Download as pdf or txt
Download as pdf or txt
You are on page 1of 114

PHYSICS OF TURBULENCE

Jean-Bernard CAZALBOU & Jérôme FONTANE

2017
Table of contents

I Theory of homogeneous turbulence 1


1 Introduction to turbulence and statistical approach 3
1.1 Definition and characterisation of the turbulent regime . . . . . . . . . . . . . 3
1.2 Statistical description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Equations for the mean flow . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Reynold stress equations . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Problematic of the statistical modelling . . . . . . . . . . . . . . . . . . . . . 9
1.4 Homogeneous turbulence and isotropic turbulence . . . . . . . . . . . . . . . 11
1.4.1 Homogeneous turbulence . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.2 Isotropic turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.3 Practical realisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Two-point correlation theory 17


2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Correlation tensors in HIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Properties of the correlation tensors in homogeneous turbulence . . . . 19
2.2.2 Properties of the correlation tensors in isotropic turbulence . . . . . . 19
2.2.3 Specific form of the correlation tensors, correlation functions . . . . . 21
2.2.4 « Conventional » correlation functions . . . . . . . . . . . . . . . . . . 21
2.3 Correlation tensors in divergence free HIT . . . . . . . . . . . . . . . . . . . . 22
2.4 The characteristic length scales . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.1 Asymptotic expression in the limit of no separation . . . . . . . . . . . 24
2.4.2 Taylor length scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Dynamics of the double correlations . . . . . . . . . . . . . . . . . . . . . . . 27
2.5.1 Linear solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5.2 Self-similar solution without viscous terms . . . . . . . . . . . . . . . . 30
2.5.3 Experimental validation . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Spectral approach in homogeneous isotropic turbulence 35


3.1 Spectral tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.3 Generic form of the spectral tensor in HIT . . . . . . . . . . . . . . . 37
3.2 Spectral density of energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3 One-dimensional spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1 One-dimensional spectra and spectral density of energy . . . . . . . . 41

iii
3.3.2 Differential relations between spectra . . . . . . . . . . . . . . . . . . . 42
3.3.3 Spectral equivalent to the first Kármán–Howarth formula . . . . . . . 43
3.3.4 Generic form of the one-dimensional spectra . . . . . . . . . . . . . . . 43
3.4 Spectral dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4.1 Physical mechanisms at stake in the spectral dynamics . . . . . . . . . 46
3.4.2 Model of balance for the Lin equation and energy cascade . . . . . . . 48
3.5 Kolmogorov hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.1 Statement of the Kolmogorov similarity hypotheses . . . . . . . . . . . 51
3.5.2 Underlying elements to the Kolmogorov hypotheses . . . . . . . . . . . 52
3.5.3 Experimental verification . . . . . . . . . . . . . . . . . . . . . . . . . 55

II Near-wall turbulence 59
4 The turbulent boundary layer 61
4.1 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.1.1 Derivation of the equations for the mean flow . . . . . . . . . . . . . . 62
4.1.2 Mean flow equations under the boundary layer assumption . . . . . . 63
4.2 The mean velocity profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.1 Absence of global self-similarity . . . . . . . . . . . . . . . . . . . . . . 66
4.2.2 Structure of the turbulent boundary layer at large Reynolds number . 67
4.3 Generalisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3.1 Local modifications of the wall conditions . . . . . . . . . . . . . . . . 75
4.3.2 The presence of an external streamwise pressure gradient . . . . . . . 75
4.4 The friction coefficient and the shear stress . . . . . . . . . . . . . . . . . . . 78
4.4.1 The friction law induced by the mean velocity profile . . . . . . . . . . 78
4.4.2 The shear stress profiles . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5 Specificity of near-wall turbulence 83


5.1 Analytical tools for analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.1 Reynolds stress equations . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1.2 Pseudo-dissipation transport equation . . . . . . . . . . . . . . . . . . 85
5.1.3 Near-wall asymptotic laws . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2 Specificities of near-wall turbulence . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2.1 Anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2.2 Importance of the viscous mechanisms . . . . . . . . . . . . . . . . . . 89
5.2.3 Energy balances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

Bibliography 96

Appendix 96
A Dynamic equations for two-point double correlations 99
A.1 Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
A.2 Variable substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
A.3 Final form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

iv
B Derivation of Kármán-Howarth equation 103
B.1 Transformation of derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
B.2 Transposition to the correlation functions . . . . . . . . . . . . . . . . . . . . 103

C Total shear stress profile for a decelerated boundary layer 107

v
vi
Part I

Theory of homogeneous
turbulence

1
Chapter 1

Introduction to turbulence and


statistical approach

1.1 Definition and characterisation of the turbulent re-


gime
Common fluid flow are driven by a mathematical model composed of a closed set of
equations (Navier-Stokes equations) and an appropriate set of initial and boundary condi-
tions. Due to its non-linear and dissipative nature, this model is likely to develop multiple
solutions, some of which can be labelled as turbulent.
The simple example of the planar channel flow will enable to remind us the conditions
of its onset as well as its very specific nature. We consider an incompressible Newtonian
fluid flowing in between two parallel flat plates separated by a distance h. The two walls are
assumed to be infinite in both x and z directions. The fluid motion results from a uniform
pressure gradient along the x direction. The problem is entirely defined by the Reynolds
number and it can be solved analytically to give the Poiseuille solution characterised by
— a stationary flow (∂/∂t = 0) ;
— a one-dimensional motion with only a single non-zero velocity component along x
~ = U (y)~ex ] ;
direction, [V
— a parabolic velocity profile U (y).
If we denote U0 the maximum velocity in the section and Um the flow rate mean velocity, one
can derive the friction coefficient at the wall as a function of the mean Reynolds number :
16 −1 2τp Um h
Cf = Re m , with Cf = and Rem = .
3 ρU02 ν

As one can see in figure 1.1, this solution — although valid whatever the value of the
Reynolds number — is not always observed : experiments agree with Poiseuille solution
only for a range of Reynolds numbers below a threshold of about 1300. For higher Reynolds
numbers, the experimentally measured friction coefficient is significantly larger than the
expected theoretical value.
At this point, one can question the nature of the observed solution. The temporal ve-
locity signal recorded at a given point in the flow brings some answers. As one can see in

3
15 15

Experiments Laminar solution


Poiseuille Turbulent solution
10 10
Cf × 103

5 5

Bifurcation

0 0
102 103 104 105 106 102 103 104 105 106
Re m Re m
(a) (b)

Figure 1.1 – Evolution of the wall friction coefficient with Reynolds number for the planar channel
flow. Poiseuille solution is a laminar solution, symbols in figure (a) correspond to various experiments
(see Dean [1]) and the dotted curve in figure (b) represent a least squares approximation of the
solution in the turbulent regime.

figure 1.2(a), the signal exhibits a strong unsteadiness : there is a « mean value » on top of
which erratic fluctuations are superimposed with a smaller amplitude. The Fourier trans-
form of the signal gives a continuous spectrum on a finite range of frequencies : a cut-off is
observable at high frequencies in figure 1.2(b). Those two characteristics — continuous spec-
trum and finite range of frequencies — constitute the safest definition of the the turbulent
regime.
In this particular flow, one can consider that :
1. two types of solutions coexist above a critical value of the Reynolds number.
2. the « laminar » solution (Poiseuille) is observable in experiments only for weak Rey-
nolds numbers.
3. the « turbulent » solution is characterised by the existence of fluctuations — which are
apparently random — on a continuous and finite range of scales. These fluctuations
are observable at a macroscopic scale (i.e. beyond the supra-molecular scale).
More generally, the transition to turbulence can occur through several « bifurcations » :
some flows are likely to give rise successively to a steady solution, then a periodic and a
doubly periodic solution before the emergence of the turbulent solution. The bifurcation
diagram [figure 1.1(b)] will then present as many distinct branches for larger values of the
Reynolds number. For a given value of the Reynolds number, the emergence of one particular
solution is driven by stability properties of the flow regarding infinitesimal disturbances.
Every flow is subjected to such perturbations through initial and boundary conditions, and
the value of the Reynolds number determines its ability to amplify or damp them. In the first
case, the unperturbed solution is stable : whatever the perturbation the flow always returns
to its initial equilibrium state. In the second case, the unperturbed solution is unstable : due
to the growth of the perturbation, the flow moves leaves its initial equilibrium state to reach

4
Instantaneous velocity U (t)

Spectral density
U

Time Frequency
(a) (b)

Figure 1.2 – Temporal evolution (a) and spectral content (b) of the velocity signal measured at
a fixed point in a turbulent flow steady in average. The spectral curve is plotted in a logarithmic
scale.

a new equilibrium state for the same Reynolds number. In this respect, the turbulent regime
corresponds to a maximum sensitivity of the flow to disturbances, the flow being sensitive
both to their existence and their characteristics : each perturbation field will give rise to a
specific solution of the flow.

1.2 Statistical description


The study of turbulent flows is complex due to their inherent properties of unsteadiness,
three-dimensionality and sensitivity to perturbations. To tackle these difficulties, the sta-
tistical approach has become early (see the seminal work of Reynolds [2]) a suited means
of engineers for both the experimental analysis and computational fluid dynamics. It relies
on the decomposition of every flow field as the sum of mean value and a fluctuation [for
instance, see the decomposition obtained from temporal averaging in figure 1.2(a)]. More
generally, the mean value is defined as the ensemble average or statistical average, obtained
from a set of realisations of the same experiment under the following form
n
1 X (i)
(1.1) F (x, y, z, t) = lim F (x, y, z, t)
n→∞ n
i=1

where F (i) (x, y, z, t) is the measured value of field F at point (x, y, z) and time t during the
ith realisation of the experiment. The statistical average, or mean field, is denoted with an
overline while the fluctuation, denoted with a lower case letter, is defined by

(1.2) F (x, y, z, t) = F (x, y, z, t) + f (x, y, z, t).

5
One can notice that both the mean value and the fluctuation can be function of space and
time.

Statistical stationarity The variable F is said to be statistically stationary when its


mean value does not depend on time. Under the hypothesis of ergodicity — impossible to
demonstrate but commonly admitted —, the ensemble average F is equal to the temporal
average :
Z
1 T
hF it (x, y, z) = lim F (x, y, z, t) dt
T →∞ T 0

Statistical homogeneity One can introduce the notion of statistical homogeneity (re-
lative to the spatial dependence of the variables), similarly to the statistical stationarity
(relative to the temporal dependence of the variables). For instance, if the mean value does
not depend on the z coordinate, the variable F is said statistically homogeneous in the z
direction. Here again, under the hypothesis of ergodicity, one can replace ensemble averaging
F by spatial averaging :
Z
1 Z
hF iz (x, y, t) = lim F (x, y, z, t) dz
Z→∞ Z 0

Properties of statistical stationarity and homogeneity enable to consider tur-


bulent flows for which the mean flow can be stationary, bidimensional, planar
even one dimensional whereas :
— each instantaneous field is necessarily three-dimensional ;
— the value of each variable measured at any given point in the flow is
necessarily unsteady.

Properties of the ensemble average Definitions (1.1) et (1.2) yield the following generic
properties :
f = 0, F + G = F + G, λF = λF , F G = F G + f g,
   
∂F ∂F ∂F ∂F
= and = ,
∂t ∂t ∂xi ∂xi
where F and G are random variables and λ is a scalar. f g is called correlation.

1.2.1 Equations for the mean flow


The scale of the turbulent motions being far larger than the supra-molecular scale, the
Navier–Stokes model remains valid in the turbulent regime. We only consider here the case
of a viscous incompressible fluid for which the equations write under the classical form :
∂Uj
(1.3) = 0,
∂xj

∂Ui ∂Ui 1 ∂P ∂Sij


(1.4) + Uj =− + 2ν with i = 1, 2, 3 ;
∂t ∂xj ρ ∂xi ∂xj

6
where Sij = (∂Ui /∂xj + ∂Uj /∂xi )/2 is the rate of strain tensor. Averaging these equations
leads to the so-called Reynolds equations, usually coined URANS equations for Unsteady
Reynods-Averaged Navier–Stokes equations :

∂U j
(1.5) = 0,
∂xj

∂U i ∂U i 1 ∂P ∂  
(1.6) + Uj =− + 2νS ij − ui uj with i = 1, 2, 3.
∂t ∂xj ρ ∂xi ∂xj

The equation (1.5) is the averaged continuity equation and shows that the mean velo-
city field is divergence free. When subtracted from equation (1.3), one gets the continuity
equation for the velocity fluctuation field of the form ∂uj /∂xj = 0. It appears that in in-
compressible regime, the instantaneous, the mean and the fluctuation velocity fields are all
divergence free. The equations (1.6) are the averaged momentum equations. Together with
the averaged continuity equation, they form a system which is not closed for mean velocity
and pressure fields, contrary to the Navier–Stokes equations. Indeed, they exhibits extra
terms composed of one-point correlations of velocity fluctuations : the Reynolds stresses
ui uj . These terms have some remarkable properties :
1. Reynolds stresses are of advective origin since they come from the statistical decompo-
sition in mean and fluctuation of the instantaneous advective terms of equation (1.4).
At that stage they appear in the averaged momentum equations under the form
uj ∂ui /∂xj . This term corresponds to an advective transport of the ui fluctuation by
the uj fluctuation.
2. For the mean flow, Reynolds stresses can be viewed as a diffusive flux ; the continuity
equation for the velocity fluctuation enables to recast them under the form ∂ui uj /∂xj
in the equation (1.6). This form justifies its designation as « stresses » by analogy
with viscous stresses and it then corresponds to a turbulent diffusion.
3. R = [Rij ] = [ui uj ] is a tensor called Reynolds stress tensor. It is symmetric and its
trace is twice the kinetic energy per mass unit of the fluctuating motion or simply
the turbulent kinetic energy : k = ui ui /2.

1.2.2 Reynold stress equations


In order to close the Reynolds equations, it is very tempting to derive the transport
equations for the extra unknowns that appeared in the mean flow momentum equations.
This is quite straightforward since the transport equations for an instantaneous fluctuation
can be obtained by simply subtracting equation (1.6) to equation (1.4). This is done in three
steps :

 [Eq : Ui ] − Eq : U i  → [Eq : ui ]
(1) ,
 [Eq : U ] − Eq : U  → [Eq : u ]
j j j

(2) ui × [Eq : uj ] + uj × [Eq : ui ] → [Eq : ui uj ] ,


(3) [Eq : ui uj ] → [Eq : ui uj ] .

7
By doing so, one obtains the transport equations for the Reynold stresses under the following
form :
∂ ui uj ∂ ui uj p
(1.7) + Uk = Pij + Dij + Dij + dνij + φij − εij ,
∂t ∂xk
where
 
∂U j ∂U i
Pij = − ui uk + uj uk . . . . . . . . . . . . . . . . . . . . . . . . . . . . is the production term ;
∂xk ∂xk
∂ ui uj uk
Dij = − . . . . . . . . . . . . . is the diffusion by velocity fluctuation ;
∂xk
 
p 1 ∂ pu i ∂ puj
Dij = − + . . . . . . . . . . . . is the diffusion by pressure fluctuation ;
ρ ∂xj ∂xi
ν ∂
dij = 2ν (ui sjk + uj sik ) . . . . . . . . . . . . . . .is the molecular or viscous diffusion ;
∂xk
 
1 ∂ui ∂uj
φij = p + . . . . . . . . . . . . . is the pressure-strain correlation term ;
ρ ∂xj ∂xi
 
∂ui ∂uj
εij = 2ν sjk + sik . . . . . . . . . . . . . . . . . . . . . . . . . . . is the dissipation tensor.
∂xk ∂xk
Viscous effects are present on the right hand side of the equation as the combination of a
term of diffusion (dνij ) and a term of dissipation (εij ). This decomposition shows well the
thermodynamic aspects of the evolution of the Reynolds tensor : it can be easily demonstra-
ted that the trace of the tensor ε represents an irreversible conversion of the turbulent kinetic
energy (at a macroscopic scale) into internal energy (at a microscopic scale). Nevertheless,
this decomposition of the viscous term is not unique and another decomposition using terms
of « pseudo » diffusion and dissipation is often used :
  
ν ∂ 2 ui uj ∂ui ∂uj
Dij = ν and ǫij = 2ν .
∂xk ∂xk ∂xk ∂xk

The sum of these two terms is strictly equal to dνij + εij , and here the diffusion term has a
conservative form (divergence).
From equation (1.7), the index contraction enables to derive the transport equation for
the turbulent kinetic energy k = ui ui /2 :
  
∂k ∂k ∂U i 1 ∂ puj ∂ kuj ∂2k ∂ui ∂ui
+ Uj = −ui uj − − +ν −ν .
∂t ∂xj ∂xj ρ ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj
| {z } | {z } | {z } | {z } | {z }
p ν ǫ
Pk Dk Dk Dk

As the Reynolds stress tensor is symmetric, the production term can also be written as
 
1 ∂U i ∂U j
Pk = − ui uj + = −ui uj Sij ,
2 ∂xj ∂xi

which shows that the turbulent kinetic energy is only produced by the mean rate of strain,
not by the mean rotation. It is clear that these equations are useless for solving the closure

8
problem since they include triple correlations. It could have been easily anticipated since,
in the same way that equations for first order momentum (U i ) bring out second order
correlations (ui uj ), the equations of second order momentum will necessarily bring out third
order correlations (diffusion terms by velocity fluctuation), and even new correlations of
second order (pressure-velocity correlations or correlations between gradients of fluctuations
in the dissipation term). The closure problem is thus never ending, expressing the fact
that statistical description leads necessarily to a loss of information. The classic analysis
of the right hand side of the turbulent kinetic energy transport equation, consisting in the
identification of source, sink and diffusion terms, leads to some interesting comments :
1. The « production » term is the only one including an interaction with the mean flow.
It corresponds to an exchange between the kinetic energy of the mean flow and the
turbulent kinetic energy (the same term is found in both equations with opposite
sign). It is usually a source term in the transport equations of turbulent kinetic and
Reynolds stresses, and a sink term in the transport equation of the mean flow kinetic
energy. It can be concluded that most of the time the fluctuating motion extracts
energy from the mean flow.
2. The dissipation term corresponds to an exchange between the turbulent kinetic energy
and internal energy. It is usually a sink term in the Reynolds stress equations and
the trace of the dissipation tensor is a source term in the transport equation of mean
internal energy.
3. The diffusion terms can all be written through a divergence form. They represent a
spatial transfer without any gain or loss when the fluxes are zero at the boundaries
of the control volume.
4. The pressure-strain correlation term in the Reynolds stress equations disappears from
the turbulent kinetic energy transport equation (its trace is nil since the velocity
fluctuation is divergence free). It represents a mechanism of energy redistribution
between components (normal stress) without any gain or loss for k.
These mechanisms are illustrated in figure 1.3.

1.3 Problematic of the statistical modelling


Modelling is a response to the necessity of getting a closed mathematical model in order
to compute turbulent flows. It results in statistical models which cannot fully represent
reality but proved their practical interest (since the beginning of the 80s for the aeronautic
industries).
The calculations are based on the resolution of the Reynolds equations in which the
Reynolds stress tensor has been somehow evaluated. There are
— first order models, when transport equations for the first order momentum of the
velocity field (U i ) are solved, and its second order momentum (ui uj ) are « modelled » ;
— second order models, when transport equations for the first and second order mo-
mentum of the velocity field are solved, and its third order momentum and other
unknown correlations are « modelled ».
The modelling assumes a scheme that links the value of the unknown correlation to all the
computed quantities. The building of realistic schemes requires to characterise the turbu-
lent motion in a synthetic way. This is what is done in the next chapters for the simplest

9
spatial
mean ✻ Dν
motion Ec = U iU i /2 k
transfers

✒ p
Dk

Pk

w2
✻✛ φ

φ ❄✲v
2

Dk

❄✛ ✒
u2 ✠ φ
k = ui ui/2

ǫ
local
transfers

molecular
Ei = Cv T motion

Figure 1.3 – Summary of energetic transfers active in the transport equation of the turbulent
kinetic energy. Transfers between different classes of motions (mean, fluctuating and molecular
motions) are local, as well as the inter-component transfers (within the fluctuating motion, φij ).
Spatial transfers correspond to different types of diffusion.

10
configuration, i.e. the homogeneous isotropic turbulence (HIT) which is defined in the next
section.

1.4 Homogeneous turbulence and isotropic turbulence


As seen just above, the description of the mean flow requires to take into account the
turbulent motion through the Reynolds stress tensor. From a formal point of view, there
exists an analogy between this modelling and the one which leads to the continuous me-
chanics hypothesis from the kinetic theory of gas. In the later, the molecular motions are
modelled at the macroscopic scale with the stress tensor. The relation between the stress
and rate of strain tensors enables to define the mechanical behaviour of the fluid « material »
and leads to the Navier–Stokes model for a Newtonian fluid. By analogy, we consider here a
« turbulent material » which mechanical behaviour could be specified by a « behaviour law »
for the Reynolds stresses. We will discuss later in the manuscript the limits of this analogy,
but it can already emphasised that the statistical properties of the fluctuating motion are far
more complex than those of the molecular motion. For this reason, the study of turbulence
is based on simplified flow configurations that enable to isolate the different mechanisms
related to the Reynolds stresses.
To this end, the simplest configurations that one can imagine are those where the dyna-
mics of the Reynolds stress is solely driven by local transfers. These configurations are the
homogeneous turbulence and the isotropic turbulence.

1.4.1 Homogeneous turbulence


Definition A fluctuating field is said to be homogeneous if all of its statistical properties
are invariant under any translation.

Reynolds stress dynamics in homogeneous turbulence As a direct consequence of


the definition of homogeneity, the spatial derivatives at all orders of all correlations are nil.
Thus the transport equations for the Reynolds stress and for the turbulent kinetic energy
simplify to :
∂ ui uj ∂k
= Pij + φij − ǫij , and = Pk − ǫ.
∂t ∂t

Necessary condition of existence The above equations show that a homogeneous field
can remain homogeneous only if the production terms (Pij and so Pk ) are constant in space.
 
It implies that the tensor of the mean velocity gradient G = ∂U i /∂xj must be constant
in space. From this point, one can introduce the following definitions of
— the Craya homogeneity : the gradient tensor is constant but non zero everywhere ;
— the Taylor homogeneity : the gradient tensor is zero everywhere,
The initial turbulent field must be homogeneous in both cases.

1.4.2 Isotropic turbulence


Definition A fluctuating field is said isotropic if all of its statistical properties are in-
variant under any rotation and any reflection symmetry. It results that an isotropic field

11
is necessarily homogeneous since any translation is a combination of two rotations. The
name « homogeneous isotropic turbulence » (HIT) is thus uselessly precise but is commonly
accepted.

Reynolds stress dynamics in HIT As a direct consequence of the properties of homo-


geneity and isotropy :
— the spatial derivatives at all orders of all correlations are nil and the tensor of mean
velocity gradient is nil as well ;
— for the second order correlations, diagonal terms are equal and off-diagonal terms are
zero 1 . In particular, one can write :

2
u2 = v 2 = w 2 = k and uv = vw = uw = 0 ;
3
2
ǫ11 = ǫ22 = ǫ33 = ǫ and ǫ12 = ǫ13 = ǫ23 = 0 ;
3
φij = 0 whatever i and j.

Thus, the transport equation for the turbulent kinetic energy simplifies to :

∂k
(1.8) = −ǫ,
∂t

which shows that, in isotropic turbulence, the amount of turbulent kinetic energy can only
decrease in time, at a rate imposed by the amount of dissipation ǫ.

1.4.3 Practical realisations

The models of homogeneous turbulence and isotropic turbulence enable to study the
behaviour of the « turbulent material » in the simplest configurations. Historically, physical
characterisation of the turbulent regime was only provided through experiments. But, since
the 1980s, the development of computing sciences made possible the direct numerical simula-
tion of turbulent flows, i.e. the solving of the 3D unsteady Navier–Stokes equations down to
the smallest scales of motion. Nowadays, the available computing power is still limiting the
numerical simulation to geometrically simple flows at Reynolds numbers which are not large
enough to be representative of industrial configurations. Both approaches — experiments
and simulations — are complementary, each of them having advantages and limitations :

1. The only second order isotropic tensor is the identity tensor, the only third order isotropic tensor is
the tensor of Levi-Civita.

12
Experiment Simulation
Interests
— « reality » — « clean » configuration
— large Reynolds numbers — any field is measurable
— complex flows

Limitations
— uncertainties of measure — numerical biases (discreti-
— imperfection in compari- sation errors, rounding er-
son with « theoretical » rors, finite box size, etc.)
configurations — small Reynolds numbers
— some fields are not measu- — « simple » flows
rable

We now will see how these two approaches can be used to characterise the various confi-
gurations of homogeneous turbulence.

Direct numerical simulation of homogeneous turbulence The homogeneity pro-


perty allows to use a finite computing domain, the only constraint being that the size of the
domain must be large enough to include the biggest structures of the flow. The common
practice consists in using a cubic domain and to impose periodic boundary conditions in
all directions to the fluctuating velocity field. « Periodicity » is one of the numerical biases
which can be easily minimised (although costing) by increasing the size of the domain.
In the case of homogeneous isotropic turbulence, the absence of any production term
imposes to initiate the simulation with a « realistic » velocity field. Some techniques based
on random selection can produce initial fields which are only getting close to real turbulence.
For this reason, the simulation starts with a transient regime during which the data generated
cannot be used for computing the correlation statistics.
The first numerical simulations of homogeneous isotropic turbulence (HIT) were perfor-
med in 1972 (Orszag & Patterson [3]), Rogallo [4] has then extended the method of Orszag
& Patterson in order to compute various elementary configurations of Craya homogeneous
turbulence :
— homogeneous turbulence in a planar deformation ;
— homogeneous turbulence in axisymmetric contraction ;
— homogeneous turbulence in a uniform shear ;
— homogeneous turbulence in solid-body rotation.
These various configurations are sketched in the table 1.1. In agreement with the result given
in paragraph 1.4.1, they are characterised by a constant value for the tensor of mean velocity
gradient. The table gives precisely the form of this tensor through the classic decomposition
in a symmetric and an antisymmetric part :
   
1 ∂U i ∂U j 1 ∂U i ∂U j
Gij = Sij + Dij with Sij = + and Dij = − ,
2 ∂xj ∂xi 2 ∂xj ∂xi

which corresponds to the rate of strain tensor (S) and the rotation tensor (D) of the mean
flow.

13
Planar deformation
Mean velocity field :

U = 0, V = Dy and W = −Dz ❄
❄ ✒
z y ✒
Strain and rotation tensors : ✻


    x ✠
0 0 0 0 0 0 ✠ ✻
S= 0 D 0 , D =  0 0 0  ✻
0 0 −D 0 0 0

Axisymmetric contraction
Mean velocity field :

U = 2Dx, V = −Dy and W = −Dz ❄ ❄


z ✛ ✠ ✠✲
y ✛ ✲
Strain and rotation tensors : ✻
✒ ✛ ✲
    ✲ ✛ ✲
2D 0 0 0 0 0 x ✒ ✒
S =  0 −D 0 , D= 0 0 0  ✻ ✻
0 0 −D 0 0 0

Uniform shear
Mean velocity field :

U = 2Sz, V =0 and W = 0 ✲

z y ✲

Strain and rotation tensors : ✒
✲ ✛
x ✛
    ✛
0 0 S 0 0 S ✛
S =  0 0 0 , D =  0 0 0 
S 0 0 −S 0 0

Solid body rotation


Mean velocity field :

U = −Ωy, V = Ωx and W = 0 ✻


Strain and rotation tensors : z y
✻ ✛
    ✒ ✲
0 0 0 0 −Ω 0 ✲
x ✛
S =  0 0 0 , D =  Ω 0 0  ✲
0 0 0 0 0 0

Table 1.1 – Different configurations for simulation of basic cases of homogeneous turbulence.

14
Experiments of homogeneous turbulence The HIT is characterised by the absence of
any production term. For this reason, the practical realisation of HIT is not straightforward.
Since the 1950s, experimentalists are using grid turbulence. As it can be seen in figure 1.4,
turbulence is generated when a uniform flow is going through a grid made of two crossed rows
of parallel bars (biplane grid). After the grid, the flow stays uniform in average (with U = U0 ,
if U0 is the upstream velocity), but it presents a significant level of turbulence (urms /U0 = 3–5
%). The absence of any production mechanism downstream of the grid implies that the level
of turbulence decreases with the downstream distance. One can obtain a statistically steady
and spatially decreasing turbulence, which appears like a weakly inhomogeneous (along U0 )
and temporally decreasing turbulence if observed in the frame moving at speed U0 . It can be
considered as HIT if, at the observation scale, the spatial decrease under diffusion is negligible
compared to the temporal decrease under dissipation. This condition is known under the
name of Taylor hypothesis or frozen field hypothesis. Practically, it is easily satisfied and it
can be admitted that time in theoretical models is analogous — given the 1/U0 factor — to
the downstream distance in the grid experiment (t ∝ x/U0 ).
From a grid turbulence experiment and assuming the Taylor hypothesis, it is possible to
get experimental realisations of the various types of homogeneous turbulence. The following
configurations can be studied :
— planar deformation where the grid is followed by a tunnel with constant section which
evolves from a rectangle elongated in one direction to a square and then to a rectangle
elongated in the other direction [figure 1.5(a)] ;
— axisymmetric contraction where the grid is followed by a wind tunnel section that
progressively and evenly decreases in both directions [figure 1.5(b)] ;
— uniform shear where the spacing between the grid bars varies along one direction
[figure 1.5(c)] ;
— solid body rotation where the grid is preceded by a honeycomb and inserted in a
cylindrical test section which is rotating [figure 1.5(d)].

15
z




✲ y

✲ ✲
x



✲ ✲
U0 U (x, y, z, t)

Figure 1.4 – Experimental configuration of a grid turbulence. The intensity of the fluctuation is
exaggerated on the velocity profile sketched right after the grid : it is usually only few percent of
the mean velocity intensity.

U = U0
✲ U0
✲ ✲
U0 ✲ ✲ U > U0
✲ ✲ ✲ ✲
✲ ✲ ✲
✲ ✲ ✲

✲ ✲ ✲ ✲
✲ ✲ ✲ ✲
✲ ✲
✲ ✲

(a) (b)
Entrée Veine
(fixe) Tronçon tournant (fixe)
z }| { z }| { z }| {

✲ ✲
✲ ✲
✲ ✲
✲ ✲
✲ ✲ U0 ✲ ❄ Ω✲
✲ ✲ ✻
✲ ✲
✲ ✲
✲ ✲
✲ ✲ | {z }

U0 ∂U /∂z = S Nid d’abeille Grille

(c) (d)

Figure 1.5 – Homogeneous turbulence experiments. (a) Planar deformation, (b) axisymmetric
contraction, (c) uniform shear, (d) solid-body rotation.

16
Chapter 2

Two-point correlation theory

The characterisation of the turbulent regime by the existence of fluctuations for all
fields and on a wide continuous range of scales naturally leads to consider the flow as
being the superposition of a population of vortical structures with various size. Besides, the
problem of modelling is related to the one-point correlations, i.e. the averaged product of
at least two fluctuations measured at one given spatial location. If these correlations bring
information about the intensity of the turbulent motion, they cannot provide information
on the size of the associated vortical structures. For this reason, we introduce the two-
point correlations : the information « distance » is added into the statistical description by
introducing a spatial separation between the points where the fluctuations appearing in the
correlation are measured.
In this chapter, it will be demonstrated that this analysis provides the definition of length
scales which are characteristic of the turbulent motion. The evolution of these length scales
is then studied in the simple configuration of an homogeneous isotropic turbulence.

2.1 Definitions
The functions of two-point correlations are defined in an orthonormal referential (O, ~ex , ~ey , ~ez ).
We consider two points P and P ′ such as
~ = X,
OP ~ ~ ′=X
OP ~ ′, ~′ − X
and ~r = X ~ ;

and three unitary arbitrary vectors α ~ and ~γ , such as


~, β

α
~ = (α1 , α2 , α3 ), ~ = (β1 , β2 , β3 )
β and ~γ = (γ1 , γ2 , γ3 ) .

One can define :


1. a double velocity two-point correlation

Qα,β (P, P ′ , t) = uα (P, t) uβ (P ′ , t) ;

2. a triple velocity two-point correlation

Sαβ,γ (P, P ′ , t) = uα (P, t) uβ (P, t) uγ (P ′ , t) ;

17
✼α~ ✼α~
✼uα ✼uα
P P p
z P uβ
✻ β~ ✰ ✰
O ✲
y ~r ~r ~r
✠ ✼α ~
x ✼uα
◆ ◆ ◆
uβ P′ P′
P′
β~ ✰ ✰ ☛ uγ
☛~γ
Qα,β (P, P ′ , t) Sαβ,γ (P, P ′ , t) L,α (P, P ′ , t)

Figure 2.1 – Definition of two-point correlation functions.

3. a mixed pressure-velocity two-point correlation


L,α (P, P ′ , t) = p(P, t) uα (P ′ , t).

The values denoted uα , uβ and uγ indicate respectively the projection along α ~ , β~ and ~γ of
the velocity fluctuation vector. The coma in subscript separates projections according to the
point where they are evaluated : at point P for projections before the coma and at point P ′
for projections after the coma. Figure 2.1 illustrates these definitions.

2.2 Correlation tensors in HIT


For two given points P and P ′ , the above definitions show that there exist a triple infinity
of correlation functions (as many as vectors α ~ , β~ and ~γ ). Under this form, it is thus a very
rich description tool but largely over-abundant. If the projections of the fluctuation fields
on the selected vectors are explicitly derived by use of the scalar product :

uα = ~v · α
~ = ui αi ,
uβ = ~v · β~ = u j βj ,
uγ = ~v · ~γ = uk γk ;
one obtains the following relations :
Qα,β (P, P ′ , t) = ui (P, t) uj (P ′ , t) × αi βj ,
Sαβ,γ (P, P ′ , t) = ui (P, t) uj (P, t) uk (P ′ , t) × αi βj γk ,
L,α (P, P ′ , t) = p(P, t) ui (P ′ , t) × αi .
These relations show that every two-point correlation can be expressed from a finite set
of them, that can be considered as the components of three tensors :
the tensor of double correlations Qi,j (P, P ′ , t) = ui (P, t) uj (P ′ , t) ;
the tensor of triple correlations Sij,k (P, P ′ , t) = ui (P, t) uj (P, t) uk (P ′ , t) ;
the tensor of the pressure-velocity correlations L,i (P, P ′ , t) = p(P, t) ui (P ′ , t).

18
Thus :

In the general case, correlation tensors


— are expressed with 7 scalar parameters ;
— bring into play 9+18+3 scalar functions (as many as the number of
independent components of the three tensors considered here).

We will now demonstrate that these expressions can be simplified further in the confi-
guration of HIT with a divergence free velocity field (i.e. isovolume evolution). For this
purpose, the consequences of these three hypothesis (homogeneity, isotropy and isovolume
evolution) will be considered next.

2.2.1 Properties of the correlation tensors in homogeneous turbu-


lence
With the hypothesis of homogeneity, the statistical properties of turbulence are inva-
riant under any geometrical translation. For the two-point correlations, it implies that the
correlation measured at two points P and P ′ is always equal to the correlation measured
−−→ −−−→
at two points P1 and P1′ if the vectors P P ′ and P1 P1′ are equal. It results in the following
simplification for the correlations :
Qi,j (P, P ′ , t) = Qi,j (~r, t), Sij,k (P, P ′ , t) = Sij,k (~r, t), L,i (P, P ′ , t) = L,i (~r, t).
This implies a kind of antisymmetry for the tensor of double correlations :

Qi,j (~r, t) = Qj,i (−~r, t),

since
Qi,j (~r, t) = ui (P, t) uj (P ′ , t)
= uj (P ′ , t) ui (P, t)
= Qj,i (−~r, t).

2.2.2 Properties of the correlation tensors in isotropic turbulence


The hypothesis of isotropy implies that all statistical properties of turbulence are inva-
riant under any rotation and reflection symmetry. Regarding the two-point correlations, this
invariance property applies for both the separation vector ~r and the vectors ~ei , ~ej et ~ek of
the coordinate system on which are projected the fluctuations. Figure 2.2 illustrates this
point with the particular case of the velocity double correlations.
Thanks to this property, one can assume without loss of generality that the separation
vector ~r is aligned with the first unitary vector ~ei . In other words, we define the value 1
of the vector subscript such that it corresponds to the direction of the separation vector ~r.
Hence, the functional dependence of the correlation tensors can be simplified by replacing
the separation vector ~r by its norm r :
Qi,j (~r, t) = Qi,j (r, t), Sij,k (~r, t) = Sij,k (r, t), L,i (~r, t) = L,i (r, t).

19



✯ r~′
~r ✯
✲ P
P
≡ P
≡ ❥
Invariance par Invariance par r~′′
Rotation Symétrie plane

Figure 2.2 – Properties of double correlations in isotropic turbulence. The dotted line on the right
part of the figure indicates the plan used for the reflection symmetry.

This convention on indexes is essential to ensure that vectors ~r, ~ei , ~ej and e~k are always
subjected together to the same rotations and symmetries. From now, this convention is
implicitly used in the expression of the correlation tensors each time that the scalar notation
r is used.
We now demonstrate that the hypothesis of isotropy leads to the symmetry of the double
correlations tensor, i.e. :

Qi,j (~r, t) = Qj,i (~r, t)

To this end, we consider the central inversion of center M defined as the middle-point of the
line segment P P ′ . This geometrical transformation is the compound of the rotation of angle
π around any given axis going through M , and of the reflection symmetry relative to the
plan perpendicular to this axis at point M . The hypothesis of isotropy enables to conclude
about the invariance of correlation Qi,j under this transformation which transforms elements
of the correlation in the following manner :

P ⇋ P′ ;
e~i → −~
ei and ui → −ui ;
e~j → −e~j and uj → −uj .

It can be deduced that

ui (P, t)uj (P ′ , t) = (−ui )(P ′ , t)(−uj )(P, t)


= ui (P ′ , t)uj (P, t) ;

which yields the result we were looking for :

Qi,j (P, P ′ , t) = Qj,i (P, P ′ , t).

20
2.2.3 Specific form of the correlation tensors, correlation functions
It can be demonstrated following Robertson [5] that the most generic form for the cor-
relation tensors if they satisfy both properties of homogeneity and isotropy is :

(2.1) L,i (~r, t) = l(r, t) × ri


(2.2) Qi,j (~r, t) = q1 (r, t) × ri rj + q2 (r, t) × δij
(2.3) Sij,k (~r, t) = s1 (r, t) × ri rj rk + s2 (r, t) × rk δij + s3 (r, t) × (ri δjk + rj δki )

Functions l, qi and si are called Robertson functions and the direct consequence is :

In isotropic homogeneous turbulence, the correlation tensors


— are expressed with only 2 scalar parameters ;
— bring into play 1+2+3 scalar functions (Robertson functions).

2.2.4 « Conventional » correlation functions


The Robertson functions do not have an obvious physical meaning. For this reason, we
introduce the conventional correlation function which are defined as :
1. two-point double correlation functions

Q1,1 (r, t) Q2,2 (r, t)


(2.4) f (r, t) = and g(r, t) = .
u′2 u′2
Function f is called longitudinal auto-correlation function and function g is called
transversal auto-correlation function.
2. two-point triple correlations functions

S11,1 (r, t) S22,1 (r, t) S12,2 (r, t)


(2.5) k(r, t) = , h(r, t) = , and q(r, t) = .
u′3 u′3 u′3
Functions k, h and q are called respectively longitudinal, crossed and mixed triple
correlation functions.
are normalised by the characteristic velocity scale u′ of the turbulent motion
These functions p
defined by u′ = 2k/3. Figure 2.3 illustrates the definitions of these five functions.

Equivalence between conventional correlation functions and Robertson functions


The Robertson and conventional functions are naturally equivalent. In order to derive the
relations between the two, one just need to write relations (2.1) to (2.3) using the convention
for the orientation of the separation vector [~r = (r, 0, 0)], which leads to

Q1,1 = q1 r2 + q2 , Q2,2 = q2 , S11,1 = s1 r3 + s2 r + 2s3 r, S22,1 = s2 r, S12,2 = s3 r;

then :
q1 f −g q2
(2.6) = and =g
u ′2 r2 u′2

21
f g

✻ ✻
~r ~r
✲ ✲ ✲ ✲
P P′ P P′

k h q

✻ ✻ ✻
~r ~r ~r

✲ ✲ ✲ ✲ ✲ ✲ ✲
P P′ P P′ P P′

Figure 2.3 – Conventional correlation functions.

s1 k − h − 2q s2 h s3 q
(2.7) = , = and =
u′3 r3 u′3 r u′3 r
Thus, the generic expression for the correlation tensors in HIT is :

Qi,j ri rj
Ri,j = = (f − g) 2 + gδij
u′2 r
Sij,k ri rj rk r rj  rk
i
Tij,k = ′3 = (k − h − 2q) 3 + q δjk + δik + h δij
u r r r r

2.3 Correlation tensors in divergence free HIT


The use of continuity equation enables to simplify even further the expression of the
correlation tensors. As seen in the previous chapter, for an incompressible fluid, the velocity
fluctuation is divergence free (or equivalently isovolume). It implies the following results :
∂L,i ∂Qi,j ∂Sij,k
(2.8) = 0, =0 and = 0.
∂ri ∂rj ∂rk
−−→
Denoting P P ′ = ~r, the invariance property under any translation allows to derive the
extremity P ′ of the vector ~r only. Indeed, one can always bring the vector −→
r back at the
~ . Then, one gets
origin of the referential with the translation of vector −OP

∂ h −−→ i ∂
L,i (P P ′ , t) = p(P, t) × [ui (P ′ , t)],
∂ri ∂x′i
∂ h −−→ i ∂
Qi,j (P P ′ , t) = ui (P, t) × [uj (P ′ , t)],
∂rj ∂x′j
∂ h −−→ i ∂
Sij,k (P P ′ , t) = ui (P, t)uj (P, t) × [uk (P ′ , t)].
∂rk ∂x′k
In the right hand side of the above relations, one can identify the divergence of the velocity
fluctuation which is zero and leads to the relations (2.8). Using expressions (2.1) to (2.3),
various relations of interdependence can be obtained between the Robertson functions :

22
0 = l
∂q1 1 ∂q2
0 = 4q1 + r +
∂r r ∂r
∂s1 2 ∂s3
0 = r + 5s1 +
∂r r ∂r
∂s2
0 = r + 3s2 + 2s3
∂r

Demonstration Deriving relation (2.1) and using (2.8), one gets :


∂ri ∂l ∂l ∂r
0=l + ri = 3l + ri .
∂ri ∂ri ∂r ∂ri
As r2 = ri ri , we can write

∂r ∂ ri ri 2ri ri
= = √ =
∂ri ∂ri 2 ri ri r
and it leads to :
∂l ri ri ∂l
0 = 3l + = 3l + r
∂r r ∂r
which can be rewritten as
1 ∂l 3
=−
l ∂r r
3
and integrated as l(r, t) = C(t)/r . Since the correlation must stay finite for no separation
(r = 0), constant C(t) is necessarily zero. Finally, we get the desired result : l = 0.
The three others relations can be demonstrated in a similar way (left for exercise to the
student), taking into account that ∂r/∂ri = ri /r, ∂ri /∂rj = δij and that rj δij = ri .

Kármán–Howarth formulas The formulas of Kármán and Howarth result from the
transposition of the previous relations to the conventional correlation functions. They can
be derived straightforwardly using the equations (2.6) and (2.7) :

r ∂f
g=f+
2 ∂r
k = −2h
r ∂h
q = −h −
2 ∂r
The direct consequence is the following minimalist expression for the correlation tensors
in divergence free HIT :

L,i = 0
 
1 ∂f r ∂f
Ri,j = − ri rj + f + δij
2r ∂r 2 ∂r
 
1 1 ∂r2 k 1 ∂r2 k k
Tij,k = 3 3k − ri rj rk + 2 (ri δjk + rj δik ) − rk δij
2r r ∂r 4r ∂r 2r

23
f (r, t)
g(r, t)
1

Figure 2.4 – Generic form of the longitudinal (f ) and transversal (g) auto-correlation functions.

From these relations, it can be concluded that in divergence free HIT, the
correlation tensors
— are expressed with only 2 scalar parameters (r and t) ;
— bring into play only 2 scalar functions (f and k).

2.4 The characteristic length scales


We now show that the auto-correlation functions give informations on the characteristic
length scales of the turbulent motion. A typical evolution of functions f and g in HIT as a
function of the separation vector r is given in figure 2.4. It can be noted that
— the absolute values of the two functions stay in between 0 and 1 (they satisfy the
Schwarz inequality) ;
— whatever the function considered, the velocity fluctuations stay correlated for small
separation distances and then goes to zero quickly with increasing r ;
— the longitudinal auto-correlation function stays positive while the transversal auto-
correlation function become negative from a given distance.
All these properties of the auto-correlation functions can be used to define a « measure »
for the size of the vortical structures characteristic of the turbulent motion.

2.4.1 Asymptotic expression of double correlations in the limit of


no separation
First, we look at the form of the auto-correlation functions in the limit of no separation,
i.e. when r → 0. To this end, the longitudinal correlation is written using an asymptotic

24
development of the fluctuation taken at point x + r when r goes to zero :
 
∂u r2 ∂ 2 u r3 ∂ 3 u 4
u(x) u(x + r) = u(x) × u(x) + r + + + O(r )
∂r r=0 2 ∂r2 r=0 6 ∂r3 r=0

∂u r2 ∂ 2 u r3 ∂ 3 u
= u2 (x) + u(x) r + u(x) + u(x) + O(r4 ).
∂r r=0 2 ∂r2 r=0 6 ∂r3 r=0

Likewise, at point x − r :

∂u r2 ∂ 2 u r3 ∂ 3 u
u(x) u(x − r) = u2 (x) − u(x) r + u(x) − u(x) + O(r4 ).
∂r r=0 2 ∂r2 r=0 6 ∂r3 r=0

The invariance under any translation implies that these two correlations are equal to each
other as well as their half sum. Then, it yields the following result :

r2 ∂ 2 u
(2.9) u(x) u(x + r) = u2 (x) + × u(x) + O(r4 ),
2 ∂r2 r=0

which proves that the longitudinal correlation is even. Dividing this expression by u′2 =
u2 (x), one gets the same property for the auto-correlation function which asymptotic deve-
lopment reads

r2 ∂ 2 f (r, t)
(2.10) f (r, t) = 1 + × + O(r4 ).
2 ∂r2 r=0

By identification with equation (2.9), one can get the expression of the second order deriva-
tive as a function of the local properties of the velocity field :

∂ 2 f (r, t) 1 ∂ 2 u
(2.11) = u .
∂r2 r=0 u′2 ∂r2 r=0

Similar developments lead to similar results for the other correlation functions and one
must retain that :

In divergence free HIT,


— the longitudinal (f ) and transversal (g) auto-correlation functions are
even functions with respect to the separation distance ;
— the longitudinal (k) and crossed (h) triple correlation functions are odd
function with respect to the separation distance ; [O(r3 )].

2.4.2 Taylor length scales


From the form of the auto-correlation functions at the origin (r → 0), a first set of length
scales can be defined. Indeed, the evenness of these functions ensure the existence of a non-
degenerated osculating parabola at the origin for both f and g. The Taylor micro-scales

25
λf and λg relative to these two functions are defined as the intersection of the osculating
parabola with the horizontal axis, see figure 2.5(a) :

1 1 ∂ 2 f (r, t) 1 1 ∂ 2 g(r, t)
(2.12) = − and = − .
λ2f (t) 2 ∂r2 r=0 λ2g (t) 2 ∂r2 r=0

Using the result obtained in the previous section, it can be rewritten as



1 1 ∂ 2 u
= − u .
λ2f (t) 2u′2 ∂r2 r=0

This form enables to evaluate theoretically the Taylor micro-scale in HIT by measuring
simultaneously the velocity fluctuation and its second order derivative for each realisation
of the flow before computing the ensemble average. In practice, things can be simplified by
noticing that
   2    2
∂2u ∂ ∂u ∂u ∂ ∂u ∂u
u 2 = u − = u − .
∂r ∂r ∂r ∂r ∂r ∂r ∂r
and that the second order derivative on the right hand side is zero under the hypothesis of
invariance under translation. Then, the relation practically used to determine the micro-scale
is :
 2
1 1 ∂u
(2.13) 2 = ′2 ,
λf (t) 2u ∂r
r=0

where only the measure of the first order derivative of the fluctuation is required to compute
the micro-scale.
Regarding the transversal micro-scale λg , it has been demonstrated earlier that functions
f and g are not independent in divergence free HIT. Using the first Kármán and Howarth
formula, it is straightforward to demonstrate that

(2.14) λf (t) = 2 λg (t).

A second way to define characteristic length scales of turbulence from the auto-correlation
functions consists in using their integral properties. Integral scales or Taylor macro-scales
are defined by
Z ∞ Z ∞
(2.15) Λf (t) = f (r, t) dr and Λg (t) = g(r, t) dr .
0 0

Λf is illustrated in figure 2.5(b). As for the micro-scales, the first Kármán–Howarth formula
provides a relation between these two definitions in HIT :

(2.16) Λf (t) = 2 Λg (t).

This equality is valid only if one assumes that f goes to zero faster than 1/r when r goes to
infinity, which is generally confirmed by experiments.

26
1 1
f (r, t)

f (r, t)
0 0
λf Λf

r r
(a) (b)

Figure 2.5 – Definition of Taylor length scales. (a) Longitudinal micro-scale, the dotted line repre-
sents the osculating parabola of the correlation function at the origin. (b) Longitudinal macro-scale,
the area between the horizontal line y = 1 and the curve for r < Λf is equal to the area between
the curve and the horizontal axis for r > Λf .

2.5 Dynamics of the double correlations


Up to now, the two-point correlation theory has been built using the continuity equation
but not the momentum transport equation. In the following, we consider the consequence of
full the Navier–Stokes model on the dynamics of the correlations. An transport equation for
the double velocity two-point correlations can be derived likewise we did for the transport
equation for the Reynolds stresses (one-point correlations). This is achieved in four steps :

(1) [Eq : ui ]P × uj P ′ + [Eq : uj ]P ′ × ui P ;


(2) Averaging ;
(3) Recasting the derivatives using rk = xk |P ′ − xk |P ;
(4) Simplifications for divergence free HIT.

The complete demonstration is given in appendices A and B. We thus get a transport


equation for the double correlations, known as the Kármán–Howarth equation :

∂Qi,j ∂ ∂ 2 Qi,j
(2.17) − (Sik,j + Sjk,i ) = 2ν
∂t ∂rk ∂rk ∂rk

It can be observed that


1. the Kármán–Howarth equation stays open (triple correlations) ;
2. the triple correlations are coming from the inertial terms ;

27
3. the Kármán–Howarth equation can be also derived for the longitudinal auto-correlation
function f (see demonstration in appendix B) :
 
∂u′2 f u′3 ∂r4 k u′2 ∂ ∂f
(2.18) − 4 = 2ν 4 r4
∂t r ∂r r ∂r ∂r

Loitsyanski invariance : a first glimpse to the evolution of the turbulence charac-


teristic length scales The equation (2.18) enables to obtain a first and very important
information regarding the physics of turbulence. Indeed, the multiplication of the equation
by r4 before the integration between zero and infinity leads to :
 Z ∞  Z ∞ 4 Z ∞  
∂ ∂r k ∂ ∂f
u′2 r4 f dr = u′3 dr + 2νu′2 r4 dr
∂t 0 0 ∂r 0 ∂r ∂r
 ∞
 ∞ ∂f
= u′3 r4 k 0 + 2νu′2 r4 .
∂r 0

The right hand side is nil if both k and ∂f /∂r go to zero faster than 1/r4 when r goes to
infinity. Under this assumption, it can be deduced that the quantity
Z ∞
u′2 J4 = u′2 r4 f dr
0

does not vary with time in an divergence free HIT. This quantity is called the Loitsyanski
integral. The quantity J4 is the momentum of order 4 for the auto-correlation function 1 and
it has the dimension of a length to the power 5. Since the turbulent kinetic energy decreases
in time [see equation (1.8)] (and so does u′ ), one can conclude that :

In divergence free HIT and under the assumption of Loitsyanski invariance,


— the characteristic velocity scale u′ decreases in time ;
1/5
— the characteristic length scale L = J4 increases in time,
′2 5
in such a way that the product u L remains constant.

In reality, the Loitsyanski invariance hypothesis has been put into question both on an
experimental and theoretical basis since the 50s (Proudman and Reid [6]). It is currently
admitted that the Loitsyanski integral diverges slowly (like tα with α ∼ 0.16–0.25) in HIT,
which implies that the above result about the increase of the characteristic length scale is
reinforced.

Local restriction of the Kármán–Howarth equation A relation between the turbu-


lent kinetic energy dissipation rate and the micro-scale can be obtained when r is going to
zero in the equation (2.18). Indeed, the equation can be rewritten as :
   
∂f ∂u′2 ∂k k ∂2f 1 ∂f
u′2 +f − u′3 − 4u′3 = 2νu′2 2 + 8νu′2 ,
∂t ∂t ∂r r ∂r r ∂r
R ∞1. nMore generally, the integral momentum of order n for the auto-correlation function, defined as
r f dr, is denoted Jn .
0

28
which gives for r = 0 :
  2
 
∂u′2 ′3 ∂k
′3 k ′2 ∂ f ′2 1 ∂f
−u − 4u = 2νu + 8νu .
∂t ∂r r=0 r r=0 ∂r2 r=0 r ∂r r=0
It has been mentioned earlier that when r goes to zero, the triple correlation function k is
proportional to r3 . This implies the cancellation of the two undetermined terms of the left
hand side. Furthermore, it can be easily demonstrated from the asymptotic development of
f [equation (2.10)] that  
1 ∂f ∂ 2 f
= .
r ∂r r=0 ∂r2 r=0
This leads to
2
∂u′2 ′2 ∂ f u′2
= 10 νu = −20 ν
∂t ∂r2 r=0 λ2f
By identification with the transport equation for turbulent kinetic energy (1.8) in HIT, it
comes :

u′2
(2.19) ǫ = 30 ν
λ2f

This relation is used by experimentalists to evaluate the dissipation rate from the measure
of the micro-scales. This is why λf is often referred as « dissipation scale » Ȧs it will be
seen further, it does not imply that this length scale is characteristic of the size of dissipative
vortices.

2.5.1 Linear solution


In the previous section, it has been seen that the Kármán–Howarth equation is open
and cannot be integrated without a « model » for the triple correlations. Nevertheless, the
simplest model, which consists in the cancellation of these correlations, is useful since it
can be physically interpreted : if one reminds the inertial origin of these terms, neglecting
them compared to the viscous terms on the right hand side appears as a valid simplification
for small turbulent Reynolds numbers. With this simplification, the equation (2.18) can be
solved analytically and the associated linear solution can be obtained from the simplified
equation :
 
∂u′2 f u′2 ∂ 4 ∂f
(2.20) = 2ν 4 r .
∂t r ∂r ∂r
It is a diffusion equation for which self-similar solutions can be sought under the form

f (r, t) = f (η) with η = r/l(t),

where l(t) is a length scale, u′ (t) being the imposed velocity scale. Derivatives with respect
to time and similarity variable are respectively denoted by the letters t and η in subscript.
Then, it can be written :
∂f rlt ∂f 1 ∂2f 1
= − 2 fη , = fη with 2
= 2 fηη .
∂t l ∂r l ∂r l

29
These expressions are introduced in the equation (2.20) in which derivatives have been
expanded before dividing by u′ u′t . It yields :
 
u ′ lt 2ν u fη
− ′ × ηfη + 2f = 2 × fηη + 4 .
ut l l ut η
This differential equation for f (η) gets a solution only if the coefficients (functions of t) are
constant (independent variables) which leads to the following compatibility conditions
2ν u′ u ′ lt
= A and =B
l2 u′t u′t l
where A and B are non-dimensional constants. At this stage, the temporal evolution of the
characteristic scales can be deduced :
r
νBt
l=2 and u′ = α t1/(2B) ,
A
where α is a dimensional constant which only depends on the initial conditions. The equation
for the auto-correlation function can now be rewritten as
 
4A
Afηη + + Bη fη − 2f = 0.
η
Under this classical form, a solution can be sought as the exponential of an even polynomial
(f is even). It gives

f = exp −η 2 by taking A = −1/5 and B = −2/5.

The final form of the solution is :

u′2 (t) = α2 t−5/2



l(t) = 8νt
 
r2
f (r, t) = exp − .
8νt

Here again, one can see that the characteristic length scales increase in time while the
turbulent kinetic energy decreases. To this respect, it can be noted that, the Loitsyanski
hypothesis assuming that the auto-correlation function is decreasing faster than 1/r4 is
verified since f decreases exponentially. This is not surprising that the product u′2 (t) l5 (t)
is constant in the solution obtained previously.

2.5.2 Self-similar solution without viscous terms


The second simplified form of the Kármán–Howarth equation for which a solution can be
obtained, corresponds to the inviscid limit for which inertial effects (triple correlations) are
predominant over viscous effects. In this case, the solution — if it exists — can be considered
as an approximation of the exact solution for large turbulence Reynolds number.
The cancellation of the viscous terms in the Kármán–Howarth equation leads to an
equation which stays open due to the triple correlations. Therefore, an exact solution is still

30
out of reach. One must accept to only look for necessary conditions to the existence of a
self-similar solution. To this end, we start from equation (2.18) in which the right-hand side
is cancelled :
∂  u′3 ∂ 
0 = u′2 f − 4 r4 k ,
∂t r ∂r
∂f ∂u′ k ∂k
= u′2 + 2f u′ − 4u′3 − u′3 .
∂t ∂t r ∂r
As done in the previous section, the self-similar solution is sought under the form

f (r, t) = f (η) and k(r, t) = k(η), with η = r/l(t).

With this change of variable, the simplified equation becomes

lt u′3 k u′3
0 = − u′2 × ηfη + 2u′ u′t × f − 4 × − kη ,
l l η l
lt lu′ k
0 = − ′ × ηfη + 2 ′2t × f − 4 − kη .
u u η
Here again, it is a differential equation in η with coefficients depending on t. The existence
of a solution requires these coefficients to be constant. Stating lt /u′ = A and lu′t /u′2 = B
one gets
B A
u′ (t) = α (t − t0 ) A−B and l(t) = β (t − t0 ) A−B .
Furthermore, the only solution that satisfies the Loitsyanski invariance, i.e. u′2 l5 = cst.,
corresponds to 2B = −5A. Finally, one gets

−10/7 2/7
u′2 (t) = α (t − t0 ) and l(t) = β (t − t0 ) .

In agreement with the the Loitsyanski invariance, the characteristic length scales are increa-
sing in time while the turbulent kinetic energy is decreasing. It can be noted that, at the
same time, the Reynolds number u′ l/ν is decreasing as Re t ∼ t−3/7 , so that this solution
does not prevent the apparition of a linear solution after a sufficiently long period of time.

2.5.3 Experimental validation


The theoretical results obtained previously can be compared to experimental data obtai-
ned with grid turbulence set-ups. For instance, considering the experimental data of Batche-
lor and Townsend [7, 8], it can be noticed that is impossible to represent the whole temporal
decrease with only one power function. The evolution of u′2 with the distance to the grid
is plotted in figure 2.6 for two series of measures. The size of the grid mesh and the mean
velocity are different for each set but the grid mesh Reynolds number (Re M = U0 M/ν) is the
same. Figure 2.6(a) shows that both experimental curves are connecting. On figure 2.6(b)
where the same data are plotted in a logarithmic scale, one can see that the exponents are
different for both the initial and the final phase :
— in the initial zone, the turbulent kinetic energy decreases like 1/xn with n of order
one (here 1.1) ;

31
2.5 10−3

U0 = 1.5 m/s and M = 0.635 cm


∝ t−1.1
2.0 U0 = 6.2 m/s and M = 0.159 cm

10−4
u′2 /U02 × 104

1.5

u′2 /U02
1.0
10−5

0.5

∝ t−1.9
0 10−6
0 200 400 600 800 1000 10 100 1000
x/M ∝ t x/M ∝ t

(a) (b)

Figure 2.6 – Energetic decrease for grid turbulence experiments of Batchelor and Townsend. (a)
Linear scales. (b) Logarithmic scales. Two series of measures are plotted, for which the grid mesh
Reynolds number is the same ReM = 650. On figure (b), one can observe the characteristic slopes of
decrease laws in both the initial phase with an exponent of 1.1 and the final phase with an exponent
of 1.9.

— in the final zone, the exponent is perceptibly larger, of the order two (here 1.9).
In the initial zone, the measured exponent has to be compared with 10/7 ≈ 1.42 obtained
for the inviscid self-similar solution corresponding to large turbulence Reynolds numbers. In
the final zone, the measured value has to be compared with 5/2 = 2.5 obtained for the linear
solution which corresponds to small turbulent Reynolds numbers. The Reynolds numbers of
the experiment are coherent with this interpretation : the Reynolds number is large in the
initial phase while it is small in the final phase.

Regarding the temporal evolution of the characteristics length scales, figure 2.7 shows
the evolution of the Taylor micro-scale as a function of the distance to the grid for the same
experiments. In figure 2.7(a), one can verify that these scales are increasing with the distance
to the grid for two values of the grid mesh Reynolds number. In figure 2.7(b) where data for
ReM = 650 are plotted in logarithmic scale, one can see the difference for the two phases
observed previously :
— in the initial phase, the value of the exponent is 0.48 ;
— in the final phase, the value is noticeably weaker (here 0.32).
These values are in quite good agreement with those derived theoretically : 2/7 ≈ 0.29 and
0.5, respectively.

32
0.5 1

0.4
∝ t0.32
0.3
λ (cm)

λ (cm)

0.2

0.1 Re M = 650
Re M = 950 ∝ t0.48

0 0.1
0 200 400 600 800 1000 100 1000
x/M ∝ t x/M ∝ t
(a) (b)

Figure 2.7 – Evolution of length scales grid turbulence experiments of Batchelor and Townsend. (a)
Linear scales. (b) Logarithmic scales. In figure (b), one can see the characteristic slopes of increase
laws in both the initial phase with an exponent of 0.48 and the final phase with an exponent of
0.32.

33
34
Chapter 3

Spectral approach in
homogeneous isotropic
turbulence

The one-point statistical analysis proved that the evolution of HIT is characterised by
an energy decrease driven by the value of the dissipation rate. The two-point correlation
theory provides informations regarding the definition and the evolution of the turbulence
characteristic length scales. Regarding the evolution, Even if no solution can be obtained for
the correlations transport equation , it appears that some elements (degenerated solutions)
support the idea that the characteristic length scales are increasing in time while the tur-
bulent kinetic energy is decreasing. This is confirmed by both grid turbulence experiments
and by direct numerical simulations of HIT. Considering that
— these evolutions apply to « characteristic » or « mean » scales ;
— the definition of the turbulent regime implies a hierarchy of vortical structures which
size span the whole range of scales from the scale of the average motion down to the
scale of dissipative structures,
one can legitimately ask the following questions :

1. Are these evolutions valid for all vortical structures whatever their size or only to
some of them ?

2. Do interactions between vortices of different size exist ?

3. How can the increase of the characteristic length scales can be explained ?

The spectral approach which is introduced in this chapter aims at answering these questions.

35
3.1 Spectral tensor
3.1.1 Definition
The spectral tensor φ is defined as the 3D Fourier transform of the tensor Q of the double
velocity two-point correlations. In HIT, it can be written as
+∞
ZZ Z
1
φij (~κ, t) = Qi,j (~r, t) exp (−i~κ · ~r) d3~r,
(2π)3
−∞

and conversely
+∞
ZZ Z
Qi,j (~r, t) = φij (~κ, t) exp (i~κ · ~r) d3~κ,
−∞

where
— ~r, is a vector in the physical space associated to an elementary volume d3~r = dr1 dr2 dr3 ;
— ~κ is a vector in the spectral space called wave vector associated with an elementary
« volume » d3~κ = dκ1 dκ2 dκ3 . The vector components scale like the inverse of a
length (wavenumbers)

3.1.2 Properties
The properties of the spectral tensor result directly from those of the correlation tensor.
The most important are determined as follow.
1. The isotropy hypothesis applied to the double correlation tensor gives
Qi,j (~r, t) = Qj,i (~r, t) ,
from which it can be deduced immediately that the spectral tensor is symmetric :

φij (~κ, t) = φji (~κ, t) .

2. The spatial homogeneity hypothesis give the following relation


Qi,j (~r, t) = Qj,i (−~r, t) ,
which can be used to demonstrate that the spectral tensor is real-valued. Indeed,
taking the conjugate 1 of the spectral tensor, one gets :
+∞
ZZ Z
1
φ∗ij (~κ, t) = 3 Q∗i,j (~r, t) exp (i~κ · ~r) d3~r.
(2π)
−∞

As the components of the correlation tensor are real, it can be inferred that
+∞
ZZ Z
1
φ∗ij (~κ, t) = 3 Qi,j (~r, t) exp (i~κ · ~r) d3~r.
(2π)
−∞

1. In general, the conjugate of a complex number Z is denoted Z ∗ .

36
The homogeneity property for Qi,j together with the variable change ~r′ = −~r lead to
+∞
ZZ Z
1
φ∗ij (~κ, t) = 3 Qj,i (−~r, t) exp (i~κ · ~r) d3~r
(2π)
−∞
+∞
ZZ Z
1   h  i  
= 3 Qj,i r~′ , t exp −i~κ · r~′ d3 r~′ .
(2π)
−∞

Hence
φ∗ij (~κ, t) = φji (~κ, t) .
Taking into account the property of symmetry demonstrated just above, one gets

φ∗ij (~κ, t) = φij (~κ, t) ,

which implies that the spectral tensor is real-valued in HIT.

3.1.3 Generic form of the spectral tensor in HIT


As done in chapter 2 for the double correlation tensor, tensor algebra enables to take
into account the invariance properties associated with homogeneity and isotropy to find a
minimalist and generic expression for the spectral tensor :

φij (~κ, t) = A(κ, t)κi κj + B(κ, t)δij

This form can be simplified further by use of the continuity equation — the isovolume
condition — which yields for the double correlation tensor

Qi,j (r, t) = 0.
∂rj
The Fourier transform of this equation leads to

iκj φij (κ, t) = 0

then, introducing the generic form of the spectral tensor in HIT

0 = A(κ, t)κi κj κj + B(κ, t)κj δij


= A(κ, t)κi κ2 + B(κ, t)κi
 
= κi A(κ, t)κ2 + B(κ, t)

from which a relation between the two scalar functions is obtained : B = −Aκ2 . The final
form of the spectral tensor in isovolume HIT is

φij (~κ, t) = A(κ, t) κi κj − κ2 δij

It appears that the spectral tensor can be expressed with a single scalar function of only
one scalar variable κ (wave vector modulus) for a given t. The direct consequence is that
the spectral tensor can be fully characterised in HIT with a « one-dimensional » spectra.

37
κ3

κ3


θ

~r

θ γ
κ1 κ2

~eγ


κ3

κ1 κ2

~eγ

~r ~κ
θ dκ

κ sin θ ~eγ

Figure 3.1 – Spectral density of energy is obtained by integration in the spectral space.

3.2 Spectral density of energy


The first one-dimensional spectrum that we introduce aims at characterising the spectral
content of the turbulent kinetic energy k, which is the local restriction of the trace of the
two-point correlation tensor, i.e.

1 1
k(t) = ui ui = lim Qi,i (~r, t) .
2 2 r→0

Its spectral content is accessible through the correlation tensor since

+∞
ZZ Z
Qi,i (~r, t) = φii (~κ, t) exp (i~κ · ~r) d3~κ.
−∞

This triple integral can be simplified thanks to the properties of the spectral tensor in HIT.
The spherical coordinate system used is illustrated in figure 3.1. The infinitesimal volume
element is then

d3~κ = dκ × κ dθ × κ sin θ dγ.

It is assumed, without any loss of generality, that the separation vector ~r is aligned along

38
the axis κ3 . It can be written :
Z +∞ Z π Z 2π  
Qi,i (r, t) = φii (κ, t)κ2 exp(iκr cos θ) sin θ dγ dθ dκ
0 0 0
Z +∞ Z π  Z 2π  
2
= φii (κ, t)κ exp(iκr cos θ) sin θ dγ dθ dκ
0 0 0
Z +∞  Z π 
2
= 2π φii (κ, t)κ exp(iκr cos θ) sin θ dθ dκ .
0 0

Then, the inner integral can be calculated using u = cos θ :


Z π Z +1
2i sin(κr)
exp(iκr cos θ) sin θ dθ = exp(iκru) du = ,
0 −1 iκr

to obtain Z +∞
sin(κr)
Qi,i (r, t) = 4π φii (κ, t)κ2 dκ,
0 κr
and finally, taking the local restriction :
Z +∞
1
k(t) = lim Qi,i (r, t) = 2π φii (κ, t)κ2 dκ.
2 r→0 0

This relation enables to define the spectral density of energy, also called energy spectrum,
E(κ, t) as follow :

Z +∞
E(κ, t) = 2π κ2 φii (κ, t) with k(t) = E(κ, t) dκ .
0

It implies that the turbulent kinetic energy is strictly equal to the integral of the spectral
density of energy over all wavenumbers.
The generic form obtained for the spectral tensor in section 3.1 can be used to relate the
two spectra E(κ, t) and A(κ, t) :

E(κ, t) = 2π κ2 φii (κ, t) = 2π κ2 A(κ, t) κi κi − 3κ2 ,

hence
E(κ, t) = −4π κ4 A(κ, t),
and the final expression of the spectral tensor using the spectral density of energy is :

E(κ, t)  κi κj 
φij (κ, t) = δ ij − .
4πκ2 κ2

3.3 One-dimensional spectra


Another possibility to relate the expression of the spectral tensor to the scalar spectrum
of a quantity with a physical meaning, consists in using the longitudinal and transversal

39
conventional correlation functions (f and g) introduced in chapter 2. Keeping in mind that
these two functions are even, their (real) one-dimensional Fourier transforms can be written
with wavenumber κ1 as follow :

Z +∞
1
F11 (κ1 , t) = u′ (t)2 f (r, t) cos(κ1 r) dr ;
π 0
Z +∞
1
F22 (κ1 , t) = u′ (t)2 g(r, t) cos(κ1 r) dr,
π 0

with the inversion relations :

Z +∞
1
f (r, t) = ′ 2 2 F11 (κ1 , t) cos(κ1 r) dκ1 ;
u (t) 0
Z +∞
1
g(r, t) = 2 F22 (κ1 , t) cos(κ1 r) dκ1 .
u (t)2

0

It is possible to relate quite easily these expressions to the spectral tensor since

Z +∞ Z +∞ Z +∞
Q1,1 (~r, t) = φ11 (~κ, t) exp (i~κ · ~r) dκ1 dκ2 dκ3
−∞ −∞ −∞
Z +∞ Z +∞ Z +∞ 
= φ11 (~κ, t) exp (i~κ · ~r) dκ2 dκ3 dκ1 ,
−∞ −∞ −∞

and, without loss of generality, the separation vector ~r is chosen to be aligned along the first
direction. Hence, one can write

Z +∞ Z +∞ Z +∞ 
Q1,1 (r, t) = φ11 (~κ, t) exp (iκ1 r) dκ2 dκ3 dκ1 ,
−∞ −∞ −∞

and, given that Q1,1 is real and even, it yields :

 
Z +∞ Z+∞
Z
Q1,1 (r, t) = 2  φ11 (~κ, t) cos (κ1 r) dκ2 dκ3  dκ1
0
−∞
 
Z +∞ Z+∞
Z
= 2 φ11 (~κ, t) dκ2 dκ3  cos (κ1 r) dκ1 .
0
−∞

By identification with the definition of the spectrum of f , the term in between brackets must
be equal to F11 . It is possible to do a similar development for the transversal conventional

40
κ3


~κ κ2

dθ ✕ ✮
✌ θ

❑ d̺

κ1

Figure 3.2 – Relation between the one-dimensional spectra and the spectral density of energy are
obtained by integration in the plane (κ2 ,κ3 ).

correlation function, which gives :

Z+∞
Z
F11 (κ1 , t) = φ11 (~κ, t) dκ2 dκ3 ;
−∞
Z+∞
Z
F22 (κ1 , t) = φ22 (~κ, t) dκ2 dκ3 .
−∞

3.3.1 Relation between the one-dimensional spectra and the spec-


tral density of energy
One-dimensional spectra can be related quite straightforwardly to the spectral density
of energy. To do so, the previous result can be rewritten as

Z+∞
Z Z+∞
Z  
E(κ, t) κ21
F11 (κ1 , t) = φ11 (~κ, t) dκ2 dκ3 = 1 − 2 dκ2 dκ3 .
4πκ2 κ
−∞ −∞

Using the polar coordinates system represented in figure 3.2 — κ2 = ̺ cos θ and κ3 =
̺ sin θ — it gives :
Z +∞ Z 2π  
E(κ, t) κ21
F11 (κ1 , t) = 1 − 2 ̺ dθ d̺,
0 0 πκ2 κ

41
then, integrating in θ and noticing that κ2 = κ21 + ̺2 and κ dκ = ̺ d̺ :
Z +∞  
E(κ, t) κ21
F11 (κ1 , t) = 1 − 2 ̺ d̺
0 2κ2 κ
Z +∞  
E(κ, t) κ21
= 1 − 2 dκ.
κ1 2κ κ

The same calculus with the spectra of the transversal correlations leads to
Z +∞   Z +∞  
E(κ, t) κ22 E(κ, t) κ23
F22 (κ1 , t) = 1 − 2 dκ and F33 (κ1 , t) = 1 − 2 dκ.
κ1 2κ κ κ1 2κ κ

The isotropy hypothesis implies that F22 = F33 = (F22 + F33 )/2, and we get the final
expressions :
Z +∞  
E(κ, t) κ2
F11 (κ1 , t) = 1 − 12 dκ ;
κ1 2κ κ
Z +∞  
E(κ, t) κ2
F22 (κ1 , t) = 1 + 21 dκ.
κ1 4κ κ

3.3.2 Differential relations between spectra


The relations derived in the previous section enable to obtain the one-dimensional spectra
from the spectral density of energy. In practice, the former are easier to measure experimen-
tally so it is interesting to invert the relations to obtain the spectral density of energy. This
can be done by deriving with respect to κ1 the expression relating F11 (κ1 , t) to E(κ, t) :
Z +∞   
∂F11 (κ1 , t) ∂ E(κ, t) κ2
= 1 − 21 dκ
∂κ1 ∂κ1κ1 2κ κ
Z +∞ Z +∞ 
∂ E(κ, t) E(κ, t)
= dκ − κ21 dκ
∂κ1 κ1 2κ κ1 2κ3
Z +∞
E(κ1 , t) E(κ1 , t) E(κ, t)
= − + κ21 3 − 2κ1 dκ,
2κ1 2κ1 κ1 2κ3

hence Z +∞
1 ∂F11 (κ1 , t) E(κ, t)
− = dκ.
κ1 ∂κ1 κ1 κ3
By deriving a second time, one gets

1 ∂ 2 F11 (κ1 , t) 1 ∂F11 (κ1 , t) E(κ1 , t)


− 2 + 2 =− ,
κ1 ∂κ1 κ1 ∂κ1 κ31

and finally

∂ 2 F11 (κ1 , t) ∂F11 (κ1 , t)


(3.1) E(κ1 , t) = κ21 2 − κ1 ,
∂κ1 ∂κ1

42
which can be rewritten under the following compact form

   
∂ 1 ∂F11 (κ1 , t) ∂ 1 ∂F11
(3.2) E(κ1 , t) = κ31 = κ1 .
∂κ1 κ1 ∂κ1 ∂κ1 κ1 ∂κ1

3.3.3 Spectral equivalent to the first Kármán–Howarth formula


We are now going to see that, as for the longitudinal and transversal conventional corre-
lation functions, the associated two one-dimensional spectra are also related by a differential
equation which can be obtained from the relations between the one-dimensional spectra and
the spectral density of energy :
Z +∞  
E(κ, t) κ21
F22 (κ1 , t) = 1 + 2 dκ
κ1 4κ κ
Z +∞   Z +∞
E(κ, t) κ21 E(κ, t) κ21
= 1 − 2 dκ + dκ
κ1 4κ κ κ1 2κ κ2
Z
1 κ2 +∞ E(κ, t)
= F11 (κ1 , t) + 1 dκ.
2 2 κ1 κ3

Using one of the intermediate results obtained in the previous section, the integral of the
right hand side can be calculated to give :

1 κ1 ∂F11 (κ1 , t)
F22 (κ1 , t) = F11 (κ1 , t) − .
2 2 ∂κ1

3.3.4 Generic form of the one-dimensional spectra


For a grid experiment, the Taylor hypothesis enables to convert temporal separation into
spatial separation (along the mean flow direction) which is used as a framework for the two-
point correlation theory. One can then access quite easily to the one-dimensional spectrum
F11 by applying a Fourier transform to the longitudinal correlation function. The relations
obtained in the previous section allow then to get in return the one-dimensional spectrum
F22 as well as the spectral density of energy E.
Figure 3.3 presents the form of the spectra obtained from the results of grid turbulence
experiments. The evolution near the origin of one-dimensional spectra and the spectral
density differ : the former have their maximum value for κ = 0. These values can be related
to the macro-scales of Taylor using the definition of F11 and F22 :
Z ∞
u′2 u′2 u′2
F11 (0, t) = f (r, t) dr = Λf , and similarly F22 (0, t) = Λg .
π 0 π π

A close look at the values at κ = 0 validates the isotropy relation Λf /Λg = 2. On the
other hand, the spectral density of energy is cancelling at the origin. This observation can
be confirmed analytically with an asymptotic development of the one-dimensional spectrum
F11 when κ → 0, before getting the asymptotic development of the spectral density of energy

43
0.0025

0.002
F11 (κ, t)

F22 (κ, t)
0.0015
E(κ, t)

0.001

0.0005

0
0 200 400 600 800 1000
κ

Figure 3.3 – Spectral density of energy and associated one-dimensional. Data are taken from a
grid turbulence experiment, Roches [9].

thanks to the differential relation (3.1). Thus one gets :


Z +∞
u′2
F11 (κ, t) = f (r, t) cos(κr) dr
π 0
Z +∞ ∞
u′2 X (−1)n (κr)2n
= f (r, t) dr
π 0 n=1
(2n)!
Z
u′2 X (−1)n κ2n +∞ 2n

= r f (r, t) dr ,
π n=1 (2n)!
|0 {z }
J2n

which is valid only if the integral momentums J2n are converging. Deriving twice this ex-
pression with respect to r and using the relation (3.1), it yields

u′2 X (−1)n (n − 1) J2n
E(κ, t) = 2 × κ2n .
π n=1 (2n − 1)!

Keeping only the first order term, the behaviour near the origin is of the form :

u′2 J4
E(κ, t) ∼ × κ4 .
κ→0 3π

On the numerator of the right hand side, the quantity u′2 J4 is the Loitsyanski integral.
When Loitsyanski invariance is verified, the form of the spectrum for small wavenumbers is
conserved in time.

44
3.4 Spectral dynamics
The evolution equation for the spectral tensor is obtained from the evolution equation
of the double correlation tensor. In HIT, the equation is

∂ ∂ ∂ 2 Qi,j (r, t)
Qi,j (r, t) − [Sik,j (r, t) + Sjk,i (r, t)] = 2ν
∂t ∂rk ∂rk ∂rk

according to the results obtained in the section 2.5. The triple correlations appear in the
Fourier transform of this equation. In HIT, the triple correlations are real odd functions.
This is why it is more convenient to work with the Fourier transforms of the functions iSij,k .
We define
+∞
ZZ Z
1
ψijk (~κ, t) = 3 iSij,k (r, t) exp (−i~κ · ~r) d3~r,
(2π)
−∞

and the inversion relation


+∞
ZZ Z
Sij,k (~r, t) = −i ψijk (r, t) exp (i~κ · ~r) d3~κ.
−∞

The Fourier transform of the evolution equation for the double correlations is


φij (κ, t) − κk [ψikj (κ, t) + ψjki (κ, t)] = −2νκ2 φij (κ, t)
∂t

Taking the trace of it (contraction of the indexes i and j), one gets the evolution equation
for the spectral density of energy :


E(κ, t) − 4πκ2 κk ψiki (κ, t) = −2νκ2 E(κ, t)
∂t

since E(κ, t) = 2πκ2 φii (κ, t). It is usually written in the following form, labelled Lin equa-
tion :

E(κ, t) = T (κ, t) − 2νκ2 E(κ, t),
∂t

where the definition of the term T (κ, t) is straightforward. This result calls two important
remarks :
1. Similarly to the transport equations for the Reynolds stress and the evolution equa-
tion for the two-point correlations, the evolution equation for the spectral density
stays an open equation. Its solving requires a modelling of the term T (κ, t) co-
ming from the triple correlations. This term, as for the opened terms of the above-
mentioned equations, has an inertial origin.
2. As it will be seen next, the integration of the equation over all the wavenumbers will
enable to give a physical meaning of the all terms on the right hand side.

45
3.4.1 Physical mechanisms at stake in the spectral dynamics
By integration over all wavenumbers, the Lin equation enables to retrieve the transport
equation for the turbulent kinetic energy :
Z +∞ Z +∞ Z +∞

E(κ, t) dκ = T (κ, t) dκ − 2νκ2 E(κ, t) dκ.
∂t 0 0 0
| {z }
∂k(t)/∂t
By identification with the transport equation 1.8 for k obtained in section 1.2.2, it appears
that the sum of the two terms on the right hand side is equal and opposite to the value of
the dissipation rate ǫ. The question is now to find the physical origin of each term considered
separately.

Dissipation spectrum
First, we focus on the term with an explicit viscous origin 2νκ2 E(κ, t). Starting from
equation (3.2)  
3 ∂ 1 ∂F11 (κ, t)
E(κ, t) = κ ,
∂κ κ ∂κ
the second integral of the right hand side can be written as
Z +∞ Z +∞  
2 5 ∂ 1 ∂F11 (κ, t)
2νκ E(κ, t) dκ = 2ν κ dκ .
0 0 ∂κ κ ∂κ

Two successive integrations by parts enable to simplify this integral, provided that ∂F ∂κ and
11

4 3
F11 go to zero faster than 1/κ and 1/κ respectively when κ → ∞ :
Z +∞ Z +∞  
∂ 1 ∂F11 (κ, t)
2νκ2 E(κ, t) dκ = 2ν κ5 dκ
0 0 ∂κ κ ∂κ
 +∞ Z +∞  
4 ∂F11 (κ, t) 1 ∂F11 (κ, t)
= 2ν κ − 2ν 5κ4 dκ
∂κ 0 0 κ ∂κ
Z +∞
∂F11 (κ, t)
= −10ν κ3 dκ
0 ∂κ
h i+∞ Z +∞
= −10ν κ3 F11 (κ, t) + 10ν F11 (κ, t) 3κ2 dκ
0 0
Z +∞
= 30ν κ2 F11 (κ, t) dκ.
0

It is thus straightforward to show that this expression is rigorously equal to the dissipation
rate in HIT. Indeed, equation (2.19) relates ǫ to the value of the longitudinal auto-correlation
function for r = 0 :
∂ 2 f (r, t)
ǫ = −15νu′2 .
∂r2 r=0
Using the definition of the one-dimensional spectrum F11
Z +∞
1
f (r, t) = ′ 2 2 F11 (κ, t) cos(κ, r) dκ ,
u (t) 0

46
it can be deduced that
Z +∞
∂ 2 f (r, t) 1
=− ′ 2 2κ2 F11 (κ, t) cos(κr) dκ.
∂r2 u (t) 0

The dissipation rate can be written under the following form

Z +∞
∂ 2 f (r, t)
ǫ = −15νu ′2
= 30ν κ2 F11 (κ, t) dκ ,
∂r2 r=0 0

which leads to the wanted result


Z +∞
ǫ= 2νκ2 E(κ, t) dκ.
0

In HIT, the value of the dissipation rate can be related to the spectral density
of energy : Z +∞
ǫ= 2νκ2 E(κ, t) dκ.
0
The function
E(κ, t) = 2νκ2 E(κ, t)
represents the dissipation spectrum. The factor κ2 shifts this spectrum — with
respect to the spectrum E(κ, t) — towards the large wavenumbers, meaning
that the most dissipative structures are small compared to the most energetic
ones.

Figure 3.4 shows the spectral density of energy obtained from a grid turbulence experiment
together with the associated dissipation spectrum. The most energetic wavenumber (κE )
and the most dissipative wavenumber (κD ) are also plotted. One can observe the shift of the
second spectrum towards large wavenumbers and, if κD is significantly larger than κE , it is
possible to distinguish two classes of vortical structures within the fluctuating motion :
— a class of big structures weakly dissipative for which κ << κD ;
— a class of small structures weakly energetic for which κ >> κE .

Inter-spectrum energy transfer

As a direct consequence of the previous result, the integral over all wavenumbers of the
term T (κ, t) is zero. This implies that if this term is positive for some wavenumbers, it will
necessary be negative for others. It will then contribute to the increase of the energy carried
by the former at the cost of the energy carried by the latest.

47
10−1

κE κD
10−2
E(κ, t)/k(t)
E(κ, t)/ǫ(t)
10−3

10−4

10−5

10−6

10−7

10−8
0.1 1 10 100 1000
κ × k 3/2 /ǫ

Figure 3.4 – Spectral density of energy and associated dissipation spectrum. One can observe
the shift of the second spectrum towards the large wavenumbers. Data are coming from a grid
turbulence experiment, Roches [9].

The nonlinear term T (κ, t) coming from the triple correlations in the evolution
equation for the spectral density of energy verifies
Z +∞
T (κ, t) dκ = 0 .
0

It does not contribute to the variation of the total turbulent kinetic and can
be seen as a term of inter-spectrum energy transfer.

Identification of T (κ, t) as the net energy transfer through the spectrum at wavenumber
κ leads to define a spectral flux of energy under the form
Z κ
W (κ, t) = − T (κ, t) dκ,
0

which represents the amount of energy transferred from the structures corresponding to
wavenumber smaller than κ towards those associated with wavenumber larger than κ.

3.4.2 Model of balance for the Lin equation and energy cascade
In homogeneous isotropic turbulence, the evolution equation for the turbulent kinetic
energy prescribes that the total amount of energy decreases in time at a rate imposed by
the dissipation rate. If we now consider the amount of energy carried by the wavenumber

48
κ, a careful observation of the different terms of the Lin equation shows that its temporal
evolution is more complex since it results from the action of two terms :
— the dissipation term, negative by definition, tends to decrease the amount of energy ;
— an extra term of inter-spectrum energy transfer, with a sign not known a priori, which
can locally lead to an increase or a decrease of the amount of energy.
The indeterminacy of the last point prevents from concluding on the net result at wave-
number κ. In these conditions, only experiments can bring further informations. Inspection
of available data as well as the consideration of the properties of both the transfer and
dissipation terms enable to get the generic trend for the turbulent kinetic energy balance
as described in the Lin equations, see figure 3.5. It gives informations on the actual energy
balance in HIT for each wavenumber. It can be observed that
1. the spectral density of energy decreases at all scales : −∂E(κ)/∂t > 0, ∀κ ;
2. the observed decrease for large scales (κ ≪ 1) results mainly from a removal of energy
through inter-spectrum transfer : T (κ) < 0 ;
3. the observed decrease for small scales (κ ≫ 1) results from the difference between
an energy input through inter-spectrum transfer (T (κ) > 0) and a removal through
viscous dissipation (−E(κ) < 0).
As plotted in figure 3.5,
R κthe curve of T (κ) shows that the spectral flux of energy is always
positive (W (κ, t) = − 0 T (κ, t) dκ > 0). It implies that, whatever the wavenumber, the
inter-spectrum transfer operates from the large structures towards the small ones. This is
called the energy cascade : the large vortices give energy to smaller ones, which in turn give
energy to even smaller ones, and so on. The end of the process consists in the « thermalisa-
tion » of the energy (the conversion into internal energy) of the smallest vortices under the
action of the viscous dissipation.

Remark The notion of energy cascade is central in the knowledge of turbulence, but one
must keep in mind that is only a statistical notion, meaning that it is true in average. For
instance, nothing prevents two vortices from merging into a bigger one. In this case, it only
constitutes a statistically rare event.

3.5 Kolmogorov hypotheses


The similarity hypotheses of Kolmogorov [10, 11] are based on the notion of quasi-
statistical equilibrium of small scales. It can be used when the energy and dissipation spectra
are sufficiently separated so that there exists a wavenumber κI such as
κE << κI << κD .
In these conditions, the Lin equation can be integrated from κI to infinity to give
Z ∞  Z ∞ Z ∞

E(κ, t) dκ = T (κ, t) dκ − 2νκ2 E(κ, t) dκ
∂t κI κI κI
| {z }
EI (t)
Z ∞
= W (κI , t) − 2νκ2 E(κ, t) dκ
κI
| {z }
ǫI (t)

49
Production (+)

− ∂E(κ)/∂t
E(κ)
T (κ)

0
Destruction (−)

0 κ
(a)

A− = A+ At = ǫ
∂E(κ)/∂t

At
A+
T (κ)

0 0
ǫ
A−
E(κ)

0 κ 0 κ
(b) (c)

Figure 3.5 – Balance model for the Lin equation. It is verified that (a) the sum of the three terms
is zero for all wavenumbers, (b) the amount of energy taken from the small wavenumbers (area A− )
is equal to amount of energy injected at large wavenumbers (area A+ ), (c) the variation rate of the
turbulent kinetic energy (area At ) is equal to the dissipation rate (ǫ).

50
Considering now that the relative position of κI to κE and κD , it can be noticed that
1. as κI << κD , the range of wavenumbers [κI , +∞[ carries most of the dissipation and
then
ǫI (t) ≈ ǫ(t)

2. as κE << kappaI , the range of wavenumbers [κI , +∞[ carries a negligible part of the
turbulent kinetic energy
∂EI (t) ∂k
− << − ;
∂t ∂t
and since ∂k/∂t = −ǫ(t), it gives

∂EI (t)
− << ǫ(t).
∂t

From the preceding, it results the « quasi-equilibrium »

(3.3) ǫ(t) ≈ W (κI , t).

This relation states that small scales — defined by κ > κI — dissipate an amount of energy
imposed by the one cascading from the energetic scales. In other words :

The amount of energy dissipated by the smallest scales adjusts itself instan-
taneously to the (comparatively) slow variations of the motion at large scales.

3.5.1 Statement of the Kolmogorov similarity hypotheses


Originally introduced with an hypothesis of local statistical isotropy for the increments
of velocity, the Kolmogorov similarity hypotheses were used shortly after for the spectral
density of energy. It is in this second framework that they will be now presented and dis-
cussed.

First similarity hypothesis of Kolmogorov This first hypothesis states that if the
Reynolds number is large enough, there exists a zone of statistical equilibrium where the
shape of the spectrum depends only on the dissipation rate (ǫ) and the fluid viscosity (ν).
From a dimensional analysis, it leads to :
 
E(κ, t) κν 3/4
=F ,
ǫ(t)1/4 ν 5/4 ǫ(t)1/4

where F is a universal function.

Second similarity hypothesis of Kolmogorov The second hypothesis completes the


first one when the Reynolds number is even larger. It states that if the Reynolds number is
large enough so that the statistical equilibrium zone exists and contains a significant range
of wavenumbers much smaller than the dissipative ones, then there exists a sub-domain

51
of this statistical equilibrium zone where the shape of the spectrum only depends on the
dissipation rate (ǫ). From a dimensional analysis, it leads to :

E(κ, t)κ5/3
=A
ǫ(t)2/3

where A is a universal constant. The previous relation can be rewritten as


 −5/3
E(κ, t) κν 3/4
=A× ,
ǫ(t)1/4 ν 5/4 ǫ(t)1/4

which shows that


F → A κ̃−5/3 when κ/κD → 0,
where κ̃ = κ ηK , and ηK = ν 3/4 /ǫ1/4 is known as the Kolmogorov length scale.

Building of the « Kolmogorov spectrum » From the hypothesis of the quasi-statistical


equilibrium of small scales and the two similarity hypotheses of Kolmogorov, it is possible to
build a energy spectrum known as the Kolmogorov spectrum. To this end, one must consider
that the population of the turbulent structures can be divided into two classes :
1. the class of the structures that carry energy for which κ is close to κE ;
2. the class of the small structures for which κ/κE is large, this class is itself divided
into two sub-classes :
(a) the inertial range where κ/κD is small
(b) the dissipative range where κ is close to κD
An extra class, sometimes called the class of very large structures, can be added to this
description, for which κ/κE is small. As a consequence, it can be admitted that the plot
of the energy spectrum in logarithmic scales is contained in a envelop constituted of three
successive lines :
— an horizontal line at the ordinate E(κE , t) = max E(κ, t) ;
— a decreasing line with a slope of −5/3 ;
— a vertical line located at a value of κ close to κD (dissipative cut-off).
Figure 3.6 illustrates such a model and lists all the different classes of structures.
Figure 3.7 presents the evolution of the spectrum as a function of the Reynolds number.
The function F to which the spectrum must connect at large wavenumbers — whatever the
Reynolds number — is plotted in figure 3.7(a). The Kolmogorov hypotheses imply that the
inertial range is widening when the Reynolds number increases. In Kolmogorov variables,
this can only be achieved through small wavenumbers, as presented in figure 3.7(b) which
shows the evolution of the spectrum when the Reynolds number increases.

3.5.2 Underlying elements to the Kolmogorov hypotheses


Scale separation at large Reynolds number The idea according to which the increase
of the Reynolds number enhance the separation between energetic structures and dissipative
ones is underlying to the Kolmogorov hypotheses. Furthermore, their main consequence
lies in the normalisation of the dissipative structures by the dissipation rate and the fluid
viscosity, i.e. the following characteristic scales

52
1
107

2 3
105 1 : Statistical quasi-equilibrium
4

103
E(κ, t)

2 : Inertial range

10 3 : Dissipative range

10−1 4 : Energetic range

10−3
10−6 10−4 10−2 1
κ̃

Figure 3.6 – Kolmogorov spectrum for different classes of vortices.

107 107
−5/3 slope −5/3 slope

105 105
one-dimensional spectra

103 103
cut-off cut-off

10 10 Re

10−1 10−1

10−3 10−3
10−6 10−4 10−2 1 10−6 10−4 10−2 1
κ̃ κ̃
(a) (b)

Figure 3.7 – Construction of the Kolmogorov spectrum. (a) Expected trend for the function F.
(b) Expected evolution of the spectrum form when the Reynolds number is increased.

53
 1/4
— ηK = ν 3 /ǫ for the length scales ;
1/4
— υK = [νǫ] for the velocities ;
— τK = [ν/ǫ]1/2 for the time.
One can easily verify that this normalisation is in agreement with the first idea. Indeed, if one
considers — as suggested by experimental results — that the integral scale characterises the
vortices that carry energy, it is possible to deduce from the equation of energy decrease (1.8)
that
u′3
ǫ∝ .
Λf
The ratio between the characteristic scales of energetic structures and dissipative ones can
be then written as :
−1/4  3/4
Λf Λf × ǫ1/4 Λf u′3/4 Λf u ′ Λf 3/4
= ∝ = = ReΛ .
ηK ν 3/4 ν 3/4 ν

It can be noticed that this ratio increases with the Reynolds number in agreement with the
initial assumption.
Furthermore, it is also possible to locate the Taylor micro-scale relatively to Λf and ηK
using the relation in HIT (2.19) :

Λ2f Λf × u′3 /ǫ Λf u ′
Since λ2f = 30νu′2 /ǫ, then 2 ∝ = = ReΛ .
λf νu /ǫ
′2 ν

The Taylor micro-scale appears as an intermediate scale (at large Reynolds number) between
the characteristic scales of the energetic structures on one side and the dissipative ones on
the other side. One must remember that

Λf 1/2 Λf 3/4
∝ ReΛ and ∝ Re Λ .
λf ηK

Inertial and local nature of the inter-spectrum transfer If we denote UI2 (t) =
−5/3
E(κI , t)κI and we take into account the fact that E(κI , t) = Aǫ(t)2/3 κI , it is possible to
demonstrate that
UI κ−1 √ −4/3
I
Re I = = Aκ̃I
ν
κ̃I is small in the inertial range and consequently the Reynolds number Re I is large. It can
be thus inferred that the active physical mechanisms are inviscid, as suggested by the name
« inertial » range. Furthermore, the self-similarity hypothesis implies that the spectrum must
be described with local scales (in κ), and this is the case for the only quantity used : the
dissipation is equal to the amount of energy which « cascades » through the spectrum at
wavenumber κ 2 . « Non-local » interactions are then excluded within the spectrum and the
inter-spectrum energy transfer is built on interactions between structures of similar size :
this is the notion of local transfer. The underlying mechanism is the vortex-stretching.
2. The relation (3.3) is valid within the entire inertial range.

54
Universality and isotropy of small scales The universality of the constant A extends
well beyond the HIT, as soon as the Reynolds number of the flow is sufficiently large. The
reason of this is that the anisotropy results from the way that energy is transferred from the
mean flow to the fluctuating motion. This process is active at large scales and, if the energy
cascade extends over a sufficiently large range of wavenumbers, the anisotropy is somehow
« forgotten » in the the transfer process towards the small scales (vortex-stretching favours
the return to isotropy). Thus, it can be admitted in the following that

At large Reynolds number, the organisation of small scales is universal and


isotropic.

It is generally admitted that the value of the Kolmogorov constant is around 1.6 (in the
range 1.45–1.8).

3.5.3 Experimental verification


Kolmogorov hypotheses The validity of Kolmogorov hypotheses as well as the associa-
ted universality of the small scales representation can be verified with experiments. Figure 3.8
shows the results of numerous experiments taken from the literature. These experiments are
not limited to HIT, one can find results obtained from various configurations of shear flows
(boundary layers, channel flow, jets, etc.) It can be verified that
— all the spectra are following the −5/3 slope (in agreement with the second hypothesis
of Kolmogorov) ;
— all the spectra are superimposed near the dissipative cut-off (in agreement with the
first hypothesis of Kolmogorov) leading to the shape universal of the function F in
this zone ;
— the size of the inertial range (along the line of −5/3 slope) increases monotonously
with the Reynolds number (as direct consequence of the two Kolmogorov hypotheses).

Spectral reconditioning Considering the characteristic scales of the turbulent motion,


the temporal evolution of a homogeneous isotropic turbulence brings into play two main
effects :
(i) the decrease of fluctuations (and also of the turbulent kinetic energy), and
(ii) the increase of the characteristic length scale.
Figure 3.9 shows how these effects are acting on the spectrum. The spectra correspond to a
grid turbulence experiment for various distances from the grid. It can be observed a decrease
of the area under the spectrum (equal to the total amount of turbulent kinetic energy)
with increasing distance from the grid. It results from a decrease of the spectral density of
energy at all scales. It seems however that the decrease observed for small wavenumbers
is proportionally less important than the one observed at large wavenumbers. Figure 3.10,
where relative variation of the spectral density between two successive distances is plotted,
confirms this point. With these data, one can observe that the variation is about twice
or three times higher for large wavenumbers than for small ones. This observation can be
related to the theoretically expected increase of the characteristic scales of turbulence : it
results from the faster disappearance of small structures. This is a statistical effect which
has nothing to do with an increase of the size of physical structures present in the flow.

55
108

106

104
 1/4
F11 (κ)/ ǫν 5

102

10−2

10−4
10−6 10−4 10−2 1
 3 1/4
κ × ν /ǫ

Figure 3.8 – Plots of Kolmogorov spectra obtained experimentally. Each symbol corresponds to a
different experiment. The flow configurations are well beyond the frame of HIT (boundary layers,
channel flow, jets, etc.) The dotted line corresponds to the slope −5/3.

10−2

10−3 t = x/U0

10−4

10−5
E(κ, t)

10−6

10−7
x/M = 30
x/M = 50
x/M = 70
10−8

10−9
1 10 100 1000 10000
κ

Figure 3.9 – Spectral density of energy decrease in HIT : the decrease is observed at all scales.
Data are coming from a grid turbulence experiment, Roches [9].

56
160

140
x/M : 30 → 50
120

100
δx/U0

x/M : 50 → 70
δE

80
− E1 ×

60

40

20

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500
κ

Figure 3.10 – Relative decrease in HIT. One can observe that small structures are decreasing faster
than big ones.

57
58
Part II

Near-wall turbulence

59
Chapter 4

The turbulent boundary layer

Industrial applications involving turbulent flows imply interactions between the fluid and
a solid wall where the transfer phenomena occurring take a great importance regarding the
prediction and the efficiency of the industrial processes :
— momentum transfers (aerodynamic drag, turbojet efficiency...)
— heat transfers (system cooling, kinetic heating...)
— mass transfers (boundary layer control, condensation/evaporation on a cold/warm
wall...).
If these questions are specific to the fields of aerodynamics and propulsion, they are also
generic to many kinds of industrial applications. The importance of these transfers between
the fluid and a solid wall requires the engineer to have good prediction tools from the
conception stage of the process. The question of a good predetermination has for a long
time relied on empirical expertise — the use of notebooks, correlations, abacus, etc. —,
but, due the increase of computing efficiency over the last decades, it is now related to the
statistic modelling of turbulence in numerical simulations.
In this respect, turbulence modelling applied to numerical simulations of near-wall tur-
bulent flows is based on the modelling of free turbulence supplemented by specific physical
aspects related to the presence of the wall. One – but not unique — of these specificity lies in
the importance of viscous effects although unusual when dealing with high Reynolds number
flows. For this reason, a turbulence Reynolds number has been introduced to measure the
relative weight of the turbulent diffusion compared to the molecular (viscous) diffusion. This
is the reason why the low Reynolds number turbulence modelling for near-wall flows has to
be distinguished from the high Reynolds number turbulence modelling for free flows at large
global Reynolds number.
In the following, we will first focus on the identification of the specific characteristics
of near-wall turbulent flows for the mean flow in this chapter before considering those of
the fluctuations in chapter 5. The description is limited here to incompressible flows since
the transposition to variable-density flows — related to large Mach numbers or mixing of
different species — has not been the subject of much specific developments but rather relies
on the use of transformed variables :
— in the equations via the introduction of density-weighted averages (Favre average 1 ) ;

1. A. Favre, J. de Mécanique, 4(3), 1965.

61
— in the analysis of the properties related to near-wall flows via the use of the Van
Driest transform 2 ,
because the general scientific consensus — or more precisely, the current scientific knowledge
on the subject — is that the incompressible case gathers most of the specificities of the near-
wall turbulence.

4.1 Governing equations


4.1.1 Derivation of the equations for the mean flow
We limit the study to incompressible Newtonian fluids. In this case, dynamic equations
are decoupled from the energy equation and they read
∂Uj
(4.1) = 0,
∂xj
 
∂Ui ∂Ui 1 ∂P ∂ 2 Ui
(4.2) + Uj = − +ν for i = 1, 2, 3.
∂t ∂xj ρ ∂xi ∂xj ∂xj
These equations state the mass and momentum balance, and together with adequate initial
and boundary conditions, they enable theoretically the access to the values of velocity and
pressure fields everywhere and at any time in the flow.
However, we know that these fields are random in space and time when the flow is
turbulent. The fields fluctuations span a broad range of scales which extent increases with
the bulk Reynolds number ReΛ , scaling as (see chapter 3)
Λf 3/4
= ReΛ .
ηK
where Λf and ηK are respectively the largest and smallest length scales of the flow. This
is why it is impossible in practice to get an exact solution of the Navier–Stokes equations
due to the limited computing capacity of computers. Most of the industrial applications are
characterised by very large Reynolds numbers and are beyond the reach of direct numerical
simulations. The alternative solution consists in solving the governing equations for the mean
values of the flow fields defined by (see chapter 1)
n=N
1 X (n)
F = lim F ,
N →∞ N
n=1

where F is the physical variable and n corresponds to a realisation of the flow. We limit here
the study to stationary and ergodic flows for which the ensemble average can be exactly
replaced by a temporal average, and then
Z
1 t0 +T
F = lim F (t)dt,
T →∞ T t
0

where t0 is arbitrary. From this definition, one can classically decompose every instanta-
neous field (F ) as the sum of a mean value (F ) and a fluctuation (f ) : F = F + f . This
2. E. R. Van Driest, J. Aero. Sci., 18(3), 1951.

62
y Ue (x)
✻ ✲






.........................................................


δ(x)



.............


x

Figure 4.1 – Schematic view with standard notations for a boundary layer over a flat plate.

decomposition is well-known and we briefly recall some of its useful properties :

F + G = F + G, λF = λF , F G = F G + f g, f = 0,

where F et G are random variable and λ a scalar.


Averaging the Navier–Stokes equations (4.2) and using the aforementioned properties,
one get straightforwardly the governing equations for the mean stationary flow, known as
Reynolds or RANS equations :

∂U j
(4.3) = 0,
∂xj

   
∂U i 1 ∂P ∂ ∂U i ∂U j
(4.4) Uj =− + ν + − ui uj for i = 1, 2, 3.
∂xj ρ ∂xi ∂xj ∂xj ∂xi

This system of equations is open since second order statistical momentum of the random
velocity field are appearing in (4.4) under a form corresponding to a diffusive flux. This
turbulent diffusion is analogous to the viscous diffusion and this is why these extra terms
are called Reynolds stresses. Closing the system requires the definition of constitutive laws
modelling the Reynolds stresses with the mean variables. This is the objective of turbulence
modelling and will be addressed in the following of the course.

4.1.2 Mean flow equations under the boundary layer assumption


The boundary layer model consist in a thin layer aligned in the streamwise direction in
which the velocity varies in the crosswise direction due to molecular diffusion. As sketched in
figure 4.1, a two-dimensional boundary layer over a flat plate develops along the wall which
normal to the y direction and the modulus of the external velocity Ue is a function of the x
coordinate only.
Assuming that the mean flow is two-dimensional and stationary, equations (4.3) and (4.4)
write as

∂U ∂V
(4.5) + = 0,
∂x ∂y

63
 
∂U ∂U 1 ∂P ∂2U ∂2U ∂u2 ∂uv
(4.6) U +V =− +ν + − − ,
∂x ∂y ρ ∂x ∂x2 ∂y 2 ∂x ∂y

 
∂V ∂V 1 ∂P ∂2V ∂ 2V ∂uv ∂v 2
(4.7) U +V =− +ν 2
+ − − .
∂x ∂y ρ ∂y ∂x ∂y 2 ∂x ∂y

Denoting δ(x) and V(x) the thickness of the boundary layer and a characteristic scale of the
crosswise velocity respectively at abscissa x, one can evaluate the magnitude of each term
of the continuity equation (4.5) :

∂U /∂x ∼ Ue /x and ∂V /∂y ∼ V/δ.

These two terms being of the same order — their sum must be zero —, it can be deduced
that V ∼ Ue δ/x and that convective terms in momentum equations (4.6) and (4.7) are of
the same order :

∂U Ue U2 ∂U Ue δ Ue U2
U ∼ Ue × = e, V ∼ × = e,
∂x x x ∂y x δ x
∂V Ue δ/x U 2δ ∂V Ue δ Ue δ/x U 2δ
U ∼ Ue × = e2 , V ∼ × = e2 .
∂x x x ∂y x δ x

Regarding viscous diffusion terms, it appears that crosswise diffusion will be dominant over
streamwise diffusion due to the thin layer assumption (δ/x ≪ 1) :

∂2U Ue ∂ 2U Ue
ν 2
∼ν 2, ν 2
∼ν 2,
∂x x ∂y δ
∂2V Ue δ ∂ 2V Ue
ν ∼ν 3 , ν ∼ν .
∂x2 x ∂y 2 δx

Similarly, the thin layer hypothesis implies that the crosswise turbulent diffusion is dominant
over the streamwise one :
∂u2 K2 ∂uv K2
∼ , ∼ ,
∂x x ∂y δ
∂uv K2 ∂v 2 K2
∼ , ∼ ,
∂x x ∂y δ

where K is a characteristic scale of the rms value for the fluctuation ui ui ). Using these
scalings, equations can be simplified accordingly.
We will now demonstrate that the static pressure stays constant throughout the boundary
layer in the crosswise direction and takes the value of the external flow Pe (x) like in the
laminar case. For this purpose, we consider the simplified momentum equations for which

64
all terms — excepting the pressure terms — can be evaluated :

∂U ∂U 1 ∂P ∂2U ∂uv
U +V = − + ν 2 −
∂x ∂y ρ ∂x ∂y ∂y
| {z } | {z } |{z}
↓ ↓ ↓
U2 Ue K2
∼ e ∼ν 2 ∼
x δ δ

∂V ∂V 1 ∂P ∂ 2V ∂v 2
U +V = − + ν 2 −
∂x ∂y ρ ∂y ∂y ∂y
| {z } | {z } | {z } |{z}
↓ ↓ ↓ ↓
U 2δ 1 (∆y P ) Ue K2
∼ e2 ∼ ∼ν ∼
x ρ δ δx δ
where (∆y P ) is a characteristic scale of the crosswise pressure variations which order of ma-
gnitude can be evaluated from the other terms in the above crosswise momentum equation :
  2    2 
δ 1 K
(∆y P ) = ρUe2 O +O +O 2 .
x Rex Ue
Using this expression in the streamwise momentum equation, the order of magnitude of the
pressure term can in turn be assessed as
    2 
1 ∂P Ue2 1  x 2 K x
− = O[1] + O +O 2 .
ρ ∂x x Rex δ Ue δ
If the streamwise pressure gradient is decomposed as the sum of the external streamwise
pressure gradient and a correction due to crosswise variation
 
1 ∂P 1 dPe (∆y P )
− = − +O
ρ ∂x ρ dx ρx
 2    2 !
1 ∂Pe Ue2 δ 1 K
= − + O +O +O 2 ,
ρ ∂x x x Rex Ue

it can be seen that the crosswise correction is at least one magnitude order (in δ/x) lower
than the external streamwise gradient. One can legitimately assume that the pressure in
the streamwise momentum equation is everywhere equal to its external value. Hence, under
the thin layer assumption only, the Navier–Stokes equations can be reduced to the Prandtl
model for the two-dimensional incompressible stationary turbulent boundary layer over a
flat plate which writes

∂U ∂V
(4.8) + = 0,
∂x ∂y

∂U ∂U 1 dPe ∂2U ∂uv


(4.9) U +V =− +ν 2 − ,
∂x ∂y ρ dx ∂y ∂y

65
where Pe is obtained from the integration of the Prandtl momentum transport equation
along y direction :
1 ∂P ∂v 2
0 = − −
ρ ∂y ∂y
Pe = P + ρv 2 .

4.2 The mean velocity profile


4.2.1 Absence of global self-similarity
In the laminar regime, the two-dimensional boundary layer flow over a flat plate without
any external streamwise pressure gradient is known to be described by a self-similar solution :
the Blasius velocity profile. The existence of such a solution states that the local (for a given
abscissa) crosswise evolution of the velocity profile is described by a universal function
(with no x dependence) when normalised by local characteristic scales. Taking respectively
the external velocity Ue (x) and the boundary layer thickness δ(x) for the velocity and length
characteristic scales, the Blasius solution enables to write
U/Ue = f (y/δ) and V /Ue = g(y/δ),
where f and g are two well-defined functions.
If it is very tempting to look also for a self-similar solution for the turbulent boundary
layer, the observation of experimental data shows that it is impossible. To illustrate this
point, it is sufficient to compare two sets of characteristic scales which can be used for
normalisation : a first one based on the usual classic scales (Ue , δ) already used in the
laminar case, or a second set built from the fluid dynamic viscosity (µ) and the wall friction
(τw ). The two time scales deduced from these two sets characterise the mean flow shear
(∂U /∂y), and they write as 3 :
δ µ
Te = and Ti = .
Ue τw
Their ratio takes the following form
Te Ue δ τw Cf
= × 2
= Reδ ,
Ti ν ρUe 2
where the wall friction coefficient Cf is simply defined by τw = 1/2 ρ Cf Ue2 . In the laminar
regime, the Blasius solution leads to
p 0.664
Reδ = 4.92 Rex and Cf = √ ,
Re x
and gives a constant value for the ratio Te /Ti . Conversely, in the turbulent regime, an
empirical law supported by experiments yields the following expression
Cf −1/4
= 0.0224 Reδ
2
3. The subscripts i and e refer respectively to an inner and an outer zone which will be defined and
described hereinafter.

66
y y/δ
δ ✻ ✻

1

Te

1 ❘
Reδ
✿ ✿
Ti Ti /Te
✲ ✲
U U/Ue
Ue 1

Figure 4.2 – Representation of the characteristic time scales Ti and Te based on the velocity profiles.
In the left figure where velocity profiles are normalised, the slope at the wall is equal to Ti /Te and
— if the boundary layer is turbulent — it decreases when the Reynolds number increases.

which implies that the ratio Te /Ti increases with the Reynolds number – and the streamwise
distance – as
Te 3/4
= 0.0224 (Reδ ) .
Ti
Thus, the laminar velocity profile can be characterised with a single time scale while the
turbulent regime exhibits at least two time scales evolving independently, preventing to find
a self-similar solution. Figure 4.2 gives a graphical representation of the characteristic time
scales Ti and Te based on the plot of the velocity profiles for increasing Reynolds numbers.

4.2.2 Structure of the turbulent boundary layer at large Reynolds


number
Despite the impossibility to find a self-similar solution for the velocity profile of the
turbulent boundary layer, it is nevertheless possible to define two regions where a partial
self-similarity exists. The connection between these two regions is also possible when the
Reynolds number is large enough, and the derivation of such an overlap condition leads to
the full characterisation of the mean flow within the turbulent boundary layer. This the
object of this section.

The surface layer


We focus first on the structure of the turbulent boundary layer with zero pressure gradient
in the region near the wall. In this region, called the surface layer, the advective terms can
be neglected in equation (4.9) due to wall adherence :
 
∂ ∂U
(4.10) ν − uv = 0.
∂y ∂y

It shows that the total shear stress τxy = ρ ν ∂U /∂y − uv is constant in this zone, which
crosswise extent still needs to be determined. This quantity is the sum of a purely viscous

67
term (ν ∂U /∂y) and the Reynolds stress (−uv). The latter is a kinematic quantity and goes
to zero at the wall while the former reaches there its maximal value (τw /ρ). It appears that,
in this zone of constant total shear stress, the fluid viscosity plays a key role especially in
the very vicinity of the wall. Hence, the flow is likely to be well characterised in this region
by the fluid dynamic viscosity (µ) and the wall friction (τw ), which were used to define Ti
in the previous subsection 4.2.1.
It is now possible to obtain the evolution of the mean velocity profile within the surface
layer by integrating the equation (4.10) :

∂U
(4.11) ν − uv = u2τ ,
∂y

where we have introduced for convenience the friction velocity uτ defined by τw = ρu2τ .
Equation (4.11) is in turn integrated to get the following relation
Z yuτ /ν
U uτ y uv  yuτ 
= + d .
uτ ν 0 u2τ ν

The variables normalised by wall characteristic scales (ν/uτ and uτ ) are usually denoted
with a sign + as a superscript, and the above equation is rewritten under the following form
Z y+
+ +
(4.12) U =y + uv + dy + .
0

Furthermore, in the vicinity of the wall where viscous shear stress is preponderant over the
turbulent shear stress, equation (4.12) degenerate in U + = y + , which define the asymptotic
expression for the mean velocity profile when y + is very small. This partial self-similarity is
valid within a sub-part of the constant shear stress zone called the viscous sub-layer.
According to equation (4.12), the generalisation of this partial self-similarity to the whole
surface layer implies that the Reynolds stress uv + is a function of y + only instead of the
couple of variables (y + , y/δ) as indicated in the conclusions of section 4.2.1. Such a depen-
dence on y + only is verified experimentally but cannot be demonstrated theoretically. It can
nevertheless be inferred from considerations on the local structure of the turbulence based
on the mechanism associated to the turbulent diffusion.
Indeed, the physical mechanism associated with turbulent diffusion can be understood
from the sketch in figure 4.3 : at a given crosswise location within the boundary layer, po-
sitive streamwise fluctuations inducing a velocity increase (u > 0) correspond mainly to the
downward motion of faster fluid particles (v < 0) coming from the upper part of the boun-
dary layer where streamwise velocity is higher. Conversely, negative streamwise fluctuations
inducing a velocity decrease (u < 0) are mainly related to upward motion of slower fluid
particles v > 0 coming from the lower part of the boundary layer where streamwise velocity
is weaker. This simple analysis is well supported by experimental data which shows that
u and v velocity fluctuations are correlated 4 and that the resulting Reynolds stress uv is
negative.
This phenomenological interpretation holds in the surface layer. Nevertheless, the confi-
nement due to the proximity of the wall tends to restrain the « population » of the efficient
p p
4. The correlation coefficient Cuv defined by uv = Cuv u2 v2 exhibits an almost constant value about
−0.5 within the boundary layer.

68
y

U (y)

induced fluctuation v<0


u<0
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .✛ . ❯. . . . . . . . . . . . . . . . . .
.......✲

u>0
v>0 induced fluctuation

Figure 4.3 – Illustration of turbulent diffusion mechanism associated with correlation uv.

vortical structures active in the turbulent diffusion mechanism. If, at a given y, structures
which size is comprised between the Kolmogorov scale 5 and the thickness of the boundary
layer cannot be excluded, one can however wonder if they all contribute equally to the corre-
lation uv. If those which size is smaller than y are unlikely to affected by the wall, those which
size lies roughly between y and δ are however very likely to be affected : they necessarily
impact the wall when they arrive at the distance y. This splat effect or kinematic blocking
induces the cancellation of the wall-normal fluctuation (v) and suppresses consequently the
creation mechanism of uv for these structures.
It allows to assume that only structures having a size smaller than y contribute to the
correlation uv. At given y, the value of uv does not depend on the boundary layer thickness
δ if much larger than y, which is the case in the surface layer when the Reynolds number
is large and it thus can be admitted that the turbulent shear stress uv does not depend on
y/δ which ensures the partial self-similarity based on the wall characteristic scales within
the whole surface layer.

There exists a region in the vicinity of the wall where the mean flow can be
described with local variables only. The influence of the global structure of the
boundary layer on this zone is indirect through the wall friction (via uτ ). For
this reason, this region — where the total shear stress is constant in the case
of a boundary layer with zero pressure gradient — is called the surface layer.
This region is subdivided into a viscous sub-layer at the immediate vicinity of
the wall and a fully turbulent zone beyond called the buffer layer.

The outer layer


Does the self-similarity based on the global characteristics scales (Ue , δ) and derived for
the laminar boundary layer still hold in the turbulent regime ? If it does, it will be a partial
self-similarity since part of or the entire surface layer must be excluded. We thus assume
5. It corresponds to the characteristic scale of the smallest structures in the turbulent flow.

69
the existence of an outer layer where the flow is characterised by global scales which have
to be determined. In order to characterise the mean velocity profile in this zone, the choice
of the length scale is straightforward : the thickness of the boundary layer δ is adapted.
However, the determination of the adequate velocity scale is more tricky since experiments
show that the natural choice of the external velocity scale Ue does not provide acceptable
results. The failure of Ue in being the adequate scale for normalisation is nonetheless easy
to understand if one keeps in mind the fact that the time scales Ti and Te — inner and
outer — evolve independently and become all the more separated when the Reynolds is
large. A direct consequence of this separation lies in the fact that the flow in the outer layer
does not « see » any more the reference U = 0 at the wall and that Ue loses its meaning as
the velocity differential across the boundary layer. Thus, the only solution is to resort to the
friction velocity uτ which is luckily both a local scale — by definition — and a global scale
as it can be seen from the von Kármán equation 6 :
Cf dθ 1 dUe
= + (2θ + δ ∗ ) .
2 dx Ue dx
This equation shows that the friction coefficient Cf — and so is uτ — is obtained through
global quantities : the momentum thickness (θ), the displacement thickness (δ ∗ ) and the
external velocity Ue .
The loss of reference for the velocity leads naturally to look for a self-similar solution
for the velocity
 defect rather than the velocity itself, i.e. considering the variable ∆U =
U − Ue . The velocity law in the outer layer will be of the following form :
U − Ue y 
=F .
uτ δ

The overlap region


The turbulent boundary layer is a thin layer structured into two regions with separate
streamwise evolution :
— the surface layer — close to the wall with η = y/δ ≪ 1 — where the mean flow
evolution is independent of the boundary layer thickness δ ;
— the outer layer — close to the external flow with y + ≫ 1 — where the mean flow
evolution is independent of the fluid viscosity.
From this description, there are two possibilities either the two regions are well separated
or exhibit an overlap. In the former case, the two regions are separated by an intermediate
zone where the mean flow evolution depends on both ν and δ. The latter case is the most
interesting since the overlap region is characterised by the two following conditions :
(4.13) y + ≫ 1 and η ≪ 1,
which imply that the mean flow evolution is independent on ν and δ. From a practical point
of view, this case corresponds to high Reynolds number turbulent flows. Indeed, using the
aforementioned expressions of Re δ and Cf , the evolution of the ratio y + /η is :
r
y+ δUe uτ Cf
= × = Reδ × ≈ 0.15 (Re δ )7/8 .
η ν Ue 2
6. In the laminar regime, the von Kármán equation is obtained by integration of the Prandtl equations
across the boundary layer, and stays valid in the turbulent regime for the mean flow.

70
The condition (4.13) corresponds to large y + /η, and it is valid as soon as the Reynolds
number is large enough. Practically, a value of 1.5 104 for Reδ is sufficient to observe an
overlap between the surface and the outer layers.
If the condition (4.13) is satisfied at a given crosswise coordinate y, then the mean flow
velocity U (y) must conform to both partial self-similarity laws and it can be written

 U − Ue
(4.14) (a) : U + = f y + , and (b) : = F (η),

which leads straightforwardly to

U + − Ue+ = f (y + ) − Ue+ = F (η) .

The continuity of the velocity gradient is obtained by differentiating both sides of the equa-
tions along y :
uτ ′ + 1
f (y ) = F ′ (η),
ν δ
before multiplying it by y to give

y + f ′ (y + ) = η F ′ (η).

The condition (4.13) of well-separated scales implies that both sides of the above equations
are independent of y + and η, and equal to a constant A. Finally, an integration leads to the
two following asymptotic laws 7 :
∆U
— in the outer layer, for « small » η : = A Log η + B ;

— in the surface layer, for « large » y + : U + = A Log y + + C.

The analysis shows that A, B and C are constant independent of x as soon as the Reynolds
number is sufficiently high to allow the existence of an overlap region between the surface
and the outer layers. Experiments confirm that the values of A and C are relatively stable :
we usually retain that 1/A = κ = 0.41 (κ is called the von Kármán constant) and C lies in
the range 5–5.2.

The existence of the logarithmic law is a main characteristic of the turbulent


boundary layer at high Reynolds number, and for this reason the overlap region
is simply called the logarithmic zone.

Summary on the complete structure of the turbulent boundary layer


The mean velocity profile is now fully characterised with its asymptotic behaviour. Spal-
ding [13] summarises it with the following implicit expression :
 
+ 1 2 1 3 −κC
(4.15) y + = U + + eκU − 1 − κU + − κU + − κU + e ,
2 6
7. These laws were first derived by Millikan [12] and it is often referred to as Millikan’s argument.

71
which degenerates to the linear law in the viscous sub-layer when U + → 0 and to the
logarithmic law in the fully turbulent zone when U + ≫ 1.
Regarding the outer layer, an analytical solution exists only for η → 0 and one must
resort to experiments to derive a more generic expression. This is what is done by Coles [14]
who proposed to write the mean velocity profile in the outer layer under the following form :
1 Π y
(4.16) U+= Log y + + C + w .
κ κ δ
The last term represents the deviation from the logarithmic law and is called the wake
component. The constant Π is the wake parameter. The wake function (w) is going to 0
when y/δ and is equal to 2 for y/δ = 1. The law for the velocity defect follows immediately :

U − Ue 1h y i
(4.17) = Log + Π (w (y/δ) − 2) .
uτ κ δ
Based on numerous experimental data, Coles suggests the following form for the wake func-
tion 8 :
y  π y
(4.18) w = 2 sin2 .
δ 2 δ
With this expression, the value of Π is about 0.55 for a turbulent boundary layer with zero
pressure gradient. It enables in return to obtain the value of the constant B appearing the
logarithmic law for the velocity defect : B = −2Π/κ ≈ −2.7.
It is now possible to give a general expression of the mean velocity profile for the whole
boundary layer :
 
+ + + 1  + 2 1  + 3 −κC
y + = U S + eκU S − 1 − κU S − κU S − κU S e ,
2 6
+ Π y
(4.19) U + = US + w .
κ δ
The mean velocity profile given by (4.19) is plotted in logarithmic scales in figure 4.4 for
Reθ ≈ 10 000 where the different regions described previously are identified, in particular
the logarithmic region where the surface layer and the outer layer are superimposed.
This velocity profile naturally satisfies both the inner and outer partial self-similarities,
which can be verified in figures 4.5 and 4.6. Figure 4.5 evidences the self-similarity in the
surface layer where the velocity profiles when normalised by the local viscous characteris-
tic scales are superimposed near the wall whatever the Reynolds number. Conversely, in
figure 4.6, the velocity profiles normalised by the global characteristic scales are superimpo-
sed near the outer of the boundary layer.
Finally, figure 4.7 shows the mean velocity profile normalised by the global characteris-
tic scales when plotted in linear scales. This representation gives a better overview of the
boundary layer structure together with the relative size of each region. Even if the extent
of the logarithmic zone is small compared to the wake region, it has a major importance
in the global structure of the turbulent boundary layer since the mean velocity rises from
8. This expression is commonly used but it brings a non-zero velocity gradient at the outer limit of
the boundary layer
 (y = δ). Granville [15] proposed to correct this problem by replacing w with w +
(y/δ)2 − (y/δ)3 .

72
30
Viscous sub-layer Buffer layer Log. layer Wake
zone

25 κ

Inner layer
20
+
+ + U = ln y + /0.41 + 5.2
U 15 U = y+

Outer layer
10

0
0.1 1 10 100 1000 10000
y+

Figure 4.4 – Representation in logarithmic scales of the mean velocity profile in the turbulent
boundary layer for Reθ ≈ 10 000.

30

Re
25

20
+

15
U

10

Reθ = 3250
5 Reθ = 7500
Reθ = 15000

0
0.1 1 10 100 1000 10000
+
y

Figure 4.5 – Evidence of the partial self-similarity in the surface layer for a turbulent boundary
layer with zero pressure gradient and increasing Reynolds numbers.

73
0

-5

-10
(U − Ue )/uτ

-15

-20 Re

Reθ = 3250
-25 Reθ = 7500
Reθ = 15000

-30
0.0001 0.001 0.01 0.1 1 10
y/δ

Figure 4.6 – Evidence of the partial self-similarity in the outer layer for a turbulent boundary layer
with zero pressure gradient and increasing Reynolds numbers..

zero to about 70 % of the external velocity Ue within less than 10 % of the boundary layer
thickness. It can also be noticed that the velocity gradient in the surface layer is much larger
than the one observed in the outer layer. This is in agreement with the high value of the
ratio Te /Ti and the loss of reference for the velocity in the outer layer.

4.3 Generalisation
The existence for the near-wall turbulent flow of a structuring into an inner and an
outer region connected by a logarithmic zone has been derived in the previous section for
a turbulent boundary layer at high Reynolds number and with zero pressure gradient. This
description actually applies to more generic wall-bounded turbulent flows than what the
specific hypotheses made would suggest. Indeed, the « useful » assumptions can be limited
only to the two following ones :
(a) The global Reynolds number characteristic of the boundary layer must be sufficiently
high so that the « viscous » length scale ν/uτ is very different from the boundary
layer thickness δ, which ensures the separation of scales.
(b) The scale separation implies the existence of region where the distance to the wall
y is large compared to the viscous length scale ν/uτ and also small compared to the
boundary layer thickness δ. In this region, the mean velocity gradient (∂U /∂y) must
be independent of both the fluid viscosity ν and the boundary layer thickness δ.
As long as these two conditions are verified, the logarithmic law is likely to apply in the
overlap region. For flows at high Reynolds number and under the thin layer assumption,
we will consider more specifically the two following cases : local modifications of the wall
conditions and the presence of an external streamwise pressure gradient.

74
1
velocity profile
0.9
0.8
0.7
Logarithmic profile
0.6
y/δ 0.5
0.4
Wake
0.3 component

0.2
0.1
0
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
U /Ue
Figure 4.7 – Structure of the mean velocity profile in the turbulent boundary layer normalised by
the global characteristic scales. The representation is in linear scales.

4.3.1 Local modifications of the wall conditions


It corresponds to the cases of the blowing-suction of the boundary layer and the presence
of a wall roughness when the associated characteristic scales are — like the viscous scales —
decoupled from the outer scales. It means that the velocity intensity of the blowing-suction
must be « weak » and the size of the wall roughness must be « small » i.e.
— if k represents the mean height of the roughness : k+ ≪ δ ;
— if vp is the blowing-suction velocity : vp ≪ uτ .
In these cases, the logarithmic law is still observed but the constants vary from their usual
values. The variation is small for the value of the von Kármán constant κ, while it is sub-
stantial for the value of the constant C, which results from an integration starting from the
wall and is thus more sensitive to the wall conditions.

4.3.2 The presence of an external streamwise pressure gradient


We consider here channel flows or accelerated/decelerated boundary layers for « mode-
rate » values of the streamwise pressure gradient. We focus first on the channel flow : it
consists of a stationary flow between two infinite parallel plane walls driven by a constant
pressure gradient in the direction parallel to the wall. Using notations illustrated in figure 4.8,
it is straightforward to show that the mean flow is one-dimensional : the velocity is aligned
along the x axis and depends only on the crosswise coordinate y. The momentum equation
along the x axis is thus reduced to
 
∂ ∂U 1 ∂P
ν − uv = .
∂y ∂y ρ ∂x

75
y


h



✲ U (y) ❄





x

z

Figure 4.8 – Representation of the channel flow.

This equation can be integrated in the crosswise direction from the lower wall up to a given
coordinate y to yield
   
∂U τw 1 ∂P
(4.20) ν − uv − = y,
∂y ρ ρ ∂x

which, for y = h, reduces to :


τw 1 ∂P
−2 = h.
ρ ρ ∂x
This relation shows that, when normalised by the outer characteristic scales uτ and h, the
pressure gradient is constant and equal to −2 for the channel flow whatever the value of the
Reynolds number. This constant, denoted βh , is defined by

h ∂ P /ρ
βh = 2 .
uτ ∂x

The fact that βh is constant whatever the value of the Reynolds number ensures, despite
the streamwise pressure gradient, the existence of the outer partial self-similarity 9 for the
channel flow within the same framework presented previously in subsection 4.2.2.
The influence of the pressure gradient within the surface layer, equation (4.20) can be
normalised by the viscous characteristic scales to give
+
∂U βh +
= 1 + uv + + y
∂y + Reτ
y 
(4.21) = 1 + uv + + βh ,
h
where Reτ = uτ h/ν. The second expression in (4.21) shows that the pressure gradient
introduces an unwanted dependence on y/h in the expression of the mean velocity gradient.
However, the first expression enables to put the influence of this term into perspective,
especially when the Reynolds number is large. Indeed, the order of magnitude of this last
9. The outer partial self-similarity is here characterised by a relation of the form (U −Umax )/uτ = F (y/h).

76
term in y/h can be evaluated for a moderate Reynolds number (Re τ = 1000) 10 in the
logarithmic region (y + = 100). In this case, βh /Reτ y + = −0.2 has to be compared with
the mean velocity gradient in absence of pressure gradient ∂U + /∂y + = 1/κ = 2.5, which
makes a small correction. It can be concluded that the y/h dependence of the mean velocity
gradient within the surface layer of a channel flow is weak, even nil in the limit of infinite
Reynolds number.
The conditions of existence of the inner and outer regions are then the same for the
channel flow than for the boundary layer with zero pressure gradient, which justifies the lo-
garithmic overlap region at high Reynolds number. The available experimental data confirm
this conclusion and, based on the review of various experimental data, Dean [1] proposed
the following values for the constant of the logarithmic law (4.16) :

κ = 0.41, C = 5.17 and Π = 0.14.

The two first values are very close to the ones measured for the turbulent boundary layer
with zero pressure gradient while the value of Π is noticeably smaller than the value of 0.55
given in subsection 4.2.2. It underlines the substantial difference between the outer regions
of each flow : the presence or not of a pressure gradient and of a free boundary.
A similar analysis can be derived for the boundary layers with an external streamwise
pressure gradient. First, the existence of the outer partial self-similarity has to be verified,
and — like in the channel flow — a normalised expression for the pressure gradient is
introduced under the form
δ ∗ ∂P /ρ
β= 2 .
uτ ∂x
It is known as the Clauser parameter and, when independent of x — i.e. constant along the
streamwise direction —, it defines the concept of a balanced boundary layer. In the near wall
region where advection is negligible compared to the diffusion mechanisms, an expression
equivalent to the relation (4.21) can be derived :
+
∂U
= 1 + uv + + p+ y +
∂y +
y
(4.22) = 1 + uv + + β ∗ ,
δ
with
ν ∂P /ρ β
p+ = = ∗+ .
u3τ ∂x δ
The normalised displacement thickness δ ∗+ can be viewed as a Reynolds number (based on
the displacement thickness and the friction velocity) which allows to conclude, like for the
channel flow, that the conditions of existence of the logarithmic overlap zone are verified for
small pressure gradients and large Reynolds numbers (small p+ ). Practically, for decelerated
boundary layers (p+ > 0), Perry et al. [16] conclude on the existence of the logarithmic
zone for values of p+ less than 0.05 11 . Practically, the large values of the Reynolds numbers
met in most of the applications are sufficient to satisfy this criterion. For example, the two
10. Dean [1] presents experimental data up to Reynolds numbers of Re τ ∼ 105 .
11. This value appears rather optimistic if one takes over the same derivation of the channel flow, but the
analysis of Perry et al. is more meticulous than the one presented here and allows to refine the value of the
proposed threshold.

77
8

Π 4
Wieghardt & Tillmann
3 Bradshaw, a = −0.15
Clauser, flow 1
2 Bradshaw & Ferriss, a = −0.255
Clauser, flow 2
Skare & Krogstadt
1 Π = 1.21(β + 0.258)0.564

0
0 5 10 15 20 25
β

Figure 4.9 – Correlation Π-β for decelerated boundary layers. Experimental data are coming from
the compilation of Coles & Hirst [20] for the 1968 Stanford conference and, for the largest pressure
gradients, from Skåre & Krogstad [18]. The empirical law is adapted from the one attributed to
Mellor & Gibson by Ligrani [19].

experiments of Samuel & Joubert [17] and of Skåre & Krogstad [18] for a turbulent boundary
layer near separation shows that the mean velocity profiles exhibit convincing logarithmic
zones.
Here again, the presence of the pressure gradient does not affect significantly the values
of the two constants of the logarithmic zone but influences mainly the outer velocity profile
through the wake parameter Π. Ligrani [19] attributes the following correlation between the
value of Π and the value of the Clauser parameter β to Mellor & Gibson :

(4.23) Π = 0.8 (β + 0.5)0.75 .

This illustrated in figure 4.9 for various experimental data of decelerated boundary layers
where the correlation (4.23) has been changed to Π = 1.21 (β + 0.258)0.564 in order to provide
a better agreement for large pressure gradients.

4.4 The friction coefficient and the shear stress


4.4.1 The friction law induced by the mean velocity profile
The equation (4.16) describes a family of mean velocity profiles function of a single para-
meter. If it is natural to choose the abscissa x as the meaningful parameter, it is practically
very difficult to define without ambiguity the origin of the x axis and it is better to use a local
quantity φ(x). The more common choice is the Reynolds number based on the momentum
thickness, and it is used to characterise the friction law Cf (Reθ ).

78
We now show that the friction law can be obtained from the mean velocity profile (4.16)
only. Indeed, it is possible to write for y = δ
1 2Π
Ue+ = Log δ + + C + ,
κ κ
or
Ue 1 δuτ 2Π
= Log +C + .
uτ κ ν κ
Introducing the friction coefficient and the Reynolds number based on the boundary layer
thickness in the above equation yields :
 −1/2  
Cf 1 1 Cf 2Π
(4.24) = Log Reδ + Log +C + .
2 κ 2κ 2 κ
This equation 4.24 introduces a one-to-one relation between Cf and Re δ . The boundary
layer thickness is very difficult to evaluate experimentally, this is why it is replaced by the
momentum thickness, defined by
Z δ  
U U
θ= 1− dy.
0 Ue Ue
When normalised by viscous scales, the equation becomes
Z δ+  
U+
Reθ = U + 1 − + dy + ,
0 Ue
which can be rewritten in :
Z Re δ √Cf /2  q 
+ +
(4.25) Re θ = U 1−U Cf /2 dy + .
0

According to equation (4.24), the friction coefficient Cf is a function of Reδ , so the mean ve-
locity profile U + = f (y + ) is sufficient to give the relation between Reθ and Re δ . This relation
associated with equation (4.25) leads to the friction law Cf (Reθ ). It is not possible to derive
its analytical expression due to the fact that the equations (4.24) and (4.25) are implicit,
but it can be solved numerically from discrete values of Reδ . It is illustrated in figure 4.10
and the numerical curve is in extremely good agreement with the experimental data of
Wieghardt & Tillmann [21]. In order to supply for the absence of an analytical expression,
several empirical correlations have been proposed, amongst which the more commonly used
is the formula of Kármán-Shoenherr :
 −1
Cf = 17.08 (log Re θ )2 + 25.11 log Re θ + 6.012 .

4.4.2 The shear stress profiles


The knowledge of the mean velocity profile enables to derive the evolution of the shear
stress within the boundary layer. We consider the total shear stress τ /ρ = (ν ∂U /∂y − uv).
Equation (4.9) can be recast under the following form
∂U 2 ∂U V 1 ∂τ
+ = .
∂x ∂y ρ ∂y

79
0.006

0.005

0.004
Cf

0.003

0.002

0.001
0 5000 10000 15000 20000
Reθ

Figure 4.10 – Friction law for the boundary layer over a flat plate with zero pressure gradient, — :
numerical resolution of (4.24) and (4.25), ◦ : experimental results of Wieghardt & Tillmann [21].

Integration in the crosswise direction from the wall to y, it gives


Z y
τ − τw ∂U
=UV +2 U dy,
ρ 0 ∂x

in which V can be eliminated using the mean continuity equation :


Z y Z y
τ − τw ∂U ∂U
(4.26) = −U dy + 2 U dy.
ρ 0 ∂x 0 ∂x
To go further, one has to assume a local self-similarity, i.e. to consider that the mean
velocity profile follows a self-similar law of the form U /Ue = f (y/δ) between abscissa x and
an abscissa infinitely close 12 . It can be deduced that

∂U  y  dδ
= Ue − 2 f ′,
∂x δ dx
and rewrite (4.26) under the following form
 Z y Z y 
τ − τw Ue2 dδ ′ ′
= 2 f y f dy − 2 y f f dy ,
ρ δ dx 0 0

which yields in a normalised form :


 Z η Z η 
τ − τw dδ
(4.27) 2
= f η f ′ dη − 2 η f f ′ dη ,
ρUe dx 0 0

12. It is equivalent to consider an asymptotic expansion of U at the first order in x.

80
0.003

0.0025

0.002
2τ /ρUe2

0.0015

0.001

0.0005

0
0 0.2 0.4 0.6 0.8 1 1.2
y/δ

Figure 4.11 – Total shear stress evolution across a turbulent boundary layer with zero pressure gra-
dient at Reθ = 7500, — : numerical integration of equation (4.29) based on the velocity profile (4.19)
with the wake function of Granville [15] (cf. note 7), ◦ : experimental results of Klebanoff [22].

or Z Z !
f f
τ − τw dδ
= f η df − 2 η f df .
ρUe2 dx 0 0

The last unknown quantity is the boundary layer thickness δ(x), which expression can be
inferred from the local self-similarity hypothesis leading to a constant value for the ratio
θ/δ. The von Kármán equation gives a relation between the momentum thickness and the
friction coefficient : dθ/dx = Cf /2, and from the definition of θ, one gets :
Z δ   Z 1
U U
(4.28) θ= 1− dy = δ f (1 − f ) dη.
0 Ue Ue 0

Introducing the relation (4.28) in equation (4.27) yields the final relation :
" Z 1 −1  Z f Z f #
τ Cf
(4.29) = 1+ f (1 − f ) dη f η df − 2 η f df .
ρUe2 2 0 0 0

This relation does not have a simple analytical solution and a numerical integration is
necessary to obtain the total shear stress profile. Figure 4.11 presents such a numerical
solution for a boundary layer at Reθ = 7500 when the mean velocity profile is given by (4.19).
It is in good agreement with the experimental data obtained by Klebanoff [22] for the same
Reynolds number.
From the total shear stress profile (4.29) and the mean velocity profile (4.19), it is possible
to derive the turbulent shear stress −uv = τ /ρ − ν ∂U /∂y and compare the contribution
of each term to the total shear stress. This is illustrated in figure 4.12 : the viscous shear

81
1
0.9
0.8
0.7 Turbulent shear stress
0.6
τ /ρu2τ

0.5
0.4
0.3 Viscous shear stress
0.2
0.1
0
1 10 100 1000 10000
y+

Figure 4.12 – Crosswise evolution of the « viscous » and « turbulent » shear stresses for a turbulent
boundary layer at Reθ = 7500. The solid line corresponds to the total shear stress.

stress is predominant near the wall, both terms are equivalent in the buffer layer and the
turbulent shear stress is dominant beyond. It has also to be noted that the constant total
shear stress zone extends up to y + ∼ 100.
The determination of the total shear stress profile from the mean velocity profile and the
von Kármán equation exposed above can be integrally transposed to the case of a boundary
layer subjected to a streamwise pressure gradient, see appendix C for the details.

82
Chapter 5

Specificity of near-wall
turbulence

5.1 Analytical tools for analysis


5.1.1 Reynolds stress equations
We briefly recall the derivation of the Reynolds stress transport equations already pre-
sented in section 1.2.2. For stress ui uj , the transport equation for each velocity fluctuation
ui and uj is obtained by the difference between the corresponding momentum transport
equation for the instantaneous and the mean velocity component. Each equation is then
multiplied by the other fluctuation (ui for uj equation and uj for ui equation) before being
summed and averaged to finally lead to :
∂ ui uj ∂ ui uj p ν
+ Uk = Pij + Dij + Dij + Dij + φij − ǫij ,
∂t ∂xk
where
 
∂U j ∂U i
Pij = − ui uk + uj uk is the production term ;
∂xk ∂xk
∂ ui uj uk
Dij = − is the diffusion by velocity fluctuations ;
∂xk
 
p 1 ∂ pui ∂ puj
Dij =− + is the diffusion by pressure fluctuation ;
ρ ∂xj ∂xi

ν ∂ 2 ui uj
Dij =ν is the viscous (molecular) diffusion ;
∂xk ∂xk
 
1 ∂ui ∂uj
φij = p + is the pressure-deformation correlation term ;
ρ ∂xj ∂xi
  
∂ui ∂uj
ǫij = 2ν is the pseudo-dissipation.
∂xk ∂xk

83
The index contraction and the summation of the normal stress transport equations yields
the transport equation for the turbulent kinetic energy per unit of mass (k = ui ui /2) :
∂k ∂k
+ Uj =
∂t ∂xj
    
1 ∂U i ∂U j 1 ∂ puj ∂ ui ui uj ∂2k ∂ui ∂ui
− ui uj + − − +ν −ν ,
2 ∂xj ∂xi ρ ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj
which includes all the physical mechanisms present in the Reynolds stress equations with
the exception of the pressure-deformation correlation term. This term corresponds to energy
exchange between the components of the Reynolds stress tensor [ui uj ] and disappears for the
equation of k in the compressible regime. The last term, noted ǫ, is the (pseudo) dissipation
of k.
As done for the mean flow equations, it is possible to simplify the above equations
for a two-dimensional stationary mean flow under the thin layer assumption. Most of the
simplifications are straightforward but it has to be noted that the two-dimensionality of
the mean flow imposes uw and vw stresses to be zero but not the normal stress w2 . The
transport equations for the non-zero components of the Reynolds stress tensor become
 
Du2 ∂U ∂  2 ∂ 2 u2 2 ∂u ∂u ∂u
(5.1) = −2uv − vu + ν + p − 2ν ,
Dt ∂y ∂y ∂y 2 ρ ∂x ∂xk ∂xk
   
Dv 2 ∂ pv ∂ 2 v2 2 ∂v ∂v ∂v
(5.2) = − v3 + 2 +ν + p − 2ν ,
Dt ∂y ρ ∂y 2 ρ ∂y ∂xk ∂xk
 
Dw2 ∂  2 ∂ 2 w2 2 ∂w ∂w ∂w
(5.3) = − vw + ν + p − 2ν ,
Dt ∂y ∂y 2 ρ ∂z ∂xk ∂xk
     
Duv ∂U ∂ pu ∂ 2 uv p ∂u ∂v ∂u ∂v
(5.4) = −v 2 − v2 u + +ν + + − 2ν ,
Dt ∂y ∂y ρ ∂y 2 ρ ∂y ∂x ∂xk ∂xk

where the material derivative takes the simple following form : D/Dt = U ∂/∂x + V ∂/∂y.
The summation of equations (5.1) to (5.3) yields the transport equation for the turbulent
kinetic energy :
 
Dk ∂U ∂ ui ui pv ∂2k
(5.5) = −uv − v + + ν 2 − ǫ.
Dt ∂y ∂y 2 ρ ∂y
The observation of these equations brings an important comment : amongst the normal
Reynolds stresses, only u2 benefits from the production 1 from the mean flow. Regarding the
two other normal stresses (v 2 and w2 ), their increase along a path line (a positive variation
of D/Dt) results either from a local energy transfer with the other components (pressure-
deformation correlation) or a spatial transfer (viscous or turbulent diffusion). In both cases,
it is a transfer intrinsic to the fluctuating motion. Finally, the pseudo dissipation terms in
equations (5.1) to (5.3) are inherently positive and are sink terms corresponding to a local
destruction of energy. This is also the case for the dissipation term in the equation of uv,
although it is not possible to demonstrate it analytically.
1. It is production in most of the cases. Considerations on the sign of uv exposed in section 4.2.2 enable
to predict that the production term is generally positive. Nevertheless, there exist few configurations for
which this term becomes locally negative, such as separations, asymmetric channels or wall jets for instance.

84
5.1.2 Pseudo-dissipation transport equation
For turbulence modelling purpose, the pseudo dissipation ǫ is generally used instead of
the true rate 2 of dissipation of the turbulent kinetic energy :
 2
ν ∂ui ∂uj
ε= + ,
2 ∂xj ∂xi
to built a characteristic length of the fluctuating motion. The reasons for this substitution are
twofold. First, the pseudo-dissipation is appearing explicitly in the transport equation for k
when the viscous diffusion term is classically written with a Laplacian. Second, it is possible
to demonstrate that both quantities ǫ and ε are strictly equivalent for a homogeneous field.
As seen in chapter 4, the turbulent shear stress uv is highly inhomogeneous in the case of
near-wall turbulence — it is also the case for the normal Reynolds stresses as it will be seen
further in this chapter — , and it is legitimate to question the equivalence between ǫ and
ε. Measuring experimentally these two quantities has been for a long time a very difficult
challenge and this is why the transposition from ε to ǫ was usually made without any solid
justification. Since then, the numerical analysis of Bradshaw & Perot [23] has filled the gap :
their DNS results show that there is a maximal difference of a few percent (∼ 2%) between
the pseudo-dissipation and the true dissipation in a conventional boundary layer. It justifies
a posteriori the commonly accepted use of the pseudo-dissipation for modelling near-wall
turbulent flows.
The transport equation for ǫ can be derived straightforwardly from the ui velocity fluc-
tuation equation when derived with respect to xj , multiplied by ∂ui /∂xj and then averaged :

∂ǫ ∂ǫ
(5.6) + Uk = Pǫ1 + Pǫ2 + Pǫ3 + Pǫ4 + Dǫt + Dǫν + Πǫ − Υ,
∂t ∂xk
where
∂ui ∂uk ∂U i ∂ui ∂ui ∂U k
Pǫ1 = −ν , Pǫ2 = −ν ,
∂xj ∂xj ∂xk ∂xj ∂xk ∂xj
∂ui ∂ 2 U i ∂ui ∂ui ∂uk
Pǫ3 = −2ν uk and Pǫ4 = −2ν
∂xj ∂xj ∂xk ∂xj ∂xk ∂xj
are interpreted as production terms, and
 
t ∂ ∂ui ∂ui
Dǫ = −ν uk is the turbulent diffusion ;
∂xk ∂xj ∂xj

∂2ǫ
Dǫν = ν is the viscous (molecular) diffusion ;
∂xj ∂xj
 
ν ∂ ∂ui ∂p
Πǫ = −2 is a correlation term with the pressure fluctuation ;
ρ ∂xi ∂xj ∂xj

∂ 2 ui ∂ 2 ui
Υ = −2ν 2 is a destruction term.
∂xj ∂xk ∂xj ∂xk
2. True refers here to the definition of dissipation in the context of thermodynamics.

85
It is difficult to relate the different terms on the right hand side of the pseudo dissipation
transport equation to a physical mechanism, except for the classical diffusion and dissipation
terms. Tennekes & Lumley [24] conclude from an analysis of vorticity fluctuation 3 equation
that the production term Pǫ4 from the turbulent motion as well as the destruction term Υ
are dominant at large turbulence Reynolds number. The magnitude order of the other terms
is generally comparable to the difference Pǫ4 − Υ and thus cannot be neglected.

5.1.3 Near-wall asymptotic laws


It is possible to derive several useful results with a near-wall asymptotic analysis of the
fluctuating fields. The generic development of the fluctuations is of the form :
 
u = a0 + a1 y + a2 y 2 + O y 3 ,
 
v = b0 + b1 y + b2 y 2 + O y 3 ,
 
w = c0 + c1 y + c2 y 2 + O y 3 ,
 
p = p0 + p1 y + p2 y 2 + O y 3 .

The wall adherence condition applies to the instantaneous velocity field as well as to the
fluctuations. It implies that the coefficients a0 , b0 and c0 are nil. The incompressibility condi-
tion imposes a divergence free velocity fluctuation field. Together with the wall adherence
condition, it leads to (∂v/∂y)y=0 = 0 which results in b1 = 0. The near-wall asymptotic
behaviour for the velocity fluctuations are then

u ∼ a1 y, v ∼ b2 y 2 , w ∼ c1 y.
y→0 y→0 y→0

It shows that the wall-normal velocity fluctuation is going to zero faster than the tangential
components. For the Reynolds stresses, one gets :

a21 + c21 2
u2 ∼ a21 y 2 , v 2 ∼ b22 y 4 , w2 ∼ c21 y 2 , uv ∼ a1 b2 y 3 , k ∼ y ,
y→0 y→0 y→0 y→0 y→0 2

where a deficit of energy is observable for the wall-normal component. It is also possible to
derive the asymptotic behaviour for the dissipation tensor. For the first component (ǫ11 ), it
gives :
 2  2  2
ǫ11 ∂u ∂u ∂u
= + +
ν ∂x ∂y ∂z
        
= a21,x y 2 + O y 3 + a21 + O[y] + a21,z y 2 + O y 3 ,

where the subscript notation (, xi ) denotes the partial derivative with respect to xi . The
other components [ǫij ] of the dissipation tensor can be obtained similarly :

ǫ11 ∼ 2ν a21 , ǫ22 ∼ 8ν b22 y 2 , ǫ33 ∼ 2ν c21 , ǫ12 ∼ 4ν a1 b2 y.


y→0 y→0 y→0 y→0

3. It is related to ǫ in a homogeneous turbulent flow.

86
And the dissipation rate of the turbulent kinetic energy is :
1  
ǫ = (ǫ11 + ǫ22 + ǫ33 ) ∼ ν a21 + c21 .
2 y→0

It shows that the dissipation rate tends towards a finite value at the wall, and it generally
corresponds to its maximal value within the boundary layer. These relations shows also that
the turbulent shear stress and the corresponding dissipation rate have the same sign near
the wall. In the vicinity of the wall, it confirms at least the fact that ǫ12 is a sink term in
the equation for uv.

5.2 Specificities of near-wall turbulence


5.2.1 Anisotropy
All the (experimental and numerical) observations of the near-wall turbulence structure
show a strong anisotropy of the Reynolds stress tensor. Figure 5.1 presents the crosswise
evolution of the normal Reynolds stresses in the case of the channel flow from the direct
numerical simulation of Kim [25] 4 . The streamwise velocity fluctuation u is dominant in the
entire boundary layer and exhibits a peak near the wall. The same quantities normalised
by the local amount of turbulent kinetic energy k are plotted in figure 5.2. They are called
structural parameters of the turbulent agitation and their evolution within the boundary
layer is remarkable : both the normal stresses and the shear stress uv exhibit a large region
where they are approximatively constant. These constants take the following values
u2 v2 w2 −uv
≈ 1, ≈ 0.4, ≈ 0.6 and ≈ 0.3
k k k k
which are usually considered to be valid in the entire sheared zone except in the vicinity of
the wall for the sub-viscous layer and the buffer layer. The value of the structural parameter
for the shear stress seems to be more universal than the three others. It is usually denoted
a1 = −uv/(2k) and takes its value in the range 0.13-0.15 for a large panel of thin shear
flows. Physically, a1 is related to the specific ability of the fluctuating motion to produce
mixing.
The origin of the anisotropy for the normal Reynolds stresses can be understood in the
light of the results presented in the previous section :
(a) First, the presence of the wall imposes a global structuring of the thin shear layer.
The streamwise normal Reynolds stress is the only one extracting energy from the
mean flows.
(b) The asymptotic analysis shows that the wall adherence condition coupled with the
continuity equation implies an energy deficit for wall-normal velocity fluctuation.
The second point has to be related to the kinematic blocking : a fluid particle associated
with a downward wall-normal velocity fluctuation v will transfer its energy to the two others
velocity fluctuation when impacting the wall. This transfer is associated in the equations to
the pressure-deformation correlation term. This observation is very specific to the near-wall
turbulence and it will be important in the building of models since the pressure-deformation
correlation terms are usually associated with a return to isotropy.
4. See also Kim et al. [26].

87
8
+
7 u2
+
v2
+
6 w2

0
0 50 100 150 200 250 300 350 400
y+

Figure 5.1 – Crosswise evolution of the normal Reynolds stresses in the channel flow at Reτ =
uτ H/ν = 400, data taken from the DNS of Kim [25].

2
u2 /k
v 2 /k
1.5 w2 /k
uv/k

0.5

-0.5
0 50 100 150 200 250 300 350 400
y+

Figure 5.2 – Crosswise evolution of the structural parameters of the turbulent motion in the channel
flow at Reτ = uτ H/ν = 400, data taken from the DNS of Kim [25].

88
105

y + = 3270
4
10

y + = 121
103

y + = 5,33
102
−1/4
F11 × ǫν 5

101
y + = 1,21
−5/3
∝ κ1
100

10−1

10−2

10−3
10−4 10−3 10−2 10−1 100 101
3/4 1/4
κ1 × ν /ǫ

Figure 5.3 – One-dimensional spetra for u as function of wavenumber κ1 obtained by Tieleman [27]
at various values of y + in a turbulent boundary layer with zero pressure gradient.

5.2.2 Importance of the viscous mechanisms


Contrary to the free turbulence case at high Reynolds number, the viscous molecular dif-
fusion mechanism plays a key role in the near-wall turbulence. As seen in subsection 4.2.2,
an immediate consequence is that the viscous diffusion cannot be neglected in the mean mo-
mentum equation where it competes equally with the turbulent diffusion within the surface
layer and is even the dominant mechanism in the viscous sub-layer. This effect is also trans-
posable to the Reynolds stress equations, the turbulent kinetic energy transport equation
and, as seen further, to the transport equation of the dissipation rate ǫ.
A second effect relative to the turbulence structure is of high importance as it presents
direct consequences for modelling purposes. It concerns the validity of the Kolmogorov hy-
pothesis relative to the spectral balance at high Reynolds number of turbulence. As seen in
section 3.5, this assumption ensures that the range of scales corresponding to the turbulent
structures extracting energy from the mean flow and the range of scales associated with

89
energy dissipation are sufficiently separated so that there exists in between an inertial range
where the associated turbulent vortical structures ensure the transfer of energy down the
scales (cascade). The Kolmogorov hypothesis also implies that the rate of energy transfer
through the inertial range is constant and equal to the energy dissipation rate at small scales.
This equality allows to use ǫ to build the characteristic scales of the vortices in the inertial
range despite the fact that dissipation is mechanism characteristic of the small scales. The
existence of the inertial range can be assessed by inspecting the spectrum of the velocity
fluctuations. More precisely, a range with a law in κ−5/3 characteristic of the inertial range
must be observed in all one-dimensional spectra. In the specific case of near-wall turbulence,
the assumption of large Reynolds number of turbulence (Re t = k 2 /νǫ) is questionable since
the fluctuations (and k) must be zero at the wall. The absence of the spectral balance in the
near-wall region can be observed in figure 5.3 where the one-dimensional spectra F11 (κ1 )
obtained by Tieleman [27] for various crosswise distances of the wall are plotted in the case
of a turbulent boundary layer with zero pressure gradient. For the largest value of y + , the
−5/3
inertial range in κ1 can be observed while it progressively shrinks with the decrease of
y + before disappearing totally for y + = 5.33 and 1.21.

5.2.3 Energy balances


The evolution of the balances for the fluctuating quantities across the boundary layer
present some specific characteristics that are worth to know, some of them being used for
tuning some constants in turbulence models. The interest for these evolutions has grown early
in the scientific community, and, since the middle of the 50s, Laufer [28] and Klebanoff [22]
presented kinetic energy balances for both the pipe flow and the boundary layer. Despite the
difficulty to measure experimentally some terms of the balance equation (in particular the
dissipation rate and the correlations implying the pressure fluctuations), those experimental
data still constitute a reference due to the relatively high values of the Reynolds numbers
at which the experiments were done. Since then, the access to quantities out of reach of
the experiments has been made possible thanks direct numerical simulations of turbulence
and these experimental observations have been confirmed and supplemented with detailed
analyses in the very near-wall region. The balances’ evolutions that will be presented in this
subsection are based on the direct numerical simulations of Spalart [29] for the boundary
layer and of Kim [25] 5 for the channel flow. The relatively small value of the turbulence
Reynolds number of these two simulations will not alter the conclusions drawn here.

Turbulent kinetic energy and normal Reynolds stresses


Figures 5.4 and 5.5 present the turbulent kinetic energy balance for a boundary layer at
Reθ = 1410 reproduced from the data of the DNS performed by Spalart [29]. The evolution
of all the terms in the equation 5.5 are reported relatively to the viscous characteristic scales
in figure 5.4 which gives informations on the balances within the surface layer. Except in the
buffer zone where all quantities exhibits significant values, there exists two degenerations of
the complete balance equation :
— in the viscous sub-layer where the energy balance reduces to molecular diffusion =
dissipation ;

5. See also Mansour et al. [30].

90
0.3

0.2

0.1

ǫ
-0.1 Pk
Dkp
Dkt
-0.2
Dkν
Dk/Dt
-0.3
0 20 40 60 80 100
+
y

Figure 5.4 – Turbulent kinetic energy balance near the wall for a boundary layer at Reθ = 1410
(data taken from the DNS of Spalart [29]. All terms are normalised by u4τ /ν.

— in the overlap region (logarithmic zone) where the energy balance reduces to produc-
tion = dissipation.
Figure 5.5, where the same quantities are reported relatively to the outer characteristic
scales, confirms the latter simplified energy balance in the overlap region. The wake zone is
characterised by a large contribution of the turbulent diffusion which is balanced by several
terms among which the main ones are the dissipation, the advection and to a lesser extent the
diffusion by pressure fluctuation. Contrary to the two first degenerated balances mentioned
previously, it does not appear to be universal since the energy balance reduces to dissipation
= turbulent diffusion in the wake zone for the channel flow.
Regarding the transport equations of the Reynolds stresses, the reader will refer to the
results of Spalart [29] and Mansour et al. [30] to get a complete description. We just limit
here to the two following remarks :
— simplified balances of the type molecular diffusion = dissipation remain valid in the
viscous sub-layer ;
— the pressure-deformation correlations contribute significantly to the balances.
In section 5.2.1, the importance of pressure-deformation correlations in the context of near-
wall turbulence has been mentioned, relatively to energy transfers which operate between
normal Reynolds stresses. In order to precise this point, the evolution of these terms is
plotted in figure 5.6 in the case the channel flow at Re τ = 400 after the DNS results of
Kim [25]. For the Reynolds stress ui ui , a positive value of φii corresponds to a energy gain
at the expense of the two other normal stresses. Thus, figure 5.6 clearly illustrates the inter-
component energy transfers in the flow : far from the wall, normal transfers operates from
the most energetic component (u2 which is the only one gaining energy from the mean flow)
towards the components in deficit (v 2 et w2 ). As we get close to the wall, the energy transfer

91
10

0
ǫ
Pk
Dkp
-5 Dkt
Dkν
Dk/Dt

-10
0.2 0.4 0.6 0.8 1 1.2
y/δ

Figure 5.5 – Turbulent kinetic energy balance in the outer layer for a boundary layer at Reθ = 1410
(data taken from the DNS of Spalart [29]. All terms are normalised by u3τ /δ.

changes at the expanse of the wall-normal component (v 2 ). It becomes in deficit in the buffer
layer (y + ≈ 12) and the transferred energy is such that the u2 component even becomes
beneficiary in the viscous sub-layer. This is a clear illustration of the blocking effect at the
wall. Comparison with figure 5.1 shows that the splat effect induces unusual energy transfers
which are not compatible with « isotropisation » mechanisms.

The dissipation rate of the turbulent kinetic energy ǫ


In section 5.1.2, we saw that the transport equation for the pseudo-dissipation is difficult
to interpret in terms of physical mechanisms with the exception of the usual terms of molecu-
lar diffusion and destruction. In particular, there are four terms identified as « production »
terms, three of which (Pǫ1 , Pǫ2 and Pǫ3 ) denotes an interaction with the mean flow while the
last one (Pǫ4 ) is related to the interaction of the turbulent motion with itself through the
physical mechanism of vortex stretching. According to Tennekes & Lumley [24], this last
production term is dominant at high turbulence Reynolds number. Figure 5.7 shows the
evolution of all terms in the equation (5.6) in the case of the channel flow at Re τ = 400 after
the DNS results of Kim [25]. Far from the wall (i.e. beyond the overlap region) and in the
viscous sub-layer, two simple degenerated balances are observed :
— production (Pǫ4 ) = destruction far from the wall, which is consistent with the analysis
of Tennekes & Lumley ;
— molecular diffusion = destruction, which is classic in the viscous sub-layer.
In between these two regions, the balance is complex since most of the terms retain significant
values.
Due to the difficulty to give a physical interpretation to all terms of equation (5.6) and
the fact that the ǫ transport equation is ill-conditioned (two dominant terms which difference

92
0.08
0.06
0.06
0
0.04

0.02 -0.06
0 10 20 30
0

-0.02

-0.04
φ11
-0.06 φ22
φ33
-0.08
0 50 100 150 200 250 300
y+

Figure 5.6 – Evolution of the pressure-deformation correlation terms appearing in the transport
equations of the normal Reynolds stresses. Data correspond to a channel flow at Reτ = 400 and are
taken from the DNS of Kim [25]. The inserted box in the upper right corner shows a zoom in the
near-wall region.

0.03
Pǫ1
0.02 Pǫ2
Pǫ3
Pǫ4
0.01
−Υ
Dǫt
0 Πǫ
Dǫν
-0.01

-0.02

-0.03
0 10 20 30 40 50
y+

Figure 5.7 – Dissipation rate balance for a channel flow at Reτ = 400 (data taken from the DNS
of Kim [25]). All terms are normalised by u4τ /ν.

93
0.03
Pǫ1 + Pǫ2
0.02 Pǫ3
−(Υ − Pǫ4 )
Dǫt
0.01
Πǫ
Dǫν
0

-0.01

-0.02

-0.03
0 10 20 30 40 50
y+

Figure 5.8 – Reconditionned dissipation rate balance for a channel flow at Reτ = 400 (data taken
from the DNS of Kim [25]). All terms are normalised by u4τ /ν.

is of the same order as the other terms), the models of this equation do not proceed term
by term but rather introduce the following groups :
— the combination of the two production terms associated with the mean velocity gra-
dients, Pǫ = Pǫ1 + Pǫ2 ;
— the combination of the two terms associated with the action of the turbulent motion
on itself, Υǫ = Υ − Pǫ4 .
The balance of ǫ is plotted according to these mergings in figure 5.8. The term Υǫ clearly ap-
pears as a destruction term compensated by the production term Pǫ in the whole boundary
layer at the exception of the viscous sub-layer, where the degenerated balance molecular
diffusion-destruction remains valid with the new definition of the destruction term Υǫ . Fi-
nally, the different terms of turbulent diffusion and the production term remains small but
significant compared to the two dominant ones.

94
Bibliography

[1] R. B. Dean, « Reynolds number dependence of skin friction and other bulk flow variables in
two-dimensional rectangular duct flow », ASME Trans. J. Fluid Eng., vol. 100, p. 215–223,
1978.
[2] O. Reynolds, « On the dynamical theory of incompressible viscous fluids and the determina-
tion of the criterion », Phil. Trans. R. Soc. Lond. A, vol. 186, p. 123–164, 1895.
[3] S. A. Orszag et G. S. Patterson, « Numerical simulation of three-dimensional homogeneous
isotropic turbulence », Phys. Rev. Letters, vol. 28, p. 76–79, 1972.
[4] R. S. Rogallo, « Numerical experiments in homogeneous turbulence », TM 81315, NASA,
1981.
[5] H. P. Robertson, « The invariant theory of isotropic turbulence », Proc. Camb. Phil. Soc.,
vol. 36, p. 209–223, 1940.
[6] I. Proudman et W. H. Reid, « On the decay of a normally distributed and homogeneous
turbulent velocity field », Phil. Trans. Roy.Soc. London A, vol. 247, p. 163–189, 1954.
[7] G. K. Batchelor et A. A. Townsend, « Decay of isotropic turbulence in the initial period »,
Proc. Roy. Soc. Lond. A, vol. 193, p. 539–558, 1948.
[8] G. K. Batchelor et A. A. Townsend, « Decay of isotropic turbulence in the final period »,
Proc. Roy. Soc. Lond. A, vol. 194, p. 527–543, 1948.
[9] P. Roches, Étude expérimentale d’un écoulement turbulent non-cisaillé soumis à un gradient
thermique. Thèse doctorat, INPT-Département de Mécanique des Fluides de l’ENSICA, 1999.
[10] A. N. Kolmogorov, « The local structure of turbulence in incompressible viscous fluid for
very large Reynolds numbers », Dokl. Akad. Nauk SSSR, vol. 30, p. 301–305, 1941.
[11] J. C. R. Hunt, O. M. Phillips et D. Williams, éds, Turbulence and Stochastic Processes.
Kolmogorov’s Ideas 50 Years on. Cambridge University Press, 1992.
[12] C. B. Millikan, « A critical discussion of turbulent flows in channels and circular tubes », in
Proc. of the 5th Int. Cong. Appl. Mech., (New York), p. 386, Wiley, 1939.
[13] D. B. Spalding, « A single formula for the law of the wall », J. Appl. Mech., vol. 28, p. 455–457,
1961.
[14] D. Coles, « The law of the wake in the turbulent boundary layer », J. Fluid Mech., vol. 1,
p. 191, 1956.
[15] P. S. Granville, « A modified law of the wake for turbulent shear layers », ASME Trans. J.
Fluid Eng., vol. 98, p. 578–580, 1976.
[16] A. E. Perry, J. B. Bell et P. N. Joubert, « Velocity and temperature profiles in adverse
pressure gradient turbulent boundary layers », J. Fluid Mech., vol. 25, p. 299, 1966.
[17] A. E. Samuel et P. N. Joubert, « A boundary layer developing in an increasingly adverse
pressure gradient », J. Fluid Mech., vol. 66, p. 481, 1974.

95
[18] P. E. Skåre et P.-A. Krogstad, « A turbulent equilibrium boundary layer near separation »,
J. Fluid Mech., vol. 272, p. 319–348, 1994.
[19] P. M. Ligrani, « Stucture of turbulent boundary layer », in Encyclopedia of Fluid Mechanics
(N. P. Cheremissinoff, éd.), p. 111, Gulf P., 1989.
[20] D. E. Coles et E. A. Hirst, éds, Computation of turbulent boundary layers-1968 AFOSR-
IFP-Stanford Conference, vol. II – Compiled Data. Stanford University, 1968.
[21] K. Wieghardt et W. Tillmann, « Flat plate flow #1400 », in Computation of turbulent
boundary layers-1968 AFOSR-IFP-Stanford Conference (D. E. Coles et E. A. Hirst, éds),
vol. II – Compiled Data, p. 98–123, Stanford University, 1968.
[22] P. S. Klebanoff, « Characteristics of turbulence in a boundary layer with zero pressure gra-
dient », TN 3178, NACA, 1954.
[23] P. Bradshaw et J. B. Perot, « A note on turbulent energy dissipation in the viscous wall
region », Phys. Fluids A, vol. 5, p. 3305, 1993.
[24] H. Tennekes et J. L. Lumley, A first course in turbulence. Cambridge, Massachusetts : MIT
Press, 1972.
[25] J. Kim. Unpublished results, 1990.
[26] J. Kim, P. Moin et R. Moser, « Turbulence statistics in fully developed channel flow at low
Reynolds number », J. Fluid Mech., vol. 177, p. 133, 1987.
[27] H. W. Tieleman, « Viscous region of turbulent boundary layer », Technical Report CER67-
68HWT21, Colorado State University-Fort Collins, 1967.
[28] J. Laufer, « The structure of turbulence in fully developed pipe flow », TN 2954, NACA,
1953.
[29] P. R. Spalart, « Direct simulation of a turbulent boundary layer up to Rθ = 1410 », J. Fluid
Mech., vol. 187, p. 61–98, 1988.
[30] N. N. Mansour, J. Kim et P. Moin, « Reynolds-stress and dissipation-rate budgets in a
turbulent channel flow », J. Fluid Mech., vol. 194, p. 15, 1988.

96
Appendix

97
Appendix A

Dynamic equations for


two-point double correlations

We want to derive the equation for the dynamics of two-point double correlations P (x) et
P ′ (x′ ) :
Qi,j (x, x′ , t) = ui (x, t)uj (x′ , t) with x = [xk ] and x′ = [x′k ]
For sake of simplification, temporal dependence is omitted in the following and we write f (x, t) = f
and f (x′ , t) = f ′ for any function f .

A.1 Derivation
Demonstration starts from the momentum equation for the fluctuation ui taken at point P :

∂ui ∂ui ∂U i ∂ui ∂ui 1 ∂p ∂ 2 ui


+ Uk + uk + uk − uk =− +ν
∂t ∂xk ∂xk ∂xk ∂xk ρ ∂xi ∂xk ∂xk

Multiplying previous equation by the fluctuation u′j taken at point P ′ , one gets

∂ui ∂ui ∂U i ∂ui ∂ui u′j ∂p ∂ 2 ui


u′j + u′j U k + u′j uk + u′j uk − u′j uk =− + νu′j ,
∂t ∂xk ∂xk ∂xk ∂xk ρ ∂xi ∂xk ∂xk

Noticing that u′j is independent of xk and using the continuity equation for the triple correlation
term, it can be written

∂ui ∂ui u′j ∂U i ∂ui uk u′j ∂ui



1 ∂puj ∂ 2 ui u′j
u′j + Uk + uk u′j + − u′j uk =− +ν ,
∂t ∂xk ∂xk ∂xk ∂xk ρ ∂xi ∂xk ∂xk
And finally, averaging, it yields :

∂ui ∂ui u′j ∂U i ∂ui uk u′j 1 ∂puj



∂ 2 ui u′j
(A.1) u′j + Uk + uk u′j + =− +ν .
∂t ∂xk ∂xk ∂xk ρ ∂xi ∂xk ∂xk

With a simple transposition (i ⇋ j, uk ⇋ u′k and xk ⇋ x′k ), one gets the equation for the u′j taken
at point P ′ multiplied by the fluctuation ui taken at point P :

∂u′j ∂ui u′j ∂Uj′ ∂ui u′k u′j 1 ∂p′ ui ∂ 2 ui u′j


(A.2) ui + Uk′ + u′k ui ′ + =− +ν ′ ′ .
∂t ∂xk′
∂xk ∂xk ′
ρ ∂xj ′
∂xk ∂xk

99
Then, summation of equations A.1 and A.2 gives :

∂Qi,j ∂Qi,j ∂Qi,j ∂Uj′ ∂U i ∂Sik,j ∂Si,jk


(A.3) + Uk + Uk′ + Qi,k ′ + Qk,j + +
∂t ∂xk ∂xk

∂xk ∂xk ∂xk ∂x′k
   
1 ∂L,j ∂Li, ∂ 2 Qi,j ∂ 2 Qi,j
=− + +ν + .
ρ ∂xi ∂x′j ∂xk ∂xk ∂x′k ∂x′k

A.2 Variable substitution


In order to get the usual form of the equation for double correlations, one must proceed to the
following substitution of variables :

 
ξk = x k x k = ξk
or reciprocally .
rk = x′k − xk x′k = ξk + rk

By omitting indices (k) for sake of simplification, on can transform the derivatives of a given function
F by applying the following formulas :

 
∂F ∂F ∂F
= x′r − x′ξ /J,
∂x ∂ξ ∂r
 
∂F ∂F ∂F
= xξ − xr /J,
∂x′ ∂r ∂ξ
 
∂2F 2 ∂2F ′ ′ ∂ F
2 2
′2 ∂ F
= x′r − 2x ξ x r + x ξ /J 2 ,
∂x2 ∂ξ 2 ∂ξ∂r ∂r 2
 
∂2F 2 ∂2F ∂2F 2
2 ∂ F
= xr − 2x ξ x r + x ξ /J 2 ,
∂x′2 ∂ξ 2 ∂ξ∂r ∂r 2

where the Jacobian is J = xξ x′r − xr x′ξ and indices stand for the derivation of the variable along
the index. The variable substitution defined previously enables one to write

(A.4) x′r = 1, x′ξ = 1, xr = 0, xξ = 1 and J = 1,

and leads to the following :

∂F ∂F ∂F
= −
∂x ∂ξ ∂r
∂F ∂F
=
∂x′ ∂r
∂2F ∂2F ∂2F ∂2F
= −2 +
∂x2 ∂ξ 2 ∂ξ∂r ∂r 2
∂2F 2
∂ F
=
∂x′2 ∂r 2

100
A.3 Final form
Using the formulas obtained in the previous section, one gets

∂Qi,j ∂Qi,j  ∂Qi,j ∂U i ∂Uj′


+ Uk + Uk′ − U k = − Qk,j − Qi,k
∂t ∂ξk ∂rk ∂ξk ∂rk
∂Sik,j ∂
− + (Sik,j − Si,jk )
∂ξk ∂rk
 
1 ∂L,j ∂L,j ∂Li,
− − +
ρ ∂ξi ∂ri ∂rj
 
∂ 2 Qi,j ∂ 2 Qi,j ∂ 2 Qi,j
+ν +2 −2 .
∂ξk ∂ξk ∂rk ∂rk ∂ξk ∂rk

One can go back to the original variable xk (since xk = ξk ) and obtain the final formulation :

∂Qi,j ∂Qi,j  ∂Qi,j ∂U i ∂Uj′


+ Uk + Uk′ − U k = − Qk,j − Qi,k
∂t ∂xk ∂rk ∂xk ∂rk
∂Sik,j ∂
− + (Sik,j − Si,jk )
∂xk ∂rk
 
1 ∂L,j ∂L,j ∂Li,
− − +
ρ ∂xi ∂ri ∂rj
 
∂ 2 Qi,j ∂ 2 Qi,j ∂ 2 Qi,j
+ν +2 −2 .
∂xk ∂xk ∂rk ∂rk ∂xk ∂rk

101
102
Appendix B

Derivation of Kármán-Howarth
equation

B.1 Transformation of derivatives


The first step consists in replacing partial derivatives along components rk of the separation
vector by its norm r in the viscous term. This is made possible because of the isotropy hypothesis :
 
∂Qi,j ∂ ∂ 2 Qi,j ∂ rk ∂Qi,j
− (Sik,j + Sjk,i ) = 2ν = 2ν
∂t ∂rk ∂rk ∂rk ∂rk r ∂r
h   i
∂ 1 ∂Qi,j 3 ∂Qi,j
= 2ν rk +
∂rk r ∂r r ∂r
h   i
rk rk ∂ 1 ∂Qi,j 3 ∂Qi,j
= 2ν +
r ∂r r ∂r r ∂r
 2 
∂ Qi,j 1 ∂Qi,j 3 ∂Qi,j
= 2ν − +
∂r 2 r ∂r r ∂r
h    i
1 ∂ ∂Qi,j ∂Qi,j
= 2ν r2 + 2r ,
r2 ∂r ∂r ∂r
hence
 
∂Qi,j ∂ 2ν ∂ ∂Qi,j
− (Sik,j + Sjk,i ) = 2 r2 .
∂t ∂rk r ∂r ∂r

B.2 Transposition to the correlation functions


One writes the corresponding equation for the trace of the correlation tensor :
 
∂Qi,i ∂Sik,i 2ν ∂ ∂Qi,i
(B.1) −2 = 2 r2 .
∂t ∂rk r ∂r ∂r
On the other hand, the general definition of Qi,j from the correlation functions f and g enables
to write
Qi,i ri ri
= (f − g) 2 + gδii = f − g + 3g = f + 2g.
u′2 r
Since the function g can be related to the function f through g = f + r (∂f /∂r) /2, one gets
 
∂f
Qi,i = u′2 3f + r ,
∂r

103
or
 
u′2 ∂f u′2 ∂r 3 f
Qi,i = 3r 2 f + r 3 = 2 .
r 2 ∂r r ∂r
Regarding triple correlations, one can do similar derivation by using the general definition of Sik,j
with the correlation function k and yields
 
Sik,i 1 1 ∂r 2 k 1 ∂r 2 k k
= 3k − ri rk ri + (ri δki + rk δii ) − ri δik
u′3 2r 3 r ∂r 4r 2 ∂r 2r
 
1 ∂r 2 k rk 1 ∂r 2 k rk
= 3k − + 2 (rk + 3rk ) − k
r ∂r 2r 4r ∂r 2r
 
1 ∂r 2 k rk
= 2k + .
r ∂r 2r

These expressions can now be injected in the different terms of the trace equation, which leads
to
— for the temporal derivative
h  i
∂Qi,i ∂ ∂f
= u′2 3f + r
∂t ∂t ∂r
 ′2 
∂u′2 f ∂ ∂u f
= 3 +r
∂t ∂t ∂r
  
1 ∂u′2 f ∂ ∂u′2 f
= 3r 2 + r3 ,
r 2 ∂t ∂r ∂t

thus :
 
∂Qi,i 1 ∂ ∂u′2 f
= 2 r3 ;
∂t r ∂r ∂t

— for the nonlinear term


 2S    
∂ ik,i ∂ k 1 ∂r 2 k
= rk 2 + 2
∂rk u′3 ∂rk r r ∂r
   
k 1 ∂r 2 k rk rk ∂ k 1 ∂r 2 k
= 3 2 + 2 + 2 + 2
r r ∂r r ∂r r r ∂r
    
1 k 1 ∂r 2 k ∂ k 1 ∂r 2 k
= 3r 2 2 + 2 +r 3
2 + 2
r2 r r ∂r ∂r r r ∂r
    
1 ∂ k 1 ∂r 2 k 1 ∂ 1 ∂r 4 k
= r3 2 + 2 = ,
r 2 ∂r r r ∂r r 2 ∂r r ∂r

thus :
 
∂Sik,i u′3 ∂ 1 ∂r 4 k
−2 =− 2 ;
∂rk r ∂r r ∂r

— for the viscous term

    
2ν ∂ ∂Qi,i 2ν ∂ ∂ u′2 ∂r 3 f
r2 = r2 .
r 2 ∂r ∂r r 2 ∂r ∂r r 2 ∂r

104
One finally gets
      
∂ ′2
3 ∂u f ∂ u′3 ∂r 4 k ∂ 2 ∂ u′2 ∂r 3 f
r − = 2ν r ,
∂r ∂t ∂r r ∂r ∂r ∂r r 2 ∂r

This equation can be integrated between 0 and r to obtain


 
∂u′2 f u′3 ∂r 4 k u′2 ∂ 1 ∂r 3 f
− 4 = 2ν
∂t r ∂r r ∂r r 2 ∂r
u′2 ∂
 ∂f 
= 2ν r + 3f
r ∂r ∂r
 h i
u′2 ∂ 1 ∂f
= 2ν r4 + 3r 3 f
r ∂r r 3 ∂r
    
u′2 ∂ 4 ∂f u′2 ∂f −3
= 2ν 4 r + 3r 3 f + 2ν r4 + 3r 3 f
r ∂r ∂r r ∂r r4
   
u′2 ∂ 4 ∂f 3 u′2 ∂f
= 2ν 4 r + 3r f − 6ν 4 r3 + 3r 2 f
r ∂r ∂r r ∂r
 
u′2 ∂ 4 ∂f 3 u′2 ∂r 3 f
= 2ν 4 r + 3r f − 6ν 4 .
r ∂r ∂r r ∂r
Hence the final result :
 
∂u′2 f u′3 ∂r 4 k u′2 ∂ ∂f
− 4 = 2ν 4 r4 .
∂t r ∂r r ∂r ∂r

Remark The presence in the transport equation for correlations Qi,j , of derivatives along the
three components of the vector r (∂/∂rk ), does not allow to simplify this equation by using the
conventional orientation of the separation vector.

105
106
Appendix C

Total shear stress profile for a


decelerated boundary layer

It is possible to derive the total shear stress profile for a boundary layer subjected to a streamwise
pressure gradient following the same steps presented in section 4.4.2. The mean momentum equation
is integrated taking into account the pressure gradient and yields
Z y
τ − τw ∂U 1 ∂P
=U V +2 U dy + y.
ρ 0
∂x ρ ∂x

The pressure gradient is also appearing in the von Kármán equation :


Cf dθ 1 dUe
= + (2θ + δ ∗ ) .
2 dx Ue dx
All calculations done, one gets the final relation
 
τ δ 2β βδ
(C.1) = 1 + (f J1 − 2J2 ) 1 + β + + (f J3 − 2J4 + η) ,
τw θ H δ∗
where H = δ ∗ /θ is the shape factor and the four integrals Ji with i = 1, 2, 3, 4 are defined by
Z f Z f Z η Z η
J1 = η df, J2 = ηf df, J3 = f dη, J4 = f 2 dη.
0 0 0 0

Figure C.1 presents the total shear stress profile in the case of a strong positive pressure gradient
obtained numerically from equation (C.1) when the mean velocity profile is given by equation (4.19)
with the corrected wake function of Granville [15]. The numerical solution is compared to the
experimental data of Skåre & Krogstadt [18] and the agreement is good. The wake parameter Π
and the Reynolds number used for the computation of the function f from equation (4.19) have
been fixed to the experimental values at x = 4,4 m : β = 19,3, Reθ = 44 420 and Π = 6,75.

107
70 18
(a) (b)
16 Experiments
60 Theory
Experiments
Theory 14
50
12
40
τ /(ρu2τ )

10
+
U

30 8

6
20
4
10
2

0 0
1 10 100 1000 10000 0 0.2 0.4 0.6 0.8 1 1.2
y+ y/δ

Figure C.1 – Comparison between theory and experiments for (a) the mean velocity profile and (b)
the total shear stress profile. Experimental data are taken from Skåre & Krogstadt [18] at x = 4,4
m. The dotted line correspond to the logarithmic law.

108

You might also like