DNS and LES of scalar transport in a turbulent

plane channel flow at low Reynolds number

Jordan A. Denev1 , Jochen Fröhlich1 , Henning Bockhorn1 ,

Florian Schwertfirm2 , Michael Manhart2
1
Institute for Technical Chemistry and Polymer Chemistry, University of Karlsruhe,
Kaiserstraße 12, D-76128 Karlsruhe, Germany,
[email protected], http://www.ict.uni-karlsruhe.de
2
Department of Hydromechanics, Technical University of Munich,
Arcisstraße 21, 80333 Munich, Germany

Abstract. The paper reports on DNS and LES of plane channel flow at
Reτ = 180 and compares these to a DNS with a higher order convection
scheme. For LES different subgrid-scale models like the Smagorinsky, the
Dynamic Smagorinsky and the Dynamic Mixed Model were used with
the grid being locally refined in the near-wall region. The mixing of a
passive scalar has been simulated with two convection schemes, central
differencing and HLP A. The latter exhibits numerical diffusion and the
results with the central scheme are clearly superior. LES with this scheme
reproduced the budget of the scalar variance equation reasonably well.

1 Introduction
Turbulent mixing of scalar quantities is a phenomenon observed in environmental
flows as well as in abundant engineering applications of chemical, nuclear power,
pharmaceutical or food industries. Their simulation requires reliable models for
turbulent mixing processes. After discretization, however, the physical and the
numerical model interact in a complex way, which is not fully understood so far.
To address these issues, this paper presents results from both, Direct Numerical
Simulations (DNS) and Large Eddy Simulations (LES) of fully developed plane
channel flow at a friction Reynolds number of Reτ = 180. This is a prototypical
flow frequently used to study physical and numerical modelling of wall-bounded
flows. The first DNS of this configuration was performed by Kim et al. [4].
In an earlier paper by the present authors [2] the impact of local grid re-
finement near the walls on the LES modelling of the flow field was investigated.
In the present paper we extend this approach and focus on the modelling of a
transported scalar.

2 Numerical methods and simulation details

The turbulent channel flow between two parallel plates is simulated for a nominal
friction Reynolds number Reτ = 180, defined by the friction velocity Uτ and the
channel half-width h. In the computation, the corresponding bulk Reynolds num-
ber Reb = 2817 was imposed by instantaneously adjusting a spatially constant
volume force in each time step, so that in fact Uτ and hence Reτ is a result of the
simulation. The computational domain in streamwise, wall-normal and spanwise
direction extends over Lx = 6.4h, Ly = 2.0h and Lz = 3.2h, respectively. Pe-
riodic boundary conditions were imposed for the streamwise and the spanwise
direction. At the walls, a no-slip condition was applied for both LES and DNS,
together with Van Driest damping for the Smagorinsky model. Dirichlet bound-
ary conditions at the walls were imposed for the scalar, i.e. C(x, 0, z) = 1.0 and
C(x, 2h, z) = −1.0.
Two different numerical codes have been applied for the present work. Both
utilize a finite volume method for incompressible fluid on block-structured grids
together with a Runge-Kutta (RK) scheme in time. A Poisson equation is solved
for the pressure-correction. The code M GLET has been developed at the TU
Munich and uses staggered Cartesian grids with a 6th order central discretization
scheme (CDS) in space and a 3rd order RK scheme [5, 8]. The code LESOCC2,
developed at the University of Karlsruhe, can handle curvilinear collocated grids
and employs discretizations of second order in space and time [3]. In the present
study, DNS results from this code were first compared to those obtained with the
higher-order discretization of M GLET . Subsequently, LES were carried out with
different subgrid-scale (SGS) models as described, e.g., in [6]. These comprise the
Smagorinsky Model (SM) with constant CS = 0.1, the Dynamic Smagorinsky
Model (DSM), and the Dynamic Mixed Model (DMM), see Table 1. The first
case in Table 1, denoted DN S − 6O (sixth order discretization in space for both
convection and diffusion), has been calculated with M GLET , all others with
LESOCC2. The LES filter is not accounted for in the present notation.
An equation governing the transport of a scalar quantity with Schmidt num-
ber Sc = 1 is also considered in the present study. The concentration C, is
regarded as passive, i.e. it does not influence the fluid flow. The unresolved tur-
bulent transport of C is modelled by an eddy diffusivity Dt = νt /Sct , where νt
is the SGS eddy viscosity and Sct the turbulent Schmidt number, here set equal
to 0.6. This also holds for the mixed model.
With the present equations and Dirichlet boundary conditions the concentra-
tion fulfils a maximum condition. The extrema are attained on the boundaries,
so that the scalar is restricted to the interval C ∈ [−1; 1] for physical reasons
(the lower bound -1 was chosen here instead of 0 for technical reasons). This
boundedness does not neccessarily carry over to the discretized solution. In nu-
merous studies bounded convection schemes are therefore applied to guarantee
the boundedness of the numerical solution which is not guaranteed with a central
scheme. In [3] and related work the HLP A scheme developed in [9] was used
for this purpose. HLP A determines the convective flux by a blending between
second-order upwinding and first-order upwinding as

U Ci + U (Ci+1 −Ci )Θi+ 21 , 0 < Θi+ 21 < 1 Ci − Ci−1
Fi+ 21 = , Θi+ 21 = , (1)
U Ci , else Ci+1 −Ci−1
where U is the velocity at the cell boundary i + 1/2. This scheme was employed
for the present study for the convection term of the concentration equation in
cases DN S and DN S − F , while still using second order CDS in the momentum
equation (see the column for the scalar convection scheme (SCS) in Table 1).

Table 1. Overview over the runs discussed. The nomenclaure is defined in the text.

Case CVtot ∆+
x y1+ ∆+
z SGS SCS tav Uτ Cτ
DNS
DNS-6O 1,407,120 9.1 0.68 7.2 - CDS-6O 544 0.064018 -0.041478
DNS 1,407,120 9.1 0.68 7.2 - HLPA 638 0.062237 -0.039232
DNS-F 10,866,960 4.5 0.34 3.6 - HLPA 537 0.061821 -0.041106
DNS-CDS 1,407,120 9.1 0.68 7.2 - CDS 745 0.062487 -0.042083
LES
HLPA-SM 258,688 29.8 0.37 14.9 SM HLPA 615 0.067434 -0.038431
CDS-SM 258,688 29.8 0.37 14.9 SM CDS 643 0.066032 -0.047632
CDS-DSM 258,688 29.8 0.37 14.9 DSM CDS 650 0.060801 -0.042553
CDS-DMM 258,688 29.8 0.37 14.9 DMM CDS 646 0.070095 -0.043067

A special feature of the code LESOOC2 is the possibility of block-wise local

grid refinement (LGR). This allows to use a fine grid close to the walls without
excessively refining in the center of the channel, so that CPU time and storage are
not overly increased. LGR is utilized near the walls up to a distance yref = h/8
+
equal to yref = 22.5 from the wall with a refinement ratio of 2 in both, x−
and z−direction. In y−direction the grid is stretched uniformly by a factor 1.03
throughout the channel. As observed in a previous study [2] the turbulent char-
acteristics of the flow exhibit some visible changes at the block-interface, and
this issue will also be addressed later in this paper. Table 1 presents information
on the numerical grids, i.e. the total number of control volumes of the entire grid,
CVtot , and the dimensionless size of the control volumes in x− and z−direction,
respectively. In case of LGR, which is used with all LES cases, this is the unre-
fined spacing used in the core of the flow. Furthermore, y1+ indicates the distance
of the wall-adjacent point from the wall.

3 Results from DNS

Statistical data for all computations in the present work have been collected over
averaging times tav larger than 540 dimensionless time units tb = h/Ub , where
Ub is the bulk velocity of the flow. Table 1 shows the results obtained for the
friction velocity and the reference concentration defined as
r  
τw D ∂hCi
Uτ = , Cτ = , (2)
ρ Uτ ∂y y=0
respectively, with D being the laminar diffusion coefficient (the turbulent dif-
fusion coefficient vanishes at the wall). In the present section four DNS cases
are compared, which allows to identify the role of the numerical discretization
scheme and the grid resolution. Case DN S − 6O is chosen as a reference case.
In [7] these data were compared with the classical ones of [4] showing excellent
agreement. The run DN S was performed with LESOCC2 on the collocated
equivalent of this grid with the second order method. Due to the lower order
these results (not reproduced here) were unsatisfactory showing deviations of
up to 18% from the reference data. Therefore, the grid was refined by a fac-
tor of 2 in each direction (case DN S − F ) to compensate for the lower-order
discretization. The comparison with DN S − 6O is presented in Figure 1. The
turbulent stresses and the time-averaged scalar match very well. The turbulent
scalar flux hu′ c′ i and the scalar variance hc′ c′ i exhibit differences, which for the
latter mainly appear in the middle of the channel.

a) 0.035
b)
<u´u´> DNS-F 0.025
-<u´c´> DNS-F
<v´v´> DNS-F
<v´c´> DNS-F
<u´u´>, <v´v´>, <w´w´>

0.03 <w´w´> DNS-F

-<u´c´> DNS-60
<u´u´> DNS-60
0.02 <v´c´> DNS-60
<v´v´> DNS-60
-<u´c´>, <v´c´>

0.025 <w´w´> DNS-60

0.02 0.015

0.015
0.01
0.01
0.005
0.005

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y
c) d)
<C> DNS-F 0.025
<c´c´> DNS-F
1 <C> DNS-60 <c´c´> DNS-60

0.02
0.8

0.015
<c´c´>
<C>

0.6

0.4 0.01

0.2 0.005

0
0 0.2 0.4 0.6 0.8 1 0
0 0.2 0.4 0.6 0.8 1
y y
Fig. 1. Comparison of the results from cases DN S − 6O and DN S − F : a) normal
turbulent stresses hu′ u′ i, hv ′ v ′ i and hw′ w′ i; b) turbulent scalar fluxes hu′ c′ i and hv ′ c′ i;
c) mean scalar hCi; d) scalar variance hc′ c′ i.
 In order to further elucidate the role of the numerical scheme, case DN S
has been repeated employing the CDS scheme of second order instead of the
HLP A scheme. These results (not depicted here for lack of space) show a clear
improvement for the scalar variance hc′ c′ i compared to case DN S, and also when
compared to case DN S − F , as the difference with respect to the reference data
in the middle of the channel decreases. The scalar flux hv ′ c′ i for DN S − CDS
matches perfectly well with DN S − 6O and the agreement of hu′ c′ i is practically
as good as for DN S − F in Figure 1. It should also be noticed that the value
for the reference scalar Cτ for this case is closer to the value of DN S − 6O than
that of DN S. This considerable improvement from case DN S to DN S − CDS
shows that for the flow considered the central differencing scheme appears clearly
superior compared to the HLP A scheme.

4 Results from LES

The results of the previous section were obtained with DNS, i.e. on fine grids
and without any turbulence model. Now we turn to LES for which numerical
and modelling errors interact in a complex way. The grid used for these LES is
much coarser in the core region of the flow (see ∆x+ and ∆z + in Table 1), while
in the vicinity of the wall (y < 1/8h) it is of similar cell size as in the DNS cases
(195,000 control volumes in the region of refinement).
The results obtained with the CDS for the convective terms of the scalar
transport equation confirm the findings of a previous paper by the authors [2],
in which a higher Reynolds number was considered, and where the Smagorinsky
model performed better than the other two models. In the present investigation,
CDS − DM M shows slightly better results than CDS − DSM . This assertion
is mainly based on the behaviour of the averaged scalar and the scalar variance
near the wall. CDS − DSM on the other hand shows the most accurate LES
value for Cτ .
To address the impact of the convection scheme, results for the cases CDS −
SM and HLP A− SM are presented together with the reference case DN S − 6O
in Figure 2. These results again show the superiority of the CDS, which is
more pronounced in the proximity of the wall. The results also demonstrate
the diffusive characteristics of the HLP A scheme. The presence of additional
diffusion is noticed in the averaged scalar distribution by an increased value and
an almost linear distribution for the region away from the wall. Furthermore,
the turbulence quantities such as the scalar fluxes and the scalar variance are
underestimated near the wall, i.e. damped by the numerical diffusion.

5 Transport equation of the scalar variance

Finally, an evaluation of the terms in the budget of the scalar variance was
carried out. In the case of the Smagorinsky model the equation for the resolved
 a) b)
0.035 <u´u´> CDS-SM
0.025 -<u´c´> HLPA-SM
<v´v´> CDS-SM <v´c´> HLPA-SM
<w´w´> CDS-SM -<u´c´> CDS-SM
<u´u´>, <v´v´>, <w´w´> 0.03 <u´u´> DNS-60
<v´v´> DNS-60
<v´c´> CDS-SM
-<u´c´> DNS-60
<w´w´> DNS-60 0.02
<v´c´> DNS-60
0.025

-<u´c´>, <v´c´>
0.02 0.015

0.015
0.01
0.01

0.005
0.005

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y
c) d)
1 0.025
<C> HLPA-SM
<C> CDS-SM
<C> DNS-60

0.8 0.02

0.6 0.015
<c´c´>
<C>

0.4 0.01

0.005 <c´c´> HLPA-SM

0.2 <c´c´> DNS-60
<c´c´> CDS-SM

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y

Fig. 2. Computations performed with LES and different numerical schemes for the
scalar: cases CDS−SM and HLP A−SM compared with the reference case DN S−6O.
a),b),c) and d) as in the previous figure.

scalar variance reads

*  ′ 2 +  
∂hCi ∂c ′ ′∂ ∂hc′ i2 ∂
′2 ′

0 = −2hc v i − 2 (D+Dt ) + hD+Dt i − ( c vi
∂y ∂xk ∂y ∂y ∂y
| {z } | {z } | {z } | {z }
PC DC TC
EC

Here, P c denotes the production by the mean concentration gradients, Ec

the scalar dissipation, Dc the diffusion transport term (comprises molecular and,
in the case of LES, eddy-diffusion) and T c turbulent transport by the normal
velocity fluctuation. In the case of dynamic LES-models, additional terms appear
in the balance due to the fluctuation of the model parameter. These are however
negligible compared to the other terms of the equation [1]. In the case of DNS
the tubulent diffusivity is omitted.
Figure 3 shows the comparison of the different terms, constituting the budget
of the scalar variance for DN S − 6O, DN S − CDS and CDS − SM . While
for the first two cases the match is very good, the case CDS − SM shows
some differences. As expected, the magnitude of the terms in the middle of the
channel is underestimated, which is due to the fact that only part of the turbulent
spectrum is resolved with LES. The values at the wall on the other hand are
quite accurately reproduced for all terms of the above equation due to the fine
grid near the walls. For CDS − SM the Figure 3b shows some artefacts at the
block boundary separating the refined and the coarse grid. They are present only
in those terms containing a derivative of a correlation term normal to the wall,
i.e. in Ec, Dc and T c. The reason is that the abrupt changes in the subgrid-filter
size and the numerical resolution cause inevitable irregularities in the turbulent
quantities at the two sides of the block boundary which modify the derivative
operator applied normal to the wall. Apart from this, the agreement between the
CDS − SM and the DNS cases is reasonably good. It can hence be concluded
that in the present case LES (case CDS − SM ) is capable of qualitatively, and
to some extent also quantitatively, reproducing the terms in the budget for the
scalar variance.

a) b)
0.6 Pc DNS-CDS 0.6 Pc LES-CDS-SM
Tc DNS-CDS Tc LES-CDS-SM
0.5 Dc DNS-CDS 0.5 Dc LES-CDS-SM
Ec DNS-CDS Ec LES-CDS-SM
0.4 Pc DNS-60 0.4 Pc DNS-60
Tc DNS-60 Tc DNS-60
Pc, Tc, Dc, Ec

Pc, Tc, Dc, Ec

0.3 Dc DNS-60 0.3 Dc DNS-60

Ec DNS-60 Ec DNS-60
0.2 0.2
0.1 0.1
0 0
-0.1 -0.1
-0.2 -0.2
-0.3 -0.3
-0.4 -0.4
0 50 100 150 0 50 100 150
y+ y+
Fig. 3. Terms in the budget of the scalar variance hc′ c′ i. All terms are normalized by
D/(Cτ2 Uτ2 ) and explained in the text. a) Case DN S − CDS (symbols) compared to
reference case DN S − 6O (lines). b) LES with CDS − SM (symbols) compared to
DN S − 6O (lines).

6 Conclusions
Different numerical and modeling issues have been studied when calculating
fluid flow and passive scalar distribution in a plane turbulent channel flow. DNS
with CDS of sixth and second order accuracy have been compared. It has been
shown, that the second-odrer scheme achieved the desired accuracy (shown by
the sixth-order CDS) only after the numerical grid has been refined twice in
each spatial direction.
Comparison of two schemes for the scalar, unbounded CDS and non-linear,
monotonous upstream-weighted HLP A showed superiority of the CDS scheme,
while the results with HLP A were found to suffer from numerical diffusion.
This is in line with the general attitude when modelling the SGS terms in the
LES-momentum equation. Usually, a non-dissipative scheme is preferred and
dissipation entirely introduced by the laminar viscous terms and the SGS model.
Additional numerical dissipation without modifying the SGS model is avoided.
The same is observed here for the scalar transport. It should however not be
concluded that the CDS is best for any LES involving a passive scalar. In other
simulations of the present authors concerned with a jet in crossflow this scheme
led to numerical instability. More appropriate schemes to maintain boundedness
of the scalar are needed.
The present study shows that LES with tangential grid refinement near the
walls delivers reasonable accuracy at low computational costs. This conclusion
is also supported by the results obtained for the budget of the scalar variance
which is reproduced reasonably well with the present LES.

Acknowledgements: This research was funded by the German Research

foundation through priority programme SPP-1141 ”Mixing devices”. Carlos Fal-
coni helped with the preparation of figures.

References
1. I. Calmet and J. Magnaudet. Large-eddy simulation of high-schmidt number mass
transfer in a turbulent channel flow. Phys. Fluids, 9(2):438–455, 1997.
2. J. Fröhlich, J. A. Denev, C. Hinterberger, and H. Bockhorn. On the impact of
tangential grid refinement on subgrid-scale modeling in large eddy simulation. In
T. Boyanov et al., editor, Numerical methods and applications, 6th International
conference NMA 2006, LNCS 4310, pages 550–557, Borovets, Bulgaria, August 2006.
Springer-Verlag Berlin Heidelberg 2007.
3. C. Hinterberger. Dreidimensionale und tiefengemittelte Large–Eddy–Simulation von
Flachwasserströmungen. PhD thesis, Institute for Hydromechanics, University of
Karlsruhe, 2004.
4. J. Kim, P. Moin, and R. Moser. Turbulence statistics in fully developed channel
flow at low Reynolds number. J. Fluid Mech., 177:133–166, 1987.
5. M. Manhart. Large-eddy simulation of high-schmidt number mass transfer in a
turbulent channel flow. Computers and Fluids, 33(3):435–461, March.
6. P. Sagaut. Large Eddy Simulation for Incompressible Flows: An Introduction, second
ed. Springer, Berlin, Oktober 2002.
7. F. Schwertfirm and M. Manhart. ADM Modelling for Semi-Direct Numerical Simu-
lation of Turbulent Mixing and Mass Transport. In J.A.C. Humphrey, T.B. Gatski,
J.K. Eaton, R. Friedrich, N. Kasagi, and M.A. Leschziner, editors, Fourth Inter-
national Symposium. on Turbulence and Shear Flow Phenomena, volume 2, pages
823–828, Williamsburg, Virginia, 2005.
8. F. Schwertfirm and M. Manhart. DNS of passive scalar transport in turbulent
channel flow at high Schmidt numbers. In K. Hanjalic, Y. Nagano, and S. Jakrilic,
editors, Turbulence, Heat and Mass Transfer 5, pages 289–292, Dubrovnik, Coratia,
2006.
9. J. Zhu. A low-diffusive and oscillation-free convection scheme. Communications in
applied numerical methods, 7:225–232, 1991.