Improving boundary-layer separation prediction by an IDDES turbulence model using a pressure-gradient sensor

This paper extends a pressure-gradient sensor from RANS to the IDDES turbulence model to improve boundary-layer separation prediction by reducing eddy viscosity and disabling the elevation term in adverse pressure-gradient regions, resulting in enhanced accuracy for stall onset and post-stall regimes across various airfoils without compromising attached-flow performance.

The Gist

Imagine you are trying to predict how a car wing (an airfoil) behaves when it's driving through the air. Sometimes the air flows smoothly over the wing, giving you lift (like a bird soaring). But if you tilt the wing too much, the air gets confused, rips away from the surface, and the wing "stalls," causing a sudden loss of lift and a lot of drag.

For decades, computer models have struggled to predict exactly when this stall happens and what happens after it. It's like trying to predict exactly when a stack of Jenga blocks will topple.

Here is a simple breakdown of what this paper did to fix that problem, using some everyday analogies.

The Problem: The "Two-Left-Shoes" Dilemma

Scientists have two main types of computer models for fluid flow:

1. The "Smooth Operator" (RANS): This model is great at predicting smooth, attached airflow (like a car cruising on a highway). But when the air starts to get messy and separate, this model gets confused and thinks the air is still sticking to the wing, even when it's not. It misses the stall.
2. The "Chaos Detective" (IDDES): This model is great at handling messy, chaotic, separated airflow (like a car spinning out of control). It sees the turbulence perfectly. However, it has a weird blind spot: it often refuses to let the air separate in the first place when the pressure gets high. It thinks the air is still glued to the wing, so it misses the start of the stall.

The Goal: The authors wanted to create a "Super Model" that is a Smooth Operator when things are calm, but instantly becomes a Chaos Detective when things get messy, without missing the moment the trouble starts.

The Solution: A "Pressure Sensor" and a "Brake"

The authors took a tool they had previously built for the "Smooth Operator" model and tried to install it into the "Chaos Detective" model.

1. The Pressure Sensor (The Smoke Detector)
They added a special sensor that acts like a smoke detector for air pressure.

• How it works: When the air pressure starts pushing back against the flow (an "adverse pressure gradient"), the sensor screams, "Hey! The air is about to peel off!"
• The Fix: When the sensor goes off, the model automatically turns down the "stickiness" of the air (eddy viscosity). Imagine the air is like honey; normally, the honey is thick and holds the flow together. The sensor tells the honey to turn into water, making it easier for the air to peel away from the wing. This helps the model predict the stall onset correctly.

2. The Hidden Brake (The Elevation Term)
Here is where it got tricky. The "Chaos Detective" model (IDDES) has a built-in safety feature designed to keep the air glued to the wing in smooth areas. It's like a hidden brake that prevents the model from getting too chaotic too early.

• The Conflict: When the "Smoke Detector" screamed "Peel off!", the "Hidden Brake" kicked in and said, "No, stay glued!" The two features fought each other, and the air refused to separate.
• The Fix: The authors realized they had to turn off the Hidden Brake specifically in the areas where the Smoke Detector was screaming. They didn't turn it off everywhere (because that would ruin the smooth flight predictions); they only turned it off where the pressure was high.

The Result: A Unified Model

By combining the Smoke Detector (to sense trouble) and the Selective Brake Release (to allow the trouble to happen), they created a unified model.

• Before: The old models were like a driver who either never hit the brakes (missing the crash) or hit them too hard (crashing too early).
• Now: The new model is like a perfect driver. It cruises smoothly, hits the brakes exactly when the road gets slippery, and handles the skid perfectly once it starts.

Testing the New Model

They tested this new "Super Model" on five different wing shapes, ranging from thin racing wings to thick wind-turbine blades. They simulated everything from gentle breezes to extreme angles where the wing is almost vertical.

• The Good News: The new model predicted the lift and drag (the forces on the wing) much better than the old models. It got the "stall" moment right, and it handled the messy "post-stall" turbulence perfectly.
• The One Catch: On some very thick wings at low speeds, the model got a little too eager to separate the air, predicting a stall slightly earlier than reality. The authors admit this isn't a flaw in their new "Super Model" logic, but rather that the "Smoke Detector" itself needs a little fine-tuning for those specific conditions.

Why This Matters

This is a big deal for engineers designing wind turbines, airplanes, and cars. Currently, they often have to run two different, expensive computer simulations to get the full picture: one for smooth flight and one for stalled flight.

This paper shows that we can now use one single model to predict the entire flight envelope—from a gentle glide to a violent stall. It's like finally having a single map that works for both driving on a highway and navigating a muddy off-road trail.

Technical Summary

Here is a detailed technical summary of the paper "Improving boundary-layer separation prediction by an IDDES turbulence model using a pressure-gradient sensor."

1. Problem Statement

Computational fluid dynamics (CFD) faces a persistent challenge in reliably predicting flow separation across all regimes, particularly for aerodynamic bodies like airfoils.

• RANS Limitations: State-of-the-art Reynolds-Averaged Navier-Stokes (RANS) models, even with pressure-gradient corrections, often fail to accurately predict post-stall and deep-stall regimes where flow is dominated by three-dimensional, unsteady phenomena (e.g., vortex shedding).
• Hybrid RANS/LES Limitations: While hybrid models like Improved Delayed Detached Eddy Simulation (IDDES) excel at capturing unsteady, massively separated flows (deep stall) and attached flows, they frequently fail to predict pressure-induced smooth-body separation (stall onset). They tend to keep the flow attached too long, leading to inaccurate predictions of stall angles and post-stall lift/drag.
• The Gap: There is a lack of a unified turbulence model that can accurately predict attached flow, stall onset, post-stall, and deep-stall regimes simultaneously without significant degradation in any specific regime.

2. Methodology

The authors propose extending a pressure-gradient sensor, originally developed for RANS models by Griffin et al., to the IDDES framework. The methodology involves two critical modifications to the baseline IDDES model (based on the $k-\omega$ SST formulation):

A. Pressure-Gradient Sensor (APG Sensor)

The model utilizes a non-dimensional parameter, $\lambda_\theta$, to identify regions of strong adverse pressure gradients (APG) sufficient to induce separation.

• Logic: When $\lambda_\theta$ falls below a critical threshold ($\lambda_{\theta,c}$), the sensor activates.
• Transition Handling: The threshold is blended between laminar and turbulent criteria using the intermittency factor ($\gamma$), allowing the model to function in both fully turbulent and transitional flow variants.

B. Model Modifications

Once the sensor is active, two specific changes are applied to the IDDES equations:

1. Eddy-Viscosity Reduction: The coefficient $a_1$ in the eddy-viscosity ($\mu_t$) formulation is reduced (from 0.31 to 0.275). This reduces momentum transport toward the wall, promoting separation.
2. Suppression of the Elevation Term ($f_e$):
   • The Issue: The baseline IDDES model includes an "elevation term" ($f_e$) in the length scale formulation designed to augment Reynolds stresses and mitigate "log-layer mismatch" in attached flows.
   • The Conflict: This $f_e$ term actively counteracts the eddy-viscosity reduction, preventing the flow from separating even when the APG sensor is active.
   • The Solution: The authors propose setting $f_e = 0$ locally only in regions where the APG sensor is active. This allows the flow to separate where needed while preserving the accuracy of the $f_e$ term in attached flow regions.

C. Computational Setup

• Solver: ExaWind (Nalu-Wind), solving 3D incompressible unsteady Navier-Stokes equations.
• Test Cases: Five airfoils (S809, NACA0012, NACA0021, DU91-W2-250, DU00-W-212) representing wind energy and aerospace applications.
• Conditions: Angles of attack (AOA) from 0° to 90°, covering Reynolds numbers from $3.9 \times 10^5$to$3.0 \times 10^6$. Both fully turbulent and transitional ($\gamma$-based) model variants were tested.

3. Key Contributions

1. Extension to IDDES: Successfully adapted the APG sensor from pure RANS to the IDDES framework, enabling detection of APG-induced separation in both fully turbulent and transitional simulations.
2. Identification of Counteracting Mechanisms: Demonstrated that eddy-viscosity reduction alone is insufficient for IDDES due to the competing effect of the elevation term ($f_e$). The paper proves that locally deactivating $f_e$ is crucial for inducing separation.
3. Unified Modeling: Developed a single model capable of accurately predicting attached flow, stall onset, post-stall, and deep-stall regimes, overcoming the specific failure modes of both baseline RANS and baseline IDDES.
4. Generalization: Validated the model across diverse airfoils, Reynolds numbers, and transition types without recalibrating model coefficients, proving the robustness of the approach.

4. Results

• Stall Onset & Post-Stall: The proposed IDDESae model (APG sensor + $f_e$ suppression) significantly improved the prediction of stall onset and post-stall lift/drag polars compared to the baseline IDDES. It correctly captured the drop in lift and rise in drag associated with separation at moderate angles of attack (e.g., 10°–30°).
• Deep Stall: The model maintained the high accuracy of the baseline IDDES in deep-stall regimes (AOA > 30°), where flow is highly unsteady and 3D.
• Attached Flow: The model did not significantly degrade attached-flow predictions (low AOA), although minor discrepancies were noted for specific low-Reynolds number, thick airfoils.
• Flow Physics: Visualizations (Q-criterion) confirmed that the modified model correctly predicted leading-edge separation and the formation of spanwise roller vortices, whereas the baseline IDDES incorrectly predicted attached flow with trailing-edge shedding.
• Limitations: For low-Reynolds number, thick airfoils (e.g., NACA0021), the APG sensor was found to activate slightly too early, causing a minor underprediction of lift in the linear regime. The authors attribute this to the underlying sensor formulation (a known RANS issue) rather than a failure of the IDDES extension.

5. Significance

This work addresses a longstanding limitation in computational aerodynamics by providing a unified turbulence modeling approach.

• Practical Impact: Engineers can now use a single IDDES-based model to simulate the entire flight envelope of an airfoil (from takeoff to deep stall) with high fidelity, eliminating the need to switch between RANS (for attached flow) and LES/DES (for separated flow) or accept poor stall predictions.
• Cost-Effectiveness: The model offers a more affordable alternative to Wall-Resolved LES or DNS while providing superior accuracy to RANS for separated flows.
• Future Direction: The study highlights that future improvements should focus on refining the pressure-gradient sensor itself to better handle low-Reynolds number flows, rather than further modifying the IDDES coupling logic.