LES vs RANS: when the extra cost actually pays off

What does "resolved" actually mean?

RANS models all turbulence scales. Every eddy, from the energy-containing integral scale down to the Kolmogorov dissipative scale, is replaced by a model. The output is a time-averaged flow field — useful for lift and drag, useless for instantaneous fluctuations.

LES takes the opposite approach: it resolves the large scales (energy-containing eddies, which are geometry-dependent and anisotropic) and only models the small scales (isotropic, universal, handled via a subgrid-scale model). The cutoff is the filter width Delta — typically proportional to the local mesh size. What you see in LES images is actual instantaneous flow, not statistical averages.

This distinction matters profoundly when flow structures interact. Vortex shedding behind a cylinder, acoustic pressure waves radiating from a jet, turbulent mixing between two streams, laminar-to-turbulent transition on a wing — all of these are governed by large, resolved-scale dynamics that RANS averages away. If the physics you care about involves coherent structures, you need LES.

The cost of resolving eddies

LES cost scales with Re^2.4 (Choi & Moin, 2012). This is brutal scaling. A typical RANS mesh for a car body might use 50 million cells and complete in 2 hours on 200 cores. The equivalent wall-resolved LES would need approximately 5 billion cells and run for weeks on thousands of cores.

The bottleneck is the near-wall region. Eddies near the wall become vanishingly small as Re increases, and resolving them requires near-DNS resolution. Wall-resolved LES cost scales with Re^1.8 — slightly better than pure DNS (Re³) but still prohibitive for industrial Re above 10⁶.

Wall-modeled LES (WMLES) solves this by using a RANS-like model in the inner layer of the boundary layer, reducing near-wall resolution requirements by 10 to 100 times. Instead of resolving the viscous sublayer and buffer layer, WMLES applies a wall model that provides the correct wall shear stress based on the velocity at the first off-wall grid point. For practical industrial LES at high Re, WMLES is the only feasible approach.

When RANS works perfectly

RANS is not obsolete. For many important engineering problems, it is not just adequate — it is the right tool:

• Attached boundary layers on streamlined bodies, where turbulence is in local equilibrium and the Boussinesq hypothesis holds
• Pressure-driven internal flows in pipes, ducts, and diffusers with mild pressure gradients
• Design optimization loops requiring 500+ CFD evaluations, where the speed of RANS makes exploration feasible
• Steady-state flows where time-averaged quantities (lift, drag, pressure recovery) are the only outputs needed
• Cases with experimental validation where the RANS model has been calibrated against measured data for the specific geometry class

When LES is mandatory

There are problems where RANS simply cannot provide the necessary physics. The extra cost of LES is not a choice — it is the only way to get the right answer:

• Aeroacoustics: far-field noise prediction requires resolving pressure fluctuations on the surface. RANS provides zero spectral information. LES + Ffowcs-Williams & Hawkings is the standard approach.
• Bluff body flows: bridges, chimneys, underwater structures — anything with strong periodic vortex shedding where the Strouhal number and fluctuating loads matter.
• Mixing and combustion: species transport, reaction rates, and pollutant formation depend on resolved turbulent structures. RANS predicts incorrect mixing rates because it smears the large-scale intermittency that drives turbulent mixing.
• Transition prediction: bypass transition (freestream turbulence penetrating the boundary layer) and separation-induced transition are fundamentally unsteady processes that RANS transition models handle poorly.
• Cavity flows and jet-in-crossflow: flows where coherent structures dominate the entire flow field, making RANS predictions qualitatively wrong.

The middle ground: hybrid RANS-LES

Hybrid approaches attempt to get the best of both worlds: RANS near walls (where it works well and LES is prohibitively expensive) and LES away from walls (where large-scale dynamics dominate). Three main approaches exist:

DES (Detached Eddy Simulation), proposed by Spalart in 1997, is the original hybrid. The Spalart-Allmaras model acts as RANS inside the boundary layer and switches to an LES Smagorinsky-like behavior in separated regions. The switch is governed by comparing the wall distance to the local grid spacing. The problem: if the mesh is accidentally too fine in the boundary layer (parallel refinement), the model can switch to LES mode prematurely inside the attached boundary layer — grid-induced separation (GIS).

DDES (Delayed DES) fixes GIS with a shield function that detects the boundary layer edge. The shield function uses the local flow topology (vorticity, strain rate, wall distance) to identify whether a cell is inside the boundary layer and prevents the LES switch there, regardless of mesh resolution.

SBES (Stress-Blended Eddy Simulation) is the ANSYS approach. Instead of a hard switch between RANS and LES, it blends the RANS and LES stress tensors using a smooth transition function. The result is more robust and avoids the numerical artifacts that can appear at the RANS-LES interface in DES.

Case study: cylinder in crossflow at Re = 3900

Flow past a circular cylinder at Re = 3900 is the canonical LES validation case. The flow features a laminar separation bubble on the cylinder surface, transition to turbulence in the free shear layer, and a fully turbulent von Karman vortex street in the wake. Experimental data from Lourenco & Shih and Ong & Wallace provide detailed velocity statistics for comparison.

RANS (k-omega SST) overpredicts the recirculation length behind the cylinder by 40%, and the drag coefficient error is approximately 15%. More critically, RANS cannot capture the unsteady nature of the wake — the solution is steady or weakly periodic with heavily damped oscillations.

LES with the Smagorinsky subgrid model captures the recirculation length within 5% of the experimental value, and the drag coefficient within 3%. The Strouhal number is predicted at St = 0.21, matching the experimental value of 0.21 exactly (RANS gives St = 0.22, a small but significant difference for VIV prediction). The velocity fluctuation profiles in the wake match Ong & Wallace data point by point.

The price: RANS completed in 2 core-hours. LES required 200 core-hours — a 100x increase. For the cylinder problem, the physics gap is so large that the extra cost is unquestionably justified. But the ratio varies by case, and not every problem shows this dramatic a difference.

Practical decision tree

Use this framework to choose your approach:

• Do you need unsteady data (spectra, noise, transient loads)? If yes, LES or DES. RANS cannot provide spectral information.
• Is the flow mostly attached, and do you only need integrated quantities (C_l, C_d, deltaP)? If yes, RANS. The time-averaged quantities will be accurate enough.
• Is there massive separation at high Re? If yes, DDES or SBES. Full LES is too expensive for the wall region, and RANS will mispredict the separation.
• Do you need to evaluate 500+ design points? If yes, RANS with a surrogate-assisted optimization approach. LES at this scale is not feasible for iterative design.
• Is aeroacoustics with far-field noise the primary output? If yes, LES with Ffowcs-Williams & Hawkings post-processing. DES can work for tonal noise but may miss broadband contributions.

Gotchas when switching to LES

Moving from RANS to LES is not just a solver switch — it is a fundamentally different workflow. Engineers coming from a RANS background routinely make these mistakes:

• Inlet boundary conditions: LES requires unsteady, physically realistic turbulence at the inlet. RANS inlet profiles (uniform velocity + turbulence intensity) are insufficient. You need synthetic turbulence generation (SEM, DFG methods) that produces correlated velocity fluctuations matching target Reynolds stresses and length scales.
• Time step: the CFL condition in LES is based on resolved eddy timescales, not mean flow. Target CFL ≈ 1 in LES regions. For URANS you could get away with CFL up to 10-20. This means your LES time step may need to be 20x smaller than what you used for URANS on the same mesh.
• Grid design: LES regions need isotropic cells (aspect ratio close to 1 in all three directions). High-aspect-ratio RANS meshes (cells 100x longer than they are tall) will not resolve large eddies properly. In the wall-normal direction, spacing is still tight for wall-resolved LES, but wall-modeled LES can use coarser spacing.
• Statistical convergence: LES statistics come from time averaging a fluctuating signal, not from residual convergence. You need 10-20 flow-through times after the initial transient dies out to accumulate converged mean and RMS statistics. This is a lot of wall clock time — plan for it.
• Validation: an LES with inadequate mesh resolution can produce results that look impressive but are quantitatively wrong. Always check that at least 80% of the turbulent kinetic energy is resolved (i.e., the subgrid contribution is less than 20%). The Pope criterion provides a formal check.

The trend in the industry is unmistakable: computing power follows Moore's law, and LES becomes feasible for more problems every year. What was done with DES a decade ago is now done with LES. But the decision to switch should always be guided by the physics of the problem, not by the availability of cores. A well-executed RANS simulation beats a poorly resolved LES every time.