CFD for the design of energy systems

Lecture 4: Statistical Modeling (RANS) of Turbulent Flows

Reference (books):
D. Wilcox, Turbulence Modeling for CFD
S. Pope, Turbulent flows

I. Reynolds Averaged Navier-Stokes (RANS) approach

a. Closure problem

In RANS approach, Reynolds decomposition is used to split the fields (velocity, pressure, …) in a mean field and
fluctuating (turbulent) field:
𝑢 = 𝑢 + 𝑢′
RANS consists then to solve the Reynolds equations to obtain the mean field (only the mean fields are
predicted by using RANS approach):
𝜕 𝑢&
=0 (from mass conservation)
𝜕𝑥&
𝜕 𝑢& 𝜕 𝑢& 𝑢9 1𝜕 𝑝 𝜕 ? 𝑢& 𝜕 𝑢& ′𝑢9 ′
+ =− +𝜈 − (from momentum conservation)
𝜕𝑡 𝜕𝑥9 𝜌 𝜕𝑥& 𝜕𝑥9 𝜕𝑥9 𝜕𝑥9
The Reynolds equations lead to an unclosed system with 10 unknown terms and only 4 equations.

The major challenge in RANS is to model the Reynolds tensor, 𝑢& ′𝑢9 ′ , to close the system.

b. First order model

The order of a model is given by the moment of statistics available. First order model is a direct modeling of the
Reynolds tensor and then only first statistics moment are available. Second order model is based on the
modeling of the transport equations of the Reynolds tensor, and then, in this case second statistics moment
are available.

Statistical modeling: RANS approach

Energetic consideration
In term of energetic balance, Reynolds tensor plays as
■ Mean kinetic energy and turbulent kinetic energy
a transfer term form the mean kinetic energy to the
Diffusion
turbulent kinetic energy (see previous lecture)
Reynolds Diffusion Diffusion
decomposition
Forcing
Total kinetic Energy Transfer
In RANS, a “loss” of energy (mean kinetic energy) has
Mean Turbulent
Forcing kinetic
Energy
kinetic
energy
then to be modeled, as the molecular dissipation
Dissipation Dissipation
process.
Dissipation

RANS Diffusion

Mean Transfer
1st order
RANS model
A popular way is then to define an artificial turbulent
Forcing
kinetic
From RANS point of view,
viscosity, 𝜈A , to mimic this additional dissipation.
Energy
transfer leads to a loss of the
Dissipation resolved (mean) kinetic energy
à as dissipation !
The Reynolds tensor is then modeled as:
12 2
𝑢& ′𝑢9 ′ − 𝑘𝛿&9 = −2𝜈A 𝑆&9
3
G G H IJ H IL
avec 𝑘 = 𝑢& ′𝑢& ′ and 𝑆&9 = + .
? ? HKL HKJ

This is the Boussinesq hypothesis. This implied two main assumptions:

- Mechanisms generating fluctuation are assumed local
- Reynolds tensor proportional to local strain rate

The Reynolds equations becomes:

To obtain a complete formulation the turbulent viscosity has to be defined.

From dimensional argument, the turbulent viscosity can be defined from a characteristic turbulent velocity and
a characteristic turbulent length:
𝜈A = 𝑢 ∗ ×𝑙P

• algebraic model: Prandtl model

With an equilibrium assumption (Production = Dissipation), we can write,
?
𝜈 A = 𝑙P 2 𝑆&9 𝑆&9
where 𝑙P has to be empirically fixed (no universal value, flow dependent).

• one-equation model
Assuming, 𝑢 ∗ = 𝐶 𝑘, a transport equation is solved for 𝑘,
𝜕𝑘 𝜕𝑘 𝑘 S/? 𝜕 𝜕𝑘
+ 𝑢9 = 2𝜈A 𝑆&9 𝑆&9 − 𝐶R + 𝜈 + 𝜈A
𝜕𝑡 𝜕𝑥9 𝑙P 𝜕𝑥9 𝜕𝑥9
et the model writes as,
𝜈A = 𝐶𝑙P 𝑘

• two-equations model: k-e model

From theory on fully developed turbulence, 𝑢 ∗ ≈ 𝑘 and
Statistical modeling: RANS approach 𝜖 ≈ 𝑢 ∗ S /𝑙P = 𝑘 S/? /𝑙P . This leads to
Modeling of the turbulent viscosity – Example: k-e model 𝐶W 𝑘 ?
𝜈A =
■ k-e model: closed system
𝜖

Transport equations are thus needed for 𝑘 and 𝜖.

5 coefficients have to be fixed to close the equations

system.

à 5 coefficients fixed from “empirical” considerations (experiments in canonical cases)

k-e model is a robust model but can fail to predict
complex flows (swirling flow, wall-bounded flow, …).
Corrected two-equations model can be proposed (RNG,
27
realizable, k-w model, …)

c. Second order model: Reynolds Stress Model (RSM)

Firs order models are based on turbulent viscosity assumption, which implies that the Reynolds stress, 𝑢& ′𝑢9 ′ ,
can be locally linked with 𝑆&9 .
To avoid this strong assumption, modeling can be consisted to directly solve transport equations for 𝑢& ′𝑢9 ′ .
à This is the RSM model.
 II. Wall modeling in RANS approach

One remaining difficulties is to accurately model wall-bounded flows.

The channel flow is an academic case of wall-bounded flows. For channel flows, the laminar velocity profile is
known as a parabolic profile,
𝜏Z 𝑦 𝑦
𝑢 𝑦 = 2−
𝜌𝜈 2 ℎ
HI
with 𝑢 the streamwise velocity, ℎ the half of the channel height and 𝜏Z = 𝜌𝜈 the wall shear stress.
H\ Z

For turbulent channel flows, the boundary layer (i.e. the flow region close to
the wall) can be divided in two regions:

- viscous sub-layer (𝑦 ] < 5), where

𝑢] = 𝑦 ]

- the logarithmic layer (𝑦 ] > 30), where

1
log 𝑦 ] + 𝐵 𝑢] =
𝜅
where 𝑢 ] and 𝑦 ] are non-dimensional quantities (expressed in wall units): 𝑢 ] = 𝑢/𝑢e and 𝑦 ] = 𝑦𝑢e /𝜈, with
𝑢e = 𝜏Z /𝜌 .

Consequences of thin boundary layer for CFD: Due to turbulent mixing properties, velocity profiles
have very stiff gradient in turbulent flows à This is a
Numerical method close to the wall: strong constraint for mesh generation.
- RANS models allow to not discretize (not mesh) turbulent fluctuations
- BUT, strong mean gradients need to be correctly solved
- Thin boundary layers are the main mesh restriction of RANS computation Wall models (or law of walls) can be then used to
avoid explicit velocity gradient computation.
distance to u distance to u
the wall the wall

In RANS application, the behavior of the model close

1st mesh
to the wall has to be also controlled. Dumping
point 1st point to be
accurate function can then be used to correct this behavior.
wall Mean velocity Mean velocity

To avoid very strong refinement (and then expensive simulation) law of the wall can
be used to avoid explicit velocity gradient computation