43rd AIAA Aerospace Sciences Meeting and Exhibit

AIAA 2005-1100

Comparison of First and Second Order Turbulence Models

for a Jet/3D Ramp Combination in Supersonic Flow
Valerio Viti* and Joseph Schetz
Department of Aerospace and Ocean Engineering,
Virginia Tech Blacksburg, VA, 24060,USA
Reece Neel
AeroSoft Inc.
Blacksburg, VA, 24060, USA

The full, unsteady three-dimensional Reynolds-Averaged Navier-Stokes Equations were

used to perform a comparative evaluation of the performance of two first-order and of one
second-order turbulence models. The flowfield chosen as the baseline test case is a Mach 4.0
turbulent flow over a three-dimensional compression ramp for which experimental data
exists. In addition to this baseline ramp case, the turbulence models were also compared in
the flowfield created by the interaction of the three-dimensional compression ramp and a
sonic normal injection. The study aims at highlighting the relative shortcomings and
advantages of first-order closure models compared to second-order models and, more
specifically, the capability of state-of-the-art turbulence modeling as implemented in a
commercial software, to produce results for complex flowfields. Both the pros and cons of
eddy viscosity models were put in evidence and discussed when compared to Reynolds Stress
models. The turbulence models selected for this study were the one-equation eddy viscosity
model of Spalart-Allmaras, the two-equation eddy viscosity k- model of Wilcox and the
second-order model proposed by Wilcox that, unlike most other Reynolds Stress models, use
the turbulence frequency instead of the dissipation of turbulent kinetic energy, to compute
the dissipation. The comparison of the turbulence models is intended as a general and partial
guideline to the choice of turbulence models for those professionals in industry that use
numerical codes with built-in turbulence models. For this reason the study made use of offthe-shelf commercial software rather than an in-house development code. The study showed
the difficulty of converging the solution of a complex flowfield when using the Reynolds
stress turbulence model and highlighted the usefulness for practical and design purposes of
eddy-viscosity models.

Nomenclature
A
CFy
CFx
CMz
Cf
Cp
dj
Fj
Fx
Fy
*

=
=
=
=
=
=
=
=
=
=

projected area of solid surface

normal force coefficient, Fy/Fj
axial force coefficient, Fx/Fj
pitching moment coefficient, Mz/(Fjdj)
skin friction coefficient
pressure coefficient
injector diameter
jet thrust
x component of the total force acting on the solid surface
y component of the total force acting on the solid surface

Postdoctoral Researcher at the University of Kentucky, 151 RGAN Bldg, Lexington, 40506 KY, [email protected],
AIAA member.

Holder of the Fred D. Durham Chair, Fellow AIAA.

Research Scientist.
Copyright 2004 by Valerio Viti

turbulent kinetic energy

freestream Mach number
jet Mach number
pitching moment acting on the solid surface
freestream static pressure
freestream static temperature

k
Ma
Maj
Mz
P
T

=
=
=
=
=
=

= inlet boundary layer thickness

= laminar viscosity
= turbulence frequency

I.

Introduction

fter the numerous developments achieved in the 70s and 80s, the science of Computational Fluid Dynamics
(CFD) and, more specifically that part that deals with the Reynolds-Averaged Navier-Stokes (RANS)
equations has touched a development plateau mainly caused by the difficulty of properly modeling turbulence. A
general turbulence model that holds good predictions for a wide range of flowfields has yet to be developed. Many
researchers argue that such a general model simply cannot be developed due to the inherent shortcomings of the
RANS equations. These same researchers are strong proponents of Large-Eddy Simulations (LES) and Direct
Numerical Simulations (DNS). However, from a practical point of view, the use of LES or DNS as design tools is
far from reality due to the high computational costs associated with these computations. For these reasons, it is
envisioned that the RANS equations associated with turbulence modeling will be the main CFD tool used by the
industry and part of the research community in the near future. As a result of this trend, there is a need to improve
the accuracy and reliability of the solutions of turbulent flowfield obtained from the RANS equations. In light of this
discussion, it seems natural to suppose the existence of an extensive benchmark database with the comparative
performance of several turbulence models for different flowfields. In reality, there is not such a database that the
major players in the turbulence field (universities, government research institutes and industry) agree upon; rather it
is left to the individual researchers to experiment with the different existing turbulence models and, somehow, create
guidelines for their use and measure their performance. Several works exist that compare the turbulence models to
experimental data or that perform a comparative study of the models. The present work is intended to add to the
literature that deals with the comparative study of turbulence models. The flowfield chosen as the baseline test case
is a Mach 4.0 turbulent flow over a three-dimensional compression ramp for which experimental data exist. In
addition to this baseline ramp case, the turbulence models were also compared in the flowfield created by the
interaction of the three-dimensional compression ramp and a sonic normal injection. The two flowfields were
selected such that they encompass those physical phenomena that turbulence models have most difficulty to simulate
properly: compressible, wall-bounded flows in strong adverse pressure gradients, separation, compressible mixing
layers,
strong
vortical
flows
produced by compression shocks,
vortex break-downs, steep pressure
Flow
Direction
Separation
gradients associated with expansion
Flat Plate
fans and compression shocks, and
Attachment of
Shock
Mach Disk
shock-boundary layer interactions. In
Counter-Rotating
particular, the study aimed at
Vortices
highlighting
the
relative
Horseshoe
High Pressure
shortcomings
of
first-order
closure
Vortex
Region
models
compared
to
second-order
Low Pressure
Injector
Region
models and, more specifically, the
capability
of
state-of-the-art
Barrel Shock Wave
turbulence modeling as implemented
(Jet Plume)
in a commercial software, to produce
Figure 1. Three-dimensional view of the jet interaction problem with a accurate results for complex
flowfields.
single circular jet injecting normally into a supersonic cross flow.
y

2
American Institute of Aeronautics and Astronautics

II.

Description of the Flowfield

In the jet interaction flowfield several fluid dynamics mechanisms are present and form a complex series of
physical phenomena. The gas is injected through a nozzle in the surface into the supersonic crossflow. The injected
gas expands through a Prandtl-Meyer expansion fan and then recompresses through a Mach disk to form a barrel
shock that acts as an obstruction to the primary flow (see Figure 1). This virtual obstruction produces a shock wave
in the primary flow. The shock wave produces an adverse pressure gradient that causes the boundary layer on the
wall to separate. High pressures are found in the separation region ahead of the injection. At the same time, a large
low-pressure region forms aft of the injector. The low-pressure region has two main effects on the forces and
moments produced by the jet on the solid surface. The first effect is to decrease the normal force on the plate. Note
that the low-pressure region effectively creates a suction behind the jet; and, even though the pressure is not
significantly lower than freestream pressure, it acts over a large area, thus creating a sensible negative force. The
second, and in many aspects most detrimental effect for most applications, is to form a couple with the high-pressure
region ahead of the jet to produce a nose-down moment about the injector. The contribution to the nose-down
moment from the low-pressure region is particularly large, since as mentioned above this region extends far aft of
the injector and therefore its moment arm is relatively large. In order to alleviate the nose-down pitching moment,
several modifications to the jet interaction flowfield have been proposed.

Mach wave

Bow shock
wave
Barrel shock
wave
plume)
Mach disk

M
P
T

Symmetry
Plane

3D
Ramp

Compression
Shock

Expansion
Fan

Turbulent
boundary layer

x
z

Injector
Flat plate

Separation

Mj
Pj
Tj

Ramp
Cross-Section

Flow
Direction

Inclined Plane
of Ramp

Figure 2 Two dimensional view of the flowfield generated by the interaction of the jet (left) and
isometric view of the ramp (right).

Jacobsen, Gallimore and Schetz2 devised the aeroramp, a design that increases the high pressure region and
decreases the size and intensity of the low pressure
region through the use of secondary jets. This
Flow
innovative configuration was studied at Virginia Tech
Direction
by Viti et al.3. However, despite the promising results,
this design needs a complex subsystem of gas supply
and it may become cumbersome to implement on
compact hypersonic vehicles. An alternative and
simpler design to the aeroramp is an actual ramp
located aft of the injector as illustrated in Figure 2.
This configuration was studied by Byun, et al.4 and is
used here to study the performance of the three
turbulence models.
Primary
Injector

Figure 3 Isometric view of the computational grid: 13

zones (mixed H- and C-type grids), 1,544,098 cells.

III.

Numerical Solver and Boundary

Conditions

The numerical solver used in this study is

AeroSofts structured flow solver GASP Version 4.
GASP was chosen because it is a mature program with a proven reliability record in simulations of turbulent flows5,
vortical flows6, jets7, shock-vortex interaction8 and jet-interaction flows9. GASP solves the integral form of the time3
American Institute of Aeronautics and Astronautics

dependent RANS equations in three dimensions. For a full description of GASP and the form of the RANS
equations it solves see Ref 10. The solution was driven to a steady-state using the implicit Gauss-Seidel scheme. The
inviscid fluxes were computed in three dimensions using
the flux-vector splitting of Roe and had 3rd order spatial
upwind biased accuracy with the Min-Mod limiter. The
Parameter
viscous terms are discretized using a standard 2nd order
Flat plate entry length, x0
7.62 cm
Computational domain
accurate central differencing scheme. An exception to this
27.7x15.2x11.4 cm
dimensions (LxHxW)
flux combination was the replacement in only one
Injector diameter, dj
0.4763 cm
computational direction of the Roe flux with the Van Leer
x0/dj
16.00
flux leaving all the other parameters unchanged to avoid
Ramp side/front angles
13.75deg
Ramp Height
0.71 cm
the carbuncle effect11.
Table 1. Flat
dimensions.

plate

and

injector

The grid design that was used in the present

calculations is the result of several optimization steps. The
design is a combination of H-type and C-type grids that
allows a near-optimal cell clustering near the injector. The
Freestream
Jet
grid size was dictated by the need to find a balance
Air, perfect gas
Air, perfect gas
Gas
between the grid refinement and the time to converge a
(=1.40)
(=1.40)
solution to a steady state. The grid was created using
4.025
1.0
M
Gridgen Version 13.312. Care was taken to ensure that the
Pt
1,120.58 kPa
3,797.62 kPa
P
7.135 kPa
2,006.21 kPa
cell adjacent to the wall was at a y+ of less than 1.0. Also,
T
70.3 K
261 K
30 points were inserted in the boundary layer at the inlet
1.65 cm
n/a
Inlet
boundary. One-dimensional hyperbolic tangent stretching
Jet mass flow
n/a
0.1159 Kg/s*
13
was used in all regions to distribute the cells along the
Jet thrust
n/a
37.5 N*
grid connectors. The injector was simulated by cells on the
* These values are for a jet discharge coefficient of 0.75.
surface of the flat plate with imposed pressure and velocity
Table 2. Summary of flow parameters for
equal to the jet total conditions. The grid was sequenced
the Virginia Tech numerical calculations.
twice by eliminating every other cell in the three spatial
directions. The grid had 1,544,098 cells. The
computational domain consisted of a six-sided box (see Figure 3). The lower plane, i.e. the plane defined by
y/d=0.00, corresponds to the solid surface of the flat plate. The no-slip condition (u=v=w=0.0) is imposed on the
solid surface along with p/y=0.0 and the adiabatic wall condition, T/y=0.0. The surface is assumed to be
smooth.
The injector is cut in the surface of the flat plate. For each case investigated, the nozzle was choked and the total
conditions of the jet were known. Therefore, sonic conditions were applied at the cells simulating the jet (MaJ=1.0,
J=*, uJ=wJ=0.0 m/s , vJ=v*, pJ=p*). The jet was assumed to have a flat step profile, i.e. no boundary layer in the
nozzle was simulated. The area of the simulated jet is smaller than the real jet, and the ratio of the two areas is equal
to the discharge coefficient (CdJ) of the real nozzle, 0.75 for the studied conditions. In this way, the viscous effects
inside the nozzle were taken into consideration and the mass flow of the simulated jet was the same as the real jet.
The flow upstream of the injector is supersonic, and a turbulent boundary layer is present. All the dependent
variables on the entry plane outside the boundary layer were assigned their respective freestream value. The
freestream turbulence intensity (TI) was assumed to be 5% since no turbulence measurements were available. The
entry boundary layer thickness, , was obtained from the Schlieren pictures of the tunnel flow. The turbulent
velocity boundary layer profile was then assumed to follow the power-law relationship with the exponent set to 1/7.
The symmetry plane is represented by the x-y plane. The symmetry boundary condition is taken such that the
primitive variables are reflected across this plane with the exception of one velocity component which is reversed.
The three remaining sides of the computational domain (the exit plane, the top surface and the longitudinal plane
opposite the symmetry plane) do not represent any physical surface. The top surface and the sidewall of the wind
tunnels were assumed to be distant enough from the injector not to interfere with the flowfield of interest. Therefore,
a first-order extrapolation boundary condition was applied to these surfaces. The iterative convergence of the
calculations was determined by checking the variation over time of several flow parameters. Convergence was
declared when the variation of the normal force, axial force and pitching moment over time was negligible. A
complete description of the computations can be found in Ref. 14.
4
American Institute of Aeronautics and Astronautics

IV.

Turbulence Models

The results presented in this paper focus on the Spalart-Allmaras model, Wilcox's 1998 k- model, and Wilcoxs
Reynolds stress model. What follows is a brief description of each turbulence model used in this paper. These
models are all implemented in GASP Version 4 and are used without any code modifications. Most of the turbulence
models can be used in conjunction with wall functions, but none of the results presented make use of them.

A. Spalart-Allmaras Model
The Spalart-Allmaras turbulence model is a one-equation model assembled using empiricism and arguments of
dimensional analysis, Galilean invariance, and selective dependence on the molecular viscosity. The model solves
for the variable, , which is similar to the eddy viscosity. The model was developed and calibrated for a certain class
of flows, which includes airfoil and wings. The model has quickly become one of the most widely used oneequation models by researchers and industry. The implementation of the Spalart-Allmaras model follows the
original paper,15 except for a compressibility correction option. The compressibility correction follows the work by
Forsyth et al.
B. k- Models
There are four types of k- models in GASP. The first represents the 1988 Wilcox k- model. For free-shear
flows, Wilcox made improvements to the 1988 model which resulted in the 1998 model.16 For wall-bounded flows,
the '88 and '98 models should perform about the same, while for free-shear flows the improvements to the '98 model
allow for spreading rate predictions. The Menter SST model17 is a blend of Wilcox's '88 model and the k- model.
This model tries to apply the '88 model to the inner wall regions of a boundary layer and a transformed k- model to
the outer boundary layer regions and free-shear layers. Menter's model is expressed in terms of k and , so it is
grouped with the k- models in GASP. The above three k- models are intended for high Reynolds number flows.
For low Reynolds number flows, GASP has a low Reynolds number version of the '98 Wilcox model.16 The
simulations presented in this paper will focus on the 1998 Wilcox model, simply referred to as k- through out the
remainder of the paper.

C. Reynolds Stress Model ( - )

The one and two-equation models previously presented make use of the Boussinesq eddy-viscosity
approximation. This assumes that the principal axes of both the Reynolds stress tensor and the mean strain-rate
tensor are coincident everywhere in the flow. The Boussinesq approximation weakens in flows with sudden changes
in the mean strain rate, flows with strong curved surfaces, flows with separation and flows with three-dimensional
features. A Reynolds stress model, in theory, will circumvent the deficiencies of the Boussinesq approximation. The
Reynolds stress model implemented in GASP Version 4 is Wilcox's stress- model.16 Unlike most Reynolds stress
models which use the dissipation rate of turbulence kinetic energy to compute the destruction of turbulence, this
model uses the turbulence frequency, . The pressure-strain correlation uses the LRR model17 which has the
advantage of not requiring wall reflection terms. Thus, the model does not require a wall distance term.

Wilcox designed the model to perform similar to the '98 k- model. The equations are very similar, along with
the wall boundary condition for
which can be used to determine wall roughness. The user inputs a nondimensional sand-grain roughness value, k+s, which is then used to set the value of at the wall. Unless otherwise
stated, all the validation cases that follow used a value of k+s = 5. A description of the implementation of the -
model in GASP can be found in Neel et al.1

In general, Reynolds stress models are not as practical as one and two-equation models due to the extra
computational cost associated with them. A Reynolds stress model solves six equations for the Reynolds stress
tensor and another equation for the dissipation. While in theory a Reynolds stress model should perform better than
a first order turbulence model, there is no guarantee.

5
American Institute of Aeronautics and Astronautics

V.

Results

This section presents the results of the computations and discusses their comparative performance. Each
calculation was checked for iterative convergence by monitoring four main parameters. The behavior of the mass
conservation residual was used as an indication of general convergence. However, more detailed data about
flowfield convergence was obtained from monitoring the behavior of the normal force on the flat plate, the pitching
moment and the drag force. In all the cases, the solution was declared converged when all of these parameters were
either not changing or oscillating around a constant value. An attempt was made to run the jet on case using the
RSTM. Using grid sequencing, solutions were attained on coarse and medium grid levels which were consistent
with the S-A and k- solutions. As the solver was run on the fine grid, the solution appeared to break down and
convergence did not seem apparent. Given the same number of iterations that converged the S-A and k- solutions,
the RSM solution was not close to being converged. Instead, the solution showed signs of transient behavior. Due to
this inability to converge the RSTM on the fine grid, no results will be presented here for the RSM jet on case. This
will instead remain a topic for further study.
The pressure distribution along the tunnel centerline for the cases with the jet-off (left) and the jet-on (right) is
shown in Figure 4. The cross-section of the three-dimensional ramp is superimposed on the pressure profiles as a
spatial reference. In the case without injection, the pressure coefficient shows the expected distribution, similar to
that encountered in a two-dimensional compression ramp. The pressure increases rapidly at the start of the ramp due
to the compression shock, and it keeps increasing more slowly along the inclined surface. At the end of the inclined
0.6

0.6

Spalart Allmaras

0.5

Spalart Allmaras

0.5

k-omega

k-omega

Reynolds Stress

Experiment, Byun, et al.

0.4

0.3

0.3

Cp

Cp

0.4

0.2

0.2

Ramp Cross Section (y-dimension not to scale)

0.1

0.1

0
-15

-10

-5

10

15

20

25

30

35

40

-15.0

-10.0

-5.0

-0.1

0.0

5.0

10.0

15.0

20.0

25.0

30.0

35.0

40.0

-0.1

x/dj

x/dj

Figure 4 Comparison of the Cp distributions along the centerline for the case without the primary jet on the
left and with the primary jet activated on the right.
0.0075

0.0075

Spalart Allmaras
k-omega
Reynolds Stress

0.005

Spalart Allmaras
0.005

k-omega

0.0025

Cf

Cf

0.0025

0
-15

-10

-5

10

15

20

25

30

35

40

-15.0

-10.0

-5.0

0.0

5.0

10.0

15.0

20.0

25.0

30.0

35.0

40.0

-0.0025

-0.0025

-0.005

-0.005

x/dj

x/dj

Figure 5 Comparison of the Cf distributions along the centerline for the case without the primary jet on the
left and with the primary jet activated on the right.
6
American Institute of Aeronautics and Astronautics

section, an expansion fan is created by the turning of the flow and the pressure is rapidly equalized back to the
freestream value. It is evident that in this relatively simple flowfield, all the turbulence models predict the pressure
distribution similarly. The right side of Figure 4 shows the pressure coefficient plot when the jet is activated. The
flowfield is more complicated than in the jet-off case and presents large recirculation regions, a barrel shock, a bow
shock, and vertical structures. In light of this, it is not surprising that the three turbulence models predict a somewhat
different pressure distribution. The SA model predicts a longer separation region than the other models. The SApredicted separation extends ahead of the injection location by 13 jet diameters. On the other hand, the k- model
shows a shorter but stronger separation region with a higher Cp peak that matches closely the pressure distribution
measured by Byun, et al.4, indicated by the triangular symbols. While the k- model captures the extent and
intensity of the separation, the SA model under-predicts it, and it shows some minor oscillatory behavior in the
region where the counter-rotating vortices form, that is between x/L of -8.0 and -4.0. In spite of these differences,
ahead of the injector, the general shape of the two distributions resembles each other, with the corresponding peaks
and troughs in the Cp plots. Aft of the injector, the predicted Cp distributions show better agreement with the
experimental data. The two models predict similarly the location and intensity of the low-pressure region behind the
jet, including the increase in pressure due to the ramp. Notice that the shape of the pressure increase produced by the
ramp in the jet-on case is very similar in shape and location to that seen in the jet-off case. Past the inclined surface
of the ramp, is where the reflected shock from the Mach disk of the jet expansion impinges on the solid surface of
the ramp (see the right side of Figure 7 for a view of this flowfield). A steep increase in Cp marks this location
where the pressure coefficient goes from negative to positive. The SA predicts the impingement location to be
slightly ahead of that predicted by the k- model. The reasons for the difference between the models are not clear.
In Figure 5, the skin friction coefficients for the two test cases are compared. As in the case for the pressure
coefficient, the predicted jet-off skin friction distributions are similar. The skin friction is constant ahead of the
compression ramp. At the start of the ramp, the compression shock is indicated by a severe dip in the skin friction
coefficient. Note how the SA model predicts a small separation (negative Cf) in the region just ahead of the ramp.
As the flow moves past the shock, it accelerates quickly and then slows to a more moderate acceleration along the
inclined section of the ramp. The expansion fan at the end of the inclined section is indicated by a spike in the skin
friction coefficient. This spike is due to the rapid acceleration of the flow that is turned back into the direction of the
freestream. Once the flow moves downstream of the expansion fan, the high Cf values of the spike are quickly
reduced and then gradually taper off toward a constant value as the flow moves along the ramp. Note that the skin
friction spikes predicted by the three turbulence models correlate well with the main features noted in the Cp
distributions.
The right side of Figure 5 shows the predicted skin friction distributions in the jet-on case. In this plot, the
differences in the predicted Cf are accentuated. The separation region ahead of the injection is larger for the SA
model. Notice the spikes and troughs that correspond to the counter-rotating vortices in the separation region. The
attachment line of these two vortices is predicted at the same location by the models, at x/L of -2.5. Also, similar to
the Cp distributions, the overall shape of the plots resembles each other, with the same number of separation and
reattachment lines. Aft of the jet, the predicted skin frictions are very similar up to the location where the reflected
shock impinges on the surface of the ramp. The k- predicts a more severe effect of the shock on the flow than the
SA model; the Cf predicted by the SA model dips to zero, indicating an incipient separation. However, the adverse
pressure gradient caused by the reflected shock does not extend downstream, and the local flow quickly recovers
momentum and speed as indicated by the steep increase in skin friction coefficient, at x/L13. Past this point, the
skin friction coefficient tapers down to an equilibrium value that, even though is not captured by this computational
domain, is expected to be the same as that of the jet-off case. The k- model too predicts a dip in the Cf at x/L 13
followed by a rapid acceleration, but the intensity of the dip is not as strong as the one predicted by the SA models.
Further understanding of the flow and of the differences in model prediction can be obtained by looking at the
mappings on the surface of the flat plate-ramp and on the plane of symmetry. Figure 7 compares the calculated
pressure coefficient mappings on the surface of the plate. The left side of the picture corresponds to the jet-off case
and the right side to the jet-on case. The top row corresponds to the SA model, the middle row to the k- model and
the bottom row to the RSTM, with the RSTM solution for the jet-on case being omitted. Looking at the left column,
it appears that all of the mappings for the jet-off case are very similar. The only noticeable difference is for the SA
model, for which the separation region seems to cover a larger area than what is predicted by the other two models.

7
American Institute of Aeronautics and Astronautics

Figure 6 Comparison of the pressure coefficient mappings on the solid surface. The left side is for the case
with the jet off and the right side is for the case with the jet on. The top row shows the results from the SA
model, the middle row from the k- model and the bottom row from the RS model.

More differences between the turbulence models can be noticed in the jet-on case. Looking at the separation
region ahead of the injector, we observe that the SA model predicts a larger separation region than the k-
turbulence models. However, the SA model fails to predict the second pressure peak right in front of the injection, as
seen in the other model. This correlates well with what was seen in the Cp plot of Figure 4, where the k- model
was found to produce the Cp plot that agreed most with the experimental one. Figure 7 also allows an analysis of the
different pressure distribution predicted far away from the centerline. This is noteworthy since, even though the SA
model predicts a longer separation ahead of the injector, its effect does not extend as far downstream and laterally as
in the case of the k- model. This difference is reflected in the force and moments coefficients shown in Table 3.
The k- model predicts the highest normal force of the three models which relates well to the large high-pressure
regions computed by the same model in Figure 6. Similarly, the large high-pressure regions ahead of the injector
explain the negative (nose-down) pitching moment predicted by the k- model, see Table 3 (c). Due to the
dependency of the moment on both the intensity of the pressure force and its spatial distribution, the comparison of
the pitching moments highlights the discrepancies between the two turbulence models.
Figure 7 is a comparison of the first spatial derivative of the density on the plane of symmetry of the flow. The
spatial derivative gives the main features that would be seen in a Schlieren photograph of the flow. The layout of the
figure is the same as Figure 7, with the mappings on the left side of the picture corresponding to the jet-off case and
those on the right side to the jet-on case. The top row corresponds to the SA model, the middle row to the k- model
and the bottom row to the RSTM. As noted for the Cp mapping, the results for the jet-off case do not show any
major difference. The compression shock at the beginning of and the expansion fan at the end of the inclined section
of the ramp are predicted similarly by the three turbulence models, with the shock angles being the same. However,
8
American Institute of Aeronautics and Astronautics

notice the higher gradients in the compression shock computed by the SA as compared to the k- model. On the
right column of Figure 7, the results for the jet-on case are shown. The comparison of the mappings shows a general
agreement between the three turbulence models, with the main flow features such as the barrel shock, the bow
shock, and the reflected shock having the same location and inclination to the freestream. However, some important
differences in the results are evident, especially in the region ahead of the injection. The SA model seems to predict
a much longer separation region that extends
to the upstream inlet boundary condition.
Notice how the SA model is predicting larger
No Jet
Jet On
regions of higher gradients than the k-
Turbulence Model
CFy
diff. %
CFy
diff. %
model, particularly in the region comprised
Spalart-Allmaras
0.21
-4.2
0.88
-18.1
between the bow shock and the leading edge
k-
0.22
0.0
1.07
0.0
of the barrel shock. A long trailing vortex
Reynolds Stress
0.22
0.9
n/a
n/a
appears to start from the barrel shock and
extend downstream to the end of the
(d) Normal Force Coefficient, CFy
computational domain. This flow feature is
No Jet
Jet On
not predicted by the k- model and it seems
to be dependent on the grid topology.
Turbulence Model
CFx
diff. %
CFx
diff. %
Spalart Allmaras

0.068

-1.4

0.056

-33.9

k-

0.069

0.0

0.084

0.0

Reynolds Stress

0.072

4.5

n/a

n/a

VI.

Computational Efficiency

The computational costs associated with

using the turbulence models in this study are
(e) (a) Axial Force Coefficient, CFx
now discussed. For the results presented in
this study, GASP solved the governing
No Jet
Jet On
equations using an implicit algorithm in
Turbulence Model
CMz
diff. %
CMz
diff. %
which the turbulence models were uncoupled
from the RANS equations. Running the
Spalart Allmaras
1.88
-1.4
-0.44
-93.8
turbulence models uncoupled reduces the
k-
1.91
0.0
-7.11
0.0
amount of computational work due to the
Reynolds Stress
1.99
4.3
n/a
n/a
smaller system size. The trade-off of running
(f) (a) Pitching Moment Coefficient, CMz
uncoupled is loss of implicit behavior, which
in theory will have an impact on solution
convergence. For most practical problems,
Table 4. Comparison of the normal force, axial force and
only a slight decrease in convergence rate is
pitching moment obtained with the three turbulence
observed when running uncoupled, which is
models. Forces are normalized by the jet thrust and
worth the trade-off of reduced CPU times per
moments are normalized by the jet thrust multiplied by the
iteration cycle. Relative CPU times are now
jet diameter.
compared for the current case for the oneequation, two-equation, and seven-equation turbulence models. These comparisons came from runs performed on an
8 node Linux cluster. Variations in CPU times can be expected if the platforms or number of processors are
changed. If we take the one-equation Spalart-Allmaras model as the baseline, then using the two-equation k- model
increases the computational cost by a factor of 1.1 and using the seven-equation RSTM increases cost by a factor of
2.7. Therefore, from a computational cost perspective, there is only a slight penalty to using k- over the SA model,
while using the RSTM more than doubles the cost of an interaction cycle. The large increase in CPU cost with the
RSTM comes from fact that both the SA and k- simulations have a max system size of 5x5 (which is due to the
RANS equations) while for the RS simulation, the max system size is 7x7 (which is from the RSTM). The
convergence rates for all three turbulence models are roughly the same, so the overall time to converge the solution
is directly dependent on the turbulence model chosen. The user should therefore be aware of the large increase in
CPU time and cost associated with using the RSTM. On the other hand, there is approximately a 10% penalty in
selecting a two-equation model over a one-equation model.

9
American Institute of Aeronautics and Astronautics

RSTM

Figure 7 Comparison of the mappings of the first spatial derivative of density on the plane of symmetry. The
left side is for the case with the jet off and the right side is for the case with the jet on. The top row shows the
results from the SA model, the middle row from the k- model and the bottom row from the RSTM.

VII.

Conclusions

Three turbulence models were tested in the Mach 4.0 flowfield created by a ramp with and without sonic
transverse injection. The turbulence models considered in this work include two first-order models, the one-equation
Spalart-Allmaras and the two-equation Wilcox 1988 k- model, and one second-order (Reynolds Stress) model, the
Wilcox - model. The jet-off flowfield represented a simpler baseline case against which the more complex jet-on
case was compared.

For the jet-off case, the three models produced similar results. The compression shock ahead of the inclined
section of the ramp as well as the expansion fan at the end of the inclined section was well predicted by all the
models. The pressure coefficient plots and mappings, the plots of the skin friction and the mappings of the first
10
American Institute of Aeronautics and Astronautics

spatial derivative of density predicted by the three models are all in good agreement. Also, the computed forces and
moments acting on the flat plate-ramp combination were similar for all the turbulence models.
For the jet-on case, the three models showed a higher level of variability. We were not able to obtain a
converged solution for the Reynolds-stress model. The Reynolds-stress computations produced a converged solution
only on the coarse and medium grid sequences. However, as the problem was run on the fine grid sequence, the
solution seemed to break down and it showed signs of numerical instability. Most of the differences between the
other two models concerned the separation region ahead of the injector. The Spalart-Allmaras model predicted the
largest separation region while the k- model predicted a smaller separation region. The best match with the
experimental values of the pressure coefficient was obtained by the k- model. In the region aft of the jet the
predicted flowfields were more in agreement, due to the fact that less steep pressure gradients are present in this
area. Comparison of force and moments showed that the k- model predicted the highest normal force, drag force,
and nose-up pitching moment. The inability to obtain a converged solution in this complex flowfield with the
Wilcoxs Reynolds-stress model highlighted the difficulty of using Reynolds stress models as compared to simpler
eddy-viscosity models. In addition to this, on average, the Reynolds-stress model took 2.7 times and the k- model
1.1 times the amount of CPU time needed by the Spalart-Allmaras model to produced a converged solution.

Acknowledgements
The authors would like to thank AeroSoft, Inc. and Bill McGrory for supporting this work and granting us the permission to use
Gasp Version 4.

References
1

D. D. Aspley, M.A. Leschziner, Investigation of Advanced Turbulence Models for the Flow in a Generic Wing-Body
Junction, Flow Turbulence, Combustion, No. 65, pp 25-55, 2001.
2
L., Jacobsen, S., Gallimore, J., Schetz, and W., OBrien, An improved Aerodynamic Ramp Injector in Supersonic
Flow, AIAA 2001-0518, Jan. 2001.
3
V. Viti, J. Schetz, and R. Neel, Numerical Studies of the Jet Interaction Flowfield with a Main Jet and an Array of
Smaller Jets, International Congress of Aeronautical Sciences, Paper ICAS 2002-4.7.1, September 2002, Toronto, Canada.
4
Y.H. Byun, K.J. Bae, S. Wallis, V. Viti, J. Schetz, Jet Interaction in Supersonic Flow with Downstream Surface Ramp,
AIAA Paper 2003-3933, 21st AIAA Applied Aerodynamics Conference, June 2003.
5
R. P. Roger and S. C. Chan, "Parameters affecting penetration of a single jet into a supersonic crossflow: A CFD Study II", AIAA Paper No. 98-0425, January 1998.
6
A. Nedungadi, M. J. Lewis ; "A Numerical Study of Fuel Mixing Enhancement Using Oblique Shock/Vortex
Interactions", AIAA Paper No. 96-2920, February 1997.
7
T. Hsieh, "Analysis of the Scaling Effects for missile Configuration with Lateral Thruster", AIAA Paper No. 99-0810,
January 1999.
8
A. Nedungadi; M. J. Lewis; "Computational Study of three-dimensional Shock-Vortex interaction", AIAA Journal of
Aircraft, Volume 34, No. 12, December 1996.
9
McDaniel, J., Glass, C., Staack, D., and Miller, C. "Experimental and Computational Comparison of an Under-expanded
Jet Flowfield", AIAA Paper No. 2002-0305, January 2002.
10
GASP 4.0 User Manual, AeroSoft, 2001, ISBN 0-9652780-5-0.
11
K. M. Peery, S. T. Imlay, Blunt-body flow simulations, AIAA paper 88-2904, 1988.
12
Gridgen Version 13.3 user manual, Pointwise, Inc., PO Box 210698, Bedford, TX 76095. 1999.
13
M. Vinokur, On one-dimensional stretching functions for finite difference calculations, Journal of Computational
Physics, Vol. 50, No. 3, pp.215-234, 1983.
14
V. Viti, Numerical Studies of the Jet Interaction Flowfield with a Main Jet and an Array of Smaller Jets, Ph.D.
dissertation, Department of Aerospace and Ocean Engineering, Virginia Tech, Blacksburg, 24060 VA, September 2002.
15
Spalart, P. R. and Allmaras, S. R., A One Equation Turbulence Model for Aerodynamic Flows," La Recherche
Aerospatiale, Vol. 1, 1994, pp. 5-21.
16
Wilcox, D. C., Turbulence Modeling for CFD, DCW Industries, 2nd ed., 1998
17
Menter, F. R., Zonal Two Equation k- Turbulence Models for Aerodynamic Flows," AIAA Paper 932906, July 1993.
18
Launder, B. E., Reece, G. J., and Rodi, W., Progress in the Development of a Reynolds-Stress Turbulence Closure,
Journal of Fluid Mechanics, Vol. 68, No. 3, 1975, pp. 537-566.
19
Neel, R. E., Godfrey, A.G., Slack, D.C, Turbulence model validation in GASP Version 4, AIAA Paper 2003-3740,
June 2003.

11
American Institute of Aeronautics and Astronautics