Professional Documents
Culture Documents
Abstract 1 Introduction
An unstructured grid algorithm for tetrahedral As the speed and memory capabilities of com-
cells has been developed for Large Eddy Simula- puters continue to increase, the prospect of per-
tion (LES). The finite volume form of the com- forming Large Eddy Simulation (LES) of engi-
pressible Navier-Stokes equations are solved for neering problems becomes more feasible. In LES,
cell-averaged conservative variables. Inviscid flux the Navier-Stokes equations are spatially filtered
computations are performed by applying a Rie- to compute the large scale flow structures, while
mann solver across each face, the values at the modeling only the small or sub-grid scale (SGS)
points on the faces being obtained by least-squares structures [4, 6]. As the small scale structures
reconstruction from the cell-averaged values. The are expected to be approximately universal in tur-
viscous fluxes and heat transfer are obtained by bulent flows, LES can potentially be applied to
application of Gauss' theorem. The numerical wider classes of problems than traditional turbu-
scheme is explicit, with second-order spatial and lence models, but at reduced computational cost
temporal accuracy. The sub-grid scale model is compared to direct numerical simulation (DNS).
of the constant coefficient Smagorinsky type, with However, whereas most of the work in the field has
van Driest damping for the viscous sublayer. been on structured meshes in the context of finite-
For channel flow, results at a Reynolds num- difference or spectral methods [7], application of
ber (based on channel height and bulk velocity) of LES to engineering problems requires methodolo-
5600 and Mach number of 0.5 are compared to in- gies suitable for complex geometries. Towards
compressible direct numerical simulations (DNS) this end, we have developed a three-dimensional
and experiments. Mean velocity and velocity fluc- Large Eddy Simulation (LES) algorithm for un-
tuations compare well with the DNS and experi- structured meshes [12, 21].
mental data. For boundary layer flow, results at a Our LES unstructured grid algorithm uses a
Reynolds number (based on the inflow boundary cell-centered storage architecture, with the cell-
layer thickness) of 20,000 and Mach 3 are com- averaged values stored at the centroid of each
pared to experimental data. Preliminary results tetrahedral cell. The inviscid fluxes are computed
for profiles of the streamwise mean velocity and by applying a Riemann solver across each face.
turbulence intensity and for the skin friction show The values of the flow variables at one or more
good agreement with experiment. points on either side of the face are obtained from
'Graduate Research Assistant, Member AIAA the cell-averaged values by function reconstruc-
okongo@jove.rutgers.edu tion. The function reconstruction is based on the
http://cronos.rutgers.edu/~okongo
'Professor, Associate Fellow AIAA the least squares method of Ollivier-Gooch [22],
knight@jove.rutgers.edu and is second-order accurate. Higher order recon-
http://www-srac.rutgers.edu/~knight struction can also be achieved using this method.
Copyright ©1998 by Nora Okong'o and Doyle D. Knight. Details of our implementation can be found in Ref.
Published by the American Institute of Aeronautics and [21]. The simulation is advanced in time using
Astronautics, Inc. with permission. second-order explicit Runge-Kutta time integra-
tion. The algorithm incorporates a constant co- features missed by their k — c model, such as recir-
efficient Smagorinsky sub-grid scale (SGS) model culation for flow in a ribbed channel. These mixed
with van Driest damping for the viscous sublayer. results illustrate the drawbacks of traditional tur-
The objective of this paper is to validate our bulence models, and hence the need for LES.
LES unstructured grid algorithm for two turbulent Several researchers have performed incompress-
shear flows. The first flow is low Mach number tur- ible LES of channel flow on structured grids us-
bulent channel flow, which is a canonical test case ing spectral methods, including Moin and Kim
for LES. We compare our results to the incom- [17], and Piomelli et al [23]. Spectral methods
pressible DNS of Kim et al [11] and experimental cannot be used in general geometries, prompting
data of Eckelmann [5] and Kreplin and Eckelmann some attempts to use finite-difference methods,
[13]. The second flow is a supersonic turbulent e.g. [1]. Most sub-grid scale (SGS) models are
boundary layer. Preliminary results are compared based on the eddy-viscosity concept. The sim-
with the conventional one-seventh power law for plest is the constant coefficient Smagorinsky sub-
the mean velocity and the experimental data of grid scale (SGS) model [26]. Although dynamic
Yanta and Crapo [29] for streamwise turbulent in- and Lagrangian methods of determining the coef-
tensity. ficient by re-filtering the flowfield have also been
used [16, 19, 30], the additional complexity ap-
pears not to justify the small improvement in the
2 Previous Work results. This is mainly because at low Reynolds
numbers most of the turbulence energy is con-
2.1 Turbulent Channel Flow tained in the resolved scales, and thus the choice
of SGS model is not critical. In our research,
Turbulent channel flow is a classic benchmark case
we opt to use the simplest (constant-coefficient)
for validating turbulent flow simulations. It has
Smagorinsky-type model.
the advantage of geometric simplicity while admit-
There are few LES computations of turbulent
ting complex three-dimensional flow structures.
channel flow using unstructured grids. Haworth
When the flow reached statistical equilibrium, the
and Jansen [8] use an unstructured hexahedral
time-averaged flow statistics are functions of the
wall-normal direction only. This flow is charac- grid finite volume method. The results compare
terized by long streamwise vortices near the walls
favorably with the DNS of Kim et al [11] for the
[5, 13, 17, 23]. incompressible flow at Reynolds number of 5600.
The work of Haworth and Jansen is similar to ours
Data for fully-developed incompressible channel
in that we both use second-order algorithms. How-
flow at a Reynolds number of 5600 (based on bulk
ever they use variable explicit/implicit character-
velocity and channel height) is available from the
istics based advection scheme, Lagrangian SGS
experiments by Eckelmann [5] and Kreplin [13].
model and hexahedral elements. Compared to the
Turbulence statistics are also available from the
DNS, their computed centerline velocity is about
DNS of Kim et al [11]. These show that the near-
5% higher (1.22 compared to 1.16). The stream-
wall streaks have a spanwise spacing of about 100
wise fluctuations are somewhat over-predicted and
wall units and a maximum in turbulence fluctua-
the spanwise and wall-normal fluctuations are
tions at a distance of 12 or 13 wall units. The fric-
under-predicted, but show the correct behavior.
tion velocity, normalized by the bulk velocity, is
Jansen [10] uses a finite-element method with an
0.0643. DNS of supersonic channel flows by Cole-
unstructured tetrahedral grid. Results for the tur-
man et al [2] agree with the incompressible results
bulence intensities are in good agreement with the
when scaled to account for mean property varia-
DNS of Kim et al [11].
tions. Their results support Morkovin's hypothe-
sis, that relationships between statistics are unaf-
fected by compressibility if the root-mean-square 2.2 Supersonic Turbulent Bound-
density fluctuations are small. ary Layer
Computations of channel flow using the k — c
model give mixed results. Nisizima and Yoshizawa The supersonic turbulent boundary layer is an im-
[20] compute channel flow using an anisotropic portant benchmark for validating LES. Accurate
k — e model. There is good agreement for the simulation of the turbulent boundary layer is nec-
mean profiles and Reynolds stress, but poor agree- essary for providing a realistic (unsteady) inflow
ment for the turbulent kinetic energy. Ciofalo and condition for more complex flow simulations such
Collins [1] report that LES is able to capture flow as shock wave turbulent boundary layer interac-
tions. There are few LES computations of super- where /j is the applied body force
sonic turbulent boundary layer flow as a problem
in its own right, rather than as part of a more com- (7)
plex simulation. Spyropoulos and Blaisdell [28]
simulated a spatially evolving Mach 2.25 boundary
layer corresponding to the experiments of Shutts
Qk = -pcp(Tuk-fuk} (9)
et al [25] using both DNS and LES. They exam-
ined a variety of finite-difference schemes and con-
ducted grid refinement studies. They employed
a modified version of the code used by Rai et al
pe 1_
= pcvT + -pUjUi + pk (11)
[24] for DNS of the same configuration. Hunt and
Nixon [9] performed a structured grid LES of a
pk = -( (12)
supersonic compression ramp which included the
inflow turbulent flat plate boundary layer. Comte
The assumptions made in deriving the energy
and David [3] performed LES of a Mach 2.5 turbu-
equation, Equation 5, are discussed in detail in
lent boundary layer over a compression ramp us-
Ref. [12]. The standard Smagorinsky sub-grid
ing a structured grid. Results show the formation
scale model is [26]:
of Gortler vortices downstream of the compression
corner.
'=£/ V Jv
QfdV (1) The two model constants are CR, the compress-
ible Smagorinsky constant, and Prt, the turbu-
where Q is the filtering function, and its Favre- lent Prandtl number. The values used are based
averaged form / is: on previous studies [6, 18]. We use CR = 0.012
for channel flow, CR = 0.0042 for boundary layer
flow and Prt — 0.4 for both flows.
/=?
P
(2)
For wall-bounded flows the constant CR is mul-
tiplied by the van Driest damping factor, 1 —
where p is the density. From the Navier-Stokes eq-
ey /26 This expression uses wall-units: y+ =
uations for the instantaneous flow variables den-
yuT/v (y measured from the nearest wall). The
sity (p), velocity in the zth coordinate direction
friction velocity is given by UT — \Jrwj pw where
( u i ) , pressure (p) and temperature (T), averaging
TW is the average wall shear stress over both walls
and filtering yield the filtered Navier-Stokes equa-
and pw is the similarly averaged density at the
tions (here written using the Einstein summation
wall. This approach is similar to that of other
notation where repeated indices denote summa- researchers [1, 15, 17].
tion): In finite volume form
{ _Q
(3)
dt ^ / QdV + I (Fi + Gj + Hk) • ndA
at Jv Jdv
+ I (Ri + Sf+ Tit) • ndA + I BdV = 0 (16)
Jdv Jv
dpe d
where
P
r\ f\ pu
a—— (Qk + qk) Uf + -z—— (Tik + &ik) U{ - fiUi (5)
Q= pv (17)
pw
= pRT (6) pe
0 averaged dependent variables stored at the cen-
/I troid of cell i are:
B= h (18)
fa Qi = 77 Qdxdydz (28)
0.02
75
TIME
DNS
LES
u'.y*
u'« 2.5 In y'-f 5.5 (DNS)
u'« 2.65 kiy't 5.8 (Exp)
Eck.hi.im (1874)
Figure 3: Planar Averages of Time Averaged Ve- Figure 6: Planar Averages of Root-Mean-Square
locity in Wall Units Spanwise Velocity Fluctuations
10
0.4
0.2
-0.2
-0.4
-0.6
-0.8
-1
0.5
Y
0.6
0.4
I 0.2
I 0
-0.4
-0.6
-0.8
0.5 0.5
Y Y
Figure 8: Planar Averages of Wall-Normal Com- Figure 11: Planar Averages of Skewness of Wall-
ponent of Streamwise Momentum Flux Normal Velocity Fluctuations
10"
10*
10''
0.5
Y
Figure 9: Planar Averages of SGS Eddy Viscosity, Figure 12: Planar Averages of Skewness of Span-
Normalized with Laminar Viscosity wise Velocity Fluctuations
11
7.5r
0.5
Y
- >*^ _ _ _ _ - ^Sj,
'• , , , , 1 , . , , 1 , • . i i . • , , i
3 0-25 0.5 0.75 1
Y
12