Journal of Wind Engineering and Industrial Aerodynamics, 35 (1990) 101-128 101

Elsevier Science Publishers B.V., Amsterdam - - Printed in The Netherlands

ANALYSING TURBULENT BACKWARD-FACING STEP FLOW

WITH THE LOWPASS-FILTERED NAVIER-STOKES
EQUATIONS

RAINER FRIEDRICH and MICHEL ARNAL

Lehrstuhl fiir StrSmungsmechanik der Technischen Universitgtt Mi~nchen Arcisstr. 21,
D-8000 Miinchen 2 (F.R.G.)
(Received February 5, 1990 )

Summary
High Reynolds number turbulent backward-facing step flow is studied using the large-eddy
simulation technique proposed by Schumann. Central differencing on an equidistant cartesian
grid and explicit time integration are chosen to predict the complicated flow structure. The com-
parison with experimental data indicates good agreement for the statistical quantities. Part of the
disagreement is certainly due to scatter in the experimental data. Nevertheless, the results of the
simulation may be further improved using globally or locally refined grids. Valuable information
on the instantaneous flow behaviour in the free-shear layer and the reattachment zone is obtained
from the simulation.

1. Introduction

Flow phenomena such as separation, recirculation, reattachment and even-

tual relaxation towards a fully developed flow appear in many technical appli-
cations such as industrial aerodynamics. The backward-facing step in a chan-
nel is a useful prototype configuration which permits one to study all these
phenomena. Although its geometry is simple and the separation line is fixed,
the accurate numerical prediction of turbulent backward-facing step flow is
still difficult. This is almost certainly due to extra strain rates in the mixing
layer and to the complex small-scale phenomena in the reattachment region.
The flow was one of the test cases for the 1980-81 A F O S R - H T T M - S t a n f o r d
Conference on Complex Turbulent Flows [ 1 ]. Up until now, even second-order
closure models fail to "predict" recirculating flows of this kind properly. At
least Peric et al. [2 ] have succeeded in showing that not all deficiencies are due
to modelling. A great deal of improvement can be achieved by using sophisti-
cated numerical techniques and high spatial resolution.
The large-eddy simulation (LES) technique provides an insightful alter-
native to the statistical modelling approach to turbulent flows. Although we

0167-6105/90/$03.50 © 1990--Elsevier Science Publishers B.V.

102

are not yet able to demonstrate that LES does a better job, we are convinced
that in the long run it will prove to be the more powerful tool. The LES tech-
nique allows us to investigate the instantaneous flow structure and to predict
statistical quantities correctly that are still not amenable to measurement (e.g.
pressure-velocity or pressure-strain correlations).
Why LES and not direct numerical simulation (DNS) which covers all as-
pects of turbulence, even the small-scale effects? The reason is that DNS faces
a serious resolution problem at high Reynolds numbers. For a Reynolds num-
ber Re~= 6480, based on the friction velocity u~ and a characteristic length h,
the necessary spatial resolution to describe the detailed flow dynamics in a
domain of size h a amounts to Re~/4 ~-0.376 × 109 grid points. Allowing that the
computational domain for the backward-facing step flow should have at least
a size of 16 × 4 × 2 in terms of the step height h in the streamwise, spanwise
and vertical directions, one quickly realizes that such grid systems are beyond
the capabilities of available computer systems. LES overcomes this resolution
barrier. It removes the smallest turbulence scales through lowpass filtering of
the basic equations and models their influence upon the filtered flow variables.
Among the different filtering techniques we prefer Schumann's volume-av-
eraging approach [3]. The filtered or grid-scale (GS) variables are the three
velocity components averaged over the three surfaces ASi of the grid volume
AV, and the volume-averaged pressure. For a cartesian grid the GS variables
are defined as

•
Ju~=
f ASj'
fJ AV
dV (1)
A,~i /,v

The averaged velocity components are perpendicular to the surface elements

ASj. The filtering process leads to a closure problem in the convective momen-
tum transport. Splitting the instantaneous velocity vector into its surface av-
erage and a deviation from the average according to
j i
U i = Ui"~-U i (2)
produces an unknown subgrid-scale (SGS) stress uiuj
' ' as follows
j t !
J1AiUj~JUiJUj"~ UiU j (3)
which, unlike the Reynolds stress, can be neglected for grid volumes with a
linear dimension of the order of the Kolmogoroff microscale.

2. The lowpass-filtered N a v i e r - S t o k e s equations

Schumann's volume averaging yields the filtered Navier-Stokes equations

in a spatially discretized form:
 103

5i~ =0 (4)
.__ ~ v--
0 t 'U i = --(~j()Ui)Uj "~ ) ~Uj" ) - O i "P+ vSiiDi~ (5)
P
Equation (5) is solved on a staggered grid. 5j indicates the central finite dif-
ference operator:

6 i ~ _ q)(xj + Axj/2) - O(xj - Axj/2) (6)

Axj

In order to close the system (4), (5) we relate the unknown SGS stress to the
filtered deformation tensor:

j UiUjt t = __ /~o (/~j/j _ ( SD~.} ) - #i,h (JD~ } + 'x~vijkUkUk,' (7)

This is Schumann's two-part eddy viscosity ansatz [3]. The fluctuating first
part assumes local isotropy in the SGS turbulence, whereas the second part
takes care of inhomogeneities in the flow field which occur particularly in the
near-wall region. The angular brackets stand for statistical averages. The last
term on the RHS of (7) ensures equality on both sides for i=j. The parame-
trization for P~nhdiffers somewhat from that in ref. 3. Details of the complete
model are given in refs. 4 and 5 together with a discussion of the action of the
two model parts, based on balance equations for the kinetic energy of the mean
GS motion and the GS fluctuations.
An explicit scheme serves to integrate (5) in time. The initial time step of
each simulation is an Euler step. The time integration is then continued with
second-order leap-frog steps with an averaging time step every 50 steps for
stability purposes. Time steps are chosen such that the Courant numbers are
always well below unity. Each time step is split into substeps. During the first
substep an approximate velocity field is computed from (5) neglecting the
pressure gradient. The pressure gradient is then determined from a 3D Poisson
equation which results from inserting the correction relation for the velocity
field into the continuity equation (4). The correction relation is used in the
final substep to make the velocity field divergence free.
The Poisson equation for the pressure is solved with the help of direct meth-
ods such as applying a fast Fourier transform in the spanwise (periodic) di-
rection and the cyclic reduction technique [6]. These methods work as long as
the computational domain is box like, but fail if it contains corners as in the
case of backward-facing step flow with an inflow portion. In this case we make
use of the capacitance matrix technique [7], which essentially reduces the
whole task to the solution of two Poisson problems.
104

3. Initial and boundary conditions

Each LES needs initial values for the filtered velocity vector in the entire
computational domain. This vector consists of a statistical and a fluctuating
part, namely*
- - rr
Lti : (U~}--U i (8)

The double prime now indicates a fluctuation in the Reynolds sense and the
overbar the above spatial filtering.
The statistical (time-averaged) part ofui is obtained from a flow simulation
with the code CHAMPION [8]. This code uses the standard k-~ model and
standard wall functions. Geometry, grid and global flow parameters are spe-
cified as for the subsequent LES. The fluctuating part u/" is constructed from
gaussian random numbers which are weighted with (2k/3) 1/2 and made diver-
gence free by applying the Poisson solver once. Continuation runs are started
from former large-eddy simulations. If changes in the global flow conditions
are desired, these are taken into account in the statistical velocity components
only.
Crucial to any LES of spatially developing flows are the boundary conditions.
The conditions at the inflow plane in particular must be carefully designed.
There, the instantaneous GS velocity vector must be specified during the whole
simulation process. Since, up until now, no experiment has been able to provide
these data, we proceed as follows.
(1) Perform an LES of a fully developed channel flow with periodic bound-
ary conditions in the streamwise direction (Reynolds number, grid, time step
and SGS model are the same as in the step flow simulation).
(2) Choose a plane perpendicular to the mean flow and store the GS velocity
vector in this plane on magnetic tape for all time steps needed.
(3) Use these data as inflow conditions for the backward-facing step flow
either in the plane of the sudden expansion or suitably upstream of the step.
In the spanwise direction we assume periodicity of the GS variables, which
is consistent with the homogeneity of turbulence in this direction.
In the outflow plane the statistical velocity components are extrapolated
linearly and simplified convection equations are solved for the velocity fluc-
tuations. In this case the convection speed is assumed to be the statistical lon-
gitudinal velocity component. These conditions guarantee a smooth transport
of turbulence structures through the outflow plane [9 ].
From the filtered m o m e n t u m equation (5) it follows that wall boundary
conditions can only be specified for the normal velocity and the two shear
stress components since the tangential velocity components are not defined in

*The upper index has been omitted for simplicity. Note that the statistical and spatial averaging
procedures are interchangeable.
 105

the wall plane itself. Following Schumann [3] we assume the wall shear stress
to be in phase with the tangential velocity component of the wall nearest grid
volume. The factor of proportionality is empirical [10]. The aim of further
work is to test conditions proposed in ref. 11 in which the wall shear stress is
correlated with the velocity of a typical near-wall flow structure.

t.8
(a)
~ -
e
Calcs.
Schmitt
Tropea

t.6

1.4

1-

N
1.2

1.0
0.0 0.2 0.4 O.S 0.8 1.0 ! .2
U/U01

40
(b)

30

:3 20

10
o
Y []
Log Law
Calcs.
FS Exp t
o CT Exp t

101 102 103 104

Z+

Fig. 1. Mean longitudinal velocity scaled with global and inner variables.
106

4. Results

In previous studies [5, 10] we have discussed LES results for backward-
facing step flow in a channel at a Reynolds number Re~ = 6480 (based on step
height h and friction velocity u~ of the incoming fully developed channel flow)

2.0 ~ - Calcs.
/ ~ e Schmitt
,8 j~..-/ ~ Tropea

1.6 -- ~ --

::I::1"4

0.0 0.2 0.4 0.6 0.8 t .0 ! .2 1.4.

UU/U012 . 1 0 -2

2.0
- Calcs.
o Schmitt
Tropea
I.a

X=-0.25

(b)

1.4

1.0
0 ! 2 3 4 5 6 7 8 9 10
~4/U012 elO -3

Fig. 2. Turbulence intensity: (a) longitudinal component; (b) vertical component.

 107

with a coarse grid of 64 X 16 X 16 uniform cells. The channel expansion ratio

was 1:2 and the spatial resolution of the 16 × 2 X 2 computational domain (in
terms of h) was A x X A y X A z = I / 4 × I / 8 × I / 8 . As a consequence of the insuf-
ficient resolution, a large portion of the turbulence energy spectrum was fil-
tered out which could not adequately be compensated for by the SGS model.
In particular, the vertical velocity fluctuations turned out to be too low, which
reduced the spreading rate of the free-shear layer and thus increased the reat-
tachment length. In comparison with the measurements of Tropea [12] and
Durst and Schmitt [13], the computed reattachment length Xr, proved to be
30% too high in this case. The Reynolds number Reo = 1.65 X l0 s of the simu-
lation compared well with the experimental values Reo= 1.1X 104 (Tropea)
and Reo= 1.13 × 105 (Durst and Schmitt). Here the Reynolds number Reo is
based on the centre-line velocity of the incoming channel flow, Uo, and the step
height h.
The present LES is performed in a 16 × 4 × 2 computational domain at the
same Reynolds number Re,= 6480 (Reo= 1.65 × 105) as before. The 128 X 32
X 32 mesh system provides a spatial resolution of A x × A y x A z = I / S x 1 / 8
× 1/16. The simulation has been advanced over more than 50 000 times steps
(50 problem times h/u,) in order to obtain the data presented below. Statis-
tical mean values were computed by averaging in the lateral direction (32 sam-
ples) and in time (50 000 samples).
We first discuss flow quantities in the inflow plane which result from an LES
of fully developed channel flow. Their behaviour undoubtedly has an influence
on the flow development downstream of the separation line. Recently, Isomoto

2.0
- Calcs.
o Schrnitt
Tropea
1.8

X=-0.25

1.6

1.4

2:
y-O
1.2

t.0
-1 . s -1 .o -o.s o.o o.s 1.0 1.5
-UW/U012 *~0 -3

Fig. 3. Reynoldsand total shear stress.

108

and Honami [14] have demonstrated that the maximum longitudinal turbu-
lence intensity of a boundary layer separating from a backward-facing step has
a strong impact on the reattachment length. It is remarkable that a 20% in-
crease in turbulence intensity can lead to a reduction in the reattachment length
of up to 24%.
In Figs. 1-3 we compare our inflow data with measurements of refs. 12 and

2oI o
Calcs.
SchmitL
A Tropea
1 5
X= 2.0
(a)

t.O

~o~

0.0
-0.4, -0.2 0.0 0.2 0,4. 0.6 O.B 1.0 1.2 1.4
U/U01

2.0
- Calcs.
o Schmitt
Tropea
t.5 X= 6.0
(b)
1.0

:Z:
~o~

0.0
-'o. 4
/
-0.2 0.0 0.2 0.4 0,6 0.8 ~.0 1.2 1.4
UIUOI

Fig. 4 ( c o n t i n u e d ) .
 109

2.0
- Calcs.
e Schrnitt
A Tropea
1.5 X= 8.0
(c)
1.0

I
~0.5

0.0
-0.4 -0.2 0.0 0.2 0.4 0.6 0.8 I .0 1.2 1.4
U/U01

2.0

- Calcs.
e Schmitt
Tropea
1.5 ¸
X=10.0
(d)

1.0

-t-

0,0
-0.4 -0.2 0.0 0.2 0.4 0.6 O.e 1.0 1.2 1.4
U/U01

Fig. 4. Profiles of the mean longitudinal velocity.

13, which were taken slightly upstream of the separation line. Obviously, there
are differences between the measured and the computed inflow data. They are
of a minor nature for the mean longitudinal velocity (Fig. 1). In fact, the dif-
ferences between the two experiments are as "large" as those between experi-
ment and computation. There are, however, greater discrepancies in the r.m.s.
110

velocity fluctuations (Fig. 2). The longitudinal component of the experiment

of ref. 13, for example, has a value of only Urms/(Uo} =0.01 at the channel
centre-line, which indicates that the flow was not fully developed before it
started separating. In contrast, the experiment of ref. 12 and the present sim-
ulation give values for UrmJ( Uo } of 0.044 and 0.04 respectively. Similar dif-
ferences are present in the vertical turbulence intensity (Fig. 2 (b)). Where

2.0

~1 - Calcs.
~_~ o Schmitt
A Tropea
1.S x=_3.0

1.0

"r-

~o~

0.0
-10 -8 -fi -4 -2 2 4.
W/UOI .iO -2

2.0

- Calcs.
o Schmitt
Tropea
I.S
X= 6.0
(b)
1.0

l-

0.0
-I0 -8 -E -4 -2
W/U01 " 1 0 -2

Fig. 5 (continued).
 111

2.0
- Calcs.
e Schmitt
Tropea
I,S X= 8.0

(c)
i.O

~o.~

0.0
-I0 -6 -4 -2 2 4
W / U O I .,iO-2

2.0
- Calcs.
o Schmitt
Tropea
|.5
X=IO.O
!
(d)
1.0

"r
~os

0.0
-tO -B -6 -+ -2 2 +
W/UOI "lO -2

Fig. 5. Profiles of the mean vertical velocity.

this component is concerned, the data of ref. 12 must be handled with caution.
Finally, the Reynolds stress of ref. 13 in Fig. 3 also clearly demonstrates that
the incoming channel flow was not fully developed in the experiment. The total
stress in the simulation, on the other hand, shows the equilibrium shape typical
for fully developed channel flow.
112

In general, the reattachment length Xr, is considered to be the most signifi-

cant parameter characterizing the back-step flow. It is frequently used as a
criterion to judge the quality of a computation. The value of Xr can be obtained
either from the mean dividing streamline < ~/> = 0, the contour line < u > -- 0 or
the mean wall shear stress distribution. We have computed X~ as that position
where < Tw> = 0 and found Xr = 7.0, non-dimensionalized with the step height

2.0

- Calcs.
o Schmitt
Tropea
1.5
X= 2.0
(a)
I.O

0 . 0 • •

I 2 3 6 5 6
UU/U012 *10 -2

2.0

- Calcs.
o Schmitt
" Tropea
1.5
×=__Ao
(b)
1.0

2=
i~0,5

0.0
I 2 3 4 5 B
UU/'U012 * I 0 -2

Fig. 6 (continued).
 113

2.0
- Calcs.
e Schmitt
A Tropea
t.5 X= 8.0
(c)
1.0

:2E
r,,~ 0 . 5

0.0
2 3 4
U U / U O 1 2 " 1 0 .2

2.0

- Calcs.
o Schmitt
Tropea
1.5
X=IO.O
(d)
1.0

!
"l-

t'-d

0.0 ,* k t2 3 4, 5 6
UU/U012 " 1 0 -2

Fig. 6. Profiles of the longitudinal turbulence intensity.

h. Tropea [ 12 ] has used the point of intersection of the < u > -- 0 line with the
lower wall to define the reattachment length and obtained the value Xr = 8.6
for a Reynolds number Reo = 1.1 X 10 4, whereas Durst and Schmitt [ 13] found
a value of Xr = 8.5 for Reo -- 1.13 X 10 ~. Considering the above discussion on the
flow state upstream of the step [13], we feel that an 18% difference in the
114

reattachment length between the experiments and our simulation is an en-

couraging result. We expect that further local or global grid refinement will
yield further improvements.
In order to test the simulation further, we have compared profiles of statis-
tical quantities with those of the two experiments [12, 13] at four positions
X = 2.0, 6.0, 8.0 and 10.0. In view of the differences among the experimental

2.0
- Calcs.
o Schmitt
A Tropea
1.5
X= 2.0

(a)
1.0

~ 0.$

0.0
0.0 0.5 1.0 I.S 2.0 2.5 3.0
WW/U012 I'1 0-2

2.0

- Calcs.
o Schmitt
Trolea
1.5
X= 6.0

1.0

"t-
~o~

0.0
0.0 0".5 1.0 1.5 2.0 2.5 3.0
W~/IUOI2 "10 -2

Fig. 7 (continued).
 115

2.0 ~ - Cslcs.
%.~ o Schmitt
~ ~ ~ Tropes
,.s ~ X= 8.0
(c)
1.0

~0.5

0.0 ~ • .
o.o o.s ,.o 1.s 2.0 2.s 3.0
WW/UOI2 ,I0 -2

2.0
Calcs.
Schmitt
Tropes
1.5 e=:
X=10.0
(d)
?.0

-r"
~o.~

0.C
0.0 0.5 1.0 1.5 2.0 2,S 3,0
WW/U012 ,I0 -2

Fig. 7. Profiles of the vertical turbulence intensity.

data, one may say that the LES does a fair job in predicting the overall behav-
iour of the mean flow (Figs. 4 and 5 ). The positions of the peak values for the
longitudinal and vertical velocities generally agree well with the experimental
results.
While there is some tendency to overpredict the maximum u-turbulence in-
116

tensities (Fig. 6 ), the opposite happens with respect to the vertical component
in the figures below. Disregarding the remarkable scatter in the experimental
data, we feel that the computed vertical turbulence intensities in Fig. 7 are
generally too low. This behaviour is a result of the intense filtering through the
grid used in the simulation, particularly near the separation point. Again, grid
refinement in the x and y directions would certainly improve the result.

2.0

- Calcs.
o Schmitt
Tropea
1.5
X = 2.0

(a)
t.O

-r-
t~4~0.5

0.0
-0.5 0.0 0.5 1.0 I.$ 2.0
-UW/UOI2 *I0 -2

2.0

- Calcs.
o Schmitt
~ Tropea
1.5
X= 6.0

1.0

-~-

~o~

f
o.0
-0.5 0.0 0.5 1.0 1.5 2,0
-UW/U012 "10 -2

Fig. 8 (continued).
 117

2.01

Calcs.
(9 Schmitt
Tropea
~.S
X= 8.0

(c)
~.0

-r-

o.0 r,r . . . .
-0.5 0.0 O.S 1,0 1.5 2.0
~UW/U012 . I 0 -2

2.0

- Calcs.
o Schmitt
Tropea
1.5
X=10.0

(d)
1.o

-r"

0.0 . . . .
-0.5 0.0 0.5 1.0 1.5 2.0
-UW/UO12 .10 "2

Fig. 8. Profiles of the Reynolds and total shear stress.

The comparison of the measured Reynolds stress in Fig. 8 with the computed
total shear stress (the viscous stress contributes only in the near-wall region)
leads to similar conclusions. The positions of the maxima are well predicted.
The values of the maxima reflect the trends of the r.m.s, velocity fluctuations
in Figs. 6 and 7.
118

We now focus on the instantaneous flow structure in terms of snapshots of

variables in different planes aligned with or perpendicular to the mean flow.
Figure 9 shows contour plots of total values of the u-, v- and w-velocity com-
ponents and of the squared fluctuating deformation tensor, which is a measure
of the dissipation rate of grid-scale-sized turbulence structures. All velocities

<_×-J+==+
co,(~) U

2 4 6 8 10 12 1¢

(b) V

. . . . . ,° ~ ~, ., ,..+ ......... ,,+-,,+ ..... , ........... •. . . . . . . . !~i- ~ ....

. . . .
x .~~Y,~~ ' ":--';"-'~:~:~5~!+'~ :+~-~i'~"~+
- ...+: + *i-: ""
+ ..,,::. . .::===:.-:.:
. . . . . ~. . . . ~_C~.~.:.%:,:.-;.-=,.,;_,.,
~.~
. ,,+:~---(

2 4 6 8 10 12 1+

(c) W

2 4 6 8 10 12 14

- - I I 9
(a) ( D i j ) -

I
r--..t

2 ¢ 6 8 10 12 14
X-AXIS

Fig. 9. C o n t o u r p l o t s o f t h e i n s t a n t a n e o u s v e l o c i t y c o m p o n e n t s a n d s q u a r e d f l u c t u a t i n g d e f o r -
m a t i o n t e n s o r : ( a ) u, m i n = - 11.7, A = 2.16, m a x = 29.3; ( b ) v, m i n - - - 10.1, ~ = 1.18, m a x - - 12.3;
( c ) w, rain = - 12.2, A = 0.96, m a x = 6.16; ( d ) (Do/j" )2, m i n = 0.24, A = 0.37, m a x = 7.33.
 119

are non-dimensionalized with the friction velocity at the inflow plane. Dashed
lines correspond to negative values for the quantities concerned.
In Fig. 9 the Y = 1.81 plane is shown. The fluid enters the domain at X = 0
through the opening 1 ~<Z~<2 and leaves it at X = 16. The plots indicate that
the mixing layer broadens downstream and reaches the lower wall near X = 7.
On the average, it is significantly inclined to the X axis upstream of the reat-
tachment line. It separates the highly unsteady separation bubble from the
incoming jet-like channel flow, which itself rather quickly loses its inflow char-
acteristics. The highest values of the velocity gradients occur in the mixing
layer. The contours of the spanwise velocity show large-scale structures typi-
cally inclined towards the X axis in the mixing layer but strongly distorted and
tilted in the reattachment region. Note that the contours of the spanwise ve-
locity correspond to the contours of the fluctuating velocity since there is no
mean flow in the Y direction. While the vertical velocity has high positive and
negative values in the free-shear layer, it is rapidly attenuated in the region
where the latter strikes the wall. It is the normal velocity constraint at the wall
which so effectively distorts the large eddies by an essentially inviscid mech-
anism. The contours of the squared fluctuating deformation tensor in Fig. 9 (d)
give a good impression of the instantaneous position and spreading rate of the
mixing layer and also show an increase in the turbulent dissipation rate down-
stream of the reattachment line observed by Chandrsuda and Bradshaw [15].
The unsteadiness of the flow is well reflected in a series of three successive
snapshots of the contour-plotted longitudinal velocity shown in Fig. 10. Above
the mixing layer, where the flow speeds are highest, large-scale turbulence
structures travel distances between 1 and 2.5 step heights in a period of 100
time steps or 0.1 problem times. The mixing layer seems to fluctuate up and
down and to splash fluid towards the lower wall near the mean reattachment
line X = 7 (Fig. 10(c)). Recall that this line is statistically stationary and
straight.
Instantaneous pictures of the line which separates locally downstream-mov-
ing fluid from upstream-moving fluid (the instantaneous reattachment line)
are given in Fig. 11. These consist of contour plots of the longitudinal velocity
in a plane very close to the lower wall (Z=0.0625). Figure 11 shows that there
are patches of fluid within the recirculation zone with positive velocities (solid
lines) which grow or decay in time. Within the time interval presented
(At= 0 . 2 h / u J , the instantaneous reattachment line moves slowly upstream
before changing its direction (not shown here). One may think of an irregular
alternate upstream and downstream motion at reattachment, or likewise, of a
splitting [ 16 ] and amalgamation of turbulent structures.
In contrast with the streamwise velocity, the v and w components are orga-
nized in much smaller flow structures within the reattachment zone. This is
clearly seen in the contour plots in a Z=0.0625 plane in Fig. 12. While the
variations in the u- and v-velocity components (both parallel to the lower wall)
120

(,J")¢'41
,
. . . .
0 2 4 6 8 10 12 14 16

0 2 4 6 8 I0 12 14 15

(b)

0 2 4 6 8 10 12 14 16
(c) X-AXIS

Fig. 10. Time development of the u-velocity component in a Y=1.8 plane: (a) t=to+0.6,
n=no+600; (b) t=to+0.7, n=no+700; (c) t=to+0.8, n=no+800 (n is the number of time
steps ).

are practically identical (compare the extremes in Figs. 11 (c) and 12 (a) ), the
vertical velocities are much smaller in the reattachment zone. This underlines
what Thomas and Hancock [17] reported, namely that the effect of the wall
constraint on the tangential components of turbulent motion is small com-
pared with the constraint on the normal component.
The next two figures represent the turbulent field in cross-sections perpen-
dicular to the main flow at positions X = 1 and X = 4. Total values of u-velocity,
the v - w vector field and the longitudinal vorticity component ~2x are used for
illustration. At X = 1 the mixing layer is clearly distinguishable in the u-veloc-
ity field {Fig. 13 (a)) as a narrow band of nearly parallel lines. Its width covers
roughly 20% of the step height and it is only weakly disturbed. The vector field
(Fig. 13 (b) ) is unremarkable and one would hardly recognize the mixing layer
by seeing this plot alone. The t~x-vorticity field (Fig. 13 ( c ) ), on the other hand,
 121

U
,:Y ,-~ :::,~--,:;,::,-,";'..
?,':,;::,;--:::,~i~ "',,L~~ J - ~-~.~-~--~

K~

0 2 4 8 O 10 )2 t4 16

:."_ "" "', ,1" ;: ,-:' ~=. - :" '-2 s.' 5>. ~ o<

0 2 4 6 8 10 12 t4 t6

2 4 6 O 10 12 14 16
X-AXIS
(c)
Fig. 11. T i m e d e v e l o p m e n t o f t h e u - v e l o c i t y c o m p o n e n t in a Z = 0.0625 plane, close to t h e lower
wall: (a) t=to, n=no; (b) t=to+O.1, n=no+lO0; (c) t=to+0.2, n = n o + 2 0 0 . E x t r e m e s for (c):
rain = - 13.8, A = 3.78, m a x - 20.2.

is much more organized, especially in the free-shear layer and the upper wall
layer where a pair of strong counter-rotating vortices (in fact, the strongest in
this plane) are observed.
In Fig. 14 the development of the mixing layer can be observed. The shear
layer is now strongly distorted by downward motions and in one case even
122

1 2 3 4 $ 6 7 8 9 I0 II 12 i~ ~ [5

(1)) l i

2 3 4 5 8 7 O 9 I0 II 12 I] 14 15
X-AXIS

Fig. 12. Contour plots of the instantaneous v- and w-velocity components in a Z=0.0625 plane:
(a) v, m i n = -14.3, A=2.89, max---- 11.7; (b) w, m i n = -8.42, A--1.41, max=4.31.

reaches the lower wall (near Y = 1 in Fig. 14(a) ). As a result, fluid is entrained
into the recirculation region and local patches of fluid with positive streamwise
momentum occur near the lower wall for a few instants (compare Fig. 11 ). The
velocity vectors in Fig. 14 (b) reveal that the intrusion of positive u-momentum
occurs at speeds of up to 8u~. These vectors also indicate swirling motions
which are confirmed by the ~x-vorticity contours (Fig. 14 (c)). In this figure a
pair of counter-rotating longitudinal vortices lie in the mixing layer itself and
their strength exceeds that of the near-wall vortices in Fig. 13 (c) by 50%.
We complete our discussion of the instantaneous behaviour of turbulent
backward-facing step flow by presenting contour surface plots of high intensity
contributions to the Reynolds shear stress ( - u" w" ). Figure 15 presents side,
top and perspective views of - u " w " contours which may be considered as
coherent turbulence-energy-producingstructures. Clearly, most of the turbu-
lence energy is created within the mixing layer upstream and shortly down-
stream of the mean reattachment line. Further evaluation of the simulation
data is necessary in order to find out which extra strain rates are the cause of
this behaviour. The figures demonstrate the rapid decay of - u" w" structures
 123

or)

X<::

r -I 4

.."-"-..,,):::- ....

0,5 1,0 1.5 2.0 2.5 3.0 3.5

(b) V - W
I z . . . . . . + . , ~ . . . .

I i ; i ~ ~ • - • . \ I I z . . . . . . / z ~ . . . . . . . .

I -/ - ' ~ , , ~ - z + . . . . . . . / ~ t . . . . . . . .

t!'ii , i;i!!i~iiiiiii!!ii~ii~Ti
.............. ' .........
co t.-~,l{ . . . . . . . . . . . . . . . . . ) \ . , ~ ,

0.5 1.0 1.5 2.0 2.5 3.0 3.5

o
(c) ~,.

co

i
r~l

0.5 1.0 I .S 2.0 2.5 3.0 3.5

Y-AXIS

Fig. 13. Instantaneous total values of u-velocity contours, v - w vectors and $2~ contours in an X = 1
plane: (a) u, m i n = - 8 . 7 2 , A=2.02, max=29.7; (b) v - w vector plot; (c) ~2x, rain= - 100, A = 11.3,
m a x - 103.
124

(a) U

i! . . . .

0.5 1,0 1.5 2.0 2.5 3.0 3,5

(c) f~x
'E.

~s ('11# --. i _ I /I 11%1.1 "----e ~ ~ ..... i----~-

O'3
<

0.5 1,0 1.5 2.0 2.5 3.0 3.5

Y-AXIS

Fig. 14. I n s t a n t a n e o u s t o t a l v a l u e s o f u - v e l o c i t y c o n t o u r s , v - w v e c t o r s a n d / 2 x c o n t o u r s i n a n X = 4
p l a n e : ( a ) u, m i n = - 12.2, A = 2.13, m a x = 28.2; ( b ) v - w v e c t o r plot; ( c ) ~2x, m i n = - 150, A = 14.7,
m a x = 116.
 125

fl .. , i ~+~ . . . . . "

1
I
< ~ " i) •

2 4 6 6 I0 Y 12

(l~)

~>~

'z'~ ~3~,,_.~ ~ ~ _ _ <.~

- ~ ~ ~ ~ .

o 2 4 5 6 10 X ~2

((:)

X " ~o X-J
12 0 x(

Fig. 15. Instantaneous stress-producing (-u"w") structures: (a) side view; (b)top view; (c)
perspective view.

after reattachment of the mixing layer and once more underline the impor-
tance of the flow behaviour within the reattachment zone. The top view in Fig.
15(b) gives an impression of the spottiness of such structures, which is in
accordance with the organization of the w" field (Fig. 12 (b) ). The destruction
126

of these structures is most probably due to the wall constraint on the w com-
ponent fluctuations.
Figure 16 shows the contour surface of the zero instantaneous longitudinal
velocity (constant-value surface: u = - 0 . 0 5 ). The surface separates the back-
flow region below it from the downstream flow above it. The irregular shape of
the surface reflects the dynamics of the vertical velocity fluctuations and in-
dicates the meandering of the instantaneous reattachment line.
Lastly, we return to two statistical results which demonstrate the capabili-
ties of the L E S technique for investigating recirculating flows. Figure 17 de-
picts the distribution of the total shear stress, which coincides with the Rey-
nolds stress (except close to the walls). It shows a maximum within the mixing
layer which is 13.6 times that of the incoming flow and a steep decrease towards
the reattachment zone. The r.m.s, pressure fluctuations in Fig. 18 show a sim-

J
J

k) 2 4 5 g ~0 12 0
'X

Fig. 16. Contour surface of the instantaneous longitudinal velocity, separating backward and for-
ward flow regions (surface value: u = - 0 . 0 5 ).

0 2 4 5 8 10 ~2 ~4 X f6

Fig. 17. Contour plot of the mean total shear stress.

0 2 4 6 8 10 12 14 X
Fig. 18. Contour plot of the r.m.s, pressure fluctuations.
 127

ilar behaviour where the position of local maxima is concerned. Again we find
a maximum in the shear layer (roughly 12 times higher than the wall r.m.s.
value of the incoming channel flow) and a decrease towards X~ = 7. Along the
bottom wall, however, the values of the pressure fluctuations reach a maximum
close to the reattachment point. This is again a consequence of the inhibition
of the vertical velocity fluctuations in the reattachment zone (compare ref.
15).

5. C o n c l u s i o n s

The results of a large-eddy simulation of the high Reynolds number back-

ward-facing step flow are reported. The overall agreement of statistical quan-
tities with experimental data was good for the two cases considered. The cal-
culated reattachment length was 18% below the values found in experiments
[12, 13]. In recent calculations which included the entry region in the calcu-
lation domain, we found a reattachment length X,= 7.3. The increase in the
mean reattachment length is encouraging and shows that the recirculating flow
has a weak influence on the incoming channel flow close to the separation line.
We feel that the reported large-eddy simulation produces data which were
close to the accuracy required by engineers at reasonable computer costs. The
simulation, with a grid resolution of 1/8 X 1/8 X 1/16 in terms of step height in
the (x, y, z) directions, requires on the order of 25 h of computer time on a
Cray X-MP to perform a complete simulation including all correlation evalu-
ations. This time increases by about 50% if an inlet portion of length 4h is
added. The increase is in large part due to the more complicated Poisson solver
which is required in this case.
The high degree of insight into the instantaneous flow behaviour that large-
eddy simulation permits, along with the possibility to compute statistical vari-
ables involving quantities which are difficult to measure (e.g. pressure fluctua-
tions ), justify the increased expenditure for LES.

Acknowledgement

The authors are grateful to the German Research Society (DFG) for their
financial support of the study.

References

1 S.J. Kline, B.J. Cantwell and G.M. Lilley (eds.), Proc. 1980-81 AFOSR-HTTM-Stanford
Conf. on Complex Turbulent Flows, Stanford University, Stanford, CA, 1982, Vols. I-III.
2 M. Peric, M. R~iger and G. Scheuerer, A finite volume multigrid method for calculating tur-
bulent flows, Proc. 7th Symp. on Turbulent Shear Flows, Stanford University, August 21-
23, 1989, Stanford University, Stanford, CA, 1989, Vol. 1, pp. 7.3.1-7.3.6.
128

3 U. Schumann, Subgrid scale model for finite difference simulations of turbulent flows in
plane channels and annuli, J. Comput. Phys., 18 (1975) 376-404.
4 R. Friedrich, On large-eddy simulation of turbulent flows, in H. Niki and M. Kawahara (eds.),
Computational Methods in Flow Analysis, Okayama University of Science, Okayama, 1988,
Vol. 2, pp. 833-843.
5 L. Schmitt and R. Friedrich, Application of the large-eddy simulation technique to turbulent
backward facing step flow, Proc. 6th Symp. on Turbulent Shear Flows, Toulouse, September
7-9, 1987, Paul Sabatier University, Toulouse, 1989, pp. 19.2.1-19.2.7.
6 U. Schumann and R.E. Sweet, A direct method for the solution of Poisson's equation with
Neumann boundary conditions on a staggered grid with arbitrary size, J. Comput. Phys., 20
(1976) 171-182.
7 B.L. Buzbee, F.W. Dorr, J.A. George and G.H. Golub, SIAM J. Numer. Anal., 8 (1971) 722-
736.
8 W.M. Pun and D.B. Spalding, A general computer program for two-dimensional elliptic flow,
HTS/76/2, Imperial College Mechanical Engineering Department, 1976.
9 K. Richter, R. Friedrich and L. Schmitt, Large eddy simulation of turbulent wall boundary
layers with pressure gradient, Proc. 6th Symp. on Turbulent Shear Flows, Toulouse, Septem-
ber 7-9, 1989, pp. 22.3.1-22.3.7.
10 L. Schmitt and R. Friedrich, Large-eddy simulation of turbulent backward facing step flow,
Notes on Numerical Fluid Mechanics, Vieweg, Braunschweig, 1988, Vol. 20, pp. 355-362.
ll U. Piomelli, J. Ferziger, P. Moin and J. Kim, New approximate boundary conditions for
large-eddy simulations of wall-bounded flows, Phys. Fluids A, 1 (1989) 1061-1068.
12 C. Tropea, Die turbulente Str(imung in Flachkan~ilen und offenen Gerinnen, Dissertation,
University of Karlsruhe, 1982.
13 F. Durst and F. Schmitt, Experimental study of high Reynolds number backward facing step
flow, Proc. 5th Symp. on Turbulent Shear Flows, Cornell University, Ithaca, NY, August 7-
9, 1985, Cornell University, Ithaca, NY, 1985.
14 K. Isomoto and S. Honami, The effect of inlet turbulence intensity on the reattachment
process over a backward-facing step, J. Fluids Eng., 111 (1989) 87-92.
15 C. Chandrsuda and P. Bradshaw, Turbulence structure of a reattaching mixing layer, J. Fluid
Mech., 110 (1981) 171-194.
16 P. Bradshaw and F.Y.F. Wong, The reattachment and relaxation of a turbulent shear layer,
J. Fluid Mech., 52 (1972) 113-135.
17 N.H. Thomas and P.E. Hancock, Grid turbulence near a moving wall, J. Fluid Mech., 82
(1977) 481-496.