Direct Numerical Simulation

June 8, 2010

In a Direct Numerical Simulation (DNS), the Navier-Stokes equations

are numerically solved without using a turbulence model. In other words, all
spatial and temporal scales of turbulence must be resolved, from the smallest
dissipative scales (Kolmogorov scales), up to the integral scale, associated
with the motions containing most of the kinetic energy. This demands a
very dense computational mesh, which, in turn, requires heavy computational
resources.
Direct Numerical Simulation of industrially-important flows is still unaf-
fordable due to hardware constraints but it has been possible in the recent
past to carry out DNS of some academic or canonical flows like isotropic
turbulence and channel flow albeit at low Reynold numbers. Still relatively
few DNS have been carried out for wall-bounded flows with adverse pressure
gradient.

Spalart & Watmuff (1993) carried out a DNS of adverse pressure gradi-
ent turbulent boundary layer at Reθ = 600. They solved the incompressible
Navier-Stokes equations over a flat plate with periodic conditions in the x−
and z− directions, where x is the streamwise and z is the spanwise direction.
A spectral method with Fourier series was used for spatial discretization in
both directions. Temporal discretization was through a second order hybrid
finite-difference scheme.

Skote & Henningson (2002) performed DNS of a turbulent boundary layer

with adverse pressure gradient. The Reynolds number Reδ = U δ/ν at the
inlet of the computational domain was 400. The highest resolution used was
720 × 217 × 256, which amounts to a total of 40 million modes. They used
a spectral method with Fourier discretization in the horizontal directions
and Chebyshev discretization in the wall-normal direction. The use of non-
periodic boundary conditions in the streamwise direction was made possible
through addition of a fringe region downstream of the physical domain. In

1
 2

the fringe region, the flow is forced from the exit of the physical domain to
its inlet. Time integration was performed using a third-order RungeKutta
method and a CrankNicholson scheme was used for the viscous terms.

Lee & Sung (2008) performed DNS of a spatially-developping adverse

pressure gradient turbulent boundary layer with an Reθ range of 300-1500.
They applied periodic boundary conditions in the spanwise direction with
no-slip condition imposed at the solid wall. Convective boundary condition
(∂u/∂t) + c(∂u/∂x) = 0 was applied at the exit, with c as the bulk mean
velocity. Terms were discretized in time using a Crank-Nicholson scheme fol-
lowed by implicit velocity-decoupling with a second-order overall accuracy.
For spatial discretization, a second-order central difference scheme with a
staggered mesh was used.

All of the above direct numerical simulations were performed with flows
over flat surfaces. To investigate the turbulence statistics and coherent struc-
tures of wall-bounded flows under strong pressure gradient with and with-
out curvature, Direct Numerical Simulation (DNS) of a converging-diverging
channel flow was performed by Marquillie et al. (2008) at Reτ = uτ h/ν =
395. Adverse pressure gradient was created through a surface bump on the
lower wall of the channel. This DNS of adverse pressure gradient flow was
conducted as part of the European project WALLTURB to meet two main
objectives. The first objective was to gather fully resolved three-dimensional
data to study structures, statistics and scaling of wall-bounded turbulence.
Secondly, it was also meant to serve as a reference for the evaluation of RANS
and LES models.
The geometry of the bump used in the simulation is the same as in
the wind-tunnel experiments of Bernard et al. (2003) at “Laboratoire de
Mécanique de Lille”. The bump was designed by Dassault Aviation for the
AEROMEMS project to replicate the flow conditions over the wing of an air-
plane at high angle of attack. The surface bump consists of concave surfaces
at the leading and trailing edges (to achieve continuity with flat surface) and
a convex surface in the middle. The wind-tunnel experiments of Bernard
et al. (2003) were conducted at a Reynolds number of upto Reθ ' 20000
(Reτ > 7000). In the following paragraphs, we shall describe a new DNS
that has been performed on the same bump geometry at Reτ = 617.
 3

Figure 1: Computational grid of the DNS at Reτ = 617 in the XY plane.

The flow is from left to right.

The DNS was performed on a domain, which is 4π long in streamwise

direction, 2 in wall-normal direction and π in spanwise direction. The grid
of the computational domain in the XY plane is shown in the figure 1. Ev-
ery 24th mesh line is shown for each direction. The spatial resolutions is
(2304 × 385 × 576). The ratio of Kolmogorov scale to the maximum mesh
size is of the order of 2 to 3 in most of the domain, goes up to 4 in the di-
verging part and up to a value of 5 very close to the wall. However, in order
to evaluate the discretisation of the near wall region, the mesh size in wall
units (∆x+ , ∆y + , ∆z + ) are more relevant. The maximum values at the inlet
+
are ∆x+ = 5.1, ∆ymin = 0.02 and ∆z + = 3.4. The global maximum values
+
are reached in the converging part of the channel with ∆x+ = 10.7, ∆ymin =
+
0.03 and ∆z = 7.4. The resolution of the computational domain compares
favorably with other contemporary simulations. Lee & Sung (2009) used a
+
resolution of ∆x+ = 12.5, ∆ymin +
= 0.17, ∆ymax = 24 and ∆z + = 5 based
on friction velocity at the inlet. The resolution in streamwise direction ∆x+
was 5.91 for Wu & Moin (2009), ≈ 13 for Skote et al. (1998) and ≈ 16.6 for
Na & Moin (1998). Thus numerical resolution for the current DNS is fine
enough to capture and accurately represent the vortical structures.

Simulations were performed using the MFLOPS3D code developped at

LML. This code was originally developped to investigate the 2D and 3D
laminar boundary layer instabilities over a bump. It was then modified to
study the effects of pressure gradient on wall-bounded turbulent flows. The
code resolves the non-dimensional Navier-Stokes equations with a Reynold
number Re = Umax h/ν = 12600 where h, the half channel height, is equal
to 1. Umax is the maximum velocity at the inlet and it is almost equal to
1. The Reynold number based on the friction velocity is Reτ = 617. In-
stead of writing the Navier-Stokes equations in curvilinear coordinates, the
wall curvature (for bump geometry) is obtained by a mathematical mapping
of the partial differential operators from physical coordinates to Cartesian
ones (Marquillie et al., 2008). The DNS code was parallelized using Message
 4

Passing Interface (MPI) and was operated using 64 vectorial processors NEC-
SX8 for a total of 160000 CPU hours at an average performance of 640 Gflops.

For spatial discretization in the streamwise direction, 4th order central

finite differences are used for 2nd derivatives and 8th order finite differences
are used for 1st derivatives. Chebyshev-collocation is used in the wall-normal
direction. The spanwise direction z is considered periodic and is discretized
using a spectral Fourier expansion. For time-integration, implicit 2nd or-
der backward Euler differencing is used. Thus, numerical code combines the
advantage of good accuracy and spectral resolution with fast integration pro-
cedure for flow simulations over smooth surfaces. DNS was integrated over
50 convective times (based on half the channel height and the maximum ve-
locity at the inlet) and 982 velocity and pressure fields were recorded every
0.06 convective time. A total of 9Tb of velocity and pressure fields were
stored in order to compute converged statistics. Further details of the code
can be found in Marquillie et al. (2008).

The pressure coefficient Cp = (P −Po )/ 21 ρUmax

2


for the present DNS and
the LML wind-tunnel experiment is shown in figure 2, where Po is the refer-
ence pressure and Umax is the maximum velocity. The order of magnitude and
evolution of the pressure gradient over the domain is almost the same as in
the experiment. Small differences are due to the different Reynolds numbers
and inlet conditions. The pressure gradient parameter P + = (dp/dx)ν/ρu3τ
for both lower and upper walls is given in figure 3. P + rises sharply after the
bump summit and goes to infinity due to the friction coefficient Cf that goes
to zero (figure 4). At s=1.5, the pressure gradient P + again becomes finite at
a value of almost 0.4 and continues to drop further downstream. The values
of P + for Lee & Sung (2008), Spalart & Watmuff (1993) and Skote et al.
(1998) were 0.026, 0.02 and 0.25 respectively. Therefore, as compared to
previous simulations, the adverse pressure gradient in the present simulation
is very strong. Generally, an adverse pressure gradient with P + > 0.15 is
regarded as a strong pressure gradient (Spalart & Watmuff, 1993).

The friction coefficient Cf = τw / 12 ρUmax2

for the DNS is shown in
figure.4, where τw is the wall shear stress. It is compared to the friction
coefficient from the experiment of Bernard et al. (2003). The flow along the
upper-wall remains attached while the flow at the lower-wall separates just
after the maximum height of the bump to form a small recirculation region.
The thickness of the recirculation region is about 20 wall units (based on
inlet quantities) and its streamwise extent is almost from x = 0.5 to 1.5 with
x = 1.5 as the flow reattachment point . The skin friction at the upper-wall
 5

decreases to a very small value (Cf ' 1.55E-03) and is close to separation.
As shown by the Cf distribution, it should be noted that the flow does not
separate at the Reynolds number of the experiment (which is much higher).
The value of Cf for both the DNS and experiment are comparable near the
summit of the bump (s = 0). They differ in the diverging part, due mostly
to the small separation bubble in the DNS. Figure 5 shows the probability
density function of the reverse flow. The recirculation region is usually de-
fined as the region with the PDF of reverse flow > 1/2.

Streamwise evolution of the mean velocity profiles over the bump is shown
in the figure 6. As is evident from the mean velocity profiles at the lower
wall, internal layers are created at a sign change in the surface curvature with
thickness δ ' 0.06 at the summit of the lower wall of the bump. Boundary
layer at the lower wall starts to grow rapidly under the effect of adverse
pressure gradient and continues to grow till the end of the computational
domain. Mean velocity profiles at the last two streamwise locations shows
that the flow has started to relax from the pressure gradient but has not yet
recovered and the effects of the pressure gradient are present till the end of
the computational domain.
 6

Figure 2: Comparison of the pressure coefficient Cp for DNS with the exper-
iment of Bernard et al. (2003)

Figure 3: Pressure gradient P + of both upper and lower walls of the

converging-diverging channel
 7

Figure 4: Comparison of the skin friction coefficient Cf for DNS with the
experiment of Bernard et al. (2003)

Figure 5: Probability Density Function (PDF) of the reverse flow over the
bump.
 8

Figure 6: Mean Streamwise velocity profiles over the bump.

Turbulent kinetic energy (k ∗ ) profiles with respect to wall-normal distance

in wall units based on the friction velocity at s = 0 are shown in figure 7
at some streamwise locations. As uτ is taken constant, profiles are just set
non-dimensional after s = 0 but should not be considered as in the wall unit
representation. This is why they are noted ’*’ instead of ’+’.
All profiles show a clear turbulence peak which starts fairly close to the
wall (n∗ = 10 at s = 0) and moves away from it as s increases (n∗ = 200
at s = 4). This displacement takes place in two phases. In the first phase
(0 ≤ s ≤ 1.5), the peak moves away and grows rapidly (upto k ∗ = 17). In
the second phase, the turbulence spreads out, the peak diminishes, flattens
and moves slowly out to occupy most of the boundary layer thickness. The
streamwise location s= -5.2 represents a nearly zero pressure gradient flow
and the near-wall peak of k ∗ has a value of only 2 as compared to almost 17
at s = 1.5 (APG region).
∗
Variation of the Reynolds shear stress (−u0 v 0 ) with wall-normal distance
(n) in the same representation as k ∗ is shown in figure 8. Like the turbulent
kinetic energy, Reynolds shear stress reaches its maximum at s = 1.5, which
is the flow reattachment point. As shown in figure 8, the Reynolds shear
stress is nearly suppressed by the favorable pressure gradient till the bump
summit at s =0 but a strong peak is created by the APG. Adverse pressure
gradient starts to strongly affect the Reynolds shear stress after s = 0 and a
large increase in the magnitude is observed together with a peak displacement
away from the wall. Like for the k ∗ , s = 1.5 is the end of the first phase of
development. Afterwards, the peak decreases, flattens out and moves away
from the wall in a manner very similar to k ∗ .
 9

18
s=-5.2
s=0.0
16 s=0.5
s=1.0
s=1.5
14 s=2.0
s=2.5
s=3.0
12 s=4.0

10
k*

0
0 100 200 300 400 500
*
n

Figure 7: Wall-normal evolution of the turbulent kinetic energy in wall units.

6
s=-5.2
s=0.0
s=0.5
5 s=1.0
s=1.5
s=2.0
s=2.5
4 s=3.0
s=4.0
-u′v′*

0 100 200 300 400 500

*
n

Figure 8: Wall-normal evolution of the Reynolds shear stress in wall units.

 10

Turbulent energy in wall-bounded flows or free shear flows is dissipated

at very small scales into heat. The characteristic length scale η, velocity
scale v and time scale τ associated with these smallest scales (or very small
scales) are called Kolmogorov scales. The expressions for Kolmogorov scales
are given as,

1/4
ν3
  ν 1/2
η= τ= v = (ν)1/4 (1)
 

where  is the dissipation rate of turbulent kinetic energy, given by,

 = 2ν < s0ij s0ij > (2)

where s0ij is the symmetric part of the fluctuating velocity gradient tensor,
given by,

1 ∂u0i ∂u0j
 
0
sij = +
2 ∂xj ∂xi

The dissipation  is shown in the figures 9 and 10 in linear and log plots
respectively. Due to the favorable pressure gradient between s = -5.2 and s =
0, a dissipation plateau is clearly visible near the wall at s = 0. This is fairly
different from a standard ZPG configuration as illustrated by the profile at
s = -5.2 on both figures, which shows a peak at the wall, as expected. After
the summit, a very specific behaviour develops. An outer peak, tightly linked
to the k ∗ peak, develops till s = 1.5 but in this case associated to a wall peak
of comparable amplitude. This double peak configuration leads to a fairly
high level of dissipation in the whole near-wall region. After s = 1.5, the
global level of dissipation falls down progressively, the outer peak flattens
∗
and moves away from the wall (together with the peak in k ∗ and u0 v 0 ) and
the profiles return progressively to a single peak at the wall but with a flatter
distribution all over the boundary layer thickness.
The wall-normal evolution of the Kolmogorov length scale η is shown in
figures 11 and 12 in linear and log plots respectively. At s = -5.2 and s =
0, η ∗ starts at 2, which is a well-admitted near-wall value in wall units for a
ZPG turbulent boundary layer. The difference is that although the pressure
gradient is almost zero at s = 0, the upstream history is fairly different from
ZPG. The consequence is a wide plateau of η ∗ = 2 near the wall, linked to
the plateau of ∗ observed in the previous figure. Further away from the wall,
 11

η∗ increases steadily in a manner which can be considered as standard in a

boundary layer. In the adverse pressure gradient part, apart from s = 0.5
where the decrease of ∗ induces a slight increase of η ∗ near the wall, the
∗
effect of the developing turbulence peak observed in k ∗ , u0 v 0 and ∗ is to
lower down η ∗ into a plateau which widens with s and occupies more or less
the width of the peak of these quantities. The value of this plateau decreases
down to η ∗ = 1 at s = 1.5 (which is the point of maximum intensity of all
∗
three peaks: k ∗ , u0 v 0 and ∗ ) and then raises slowly up towards a two step
plateau with slightly different values in the very near-wall and more outer
region. In all cases, the plateau converts to a steadily growing behaviour in
the outer part of the boundary layer, this with a comparable slope but with
a different value at origin, depending upon the s location considered. This
value at origin in fact decreases monotonically with s, showing the progressive
outer influence of the near-wall changes in η ∗ .
In this chapter, only the statistics relevant to our study (of vortical struc-
tures) have been presented. The complete budget of the Reynold stresses
was computed and part of the statistics are presented in Laval & Marquil-
lie (2009). The DNS database being highly resolved, it represents a real
opportunity to study coherent structures in the adverse pressure gradient
wall-bounded flow. A detection of coherent structures on the current DNS
database is presented in the next chapters.
 12

0.45
s=-5.2
s=0.0
0.4 s=0.5
s=1.0
s=1.5
0.35 s=2.0
s=2.5
s=3.0
0.3 s=4.0

0.25
ε*

0.2

0.15

0.1

0.05

0
50 100 150 200 250 300 350 400
*
n

Figure 9: Wall-normal evolution of the dissipation in wall units.

0.45
s=-5.2
s=0.0
0.4 s=0.5
s=1.0
s=1.5
0.35 s=2.0
s=2.5
s=3.0
0.3 s=4.0

0.25
ε*

0.2

0.15

0.1

0.05

0
10 100
*
n

Figure 10: Wall-normal evolution of the dissipation in wall units.

 13

9
s=-5.2
s=0.0
8 s=0.5
s=1.0
s=1.5
7 s=2.0
s=2.5
s=3.0
6 s=4.0

5
η*

0
50 100 150 200 250 300 350 400 450 500
*
n

Figure 11: Wall-normal evolution of the Kolmogorov length scale.

9
s=-5.2
s=0.0
8 s=0.5
s=1.0
s=1.5
7 s=2.0
s=2.5
s=3.0
6 s=4.0

5
η*

0
1 10 100
*
n

Figure 12: Wall-normal evolution of the Kolmogorov length scale.

 14

References
Bernard, A., Foucaut, J. M., Dupont, P. & Stanislas, M., 2003
Decelerating boundary layer : a new scaling and mixing length model.
AIAA Journal 41 (2), 248–255.

Laval, J.-P. & Marquillie, M., 2009 Direct numerical simulations of

converging-diverging channel flow. In Stanislas, M., editor, Progress in
Wall Turbulence: Understanding and Modeling (to appear), Lille, France.
Springer.

Lee, J.-H. & Sung, H., 2008 Effects of an adverse pressure gradient on a
turbulent boundary layer. Int J. of Heat and Fluid Flow 29, 568–578.

Lee, J.-H. & Sung, H., 2009 Structures in turbulent boundary layers sub-
jected to adverse pressure gradients. J. Fluid Mech 639, 101–131.

Marquillie, M., Laval, J.-P. & Dolganov, R., 2008 Direct numerical
simulation of separated channel flows with a smooth profile. J. Turbulence
9 (1), 1–23.

Na, Y. & Moin, P., 1998 Direct numerical simulation of a separated tur-
bulent boundary layer. J. Fluid Mech. 374, 379–405.

Skote, M., Henningson, D. & Henkes, R., 1998 Direct numerical sim-
ulation of self-similar turbulent boundary layers in adverse pressure gradi-
ents. Flow, Turbulence and Combustion 60, 47–85.

Skote, M. & Henningson, D. S., 2002 Direct numerical simulation of

separating turbulent boundary layers. J. Fluid Mech. 471, 107–136.

Spalart, P. R. & Watmuff, J. H., 1993 Experimental and numerical

investigation of a turbulent boundary layer with pressure gradients. J.
Fluid Mech. 249, 337–371.

Wu, X. & Moin, P., 2009 Direct numerical simulation of turbulence in a

nominally zero-pressure-gradient flat-plate boundary layer. J.Fluid.Mech
630, 5–41.