with this "shadow guessing" game.

---

Problem 1: The Shadow is a Bad Clue (The "Ill-Posed" Problem)

The Analogy:
Imagine you see a shadow on the wall that looks like a long, thin rectangle.

• Could it be a flat sheet of paper held up?
• Could it be a long, thin pipe?
• Could it be a flat board with a hole in it?

All three objects cast almost the same shadow from one specific angle. If you only look at the shadow from one angle, you can't know for sure which object is actually there. You might guess the pipe, but it's actually the sheet of paper.

The Paper's Finding:
The researchers found that if you try to identify an airplane wing using wind data from just one speed or angle, the problem is "ill-posed." There are too many possible answers, and the computer might guess the wrong shape.

The Solution:
They discovered that if you look at the shadow from multiple angles (like walking around the object and looking at it from the front, side, and top), the mystery is solved. By combining wind data from three different angles of attack, the computer can pinpoint the exact shape of the wing with much higher accuracy. It's like having a 3D scanner instead of a 2D camera.

---

Problem 2: The "Translator" is Lying (Turbulence Closures)

This is the most surprising part of the paper. To predict how air moves, computers use mathematical shortcuts called Turbulence Closures. Think of these as "translators" that turn complex, chaotic wind physics into simple math the computer can solve.

There are different translators (models) available, such as the S-A model, the k-ω SST model, and the k-ε model.

The Analogy:
Imagine you are trying to guess a person's height by measuring the length of their shadow.

• Translator A says: "The shadow is 5 feet long, so the person is 6 feet tall."
• Translator B says: "The shadow is 5 feet long, so the person is 10 feet tall."

Both translators agree on the shadow length (the forward prediction), but they disagree wildly on what caused it (the sensitivity).

The Paper's Finding:
The researchers tested these different "translators" to see if they could correctly identify the wing shape.

• When they used the same translator to both create the "shadow" data and guess the shape, they got a result very close to the real wing.
• When they used different translators (e.g., creating data with Translator A but guessing the shape using Translator B), the result was a disaster. The computer guessed a wing shape that was completely wrong—sometimes off by a huge margin (orders of magnitude).

The "Sensitivity" Surprise:
The paper introduces a new concept called Sensitivity Consistency.

• Predictive Accuracy: Does the model predict the wind speed correctly? (Yes, all models were okay at this).
• Sensitivity Consistency: If you change the wing shape by a tiny bit, does the model correctly predict how the wind will change?

The researchers found that even if two models predict the wind speed perfectly, they might disagree on how the wind reacts to a shape change. It's like two doctors who both agree a patient has a fever, but one thinks it's caused by a virus and the other thinks it's caused by a broken leg. If you treat the wrong cause, the patient gets worse.

---

Why This Matters

In the past, engineers only cared if their models could predict the wind speed correctly (Forward Prediction). This paper says: "That's not enough."

If you are designing a new airplane wing using a computer, you need a model that not only predicts the wind correctly but also correctly understands cause and effect. If the model gets the "cause and effect" wrong, your optimization will lead you to build a wing that looks great on the computer but fails in real life.

The Takeaway

1. Don't guess from one angle: To identify a shape from wind data, you need to look at it from many different angles.
2. Check your translator: When using computer models to design shapes, you can't just pick any turbulence model. You must ensure the model is "sensitive-consistent"—meaning it understands how small changes in shape lead to changes in the wind. If the model gets this wrong, your design will be wrong, even if the math looks perfect.

In short: To build the perfect wing, you need the right map (the model) and you need to look at the terrain from every possible direction.

Technical Summary

1. Problem Statement

The paper addresses the inverse identification of airfoil shapes based on sparse flow measurements, specifically the velocity field in the wake of the airfoil. This is formulated as a PDE-constrained optimization problem where the goal is to recover the geometric parameters ($\theta$) of an airfoil such that the simulated flow field ($F_M(\theta)$) matches observed data ($u_{obs}$).

The study highlights two critical challenges in this domain:

1. Ill-posedness: Inverse problems often lack a unique solution. The authors demonstrate that identifying a shape from a single flow condition (e.g., one Angle of Attack) is highly ill-posed, leading to significant reconstruction errors.
2. Model Discrepancy (Inverse Crime): In practical CFD, the forward model ($F_M$) relies on Reynolds-Averaged Navier-Stokes (RANS) equations with turbulence closures (e.g., Spalart–Allmaras, $k-\omega$ SST, $k-\epsilon$). These models are approximations of true physics. If the turbulence closure used in the optimization differs from the one that generated the "true" data (or if the closures themselves are inconsistent), the inferred geometry will be biased, even with noise-free data.

2. Methodology

The authors employ a rigorous computational framework to investigate these issues:

• Forward Simulation:

  • Solver: OpenFOAM (finite-volume method).
  • Governing Equations: Incompressible, steady RANS equations.
  • Turbulence Models: Three distinct closures are tested: Spalart–Allmaras (S-A), $k-\omega$ SST, and $k-\epsilon$.
  • Conditions: Simulations are run at a Reynolds number of $Re = 1 \times 10^5$ across three Angles of Attack (AoA): $0^\circ, 5^\circ, 10^\circ$.
  • Data Source: The "target" wake signatures are generated using the S-A model for a target airfoil (NACA 16021).

• Inverse Optimization Framework:

  • Objective Function: A weighted sum of the $L_2$ error in the wake velocity field, drag coefficient ($C_D$), and lift coefficient ($C_L$).
  • Parameterization: The airfoil geometry is controlled via Free-Form Deformation (FFD) with a $2 \times 10$ grid of control points.
  • Optimization Algorithm: A gradient-based approach using the discrete adjoint method. The gradient $\frac{dJ}{d\theta}$ is computed efficiently using the differentiable solver DAFoam, which employs a Jacobian-free Krylov method and automatic differentiation.
  • Constraints: Geometric constraints are enforced to ensure physical meaningfulness, including symmetry, relative volume, pointwise thickness, leading-edge radius, and surface curvature.

3. Key Contributions

1. Mitigation of Ill-Posedness: The study demonstrates that while single-condition inverse identification is ill-posed, incorporating multiple wake signatures (from different AoAs) significantly alleviates this issue, leading to more accurate shape recovery.
2. Quantification of Closure Inconsistency: The paper provides systematic evidence that using a turbulence closure inconsistent with the data generation model leads to order-of-magnitude differences in the reconstructed airfoil shape.
3. Introduction of Sensitivity Consistency: The authors introduce the concept of sensitivity consistency as a critical criterion for turbulence models. They argue that a model must not only predict flow fields accurately (forward fidelity) but also provide physically consistent gradients (sensitivity) for inverse design.
4. Adjoint Analysis of Model Discrepancy: The study directly compares geometric sensitivities ($\partial F_M / \partial \theta$) across different turbulence models, revealing that models with similar forward predictions can yield drastically different gradients (up to 250% difference).

4. Key Results

• Single vs. Multiple Conditions:

  • Identifying the NACA 16021 airfoil using a single AoA resulted in relative $L_2$ errors ranging from 5.77% to 11.0%, depending on the angle.
  • Using multiple AoAs ($0^\circ, 5^\circ, 10^\circ$) reduced the error to 3.15%, confirming that multi-condition data constrains the solution space effectively.

• Impact of Turbulence Closure Consistency:

  • When the optimization used the consistent closure (S-A, same as data generation), the shape error was 3.15%.
  • When using inconsistent closures:
    • $k-\omega$ SST: Resulted in a 16.5% error.
    • $k-\epsilon$: Resulted in a 21.7% error.
  • This indicates that closure inconsistency can degrade reconstruction accuracy by more than an order of magnitude compared to the consistent case.

• Sensitivity Analysis:

  • For a fixed geometry perturbed by $\pm 10\%$, the three turbulence models predicted wake mismatch ($\Delta U^2$) and aerodynamic coefficients with significant variance.
  • Crucially, the adjoint gradients (sensitivity of the objective to shape changes) differed by 70% to 250% between models.
  • In some regions, the inconsistent models predicted gradients with opposite signs compared to the reference S-A model, implying that the optimization would move the shape in the wrong direction, leading to fundamentally different (and incorrect) recovered geometries.

5. Significance and Implications

• Redefining Model Accuracy: The paper challenges the traditional view that turbulence models should be evaluated solely on forward predictive accuracy (matching mean flow or forces). It argues that for inverse design and optimization, sensitivity consistency is equally vital.
• Guidance for Data-Driven Modeling: As the field moves toward data-driven or hybrid turbulence models, the authors suggest that these models must be trained and validated not just on flow fields, but also on their adjoint behavior (gradients).
• Uncertainty Quantification: The findings highlight the need for new uncertainty quantification frameworks that explicitly account for model-form errors in inverse problems.
• Practical Design: For engineers using CFD for shape optimization, the choice of turbulence closure is not merely a numerical detail; it is a fundamental determinant of the final design's validity. Using a closure that is inconsistent with the physical reality (or the data source) can lead to "predictively correct but inversely inconsistent" designs.

In conclusion, this work establishes that turbulence-closure consistency is a prerequisite for reliable aerodynamic shape identification. It proposes a new paradigm where turbulence models are assessed by their ability to provide consistent sensitivities, ensuring that optimization trajectories lead to physically correct solutions.