DDES (Delayed Detached Eddy Simulation) Turbulence Model

DES (Detached Eddy Simulation) uses RANS in attached boundary layers and LES in separated wakes — a cost-effective hybrid that's been the standard for bluff-body aerodynamics since 1997. But the original DES97 had a serious flaw: if the mesh was fine enough in the boundary layer region, it could accidentally trigger LES mode there. When that happens, turbulent eddy viscosity drops dramatically — a phenomenon called Modeled Stress Depletion (MSD) — and the boundary layer separates prematurely. DDES, proposed by Spalart, Deck, Shur, Squires, Strelets, and Travin in 2006, adds a "delaying function" $f_d$ that explicitly keeps the model in RANS mode inside the boundary layer based on flow physics, not just mesh size. This single addition dramatically improves robustness on fine meshes.

On a car body or aircraft wing, the boundary layer should stay attached along the sides and lifting surfaces. If MSD triggers false separation there, you see excessive drag — drag coefficients 20–50% higher than wind tunnel data — incorrect lift predictions, and acoustic noise that doesn't match reality. The effect is especially pronounced on fine meshes at high Reynolds numbers, exactly the conditions you'd choose for a high-accuracy simulation. Ironically, refining the mesh made DES97 less accurate — a deeply counterintuitive behavior that motivated the development of DDES.

In DES97, the turbulence length scale switches based on:

$d_w$ is wall distance, $\Delta = \max(\Delta x, \Delta y, \Delta z)$ is the local mesh size, and $C_{DES} \approx 0.65$ for the Spalart-Allmaras base model. When the mesh is fine enough that $C_{DES}\Delta < d_w$ inside the boundary layer, the model switches to LES mode unintentionally. The LES SGS eddy viscosity is much lower than RANS eddy viscosity — perhaps by a factor of 10–100 — so turbulent Reynolds stresses are depleted. The boundary layer then behaves as if the effective Reynolds number were much lower, causing early separation. This is Modeled Stress Depletion: the model literally loses the modeled stress that was holding the boundary layer attached.

DDES introduces the delaying function $f_d$ to modify the length scale:

where:

Here $\nu_t$ is the turbulent kinematic viscosity, $\nu$ is the molecular kinematic viscosity, $U_{i,j}$ are the velocity gradient components, and $\kappa = 0.41$ is the von Kármán constant. Inside the boundary layer, the velocity gradient $|\partial u/\partial y|$ is large, making $r_d \approx 1$, so $f_d \approx 0$ and $\tilde{d}_{DDES} \approx d_w$ — RANS mode is preserved. In the separated wake, velocity gradients are small, $r_d \to 0$, $f_d \approx 1$, and the length scale reverts to $C_{DES}\Delta$ — LES mode activates. The transition is physically based and smooth, completely independent of mesh resolution in the boundary layer.

Here's the practical tradeoff for automotive external aerodynamics at $Re = 10^6$:

Yes — the LES region has much tighter requirements than the RANS region. A common mistake is using the same numerical settings for DDES as for steady RANS:

• Convection scheme: Pure upwind is fine for RANS, but the LES region requires central differencing (Linear) or a bounded central scheme (Bounded Central Differencing in OpenFOAM, Central Differencing in Fluent) to minimize numerical dissipation. Excessive upwinding in the LES region kills resolved turbulence by over-diffusing the vortical structures.
• Time integration: DDES is inherently unsteady. Use second-order or higher temporal schemes (Crank-Nicolson with off-centering coefficient 0.9, or 2nd-order backward Euler in OpenFOAM).
• Time step: Keep Courant number Co ≤ 0.5–1.0 in the LES region. Use adaptive time stepping.
• Blend factor: In OpenFOAM, the limitedLinear or blended (linearUpwindV) scheme can be set to blend central/upwind based on local cell quality — useful near low-quality mesh regions.

The grey area is the transition zone where the model switches from RANS to LES. In RANS mode, turbulence is fully modeled (high eddy viscosity). When the model switches to LES mode, the modeled eddy viscosity drops — but resolved turbulent fluctuations don't immediately appear to replace it, because the resolved turbulent kinetic energy in the LES region is essentially zero at the transition point. This results in an unphysical "dead zone" where both modeled and resolved turbulence are low — the turbulence is too low on both accounts. The consequences: incorrect reattachment location, underestimated Reynolds stresses in the near-wake, and inaccurate force coefficients. IDDES (Improved DDES) addresses this by integrating a synthetic turbulence injection mechanism at the RANS-LES interface. The SEM (Synthetic Eddy Method) seeds turbulent fluctuations at the interface, bridging the transition and reducing grey area errors significantly.

Automotive aero is the most common DDES application. Here's the standard workflow:

1. Mesh design: y+ ≈ 1 with fine prism layers on all body surfaces. Extra refinement on the A-pillar, door mirrors, underbody diffuser, and tail — areas prone to separation. Wake refinement box extending 5–10 body lengths behind the vehicle with near-isotropic cube cells.
2. RANS initialization: First run a converged steady-state RANS (k-ω SST) to get a reasonable initial field. Run for 1000–2000 iterations or until drag converges.
3. DDES restart: Restart the DDES unsteady simulation from the RANS solution. The warm start avoids the long initial transient from a uniform field initialization. Switch convection to bounded central and time to 2nd-order backward.
4. Initial transient: Discard the first 2–5 flow-through times ($L/U$) of DDES results as the resolved turbulence develops from the initial RANS field. This is the "spin-up" phase.
5. Statistical sampling: Average over at least 5–10 $L/U$ flow-through times to get converged mean drag, lift, and pressure coefficients. For pressure spectra and acoustic analysis, longer averaging is needed.
6. fd field check: Visualize $f_d$ to verify boundary layer protection. Inside attached boundary layers, $f_d$ should be $\approx 0$.

Production automotive DDES meshes typically range from 50M to 500M cells depending on the level of detail. A 1:4 scale wind-tunnel-equivalent simulation (quarter-scale model, 10 m domain) with DDES typically uses 80–150M cells. Full-scale vehicle with detailed underbody and wheel well resolution: 200–500M cells. At 500M cells on a 1000-core compute cluster, a single DDES run for 5 flow-through times takes approximately 48–72 hours of wall time. This cost explains why most OEM programs run DDES on specific design candidates rather than in the design exploration phase.

Here's a comparison of DDES support across major platforms:

The DES family has continued to evolve rapidly after DDES:

• IDDES (Improved DDES, Shur et al. 2008): Integrates Wall-Modeled LES (WMLES) capability into the DDES framework. IDDES automatically switches between RANS-shielded DES mode (for attached BL) and WMLES mode (for resolved BL with coarse near-wall mesh), and seeds synthetic turbulent fluctuations at the RANS-LES interface. Currently the recommended variant for most industrial applications.
• XLES (eXtra-LES): A more general RANS/LES blending framework with a user-tunable blending function, allowing more aggressive LES activation in specific regions.
• SLA (Stress-Limiting Approach, 2023): A recent DDES variant that limits the modeled stress in the RANS region based on the local strain rate, improving predictions for cases with strong mean shear in the LES region.
• ML-based RANS/LES switching: Neural networks trained on DNS/LES databases predict the optimal blending coefficient $f_d$ based on flow features, replacing the empirical tanh function with a data-driven alternative.

Systematic diagnostic steps:

1. Visualize the $f_d$ field: Inside attached boundary layers, $f_d$ should be close to 0. If you see $f_d \approx 1$ inside the boundary layer, the delaying function is failing — this is the primary indicator of residual MSD.
2. Check prism layer resolution: DDES needs y+ ≈ 1 and sufficient prism layers to resolve the boundary layer in RANS mode. If y+ is too high (wall functions mode), the $r_d$ sensor may not correctly detect the boundary layer region.
3. Verify RANS warm-up is adequate: If you started DDES from a poor initial field (e.g., zero velocity), the RANS initialization may not have converged before switching to DDES, causing incorrect initial $f_d$ distribution.
4. Check turbulence model constants: For k-ω SST-DDES, ensure the SST blending coefficients ($F_1$, $F_2$) are behaving correctly near the wall. Non-standard turbulence model setups can interfere with the $r_d$ sensor.
5. Switch to IDDES: If premature separation persists despite correct $f_d$ in the boundary layer, the grey area problem may be contributing. IDDES with synthetic turbulence injection significantly reduces this effect.