RANS Turbulence Modeling — Overview

Regular Navier-Stokes is exact — it describes every turbulent eddy down to the millimeter. But at Re = 10⁶, those eddies span scales from meters down to 0.1 mm. Resolving all of them (DNS) would require $10^{11}$ cells. Nobody does that for a car or a ship.

RANS takes the velocity field and splits it into a time-averaged mean $\bar{U}_i$ and a fluctuation $u'_i$:

$$u_i(x,t) = \bar{U}_i(x) + u'_i(x,t), \qquad \bar{U}_i = \lim_{T\to\infty}\frac{1}{T}\int_0^T u_i\,dt$$

Substituting into N-S and averaging eliminates the fluctuation terms — except for one stubborn nonlinear group: $-\rho\overline{u'_i u'_j}$, the Reynolds stress tensor. This 6-component symmetric tensor represents momentum transfer by turbulent eddies. Modeling it is what all RANS turbulence models are about. The resulting equations describe only the smooth time-mean flow — exactly what engineers usually want.

Yes — and that's a known limitation. Steady RANS completely discards time history; you get one snapshot of the mean flow. For most industrial use cases — does this heat exchanger meet target pressure drop? Does this aircraft wing generate enough lift at cruise? — the mean flow is precisely what you need.

But for phenomena like vortex shedding behind a bluff body, buffet on a transonic wing, or combustion dynamics, you need either unsteady RANS (URANS) or LES. URANS solves the RANS equations with a time-marching scheme, capturing low-frequency large-scale unsteadiness while still modeling all turbulent fluctuations. It's a pragmatic middle ground used widely in turbomachinery and offshore structures.

Pure economics. Consider a full-car external aerodynamics simulation at highway speed (Re ≈ 3 × 10⁷). A well-converged steady RANS run — say with k-ω SST — might use 20 million cells and finish in 8 hours on a 64-core workstation. An equivalent wall-modeled LES would need 500 million cells, require time-accurate integration for several flow-through times, and run for 2–3 weeks on a 2,000-core HPC cluster. The RANS result predicts drag to within about 5–10% of wind-tunnel data. LES might get to 2–3% error. For design-space exploration — comparing 10 variants of a bumper shape — RANS wins decisively on cost.

The breakdown by computational cost, roughly:

• RANS: $\sim N^{1.3}$ scaling with mesh size; hours per case
• URANS: 5–20× RANS cost; days per case
• Wall-modeled LES (WMLES): $\sim Re^{1.8}$; weeks on HPC
• Wall-resolved LES: $\sim Re^{2.4}$; years (research only)
• DNS: $\sim Re^3$; practically impossible at industrial Re

The clearest framework I use is three questions:

1. Do you have significant boundary layer separation or adverse pressure gradients? If yes — aerodynamics with possible stall, diffuser with angle > 15°, turbine blade pressure side near trailing edge — use k-ω SST. It explicitly limits eddy viscosity in APG regions. k-ε will consistently predict delayed separation.

2. Is the flow strongly swirling or does it have curved streamlines? Think cyclone separators, swirl burners, draft tube of a Francis turbine. Here eddy viscosity models fail because turbulence is highly anisotropic. Use RSM.

3. Is it a simple internal flow — ducts, pipes, heat exchangers — with no major separation? Then k-ε Realizable or even standard k-ε is perfectly adequate and will converge faster and more robustly than SST.

Spalart-Allmaras is excellent for external aerodynamics with attached boundary layers — cruise conditions for wings and fuselages, helicopter blade near hover, high-lift devices with moderate flap angles. It was purpose-built for these flows and has been extensively validated against NASA wind-tunnel databases. One equation makes it fast to converge. The key restriction is: do not use it for bluff bodies, round jets, massively separated flows, or any problem where bulk turbulence generation matters away from walls. For example, running S-A on a turbomachinery stage with strong secondary flows will give noticeably wrong results. In that case, SST or RSM is more appropriate.

y+ = 50 is in the log-law region, which is fine for standard wall functions with k-ε. The log-law region spans roughly $30 < y^+ < 300$. Here the velocity profile follows:

$$u^+ = \frac{1}{\kappa}\ln(y^+) + B, \qquad \kappa \approx 0.41,\; B \approx 5.0$$

Wall functions apply this relation algebraically at the first cell, bypassing the need to resolve the viscous sublayer ($y^+ < 5$) and buffer layer ($5 < y^+ < 30$). For k-ω SST with automatic wall treatment — which is the default in modern OpenFOAM and Fluent — the model adapts smoothly between log-law behavior (y+ > 30) and low-Re behavior (y+ ≈ 1). If you have y+ around 50, SST with automatic wall treatment handles it gracefully. The danger zone is $5 < y^+ < 30$ with standard wall functions — neither assumption is valid there.