International Journal of Heat and Fluid Flow 27 (2006) 585–593

www.elsevier.com/locate/ijhff

A synthetic-eddy-method for generating inflow conditions

for large-eddy simulations
a,*
N. Jarrin , S. Benhamadouche b, D. Laurence a, R. Prosser a

a
School of Mechanical, Aerospace and Civil Engineering, The University of Manchester, Manchester, UK
b
Département Mécanique des Fluides et Transferts Thermiques, Electricité de France, Chatou, France

Available online 30 March 2006

Abstract

The generation of inflow data for spatially developing turbulent flows is one of the challenges that must be addressed prior to the
application of LES to industrial flows and complex geometries. A new method of generation of synthetic turbulence, suitable for complex
geometries and unstructured meshes, is presented herein. The method is based on the classical view of turbulence as a superposition of
coherent structures. It is able to reproduce prescribed first and second order one point statistics, characteristic length and time scales, and
the shape of coherent structures. The ability of the method to produce realistic inflow conditions in the test cases of a spatially decaying
homogeneous isotropic turbulence and of a fully developed turbulent channel flow is presented. The method is systematically compared
to other methods of generation of inflow conditions (precursor simulation, spectral methods and algebraic methods).
 2006 Elsevier Inc. All rights reserved.

Keywords: Large-eddy simulation; Synthetic turbulence; Inflow boundary conditions; Spatially decaying isotropic turbulence; Channel flow

1. Introduction bulent inlet boundary conditions. From an engineering

point of view, almost all flows of industrial interest are spa-
It is widely accepted that the specification of realistic tially evolving flows. The need for methods to generate
inlet boundary conditions play a major role in the accuracy realistic unsteady inflow conditions is then mandatory to
of a numerical simulation. For RANS approaches, only use LES in industrial cases with a certain degree of confi-
mean profiles for the velocity and the turbulence variables dence in the results. It has been shown that the results of
need to be prescribed, which makes the definition of inflow DNS or LES, particularly in the cases of a plane jet (Klein
data comparatively straight-forward. For large-eddy and et al., 2003), a spatially developing boundary layer (Lund
direct-numerical simulations, the generation of inflow data et al., 1998) or a backward facing step (Jarrin et al.,
is much more of an issue. The early applications of LES 2003) are very sensitive to upstream inflow conditions.
were simple temporally developing flows where the simula- The most accurate technique to impose the inlet bound-
tion generated its own inflow data using periodic boundary ary conditions of a LES consists of obtaining inflow data
conditions. With the progress of the computer powers, LES from a precursor simulation. The inflow represents exactly
has been increasingly used to simulate spatially developing the large scale eddies of the flow at the inlet and there is no
flows, which require the specification of instantaneous tur- or very little transient downstream the inlet. This technique
has two major drawbacks. Firstly, it is restricted to simple
cases where the flow at the inlet of the computational domain
*
Corresponding author. Tel.: +44 161 200 3730; fax: +44 161 200 3723. can be regarded as a fully developed turbulent flow (Kalten-
E-mail addresses: n.jarrin@postgrad.manchester.ac.uk (N. Jarrin),
bach et al., 1999) or a spatially developing turbulent bound-
sofiane.benhamadouche@edf.fr (S. Benhamadouche), dominique.lauren-
ce@manchester.ac.uk (D. Laurence), robert.prosser@manchester.ac.uk ary layer (Lund et al., 1998) and therefore lacks generality. In
(R. Prosser). the scope of performing embedded LES or hybrid RANS/

0142-727X/$ - see front matter  2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.ijheatfluidflow.2006.02.006
586 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593

LES, this approach is not suitable either. Secondly, it entails of this method is that it is particularly suitable for non-
a heavy extra computational load and requires important homogeneous turbulence and complex inlet meshes as the
storage capacities. Thus the research effort seems to head size of the filter can easily be varied across the flow. Unfor-
towards methods of generation of synthetic turbulence tunately it can become very expensive as the mesh is
which can be both faster and more flexible. These methods refined.
try to generate an inflow signal which is as close as possible To obtain more insight into the flow physics, Druault
to the real flow by matching a reduced set of statistics. The et al. (2004) used a proper-orthogonal-decomposition of
statistics typically available are the mean velocity, the turbu- a turbulent signal coming from an external source and used
lent kinetic energy, the dissipation rate and sometimes the this signal as an inflow condition for LES calculation. Even
full Reynolds stress tensor profiles. though this technique cannot be applied systematically for
A basic technique to generate turbulent inflow data is to general flows as it requires a previous realization of the
take a mean velocity profile with superimposed random flow, it is interesting to note that a better simulation of
fluctuations. The data generated do not exhibit any spatial the large scale coherent structures of the flow at the inlet
or temporal correlations. The energy generated is also uni- enables a better simulation of the downstream flow.
formly spread over all wave numbers and, due to a lack of The idea behind the method presented in this paper is to
large scale energy-containing structures, the pseudo turbu- directly focus on prescribing coherent structures rather
lence is quickly dissipated (Jarrin et al., 2003). than reverting to spectral methods. It is a generalisation
A standard method to give some spatial and temporal of the previous work of Jarrin et al. (2003) which used
correlations to the generated data is to create time series streamwise vortices to trigger turbulence downstream of
of velocity fluctuations by performing an inverse Fourier the inlet in an LES calculation. The method presented
transform for prescribed spectral densities (Lee et al., herein is easy to implement, fast to run and performs well
1992 or Kondo et al., 1997). The amplitude of the Fourier on any geometry and any kind of flow. The data generated
modes are computed from the model spectrum and their exhibit very good physical properties such as; first and sec-
phase is drawn randomly. The missing phase information ond order one point statistics, prescribed length scales, time
of the real turbulent eddies makes these methods less accu- scales and the shape of the autocorrelation function.
rate than the ones employing a precursor simulation and a
transition section downstream of the inlet is necessary for 2. Numerical procedure
the flow to become realistic again. The main objective of
methods of generation of synthetic turbulence is to reduce 2.1. Inflow generation method
this transition section to be able to use shorter domains and
hence reduce the cost of the simulation. Even though these The method is based on the classic view of turbulence as
methods were applied with success for the simulation of a superposition of coherent structures. Coherent structures
both isotropic homogeneous turbulence (Lee et al., 1992) will be generated over the inlet plane of our calculation and
and flow over a backward facing-step (Le et al., 1997), they will be defined by a shape function that encompasses the
have several drawbacks which make them unsuitable for structure’s spatial and temporal characteristics.
industrial purposes. Indeed they are derived to generate We start with the one-dimensional case, where a one
periodic signals on uniform meshes. It is not clear if a Fou- component velocity signal is to be generated on the interval
rier decomposition of the inflow signal can be used in the [a, b]. fr(x) is the shape function of the turbulent spot,
case of non-homogeneous turbulence, or at least it does which has a compact support on [r, r] and satisfies the
not have any justification. On complex unstructured inlet normalization condition
meshes where the fast Fourier transform cannot be used, Z
they become expensive and hence not appropriate. Adapta- 1 D=2 2
f ðxÞdx ¼ 1 ð1Þ
tions of these spectral methods able to tackle industrial D D=2 r
issues have been proposed by Smirnov et al., 2001) and
where D = b  a + 2r. Each turbulent spot i has a position
then Batten et al. (2004). These authors use different spec-
xi (defining its physical location) and a length scale ri
tra rescaled by some turbulent variables at different loca-
(defining its spectral content). For the sake of simplicity,
tions across the flow and limit the number of modes
we retain a constant ri for the moment. The issue of non-
simulated to reduce the computational cost of generation
constant length scale r will be tackled later. Each spot is as-
of the inflow data. Keating et al. (2004) noted that on the
signed a sign ei. Thus the contribution u(i)(x) of turbulent
plane channel flow at Re* = 400 the method of Batten
spot i to the velocity field is
et al. (2004) produces a transition region downstream of
the inlet of about 20 channel half height before the flow uðiÞ ðxÞ ¼ ei fr ðx  xi Þ ð2Þ
becomes fully turbulent.
Another family of techniques for arbitrary inlet meshes where ei is a random step of value +1 or 1 and xi is drawn
is to filter random data on the inlet mesh (Klein et al., randomly on the interval [a  r, b + r]. The synthetic ed-
2003). Gaussian filters have been used to generate inflow dies are generated on an interval larger than [a, b] so that
data with spatial and temporal correlations. An advantage the boundary points can be surrounded by eddies. The
 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593 587

0 pffiffiffiffiffiffiffi 1
velocity signal at a point x is the sum of the contribution of R11 0 0
B R =a pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi C
all synthetic eddies on the domain. For N eddies it reads @ 21 11 R22  a221 0 A ð7Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 X N
R31 =a11 ðR32  a21 a31 Þ=a22 R33  a31  a322 2
uðxÞ ¼ pffiffiffiffi ei fr ðx  xi Þ ð3Þ
N i¼1 The length scale in the flow can also be varied. This is
The number of eddies on the domain can be set to (b  a)/r clearly one advantage of our method compared to spectral
which ensures that the plane remains statistically covered methods. The compact support of the spots enables us to
with turbulent spots. have different length scales in different parts of the flow
It can readily be shown then that our signal is of zero which might be of great interest to simulate wall flows.
mean, unit variance and that the two-point autocorrelation However, it must be mentioned that a strong variation of
function reads the length scale leads to a deviation of the correlation func-
Z tion from the analytical formulae Eq. (4) which assumes
1 D=2 constant length scales. Further distortions can result from
Ruu ðrÞ ¼ fr ðxÞfr ðx þ rÞdx ð4Þ
D D=2 the rescaling procedure Eq. (6), but are normally small
(Smirnov et al., 2001). The structures of the flow can also
Thus inflow data with any autocorrelation function can be
be controlled, for instance to generate streamwise vortices
generated. However inverting Eq. (4) to obtain fr from Ruu
in the near wall layer of a channel while more isotropic ed-
involves performing a complex deconvolution procedure.
dies are imposed at the center. In the following, the method
Details of such a procedure can be found in di Mare and
will thus be referred to as the synthetic-eddy-method
Jones (2005). Herein the simpler approach of choosing fr
(SEM). Only simple turbulent spots with a Gaussian decay
a priori is chosen.
of the intensity of the spots are going to be considered here-
The generalisation of the 1D procedure to the genera-
after. In this case, the shape function of the synthetic-eddy
tion of 2D time dependent fluctuations is straight-forward.
reads
The eddies are now three-dimensional structures with com-
pact three-dimensional supports on [rx, rx; ry, ry; rz, fj ðxÞ ¼ f ðkxkÞ ð8Þ
rz], satisfying a three-dimensional normalization condition where f is a Gaussian function.
of the same type as Eq. (1). The inlet plane is located at
x = 0 and it has dimensions [0, Lz] · [0, Ly]. The position 2.2. Flow solver
(xi, yi, zi) of synthetic-eddy i is drawn randomly on [rx,
rx] · [ry, Ly + ry] · [rz, Lz + rz]. The eddies are con- Code_Saturne, a collocated finite volume code for com-
vected through the inlet plane with a reference velocity plex geometries (Archambeau et al., 2004) and (Benhama-
scale U0 using Taylor’s frozen turbulence hypothesis douche and Laurence, 2003) is used to solve the
xi(t + dt) = xi(t) + U0 dt. Once xi(t) > rx the eddy is regen- incompressible Navier–Stokes equations. Velocity and
erated upstream at x = rx and convected again through pressure coupling is ensured by a prediction/correction
the inlet plane. The signal at a point x and a time t in method with a SIMPLEC algorithm. The collocated dis-
the inlet flow plane reads cretization requires a Rhie and Chow (1982) interpolation
1 X N in the correction step to avoid oscillatory solutions. A sec-
u0j ðx; tÞ ¼ pffiffiffiffi eij fj ðx  xi ðtÞÞ ð5Þ ond order centered scheme (in space and time) is used.
N i¼1
where eij is the sign of vortex i on component j and are inde- 3. Spatially decaying isotropic turbulence
pendent random steps of values +1 or 1. The number of
turbulent spots N on the inlet plane can be approximated The mean flow is in the positive x direction. The mesh
by Sp/Ss where Sp is the surface of the inlet plane and Ss dimensions are 6p · 2p · 2p, respectively, in the x, y and
the surface of the support of a turbulent spot, which en- z directions. The mesh is homogeneous in all three direc-
sures the inlet plane remains statistically covered with syn- tions and has 192 · 64 · 64 cells. Periodic boundary condi-
thetic eddies. tions are used in the y and z direction. The Smagorinsky
The independence of the rotation sign ensures that our constant is set to its theoretical value CS = 0.18.
inflow signal satisfies the condition ui uj ¼ dij . If the Rey- The independent parameters defining the calculation are
nolds stress tensor Rij and the mean velocity profile ui are the mean streamwise velocity U0 = 20 m/s, the viscosity
known a priori from previous experiments, DNS or RANS m = 3.5 · 104 m2/s, the turbulent energy k = 3/2 m2/s2
calculations, our signal can be transformed to match these and the integral length scale L defined as the distance at
statistics (Lund et al., 1998). The final velocity field ui is which the two-point correlation has the value 0.01. The
then reconstructed from the vortex field u0i according to first three parameters are kept constant which leaves only
ui ¼ ui þ aij u0j ð6Þ one free parameter, the length scale L that defines the Rey-
nolds number. Simulations carried out using different
where aij is obtained from the prescribed Reynolds stress length scales and different methods are listed in Table 1.
tensor and reads The time step is dt = dx/U0. Statistics are averaged over
588 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593

Table 1 2.0
Computations for the spatially decaying isotropic turbulence case
(a) RAND
Case Method Details
SPE0
RAND Basic random X 1.5
SEM0
DIG0 Digital filtering Gaussian filter, L = 0.4 DIG0
2
SPE0 Spectral method k 2 eðk=k0 Þ spectrum, L = 0.4
SEM0 SEM Gaussian spots, L = 0.4 1.0

k
SEM1 SEM Gaussian spots, L = 0.2
SEM2 SEM Gaussian spots, L = 0.8
0.5

two flow-through domain time and over the y and z

directions. 0.0
0 3 6 9 12 15 18
The results obtained using a basic random procedure x
RAND, the synthetic-eddy-method SEM0, the digital
filtering technique of Klein et al. (2003) DIG0 and the 0.0
spectral method of Lee et al. SPE0 are first compared. (b)
SEM0 uses 3D isotropic turbulent spots with a Gaussian
–0.1
decay of the spots (Eq. (8)). The synthetic eddies are
convected by the mean flow U0. DIG0 uses Gaussian
filters. SPE0 uses an isotropic energy spectrum of the form –0.2

S
(k 2 expðk 2 =k 20 Þ. The size of the structures in SEM0, SPE0
and DIG0 can be controlled so that the three methods pro-
duce inflow data with the same correlations and energy –0.3
spectrum (Fig. 1). SEM0, SPE0 and DIG0 thus only differ
by the procedure chosen to generate the inflow data. It can
be seen on Fig. 2 that the theoretical value of the skewness –0.4
0 3 6 9 12 15 18
is reached after the same distance downstream of the inlet x
for SEM0, SPE0 and DIG0 and that the turbulent energy
Fig. 2. Downstream evolution of the (a) turbulent kinetic energy and (b)
decays at similar rates. The large eddy turn-over time scale velocity derivative skewness for the spatially decaying isotropic
(LETOT) evaluated as T = L/u 0 = 0.4 s (L = 0.4 m, turbulence.
u 0 = 1 m/s) corresponds to a convection distance of
x = 8 m (U0 = 20 m/s). Thus the skewness is established
after one LETOT. The kink on all k profiles from x = 0 have any spatial or temporal correlation. Two neighbour-
to x = 0.5 is most certainly a numerical artifact but not ing points are uncorrelated, which leads to high velocity
linked to incompressibility (even divergence-free structures gradients and high dissipation. The energy cascade does
generate a similar kink). not establish as there is too much energy in the high wave-
RAND does not recover the correct skewness even by number part of the spectrum and the skewness without
the end of the domain and all the energy is dissipated after energy transfer cannot converge towards its theoretical
a few cells downstream the inlet. The inflow data does not value for the case. On Fig. 3, the evolution of the one-
dimensional longitudinal energy spectrum confirms this
interpretation. SEM0 has too little energy at high wave-
2 number due to the Gaussian decay of the spots and energy
10
is cascaded from the low wavenumber part of the spectrum
1
to the high wavenumber part. On the contrary, RAND has
10 too much energy at high wavenumbers and this energy is
0
dissipated. The spectra in both RAND and SEM0 tend
10 towards a classic filtered 5/3 spectrum for LES.
Ek(k)

Different sizes of turbulent spots are now generated at

–1 RAND
10
SPE0 the inlet. Calculations SEM1 has inflow fluctuations with
SEM0 turbulent spots twice smaller than SEM0, whereas SEM2
–2
10 DIG0 has turbulent spots twice bigger than SEM0. The one-
dimensional longitudinal energy spectrum of SEM0,
10
–3
SEM1, SEM2 and RAND is given on Fig. 4. These differ-
0 1 2
10 10 10 ent spectra produce inflow data with different values of the
k dissipation at the inlet. The bigger the spots are, the lower
Fig. 1. One-dimensional longitudinal energy spectrum of the inflow data the dissipation is and the slower the decay of the turbulent
for spatially decaying isotropic turbulence. kinetic energy is as dk/dt = e. As smaller spots are gener-
 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593 589

0 2
10 10
(a) (a)
1
10
–1
10 0
10

E(k)
E(k)

x=0.05 –1 RAND
x=1.5 10
10
–2 SEM0
x=6.28 SEM1
x=12.57 –2
10 SEM2

–3 –3
10 0 1 2
10 0 1 2
10 10 10 10 10 10
k k
2.0
(b) (b)
0
10 1.5
E(k)

–2
10 1.0

k
x=0.05
x=1.5
–4 x=6.28
10 x=12.57 0.5

–6
10 0 1 2 0.0
10 10 10 0 3 6 9 12 15 18
k x
Fig. 3. Development of the longitudinal one-dimensional energy spectrum Fig. 4. (a) One-dimensional longitudinal energy spectrum of the inflow
for the spatially decaying isotropic turbulence. (a) Calculation RAND, (b) data and (b) downstream evolution of the turbulent kinetic energy for
Calculation SEM0. spatially decaying isotropic turbulence.

ated, the inflow data becomes less and less correlated, the
dissipation increases and the decay rate tends towards the Table 2
decay rate of RAND. Computations for the turbulent channel flow case at Re* = 395
Computation Method Details
4. Plane channel flow PREC Prec calculation X
RAND Basic random X
2

The quality of a synthetic turbulent inlet methodology is SPEC Spectral method j2 eðj=j0 Þ ; L ¼ 0:4
SEM0 SEM Gaussian spots, L = 0.4
measured by its capacity to maintain and/or produce self-
SEM1 SEM Gaussian spots, L = 0.8
sustaining turbulence after the shortest possible develop- SEM2 SEM Gaussian spots, L = 0.2
ment period. The theoretical distance of development after SEM3 SEM ui uj ¼ 2=3kdij ; L ¼ k 3=2 =e
which a laminar flow entering a channel is considered as
fully turbulent is more than 110d where d is the channel
half width. It was reported (Le et al., 1997) that about value (CS = 0.065) with Van Driest near-wall damping.
10d were needed to recover correct intensity levels for a Simulations carried out using different inflow data are
DNS of a turbulent boundary layer with an inflow data listed in Table 2.
generated with a spectral method. The precursor simulation PREC uses periodic boundary
The chosen Reynolds number of Re* = 395 for this test conditions in the streamwise direction. It was carried out
case, in combination with a fairly coarse mesh makes the on a domain of dimensions 6d · 2d · 3d. The mesh is the
case more challenging (real LES rather than quasi DNS). same as the one used for the spatially developing flow in
The mesh dimensions are 24d · 2d · 3d (respectively in order to avoid interpolation at the interface. Velocity fields
the x, y and z direction) to allow a fully developed flow from a plane perpendicular to the mean flow were stored
to establish from the inlet. The number of cells is 160 · and injected at the inlet of the spatially developing simula-
30 · 30 and Dx+ = 60, Dy þ þ +
mean ¼ 24, Dy min ¼ 1, Dz = 40. tion. The basic random procedure RAND generates series
Periodic boundary conditions are used in the spanwise of independent random numbers at each point of the inlet
direction and a no-slip boundary condition is used at the mesh at each time step. SEM0 uses the SEM with 3D iso-
walls. The Smagorinsky constant is set to its recommended tropic spots of constant size across the domain. As a first
590 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593

approximation the radius of the spot is chosen as r = 2Dz 1.2

(hence L = 2r = 4Dz = 0.4). The synthetic eddies are con-

1.0
vected by the bulk velocity UB in Eq. (5). The spectral PREC
RAND
method of Lee et al. (1992) with a constant Gaussian spec- 0.8 SPEC
trum across the flow is used at the inflow for SPE0. The SEM0
0.6
spectrum is chosen so that SEM0 and SPE0 produce inflow
data with the same integral length scale. All the simulations 0 3 6 9 12 15 18 21 24

use the same mean velocity and Reynolds stresses profiles Fig. 6. Downstream development of the friction coefficient at the bottom
obtained from the periodic calculation. wall for the channel flow case using different inflow generation methods.
Calculation PREC is used as a baseline for comparison
with other methods. Fig. 5 shows the downstream evolu-
tion of isoprofiles of Q = X2  S2 and Fig. 6 the evolution 2.0
of the friction coefficient. As expected PREC does not pro-
duce any transient downstream of the inlet. On the con- PREC
RAND
trary the intensity of the fluctuations imposed with the 1.5
SPEC
basic random procedure at the inlet of RAND seems to SEM0
decay and the friction coefficient decreases which shows (a)
1.0
that the flow tends to laminarize. The basic random 0 3 6 9 12 15 18 21
method does not manage to produce self-sustaining turbu- 2.0

lence for this case. It can be seen on Fig. 5 that the SEM
succeeds in generating large scale eddies which are sus-
tained downstream of the inlet. The evolution of the skin 1.5

(b)
1.0
0 3 6 9 12 15 18 21
2.0

1.5

(c)
1.0
0 3 6 9 12 15 18 21
2.0

1.5

(d)
1.0
0 3 6 9 12 15 18 21

Fig. 7. Downstream development of the Reynolds stresses profiles on the

top half of the channel for different inflow generation methods, (a)
x þ u2 =3:5, (b) x þ v2  4, (c) x þ w2  2:2 and (d) x þ uv  3:5.

friction coefficient (Fig. 6) and of the Reynolds stresses

(Fig. 7) give more insight into the evolution of the turbu-
lence downstream of the inlet for SEM0. The skin friction
decreases up to x = 3d where it reaches a minimum. This is
caused by a drop of the levels of intensity of all the Rey-
nolds stresses. Then realistic correlations and structures
establish which enhance the mixing of momentum at the
wall. The skin friction increases again and reaches its fully
developed initial value by the end of the domain. The same
behaviour can be observed for SPEC. The large scale struc-
Fig. 5. Isoprofiles Q = 3 for the plane channel flow using different inflow tures introduce at the inlet of SEM0 and SPEC manage to
generation methods. From top to bottom computation RAND, SEM0, evolve towards realistic structures specific to the flow after
SPEC and PREC. a transition period where they adapt to the Navier–Stokes
 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593 591

2.0

PREC
1.5 SEM1
SEM2
SEM0
(a)
1.0
0 3 6 9 12 15 18 21
2.0

1.5

(b)
1.0
0 3 6 9 12 15 18 21
2.0

Fig. 8. Isoprofiles Q = 3 for the plane channel flow using the SEM with 1.5
different sizes of eddies. Computation SEM1 (top) is with eddies larger
than in the reference calculation SEM0, computation SEM2 (bottom) is
with smaller ones. (c)
1.0
0 3 6 9 12 15 18 21
2.0

equations and the numerical scheme and by the end of the

domain we have fully developed turbulence. 1.5
Different sizes of eddies will now be tested to try to
reduce the development section. Fig. 8 shows isoprofiles (d)
of Q for the larger SEM1 and smaller SEM2 inlet struc- 1.0
0 3 6 9 12 15 18 21
tures. SEM2 has structures of the size of the structures of
the real flow at the wall, whereas SEM1 has structures of Fig. 10. Downstream development of the Reynolds stresses profiles
the size of the real flow in the middle of the channel. The downstream the inlet on the top half of the channel Re* = 395 for different
sizes of eddies (a) x þ u2 =3:5, (b) x þ v2  4, (c) x þ w2  2:2 and (d)
structures generated with SEM2 seem too small to generate x þ uv  3:5.
fully developed turbulence by the end of the domain. These
tend to decay rather than to evolve towards larger scales.
On the contrary SEM1 starts with too large structures
SEM1 and SEM2. All the stresses on Fig. 10 are vanishing
but in the second half of the channel these are seen to adapt
towards the end of the domain for SEM2. It can be seen
(or possibly generate new structures) with correct length
that all the stresses reach fully developed turbulence levels
scales. This confirms the results obtained in isotropic tur-
at the wall faster for SEM0 than for SEM1. On the con-
bulence where energy at low wavenumbers is more easily
trary, the structures in the middle of the channel of
cascaded into high wavenumbers than the opposite. The
SEM0 are smaller than those of SEM1 so they are dissi-
evolution of the friction coefficient is given on Fig. 9. The
pated faster and SEM0 lacks energy in the middle of the
large scale structures generated at the inlet of SEM1 are
channel. This stresses the importance of having different
not adapted to the flow at the wall and the drop in the fric-
sizes of structures in different regions of the flow.
tion coefficient is more important than for SEM0. However
Finally, a calculation using only RANS information is
by the end of the domain the large eddies finally manage to
used. A k  e calculation of the channel flow at Re* =
break up into suitable structures for the flow and the fric-
395 is run and profiles of k and e from this calculation
tion recovers its initial value. Fig. 10 shows the evolution of
are used as the only source of information for the SEM.
the Reynolds stresses downstream of the inlet for SEM0,
Instead of the resolved Reynolds stress tensor obtained
from the periodic LES, the model Reynolds stress tensor
obtained from the k  e calculation is used ui uj ¼
1.10
2=3kdij  2mt S ij . The size of the structures is calculated
1.00 using a turbulent length scale k3/2/e. In order for the syn-
thetic eddies to be properly discretized by the inlet mesh,
0.90 SEM0
SEM1
the radius of the eddies reads r = max(Dz, k3/2/e). On
0.80 SEM2 Fig. 11(a) the downstream evolution of the friction coeffi-
0.70
cient shows a very short transient and an error of less than
0 3 6 9 12 15 18 21 24 5% is attained after only 3d. This value even increases
Fig. 9. Downstream development of the friction coefficient cf(x)/cf(0) at just downstream of the inlet. This can be explained by
the bottom wall of the channel for different sizes of eddies. the higher value of the wall-normal Reynolds stress v2
592 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593

1.2 inflow structures which can be properly discretized by the

1.1
inlet mesh even if it means overestimating their size.
The generation of synthetic turbulence with reduced
1.0 information coming from a RANS simulation gives a very
0.9
short transition section downstream of the inlet for the
(a) channel flow case and allows us to use very short domains,
0.8
0 3 6 9 12 15 18 21 24 thus reducing the computational cost. the synthetic-
2.0 eddy-method is therefore very promising to force synthetic
turbulence at the interface of a hybrid RANS/LES flow
solver. The method is currently being extended to represent
1.5 PREC
SEM3
turbulence generation parallel to the wall for RANS/LES
simulations of high Reynolds number channel flows on
(b)
coarse grids.
1.0
0 3 6 9 12 15 18 21
2.0
Acknowledgements

1.5
The work was partially supported by the EU project
DESider, which is a collaboration between Alenia, ANSYS–
(c) AEA, Chalmers University, CNRS–Lille, Dassault, DLR,
1.0 EADS Military Aircraft, EUROCOPTER Germany,
0 3 6 9 12 15 18 21
EDF, FOI–FFA, IMFT, Imperial College London, NLR,
Fig. 11. Downstream development of (a) the friction coefficient cf(x)/cf(0) NTS, NUMECA, ONERA, TU Berlin, and UMIST. The
at the bottom wall of the channel, (b) x þ u2 =3:5 and (c) x þ v2  4. project is funded by the European Community represented
by the CEC, Research Directorate–General, in the 6th
Framework Programme, under Contract No. AST3-CT-
2003-502842.
generated in SEM3 (as erroneously predicted by the k  e
model) which strongly enhances the mixing of momentum
References
at the wall. It can be seen on Fig. 11(b) that u2 which is
underestimated at the inlet reaches fully developed value Archambeau, F., Mehitoua, N., Sakiz, M., 2004. Code saturne: a finite
after only 3d. The excess of v2 at the inlet is quickly dissi- volume code for the computation of turbulent incompressible flows.
pated and after 3d too, the fully developed turbulent pro- International Journal on Finite Volumes 1 (1).
files are recovered (Fig. 11(c)). The importance of the Batten, P., Goldberg, U., Chakravarthy, S., 2004. Interfacing statistical
wall-normal velocity fluctuations has already been stressed turbulence closures with large-eddy simulation. AIAA Journal 42 (3),
485.
by Keating et al. (2004) and Spille-Kohoff and Kaltenbach Benhamadouche, S., Laurence, D., 2003. LES, coarse LES, and transient
(2001). RANS comparisons of the flow across a tube bundle. International
Journal of Heat and Fluid Flow 24, 470–479.
5. Conclusion di Mare, L., Jones, W.P. 2005. Algebraic and operators methods for
generation of inflow data for LES and DNS, TSFP4. In: 4th
International Symposium on Turbulence and Shear Flow Phenomena,
A new method for generating turbulent inlet boundary vol. 2, Williamsburg, VA, USA, pp. 687–692.
conditions has been developed and compared to existing Druault, P., Lardeau, S., Bonnet, J.-P., Coiffet, F., Delville, J., Lamballais,
methods on two test cases. The method generates synthetic E., Largeau, J.-F., Perret, L., 2004. Generation of Three-Dimensional
eddies on the inlet plane. Each eddy is represented by a spe- Turbulent Inlet Conditions for Large-Eddy Simulation. AIAA Journal
cific shape function which describes its spatial and tempo- 42 (3), 447.
Jarrin, N., Benhamadouche, S., Addad, Y., Laurence, D. 2003. Synthetic
ral characteristics. The method is able to reproduce specific turbulent inflow conditions for large eddy simulation. In: Proceedings,
first and second order one point statistics as well as auto- 4th International Turbulence, Heat and Mass Transfer Conference,
correlation functions. Antalya, Turkey.
Compared to the random method, the new approach Kaltenbach, H.-J., Fatica, M., Mittal, R., Lund, T.S., Moin, P., 1999.
can produce spatial and temporal correlations which pro- Study of flow in a planar asymmetric diffuser using large-eddy
simulation. Journal of Fluid Mechanics 390, 151–185.
duce fully developed turbulence in a channel flow a few Keating, A., Piomelli, U., Balaras, E., Kaltenbach, H.-J., 2004. A priori
diameters downstream of the inlet. It gives similar results and a posteriori tests of inflow conditions for large-eddy simulation.
to the spectral method of Lee et al. (1992) for the two test Physics of Fluids 16 (12), 4696–4706.
cases computed but is more appropriate for complex geom- Klein, M., Sadiki, A., Janicka, J., 2003. A digital filter based generation of
etries on unstructured mesh. inflow data for spatially developing direct numerical or large eddy
simulations. Journal of Computational Physics 186, 652–665.
The importance of the size of the structures at the inflow Kondo, K., Murakami, S., Mochida, A., 1997. Generation of velocity
has been shown both on the isotropic turbulence and on fluctuations for inflow boundary conditions of LES. Journal of Wind
the channel test case. It appears to be essential to generate Engineering and Industrial Aerodynamics 67&68, 51–64.
 N. Jarrin et al. / Int. J. Heat and Fluid Flow 27 (2006) 585–593 593

Le, H., Moin, P., Kim, J., 1997. Direct numerical simulation of turbulent Smirnov, A., Shi, S., Celik, I., 2001. Random flow generation technique
flow over a backward-facing step. Journal of Fluid Mechanics 330, for large eddy simulations and particle-dynamics modeling. Journal of
349–374. Fluids Engineering 123, 359–371.
Lee, S., Lele, S., Moin, P., 1992. Simulation of spatially evolving Spille-Kohoff, A., Kaltenbach, H.-J. 2001. Generation of turbulent inflow
compressible turbulence and the application of Taylors hypothesis. data with a prescribed shear-stress profile. In: Liu, C., Sakell, L.,
Physics of Fluids, 1521–1530. Beutner, T. (Eds.), Third AFOSR International Conference on DNS/
Lund, T., Wu, X., Squires, D., 1998. Generation of turbulent inflow data LES Arlington, TX, 5–9 August 2001, DNS/LES Progress and
for spatially-developing boundary layer simulations. Journal of Challenges. Greyden, Colombus, OH.
Computational Physics 140, 233–258.