PHD 2000 Stefan Wallin

Engineering turbulence modelling for CFD with a focus
on explicit algebraic Reynolds stress models

by
Stefan Wallin
February 2000
Technical Reports from
Royal Institute of Technology
Department of Mechanics
S-100 44 Stockholm, Sweden
Typsatt i AMS-LATEX med KTH’s thesis-stil.
Akademisk avhandling som med tillstånd av Kungliga Tekniska Högskolan i

Stockholm framlägges till offentlig granskning för avläggande av teknologie dok-
torsexamen fredagen den 25 februari 2000 kl 10.15 i Kollegiesalen, Administra-
tionsbyggnaden, Kungliga Tekniska Högskolan, Valhallavägen 79, Stockholm.
Stefan
c Wallin 2000
Norstedts Tryckeri AB, Stockholm 2000
Engineering turbulence modelling for CFD with a focus
on explicit algebraic Reynolds stress models
Stefan Wallin 2000
Department of Mechanics, Royal Institute of Technology
SE-100 44 Stockholm, Sweden
Abstract
Modelling of complex turbulent flows using explicit algebraic Reynolds stress
and scalar flux models (EARSM and EASFM) has been considered. The eddy-
viscosity or eddy-diffusivity assumption has been replaced by a more general con-
stitutive relation for the second-order correlation in the Reynolds averaged equa-
tions. This relation has been derived using a formal approximation of the corre-
sponding second-order transport model equations in the weak-equilibrium limit.
The proposed EARSM is an exact solution of the implicit algebraic Reynolds
stress model (ARSM) for two-dimensional mean flows and a reasonable approx-
imation also in three-dimensional mean flows where the fully three-dimensional
tensorial form is kept. Asymptotically correct near-wall treatment, extension
to compressible mean flows and approximations of the neglected advection and
diffusion terms are proposed. The resulting model behaves well in a number
of different engineering and generic test cases giving significant improvements
compared to standard eddy-viscosity models and the computational effort was
comparable to standard two-equation models. The proposed EASFM is an exact
solution of the corresponding implicit algebraic model in both two- and three-
dimensional mean flows. Á priori tests show good model behaviour in homoge-
neous shear flows, channel and wake flow for the scalar flux vector.
Descriptors: Turbulence model, Reynolds stress, passive scalar flux, explicit

algebraic models, computational fluid dynamics, aerodynamics, nonequilibrium
turbulence, weak-equilibrium assumption.
Preface
This thesis considers the modelling of turbulent flows using explicit algebraic
Reynolds stress and scalar flux models. The thesis is based on and contains the
following papers.
Paper 1. Wallin, S. & Johansson, A. V. (2000) An explicit algebraic

Reynolds stress model for incompressible and compressible turbulent flows. J.
Fluid Mech. 403, 89–132.
Paper 2. Lindblad, I.A.A., Wallin, S., Johansson, A.V., Friedrich, R.,

Lechner, L., Krogmann, P., Schülein, E., Courty, J.-C., Ravachol, M.
& Giordano, D. (1998) A prediction method for high speed turbulent sepa-
rated flows with experimental verification. AIAA Paper No. 98-2547.
Paper 3. Wallin, S. & Girimaji, S.S. (2000) Evolution of an isolated turbu-

lent trailing vortex. Accepted for publication in AIAA J.
Paper 4. Wallin, S., Wang, D., Berggren, M. & Eliasson, P. (2000) A

computational study of unsteady turbulent buffet aerodynamics. To be submit-
ted.
Paper 5. Wikstrm, P.M., Wallin, S. & Johansson, A.V. (2000) Deriva-

tion and investigation of a new explicit algebraic model for the passive scalar
flux. Accepted for publication in Phys. Fluids.
Paper 6. Wallin, S. (1999) An efficient explicit algebraic Reynolds stress k–ω

model (EARSM) for aeronautical applications. FFA TN 1999-71.
Paper 7. Eliasson, P. & Wallin, S. (1999) A robust and positive scheme for
viscous, compressible steady state solutions with two-equation turbulence mod-
els. FFA TN 1999-81.
The papers are here re-set in the present thesis format, and some minor differ-
ences are present as compared to published versions. The first part of the thesis
is both a short introduction to the field and a summary of the most important
v
vi PREFACE
results presented in the papers given above. The main reults of the papers are
presented in the context of a general introduction to the field of engineering
turbulence modelling.
In memory of my father
Contents
Preface v
Chapter 1. Introduction 1
Chapter 2. Basic concepts in turbulence modelling 3

2.1. Governing equations and Reynolds averaging 3
2.2. Eddy-viscosity models 5
2.2.1. Algebraic and one-equation models 5
2.2.2. Two-equation models 6
2.3. Reynolds stress models 8
2.3.1. Differential Reynolds stress models (DRSM) 9
2.3.2. Algebraic Reynolds stress models (ARSM) 10
2.3.3. Explicit algebraic Reynolds stress models (EARSM) 11
2.3.4. Nonlinear eddy-viscosity models (NL-EVM) 13
2.4. Scalar flux models 14
Chapter 3. The proposed modelling approach 16

3.1. The weak equilibrium assumption 16
3.1.1. The consistency condition 17
3.1.2. Slowly distorted turbulence 19
3.1.3. Streamline curvature effects 19
3.1.4. Pressure-strain rate model 21
3.2. Near-wall treatment 24
3.3. Numerical behaviour and some applications 26
3.3.1. Effective eddy viscosity 26
3.3.2. Positivity 26
3.3.3. Compressible flow applications 27
3.4. Time-dependent turbulence 29
3.5. Passive scalar modelling 30
Chapter 4. Conclusions and outlook 33
Acknowledgments 35
Bibliography 36
ix
x CONTENTS
Paper 1. 43
Paper 2. 97
Paper 3. 119
Paper 4. 145
Paper 5. 163
Paper 6. 195
Paper 7. 215
CHAPTER 1
Introduction
Two brothers on a windy field on one of the barrier islands of North Carolina
are getting ready for the first test flight of the day. It is December 17 (a special
date in the history of fluid mechanics) and the conditions are perfect. It is
Orville’s turn to control the airplane. The engine starts and he takes off for a 36
m flight and Wilbur and Orville Wright became historic with the first controlled
flight.
This event in 1903 marks the beginning of the amazing history of aeronautics
during the 20th century. It was rather the pioneering work of Wilbur and Orville
in flight mechanics and aerodynamic research preceding the historic flight that
had the largest scientific value. Carrying out tedious experiments in a primitive
wind tunnel they found that the relation for the lift force that they used was
basically wrong and, moreover, they discovered wing profiles that produced much
more lift than the previously used circular arc shapes. These experiments during
the first years of the century became the necessary basis for the successful flight.
Since then wind tunnel techniques have been greatly improved. Most of the
airplanes today are designed based on data from wind tunnel experiments and
empirical correlations derived from such experiments and from experiences of
previous airplane designs. It is not until the last few decades that computational
fluid dynamics (CFD) has begun to be an alternative (or rather a complement)
to expensive wind tunnel experiments. Earlier computational methods had se-
vere physical restrictions, and were typically based on potential methods and
boundary layer methods. These methods were, and still are, mainly used for ap-
proximative concept studies rather than for complete validations of final designs.
The major problem in CFD is the strong nonlinearity in the governing equa-
tions along with the corresponding nonlinearity in the dynamics of the physical
reality. The nonlinearity causes anomalies in the flow, such as discontinuities
(shock waves), and that small disturbances in some critical points may cause
major effects in the flow. However, progress in numerical methods and com-
puter performance have solved many of these problems and today it is possible
to, with at least some confidence, obtain a numerical solution of the flow field
around a complete aircraft configuration including all major physical effects, such
as e.g., shock waves, turbulent boundary layers and vortices. In some applica-
tions, such as the reentry of space vehicles back to earth, representative wind
tunnel experiments are not possible and one must rely on CFD methods.
1
2 1. INTRODUCTION
There is, however, one major physical approximation still present in the
present CFD methods namely the modelling of the turbulence. The flow in the
boundary layer closest to the solid walls of all moving bodies, including airplanes,
cars, trains, ships and whales are mostly turbulent. This is also the case in most
internal flows, such as pipe flows and flows in engines and pumps. This means
that the flow is chaotic and stochastic with eddies in a wide range of sizes. This
can quite literally be felt in the wind on a stormy day. The turbulent boundary
layer very much controls the performance of the airplane, such as drag and
maximum lift forces, which set the fuel consumption and landing and take-off
speeds. Thus, it is crucial to be able to accurately predict, and possibly also to
control, the turbulent boundary layer.
The equations that govern the turbulent flow are known, so in principal the
turbulent field may be computed directly in a, so called, direct numerical sim-
ulation (DNS). If the Reynolds number is large (it almost always is), i.e. the
internal viscosity of the fluid is small, the smallest turbulent eddies are very
small compared to the largest eddies and the computational effort to capture all
the dynamics is enormous. Years of computational time on the fastest super-
computers is needed to simulate the turbulence in just a small strip (say 10 × 1
cm2 ) on a typical airplane wing. For that reason the turbulence needs to be
modelled. Hence the turbulent flow is described by statistical methods, just like
the meteorologist talking about the mean wind speed and direction even if the
wind in the gusts are varying wildly.
Turbulence research and modelling efforts go back to the late 19th century
with landmark events such as the first attempt to model turbulence in terms
of a turbulent viscosity proposed by Boussinesq[5] (1877) contemporary with
the Reynolds[40] pipe experiments in 1880. First, the turbulent viscosity was
assumed to be constant and not until 1925 the more useful mixing-length idea
was proposed by Prandtl[38] from which the well known, and in CFD much used,
Baldwin & Lomax[3] model originates. Other pioneering work in more advanced
models are the first two-equation model by Kolmogorov[27] (1942) and the first
Reynolds stress transport model by Rotta[46] (1951). It was not until the 70’s
the modelling efforts started to make real progress when the computers started
to be useful for CFD applications. Worth notice are the Jones & Launder[22]
K–ε model and the Launder, Reece & Rodi[29] RST model.
Engineering turbulence flows are often, by nature, complex. This means that
different turbulence effects are present simultaneously and are interacting. The
most important demand on an engineering turbulence model is thus generality.
Complex flows often mean three-dimensional complex geometries and another,
equally strong demand is that the computational effort must be limited. The ma-
jor contribution of this work is the extension of validity for a class of turbulence
models that is affordable for engineering CFD methods.
CHAPTER 2
Basic concepts in turbulence modelling
2.1. Governing equations and Reynolds averaging

Compressibility is an important aspect of many engineering flows. However,
most of the basic turbulence theories, modelling developments, and experimen-
tal as well as numerical simulations are made for incompressible flows. Simple
extensions of incompressible turbulence models to account for mean-density vari-
ations are successfully used as long as the Morkovin’s hypothesis holds, that is
when the compressibility effects due to turbulent fluctuations are negligible.
For attached boundary layers, compressibility effects due to turbulent fluc-
tuations begin to be important for a Mach number around five, but in free
shear flows these effects become important at much lower Mach numbers. For
a detailed description of the intricate aspects of compressibility effects due to
turbulent fluctuations, please consult Friedrich[11].
To illustrate the basic approaches for turbulence modelling we will here
restrict ourself to incompressible flows. The extension to compressible flows is
described in Paper 1 (see also Johansson & Burden[19]). Turbulent flows are
supposed to be governed by conservation of mass and momentum through the
Navier–Stokes equations, which read
∂
ui
= 0 (1)
∂xi
∂
ui ∂
ui 1 ∂ p ∂
j
+u = − + (2ν
sij ) (2)
∂t ∂xj ρ ∂xi ∂xj
where ui and p are the instantaneous velocity and pressure fields. ρ and ν are
the (constant) density and kinematic viscosity respectively. The instantaneous
strain rate tensor is defined as sij ≡ ( j,i )/2.
ui,j + u
These equations completely describe the turbulent field and may, in prin-
cipal, be solved directly in a so called direct numerical simulation (DNS). The
smallest length and time scales in a turbulent field then need to be resolved,
and thus the computational effort grows rapidly with increasing Reynolds num-
ber, which relates the largest or geometrical scales to the smallest or viscous
scales. Therefore, direct numerical simulation is mainly used as a research tool
for studying turbulence in detail and for calibration of turbulence models. The
simulations are typically run for geometrically simple generic cases like different
kinds of homogeneous turbulence, channel and pipe flows, or boundary layers.
3
4 2. BASIC CONCEPTS IN TURBULENCE MODELLING
In practically all engineering cases, the turbulence needs to be considered

using a statistical approach. The flow-field variables are then decomposed into
mean and fluctuating parts such as u i = Ui +ui and p = P +p where Ui and P are
defined as the ensemble average over a large number of independent realizations
of the flow field, Ui ≡ u i and P ≡ p. ui and p are the fluctuating parts, for
which ui = 0 and p = 0.
The Reynolds averaged Navier–Stokes (RANS) equations are formed by tak-
ing the mean of the Navier–Stokes equations using the decomposition defined
above
∂Ui
= 0 (3)
∂xi
D Ui 1 ∂P ∂
= − + (2νSij − ui uj ) (4)
Dt ρ ∂xi ∂xj
where the mean strain rate tensor Sij ≡ (Ui,j + Uj,i )/2. The notation D/Dt ≡
∂/∂t + Uj ∂/∂xj is used to denote the rate of change following the mean flow.
The momentum equation for the mean flow field contains an additional term
that originates from the average of the nonlinear term. That term is called the
Reynolds stress −ρui uj and is felt by the mean field as an additional stress due
to the turbulent fluctuations. In incompressible flows the density, ρ, is constant
and gives no additional information and, thus, it is common to denote only the
velocity correlation as the Reynolds stress.
Turbulence modelling in engineering CFD methods is simply a matter of
finding an expression for the Reynolds stress tensor in terms of known quantities.
Its complexity ranges from purely algebraic relations to relations that consider
one or several extra transport equations for different turbulent correlations.
The transport equation for the Reynolds stress tensor may be derived from
the Navier–Stokes equations
D ui uj
= Pij − εij + Πij + Dij . (5)
Dt
The terms represent production, dissipation, pressure-strain rate, and diffusion
(molecular and turbulent), respectively. The exact expressions for these are
given by, e.g., Johansson & Burden[19]. One should particularly note that the
production term is explicit in the Reynolds stresses, whereas the other terms
need to be modelled.
The transport equation for the turbulent kinetic energy, K ≡ ui ui /2, is
derived by taking half of the trace of the transport equation for the Reynolds
stress tensor
DK
= P − ε + D(K) . (6)
Dt
The production, dissipation and diffusion (molecular and turbulent) terms on
the right-hand side are given by half the trace of the respective terms in the
Reynolds stress transport equation.
2.2. EDDY-VISCOSITY MODELS 5
The trace of the pressure-strain term is zero so that it does not contribute
to the energy equation. The Reynolds stresses may also be expressed in terms
of the Reynolds stress anisotropy tensor
ui uj 2
aij ≡ − δij (7)
K 3
which is symmetric and traceless (aij = aji and aii = 0).
2.2. Eddy-viscosity models

The analogy between the viscous stress generation caused by fluctuations on
the molecular level and the generation of turbulent stresses caused by macro-
scopic velocity fluctuations lead historically to an eddy (or turbulent) viscosity
formulation or the Reynolds stress. The first attempt to model the Reynolds
stresses was made by Boussinesq[5] who introduced an eddy viscosity, νT , in
complete analogy with the molecular viscosity for a Newtonian fluid. The orig-
inal Boussinesq assumption or the eddy viscosity assumption related the shear
component of the Reynolds stress tensor in nearly parallel shear flows to the
cross stream mean velocity gradient. This approach was later generalized for
modelling turbulent flows in general geometries. The Reynolds stress tensor is
then related to the mean flow field through
2
ui uj = −2νT Sij + Kδij (8)
3
The last term is often included in the pressure term. In that way a somewhat
modified pressure is obtained, but is a necessary approach in, e.g., simple alge-
braic eddy-viscosity models where K is not available.
The eddy viscosity may be interpreted as a diffusivity coefficient (of momen-
tum) related to the velocity scale V and length scale L of the large energetic
turbulent eddies (νT ∼V L). By applying the eddy-viscosity hypothesis, turbu-
lence modelling reduces to a matter of modelling the eddy viscosity νT .
2.2.1. Algebraic and one-equation models. Models where the eddy
viscosity is completely determined in terms of the local mean flow variables are
referred to as zero-equation or algebraic models. In these models the turbulent
velocity scale is related to the mean flow velocity or vorticity and the length
scale is related to some geometrical length, for example wall distance or wake
thickness. Well known and much used is the Baldwin–Lomax[3] model.
In one-equation models, one of the two turbulence scales (or a combination
of both) is determined from a transport equation. As for algebraic models addi-
tional information from the mean flow field or geometrical measures is needed.
Many one-equation models are based on the transport equation for the turbulent
kinetic energy originally proposed by Prandtl[39]. The one-equation model of
Spalart & Allmaras[58] is based on a transport equation for the eddy viscosity
itself. This model has attracted considerable attention recently and has been
shown to perform reasonably well in typical aeronautical applications.
2.2.2. Two-equation models. The major deficiency in algebraic and one-

equation models is that the turbulence scales need to be related to the global
flow or geometrical scales. Many of the ad hoc corrections and relations present
in such models are caused by this ‘incompleteness’ of the model. A complete
model is characterized by the feature that the Reynolds stresses are completely
determined from the local state of the mean flow and mean turbulence quantities.
No global measures are thus needed such as, e.g., the wall distance, boundary
layer thickness or free stream velocities.
The lowest modelling level that is complete in that sense is the two-equation
modelling approach where the turbulence scales are determined from two trans-
port equations. The natural choice for the turbulent velocity scale is the square
root of the turbulent kinetic energy K and the turbulent length scale is de-
termined from K and the auxiliary quantity, i.e. ε in the K–ε model where
L∼K 3/2 /ε. Auxiliary quantities other than ε have been used, e.g., ω ∼ ε/K and
τ ∼ K/ε together with K.
The exact transport equation for K is given by (6) where the different terms
need to be modelled in terms of known quantities. The production term needs
no further modelling after applying the eddy-viscosity assumption
P = 2νT Sij Sji . (9)
The turbulent diffusion is usually modelled using the gradient diffusion assump-
tion which gives the total diffusion

∂ νT ∂K
D(K) = ν+ (10)
∂xj σK ∂xj
where the diffusivity is related to the turbulent viscosity through the turbulent
Schmidt number σK .
The dissipation rate ε is modelled in form of an additional transport equation
for ε or some other quantity Z related to K and ε, such as Z = K m εn . The
standard structure of the transport equation for Z is
DZ Z
= (CZ1 P − CZ2 ε) + D(Z) (11)
Dt K
The diffusion term usually reads

∂ νT ∂Z
D(Z) = ν+ (12)
∂xj σZ ∂xj
but additional, so called cross diffusion terms, could be added
∂K ∂Z
+CZ3 . (13)
∂xj ∂xj
Such terms result from straightforward transformations between different bases
for Z, e.g., between ε and ω.
2.2. EDDY-VISCOSITY MODELS 7
Finally the eddy viscosity is related to K and Z, for instance through

K2 K
νT = Cµ or νT = (14)
ε ω
in case of a K–ε or a K–ω model respectively.
In wall bounded flows the very near-wall region (y + 50) needs special atten-
tion. Typically, the coefficients in the Z equation and Cµ need to be multiplied
with so called damping functions based on y + , Rey or ReT or a combination of
these nondimensional numbers
√
yuτ y K K2
y ≡
+
Rey ≡ ReT ≡ (15)
ν ν εν

where y is the wall distance and uτ ≡ τw /ρ is the wall friction velocity. There
are, however, models that are defined completely without any near-wall damping
functions, e.g. the Wilcox[66] standard K–ω model.
Another alternative is the use of wall function boundary condition which
means that the boundary conditions are set outside of the viscous sub-layer. The
very near-wall stiffness is then avoided and a considerable amount of computa-
tional grid points and computational effort are saved. However, the numerical
solutions are sensitive to the distance to the first computational point above the
surface and the law of the wall does not always hold, e.g., near boundary layer
separation.
The standard two-equation models are insensitive to rotation. The K–Z
equations and the stress-strain relation contain no explicit dependence on the
rotation rate tensor Ωij ≡ (Ui,j − Uj,i )/2 and thus the equations expressed in a
rotating coordinate system are identical to the ones written in an inertial system.
Also the effect of local rotation is lost, that could be related to the absence of
rotation near stagnation points and the excessive rate of rotation within vortices.
This is a major deficit of standard eddy-viscosity two-equation models and one
of the reasons for the fact that such models fail in predicting the correct rate of
production in complicated turbulent flows.
Moreover, even in the simplest possible equilibrium shear flows, or in the
log-layer of boundary layers, the eddy-viscosity assumption completely fails in
predicting all but one of the components in the Reynolds stress anisotropy ten-
sor. However, the well predicted component is the only one that is of primarly
importance in thin shear layers.
Another deficit is that two-equation eddy-viscosity models may produce un-
physical values of the Reynolds stresses in nonequilibrium flows where the pro-
duction to dissipation ratio is large, P/ε1. Any normal component of the
Reynolds stress tensor must be positive u2α ≥ 0 regardless of the choice of unit
(α)
vector ei . This is the so called realizability of turbulence models and gives
that the components of the anisotropy tensor are bounded. The Reynolds stress
anisotropy for the K–ε model may be expressed as aij = −2Cµ (K/ε)Sij . In
rapidly strained turbulence the term (K/ε)Sij may become large, which ob-
viously may result in unphysical values of the anisotropy. This could lead to
unphysical growth of turbulence in regions where low-level free-stream turbu-
lence interacts with flow distortions such as around stagnation points and near
shocks in compressible flows.
To avoid unphysical growth of the turbulence in practical computations, the
production is often limited by the dissipation rate P<Cε where C is some suf-
ficiently high number, typically around 10. The problem with this form is that
in rapidly growing turbulence the growth rate should be independent of the dis-
sipation rate. An alternative limiter for the production is described in Paper 6

where P<K Sij Sji , which is based on realizability constraints. A similar effect
is obtained with the Menter SST K–ω model[32], where the Bradshaw assump-
tion is used. Here, the shear anisotropy a12 is limited to −0.3 for P/ε ratios
greater than unity, which has been observed in adverse pressure gradient bound-
ary layers and also in homogeneous shear flows. This model has been found to
perform well for typical aeronautical applications.
However, in many engineering flows the boundary layers are attached with-
out strong curvature or rotational effects, and thus the eddy-viscosity assumption
has been, and still is, relatively successful. The absolute majority of turbulence
models used in industrial CFD methods are based on the eddy-viscosity assump-
tion.
2.3. Reynolds stress models

Eddy-viscosity models perform reasonably well in attached boundary layer
flows as long as only one component of the Reynolds stress tensor is of significant
importance. In these cases one could consider the eddy viscosity as a representa-
tive for the significant Reynolds stress component. However, if the flow becomes
more complicated the eddy-viscosity assumption fails and, thus, there is not
much hope for a more general validity of the eddy-viscosity approach.
Models that are based on the exact transport equation for the Reynolds
stress tensor (5) contain much more of the fluid mechanics needed in compli-
cated turbulent flows. The advection and diffusion terms account for transport
effects for the individual Reynolds stress components, whereas eddy-viscosity
two-equation models only consider such effects for the trace of the Reynolds
stress tensor. These effects are of significant importance for the development of
the trace of the Reynolds stress tensor, i.e., the kinetic energy, as well as the
individual components in nonequilibrium flows where P/ε1.
More important are the different local source terms. No modelling is needed
for the production term, Pij , which is a significant improvement compared to
the modelling of the trace of the production present in eddy-viscosity models de-
scribed in the previous section. The intercomponent energy transfer, represented
by Πij needs to be modelled, but already the simplest possible models represent
2.3. REYNOLDS STRESS MODELS 9
improvements compared to eddy-viscosity two-equation models in which this

effect is not present at all.
2.3.1. Differential Reynolds stress models (DRSM). In differential

Reynolds stress models, or Reynolds stress transport (RST) models, all different
terms in (5) are kept or modelled which results in a transport equation for every
individual Reynolds stress component. In general three-dimensional mean flows
this implies six equations due to symmetry in the Reynolds stress tensor.
The dissipation rate tensor εij is usually decomposed into an isotropic part
and a deviation from that, εij = ε(eij + 2δij /3). First, the total dissipation rate
ε is modelled through a transport equation, similar to the ε equation in the K–ε
models. Also here other alternatives to ε, such as ω or τ exist. The dissipation
rate anisotropy eij is typically explicitly modelled in terms of the Reynolds stress
anisotropy or included into the modelling of the pressure strain rate. Turbulent
diffusion, Dij , may be modelled using the generalized gradient diffusion model
by Daly & Harlow[10].
The focus of the modelling in this context is that of the pressure strain
rate. The classical decomposition into two parts, the rapid and slow parts,
is based on the formal solution of the Poisson equation for the pressure field.
The rapid part responds directly to changes in the mean flow field while the
source term for the slow part does not contain any mean flow field information.
Originally these terms were modelled linearly in the Reynolds stress tensor, see
Rotta[46] and Launder, Reece & Rodi[29]. More recently higher order models
that are nonlinear in the Reynolds stresses have been proposed mainly in order
to satisfy strong realizability, which is not possible for linear models, see Sjögren
& Johansson[56] and Sarkar, Speziale & Gatski[60].
The most general quasi-linear model for the pressure-strain rate and dissi-
pation rate anisotropy eij lumped together reads

Πij 1 P
− eij = − C10 + C11 aij + C2 τ Sij
ε 2 ε

C3 2 C4
+ τ aik Skj + Sik akj − akl Slk δij − τ (aik Ωkj − Ωik akj ) , (16)
2 3 2
for example, see Girimaji[13]. τ = K/ε is the turbulent timescale. Classical

linear models, such as the Launder, Reece & Rodi[29] (LRR) or the simplified, so
called isotropization of production (IP) model (see also Naot[35]) may be written
in this form. Also the Sarkar, Speziale & Gatski[60] (SSG) model linearized
around equilibrium homogeneous shear flows may be expressed in this form, see
for example Gatski & Speziale[12] for linearization of the SSG model. The C
coefficients for these models are given in Table 2.1. By quasi-linear pressure-
strain models we here mean that the model is linear in the Reynolds stress
anisotropy, but may contain the scalar nonlinearity aij P/ε ≡ −τ akl Slk aij .
Table 2.1 The values of the C coefficients for different quasi-linear

pressure-strain models
C10 C11 C2 C3 C4
Original LRR 3.0 0 0.8 1.75 1.31
Recalibrated LRR (W&J) 3.6 0 0.8 2 1.11
Linearized SSG 3.4 1.8 0.36 1.25 0.40
An alternative to using the Reynolds stress transport equations (5) is to re-

formulate the equations in terms of the Reynolds stress anisotropy and the tur-
bulent kinetic energy. The transport equation for the Reynolds stress anisotropy
reads

D aij (a) ui uj ui uj
K − Dij = Pij − P − εij − ε + Πij , (17)
Dt K K
where diffusion of the anisotropy is defined as
(a) Dij ui uj (K)
Dij ≡ − D . (18)
K K2
Inserting the quasi-linear pressure-strain rate model (16) into (17) gives

D aij P ∗

− aik Ω∗kj − Ω∗ik akj
(a)
τ − Dij = A0 A3 + A4 aij + A1 Sij
Dt ε

∗ ∗ 2 ∗
+ A2 aik Skj + Sik akj − akl Slk δij (19)
3
where the strain and rotation rate tensors are normalized by the turbulent
∗
timescale, Sij = τ Sij and Ω∗ij = τ Ωij . The A coefficients are related to the C
coefficients through
C4 3C2 − 4 C3 − 2 2 − C10 −C11 − 2
A0 = − 1, A1 = , A2 = , A3 = , A4 = .
2 3A0 2A0 2A0 2A0
(20)
The transport equation for the Reynolds stress anisotropy may be written
in the following symbolic form
∗
tr(aij ) = fij (akl , Skl , Ω∗kl ) (21)
where tr(aij ) represents the advection and diffusion of the Reynolds stress anisotropy
and fij the production, dissipation and redistribution terms.
2.3.2. Algebraic Reynolds stress models (ARSM). In flows where

the anisotropy varies slowly in time and space, the transport equation for the
Reynolds stress anisotropy tensor is reduced to an implicit algebraic relation.
Also in many inhomogeneous flows of engineering interest the flow is steady and
the advection and diffusion of the Reynolds stress anisotropy may be neglected.
This is equivalent to the assumption by Rodi[42, 43] that the advection and
diffusion of the individual Reynolds stresses scale with those of the turbulent
kinetic energy. The set of transport equations for the Reynolds stress anisotropy
is then reduced to an algebraic equation system, which in the case of quasi-linear
pressure-strain models may be written as

P ∗

0 = A3 + A4 aij + A1 Sij − aik Ω∗kj − Ω∗ik akj
ε

∗ ∗ 2 ∗
+ A2 aik Skj + Sik akj − akl Slk δij . (22)
3
This system is implicit in the Reynold stress anisotropy. It is also nonlinear

in the Reynolds stresses, even with linear modelling of the pressure-strain rate
∗
term, because of the term (P/ε)aij ≡ − akl Slk aij . The set of equations may be
written in the following symbolic form
∗
0 = fij (akl , Skl , Ω∗kl ) . (23)
and may, in principle, be solved for any kind of model for the pressure-strain
rate and dissipation rate tensors.
2.3.3. Explicit algebraic Reynolds stress models (EARSM). The

system of equations (22) has been found to be numerically and computationally
cumbersome since there is no diffusion or damping present in the equations. In
many applications the computational effort has been found to be excessively
large and the benefit of using ARSM instead of the full transport form is then
lost. In a pioneering work of Pope[36] (1975) an explicit form was proposed
for two-dimensional flows, but the approach did not attract significant attention
until the beginning of the 90’s. The work in this area of algebraic Reynolds
stress modelling has been focused on finding explicit expressions. EARSMs are
much more numerically and computationally robust and have been found to be
comparable to standard two-equation models in computational effort.
So far explicit solutions of algebraic Reynolds stress models have been re-
stricted to linear or quasi-linear pressure-strain models where the ARSM equa-
tions can be written as (22). The procedure proposed by Pope[36] to obtain
an explicit form was to avoid the nonlinearity by considering P/ε as an extra
unknown. The resulting linear equation system can then be written as
∗
0 = Lij (akl , Skl , Ω∗kl , P/ε) . (24)
and may, in principal, be solved directly. However, by the inspection of (22)

one realizes that the anisotropy tensor is dependent on only two other tensors,
∗
Sij and Ω∗ij , which can be used to form a complete base for the anisotropy.
Following Spencer & Rivlin[59] and Pope[36] the complete expression consists of
ten tensorially independent groups

a = β1 S

1 1
+ β2 S2 − IIS I + β3 Ω2 − IIΩ I + β4 (SΩ − ΩS)
3 3

2 2
+ β5 S Ω − ΩS2 + β6 SΩ2 + Ω2 S − IV I
3

2
+ β7 S2 Ω2 + Ω2 S2 − V I + β8 SΩS2 − S2 ΩS + β9 ΩSΩ2 − Ω2 SΩ
3
2 2
+ β10 ΩS Ω − Ω S Ω .
2 2
(25)
The β coefficients may be functions of the five independent invariants of S and
Ω, which can be written as
IIS = tr{S2 }, IIΩ = tr{Ω2 }, IIIS = tr{S3 }, IV = tr{SΩ2 }, V = tr{S2 Ω2 }.
(26)
Other scalar parameters may also be involved. For simplicity, we have used
boldface for denoting second-rank tensors and tr{} is the trace. We have also
omitted the ∗ , so that S≡Sij
∗
and Ω ≡ Ω∗ij .
The ARSM equation system (22) may now be solved by mapping every term
onto this base and solve the linear equation system for the β coefficients. The
β coefficients are thus completely determined in terms of the coefficients in the
basic RST model. Pope[36] derived a solution for two-dimensional mean flows,
where only the three terms associated to β1 , β2 and β4 are needed. In three-
dimensional mean flows Taulbee[62] obtained a solution for a specific choice of
ARSM where A2 = 0 and Gatski & Speziale[12] determined the solution for the
general quasi-linear ARSM. Also more general nonlinear models may be mapped
onto the base (25) but the corresponding equation system for the β coefficients
is then no longer linear.
Now, we return to the assumption by Pope of leaving the production to
dissipation ratio as an unknown quantity. The solution may be written in the
following form
∗
aij = Aij (Skl , Ω∗kl , P/ε) . (27)
∗
Remembering that P/ε≡ − aij Sji the following equation must hold for consis-
tency
∗
P/ε = −Aij (Skl , Ω∗kl , P/ε) Sji
∗
. (28)
Models that fulfill this so called consistency condition may be considered as ‘self
consistent’. The condition (28) may be fulfilled implicitly by some iteration
procedure or as a part of the solution process, as Pope[36] and Taulbee[62] sug-
gest. Self consistency may also be obtained explicitly by expanding and solving
(28), which results in a third-order polynomial equation with an exact solution
in two-dimensional mean flows. That approach was taken by Girimaji[13] and
Johansson & Wallin[20]. The latter also extended the solution with an approxi-
mation valid also in three-dimensional mean flows, see Paper 1.
A different approach was taken by Gatski & Speziale[12] where the equi-
librium value for P/ε in homogeneous shear flows was taken as a universal con-
stant. This model is thus only exactly self consistent in equilibrium homogeneous
shear flows. The major motivation behind that assumption was that the basic
(a)
equilibrium assumption of neglecting Daij /Dt and Dij is strictly valid only in
equilibrium flows, and thus it might be considered as inconsistent to extend P/ε
to beyond equilibrium flows. The assumption of a constant P/ε resulted in a
model with the wrong asymptotic behaviour in rapidly distorted flows and that
also may become singular under some conditions. Additional corrections were
therefore needed, see Gatski & Speziale[12] and Speziale & Xu[61].
More recently it has become quite well accepted that the consistency condi-
tion improves the predictive performance and that it is important for avoiding
numerical problems, see e.g. Rumsey, Gatski & Morrison[47], Jongen, Machiels
& Gatski[24] and the analysis of the consequences of the consistency condition by
Jongen & Gatski[25]. Approximate self consistency was obtained in the model
proposed by Rung et al.[48], which performs similarly to fully self-consistent
models.
Extensions of EARSMs to account for anisotropic dissipation rate have been
proposed by Xu & Speziale[68] and extended to inhomogeneous flows by Jongen,
Mompean & Gatski[26]. Some improvements compared to the basic EARSM
was reported for S-duct flow using this composite model.
2.3.4. Nonlinear eddy-viscosity models (NL-EVM). A different fam-

ily of models, that has the same principal form (25) as EARSMs, are the non-
linear eddy-viscosity models (NL-EVM). A standard eddy-viscosity model is re-
called by setting β1 = −2Cµ and β2−10 = 0 in (25). By adding additional terms
from the general constitutive relation (25) and by letting the β coefficients be
functions of the invariants (26) more of the turbulence physics may be captured.
The method is typically to relate each term, or group of terms, to a specific
generic flow case, to find a suitable functional form for the β coefficients and cal-
ibrate the coefficients. Also more analytical considerations, such as realizability,
may be considered when formulating the β-functions in terms of the invariants.
When formulating a NL-EVM one should consider that the components of
the Reynolds stress anisotropy tensor are restricted to order of unity due to
realizability constraints. Eddy-viscosity models with constant Cµ (or constant
β1 ) become unrealizable for large strain rates since the anisotropy is proportional
to the normalized strain rate, aij ∼ σ, where σ is some representative measure
of the normalized deformation rate, e.g. σ ∼ τ ∂Ui /∂xj . Adding higher-
order terms makes this problem worse if the corresponding β coefficients are
constant. For example, second-order terms give aij ∼ σ 2 and third-order
terms aij ∼ σ 3 .
The β coefficients should, therefore, be functions of the strain rate so that

for large strain rates they become inversely proportional to the strain rate to a
power at least equal to the degree of the corresponding term. An example of
such a form is
0 0
β10 β2−4 β5−6
β1 =
1 + β11 σ
, β2−4 = 1 σ2
1 + β2−4
, β5−6 = 1 σ3 , · · ·
1 + β5−6
(29)
where the constants βαβ need to be calibrated. Examples of this kind of model are
Shih, Zhu & Lumley[53] and Craft, Launder & Suga[9]. Self-consistent EARSMs
naturally fulfill this condition and are, thus, realizable in this respect.
2.4. Scalar flux models

In turbulent flows where transport of some additional scalar θ is present
a velocity–scalar correlation ui θ appears in the governing Reynolds averaged
equations. The additional scalar may be the temperature in compressible flows,
species concentrations in combustion flows or pollutant concentration in atmo-
spheric or ocean flows. The Reynolds decomposition of the scalar θ into the
average Θ and the turbulent fluctuation θ reads θ = Θ + θ, and the Reynolds
averaged equation (in constant density flows) becomes

∂Θ ∂Θ ∂ ∂Θ
+ Uj = α − uj θ (30)
∂t ∂xj ∂xj ∂xj
where α is the molecular diffusivity.
The modelling of the scalar flux has many similarities with the modelling
of the Reynolds stresses. The, by far, most used model is the eddy-diffusivity
model or gradient flux model that relates the scalar flux vector to the mean
scalar gradient by an eddy diffusivity αT
∂Θ
ui θ = −αT . (31)
∂xi
The eddy diffusivity is related to the eddy viscosity by αT = νT /ScT where the
Schmidt number (or Prandtl number if Θ is the temperature) is a model constant
or sometimes a function of the ratio between the scalar (thermal) and dynamic
timescales.
In reality the scalar flux vector is not aligned with the mean scalar gradient
vector and, thus, the eddy-diffusivity approach fails in predicting the flux com-
ponents in the other directions. These components are often of the same order
as the component in the gradient direction. However, in thin shear layers, such
as attached boundary layers, these effects are of minor importance for the mean
scalar field and the eddy-diffusivity approach may be used.
In more complicated flow situations higher level modelling is needed. Ana-
logous to Reynolds stress transport modelling, a transport model for the flux
2.4. SCALAR FLUX MODELS 15
vector may symbolically be written as

D ui θ
− Di = Pθi + Πθi − εθi . (32)
Dt
The right-hand side may in general be assumed to be a function of the Reynolds
stress tensor ui uj , the scalar flux vector ui θ, the gradients of the mean velocity
∂Ui /∂xj and scalar ∂Θ/∂xj , the turbulent kinetic energy K and its dissipation
rate ε, and the scalar flux variance Kθ and its dissipation rate εθ . Additional
equations for Kθ and εθ are thus needed together with transport equations for
ui uj and ε.
The production term Pθi is explicit in the primary quantities and needs no
further modelling. The pressure-scalar gradient correlation Πθi and the destruc-
tion term εθi need modelling and are often lumped together[28, 50, 54]. The
diffusion term Di is usually modelled using the Daly & Harlow[10] gradient-
diffusion model.
There is a renewed interest in algebraic models which are obtained from
the transport equations using some equilibrium assumption. The most common
approach is the weak equilibrium assumption, where the advection and diffusion
√
of the normalized scalar flux ui θ/ KKθ is neglected, rather than the scalar flux
itself[1, 2, 16, 45, 52, 54]. A different approach is taken by Shabany & Durbin
[50] where the advection and diffusion of the normalized dispersion tensor Dij ,
defined as ui θ = −Dij (K 2 /ε)∂Θ/∂xj , is neglected.
CHAPTER 3
The proposed modelling approach
Engineering problems are typically characterized by complexity in that dif-

ferent physical effects are present simultaneously and are interacting. Generality
of the turbulence models used for such problems is thus of major importance.
Two-equation eddy-viscosity turbulence models are today widely used in in-
dustrial CFD codes and have been found to be reasonably numerically robust
and computationally efficient. The objective with the present study is to ex-
tend the generality of two-equation models while retaining the computational
efficiency of standard models.
The model, or class of models, derived and evaluated in Paper 1 is the basis
of the present study. Models published by others in this field have similarities
and also differences compared to the present model. In this chapter motivations
of the different choices are given and the different approaches are explained in
some detail.
3.1. The weak equilibrium assumption

The classical definition of equilibrium turbulence is the condition when the
production rate of turbulent energy at the larger scales is balanced by the dissi-
pation rate at the smallest scales. This means that neither the mean flow scales
nor the turbulence scales vary significantly in time or space. This condition can
never exactly be fulfilled since the turbulence in a constant mean-flow field is
always evolving in time. However, the condition may be satisfied in the sense
that the energy produced by the larger scales is approximately balanced by the
dissipation rate at the smaller scales. The energy is said to cascade down from
the large to the small scales. For sufficiently large Reynolds numbers there is
an intermediate range of scales, the inertial subrange for which the net rate of
change is negligible. This is the basis for the Kolmogorov’s universal equilibrium
theory, see ,e.g., Tennekes and Lumley[63]. Let us call this perfect equilibrium.
Another definition of equilibrium is that the evolution of the turbulent scales
should be similar to that of the mean-flow scales. A measure of the ratio between
the mean-flow and turbulence scales is the production to dissipation ratio P/ε or
the normalized deformation rate σ ≡ SK/ε where S is some measure of the mean
flow deformation rate. In such strong equilibrium flows P/ε or σ is constant in
time and space. This is the case in free shear flows after some initial transients
16
3.1. THE WEAK EQUILIBRIUM ASSUMPTION 17
0.0
-0.1
-0.2 hom. shear
-0.3
a12
-0.4 log-layer
-0.5
-0.6
0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ
Figure 3.1 The a12 anisotropy versus strain rate σ for parallel shear
flow. The Johansson & Wallin[20] self-consistent model ( ) com-
pared with fixed P/ε = 1 ( ), with the diffusion model ( ), and
with a standard eddy-viscosity model ( ).
and also in the log-region of wall bounded flows, in the latter only P/ε, not σ,
is constant. In these flows P/ε is of the order of unity.
The assumption introduced by Rodi[42, 43] for obtaining an algebraic rela-
tion for the Reynolds stresses is referred to as the weak equilibrium assumption,
characterized by a Reynolds stress anisotropy tensor that is constant in time and
space, or
D aij (a)
− Dij = 0. (33)
Dt
The relaxation of some nonequilibrium initial state where P/ε > 1 is character-
ized by a rapid and slow time scale[15]. The anisotropy relaxes rapidly to some
quasi-equilibrium state for prescribed mean flow and turbulence scales. The
weak equilibrium assumption (33) could not be expected to capture this initial
process. After the initial transient the turbulence scales adjust slowly with the
mean flow where the anisotropy is nearly in local equilibrium. This slow process
could be reasonably well captured using the algebraic Reynolds stress approach
since the transport effects on the turbulence scales are captured through the
transport equations for, e.g., K and ε.
3.1.1. The consistency condition. Since the weak equilibrium assump-

tion (33) may be valid also in flows that depart from the strong equilibrium
condition (for P/ε 1), it is natural to invoke the consistency condition (28).
The effect of the consistency condition is illustrated in Figure 3.1 where the
shear component a12 of the Reynolds
stress anisotropy tensor is plotted versus
the normalized strain rate σ ≡ IIS /2 in parallel flow (IIS = −IIΩ ). It is
18 3. THE PROPOSED MODELLING APPROACH
25
20
15
K
10
0
0 2 4 6 8 10
T
Figure 3.2 Time evolution of the turbulent kinetic energy in rapidly

sheared homogeneous flow, SK/ε = 50. Eddy-viscosity model ( ),
the Johansson & Wallin[20] self-consistent EARSM ( ) and the
Gatski & Speziale[12] EARSM ( ) compared with RDT ( ❛).
clearly seen that the consistency condition makes a significant difference away
from the equilibrium choice for P/ε. The differences in this σ-range between the
two EARSM approaches are, however, small compared to the difference between
EARSM and the eddy-viscosity model that gives unrealizable anisotropies for
large strain rates.
The asymptotic behaviour for large strain rates in parallel flow can be in-
vestigated by letting σ → ∞. The production to dissipation ratio then becomes
P/ε ∼ σ and β1 ∼ Cµeff ∼ 1/σ. The asymptotic behaviour is consistent with the
asymptotic characteristics of homogeneous shear flow. With a model assumption
of a constant production to dissipation ratio the wrong asymptote, β1 ∼ 1/σ 2 ,
is obtained, as also noticed by Speziale & Xu[61].
To illustrate the behaviour for large shear rates the model is tested in ho-
mogeneous shear flow at high initial shear rate (SK/ε = 50), see figure 3.2. As
expected, the eddy-viscosity model fails, giving a far too high growth rate. In-
teresting to note is that also the Gatski & Speziale[12] model with fixed P/ε fails
because of the wrong asymptotic behaviour. This flow is a case where one should
expect differences between the algebraic approach and the full differential model
due to the fact that the anisotropies undergo a temporal evolution (∂aij /∂t = 0)
in the development towards an asymptotic state. Moreover, the LRR model
gives quite poor predictions of this case when used in a differential form. The
very good predictions of the present EARSM can thus be regarded as a bit fortu-
itous. Nevertheless, the self-consistent approach gives a model with the correct
asymptotic behaviour, which is a prerequisite for reasonable predictions in the
limit of high shear. It is important to make clear that the proposed model is not
intended for these extreme high shear rates and the normal stress components
are not as well predicted as the turbulent kinetic energy. It is, however, an im-
portant step towards a more general engineering model that the model is able
to give reasonable results also in extreme flow cases.
3.1.2. Slowly distorted turbulence. In slowly distorted turbulence where

P/ε 1, such as in the outer parts of boundary layers and free shears or at
the centre of channel, wake, or jet flows, the source terms are weak compared to
the advection or diffusion terms. In such cases the weak equilibrium assumption
does not hold and (E)ARSMs that do not consider this behave worse than eddy-
viscosity models with constant Cµ . In many flows, such as wall bounded flows,
this deficit does not influence the mean flow much, whereas it is more important
in some other flows like free shear flows, jets, and wakes.
The slow distortion problem was noticed by Taulbee[62], who suggested an
alternative method of obtaining the algebraic relation based on an approximation
of the usually neglected advection. Improvements were reported in cases where
the advection was important. Another alternative is proposed in Paper 1 where
an approximation of the turbulent diffusion was included, which improved the
behaviour in cases with nonnegligible diffusion. These approaches are somewhat
unsatisfying because they do both give effects in both advection dominated and
diffusion dominated flows although they both are motivated from only one of
these effects. These ad hoc corrections do, however, give some of the wanted
effects, but more work is needed on this specific topic. The effect of the diffusion
approximation in Paper 1 is shown in Figure 3.1 where we can see that the
corrected model behaves much like the eddy-viscosity model for small strain
rates.
3.1.3. Streamline curvature effects. Turbulent flows over curved sur-

faces, near stagnation and separation points, in vortices and in rotating frames
of reference are all affected by streamline curvature effects. Strong curvature
effects form a major cornerstone problem also at the Reynolds stress transport
modelling level, and pressure-strain rate models that are able to accurately cap-
ture strongly rotating turbulence are rare. In more moderate situations the
SSG[60] model, and derivations thereof, show rather good behaviour in rotat-
ing flows such as rotating homogeneous shear flows, see Gatski & Speziale[12].
Standard eddy-viscosity models fail in describing effects of local as well as global
rotation and thus completely fail.
In Paper 3, a schematic wingtip vortex was studied where the SSG model
behaves qualitatively correct while the standard eddy-viscosity K–ε model highly
overestimates the turbulence diffusion. More surprisingly, it was found that also
the EARSM based on the SSG model completely failed. The reason for this is
that the advection term in the anisotropy transport equation (17) is significant
and cannot be neglected in strongly curved flows, also noticed by Rumsey, Gatski
& Morrison[47].
A schematic vortex, where the axial ẑ and azimuthal θ̂ directions are ho-
mogeneous, may be used for illustrating this mechanism. In the homogeneous
directions all scalar quantities are constant and from continuity the mean ve-
locity has no radial r̂ component and thus the advection D/Dt is zero for any
scalar (neglecting the small ∂/∂t term). This means that the advection of the
anisotropy invariants, IIa ≡ aij aji and IIIa ≡ aij ajk aki , are zero. However, the
principal axes of the anisotropy tensor varies in the azimuthal direction and thus
also the individual components aij varies and Daij /Dt = 0.
Advection of the anisotropy tensor (or any symmetric tensor) may be written
in the following form
D aij ∂aij ε (r) (r)

= + G (aij ) − aik Ωkj − Ωik akj (34)

Dt ∂t K
where G contains spatial derivatives evaluated in a curvilinear coordinate system,

(r)
see Sjögren[55]. Ωij is an antisymmetric tensor that represents the rotation
rate of the coordinate system metrics following the mean flow. In a Cartesian
(r)
coordinate system Ωij = 0 and G = Uk ∂/∂xk . If the generic vortex is expressed
in a streamline based coordinate system (cylindrical) the rotation tensor reads
V (r)
Ω(r) = θ̂r̂ − r̂θ̂ (35)

r
and the spatial derivative of the anisotropy components becomes zero G = 0.

The advection term may thus be exactly included in the EARSM formulation
(r)
for this case. Here it becomes clear that Ωij represents the local rotation rate
of the anisotropy tensor following the streamline.
The last term in (34) may be fully accounted for and included into the
EARSM solution without increasing the complexity of the solution. This was
done in Paper 3 for the generic vortex and it was found that the predictions
using the extended EARSM were much improved. The results were in line with
the corresponding full RST models. In general three-dimensional mean flows
the problem is to determine a streamline-based coordinate system for obtaining
(r)
Ωij . Girimaji[14] proposed to use the local acceleration vector, a, and the
rate of change of that, ȧ, following the mean flow in obtaining the base of the
curvilinear coordinate system. The algebra becomes very tedious in general
three-dimensional mean flows involving also gradients of the coordinate system
metrics (i.e. gradients of a and ȧ) and such a method is thus of limited practical
use.
A more simple and straightforward alternative could be to directly formulate
(r) (r)
the Ωij tensor in terms of the a and ȧ vectors by recalling that Ωij represents
the rotation rate of the coordinate system metrics following the mean flow. The
local rotation vector may be obtained as

a × ȧ
ω= (36)
a2
and the rotation tensor reads
(r) ȧi aj − ai ȧj
Ωij = −eijk ωk = . (37)
ak ak
This new approach is identical to the more complete relation above in circular
flows, such as the generic vortex. It also behaves well in linearly accelerated or
decelerated flows, such as on the symmetry line of a stagnation flow since a and
ȧ have the same direction and thus a × ȧ = 0. This approach has not yet been
tested in general three-dimensional mean flows.
The acceleration vector and its rate of change fulfill Galilean invariance, that
is independency of solid-body motion of the frame of reference. However, any
incompressible flow field should also be independent of a superimposed solid-
body constant acceleration, according to Spalart & Speziale[57], except for a
modified pressure field. The proposed modification must thus be used with
caution in accelerated frames of reference. Extensions of EARSMs for including
approximations of the usually neglected transport terms could never be expected
to be completely general, but could anyway be motivated by improved model
performance in a reasonably wide class of flows.
If the curvature effects are smaller the advection term is of lesser importance
and the standard EARSM approach may be used successfully. The stabilizing
and destabilizing effects in curved wall bounded flows or in a rotating channel
are naturally incorporated in also the standard EARSM. Launder[31] stated that
cubic terms are needed in nonlinear eddy-viscosity models for capturing these
effects, but, in the proposed EARSM these effects are actually captured already
with the linear term because of the rotation rate dependency in the β1 (or Cµeff )
coefficient.
3.1.4. Pressure-strain rate model. The method to derive an EARSM
described in Paper 1 requires that the pressure-strain rate model is quasi-linear,
which means that it can be written on the form given in (16). The linearized
SSG and the LRR models may be written on this form and are thus basically
very similar. However, the two models behave rather differently when used in a
full RST model and the SSG model has been found to perform better than the
LRR in a number of situations.
The EARSM proposed in Paper 1 was based on a somewhat recalibrated
LRR model for the pressure-strain rate which was motivated by the simplifica-
tion in three-dimensional mean flows, as compared to the general expression.
With this recalibrated LRR model five of the ten tensor groups in (25) vanish.
Moreover, the resulting EARSM is guaranteed not to become singular for any
parameter choice which cannot be proven for the general case. The proposed
EARSM in Paper 1 was demonstrated to give reasonable predictions in different
Table 3.1 The values of the A-coefficients for different quasi-linear

pressure-strain models
A0 A1 A2 A3 A4
Original LRR −0.34 1.54 0.37 1.45 2.89
Recalibrated LRR (W&J) −0.44 1.20 0 1.80 2.25
Linearized SSG −0.80 1.22 0.47 0.88 2.37
flows and the choice of the recalibrated LRR model does not significally degener-
ates the results. However, in Paper 3 the wing tip vortex was significally better
predicted using the SSG based EARSM rather than the LRR based EARSM.
The reason for the LRR based EARSM to behave similarly to SSG in many
cases is mainly because of the recalibration of the original LRR model. The
value of c2 in the rapid pressure–strain model was originally suggested to be
0.4 by Launder et al.[29], but more recent studies have suggested a higher value
close to 5/9[30, 51]. By setting c2 = 5/9 one obtains the simplification in three-
dimensional mean flows. The Rotta coefficient, c1 , was originally set to 1.5,
but also that has more recently been recalibrated and is here set to 1.8. This
recalibration gives a model that behaves well in the log-layer in wall bounded
flows as well as in homogeneous shear flows. The normal components in the
streamwise and spanwise directions, a11 and a33 , are, however, predicted better
using the SSG model. That has been found to be of minor importance since
these components, in contrast to a12 and a22 , do not influence the mean flow or
the turbulent scale equations in parallel flows.
The full quasi-linear RST model may be written in terms of five model coef-
ficients, A0−4 in (19). In the corresponding ARSM (22) one of these coefficients,
A0 , may be eliminated which means that the ARSM approximations of differ-
ent RST models may be identical. The A-coefficients are listed in Table 3.1.
Comparing the A1−4 -coefficients that defines the EARSM, one can see that the
A1 and A4 coefficients are very similar between the recalibrated LRR and the
linearized SSG. These coefficients have the strongest impact on the a12 compo-
nent in parallel flows and thus the two models behave similarly. More important
is that the A0 -coefficient differs by almost a factor of two and since that fac-
tor multiplies the complete right-hand side in the transport equation (19) the
RST models behave very differently in nonequilibrium flows. This is also the
case when some part of the advection or diffusion terms in (19) is included into
the EARSM, which is the case in the vortex computations in Paper 3. The
A0 -coefficient then enters into the EARSM and may cause important differences
that explain the differences between the EARSM results based on LRR and SSG.
Actually, in Paper 3, the vortex was recomputed using the LRR-based EARSM
but with the A0 -coefficient changed to that of the SSG model resulting in much
closer agreement with the SSG based EARSM.
4.0 4.0
rot = 0 rot = 1/4
K
K
2.0 2.0
0.0 0.0
0.0 5.0 10.0 0.0 5.0 10.0
T T
4.0 4.0
rot = 1/2 rot = -1/2
K
2.0 2.0
0.0 0.0
0.0 5.0 10.0 0.0 5.0 10.0
T T
Figure 3.3 Computed evolution of the turbulent kinetic energy K in

rotating homogeneous shear flow compared to large eddy simulation
( ❛), Bardina et al.[4]. Eddy-viscosity model ( ), Wallin & Johans-
son basic EARSM ( ), modified EARSM with A0 = −0.8, ( )
and A0 → ∞ ( ), and EARSM based on SSG ( ). SK/ε = 3.4
initially.
Rotating homogeneous shear flow is another example in which the A0 -coeffi-

cient may cause differences, see Figure 3.3. The proposed EARSM was computed
using different values for the A0 -coefficient, the basic model with A0 = −0.44,
the modified model with A0 = −0.8 (same as for SSG) and the model without
advection correction (A0 → ∞). If was only the modified EARSM with A0 =
−0.8 that gives reasonable predictions comparable to EARSM based on SSG.
To conclude, if it is based on the recalibrated LRR the EARSM is simplified
in three-dimensional mean flows without major deficits in model predictions
compared to EARSM based on SSG. However, the A0 -coefficient needs to be
recalibrated if some part of the advection or diffusion terms is included into the
EARSM. The full three-dimensional form of the EARSM has been found to be
important in fully developed pipe flow rotating around the symmetry axis in
order to capture the secondary swirl flow, see Paper 1.
8.0
7.0
6.0
5.0
u iu j 4.0
3.0
2.0
1.0
0.0
-1.0
0. 20. 40. 60. 80. 100. 120. 140.
y+
Figure 3.4 The Reynolds stresses in channel flow. Comparison of

the Wallin & Johansson near-wall correct EARSM ( ) with DNS
data (symbols) of Moser et al.[33]. From top to bottom: uu, ww, vv
and uv components. The predicted Reynolds stresses were evaluated
by use of the DNS data for the mean flow, K and ε fields.
3.2. Near-wall treatment

The near-wall region usually needs special treatment in turbulence models
through the use of wall-damping functions. This is the case also for EARSMs, but
the situation is somewhat better compared to standard eddy-viscosity models.
The very near-wall behaviour is studied by using channel DNS data of Moser,
Kim & Mansour[33] at Reδ ≈ 7800. The mean velocity, K and ε profiles obtained
from the DNS data have been used to compute the modelled anisotropy, which
is compared to the anisotropy determined directly from the DNS data.
The modelled a12 anisotropy without any near-wall corrections has been
computed from the channel DNS data. It was found in Paper 1 that a12 is
nearly constant as the wall is approached while DNS data exhibit a behaviour
similar to an exponential decay. The obvious choice of ‘wall damping function’
is of van Driest type f1 = 1 − exp (−y ∗ /A+ ), which gives almost perfect agree-
ment with the DNS data. A standard eddy-viscosity K–ε model, without any
near-wall damping function, gives strongly negative values for a12 in the buffer
layer, almost down to −2. Note that this is well outside the range of physically
realizable values, that are limited to be between ±1. The K–ε model cannot
be correctly damped towards the wall as easily as the EARSM, which is much
better suited to be integrated down to the wall. This is due to the correct as-
ymptotic model behaviour for large strain rates, giving a balanced a12 anisotropy
(cf. figure 3.1).
3.2. NEAR-WALL TREATMENT 25
30.0
20.0
10.0
0.0
0.00 0.25 0.50 0.75
U
Figure 3.5 The velocity profile for an adverse pressure gradi-

ent boundary layer (U∞ ∼ x−0.25 ). Computations with standard
Wilcox[66] K–ω ( ), Chien[7] K–ε ( ), EARSM based on K–ω
and y + ( ) or y ∗ ( ) and the Hanjalić[17] RST model ( ).
Comparisons with DNS data ( ❛), (M. Skote, personal communication)
Also the normal components of the Reynolds stress tensor are fitted to the
DNS data by applying the same damping function f1 to the β2 and β4 coefficients
and also considering the correct anisotropies at the wall, for example a22 = −2/3
at the wall. All Reynolds stress components are, thus, well fitted to the DNS
data, see Figure 3.4.
The damping function, f1 , is formulated in terms of y ∗ defined as y + ≡ yuτ /ν
in the original van Driest function. The scaling with the local wall skin friction
is, however, not valid in flows near separation and√reattachment and the y ∗ is
defined in terms of the Reynolds number Rey ≡ Ky/ν so that y ∗ ≈ y + for
y + 100 in zero-pressure gradient boundary layers. The importance of the y ∗
choice is illustrated in Figure 3.5 where the models based on y + fail (the Chien
K–ε model uses damping functions based on y + ). For this case it is obvious that
the y + scaling degenerates the model performance and should be avoided.
An alternative to the asymptotically correct near-wall treatment is presented
in Paper 6 in an approach without any damping functions or wall distance de-
pendence. That is possible if the EARSM is based on the Wilcox[66] K–ω model
or derivates of that, such as the Menter BSL model[32]. These models reproduce
the log-layer without near-wall corrections and, thus, the EARSM approach may
be invoked without the wall damping function f1 . Obviously this model does
not reproduce the correct near-wall behaviour for the Reynolds stresses, but that
may be of lesser importance in high Reynolds number flows and this approach
is an alternative that somewhat simplifies the formulation.
3.3. Numerical behaviour and some applications

The major argument for EARSMs is to increase the validity and generality
of two-equation models without significantly increased computational cost, so
the numerical treatment is one of the cornerstones in this study in spite of the
relative small attention to this in the attached papers.
3.3.1. Effective eddy viscosity. The Reynolds stresses expressed by the

EARSM may be written in terms of an effective eddy viscosity νT and an extra
(ex)
anisotropy aij
2 ∗ (ex)
ui uj = Kδij − 2νT Sij + Kaij (38)
3
where the effective turbulent viscosity reads
1
νT = − (β1 + IIΩ β6 ) Kτ (39)
2
and the extra anisotropy

1
a(ex)
= β3 Ω − IIΩ I + β4 (SΩ − ΩS)
2
3

2
+ β6 SΩ + Ω S − IIΩ S − IV I + β9 ΩSΩ2 − Ω2 SΩ .
2 2
(40)
3
The β-coefficients are given in Paper 1.
This formulation is identical to that in (25) and the reason for invoking the
eddy viscosity is that many CFD solvers already have different eddy-viscosity
two-equation models implemented so the model implementation becomes easier
with a formulation in terms of an eddy viscosity. It has also been found that the
extra anisotropy may be treated in a fully explicit way without degenerating the
overall performance of the solution process.
3.3.2. Positivity. It is well known that positivity of the turbulence vari-

ables (K, ε) at any time during iteration to a steady-state solution is crucial. Ad
hoc methods, like a lower limit of the turbulent variables, often prevent the con-
vergence or cause divergence of the iterative process which means that positivity
must be maintained by the numerical scheme. First, the spatial scheme must
suppress spurious oscillations that may give negative values, which is obtained
by imposing TVD-like conditions[23]. Then the time integration method, or it-
erative process, must guarantee positivity. The CFD solver EURANUS[41] used
in this study uses explicit pseudotime marching to steady state, so the positive
time integration described here will be focused on explicit methods.
In Paper 7 a novel approach is suggested for updating the turbulence vari-
ables guaranteeing positivity. The approach is based on an estimate of the spec-
tral radius of the complete turbulent equations and produces an underrelaxed
3.3. NUMERICAL BEHAVIOUR AND SOME APPLICATIONS 27
update of the turbulence variables q. This may be written in the following form
∆tR(q)
q n+1 = q n + . (41)
1 − min ∆tR(q) /q n , 0
The underrelaxation depends on the local residual R(q) = ∂q/∂t and is significant
only in those regions where the residual is large compared to the variable itself
and is only active if the variable is decreasing. The method does not affect the
asymptotic convergence rate and is most important initially.
Multigrid methods are important for convergence acceleration, especially for
explicit methods, and are considered in Paper 7. The update of the turbulence
variables from the prolongated correction may result in spurious oscillations that
may cause negative values locally. Also here the method for positive updating
(41) is used.
3.3.3. Compressible flow applications. A straightforward extension of

the proposed EARSM for compressible mean flow effects is described in Paper 1.
Compressible effects due to turbulent fluctuations have not been considered. The
extended EARSM was tested on a supersonic boundary layer at Mach 5 inter-
acting with an impinging shock of different strengths (Schülein, Krogmann, &
Stanewsky[49]). The proposed EARSM was used with both the Chien[7] K–ε and
the Wilcox[66] K–ω models as platforms and substantial differences between the
baseline two-equation models were found. The EARSM approach significantly
improved the prediction of the size of the separation zone over the correspond-
ing eddy-viscosity models, which severely underpredicted the separation length.
Also the influence of different shock strengths was accurately captured, rang-
ing from almost separated for the weakest shock to a large separation region
for the strongest shock. The Chien K–ε models, both the eddy-viscosity and
the EARSM version, severely overpredicted the skin friction coefficient in the
reattachment region. The K–ω models behaved much better in that respect.
Detailed results are reported in Paper 2.
Figure 3.6 shows the computational results for the two-dimensional RAE2822
aerofoil profile using the Wallin & Johansson EARSM compared to the Wilcox[67]
K–ω model. The EARSM approach clearly improves the prediction of the posi-
tion of the shock and the results are very much in line with differential Reynolds
stress computations by Hellström, Davidsson & Rizzi[18] for the same conditions
and geometry.
Computations are made where the damping function f1 in the EARSM is
formulated in terms of y + as well as y ∗ and the figure shows no major differences
between these approaches except in the separated region where the y + formu-
lation gives a somewhat larger negative skin friction. The figure also shows a
computation using the proposed EARSM, together with the Wilcox[66] K–ω
model, without any damping functions whatsoever. Even if this model is unable
to reproduce the correct near-wall behaviour for the turbulence quantities, the
-1.5
-1.0
-0.5
Cp
0.0
0.5
1.0
0.0 0.2 0.4 0.6 0.8 1.0
x/c
0.010
0.006
Cf
0.002
-0.002
-0.006
0.0 0.2 0.4 0.6 0.8 1.0
x/c
Figure 3.6 Wall pressure and skin friction coefficients for the
RAE2822 wing profile (M = 0.754, α = 2.57◦ and Re = 6.2 × 106 ).
Predictions using Wilcox[67] K–ω ( ) and the Wallin & Johansson
EARSM based on K–ω with damping function based on y + ( ),
y∗ ( ) or without any damping functions ( ), compared to ex-
perimental data ( ❛) by Cook et al.[8]. The geometry is the measured
one including a camber correction.
mean velocity profiles are well reproduced and there are no major differences
between the EARSM with or without damping functions.
The convergence history is shown in figure 3.7 where we can see that there
are no major differences in convergence rate between the different approaches.
Actually, the proposed EARSM converges to a somewhat lower residual than the
corresponding eddy-viscosity model for this case. The numerical parameters are
the same for these two cases and the computational time for the 5000 iteration
steps is 6% higher for the EARSM computation. This case was found not to
be completely numerically stationary which results in the residual ‘hanging’ as
for the K–ω model. The fluctuations are, however, very small and could not be
seen in the solution. In other cases without separation the convergence curves
are even closer to each other, and the convergence rates are in general faster
than for the case shown in figure 3.7.
3.4. TIME-DEPENDENT TURBULENCE 29
0.0
-2.0
rho_rms
-4.0
-6.0
0. 2000. 4000. 6000.
iter
Figure 3.7 Convergence history for the RAE2822 wing profile.

Wilcox[67] K–ω ( ) compared with the Wallin & Johansson
EARSM based on K–ω ( ). The computational time is increased
by 6% by using EARSM. Three levels of full multigrid is used.
A further example with a three-dimensional transonic supercritical wing was

computed using the standard Wilcox[66] K–ω model and the EARSM based on
that, see Paper 6. Again, the predicted shock position was found to be improved
compared to the eddy-viscosity model. This case illustrates the benefit of the
proposed model which is also viable in expensive three-dimensional cases, here
with almost one million grid points. Most importantly, there is no substan-
tial increase in the computational cost compared to the, in many cases, robust
standard K–ω model.
3.4. Time-dependent turbulence

The mean flow is considered as unsteady when the time scale of the unsteadi-
ness is much larger than the characteristic integral time scale of the turbulence.
In that case the turbulence energy spectrum is well separated from the unsteadi-
ness. Thus, the turbulence may be modelled while the mean flow unsteadiness
is left to be resolved in the unsteady RANS solution. This is known as the
quasi-steady approach where a standard RANS turbulence model, with the time
derivatives included, may be used. No additional modification of the turbulence
model is in principal needed due to the unsteadiness.
If there is no clear separation between the turbulence scales and the unsteadi-
ness, then standard RANS turbulence models should not be used. Therefore this
class of problems needs to be tackled using large eddy simulations (LES) where
subgrid-scale models are used to represent the unresolved stresses. The Reynolds
averaged and the filtered Navier-Stokes equations are in principal similar, but
the unknown correlation ui uj is modelled in completely different ways in RANS
Table 3.2 Computed reduced frequency for unsteady transonic air-

foil flow compared to experiment by McDevitt[34].
model EARSM K–ω B–L Exp.

reduced frequency 0.493 0.441 0.461 ∼ 0.49
and LES. A RANS model could not be used as a subgrid-scale model, mainly
because a RANS model is unable to capture the decreased subgrid-scale energy
with decreased filter width. Actually, it gives the opposite trend of increased
eddy viscosity.
Unsteady turbulent flows are often by nature, or forced to be, periodic and
the mean is thus conveniently defined as a phase average. The flow variables
are decomposed into three parts, φ = φ + φ” + φ, where φ is the time-averaged
value, φ” is the periodic component and φ the turbulent fluctuations. The phase-
averaged variable is then defined as φ ≡ φ + φ”.
In the quasi-steady approach, RANS turbulence models will be used without
any explicit modifications due to the unsteadiness. However, the unsteadiness
of the mean flow poses some additional requirements on the turbulence model,
or will emphasize existing requirements on models for steady flows even more.
The following general requirements can thus be identified: (i) No y + or log-law
dependency, (ii) correct near-wall asymptotic behavior, (iii) good prediction of
nonequilibrium turbulence and (iv) good prediction of boundary layer separation.
The Wallin & Johansson EARSM is thus a good candidate also for modelling
unsteady turbulent flows.
The periodic self-excited turbulent flow around an 18-percent thick circular-
arc airfoil at transonic speeds has been computed in Paper 4 using three different
turbulence models: the Baldwin & Lomax (B–L) algebraic model[3], the Wilcox
K −ω model[67], and the Wallin & Johansson EARSM. The EARSM predicts the
unsteadiness due to shock and boundary layer separation better than the eddy-
viscosity models. The EARSM predicts the frequency of self-exited unsteady
transonic flow to high accuracy, see Table 3.2, and the motion of the shock
agrees well with the experimental data.
3.5. Passive scalar modelling

In Paper 5 an explicit algebraic relation for the scalar flux is proposed. This
is intended to be used in an explicit algebraic approximation approach for the
transport equations for the Reynolds stress tensor and scalar flux vector. The
relation for the scalar flux vector is derived by neglecting the advection and
√
diffusion of the normalized scalar flux ui θ/ KKθ . This gives in general an
implicit, nonlinear set of algebraic equations. The nonlinearity in the algebraic
equations for the normalized scalar fluxes may be eliminated directly by using a
nonlinear term in the model of the pressure scalar-gradient correlation and the
3.5. PASSIVE SCALAR MODELLING 31
Table 3.3 Predicted scalar fluxes in homogeneous shear flow using

the WWJ model proposed in Paper 5 compared to the DNS data of
Rogers et al.[44]. Case 1, 2 and 3 corresponds to a scalar gradient in
the streamwise, cross-stream and spanwise directions respectively.
Case 1 Case 2 Case 3

uθ vθ uθ vθ wθ
DNS data −2.41 0.45 0.94 −0.36 −0.67
WWJ model −2.05 0.42 1.13 −0.46 −0.70
−2
0 0.2 0.4 0.6 0.8 1
η
Figure 3.8 Predicted scalar fluxes in the channel flow using the
WWJ model ( ) compared to DNS data for −uθ (∗) and vθ (◦) of
Wikström et al.[64]. The predicted fluxes were evaluated from (42)
by use of the DNS data.
destruction and thus results in an algebraically simple model for both two- and
three-dimensional mean flows. The resulting model may be written as
K ∂Θ
ui θ = −A−1
ij uj uk (42)
ε ∂xk
where the tensor A−1 ij is an explicit function of the normalized mean velocity
gradients τ ∂Ui /∂xj , the production to dissipation ratio P/ε and the ratio of the
scalar to dynamical timescales r.
The proposed model is, however, not complete since transport equations for
the scalar variance θ2 ≡ 2Kθ and its dissipation rate εθ are needed for obtaining
the timescale ratio r. The modelling of these equations is not considered in
Paper 5. These issues are investigated to some extent in Johansson & Wikström
[21].
The performance of the present model is investigated in three different flow
situations in Paper 5. These are homogeneous shear flow with an imposed mean
0.06
0.04
0.02
−0.02
−0.04
−0.06
−0.08
−0.1
0 0.5 1 1.5 2 2.5
η
Figure 3.9 Predicted scalar fluxes in the cylinder wake using the
WWJ model ( ) compared to experimental data for −uθ (∗) and
vθ (◦) of Wikström et al.[65]. The predicted fluxes were evaluated
from (42) by use of the experimental data.
scalar gradient (Table 3.3), turbulent channel flow (Figure 3.8), and the flow field
downstream a heated cylinder (Figure 3.9). DNS data are used to analyse the
passive scalar flux in the homogeneous shear and channel flow cases, and exper-
imental data are used in the case of the heated cylinder wake. Good predictions
in all three cases are demonstrated. It is interesting to note that, both for the
channel flow and the heated cylinder wake, uθ and vθ are of the same order
of magnitude. This is well captured by the EASFM, whereas eddy-diffusivity
models would predict a zero uθ.
CHAPTER 4
Conclusions and outlook
The ultimate goal in turbulence modelling is to obtain a universally valid

model. Since that is, by far, out of reach, a more realistic goal must be set
and could be what Pope[37] referes to as an optimal turbulence model. This he
defines as the best particular model within each class of models. For a given
such class of models, the validity and generality should be increased to the limit
of the applicability of that class. In this effort one really needs to be aware of
the physical limitations for different types of models.
This work fits into the efforts to increase the validity and generality of two-
equation models during the last decade by abandoning the eddy-viscosity hy-
pothesis. Bradshaw[6] writes ”Using eddy viscosity is like keeping one foot on
the bottom when learning to swim – a restriction, but helpful if one does not
venture into deep water”, an argument that can be used both for keeping and
leaving the eddy-viscosity hypothesis. The improvements result in more general
constitutive relations between the second moments (ui uj and ui θ) and the mean
flow.
The proposed models are the results of careful and well-motivated approx-
imations of the corresponding second-order transport model equations. The
major advantage of this approach compared to more empirical nonlinear exten-
sions to eddy-viscosity models is that the model coefficients and behaviour of the
full second-order model is inherited in the algebraic form. This means that it is
possible to estimate the validity of the model prediction based on the expected
behaviour of the full second-order model and the validity of the approximations.
Empirical extensions of eddy-viscosity models have many degrees of freedom and
may be adjusted to a large number of different test cases. These kind of models
could, however, not with confidence be extended outside of the range of the set
of validation and calibration cases used for finding appropriate functional forms
and calibration of the different terms.
The EARSM proposed in Paper 1 is derived from the recalibrated LRR
model. The method, at least for two-dimensional mean flows, are general and
the EARSM could in principal be derived also from the linearized SSG model.
There is actually a strong argument for deriving the EARSM from the SSG
model and that is that the SSG model has previously been found to perform
better than LRR in a number of cases, especially in nonequilibrium and rotating
flows. The LRR used in this study was, however, found to give results comparable
33
34 4. CONCLUSIONS AND OUTLOOK
to the SSG when used in the algebraic forms and after recalibration of the LRR
model coefficients. An explanation of why the algebraic forms perform more
similarly than the corresponding full differential forms has been proposed. The
slight degeneration in model performance using EARSM derived from LRR is
well balanced by the fact that the full three-dimensional form of the proposed
EARSM is simple enough to be useful only for the EARSM derived from LRR.
The full three-dimensional form is necessary for capturing, for instance, the
secondary swirl in axially rotating pipes.
In the case of strong streamline curvature, the weak-equilibrium assumption
must be derived in a streamline based coordinate system. In some specific cases,
such as symmetric vortices and rotating homogeneous shear flows, the streamline
based coordinate system may explicitly be defined and the advection term may
be exactly included into the EARSM form. For these specific cases the proper
treatment of the advection term has been found to be of crucial importance.
When approximations of the advection term is included into the EARSM formu-
lation, the proposed EARSM fails because of the basic LRR model. However, by
an ad hoc modification of the coefficient preceding the complete right-hand side
of the anisotropy transport equation (19), it is possible to get a model behaviour
similar to EARSM derived from SSG in these kind of flows and also keeping the
simplified three-dimensional form.
One should here start to consider if it is worth the effort to include approx-
imations of the transport terms from the full RST model within the EARSM
form instead of leaving this class of models for full RST models where such ef-
fects enters naturally. The only, and so far completely relevant, argument for
EARSM is the troublesome numerical treatment and computational cost of full
RST models. One should, thus, put more efforts on improvements of the numer-
ical methods for full RST models and entice specialists in numerical methods to
leave the island of constant coefficients for the world of these strongly nonlinear
problems. Hopefully, one would expect that improvements in numerical methods
and computer performance would decrease the importance of the argument for
EARSM in the near future and, thus, the importance of RST models in engineer-
ing CFD methods could be expected to increase. Also in such a situation it is
worthwhile to continue the struggle for the search of the ‘optimal’ model within
both the class of two-equation models and that of full differential Reynolds stress
models.
Acknowledgments
I wish to thank my supervisor Professor Arne Johansson for introducing me

into the field of turbulence modelling. During our numerous discussions, which
he always had time for, Arne guided me with his great knowledge in this field.
The atmosphere at the department has been inspiring and friendly. I would
especially like to thank Doctor Petra Wikström for a fruiful collaboration on
passive scalars and Martin Skote for providing me with DNS data.
I also wish to thank Doctor Ingemar Lindblad, head of my department at
FFA, for believing in my ideas and supporting me even when I seemed to have
disappeared into the boundary layer. The professionalism and wide range of
experience of my colleagues at FFA were the prerequisite and basis of my work.
I particularly want to thank Doctor Peter Eliasson for helping me with numerical
treatment and implementation issues.
I would like to thank Doctor Torbjörn Sjögren for our discussions concerning
representation theory and automatic code generation. Professor Sharath Giri-
maji is kindly acknowledged for fruitful discussions and for making the visit of
me and my family at ICASE very pleasant.
Finally, I want to thank Ingela for always supporting me. Tack Tommy och
Cecilia för att ni är så mysiga.
Financial support from FFA, ESA, EU, NUTEK and FMV is gratefully
acknowledged.
35
Bibliography
[1] Abe, K., Kondoh, T. & Nagano, Y. 1996 A two-equation heat transfer model reflecting
second-moment closures for wall and free turbulent flows. Int. J. Heat and Fluid Flow 17,
228-237.
[2] Adumitroaie, V., Taulbee, D.B. & Givi, P. 1997 Explicit Algebraic Scalar-Flux Models
for Turbulent Reacting Flows. A. i. ch. e. J. 8, 1935-2147.
[3] Baldwin, B. & Lomax, H. 1978 Thin Layer Approximation and Algebraic Model for
Separated Turbulent Flow. AIAA Paper 78-257.
[4] Bardina, J., Ferziger, J.H. & Reynolds, W.C. 1983 Improved turbulence models based
on large-eddy simulation of homogeneous, incompressible turbulent flows. Stanford Uni-
versity Tech. Rep. TF-19.
[5] Boussinesq, J. 1877 Essai sur la théorie des eaux courantes. Mém. prés. Acad. Sci., 3rd
edn, Paris XXIII, 46.
[6] Bradshaw, P. 1997 The best turbulence models for engineers. In Modeling complex tur-
bulent flows, Eds. M.D. Salas, J.N. Hefner, L. Sakell, Kluwer (1999).
[7] Chien, K.Y. 1982 Predictions of Channel and Boundary-Layer Flows with a Low-
Reynolds-Number Turbulence Model. AIAA J. 20, 33–38.
[8] Cook, P.H., MacDonald, M.A. & Firmin, M.C.P. 1979 Aerofoil 2822 - Pressure distri-
butions, boundary layer and wake measurements. AGARD AR 138.
[9] Craft, T.J., Launder, B.E. & Suga, K. 1993 Extending the applicability of eddy viscos-
ity models through the use of deformation invariants and non-linear elements. In Proc. 5th
Int. Symp. on Refined Flow Modelling and Turbulence Measurements, 125–132, Presses
Ponts et Chaussées, Paris.
[10] Daly, B.J. & Harlow, F.H. 1970 Transport equations in turbulence. Phys. Fluids 13,
2634–2649.
[11] Friedrich, R. 1998 Modelling of turbulence in compressible flows. Chapter 5 of Transi-
tion, Turbulence and Combustion Modelling, Eds. A. Hanifi, P.H. Alfredsson, A.V.
Johansson and D.S. Henningson. Kluwer.
[12] Gatski, T.B & Speziale C.G. 1993 On explicit algebraic stress models for complex
turbulent flows. J. Fluid Mech. 254, 59–78.
[13] Girimaji, S.S. 1996 Fully-explicit and self-consistent algebraic Reynolds stress model
Theor. and Comp. Fluid Dyn. 8 387–402 (also ICASE Report No. 95-82).
[14] Girimaji, S.S. 1997 A Galilean invariant explicit algebraic Reynolds stress model for
turbulent curved flows Phys. Fluids. 9 1067–1077 (also ICASE Report No. 96-38).
[15] Girimaji, S.S. 1997 Development of algebraic Reynolds stress model for non-equilibrium
turbulence. In Modeling complex turbulent flows, Eds. M.D. Salas, J.N. Hefner, L. Sakell,
Kluwer (1999).
[16] Girimaji, S.S. & Balachandar, S. 1998 Analysis and modeling of boyancy-generated
turbulence using numerical data. Int. J. Heat Mass Transfer 41, 915-929.
36
BIBLIOGRAPHY 37
[17] Hanjalić, K., Jakirlić, S. & Hadžić, I. 1995 Computation of oscillating turbulent flows
at transitional Re-numners Turbulent Shear Flows 9 pp 323–342. Eds F. Durst, N. Kasagi
& B.E. Launder. Springer-Verlag, Berlin.
[18] Hellström, T., Davidson, L. & Rizzi, A. 1994 Reynolds stress transport modelling
of transonic flow around the RAE2822 airfoil. AIAA 94-0309 32nd Aerospace Sciences
Meeting, Reno, Jan 1994
[19] Johansson, A.V. & Burden, A.D. 1998 An introduction to turbulence modelling. Chap-
ter 4 of Transition, Turbulence and Combustion Modelling, Eds. A. Hanifi, P.H.
Alfredsson, A.V. Johansson and D.S. Henningson. Kluwer.
[20] Johansson, A.V. & Wallin, S. 1996 A new explicit algebraic Reynolds stress model.
Proc. Sixth European Turbulence Conference, Lausanne, July 1996, Ed. P. Monkewitz,
31–34.
[21] Johansson, A.V. & Wikström, P.M. 2000 DNS and modelling of passive scalar transport
in turbulent channel flow with a focus on scalar dissipation rate modelling. To appear in
ERCOFTAC J.
[22] Jones, W.P. & Launder, B.E. 1972 The prediction of laminarization with a two-equation
model of turbulence. Int. J. Heat Mass Trans. 15, 310–314.
[23] Jongen, T. & Marx, Y.P. 1997 Design of an unconditionally stable positive scheme for
the K–ε and two-layer turbulence models. Computers & Fluids 26, 469–487.
[24] Jongen, T., Machiels, L. & Gatski, T.B. 1998 Predicting noninertial effects with linear
and nonlinear eddy-viscosity, and algebraic stress models. Flow Turbul. Combust. 60 (2),
215–234.
[25] Jongen, T. & Gatski, T.B. 1998 A new approach to characterizing the equilibrium
states of the Reynolds stress anisotropy in homogeneous turbulence. Theoret. Comput.
Fluid Dynamics 11, 31–47.
[26] Jongen, T., Mompean, G. & Gatski, T.B. 1998 Predicting S-duct flow using a composite
algebraic stress model. AIAA J. 36, 327–335.
[27] Kolmogorov, A.N. 1942 Equations of turbulent motions of an incompressible fluid. Izves-
tia Academy of Sciences, USSR; Physics, 6, 56–58.
[28] Launder, B.E. 1978 Heat and mass transport. In Turbulence, P. Bradshaw (ed.), Springer,
Berlin, 231-287.
[29] Launder, B.E., Reece, G.J. & Rodi, W. 1975 Progress in the development of a Reynolds-
stress turbulence closure. J. Fluid Mech. 41, 537–566.
[30] Lumley, J.L. 1978 Computational modeling of turbulent flows. Adv. Appl. Mech. 18,
123–176.
[31] Launder, B.E. 1996 Advanced turbulence models for industrial applications Chapter 5
of Turbulence and Transition Modelling, Eds. M. Hallbäck, D.S. Henningson, A.V.
Johansson and P.H. Alfredsson. Kluwer.
[32] Menter, F.R. 1994 Two-equation eddy-viscosity turbulence models for engineering ap-
plications. AIAA J. 32, 1598–1604.
[33] Moser, R.D, Kim J. & Mansour N.N. 1998 DNS of Turbulent Channel Flow up to
Reτ = 590. Phys. Fluids 11, 943–945.
[34] McDevitt, J. 1979 Supercritical Flow about a Thick Circular-Arc Airfoil. NASA TM-
78549.
[35] Naot, D., Shavit, A. & Wolfstein, M. 1973 Two-point correlation model and the
redistribution of Reynolds stresses. Phys. Fluids 16, 738–743.
[36] Pope, S.B. 1975 A more general effective-viscosity hypothesis J. Fluid Mech. 72, 331–340.
[37] Pope, S.B. 1997 A perspective on turbulence modeling. In Modeling complex turbulent
flows, Eds. M.D. Salas, J.N. Hefner, L. Sakell, Kluwer (1999).
[38] Prandtl, L. 1925 Über die ausgebildete Turbulenz. ZAMM 5, 136–139.
38 BIBLIOGRAPHY
[39] Prandtl, L. 1945 Über ein neues Formelsystem für die ausgebildete Turbulenz. Nachr.
Akad. Wiss. Göttingen. Math.-Phys. Kl. 1945, 6–19.
[40] Reynolds, O. 1883 An experimental investigation of the circumstances which determine
whether the motion of water shall be direct or sinuous, and of the law of resistance in
parallel channels. Philos. Trans. R. Soc. London 174, 935–982.
[41] Rizzi, A., Eliasson, P., Lindblad, I., Hirsch C., Lacor, C., Haeuser, J. 1993 The
engineering of multiblock/multigrid software for Navier-Stokes flows on structured meshes
Computers Fluids 22, 341–367.
[42] Rodi, W. 1972 The prediction of free turbulent boundary layers by use of a two equation
model of turbulence. Ph.D. thesis, University of London.
[43] Rodi, W. 1976 A new algebraic relation for calculating the Reynolds stresses. Z. angew.
Math. Mech. 56, T219–221.
[44] Rogers, M.M., Moin, P. & Reynolds, W.C. 1986 The structure and modeling of the
hydrodynamic and passive scalar fields in homogeneous turbulent shear flow. Dept. Mech.
Engng. Rep. TF-25, Stanford University, Stanford, California.
[45] Rogers, M.M., Mansour, P. & Reynolds, W.C. 1989 An algebraic model for the tur-
bulent flux of a passive scalar. J. Fluid Mech. 203, 77-101.
[46] Rotta, J.C. 1951 Statistische Theorie nichthomogener Turbulenz. Z. Phys. 129, 547–572.
[47] Rumsey, C.L., Gatski, T.L. & Morrison, J.H. 1999 Turbulence model predictions of
extra-strain rate effects in strongly-curved flows. AIAA Paper 99-157
[48] Rung, T., Lübcke, H., Franke, M., Xue, L., Thiede, F. & Fu, S. 1999 Assessment
of explicit algebraic stress models in transonic flows. In Proc. Eng. turb. modelling and
experiments 4, Ajaccio, France, May 1999, Elsevier.
[49] Schülein, E., Krogmann, P. & Stanewsky, E. 1996 Documentation of Two-Dimensional
Impinging Shock/ Turbulent Boundary Layer Interaction Flow. DLR Report: DLR IB 223
- 96 A 49.
[50] Shabany, Y. & Durbin, P.A. 1997 Explicit algebraic scalar flux approximation. AIAA
J. 35, 985-989.
[51] Shabbir, A. & Shih, T.H. 1992 Critical Assessment of Reynolds Stress Turbulence Models
Using Homogeneous Flows NASA TM 105954, ICOMP-92-24, CMOTT-92-12.
[52] Shih, T.H. & Lumley, J.L. 1993 Remarks on Turbulent Constitutive Relations. Mathl.
Comput. Modelling 18(2), 9-16.
[53] Shih, T.H., Zhu, J. & Lumley, J.L. 1992 A Realizable Reynolds Stress Algebraic Equa-
tion Model NASA TM 105993, ICOMP-92-27, CMOTT-92-14.
[54] Shih, T.-H. 1996 Constitutive relations and realizability of single-point turbulence clo-
sures Chapter 4 of Turbulence and Transition Modelling, Eds. M. Hallbäck, D.S.
Henningson, A.V. Johansson and P.H. Alfredsson. Kluwer.
[55] Sjögren, T. 1997 Development and Validation of Turbulence Models Through Experi-
ment and Computation Doctoral thesis, Dept. of Mechanics, KTH, Stockholm, Sweden.
[56] Sjögren, T. & Johansson, A.V. 1999 Development and calibration of algebraic non-
linear models for terms in the Reynolds stress transport equations To appear in Phys.
Fluids
[57] Spalart P.R. & Speziale, C.G. 1999 A note on constraints in turbulence modelling.
J. Fluid Mech. 391, 373–376.
[58] Spalart P. & Allmaras, S.R. 1992 A one-equation model for aerodynamic flows. AIAA
paper 92-439.
[59] Spencer, A.J.M. & Rivlin, R.S. 1959 The theory of matrix polynomials and its appli-
cation to the mechanics of isotropic continua Arch. Rat. Mech. Anal. 2, 309–336.
[60] Speziale, C.G., Sarkar, S. & Gatski, T.B. 1991 Modelling the pressure-strain cor-
relation of turbulence: an invariant dynamical systems approach. J. Fluid Mech. 227,
245–272.
BIBLIOGRAPHY 39
[61] Speziale, C.G. & Xu, X.-H. 1996 Towards the development of second-order closure mod-
els for nonequilibrium turbulent flows Int. J. Heat and Fluid Flow 17, 238–244.
[62] Taulbee, D.B. 1992 An improved algebraic Reynolds stress model and corresponding
nonlinear stress model. Phys. Fluids A 4, 2555–2561.
[63] Tennekes, H. & Lumley, J.L. 1972 A first course in turbulence. The MIT Press, Cam-
bridge, MA, and London, England.
[64] Wikström, P.M. & Johansson, A.V. 1998 DNS and scalar-flux transport modelling in a
turbulent channel flow. In proc. Turbulent Heat Transfer II, May 31-June 5, Manchester.
[65] Wikström, P.M., Hallbäck, M., & Johansson, A.V. 1998 Measurements and heat-flux
transport modelling in a heated cylinder wake. Intl. J. Heat Fluid Flow 19, 556–562.
August 1998.
[66] Wilcox, D.C. 1988 Reassessment of the scale-determining equation for advanced turbu-
lence models. AIAA J. 26, 1299–1310.
[67] Wilcox, D.C. 1994 Simulation of Transition with a Two-Equation Turbulence Model.
AIAA J. 32, 247–255.
[68] Xu, X.-H. & Speziale, C.G. 1996 Explicit algebraic stress model of turbulence with
anisotropic dissipation. AIAA J. 34, 2186–2189.
1
Paper 1
AN EXPLICIT ALGEBRAIC REYNOLDS STRESS MODEL
FOR INCOMPRESSIBLE AND COMPRESSIBLE TURBULENT
FLOWS
Stefan Wallin
The Aeronautical Research Institute of Sweden (FFA)
Box 11021, SE-16111 Bromma, Sweden
Arne V. Johansson
Department of Mechanics, KTH, SE-100 44 Stockholm, Sweden
Abstract. Some new developments of explicit algebraic Reynolds stress

turbulence models (EARSM) are presented. The new developments include
a new near-wall treatment ensuring realizability for the individual stress
components, a formulation for compressible flows, and a suggestion for
a possible approximation of diffusion terms in the anisotropy transport
equation. Recent developments in this area are assessed and collected into
a model for both incompressible and compressible three-dimensional wall-
bounded turbulent flows. This model represents a solution of the implicit
ARSM equations, where the production to dissipation ratio is obtained as
a solution to a nonlinear algebraic relation. Three-dimensionality is fully
accounted for in the mean flow description of the stress anisotropy. The
resulting EARSM has been found to be well suited to integration to the wall
and all individual Reynolds stresses can be well predicted by introducing
wall damping functions derived from the van Driest damping function. The
platform for the model consists of the transport equations for the kinetic
energy and an auxiliary quantity. The proposed model can be used with
any such platform, and examples are shown for two different choices of the
auxiliary quantity.
1. Introduction
Standard two-equation models are still dominant in the context of industrial
flow computations. In flows with strong effects of streamline curvature, adverse
pressure gradients, flow separation or system rotation, such models fail to give
accurate predictions. Turbulence models based on the transport equations for
the individual Reynolds stresses have the natural potential for dealing with, for
example, the associated complex dynamics of inter-component transfer. The
Boussinesq hypothesis may in this context be said to be replaced by transport
equations for the individual Reynolds stress anisotropies. As yet, there are non-
trivial numerical aspects of flow computations with such models in complex
flow situations. This represents an active area of research. In parallel with such
efforts there has been a considerable renewed interest in various forms of algebraic
approximations of the anisotropy transport equations. In the present work some
new developments are presented for explicit formulations of algebraic Reynolds
43
44 AN EXPLICIT ALGEBRAIC REYNOLDS STRESS MODEL
stress models. The motivation for this work is the general need for improvements
in the prediction of complicated turbulent flows using the platform of existing
CFD prediction tools based on the two-equation modelling level.
1.1. Algebraic Reynolds stress models. The classical algebraic Reynolds

stress model (ARSM), Rodi[26, 27], was developed from the modelled Reynolds
stress transport (RST) equation by assuming that the advection minus the dif-
fusion of the individual Reynolds stresses can be expressed as the product of the
corresponding quantity for the kinetic energy, K, and the individual Reynolds
stresses normalized by K. This results in an implicit relation between the stress
components and the mean velocity gradient field that replaces the Boussinesq
hypothesis. Since the algebraic Reynolds stress model is determined from the
modelled Reynolds stress transport equation no additional model constants are
needed and the basic behaviour and experiences of the particular RST model
will be inherited. The linearity of the Boussinesq hypothesis excludes any de-
pendence on the rotational (antisymmetric) part of the mean velocity gradient
tensor. An ARSM approach here represents a systematic method of constructing
a nonlinear stress relationship that includes effects of the rotational part of the
mean velocity gradient tensor. Despite this definite improvement one should,
however, keep in mind that the ARSM can never represent the transport effects
as well as a full RST model, which always should be expected to give a more
correct description of the turbulence. Also, the RST models have limitations in
predicting turbulence for general and complicated flows, especially in describing
the effects of rotation.
The traditional ARSM idea is equivalent to neglecting advection and diffusion
terms in the exact transport equation for the Reynolds stress anisotropy, aij ,
defined as aij ≡ ui uj /K − 2δij /3. The advection term is indeed exactly zero for
all stationary parallel mean flows, such as fully developed channel and pipe flows.
For inhomogeneous flows the assumption of negligible effects of diffusion in the
aij equation can cause problems, particularly in regions where the production
term is small or where the inhomogeneity is strong. However, the Rodi ARSM
assumption incorporates in a natural way not only effects of rotation but also
effects of streamline curvature and three-dimensionality of the flow and has been
found to be a reasonable approximation of the full differential RST equations
in a number of flow situations, in many respects superior to the eddy-viscosity
hypothesis.
To illustrate the natural way in which system rotation, and rotational effects
in general, enter in this type of formalism we may take as the starting point
the transport equation for the Reynolds stress anisotropy tensor in a rotating
Cartesian coordinate system (the formulation of algebraic Reynolds stress models
for rotating and curvilinear coordinate systems has recently been studied in some
STEFAN WALLIN AND ARNE V. JOHANSSON 45
detail by Sjögren[35])
(K)

D aij ∂Tijl ui uj ∂Tl ui uj (a)
K − − =− (P − ε) + Pij − εij + Πij + εCij
Dt ∂xl K ∂xl K
(1)
(K)
where −Tijl and −Tl are the fluxes (turbulent and molecular) of the Reynolds
stress and turbulent kinetic energy, respectively. The dissipation rate tensor, εij ,
and the pressure strain, Πij , need to be modelled whereas the production terms,
(a)
Pij and P = Pii /2, and the Coriolis term, Cij , do not need any modelling since
they are explicit in the Reynolds stress tensor. In a non-rotating coordinate
system the Reynolds stress production term is normally written as
Pij = −ui uk Uj,k − uj uk Ui,k , (2)
where Ui,j denotes the mean velocity gradient tensor. In a rotating system
it is convenient to split the mean velocity gradient tensor into a mean strain
and a mean rotation tensor. We will here let Sij and Ωij denote these tensors
normalized with the turbulent timescale, τ ≡ K/ε,
τ τ
Sij = (Ui,j + Uj,i ) , Ωij = (Ui,j − Uj,i ) . (3)
2 2
A consistent formulation of (1), valid also in the rotating system, can then be
obtained by replacing the mean rotation tensor by the absolute rotation tensor
Ω∗ij = Ωij + Ωsij , (4)
where
Ωsij = τ /jik ωks (5)
and ωks is the constant angular rotation rate vector of the system. This procedure
illustrates the origin of the two parts of what normally is referred to as the
Coriolis term in the aij equation. In this way the first part is included in the
production term, that now (normalized with ε) can be expressed as
Pij 4
= − Sij − (aik Skj + Sik akj ) + aik Ω∗kj − Ω∗ik akj . (6)
ε 3
The second part of the Coriolis term arises from the transformation of the
(a)
advection term. This part (normalized by ε) is denoted by Cij in (1), and can
be expressed as
(a)
Cij = aik Ωskj − Ωsik akj . (7)
The ARSM assumption results in the following implicit algebraic equation for
aij :
ui uj (a)
(P − ε) = Pij − εij + Πij + εCij , (8)
K
the structure of which, of course, depends on the choice of the models for εij
and Πij . For the present modelling purpose we choose an isotropic assumption
for the dissipation rate tensor,
2
εij = εδij , (9)
3
and the Rotta model[28] for the slow pressure strain
(s)
Πij = −c1 εaij . (10)
For the rapid pressure–strain rate we choose the general linear model of Launder,
Reece & Rodi[17], which for a non-rotating system normally is written as

(r) c2 + 8 2 30c2 − 2
Πij = − Pij − Pδij − K (Ui,j + Uj,i )
11 3 55

8c2 − 2 2
− Dij − Pδij (11)
11 3
where Dij = −ui uk Uk,j − uj uk Uk,i .
A simple way of obtaining a consistent, frame-independent formulation of the
rapid pressure–strain rate model is to apply the same methodology as for the
production term. This gives
(r)
Πij 4 9c2 + 6 2
= Sij + aik Skj + Sik akj − akm Smk δij
ε 5 11 3
7c2 − 10
+ aik Ω∗kj − Ω∗ik akj . (12)
11
From (8) we then obtain the implicit algebraic equation for the Reynolds
stress anisotropy tensor in the form

P 8 7c2 + 1 R
c1 − 1 + a = − S+ aΩ − ΩR a
ε 15 11

5 − 9c2 2
− aS + Sa − tr{aS}I (13)
11 3
In equation (13) a, S and Ω denote second-rank tensors, and I is the identity

matrix. The inner product of two matrices is defined as (SS)ij ≡ S2 ij ≡ Sik Skj
and tr{} denotes the trace. This notation will be kept through this paper. One
should note that (13) represents a nonlinear relation since P/ε ≡ −tr{aS}.
It is interesting to note that the ‘effective’ mean rotation rate tensor, ΩRij ,
depends on the choice of model:
∗ 11 7c2 + 12 s
ΩR
ij = Ωij + Ωs = Ωij + Ω . (14)
7c2 + 1 ij 7c2 + 1 ij
The ARSM approximation of the aij transport equation with this approach is
then equivalent to neglecting the advective term (and diffusion) in the chosen
rotating coordinate system. Hence, the adequacy of the ARSM approach is
coupled to the choice of a coordinate system where the omission of advection
terms in the aij equation can be justified. The choice of coordinate system is
not at all trivial in strongly curved flows and the coordinate direction is not in
general aligned with the flow direction. There are, however, methods to construct
ARSMs that generally neglect only the advection term in the streamline direction
(see e.g. Girimaji[10] and Sjögren[35]).
As is seen from equation (13) the treatment of system rotation is quite straight-
forward. The superscript R on the mean rotation rate tensor will be dropped in
the following.
The turbulent kinetic energy, K ≡ ui ui /2, and its dissipation, ε, are deter-
mined from transport equations
(K)
DK ∂Tl
+ = P − ε, (15)
Dt ∂xl
(ε)
D ε ∂Tl ε
+ = (Cε1 P − Cε2 fε ε) . (16)
Dt ∂xl K
There is no direct influence of system rotation on these equations. A substantial

improvement over the eddy-viscosity model equations is that the production
(K) (ε)
term, P, does not need to be modelled. The transport terms, Tl and Tl , are
usually modelled using gradient diffusion with the diffusivity coefficient expressed
with the aid of the eddy viscosity, but here an improvement can also be achieved
by using the knowledge of the complete Reynolds stress tensor. The gradient-
diffusion model proposed by Daly & Harlow[3] applied to the turbulent kinetic
energy and its dissipation gives
K ∂K K ∂ε
= −cs
(K) (ε)
Tl ul um , Tl = −cε ul um . (17)
ε ∂xm ε ∂xm
Launder, Reece & Rodi[17] recommend cs = 0.25 and cε = 0.15.
1.2. Explicit algebraic Reynolds stress models. The implicit relation for
a in the ARSM equations has been found to be numerically and computation-
ally cumbersome since there is no diffusion or damping present in the equation
system. In many applications the computational effort has been found to be
excessively large and the benefits of using ARSM instead of the full Reynolds
stress model are then lost. An explicit algebraic Reynolds stress model, EARSM,
where the Reynolds stresses are explicitly related to the mean flow field is much
more numerically robust and has been found to have almost a negligible effect
on the computational effort as compared to a K–ε model.
The most general form for a in terms of S and Ω consists of ten tensorially
independent groups to which all higher-order tensor combinations can be reduced
with the aid of the Caley–Hamilton theorem:

1 1
a = β1 S + β2 S2 − IIS I + β3 Ω2 − IIΩ I + β4 (SΩ − ΩS)
3 3

2 2
+ β5 S Ω − ΩS2 + β6 SΩ2 + Ω2 S − IV I
3

2
+ β7 S2 Ω2 + Ω2 S2 − V I + β8 SΩS2 − S2 ΩS
3

+ β9 ΩSΩ − Ω SΩ + β10 ΩS2 Ω2 − Ω2 S2 Ω .
2 2
(18)
The β-coefficients may be functions of the five independent invariants of S and
Ω, which can be written as
IIS = tr{S2 }, IIΩ = tr{Ω2 }, IIIS = tr{S3 }, IV = tr{SΩ2 }, V = tr{S2 Ω2 }.
(19)
Other scalar parameters may also be involved.
The independence of the ten groups and the completeness of expression (18)
is further discussed in Appendix B. Shih & Lumley[34] include also an eleventh,
sixth-order group. The exact expression for such a group in terms of the lower-
order groups is also given in Appendix B. In two-dimensional mean flows, there
are only three independent tensor groups, i.e. the β1,2,4 groups, and two inde-
pendent invariants, IIS and IIΩ .
Formulation of expression (18) in terms of S and Ω follows Pope[24] and gives
a much more compact form than by using the mean velocity gradients directly.
It is also more cumbersome to compute matrix products involving Ui,j due to
the full matrix (eight independent elements) while Sij and Ωij contain only five
and three independent elements respectively. Also, the introduction of a rotating
coordinate system is simplified here.
The main problem in obtaining an explicit relation for the anisotropy is that
of determining the β-coefficients. A possibility would be to calibrate these from
some chosen set of ‘basic flows’. Shih, Zhu & Lumley[32] partially adopted
this approach combined with conditions related to realizability and restraints
on correct behaviour in rapid distortion limits. A more traditional approach
is to derive an explicit form from an implicit a-relation based on established
models for the terms in the Reynolds stress (or its anisotropy) tensor transport
equation, equivalent to equation (13). Pope[24] was the first to propose using
the ten tensor groups to form a consistent explicit relation. He also derived a
relation for two-dimensional mean flows leaving the production to dissipation
ratio (P/ε) implicit. This approach was later extended and solved for three-
dimensional mean flows by Taulbee[41] for the special case of c2 = 5/9 and by
Gatski & Speziale[7] for a general linear pressure–strain model.
The nonlinearity of equation (13) (P/ε ≡ −tr{aS}) forms a major obstacle
for this approach and the studies published so far have circumvented the problem
by letting P/ε = −tr{aS} be implicit during the solution of (13), adopted by
Pope[24] and Taulbee[41] or the approach used by Gatski & Speziale[7] where
they used the asymptotic equilibrium value for P/ε as a universal constant.
The nonlinear system of equations is, however, conveniently solved in the
form of a linear system of (five) equations complemented by a nonlinear scalar
equation for P/ε. For two-dimensional mean flows, Johansson & Wallin[15] and
Girimaji[8, 9] have independently shown that this equation has a closed and fully
explicit solution that can be expressed in a compact form. In the present work it
is shown that good approximations are easily found for general three-dimensional
mean flows. The complexity of the solution is substantially reduced by setting
the coefficient c2 = 5/9 (see e.g. Taulbee[41]).
The removal of the need for ad hoc relations for P/ε represents a substantial
improvement for this type of modelling. A constant P/ε gives wrong asymptotic
behaviour for large strain rates, also noticed by Speziale & Xu[40], while the
fully consistent solution of the nonlinear equation system automatically fulfils
the correct asymptotic behaviour. Also, in the very near-wall region, the cor-
rect asymptotic behaviour implies that all individual Reynolds stresses can be
satisfactorily represented simply by introducing a van Driest damping function.
A straightforward extension to compressible flow is derived in which the mean
density variations are taken into account. This model is applied to a complex
flow situation with a shock–boundary layer interaction. An extension of the
model to account in a simple way for the neglected turbulent transport of the
anisotropies is also considered.
2. Formulation of an explicit algebraic model (EARSM)

The value of c2 in the rapid pressure–strain model was originally suggested to be
0.4 by Launder et al.[17], but more recent studies have suggested a higher value
close to 5/9[18, 31]. This means that the last term in equation (13) is of quite
small influence, also noticed by Taulbee[41]. Setting c2 = 5/9 one obtains the
simplified but still implicit equation

P 8 4
c1 − 1 + a = − S + (aΩ − Ωa) . (20)
ε 15 9
System rotation can easily be accounted for by substituting Ω with ΩR = Ω +

(13/4) Ωs according to (14) where Ωs is given by (5) (see section 1). It will be
shown later that the removal of the last term in equation (13) gives a substantial
simplification of the solution, especially in three-dimensional mean flow. The
Rotta coefficient, c1 , is here set to 1.8.
The simplified but implicit algebraic Reynolds stress equation (20) is rewritten
in the following form:
6
N a = − S + (aΩ − Ωa) (21)
5
where N is closely related to the production to dissipation ratio (P/ε = −tr{aS}),

9P
N = c1 + (22)
4ε
and
9
c1 = (c1 − 1) . (23)
4
The procedure to solve this equation is the following: First, the general form
for the anisotropy, equation (18), is inserted into the simplified ARSM equation
(21) where N is not yet determined. The resulting linear equation system for the
β-coefficients is then solved by using the fact that higher-order tensor groups can
be reduced with the aid of the Cayley–Hamilton theorem where the ten groups in
the general form (18) form a complete basis. The β-coefficients are now functions
of the production to dissipation ratio, P/ε, or N . The final step is to formulate
and solve the nonlinear scalar equation for N or P/ε.
2.1. Solution of the simplified ARSM for two-dimensional mean flows.
For two-dimensional mean flows the solution is reduced to only two non-zero
coefficients, which can be expressed as
6 N 6 1
β1 = − , β4 = − . (24)
5 N 2 − 2IIΩ 5 N 2 − 2IIΩ
It is clearly seen that the denominator, N 2 − 2IIΩ , cannot become singular
since IIΩ is always negative. It will be shown later that an explicit formulation of
any quasi-linear ARSM is non-singular if the corresponding nonlinear equation
for the production to dissipation ratio is solved, as also noticed by Girimaji[8, 9].
The nonlinear equation for N in two-dimensional mean flow can be derived
by introducing the solution of a for two-dimensional mean flow in the definition
of N . The resulting equation is cubic

27
N 3 − c1 N 2 − IIS + 2IIΩ N + 2c1 IIΩ = 0, (25)
10
and can be solved in a closed form with the solution for the positive root being

1/3

c1 + P1 + P2 + sign P1 − P2 | P1 − P2 |1/3 , P2 ≥ 0
3
N= c1 2 1/6 1 P1
+ 2 P1 − P2 cos arccos 2 , P2 < 0
3 3 P1 − P2
(26)
where the arccos function should return an angle between 0 and π and
3
c1 c1
2 2
9 2 9 2
P1 = + IIS − IIΩ c1 , P2 = P1 − 2
+ IIS + IIΩ .
27 20 3 9 10 3
(27)
It can easily be shown that N remains real and positive for all possible values of
IIS and IIΩ . The production to dissipation ratio may then be found from (22).
3.0
2.5 0.5
1.0
1.5
2.0
2.0
ω
2.5
1.5
1.0
0.5
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ
Figure 1. Isolines of the production to dissipation ratio for the cur-
rent model.
3.0
eddy viscosity
2.5
2.0
hom. shear
P/ε
1.5
log-layer
1.0
0.5
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ
Figure 2. Production to dissipation ratio versus strain rate σ for
different rotation ratios ω/σ. The current model ( ) compared to
an eddy-viscosity model ( ).
The above solution for two-dimensional flows was also given in [15] and is very
similar in structure to that of Girimaji[8, 9] who also derives an explicit solution
for P/ε although for a somewhat more general set of models that yields β2 = 0
(cf the solutions in Appendix C).
2.1.1. Illustration of the behaviour of the proposed model. Figures 1 and 2 illus-
trate the behaviour of the solution for P/ε. We note that it is zero for all cases
with σ = 0, i.e. irrespective
of the value of ω, where σ and ω are defined as
σ ≡ IIS /2 and ω ≡ −IIΩ /2 and that the P/ε ratio decreases monotonically
with increasing influence of rotation. For all parallel shear flows σ = ω.
Homogeneous shear flow is a classical corner stone case for calibration of tur-
bulence models. Tavoularis & Corrsin[43] have experimentally shown that the
asymptotic value of SK/ε ≈ 6 corresponding to σ = ω = 3. In the experiments
Table 1. The computed anisotropy in the log-layer using the current

model assuming balance between turbulence production and dissipa-
tion compared to channel DNS data by Kim[16].
a12 a11 a22 a33 σ

DNS −0.29 0.34 −0.26 −0.08 1.65
Current model −0.30 0.25 −0.25 0.00 1.69
Table 2. The anisotropy in asymptotic homogeneous shear flow us-

ing the current model with σ = 3.0 compared to measurements by
Tavoularis & Corrsin[43].
a12 a11 a22 a33 P/ε

Experiment −0.30 0.40 −0.28 −0.12 1.8
Current model −0.30 0.31 −0.31 0.00 1.8
the production to dissipation ratio was found to be approximately 1.8, marked

as a circle in figure 2. The present model exactly replicates that result. Further-
more, in the log-layer of a boundary layer we know that the production balances
the dissipation rate, P = ε, which is obtained with the proposed model when the
strain rate σ = 1.69. This is within the range of σ values found in the log-layer
of the DNS data for channel flow [16]. This is also consistent with an effective
Cµ = 0.09 which gives σ = 1.67, also marked as a circle in the figure.
The anisotropies for the two cases are compared in tables 1 and 2. Most
important is to correctly predict the a12 anisotropy since this is the only com-
ponent of the anisotropy tensor that contributes to the turbulent production in
parallel flows. The a22 component is also important since this is the only term
that contributes to the turbulent diffusion term. The tables show that a12 and
a22 are well predicted for the two different cases. The a11 and a33 components
are, however, not as well predicted due to the simplification of setting c2 = 5/9
since this implies that a33 = 0. The inclusion of c2 = 5/9 will, however, not be
sufficient if one wishes to improve the prediction of the anisotropies, since the
particular choice of c1 and c2 used here is the only combination that predicts
both a12 and a22 correctly in these two different cases. The good behaviour
of the model for these two very different cases justifies the particular choice of
model constants, c1 and c2 , in the pressure–strain model.
A better tuning of the different anisotropies could possibly be achieved by
using the most general quasi-linear pressure–strain model, where the coefficients
may be functions of the production to dissipation ratio. This will be further
discussed in section 2.3.
The asymptotic behaviour for large strain rates in parallel flow can be investi-
gated by letting σ = ω → ∞. The production to dissipation ratio then becomes
P/ε ∼ σ and β1 ∼ 1/σ. This asymptotic behaviour for the β1 -coefficient (equiv-
alent to −2Cµ ) is of particular interest since it ensures good model behaviour
in the very near-wall buffer- and viscous sub-layers as will be shown later. The
asymptotic behaviour is also consistent with the asymptotic characteristics of
homogeneous shear flow. With an erroneous model assumption of a constant
production to dissipation ratio, P/ε, the wrong asymptote, β1 ∼ 1/σ 2 , is ob-
tained, as also noticed by Speziale & Xu[40], while the solution of the nonlinear
equation for P/ε automatically gives the correct asymptotic behaviour.
To illustrate the behaviour for large shear rates the model is tested in homo-
geneous shear flow at high initial shear rate (SK/ε = 50), see figure 3. This flow
is a case where one should expect differences between the algebraic approach
and the full differential model due to the fact that the anisotropies undergo a
temporal evolution (∂aij /∂t = 0) in the development towards an asymptotic
state. Moreover, the Launder et al.[17] model gives quite poor predictions of
this case when used in a differential form. The very good predictions of the
present EARSM can thus be regarded as a bit fortuitous. Nevertheless, the self-
consistent approach gives a model with the correct asymptotic behaviour, which
is a pre-requisite for reasonable predictions in the limit of high shear. It is im-
portant to make clear that the proposed model is not intended for these extreme
high shear rates and the normal stress components are not as well predicted as
the turbulent kinetic energy. It is, however, an important step towards a more
general engineering model that the model is able to give reasonable results in
extreme flow cases also.
In flows with an adverse pressure gradient, the production to dissipation ra-
tio is greater than 1 and eddy-viscosity models with constant Cµ overestimate
the turbulent viscosity or the a12 anisotropy. Bradshaw’s assumption, which is
adopted by Menter[20] in the shear stress transport (SST) model, forces the a12
anisotropy to be constant for P/ε ratios greater than unity, which gives β1 ∼ 1/σ.
This is fulfilled in the limit of large strain rates by the proposed model, which
also gives a nearly constant a12 anisotropy in boundary layers with small pres-
sure gradients. This can be seen in figure 4 where the a12 anisotropy versus the
strain rate is shown for parallel flows. The a12 anisotropy computed from the
proposed model is nearly constant here for a wide range of strain rates including
both the log-layer and asymptotic homogeneous shear. The eddy-viscosity as-
sumption gives a12 = −2Cµ σ and has a totally different behaviour, shown in the
figure, and becomes physically unrealizable for large strain rates, σ > 5.56. This
is, however, avoided by the SST limitation on the anisotropy which nearly coin-
cides with the proposed model for strain rates larger than those in the log-layer.
This feature of the proposed model ensures an improved behaviour in boundary
layers with pressure gradients as compared to standard eddy-viscosity models.
25
20
15
K
10
0
0 2 4 6 8 10
T
Figure 3. Time evolution of the turbulent kinetic energy in rapidly

sheared homogeneous flow, SK/ε = 50. Eddy-viscosity model ( ),
the proposed EARSM ( ) and the Gatski & Speziale[7] EARSM
( ) compared with RDT ( ❛).
0.0
-0.1
-0.2 hom. shear
-0.3
a12
-0.4 log-layer
-0.5
-0.6
0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ
Figure 4. Predicted a12 anisotropy versus strain rate σ for parallel
flow. The current model ( ) compared to the eddy-viscosity model
( ) and the Menter[20] SST model ( ).
Since equation (18) represents the general form of the anisotropy, different
explicit algebraic Reynolds stress models can be compared by studying the be-
haviour of the β-coefficients. In two-dimensional mean flows the β-coefficients
can be illustrated as iso-curves in the σ–ω plane. In figure 5 the β1 -coefficient
(equivalent to −2Cµ ) is shown for the Taulbee[41], Gatski & Speziale[7] and
Shih, Zhu & Lumley[32] models. There are substantial differences between the
different models for the β1 coefficient and a similar behaviour can also be seen
for the other coefficients. Figure 5(a) is very similar to the results of Pope[24]
although the coefficients in the pressure–strain model were somewhat different.
(a) (b)
3.0 3.0 -0.1
-0.1
2.5 2.5
2.0 2.0
ω
ω
1.5 1.5
-0.2
-0.2
1.0 1.0
-0.3
0.5 0.5
-0.3
0.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ σ
(c) (d)
3.0 -0.1 3.0
2.5 2.5
2.0 2.0
ω
1.5 1.5 -0.2
1.0 -0.2 1.0

-0.3
0.5 0.5
0.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ σ
Figure 5. The behaviour of the β1 -coefficient in the σ–ω-plane for:
(a) the implicit ARSM and the present EARSM, (b) Taulbee[41], (c)
Gatski & Speziale[7] and (d) Shih et al.[32].
One should of course bear in mind that the underlying approaches are chosen
differently in the different models. This only partly explains the different be-
haviours though. The Taulbee[41] model could, however, be directly compared
to the implicit ARSM since the basic approach is the Launder et al.[17] model
with similar choices of the model coefficients. The classical ARSM assumption
is, however, not asymptotically correct for small strain rates and Taulbee thus
makes a different approximation in that limit. That difference is clearly seen in
the figure. The approximation imposed by Taulbee is motivated by the neglected
advection term and gives improved predictions in developing homogeneous shear
flows where the advection is important for small times. However, the prediction
of fully developed channel flow, where the advection is zero, is also affected by
the Taulbee assumption. For that reason this approach is not adopted, even
though it improves the predicted shear stresses near the channel centre. A simi-
lar effect can be obtained by an approximate inclusion of turbulent diffusion (of
aij ) effects, as will be discussed in section 5.
The benefits of a direct, explicit solution of the production to dissipation ratio

can be further illustrated by considering rotating plane channel flow. The mean
flow is here described by U (x2 ) δi1 and the system rotation is taken as ω s δi3 .
The mean strain rate tensor is unaffected by system rotation, implying that the
effective IIS for the calculation of P/ε is the same as for a non-rotating channel
flow. The ‘effective’ second invariant of the mean rotation tensor (see equation
(14)) becomes
2
1 2 13 s 2 13 s
IIΩ = τ − (U ) − 6
R 2
(ω ) + 2 ω U . (28)
2 4 4
We note that, whereas the first two terms have the same sign on both sides
of the channel, the last term has an alternating sign. For moderate rotation
rates the model will, thus, predict a decrease of P/ε on the side of the channel
where IIΩ is increased, and vice versa. This is in agreement with the observed
behaviour of a stabilized and a destabilized side in rotating channel flow. Due to
the quadratic term (in ω s ) in the expression for IIΩ the model will also predict
a global decrease in P/ε for very large rates of rotation, as could be expected
from comparison with stability theory (see e.g. Matsson & Alfredsson[19]).
2.2. Solution of the simplified ARSM for three-dimensional mean flow.

For general three-dimensional mean flows the solution for the β-coefficients can
be written as

N 2N 2 − 7IIΩ 12N −1 IV
β1 = − , β3 = − ,
2 Q Q (29)
2 N − 2IIΩ 6N 6
β4 = − , β6 = − , β9 = ,
Q Q Q
where all the other coefficients are identically zero. The denominator
5 2
Q= N − 2IIΩ 2N 2 − IIΩ (30)
6
is also here clearly seen always to remain positive since IIΩ always is negative.
The nonlinear equation for N or the corresponding equation for P/ε is for ex-
ample obtained by introducing the above solution (29) for a into the definition
of N . The resulting equation is of sixth order and reads

27 5 5
N − c1 N −
6 5
IIS + IIΩ N 4 + c1 IIΩ N 3
10 2 2

189 81 81
+ IIΩ2 + IIS IIΩ − V N 2 − c1 IIΩ2 N − IV 2 = 0. (31)
20 5 5
This equation reduces to the two-dimensional cubic equation by recalling that in

two-dimensional mean flows there are only two independent invariants, whereas
IIIS = IV = 0 and V = IIS IIΩ /2. Equation (31) cannot be solved in a
closed form but the solution of N for the cubic equation (25) applicable in
two-dimensional mean flow, Nc , can be used as a first approximation. A further
improvement of the approximation of N is also possible by making a perturba-

tion solution of the three-dimensional equation. This is done by perturbing the
√
IV and V invariants around the two-dimensional solution, that is IV = φ1 and
V = IIS IIΩ /2 + φ2 , assuming that φ1 and φ2 are independent. Putting this into
the three-dimensional equation (31) and sorting in powers of φ1 and φ2 we get

162 φ1 + φ2 Nc2
N = Nc + + O φ21 , φ22 , φ1 φ2 (32)
D
where the denominator, D, is given by

1
D = 20Nc Nc − c1 − IIΩ 10Nc3 + 15c1 Nc2 + 10c1 IIΩ2
4
(33)
2
where (25) has been used to rewrite the expression for D. From (33) it is obvious
that D will always remain positive, since Nc ≥ c1 (see Appendix C).
One could also consider keeping P/ε or N implicit during the iteration pro-
cedure to a steady-state solution and thus avoid any further approximations.
It is, however, not known how this would affect the stability of the numerical
method and should be avoided. There could also be problems associated with
the existence of multiple roots, especially for three-dimensional cases.
2.2.1. Example of a three-dimensional mean flow: rotating pipe. Fully devel-

oped turbulent flow in a circular pipe rotating around its length axis is an in-
teresting case, since it represents a three-dimensional flow that can be described
with only one spatial coordinate, r in a cylindrical coordinate system (r̂, θ̂, ẑ).
If the flow is laminar, the tangential velocity, Uθ , varies linearly with the radius,
r, like a solid-body rotation. In turbulent flow, on the other hand, the tangential
velocity is nearly parabolic, which cannot be described with an eddy-viscosity
turbulence model. The fully three-dimensional form of the proposed EARSM is
needed to capture this behaviour. Limited forms with only second-order terms
is not sufficient, as will be shown below.
The Navier–Stokes equation in the tangential direction can be written as
2
d Uθ 1 dUθ Uθ d Karθ
ν + − 2 = (Karθ ) + 2 . (34)
dr2 r dr r dr r
After two integrations, the tangential velocity can be expressed as

R
r r Karθ
Uθ (r) = Uθ (R) − du (35)
R ν r u
where R is the radius of the tube. The first term corresponds to the linear Uθ
profile while the second term is the correction that may give a parabolic-like
profile if the arθ anisotropy is positive.
In this particular flow the strain- and rotation-rate tensors are evaluated in
an inertial frame and read
 dUz 
dr − r
dUθ Uθ
0 dr
1  dUθ
S= τ dr − Uθ
r 0 0  (36)
2 dUz
dr 0 0
and
 
0 − dU
dr −
θ Uθ
r − dU
dr
z
1 
Ω= τ dUθ
dr + Uθ
r 0 0 . (37)
2 dUz
dr 0 0
The terms that contribute to the arθ component in the general expression (18)
are the terms associated with the β1 , β5 , β6 and β10 -coefficients. In the case of
a linear Uθ profile, Srθ is zero and the contribution from the β1 -term vanishes.
This is consistent with the behaviour for an eddy-viscosity model.
We see from relation (35) that arθ should vanish rapidly for increasing Re. To
verify that the present EARSM is consistent with such a behaviour we start with
the EARSM solution for a flow that may deviate from solid-body rotation. For
the present EARSM β5 and β10 are zero, which means that here only the linear

term (β1 S) and the β6 -term (associated with the SΩ2 + Ω2 S − 23 IV I -group)
contribute to arθ :
2
1 dUθ Uθ 1 dUθ Uθ dUθ Uθ
arθ = β1 τ − − β6 τ 3
− +
2 dr r 4 dr r dr r
2
1 dUθ dUz
− β6 τ 3 . (38)
4 dr dr
For solid body rotation only the third term remains and we see that the arθ
component will vanish only if the axial mean velocity gradient dUz /dr vanishes.
This gradient does indeed vanish in the limit of infinite Reynolds number, but
only slowly1 with increasing Reynolds numbers while the relation (35) indicates
a more rapid decay rate (arθ → 1/Re).
The β1 and β6 coefficients in (38) are both negative, so the term with the axial
mean velocity gradient (the last term) gives a positive contribution to arθ which
drives the azimuthal velocity Uθ (r) towards a more parabolic-like profile (where
dUθ /dr > Uθ /r). The second term enhances this trend but what is interesting
is that the β1 -term (the first term) has the opposite sign and has the possibility
of balancing the arθ anisotropy component depending on the exact form of the
β-coefficients.
1 The ‘two-layer hypothesis’ ⇒ (UCL − Uz )/uτ = f ((R − r)/R) ⇒

(−(R/UCL )(dUz /dr))2 = (−(uτ /UCL )f )2 ∼ Cf where it is also well known that Cf
slowly decreases with increasing Reynolds number and that the mean velocity profile in a pipe
approaches a top-hat in the infinite Re limit (see e.g. Schlichting[30], chap. 20).
Hence, the required consistency of a rapidly vanishing arθ is indeed obtained

as a solution to the present EARSM, and is in fact associated with a non-zero
deviation from solid-body rotation for large but finite Reynolds numbers.
To clearly see this we may need to take a closer look at (38) by inserting the
present EARSM solutions for β1 , β6 . This gives
2
Uθ dUθ dUz
arθ ∼ 7 −
r dr dr
2 2
dUθ Uθ 81 P dUθ Uθ
− − C1 − 1 + + + (39)
dr r 4τ 2 ε dr r
and balance is obtained for a parabolic-like profile. Hence, the EARSM solution
implies an anisotropy that vanishes much more rapidly than dUz /dr. As is
obvious from (39) we also see that in the limit of infinite Re where dUz /dr → 0
the vanishing arθ is associated with an EARSM solution that gives a solid-body
rotation.
The rotation also affects the axial velocity component, which becomes less full,
i.e. more parabolic. That effect enters mainly through the rotation dependence
in the β1 -coefficient and could thus be captured also by a linear eddy-viscosity
model where the Cµ -coefficient is dependent on the rotation rate.
A fully developed turbulent rotating pipe flow has been computed using the
proposed model. The Reynolds number is 20 000 based on the mean flow ve-
locity and tube diameter. Three different rotation ratios Z = 0, 0.5 and 1 were
computed where Z = Uθ (R) /Um , i.e. the wall angular velocity divided by the
axial bulk velocity. The results are compared to the experiment by Imao, Itoh
& Harada[14]. The turbulence models used are the Chien[1] K–ε model and the
proposed EARSM including the near-wall formulation in section 3. The EARSM
is based on both the K–ε and K–ω models discussed in section 3.4.
The three-dimensional form of the model has been used and the influence of
the approximation of N was assessed by computing the case by using both the
zeroth- (Nc ) and first-order perturbation solution of N . The former is given by
(26) and labelled ‘EARSM 0’ in the figures while the latter ‘EARSM 1’ is given
by (32).
Figure 6 shows the predicted axial velocity for Z = 1 using the Chien K–ε
model alone and as the platform for EARSM calculations. The original Chien
eddy-viscosity K–ε model is seen to be completely insensitive to rotation, while
the EARSM predictions agree well with the experimental results. It is also
seen that the different approximations of N only have a minor influence on the
predicted velocity profile. In figure 7 predictions for different values of Z are
shown for the K–ω model as the platform of the EARSM. The calculated results
are seen to capture well the trend with increasing rate of rotation. The EARSM
predictions with the K–ε and K–ω platforms are quite similar except in a region
close to the wall. The angular velocity (figure 8) is also seen to be reasonably well
1.6
1.2
Uz / Umean
0.8
0.4
0.0
0.0 0.2 0.4 0.6 0.8 1.0
r/R
Figure 6. Axial velocity in rotating pipe flow for rotation ratio

Z = 1. Computations with Chien K–ε model ( ) and the cur-
rent EARSM based on Chien K–ε, EARSM 0 ( ) and EARSM 1
( ), compared to experiment by Imao et al.[14] ( ).
predicted with the different EARSM formulations, among which the differences
are small.
The prediction of the velocity could, however, be improved by the inclusion
of the neglected other cubic term associated with β5 (see (18)) and tuning the
coefficients for this special case, but it is worth noticing that the results here are
obtained without any tuning whatsoever. Moreover, three-dimensional effects
driven by turbulence are in most cases quite weak compared to three-dimensional
effects driven by mean momentum forces. The rotating pipe is in this case very
extreme since the three-dimensional effects are purely turbulence driven.
The zeroth- and first-order solutions of N are only different approximations
of the exact solution for N or P/ε. The error can be investigated by computing
the P/ε ratio using the different approximations of N from a given flow field.
Figure 9 shows P/ε with N evaluated from the first- (EARSM 0) and second-
(EARSM 1) order solutions of N given by (26) and (32) respectively, compared
to the exact solution. The mean flow invariants of these expressions were taken
from a fixed mean flow field, which was the solution with EARSM 0 based on
K–ε for Z = 1. We can see a substantial difference between the zeroth- and
first-order solutions and also that the first-order solution is quite close to the
exact one. As seen from the previous figures this difference has still a quite
small influence on the computed velocity profiles.
2.3. Solution of the general quasi-linear ARSM equation. So far, the

explicit solution of the ARSM resulting from the special case of the Launder et
al.[17] model with c2 = 5/9 has been analysed. The solution procedure described
in this section can, however, be applied to any ARSM that has been derived from
a linear or quasi-linear pressure–strain model. Quasi-linear here means that the
1.6
1.2
Uz / Umean
0.8
0.4
0.0
0.0 0.2 0.4 0.6 0.8 1.0
r/R
Figure 7. Axial velocity in rotating pipe flow for different rotation

ratios. Computations with the current EARSM (EARSM 0) based
on K–ω (lines) compared to experiment by Imao et al.[14] (symbols)
for Z = 1 ( ) ( ), Z = 0.5 ( ) ( ❛) and Z = 0 ( ) ( ).
1.00
0.80
Uθ / U wall
0.60
0.40
0.20
0.00
0.0 0.2 0.4 0.6 0.8 1.0
r/R
Figure 8. Angular velocity in rotating pipe flow for rotation ratio

Z = 1. Computations with Chien K–ε model ( ) and the current
EARSM based on Chien K–ε, EARSM 0 ( ) and EARSM 1 ( ),
and the current EARSM (EARSM 0) based on K–ω ( ) compared
to experiment by Imao et al.[14] ( ).
pressure–strain model must be tensorially linear in a but may contain terms like
tr{aS}a. The resulting ARSM may be written

2
N a = −A1 S + (aΩ − Ωa) − A2 aS + Sa − tr{aS} (40)
3
where
P
.N = A3 + A4 (41)
ε
The solution of (40) for three-dimensional mean flow was first derived by
Gatski & Speziale[7] where they considered the production to dissipation ratio
2.00
1.50
P/ε
1.00
0.50
0.00
0.0 0.2 0.4 0.6 0.8 1.0
r/R
Figure 9. Production to dissipation ratio in rotating pipe flow for

different approximation levels for N : EARSM 0 ( ) and EARSM 1
( ) compared to the exact ( ) solution. The mean flow from the
EARSM 0 K–ε solution for Z = 1 was used.
as a universal constant which is equivalent to A4 = 0. The fully consistent

solution for two-dimensional mean flow was first derived by Girimaji[8, 9] where
the production to dissipation ratio was obtained as part of the solution.
The solution of (40) for three-dimensional mean flow is given in Appendix C.
The algebra in that case is considerably more complex than for the simplified
ARSM with c2 = 5/9. In this section we restrict our attention to two-dimensional
mean flow for which the solution for the β-coefficients is
A1 N A1 A2 A1
β1 = − , β2 = 2 and β4 = − . (42)
Q Q Q
The denominator
2
Q = N 2 − 2IIΩ − A22 IIS (43)
3
for this case consists of both positive and negative terms but it can be shown
that it is strictly positive due to the fact that N is a function of IIS and IIΩ (see
Appendix C).
The Launder et al.[17] model (LRR) is a special case with the A1 to A4 coef-
ficients given in table 3. Also, the SSG model by Speziale, Sarkar & Gatski[39]
may be expressed in this form if it is linearized according to Gatski & Speziale[7].
The SSG pressure–strain model reads

Π C1 C∗ P C∗
= − + 1 a + C3 − 3 IIa S
ε 2 2 ε 2

C4 2 C5 C2 1
+ aS + Sa − tr{aS}I − (aΩ − Ωa) + a2 − IIa I ,(44)
2 3 2 4 3
where IIa = tr{a2 } and the coefficients are
4
C1 = 3.4, C1∗ = 1.8, C2 = 4.2, C3 = , C3∗ = 1.30, C4 = 1.25, C5 = 0.40.
5
(45)
The linearized SSG model used by Gatski & Speziale[7] is then obtained by
neglecting the quadratic anisotropy term and for the IIa invariant they used the
equilibrium value predicted by the SSG model for two-dimensional homogeneous
turbulence. This results in the following set of coefficients:
C1 = 3.4, C1∗ = 1.8, C2 = 0, C3 = 0.36, C3∗ = 0, C4 = 1.25, C5 = 0.40,

(46)
and A1 to A4 are given in table 3. Gatski & Speziale[7] based their EARSM
on this linearized SSG pressure–strain model and implied an additional approx-
imation in order to avoid the nonlinearity in the ARSM equation system. The
approximation was to use the asymptotic value for the production to dissipation
ratio as a universal constant
P Cε2 − 1
= (47)
ε Cε1 − 1
which is equivalent to the A1 to A4 -coefficients given in table 3 for Gatski &

Speziale without regularization. The approximation made by Gatski & Speziale[7]
is quite severe, especially since the A4 -coefficient is zero which means that the
denominator Q given by (43) may become singular. However, in their final ex-
pression they have regularized the denominator to avoid the singular behaviour.
That form of the model, however, could not exactly be expressed through the
A1 to A4 -coefficients.
The predicted anisotropies in the log-layer and for the asymptotic homoge-
neous shear flow are listed in tables 4 and 5. The linearized SSG model is able
to predict all individual Reynolds stresses fairly well for the two different cases.
The Gatski & Speziale[7] model gives the same result for the asymptotic homo-
geneous shear but fails in the log-layer since the approximation of the production
to dissipation ratio used by them is only consistent for asymptotic shear flows.
It should be noted here that the regularization in the Gatski & Speziale model
does not influence the solution for these two cases.
The proposed model based on the LRR pressure–strain model is a reasonable
choice here due to its simplicity in three-dimensional mean flows, and since the
most important anisotropies, namely a12 and a22 , are fairly well predicted for
both the log-layer and for the asymptotic homogeneous shear flow. If better
predictions of the a11 and a33 anisotropies are needed, the linearized SSG may be
considered as an alternative but one should realize that this model is considerably
more complex for three-dimensional mean flows.
Table 3. The coefficients in the general ARSM for different models.
A1 A2 A3 A4
88 5−9c2 11(c1 −1) 11
LRR 15(7c2 +1) 7c2 +1 7c2 +1 7c2 +1
Current model (c1 = 1.8, c2 = 5/9) 1.20 0 1.80 2.25
Original LRR (c1 = 1.5, c2 = 0.4) 1.54 0.37 1.45 2.89
linearized SSG 1.22 0.47 0.88 2.37
Gatski & Speziale w/o regularization 1.22 0.47 5.36 0
Table 4. The predicted anisotropy in the log-layer using different

models assuming balance between turbulence production and dissi-
pation compared to channel DNS by Kim[16].
a12 a11 a22 a33 σ

DNS −0.29 0.34 −0.26 −0.08 1.65
Current model (c1 = 1.8, c2 = 5/9) −0.30 0.25 −0.25 0.00 1.69
Original LRR (c1 = 1.5, c2 = 0.4) −0.36 0.26 −0.20 −0.06 1.40
Linearized SSG −0.32 0.36 −0.26 −0.10 1.59
Gatski & Speziale w/o regularization −0.29 0.22 −0.16 −0.06 1.75
Gatski & Speziale −0.28 0.22 −0.16 −0.06 1.76
Table 5. The predicted anisotropy in homogeneous shear flow us-

ing different models with σ = 3.0 compared to measurements by
Tavoularis & Corrsin[43].
a12 a11 a22 a33 P/ε

expr −0.30 0.40 −0.28 −0.12 1.8
Current model (c1 = 1.8, c2 = 5/9) −0.30 0.31 −0.31 0.00 1.8
Original LRR (c1 = 1.5, c2 = 0.4) −0.38 0.32 −0.25 −0.07 2.3
Linearized SSG −0.32 0.41 −0.30 −0.11 1.9
Gatski & Speziale w/o regularization −0.32 0.41 −0.30 −0.11 1.9
Gatski & Speziale −0.31 0.41 −0.30 −0.11 1.9
3. Near-wall treatments
In the model presented so far no special attention has been given to the very
near-wall region. To obtain the correct behaviour in this region it needs to
be modified in a similar way to low Reynolds number two-equation turbulence
models. An important difference compared to eddy-viscosity models is that the
effective Cµ or β1 in the proposed model is not a constant and, as has been
shown by Wallin & Johansson[45], will adjust to the near-wall flow in a more
natural way than is possible with eddy-viscosity models.
The turbulence timescale τ = K/ε, which is used to scale the strain- and
rotation-rate tensors goes to zero as the wall is approached. A more appropriate
expression for the timescale was proposed by Durbin[4] and reads

K ν
τ = max , Cτ . (48)
ε ε
This is just the usual timescale with a lower bound given by the Kolmogorov
scale. Durbin[4, 4] uses Cτ = 6.0 which will be kept in this study also.
In the very near-wall region of any shear flow the presence of the solid bound-
ary will enforce a nearly parallel flow except in the immediate vicinity of a sepa-
ration or stagnation point. A fully developed channel flow is exactly parallel and
will be used to formulate and calibrate the very near-wall correction. A specific
coordinate system will be used in the derivation of the near-wall correction but
the final form is invariant to coordinate system.
In a channel flow, as well as in all parallel flows, we can simply express the
anisotropy in terms of the two-dimensional β-terms and the non-dimensional
shear, σ, which in parallel flow reads
1 K dU
σ= . (49)
2 ε dy
The invariants can then simply be expressed as IIS = 2σ 2 and IIΩ = −2σ 2 and
the anisotropy becomes

1 1
a12 = σβ1 , a11 = σ 2
β2 − 2β4 , a22 = σ 2
β2 + 2β4 . (50)
3 3
By letting U in (49) represent the velocity along the limiting streamline at
the wall the expressions (50) can also be said to be approximately valid for
three-dimensional near-wall boundary layer flows.
The very near-wall behaviour is studied by using channel DNS data by Kim[16]
at Reδ ≈ 7800 or Reτ ≈ 395. The mean velocity, K and ε profiles obtained from
the DNS data have been used to compute the modelled anisotropy, which is
compared to the anisotropy determined directly from the DNS data.
The near-wall asymptotic behaviour can be written as
2
u+
rms = au y + + bu y + + ··· ,
2 3
+
vrms = av y + + bv y + + ··· ,
2
+
wrms = aw y + + bw y + + ··· , (51)
2 3
K+ = aK y + + bK y + + ··· ,
3 4
−uv + = auv y + + buv y + + ··· .
So, Lai & Zhang[36] summarize these near-wall asymptotic coefficients, au , av , ...,
for different experimental and numerical near-wall turbulence studies of flat
plates, channels and pipes at different Reynolds numbers. These coefficients
together with the DNS data by Kim[16] have been used to calibrate the coeffi-
cients in the near-wall corrections.
3.1. The shear component of the Reynolds stress. In a channel flow, the
mean flow is only directly affected by the Reynolds shear stress, uv, so let us
start by looking at the model of a12 .
The modelled a12 anisotropy without any near-wall corrections is nearly con-
stant as the wall is approached while DNS data exhibit a behaviour similar to an
exponential decay (see figure 10). The obvious choice of ‘wall damping function’
is of van Driest type

y+
f1 = 1 − exp − + . (52)
A
Also shown in figure 10 is the a12 behaviour predicted by a standard K–ε
model, without any near-wall damping function, which gives strongly negative
values near the wall, almost down to −2. Note that this is well outside the range
of physically realizable values, that are limited to be between ±1. The K–ε
model cannot be correctly damped towards the wall as easily as the EARSM
and the EARSM is therefore much better suited to be integrated down to the
wall. This is due to the fact that the β1 coefficient is not a constant, as in the
eddy-viscosity hypothesis, but a function of the mean flow strain rate. In the
very near-wall region the strain, normalized by the turbulent timescale, becomes
large but the β1 -coefficient goes to zero for large strain rates giving a balanced
a12 anisotropy (cf. figure 4).
The slope of the a12 anisotropy at the wall can be evaluated from the near-
wall asymptotic behaviour to be da12 /dy + = −auv /aK . The constant A+ varies
between 18 and 37 in the data summarized by So et al.[36]. By choosing A+ = 26,
which also is the standard value in the van Driest function, a good fit to the
DNS data is obtained according to figure 10, which shows the corresponding uv
Reynolds stress. The low Reynolds number coefficient can now be determined
as
β1,low-Re = f1 β1 , (53)
where β1 is the high Reynolds number coefficient obtained from the solution in
section 2.
3.2. The normal components of the Reynolds stress. The correct near-
wall behaviour for the normal Reynolds stresses is then ensured through a correct
behaviour of the β2 - and β4 -coefficients. The near-wall asymptotic behaviour of
the a11 and a22 anisotropy is
u2 2 au 2 2
a11 = − = − + O y+ ,
K 3 aK 3
v2 2 av 2 + 2 2 2
2
a22 = − = y − = − + O y+ . (54)
K 3 aK 3 3
0.00 0.00
-0.05 -0.15
-0.10 -0.30
-0.15 -0.45
uv
a12
-0.20 -0.60
-0.25 -0.75
-0.30 -0.90
-0.35 -1.05
0. 20. 40. 60. 80. 100. 120. 140. 0. 20. 40. 60. 80. 100. 120. 140.
y+ y+
Figure 10. a12 anisotropy (left) and uv stress (right) in channel

flow. Comparison of the current EARSM with near-wall correction
( ) with DNS data ( ❛) by Kim[16]. The computed a12 anisotropy
using the current EARSM without near wall correction ( ) and the
eddy-viscosity model ( ) is also shown. The predicted anisotropy
was evaluated by use of the DNS data for the S and Ω fields.
The modelled a11 and a22 anisotropies, without any wall corrections, are lim-
ited near the wall, so a blending of the near-wall asymptote and the outer solution
can be done. For simplicity, the same blending function, f2 , is used for the both
anisotropies, which become

2
a11,low-Re = f2 a11 + (1 − f2 ) B2 − ,
3

2
a22,low-Re = f2 a22 + (1 − f2 ) − , (55)
3
where B2 = au 2 /aK . Equation (54) states that a22 has zero slope at the wall
so the function f2 must also have zero slope. A simple choice that satisfies this
criterion is f2 = f1 2 . The low Reynolds number coefficients then become
3B2 − 4 B2
β2,low-Re = 2
1 − f1 2 , β4,low-Re = f1 2 β4 − 2 1 − f1 2 , (56)
2σ 4σ
where β4 is the high Reynolds number coefficient obtained from the solution in
section 2.
The constant B2 can be evaluated using the data summarized by So et al.[36]
and varies between 1.56 and 1.84. By choosing B2 = 1.8 a good fit to the DNS
data is obtained (see figures 11 and 12).
3.3. Extension to general flows. The near-wall correction described so far

was obtained for a special case, namely parallel two-dimensional flows, and can-
not directly be generalized to three-dimensional mean flows. Some singulari-
ties in flows near separation must be considered and the formulation needs to
be written in a coordinate-system-invariant form. These additional extensions
1.4 0.0
1.2 -0.1
1.0 -0.2
0.8 -0.3
a11
a22
0.6 -0.4
0.4 -0.5
0.2 -0.6
0.0 -0.7
0. 20. 40. 60. 80. 100. 120. 140. 0. 20. 40. 60. 80. 100. 120. 140.
y+ y+
Figure 11. a11 (left) and a22 (right) anisotropies in channel flow.
Comparison of the current EARSM with ( ) and without ( )
near-wall correction with DNS data ( ❛) by Kim[16]. The predicted
anisotropy was evaluated by use of the DNS data for the S and Ω
fields.
10.5 3.5
9.0 3.0
7.5 2.5
vv and ww
6.0 2.0
uu
4.5 1.5
3.0 1.0
1.5 0.5
0.0 0.0
0. 20. 40. 60. 80. 100. 120. 140. 0. 20. 40. 60. 80. 100. 120. 140.
y+ y+
Figure 12. Reynolds stresses in channel flow. Comparison of the

current EARSM, u2 ( ) (left) and v 2 ( ), w2 ( ) (right),
❛
with DNS data, ( ) and ( ) by Kim[16]. The predicted anisotropy
was evaluated by use of the DNS data for the S and Ω fields.
must be formulated such that the original form above is retained for parallel
two-dimensional flows.
So far, the near-wall correction is described in terms of σ, which is defined in
(49) for a specific coordinate system. By using IIS = 2σ 2 , a coordinate-system-
invariant form of the near wall correction can be obtained. In separated flow
the shear rate σ may become small, leading to a singular behaviour of the near
wall corrections to the β2 - and β4 -coefficients. To avoid this problem, the shear
rate in the denominator of the near-wall correction is limited to the equilibrium
shear rate where the turbulence production balances the dissipation rate. The
near-wall shear rate is always larger than the equilibrium one in zero pressure
gradient flows and the limiter will only be active in flows near separation. The
anisotropy model including the near-wall formulation then reads

3B2 − 4 1
a = f 1 β1 S + 1 − f 1 2
S − IIS I
2
max (IIS , IISeq ) 3

B 2
+ f 1 2 β4 − 1 − f 1 2 (SΩ − ΩS) (57)
2 max (IIS , IISeq )
where IISeq is the equilibrium value obtained by setting P = ε in the two-

dimensional solution and reads IISeq = 405c21 / (216c1 − 160) ≈ 5.74 for c1 = 1.8.
The β-coefficients are given by equation (24) and f1 is given by equation (52).
Please observe that even though the high Reynolds number β2 -coefficient is zero
the tensor group associated with the β2 -coefficient does not vanish.
The straightforward extension to three-dimensional flow reads

3B2 − 4 1
a = f 1 β1 S + 1 − f 1 2 S2
− IIS I
max (IIS , IISeq ) 3

1 B2
+ f12 β3 Ω2 − IIΩ I + f1 2 β4 − 1 − f1 2 (SΩ − ΩS)
3 2 max (IIS , IISeq )

2
+ f1 β6 SΩ2 + Ω2 S − IV I + f12 β9 ΩSΩ2 − Ω2 SΩ (58)
3
This form reduces to (57) for two-dimensional mean flows.

The near-wall correction is strictly valid only for parallel two-dimensional
mean flow but the very near-wall flow is, however, near parallel and two-dimen-
sional also for quite complex three-dimensional flow fields so the equation (58)
can be used as a first approximation. The correction can be extended to also be
valid for flows near stagnation points like separation and reattachment points,
for flows over curved surfaces and also for three-dimensional mean flows. This
extension of the near-wall correction is, however, outside the scope of this paper
but will be addressed in future studies.
3.4. Lengthscale determining equation. So far, we have proposed a new

explicit constitutive relation between the mean flow strain rate and the Reynolds
stresses including near-wall treatments. The choice of velocity- and length-scale
determining equations has not yet been discussed. The turbulent kinetic energy,
K, is a natural choice for determining the turbulent velocity-scale since this
equation needs no further modelling. The dissipation of the turbulent kinetic
energy, ε, is the most commonly used quantity for determining the turbulent
length-scale but other alternatives are also possible like τ and ω.
Common to all length-scale determining equations is that they need a lot
of ad hoc modelling, substantially more than the K-equation. The different
terms are quite empirical and are tuned for specific flow cases. There are many
different K–ε models with the major differences in the near-wall modelling in
the ε-equation and in the eddy-viscosity relation. Most of the ε-models are also
tuned and calibrated together with the eddy-viscosity assumption, which gives
a poor description of the near-wall anisotropy. One should thus not be surprised
if the proposed EARSM together with an existing length-scale model does not
give any improvements and perhaps also worse capabilities in predicting basic
wall-bounded flows.
The correct methodology to obtain a length-scale determining equation is to
develop it from scratch together with the proposed EARSM, so that one avoids
the risk of inheriting terms that are needed to balance the errors introduced by
the eddy-viscosity assumption. This is, however, outside the scope of this paper
but will be addressed in future studies. For an illustration of the capability of
the proposed model and to obtain an indication of how different length-scale
determining equations act, we have tested the proposed model together with the
well known Chien[1] K–ε model and the Wilcox[48] K–ω model for the fully
developed channel case simulated by Kim[16].
In these computations we did not use the Daly & Harlow model, equation (17),
of the diffusion terms in the K and ε (or ω) equations. The approach was to use
the standard eddy-viscosity modelling of the diffusion term but with an effective
Cµ evaluated from the EARSM such as Cµeff = −f1 β1 /2. The differences between
the Daly & Harlow and the eddy-viscosity approaches were found to be small
for the K–ε model as long as the effective Cµ is used rather than a constant. In
case of the K–ω model, the turbulent diffusion is complete different compared
to that of the K–ε model and thus the Daly & Harlow approach needs to be
recalibrated. The eddy-viscosity approach could, however, directly be used also
here as long as the effective Cµ (or β ∗ ) is used. In three-dimensional mean flows
the effective Cµ is Cµeff = −f1 (β1 + IIΩ β6 )/2
Figures 13 to 18 show the result of the computations with the two different
two-equation models based on the proposed EARSM as well as with the original
eddy-viscosity assumption. The EARSM has been used both with the original
and modified (‘mod’) length-scale equations.
If we first look at the velocity profiles in figure 13 one finds that the additive
constant B in the log-law is much too high when the EARSM is used. This
clearly illustrates the need for recalibration discussed above. The figure also
shows the results when the length-scale equation is tuned or modified to better
match the log-law. In the K–ε model the definition of ε in Chien’s model has
been changed to
2νK
ε = ε̃ + 2
exp −Ck y + , (59)
y
where the constant Ck = 0.04. This means that the original ‘wall dissipation’ is
multiplied by an exponential function to give a more rapid decay of the modifi-
cation near the wall. In the Wilcox K–ω model, the constant Rβ was increased
24.0 24.0
16.0 16.0
U+
U+
8.0 8.0
0.0 0.0
0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0
log y+ log y+
Figure 13. The streamwise velocity in channel flow. Computations

based on Chien K–ε (left) and K–ω (right) using eddy-viscosity ( )
and the current EARSM with the original ( ) and modified ( )
ε and ω equations compared with DNS data ( ❛) by Kim[16] and the
log-law ( ).
6.0 6.0
4.0 4.0
k+
k+
2.0 2.0
0.0 0.0
0. 40. 80. 120. 0. 40. 80. 120.
y+ y+
Figure 14. The turbulent kinetic energy in channel flow. Computa-

tions based on Chien K–ε (left) and K–ω (right) using eddy-viscosity
( ) and the current EARSM with the original ( ) and modified
( ) ε and ω equations compared with DNS data ( ❛) by Kim[16].
from 8 to 10 to obtain the desired behaviour. For a more thorough comparison

between K–ε and K–ω models the reader may wish to consult Menter[21].
It is also interesting to note the clear improvement in the K-profiles with the
proposed EARSM both for the K–ε and K–ω models in figure 14. The near-
wall dissipation ratio shown in figure 15 illustrates the need for a better near-wall
modelling of the length-scale equation. In figures 16 to 18 we can see the good
near-wall behaviour of the proposed EARSM which seems to be quite insensitive
to the choice of basic two-equation model. In figure 16 we can also clearly see
how poorly the a12 anisotropy is predicted with the eddy-viscosity models.
0.24 0.24
0.16 0.16
ε+
ε+
0.08 0.08
0.00 0.00
0. 40. 80. 120. 0. 40. 80. 120.
y+ y+
Figure 15. The dissipation of turbulent kinetic energy in channel

flow. Legend as in figure 14.
0.0 0.0
-0.1 -0.1
a12
a12
-0.2 -0.2
-0.3 -0.3
0. 40. 80. 120. 0. 40. 80. 120.
y+ y+
Figure 16. The a12 anisotropy in channel flow. Legend as in figure 14.
0.0 0.0
-0.4 -0.4
uv+
uv+
-0.8 -0.8
-1.2 -1.2
0. 40. 80. 120. 0. 40. 80. 120.
y+ y+
Figure 17. The uv Reynolds stress in channel flow. Legend as in

figure 14.
8.0 8.0
6.0 6.0
uu+ and vv+
uu+ and vv+

4.0 4.0
2.0 2.0
0.0 0.0
0. 40. 80. 120. 0. 40. 80. 120.
y+ y+
Figure 18. The normal Reynolds stresses in channel flow. Legends

as in figure 14.
The rotating pipe flow discussed in section 2.2.1 was computed with these
two models (K–ε and K–ω) as platforms for EARSM. Figures 6 and 7 show that
differences are seen in a region close to the wall, where the K–ε based model
shows a higher wall shear stress than the K–ω based model. The latter correctly
captures the trend of a decreasing skin friction with increasing rate of rotation,
whereas the K–ε based EARSM gives the opposite trend.
3.5. Alternative near-wall scaling. The damping function, f1 given by equa-

tion (52), is formulated in terms of y + . The scaling with the local wall skin fric-
tion is not valid in √
flows near separation and reattachment. Other possibilities
than y + are Rey ≡ Ky/ν or the turbulent Reynolds number Ret ≡ K 2 /νε (see
e.g. Wilcox[47]).
The approach is here to formulate an alternative scaling y ∗ in terms of Rey or
Ret so that y ∗ ≈ y + for y + ≤ 100 in channel flows. Rey and Ret were computed
from DNS data at different Reynolds numbers ranging from Reτ = 150 to 650 in
channel flows (Moser, Kim & Mansour[22]) and zero pressure gradient boundary
layers (Spalart[37]). It was found that Rey was increasing from the wall more
linearly with less scattering compared to Ret . Moreover, it should not pose
any major problems to compute the wall distance y needed for Rey if that is
defined as the distance to the closest wall point. The following form for y ∗ is
thus proposed:

y ∗ = Cy1 Rey + Cy2 Re2y . (60)

The Rey -term is motivated by the fact that the near-wall asymptotic behaviour
for Rey is ∼ y 2 . The Re2y -term is artificially introduced to obtain a near linear
relation in the buffer region also. With Cy1 = 2.4 and Cy2 = 0.003 good agree-
ment with y + for channel and zero pressure gradient boundary layer flows at
different Reynolds numbers is obtained, see figure 19.
150. 1.20
100. 0.80
y* f1
50. 0.40
0. 0.00
0. 50. 100. 150. 0. 50. 100. 150.
y+ y+
Figure 19. (left) The near-wall scaling y ∗ compared to y + . (right)

The damping function f1 using y ∗ or y + . The original form using y +
( ) is compared to channel DNS data (Moser et al.[22]) at different
Reynolds numbers Reτ = 180, 395 and 590 ( ) and zero pressure
gradient boundary layer DNS data (Spalart[37]) at Reτ = 150, 325
and 650 ( ). y ∗ increases with increasing Reynolds number.
In the damping function f1 , y + is then replaced by y ∗

y∗
f1 = 1 − exp − + . (61)
A
In figure 19 the damping function f1 based on y ∗ is compared to f1 based on
y + . The correspondence is good except for the lowest Reynolds numbers.
In some situations the y + scaling may worsen the computational results. In
a forthcoming study by M. Skote adverse pressure gradient boundary layers are
studied by DNS. For the highest pressure gradient studied, with the pressure
gradient similarity parameter m = −0.25 (U∞ ∼ xm ), the difference between
the y + and y ∗ approaches are significant here, see figure 20. For this case it is
obvious that the y + scaling degenerates the model performance and should be
avoided.
In other cases, where the Reynolds number is higher, the near-wall scaling is
not as critical. Figure 21 shows the computational results for the two-dimensional
RAE2822 aerofoil profile using the proposed EARSM compared to the Wilcox[48]
K–ω model. The flow in this case is compressible and the formulation is modified
in a manner described in section 4. The EARSM approach clearly improves the
position of the shock and the results are very much in line with differential
Reynolds stress computations by Hellström, Davidsson & Rizzi[12] for exactly
the same conditions and geometry.
The damping function in the EARSM is formulated in terms of y + as well
as y ∗ and the figure shows no major differences between these approaches ex-
cept in the separated region where the y + formulation gives a somewhat larger
negative skin friction. In the figure a computation using the proposed EARSM
0.004 30.0
0.003
20.0
y
Cf
0.002
10.0
0.001
0.000 0.0
150. 200. 250. 300. 350. 0.00 0.25 0.50 0.75
X U
Figure 20. (left) The wall skin friction and (right) the velocity pro-
file for an adverse pressure gradient boundary layer (U∞ ∼ x−0.25 ).
Computations with standard Wilcox[46] K–ω ( ), Chien K–ε
( ), EARSM based on K–ω and y + ( ) or y ∗ ( ) and the
Hanjalić[11] RST model ( ). Comparisons with DNS data ( ❛),
Skote (private communication).
without any damping functions whatsoever is also shown. For that choice, the
standard Wilcox[46] K–ω, also without damping functions, must be used as the
platform. This combination gives, however, incorrect near-wall behaviour for the
turbulence quantities but the mean velocity profiles are well reproduced. That
is also seen in the figure where no major differences between the EARSM with
or without damping functions are present.
The convergence history is shown in figure 22 where we can see that there
are no major differences in convergence rate. Actually, the proposed EARSM
converges to a somewhat lower residual than the corresponding eddy-viscosity
model for this case. The numerical parameters are the same for these two cases
and the computational time for the 5000 iteration steps is 6% higher for the
EARSM computation. This case was found to be not completely numerically
stationary which results in the residual ‘hanging’ as for the K–ω model. The
fluctuations are, however, very small and could not be seen in the solution. In
other cases without separation the convergence curves are even closer to each
other, and the convergence rates are in general faster than for the case shown
in figure 22. The computational results were obtained using the EURANUS
code[25] which is an explicit time-stepping multigrid and multiblock Navier–
Stokes solver. The grid convergence was assessed by repeating the computation
on a coarser grid.
A further example is a three-dimensional transonic supercritical wing (figure
23). This case is computed using the standard Wilcox[46] K–ω model and the
EARSM based on that. Since that K–ω model has no damping functions the
EARSM without damping functions must also be used for consistency. Again,
-1.50 0.010
-1.00
0.006
-0.50
Cp Cf
0.002
0.00
-0.002
0.50
1.00 -0.006
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
x/c x/c
Figure 21. Wall pressure and skin friction coefficients for the
RAE2822 wing profile (M = 0.754, α = 2.57◦ and Re = 6.2 · 106 ).
Predictions using Wilcox[48] K–ω ( ) and the current EARSM
based on K–ω with damping function based on y + ( ), y ∗ ( )
or without any damping functions ( ), compared to experimental
data ( ❛) (Cook et al.[2]). The geometry is the measured one including
a camber correction.
0.0
-2.0
rho_rms
-4.0
-6.0
0. 2000. 4000. 6000.
iter
Figure 22. Convergence history for the RAE2822 wing profile.

Wilcox K–ω ( ) compared with the current EARSM based on
K–ω ( ). The computational time is increased by 6% by using
EARSM. Three levels of full multigrid is used.
the predicted shock position is improved compared to the eddy-viscosity model.

This case illustrates the benefit of the proposed model which is also viable in ex-
pensive three-dimensional cases, here with almost one million grid points. Most
importantly, there is no substantial increase in the computational cost compared
to the, in many cases, robust standard K–ω model.
-1.40
-0.60
Cp
0.20
1.00
0.00 0.25 0.50 0.75 1.00
x/c
Figure 23. Wall pressure coefficient Cp for the LANN-wing (M =

0.82, α = 2.59◦ and Re = 5.4×106 ) at the spanwise position of 0.474.
Predictions using Wilcox[46] K–ω ( ) and the current EARSM
( ) based on K–ω without wall damping functions. Comparison
with experimental data ( ❛) (Horsten et al.[13]).
4. Compressibility
Many turbulent flow applications are within the Mach number regime where
compressible effects must be considered. Also, in low speed flow compressible ef-
fects may be important due to local heating or cooling of the flow. Compressible
turbulent flow may be classified according to Friedrich[6] into flows with van-
ishing compressibility effects due to turbulent fluctuations and flows where such
effects play a significant role. Friedrich[6] also states that compressibility effects
due to turbulent fluctuations might be important in hypersonic, high Mach num-
ber, wall-bounded flows and in mixing layers at high convective Mach numbers.
The understanding of such flows is poor and models capable of distinguishing the
principal differences between wall-bounded and free shear flows are not well de-
veloped. The compressibility effects, due to turbulent fluctuations, increase the
anisotropy of the Reynolds stress tensor (see Friedrich[6]) and depend themselves
on the anisotropy, which means that algebraic Reynolds stress models are better
suited than eddy viscosity models to act as a basis for improving the prediction
of compressibility effects.
In wall-bounded flows with Mach numbers below 5 compressibility effects due
to turbulent fluctuations may be neglected and the effect of compressibility enters
into the problem essentially only through the mean flow compressibility. In this
study we will restrict the attention to this class of flows and a straightforward
compressibility extension of the incompressible model will be made.
First, the stress anisotropy and the turbulent kinetic energy must be redefined
as aij ≡ ρui uj /ρK − 2δij /3 and K ≡ ρui ui /2ρ where ρ is the local mean density
of the fluid. The trace of the strain is not zero for compressible flow. We may
instead use a somewhat redefined normalized strain rate tensor
τ D
Sij ≡ (Ui,j + Uj,i ) − δij , (62)
2 3
where the normalized dilatation of the mean flow is defined as D ≡ τ Uk,k .

The redefinition of S to have zero trace allows us to make use of the in-
compressible solution process. The redefinition also implies, however, that the
S33 -component is non-zero for two-dimensional mean flow in the (x, y)-plane and
that the simplifications for two-dimensional mean flow are not strictly valid for
compressible flow. This will be addressed as a special case later in this section.
The general linear model of the LRR rapid pressure–strain rate model for
incompressible flow does not have zero trace in compressible flow, so the model
needs to be generalized according to Vandromme[44] and reads

(r) c2 + 8 2 30c2 − 2 2
Πij = − Pij − Pδij − ρK Ui,j + Uj,i − Ul,l δij
11 3 55 3

8c2 − 2 2
− Dij − Pδij (63)
11 3
where Pij = −ρui uk Uj,k − ρuj uk Ui,k and Dij = −ρui uk Uk,j − ρuj uk Uk,i . The
incompressible models for the slow pressure–strain and the dissipation tensor can
also be used here. The general ARSM for compressible flow can now be written
as

6c2 + 4 8 7c2 + 1
c1 − 1 − D − tr{aS} a = − S + (aΩ − Ωa)
11 15 11

5 − 9c2 2
− aS + Sa − tr{aS}I (64)
11 3
which is identical to the incompressible form, equation (13), except for the di-
latation term 0n the left-hand side and the different definitions given above. The
same conclusion is valid for the simplified equation obtained by setting c2 = 5/9,
which reads

2 8 4
c1 − 1 − D − tr{aS} a = − S + (aΩ − Ωa) . (65)
3 15 9
The solution of the simplified compressible ARSM equation is the same as the
incompressible solution except for the definition of the c1 -coefficient, which for
the compressible case is

9 2
c1 = c1 − 1 − D (66)
4 3
bearing in mind that Sij here is defined by (62).

4.1. Compressible two-dimensional mean flow. An approximation of the

compressible ARSM equation can be derived by an expansion around a so-
lution obtained by use of the two-dimensional simplification valid in incom-
pressible flow. Let us first define a compressible two-dimensional strain rate
2D
as Sij ≡ (τ /2) (Ui,j + Uj,i ) − Dδij
2D 2D
/2 where δij ≡ δij except that δ33
2D
= 0. The
compressible ARSM, equation (65), can now be written as

2 8 4 2D
c1 − 1 − D − tr{a S } a2D = − S2D +
2D 2D
a Ω − Ωa2D (67)
3 15 9
with the solution a2D , which is derived using the two-dimensional solution pro-
cess for incompressible flow. The true anisotropy, a, is then related to the solu-
tion a2D , as described below.
The difference between S and S2D is of O (D) which can be assumed small
except in a shock wave. For the same reason the difference between a and a2D
can also be assumed to be of O (D) and with the simplest tensorial form that
relates the two-dimensional and three-dimensional anisotropies, leading to the
following assumption. Let

2
a = a + βa D I − I + O D 2 ,
2D 2D
(68)
3
and subtract equation (67) from equation (65). Sorting terms in powers of D
gives us the first-order solution in D
4
βa = − . (69)
15 (c1 − 1 − tr{a2D S2D })
In obtaining this solution the following relations have been used:

D 2D 2
S = S−
2D
I − I
2 3
2
tr{a S } = tr{aS} + O D
2D 2D
(70)

a Ω − Ωa2D = aΩ − Ωa + O D2 .
2D
The zeroth-order solution, a2D , is obtained as described for the incompressible

case, with the only difference being that c1 is now given by (66). The full solution
is then given by

2D 2
a = β1 S + β4 S Ω − ΩS
2D 2D
+ βa D I − I + O D 2
2D
(71)
3
with βa from (69).
4.2. Shock/boundary layer interaction. The compressible form of the pro-

posed model together with the near-wall formulations has been tested on a
shock/boundary layer interaction. A turbulent boundary layer at Mach 5 on
a flat plate interacts with an oblique shock from a shock generator above the
plate, see figure 24. The flow deflection angle is 10◦ which gives a strong enough
999 9999999999999999 Computational

999 9999999999999999
Shock generator domain
9999999999999999
9999999999999999 Expansion fan
9999999999999999
Incoming shock Reflected shock
Boundary layer Interaction
999
999999999999999999999999999999999
999
999999999999999999999999999999999
Figure 24. Impinging shock at Mach 5. Experimental and compu-
tational setup.
shock to cause a boundary layer separation. The experiment by Schülein, Krog-

mann & Stanewsky[29] was performed at DLR in Göttingen. Figure 25 shows
the computed wall pressure and skin friction coefficient along the flat plate com-
pared to the experimental data. The turbulence models used are the Chien[1]
K–ε model and the Wilcox[48] K–ω model based on the original eddy-viscosity
assumption and also based on the proposed EARSM. The computational results
were obtained using the EURANUS code[25]. The grid convergence was assessed
by repeating the computation on a coarser grid.
The size of the separated region is underestimated by the two eddy viscos-
ity models whereas a correct separation length is obtained when the models are
based on the proposed EARSM. This can be seen both in the skin friction be-
haviour and in the wall pressure distribution. The computed wall pressure is,
however, somewhat shifted downstream.
The Chien K–ε model strongly over-predicts the skin friction downstream
of the reattachment point, which is typical for many low Reynolds number K–ε
models. This behaviour is inherited by the proposed EARSM based on the Chien
K–ε model. The K–ω models give much better skin friction predictions.
5. Diffusion term
In regions of the flow where the production to dissipation ratio is small, the
assumption of negligible effects of advection and diffusion of the anisotropy may
cause problems, also noticed by Taulbee[41]. In the outer-most part of a bound-
ary layer and in the centre of a turbulent channel flow the magnitude of the β1
coefficient may thus be too large leading to an overestimation of the uv Reynolds
stress. The behaviour of the model can be analysed by looking at the effective
Cµeff ≡ −β1 /2 when the strain rate goes to zero, which for the simplified ARSM
reads
3
Cµeff (σ → 0) = . (72)
5c1
For the proposed model constants Cµeff ≈ 0.33 which is far too high.
45. 0.010
30. 0.006
P wall [kPa]
Cf
15. 0.002
0. -0.002
0.30 0.36 0.42 0.48 0.30 0.36 0.42 0.48
x [m] x [m]
Figure 25. (left) Wall pressure and (right) skin friction coefficient
for the impinging shock at Mach 5 with the flow deflection angle
β = 10◦ . Comparison of the predictions using Chien K–ε ( ),
Wilcox K–ω ( ) and the current EARSM based on K–ε ( )
and K–ω ( ) with the experiment by Schülein et al.[29] ( ❛).
In this section we will discuss the possibility of including a correction to the

ARSM equations formulated so that the proposed EARSM solution process can
be retained. The transport equation for the Reynolds stress anisotropy is given
(K)
by (1) where Tijl and Tl are the flux terms for the Reynolds stress tensor and
turbulent kinetic energy. A simple model for the neglected diffusion term may
be written as
(K) (K)
∂Tijl ui uj ∂Tl ∂Tl
− = CD aij , (73)
∂xl K ∂xl ∂xl
which gives a modification of the c1 -coefficient in the simplified ARSM equation
(21). The modified coefficient reads

(K)
9 CD ∂Tl
c1 = c1 − 1 − . (74)
4 ε ∂xl
There are, however, numerical problems associated with this form due to the
inclusion of the second derivative of K. This could be avoided if the term that
balances the turbulent flux term in the K-equation (15) is used to approximate
the turbulent flux of K:
(K)
∂Tl
≈ P − ε. (75)
∂xl
The advection of K is neglected here. Moreover, the extra term worsens the
model behaviour for large strain rates and can even lead to a singular behaviour.
The correction thus needs to be switched off for P/ε > 1. The definition of the
c1 -coefficient then becomes

9 P
c1 = c1 − 1 + CD max 1 − , 0 . (76)
4 ε
The problem with this form is that it is implicit in the P/ε ratio and that the
max function makes it impossible to solve. The diffusion model will thus be
further approximated by using P/ε = −β1 IIS . If β1 is approximated with β1eq
where P = ε, the correction is guaranteed to be zero in the log-layer and for
higher strain rates. The definition of the c1 -coefficient now becomes
9
c1 = [c1 − 1 + CD max (1 + β1eq IIS , 0)] (77)
4
where
6 N eq
β1eq = − (78)
5 (N eq )2 − 2IIΩ
and
9c1
N eq = . (79)
4
The constant CD can now be estimated by looking at the effective Cµ for
zero strain rates in (72). With CD = 2.2 the effective Cµ becomes Cµeff = 0.09
which is close to what one should expect. The a12 anisotropy with the proposed
diffusion model is compared to the basic EARSM in figure 26 for parallel flow
(cf. figure 4). The inclusion of the diffusion model makes the proposed EARSM
behave almost the same as the SST model for this case. In figure 27 we can see
an improvement of the a12 anisotropy prediction near the centre of the channel
simulated by Kim[16]. The actual difference in the predicted mean velocity
profile is quite small. This improvement is similar in character to that obtained
by the correction introduced by Taulbee[41] although the two approaches are
motivated differently.
The derivation of the diffusion correction is based on a number of ad hoc
arguments and should be seen only as a possible extension of the model to flow
situations where diffusion may be of substantial importance.
6. Concluding remarks
For two-dimensional mean flows the new proposed EARSM represents an exact
solution of the implicit ARSM relation for the anisotropy tensor, but also a good
approximation for three-dimensional mean flows. The fact that it fully accounts
for three-dimensional effects gives it a natural predictive capability for complex
flows. This was demonstrated here by capturing the correct trends for axially
rotating pipe flow. Standard two-equation models predict a solid-body rotation
with an unaffected axial velocity profile. In reality the azimuthal velocity has
a variation that is close to parabolic and the rotation causes the axial velocity
profile to be less full, i.e. more parabolic. To capture these features the model
description of the inter-component transfer is crucial. For instance, the simplified
0.0
-0.1
-0.2 hom. shear
-0.3
a12
-0.4 log-layer
-0.5
-0.6
0.0 0.5 1.0 1.5 2.0 2.5 3.0
σ
Figure 26. The a12 anisotropy versus strain rate σ for parallel flow.
The current model without ( ) and with ( ) diffusion model
compared to an eddy-viscosity model ( ) and the Menter[20] SST
model ( ).
0.00
-0.05
-0.10
-0.15
a12
-0.20
-0.25
-0.30
-0.35
0. 100. 200. 300. 400.
y+
Figure 27. The a12 anisotropy in channel flow. Comparison of the

current EARSM without ( ) and with ( ) diffusion model with
the eddy-viscosity model ( ) and DNS data ( ❛) by Kim[16].
linear rapid pressure–strain model of Launder et al.[17] and Naot, Shavit &
Wolfstein[23], usually referred to as the isotropization of production model, does
not have the necessary ingredients. The full linear model, on the other hand,
together with linear models for the other terms, was demonstrated here to be
sufficient to capture the main features of the flow in the axially rotating pipe.
In general the predictive capability of the proposed EARSM for rotating flows
is substantially better than standard two-equation models and than EARSM
formulations including only terms up to second order in S and Ω.
The extension to compressible flows is done in a simple way in the present
EARSM in which the equations are expressed by use of Favre averages. Com-
pressibility of the mean flow is accounted for, but no explicit compressibility
corrections are added to the ‘platform’ (K–ε, K–ω . . . ) equations. With some
minor redefinitions the same ARSM relation for aij as for incompressible flow
could be used also here. This approach for wall-bounded turbulent flows is ad-
equate for Mach numbers up to about 5. It was shown to capture the essential
features of the complex interaction between an inclined shock and a turbulent
boundary layer at Mach 5. In particular it describes the skin friction and sep-
aration length much better than standard two-equation models. Compared to
the standard two-equation models on which the present EARSM is based it also
gives a substantial improvement for the prediction of the skin friction variation in
incompressible boundary layers with adverse pressure gradient (detailed results
not reported here).
Poor prediction of the effects of rotation and the under-prediction of separa-
tion tendency in adverse pressure gradient boundary layers are well known defi-
ciencies of closures based on the eddy-viscosity hypothesis. The present EARSM
substantially improves these two aspects, and reduces the need for wall damping.
A simple way to obtain the correct near-wall limits of the anisotropies was con-
structed here on the basis of a van Driest type of damping function. For channel
flow this was shown to give a very accurate description, and a simple method to
retain numerical stability for situations with separated
√ flows was demonstrated.
A form of wall damping function based on Rey ≡ Ky/ν was shown to be an
attractive alternative to forms based on y + in flows with (or near) separation.
The reduced need for wall damping is coupled to the fact that the present
EARSM automatically predicts a production to dissipation ratio with correct
asymptotic behaviour for large strain rates. For instance, in parallel flows this
correctly gives an asymptotically constant shear stress anisotropy component
for large strains (shear rates). This behaviour is not a natural part of eddy-
viscosity based models, but was incorporated in a somewhat ad hoc manner by
Menter[21]. In the present model this behaviour emanates naturally from the
underlying modelling of the terms in the RST equations and the direct solution
of the production to dissipation ratio.
To further illustrate the behaviour for large strain rates the proposed model
was tested in homogeneous shear flow at high initial shear rate. The good predic-
tions that was obtained for this case can, however, be regarded as a bit fortuitous
in the light of the fact that the basic ARSM approximation is somewhat ques-
tionable for a case with a significantly non-zero left-hand side of the aij transport
equation (∂aij /∂t = 0). Moreover, the Launder et al.[17] model does not give
accurate predictions of this case when used in a differential form. It is, however,
a fact that the self-consistent approach (i.e where the production to dissipation
ratio is solved for as part of the total EARSM-solution) gives a model with the
correct asymptotic behaviour, which is a pre-requisite for reasonable predictions
in the limit of high shear. The present model is not constructed specifically to
incorporate such effects. It is, however, an important step towards a more gen-
eral engineering model that the model is also able to give reasonable results in
extreme flow cases.
Perhaps the most important feature of the proposed model is the numeri-
cal behaviour. The computational cost is not significantly increased compared
to standard eddy-viscosity two-equation models and the general numerical be-
haviour is almost the same. Implementation of this model into flow solvers with
existing eddy-viscosity two-equation models should not pose major problems.
The model may be formulated in terms of an effective eddy viscosity with an
additional correction that may be treated fully explicitly (see Appendix A).
The authors would like to thank Dr Torbjörn Sjögren and Dr Magnus Hallbäck
for many helpful discussions, especially concerning the treatment of rotating co-
ordinate systems. The first author would also like to thank Dr Sharath Girimaji
for many helpful discussions, especially concerning non-equilibrium turbulence.
The authors gratefully acknowledge funding of this study from the European
Space Agency within a project for improving the modelling of turbulent flows
related to hypersonic lifting vehicles and would like to acknowledge Dr Ingemar
Lindblad for leading this project.
Appendix A. Summary of the proposed model

The Reynolds stresses may be written in terms of an effective Cµ -coefficient
which is mathematically identical with the formulation in section 2:

2 (ex)
ui uj = K δij − 2Cµeff Sij + aij , (80)
3
where the effective Cµ -coefficient is
1
Cµeff = − f1 (β1 + IIΩ β6 ) , (81)
2
(ex)
and the extra anisotropy aij then becomes

3B2 − 4 1 1
a(ex) = 1 − f1 2
+ f1 β3 Ω − IIΩ I
2 2
S − IIS I
2
max (IIS , IISeq ) 3 3

B2
+ f 1 2 β4 − 1 − f 1 2 (SΩ − ΩS)
2 max (IIS , IISeq )

2
+ f1 β6 SΩ2 + Ω2 S − IIΩ S − IV I + f12 β9 ΩSΩ2 − Ω2 SΩ . (82)
3
Here a, S and Ω denote second rank tensors, tr{} denotes the trace and I is
the identity matrix. The inner product of two matrices is defined as (SS)ij ≡
2
S ij ≡ Sik Skj . The normalized mean strain and rotation tensors are defined
as
τ τ
Sij =
(Ui,j + Uj,i ) , Ωij = (Ui,j − Uj,i ) , (83)
2 2
where the turbulent timescale is defined by

K ν
τ = max , Cτ . (84)
ε ε
The invariants are defined by
IIS = tr{S2 }, IIΩ = tr{Ω2 }, IV = tr{SΩ2 }, V = tr{S2 Ω2 }, (85)
and
405c21
IISeq = . (86)
216c1 − 160
By introducing an effective Cµ -coefficient one can easily introduce this level
of modelling into flow solvers with existing two-equation eddy-viscosity models
by setting
νt = Cµeff Kτ. (87)
(ex)
The contribution from the extra anisotropy aij may now be added as fully
explicit additional terms in the equations.
The β-coefficients are given by

N 2N 2 − 7IIΩ 12N −1 IV
β1 = − , β3 = − ,
2 Q Q (88)
2 N − 2IIΩ 6N 6
β4 = − , β6 = − , β9 = ,
Q Q Q
with the denominator
5 2
Q= N − 2IIΩ 2N 2 − IIΩ . (89)
6
For most purposes it is sufficient to take N = Nc where

1/3

c1 + P1 + P2 + sign P1 − P2 | P1 − P2 |1/3 , P2 ≥ 0
3
Nc = c1 2 1/6 1 P1
+ 2 P1 − P2 cos arccos 2 , P2 < 0
3 3 P1 − P2
(90)
with
3
c1 c1
2 2
9 2 9 2
P1 = + IIS − IIΩ c1 , P2 = P12 − + IIS + IIΩ (91)
27 20 3 9 10 3
and
9
c1 =
(c1 − 1) . (92)
4
An additional term can be added to Nc in order to improve the accuracy for
three-dimensional mean flows (see section 2.2).
The simplifications for two-dimensional mean flow are

6 Nc 6 1
β1 = − , β4 = − , β3 = β6 = β9 = 0. (93)
5 Nc2 − 2IIΩ 5 Nc2 − 2IIΩ
The damping function reads

f1 = 1 − exp −Cy1 Rey − Cy2 Re2y , (94)
where
√
Ky
Rey = . (95)
ν
Finally, the five model constants are
2.4 0.003
Cτ = 6.0, c1 = 1.8, B2 = 1.8, Cy1 = , Cy2 = . (96)
26.0 26.0
Appendix B. The completeness of the a(S, Ω) expression with ten

terms
Based on the derivations of Spencer & Rivlin[38] for matrix polynomials Pope[24]
concluded that there are ten independent terms in the complete expression
a(S, Ω). One should here restrict the meaning of independence to polynomial
independence as was also pointed out by Taulbee, Sonnenmeier & Wall[42]. The

highest-order term ΩS2 Ω2 − Ω2 S2 Ω in expression (18) is of order 5 and ex-
tension 3.
The S and Ω tensors can appear in powers up to 2. A term Sm Ωn is then
said to be of power m + n and extension 2, etc. Shih & Lumley[34] (see also
Shih[33]) also include a term of power 6 and extension 4. They use a notation
with the mean velocity gradient tensor and its transpose, but with the present
notation such a term can be written as
2
SΩ2 S2 Ω − ΩS2 Ω2 S − VI I, (97)
3
where the invariant VI equals tr{SΩ2 S2 Ω}. One may first note that the invariant
VI is of third power in Ω, which means that it cannot be expressed in terms of a
polynomial in the other invariants, which are of order zero or two in Ω. Hence,
VI contains sign information that is not contained in the other invariants. We
can derive the following relation for its square:
IIΩ 3 IIΩ 2
VI 2 = 9IIS 3 + 2IIIS 2 + −3V IIS 2 + IV IIS IIIS
144 6
IIΩ 1 1
+ 10 V IIS − 4V IV IIIS − IV IIS − V 3 + V IV 2 IIS − IV 3 IIIS .
2 2 2
8 2 3
In the present modelling the invariant VI will not appear since it would arise
from the product of the highest (fifth-order) term and S, and even in the general
solution (for c2 = 5/9) the coefficient β10 = 0.
We can furthermore express the sixth-order term as the following combination

of lower-order terms

2 2 1
SΩ2 S2 Ω − ΩS2 Ω2 S − VI I = V + IIS IIΩ (SΩ − ΩS)
3 3 12
2 1 1
+ 3 IV S Ω − ΩS − 6 IIΩ SΩS − S ΩS + 6 IIS ΩSΩ2 − Ω2 SΩ .
1 2 2 2
By use of the generalized Cayley–Hamilton theorem it is fairly straightforward

to show that no other independent sixth- or higher-order term exists (see e.g.
Shih & Lumley[34].
Appendix C. Solution of the general quasi-linear ARSM equation

The implicit general quasi-linear ARSM equation is rewritten similarly to the
simplified equation (21)

2
N a = −A1 S + (aΩ − Ωa) − A2 aS + Sa − tr{aS} , (98)
3
where
P
N = A3 + A4. (99)
ε
The simplified equation is now a special case of (98) with the coefficients A1 =
6/5, A2 = 0, A3 = c1 and A4 = 9/4. For the general case we require that A1 ,
A3 and A4 are positive.
C.1. Solution for two-dimensional mean flow. The solution for two-dimensional
mean flow is
A1 N A1 A2 A1
β1 = − , β2 = 2 , β4 = − , (100)
Q Q Q
where the denominator
2
Q = N 2 − 2IIΩ − A22 IIS . (101)
3
The equation for N is, for two-dimensional mean flow,

2 2 1 2
N − A3 N −
3 2
A1 A4 + A2 IIS + 2IIΩ N + 2A3 A IIS + IIΩ = 0,
3 3 2
(102)
with the solution

1/3

A3 + P1 + P2 + sign P1 − P2 | P1 − P2 |1/3 , P2 ≥ 0
3
N= A3 2 1/6 1 P1
+ 2 P1 − P2 cos arccos 2 , P2 < 0
3 3 P1 − P2
(103)
where

A3 2 A1 A4 2 2
P1 = + − A2 2 IIS − IIΩ A3 (104)
27 6 9 3
and
3
A3 2 A1 A4 2 2
P2 = P12 − + + A2 2 IIS + IIΩ . (105)
9 3 9 3
Let us investigate whether the denominator in equation (101) can become
zero. Equation (102) can be rewritten, by using equation (101), as
Q (N − A3 ) = A1 A4 IIS N, (106)
which shows that Q > 0 if N > A3 . The problem is then to show that N > A3 .
Let us first look at the special case when IIS = 0. Equation (102) can then be
written

(N − A3 ) N 2 − 2IIΩ = 0, (107)
with the only real root N = A3 . Differentiating equation (102) with respect to
IIS gives
∂N 2 2
3A (N − A3 ) + A1 A4 N
= 2 , (108)
∂IIS 3 N − 3 A3 N − A1 A4 + 23 A22 IIS − 2IIΩ
2
which is positive if N ≥ A3 . When IIS = 0 (108) can be written as

∂N A1 A3 A4
= 2 > 0. (109)
∂IIS IIS =0 A3 − 2IIΩ
So far we know that N = A3 for IIS = 0 and that N increases until (108) changes
sign. By using (102) IIΩ can be eliminated in (108) which then becomes
2
∂N 2 2
3 A2 (N − A3 ) + A1 A4 N (N − A3 )
= 2 . (110)
∂IIS 2N (N − A3 ) + A1 A3 A4 IIS
From this form it is obvious that the denominator is strictly positive and that
∂N/∂IIS = 0 for IIS = 0 only (where N = A3). Hence, the numerator is strictly
positive and N increases for all IIS , showing that N ≥ A3 for all IIS . The
corresponding relation for the simplified ARSM equation is N ≥ c1 .
The denominator Q is thus always positive which guarantees a non-singular
solution.
C.2. Solution for three-dimensional mean flow. The solution of the linear
equation system where N is assumed as known can be formulated as (see Gatski
& Speziale[7])

N βλ = −A1 δ1λ + Jλγ βγ − A2 Hλγ βγ (111)
γ γ
or in the standard form for linear equation systems
(N δγλ − Jγλ + A2 Hγλ ) βλ = −A1 δ1γ (112)

where the matrixes are given below for three-dimensional mean flow
 
0 1
3 IIS − 23 IIΩ 0 0 3 IV − 3 V
2 1
0 0 0
2 
 0 0 0 0 2IIΩ IV 0 0 0 
 
0 0 0 0 0 IIS 13 IIIS 0 0 0 
 
0 0 0 0 12 IIS 0 0 1 1 
3 IIIS −IV − 3 V − 6 IIS IIΩ 
1

0 0 0 1 0 0 0 0 −IIΩ − 3 IV
2 
H=
0 1
,

 0 1 0 0 0 2 IIS 0 0 0 
0 −1 
 0 0 0 0 0 0 0 0 
 
0 0 0 0 −1 0 0 0 0 1
II Ω 
 3 
0 0 0 0 0 0 0 0 0 − 13 IIS 
0 0 0 0 0 0 0 0 −2 0
(113)
 
0 0 0 −IIΩ 0 0 0 2V − IIS IIΩ IIΩ2 0
0 −IIΩ −2IV IIΩ2 
 0 0 0 0 0 0 
 
0 0 0 0 −2IIS 0 0 0 −2IV 2IIS IIΩ − 2V 
 
1 0 0 0 0 1
II 0 0 0 0 
 2 Ω 
0 1 0 0 0 0 1
II 0 0 0 
J=
0
2 Ω .

 0 0 3 0 0 0 IIS −2IIΩ 0 
0 −2IIΩ 
 0 0 0 3 0 0 0 0 
 
0 0 0 0 0 0 0 0 0 0 
 
0 0 0 0 0 −1 0 0 0 0 
0 0 0 0 0 0 −1 0 0 0
(114)
The solution for three-dimensional mean flow is
1
β1 = − A1 N 30A2 IV − 21N IIΩ − 2A2 3 IIIS + 6N 3 − 3A2 2 IIS N /Q,
2

β2 = −A1 A2 6A2 IV + 12N IIΩ + 2A2 3 IIIS − 6N 3 + 3A2 2 IIS N /Q,

β3 = −3A1 2A2 2 IIIS + 3N A2 IIS + 6IV /Q,

β4 = −A1 2A2 3 IIIS + 3A2 2 N IIS + 6A2 IV − 6N IIΩ + 3N 3 /Q,
β5 = 9A1 A2 N 2 /Q,
β6 = −9A1 N 2 /Q,
β7 = 18A1 A2 N/Q,
β8 = 9A1 A2 2 N/Q,
β9 = 9A1 N/Q,
β10 = 0, (115)
References 91
with the denominator

15 7 2
Q = 3N + −5
IIΩ − A2 IIS N 3 + ( 21 A2 IV − A2 3 IIIS ) N 2
2 2
2
+ ( 3 IIΩ − 8 IIS IIΩ A2 2 + 24 A2 2 V + A2 4 IIS 2 ) N + A2 5 IIS IIIS
2
3
+ 2 A2 3 IV IIS − 2 A2 3 IIΩ IIIS − 6 IV A2 IIΩ . (116)
By substituting N = 1, A1 = 1 and A2 = 1 one gets the same ARSM equation

as Gatski & Speziale[7] and also an identical solution.
The equation for N can also be derived for three-dimensional mean flow and
is a sixth order polynomial equation, as for the simplified case. The expression
is, however, complicated and of small practical interest. As for the solution of
the simplified ARSM equation, the two-dimensional solution of N can also be
used here as a first approximation.
References
[1] Chien, K.Y. 1982 Predictions of channel and boundary-layer flows with a
low-Reynolds-number turbulence model. AIAA J. 20, 33–38.
[2] Cook, P.H., MacDonald, M.A. & Firmin, M.C.P. 1979 Aerofoil 2822
- Pressure distributions, boundary layer and wake measurements. AGARD AR
138.
[3] Daly, B.J. & Harlow, F.H. 1970 Transport equations in turbulence. Phys.
Fluids 13, 2634–2649.
[4] Durbin, P.A. 1993 Application of a near-wall turbulence model to boundary
layers and heat transfer. Int J. Heat Fluid Flow 14, 316–323.
[5] Durbin, P.A. 1995 Separated flow computations with the K–ε–v 2 model.
AIAA J. 33, 659–664.
[6] Friedrich, R. 1996 Compressibility effects due to turbulent fluctuations.
Oral presentation at Sixth European Turbulence Conference, Lausanne, July
1996, submitted to Journal of Appl. Sci. Res.
[7] Gatski, T.B & Speziale C.G. 1993 On explicit algebraic stress models for
complex turbulent flows. J. Fluid Mech. 254, 59–78.
[8] Girimaji, S.S. 1996 Fully-explicit and self-consistent algebraic Reynolds
stress model Theor. and Comp. Fluid Dyn. 8 387–402 (also ICASE Report
No. 95-82).
[9] Girimaji, S.S. 1996 Improved algebraic Reynolds stress model for engineer-
ing flows In Engineering Turbulence Modelling and Experiments 3 pp 121–129.
Eds W. Rodi and G. Bergeles. Elsevier. Science B.V.
[10] Girimaji, S.S. 1997 A Galilean invariant explicit algebraic Reynolds stress
model for turbulent curved flows Phys. Fluids. 9 1067–1077 (also ICASE Report
No. 96-38).
92 References
[11] Hanjalić, K., Jakirlić, S. & Hadžić, I. 1995 Computation of oscillating

turbulent flows at transitional Re-numbers Turbulent Shear Flows 9 pp 323–
342. Eds F. Durst, N. Kasagi & B.E. Launder. Springer-Verlag, Berlin.
[12] Hellström, T., Davidson, L. & Rizzi, A. 1994 Reynolds stress trans-
port modelling of transonic flow around the RAE2822 airfoil. AIAA 94-0309
32nd Aerospace Sciences Meeting, Reno, Jan 1994.
[13] Horsten, J.J., den Boer, R.G. & Zwaan, R.J. 1983 Unsteady tran-
sonic pressure measurements on a semi-span wind tunnel model of a transport-
type supercritical wing (LANN model). AFWAL-TR-83-3039.
[14] Imao, S., Itoh, M. & Harada, T. 1996 Turbulent characteristics of the
flow in an axially rotating pipe J. Heat Fluid Flow 17, 444–451.
[15] Johansson, A.V. & Wallin, S. 1996 A new explicit algebraic Reynolds
stress model. Proc. Sixth European Turbulence Conference, Lausanne, July
1996, Ed. P. Monkewitz, 31–34.
[16] Kim, J. 1989 On the structure of pressure fluctuations in simulated turbu-
lent channel flow. J. Fluid Mech. 205, 421–451.
[17] Launder, B.E., Reece, G.J. & Rodi, W. 1975 Progress in the develop-
ment of a Reynolds-stress turbulence closure. J. Fluid Mech. 41, 537–566.
[18] Lumley, J.L. 1978 Computational modeling of turbulent flows. Adv. Appl.
Mech. 18, 123–176.
[19] Matsson, O.J.E. & Alfredsson, P.H. 1990 Curvature- and rotation-
induced instabilities in channel flow. J. Fluid Mech. 210, 537–563.
[20] Menter, F.R. 1993 Zonal Two Equation K–ω Turbulence Models for Aero-
dynamic Flows. AIAA 93-2906
[21] Menter, F.R. 1994 Two-equation eddy-viscosity turbulence models for
engineering applications. AIAA J. 32, 1598–1604.
[22] Moser, R.D, Kim J. & Mansour N.N. 1998 DNS of Turbulent Channel
Flow up to Reτ = 590. Phys. Fluids 11, 943–945.
[23] Naot, D., Shavit, A. & Wolfstein, M. 1973 Two-point correlation
model and the redistribution of Reynolds stresses. Phys. Fluids 16, 738–743.
[24] Pope, S.B. 1975 A more general effective-viscosity hypothesis J. Fluid
Mech. 72, 331–340.
[25] Rizzi, A., Eliasson, P., Lindblad, I., Hirsch C., Lacor, C.,
Haeuser, J. 1993 The engineering of multiblock/multigrid software for Navier–
Stokes flows on structured meshes Computers Fluids 22, 341–367.
[26] Rodi, W. 1972 The prediction of free turbulent boundary layers by use of
a two equation model of turbulence. Ph.D. thesis, University of London.
[27] Rodi, W. 1976 A new algebraic relation for calculating the Reynolds
stresses. Z. angew. Math. Mech. 56, T219–221.
[28] Rotta, J.C. 1951 Statistische Theorie nichthomogener Turbulenz. Z. Phys.
129, 547–572.
[29] Schülein, E., Krogmann, P. & Stanewsky, E. 1996 Documentation
References 93
of Two-Dimensional Impinging Shock/ Turbulent Boundary Layer Interaction

Flow. DLR Report: DLR IB 223 - 96 A 49.
[30] Schlichting, H. 1979 Boundary-Layer Theory, 7th ed., McGraw-Hill, New
York.
[31] Shabbir, A. & Shih, T.H. 1992 Critical Assessment of Reynolds Stress
Turbulence Models Using Homogeneous Flows NASA TM 105954, ICOMP-92-
24, CMOTT-92-12.
[32] Shih, T.H., Zhu, J. & Lumley, J.L. 1992 A Realizable Reynolds Stress
Algebraic Equation Model NASA TM 105993, ICOMP-92-27, CMOTT-92-14.
[33] Shih, T.-H. 1996 Constitutive relations and realizability of single-point tur-
bulence closures Chapter 4 of Turbulence and Transition Modelling, Eds.
M. Hallback, D.S. Henningson, A.V. Johansson and P.H. Alfredsson. Kluwer.
[34] Shih, T.-H. & Lumley, J.L. 1993 Remarks on turbulent constitutive
relations Math. Comput. Modelling, 18, 9–16.
[35] Sjögren, T. 1997 Development and Validation of Turbulence Models
Through Experiment and Computation Doctoral thesis, Dept. of Mechanics,
KTH, Stockholm, Sweden.
[36] So, R.M.C., Lai, Y.G. & Zhang, H.S. 1991 Second-Order Near-wall
Turbulence Closures: A Review. AIAA J. 29, 1819–1835.
[37] Spalart P.R. 1988 Direct simulation of a turbulent boundary layer up to
Rθ = 1410. J. Fluid Mech. 187, 61–98.
[38] Spencer, A.J.M. & Rivlin, R.S. 1959 The theory of matrix polynomials
and its application to the mechanics of isotropic continua Arch. Rat. Mech.
Anal. 2, 309–336.
[39] Speziale, C.G., Sarkar, S. & Gatski, T.B. 1991 Modelling the
pressure–strain correlation of turbulence: an invariant dynamical systems ap-
proach. J. Fluid Mech. 227, 245–272.
[40] Speziale, C.G. & Xu, X.-H. 1996 Towards the development of second-
order closure models for nonequilibrium turbulent flows Int. J. Heat and Fluid
Flow 17, 238–244.
[41] Taulbee, D.B. 1992 An improved algebraic Reynolds stress model and
corresponding nonlinear stress model. Phys. Fluids A 4, 2555–2561.
[42] Taulbee, D.B., Sonnenmeier, J.R. & Wall, K.M. 1994 Stress relation
for three-dimensional turbulent flows Phys. Fluids 6, 1399–1401.
[43] Tavoularis, S. & Corrsin, S. 1981 Experiments in nearly homogeneous
turbulent shear flow with a uniform mean temperature gradient. Part I. J. Fluid
Mech. 104, 311–347.
[44] Vandromme, D. 1992 Introduction to the Modeling of Turbulence von
Karman Institute for Fluid Dynamics, Lecture Series 1991-02.
[45] Wallin, S. & Johansson, A.V. 1996 A new explicit algebraic Reynolds
stress turbulence model including an improved near-wall treatment. Proc.
Flow Modeling and Turbulence Measurements VI, Tallahassee FL, Chen, Shih,
94 References
Lienau & Kung (eds), 399–406.

[46] Wilcox, D.C. 1988 Reassessment of the scale-determining equation for
advanced turbulence models. AIAA J. 26, 1299–1310.
[47] Wilcox, D.C. 1993 Turbulence modeling for CFD. DCW Industries Inc.,
La Cañada, California.
[48] Wilcox, D.C. 1994 Simulation of transition with a two-equation turbu-
lence model. AIAA J. 32, 247–255.
Paper 2
2
A PREDICTION METHOD FOR HIGH SPEED TURBULENT SEPARA-
TED FLOWS WITH EXPERIMENTAL VERIFICATION
I. A. A. Lindblad (Head Comp. aerodynamics dept.), S. Wallin (Research Scientist)
The Aeronautical Research Institute of Sweden (FFA), Bromma, Sweden
A. V. Johansson, (Professor)
The Royal Institute of Technology (KTH), Stockholm, Sweden
R. Friedrich (Professor), R. Lechner (PhD student)
Technische Universität München (TUM), Garching, Germany
P. Krogmann (Senior Research Scientist), E. Schülein (Research Scientist)
Deutsches Zentrum für Luft- und Raumfahrt (DLR), Göttingen, Germany
J.-C. Courty (Senior Research Scientist), M. Ravachol (Research Scientist)
Dassault Aviation (DAA), St Cloud, Paris, France
D. Giordano (ESA Technical Project Officer)
ESA/ESTEC, Noordwijk, The Netherlands
Abstract. It is well known that the level of current engineering computational

aerodynamic analysis for turbulent hypersonic separated flows is not satisfactory.
Computations using standard two-equation turbulence models do not give suffi-
ciently accurate predictions. Existing compressibility corrections do not provide
general improvements. In fact, they seem to render predictions worse, at least for
Mach numbers up to five. Possible improvements using Explicit Algebraic Turbu-
lence Models (EARSM) have been investigated. This class of models allow for a
more general coupling between the mean field and the turbulent Reynolds
stresses, compared to the standard two-equation models based on the Boussinesq
hypothesis. A new class of EARSMs is proposed and compared to results of exist-
ing standard models. It is demonstrated that the new proposed EARSM gives
improved prediction over standard two-equation models and previously published
EARSMs for validation cases relevant for high speed separated flows. A new
complete EARSM turbulence model based on the k-ω two-equation model is rec-
ommended as giving the overall best prediction of the validation test cases. As
part of the project a new series of detailed validation experiments, with turbulent
boundary layer-shock wave interaction on a flat plate at Mach 5, with different
degree of flow separation, have been performed.
1. Introduction
1.1. Problems in hypersonic aerodynamics. Future space transportation sys-

tems, in particular reusable vehicles, will have to be more aerodynamically optimised
than current designs. Currently novel space transportation concepts are under consider-
ation, in Europe and elsewhere. The detailed evaluation of the technical feasibility and
economy of the concepts as well as technical development of a selected design must be
97
98 PREDICTION METHOD, EXPR. VERIFICATION, HIGH SPEED TURBULENT SEPARATED FLOWS
done to a large extent by CFD methods, in part because full scale flight conditions can-
not in general be reproduced in ground based experimental facilities.
External as well as internal flows are of importance. Heat loads of complex attached
and separated flows must be determined with sufficient accuracy. Current industrial
standard CFD methods in some cases fall short of supplying a reliable solution to the
given analysis problem. In many instances the shortcomings are related to the turbu-
lence modeling in Reynolds Averaged Navier-Stokes methods. In the current project we
have mainly dealt with the problem of predicting turbulent flow separation in high-
speed flows. Typical applications are air intakes of hypersonic airbreathing vehicles or
flow over control surfaces, with hinge line flow separation.
1.2. Project description. Turbulence modeling in high-speed flows is a vast subject.

Our approach has been that a more general modeling, including the necessary flow
physics, will be able to treat a wider range of flow problems, also in areas of ’extrapola-
tion’ beyond the regime of calibration of the model. With this general idea turbulence
modeling has been reviewed (Lindblad et al., 1995, chs 3-4). As a result of this review a
class of turbulence models, Explicit Algebraic Reynolds Stress Models (EARSMs) was
selected for further study and development.
The EARSM is formulated together with a base-line two equation turbulence model.
In the project three partners (FFA, TUM, DAA) used different base-line models, which
were all modified to be EARSM. To compare results and to judge the most promising
route of development the codes were first tested on common schematic test cases (Wal-
lin et al., 1996). In order to ensure that the codes gave the same results with the same
models and assumptions, test cases assuming laminar flow or using a simple algebraic
turbulence model were computed. The results using the different base-line models were
subsequently compared. All base-line models were modified, to include published
EARSMs.
Based on the experiences of implementation and validation of existing EARSMs a
new anisotropy formulation was developed (Johansson and Wallin, 1996; Wallin &
Johansson, 1997). This model was implemented in the codes on the basis of the differ-
ent two-equation turbulence models and tested. One variant was selected as giving the
overall best results.
A set of test cases was selected for the project, after an extensive review of suitable
flow cases for the validation of turbulence models for the selected classes of target
problems (Lindblad et al. 1995; Wallin et al. 1996). These were mainly schematic two-
dimensional cases related to the selected target problems, used for model development
and validation. A three-dimensional test case (Rodi & Dolling, 1992) was selected for
the demonstration of the generality of the model development and implementation.
One project goal was to contribute to the database of validation data for turbulence
models at high speed. In the literature survey a lack of schematic test cases for hyper-
sonic shock wave-turbulent boundary layer interaction was identified. A set of test
INGEMAR A. A. LINDBLAD ET AL. 99
experiments, with a shock wave incident on flat plate boundary layer at Mach 5 was
conducted by DLR (Krogmann & Schülein, 1996). This was used as the main valida-
tion case in the project. The results are summarised in section 4.
Selected project computational results, demonstrating the main findings, are ana-
lysed in section 5. The main conclusions of the turbulence model development and val-
idation, as well as the lessons learned from the experimental campaign are summarised
in section 6.
2. Problems in turbulence modeling for high speed flows
2.1. The required level of modeling. In the review of turbulence modeling for
high speed flows (Lindblad et al., 1995, ch. 4) the specific problems of turbulence mod-
els for compressible flows were investigated. It was concluded that intrinsic compressi-
bility effects on the turbulence had a significant influence on free shear flows. In wall
bounded flows the main influence of compressibility (for Mach numbers less than five)
lies in the interaction between the turbulence and the compressible mean field. The
importance of the intrinsic compressibility effects on turbulence is measured by a tur-
bulent Mach number. Recent evidence (Sarkar, 1995, Blaisdell et al. 1991) points to the
gradient Mach number M g = S ( λ ⁄ c ) as the representative compressible parameter.
Here S is the mean shear rate, λ is a representative integral length scale and c is the
mean speed of sound.
In cases where the turbulent Mach number (or gradient Mach number) is large,
explicit compressibility terms should be added to the turbulence model equations.
Attempts have been made at modeling pressure-dilatation correlation and dilatational
dissipation rate for compressible flows (e.g. Sarkar et al., 1991; Zeman, 1990). How-
ever, these do in general not agree well with experimental or DNS results, in particular
not for wall-bounded flows.
As noted above the main target applications in the current project feature flow sepa-
ration of wall bounded flows. We have concluded that the most fruitful direction of
development is to improve on incompressible turbulence models, formulated for com-
pressible flow by applying Favre averaging to the equations. A survey of incompressi-
ble turbulence models was performed, aimed at the question of prediction of separated
flows. It was concluded that in order to seek a general improvement in prediction of a
wide class of separated flows in complex configurations, it is necessary to work with
Reynolds Stress turbulence models. In general prediction of complex, separated flows
requires a realistic model for the anisotropy of the Reynolds stress tensor, appearing in
the averaged Navier-Stokes equations. In principle it would be advantageous to derive a
transport equation for the independent stress components, leading to a differential Rey-
nolds Stress Model (DRSM). The practical application of DRSM is still in early stages.
The implementation and calibration of DRSM models are quite complex for the types
of target problems envisioned for the current project. The computational cost and prob-
lems with numerical robustness precludes the use of DRSM models for engineering
purposes in the near future.
In two-equation models the transport equations for individual Reynolds stresses are
replaced by a simple algebraic proportionality between the Reynolds Stress anisotropy,
a ij and the mean flow strain rate S ij (Boussinesq’s hypothesis). This is a quite simplis-
tic model for the Reynolds stress components and is formally valid only for the shear
stress component, uv, in near parallel flows in equilibrium. Despite this, the standard
two-equation turbulence models work quite well in many engineering flow problems,
attesting to the fact that these are often close to equilibrium. A possible route for
improving on the known deficiencies of two-equation models in prediction e.g. flow
separation, is to generalise the relation between the Reynolds stress anisotropy and the
mean flow strain and rotation rate
f ( a ij, S ij, Ω ij ) = 0 (1)
This is the idea of the Algebraic Reynolds Stress Models (ARSM), pioneered by
Rodi (1972, 1976). The first of these models were formulated as implicit relations to be
solved as part of the iteration process. The use of ARSM has been slowed by the
numerical difficulties, as serious as for DRSMs, which are due to the implicit and non-
linear algebraic relations to be solved in the ARSM case. This has motivated the devel-
opment of Explicit Algebraic Reynolds Stress Models (EARSM) in which the Rey-
nolds Stress anisotropy is expressed as an explicit function of the mean flow strain and
rotation rate tensors,
a ij = a ij ( S ij, Ω ij ) . (2)
The EARSM formulation allows a direct evaluation of the Reynolds stress terms
once the mean field is known. This is quite amenable to the numerical procedures of
flow solvers and makes EARSM models as robust as two-equation turbulence models.
2.2. Development of an EARSM anisotropy model. In the current project

EARSMs have been reassessed and a novel development has been made. Published
models (Shih, Zhu & Lumley, 1992; Gatski & Speziale, 1993) were implemented and
tested for reference. The EARSM concept has been reassessed and a new model devel-
oped (Johansson & Wallin, 1996; Wallin & Johansson, 1996). The following main
developments have been made.
• The development of the complete explicit expression a ij from (2) in terms of a
series of nonlinear combinations of S ij and Ω ij . The coefficients of this series have
been determined to make the EARSM anisotropy model an exact solution of the
ARSM equations for two-dimensional flows and an approximate solution for three-
dimensional flows.
• A calibration of the near wall behaviour of the new EARSM. A low Reynolds
number formulation has been created in which it has been attempted to allow the
Reynolds stress anisotropy to behave according to evidence from experimental and
DNS results.
The classical algebraic Reynolds stress model (ARSM) developed by Rodi (1976)
based on the general linear pressure strain model by Launder et.al. (1975) was the start-
ing point for the new anisotropy model, which is described in detail by Wallin &
Johansson (1997). The ARSM is a non-linear relation between the Reynolds stress ani-
sotropy tensor, a ij = ρu i u j ⁄ ρk – 2δ ij ⁄ 3 , and the mean velocity strain and rotation
rate tensors, S ij and Ω ij . By making reasonable choices of constants of the rapid pres-
sure strain model the resulting ARSM is simplified and reads
 C – 1 – 2--- D – a S  a = – -----
8 4
- S + --- ( a Ω – Ω ik a kj ) . (3)
 1 3 kl lk ij 15 ij 9 ik kj
where C1 is the Rotta constant. Several attempts have been made at solving this equa-
tion to obtain an explicit relation between the anisotropy tensor and the mean flow field
which is an explicit algebraic Reynolds stress model (EARSM), see e.g. Pope (1975),
Taulbee (1992) and Gatski & Speziale (1993). The non-linearity in a ij forms a major
obstacle in the solution process but recently it has independently been shown by
Johansson & Wallin (1996) and Girimaji (1995) that this equation can be solved exactly
for two-dimensional mean flows and also that good approximations exists also for
three-dimensional mean flows (Wallin & Johansson 1996 and 1997). The correct treat-
ment of the non-linear term is of crucial importance for the asymptotic model behav-
iour for large strain rates and represents a substantial improvement with this type of
modeling. The resulting EARSM has been found to be well suited for integration to the
wall. The new model development includes a new near wall treatment ensuring realiza-
bility for the individual stress components which can be well predicted by introducing
one simple wall damping function of ‘van Driest’ type.
A two-equation turbulence model is used as the basis on which to build the EARSM.
The basis for the two-equation models is the derivation of transport equations which
give an estimate of the turbulent velocity and length scales. The selection of the
dependent variables is somewhat arbitrary. The equation for the turbulent kinetic
energy, k , gives an estimate for the velocity scale. The formulation of the length scale
equation is less obvious and formulations based on the dissipation, ε , or on another
quantity such as ω ∼ ε ⁄ k or τ ∼ k ⁄ ε have been developed. In principle these formula-
tions are equivalent. Different detailed modeling, calibration of constants and near wall
modifications make the respective formulations show different behaviour, in particular
in regions of separated flow. This nonuniqueness in the formulation carries over to the
EARSMs. In the project we have taken the modest route of implementing and testing
the new EARSM based on the different two-equation formulations of the project part-
ners. In this way it was possible to quantify the effect of the EARSM modification of
different formulations of the length scale equation. Also the different grid and numeri-
cal methods strategies used have been utilized to increase the experience in using the
new models. Based on the experience of the implementation of the new EARSM based
on the different base-line two-equation models a recommended choice of complete
model was made.
3. Models used for validation
3.1. Codes and base-line turbulence models. We here present a summary of the
codes and base-line turbulence models used as a basis for the development of an
improved prediction methodology. The codes and turbulence models used in the project
are described in detail in Wallin et. al (1996).
FFA has used the EURANUS code (Rizzi et al., 1993), which is a multi-block struc-
tured three-dimensional Navier-Stokes solver. The governing equations of compressi-
ble flow are in Favre averaged form. The base-line two-equation turbulence model used
is the k – ε model of Chien (1982). The dependent variables are stored in cell centers
and the code employs, in this project, second order upwind symmetric TVD scheme
with Van Leer limiters for the convective terms and purely central approximation for
the viscous fluxes. The equations are converged with a Runge-Kutta time stepping pro-
cedure with multigrid acceleration.
TUM has used the structured NSFLEX code (Eberle, 1985 and Schmatz, 1989), in a
two-dimensional version. The governing equations are Favre averaged. As a base-line
model TUM has implemented the k – ω model of Wilcox (1991) with low-Reynolds
number corrections proposed by Wilcox (1993), (1994). The dependent variables are
stored in cell centers. The equations are transformed to nonorthogonal, curvilinear
coordinates. The convective fluxes are approximated by a hybrid local characteristic
(Eberle, 1985) and a Steger-Warming scheme (Schmatz et al., 1991). Viscous fluxes are
approximated by central differences. The time-stepping is first order implicit employ-
ing a point Gauss-Seidel relaxation technique.
DAA uses the unstructured FEM Navier-Stokes solver VIRGINI. The governing
equations are expressed in terms of entropy variables. The base-line two-equation tur-
bulence model is the k – ε model (Launder & Spalding, 1974). The near wall region is
treated by the two-layer model of Patel and Chen (1991). The equations are solved
using the Galerkin least-squares method. The equations are converged to steady state
using an implicit iterative method using the GMRES algorithm.
3.2. Validation procedure. The development and validation procedure was made in
a step-by-step fashion, described in detail in Wallin et al. (1996). The different Navier-
Stokes codes were first compared on a laminar flow case (flat plate at Mach 2) to ensure
that the numerical errors of the methods and the grids used were under control. As a
second step results using the same turbulence model (Baldwin-Lomax) were compared
to ensure that a simple turbulence model did not introduce errors. In a third step results
using the different base-line two-equation models were compared. The differences
between the results using different base-line models were judged to be in line with
known differences in predictions using the models. These differences carry over to the
main computational phase of the project (section 5).
4. The validation experiment
4.1. Wind Tunnel Facility. For the validation of numerical computations the Ludw-
ieg Tube Tunnel Facility RWG of DLR Göttingen with Mach number five was chosen
because of its high Reynolds number capability (maximum unit Reynolds number of
50x106 m-1) and documented flow quality. The RMS fluctuations of mass flow and pitot
pressure are of the order of 1%.
4.2. Experimental model and test setup. As test case the zero pressure gradient
flat plate turbulent boundary layer interfering with an impinging two-dimensional
oblique shock wave of different shock strength at Mach five and a freestream unit Rey-
nolds number of approximately 40x106 m-1 was defined.
The test set-up, shown in Figure 1, consisted of a basic flat plate 400 mm wide and
500 mm in length with sharp leading edge and rows of static pressure tabs in axial
direction close to the center-line and in lateral direction immediately upstream of the
anticipated shock impingement location. As shock generator served a second flat plate
at incidence mounted above the basic plate such that for all incidences tested the invis-
cid shock/boundary layer interaction occurred at the same position at x = 350 mm
downstream of the leading edge (Rex ≈ 14x106 m-1). This Reynolds number was con-
sidered to be high enough to guarantee a fully developed turbulent boundary layer at the
shock impingement position. The plate incidences tested were ß = 0˚, 6˚, 8˚, 10˚, 12˚
and 14˚, giving an extensive database of shock-boundary layer interaction at varying
shock strength. For the project the 0˚, 6˚, 10˚ and 14˚ cases were mainly used.
Initially, some preliminary qualitative oil flow visualizations were conducted on the
basic plate at the shock interaction location in order to demonstrate the two-dimension-
ality of the flow in this region, the shock generator being at ß = 10˚. A straight separa-
tion line caused by the impinging shock could be observed over a considerable distance
across the flat plate center line, verifying the anticipated two-dimensional turbulent
flow field.
Furthermore, the basic plate was equipped in the axial direction with ten and later-
ally with three additional ports for mounting a miniature flattened pitot pressure probe
for boundary layer profile measurements. The three lateral ports were applied in the
course of the experiments in order to demonstrate that the mean flow in the interaction
region was really two-dimensional.
FIGURE 1: A schematic of the test model with measurement stations on the flat
plate.
4.3. Experimental results. The first boundary layer profile measurements

upstream and downstream of the anticipated shock impingement location as well as
static pressure distribution measurements were performed on the basic plate without the
shock generator installed, ß = 0˚. The velocity profiles show self-similarity and fit well
to the wall-wake law. From the measured velocity profiles the relevant boundary layer
properties, thickness d, displacement thickness d*, momentum thickness q and skin
friction coefficient Cf are determined. Figure 2 shows the boundary layer parameters
together with skin friction and static pressure distribution. The accuracy of determining
the thickness parameters of the boundary layer may be demonstrated by the four points
taken laterally at x = 316 mm.
At the end of all measurements a newly developed oil film interferometry measure-
ment technique for determining the local skin friction was applied for the first time in
RWG. The results (marked as Optic/LISF) are compared in Figure 2 with those values
derived from the velocity profile measurements. It is seen that at x ≈100 mm the skin
friction increase due to transition is completed and that downstream the boundary layer
may be considered as turbulent.
FIGURE 2: Boundary layer on the flat plate.
The static pressure distribution is fairly constant at the level of freestream pressure
within an uncertainty of less than 1 % up to a distance of x = 425 mm. Only on the rear
of the plate the pressure drops somewhat below the freestream pressure which may be
attributed to an upstream influence of the expansion at the trailing edge over the sub-
sonic part of the relatively thick boundary layer.
With the impinging shock generated by the shock generating plate set to an inci-
dence of ß = 6˚ no separation is observed. The static pressure starts slightly upstream of
the inviscid shock impingement position to rise up to the inviscid level.
The case of an impinging shock generated by the shock generator at ß = 10˚ inci-
dence was defined as the main test case for the numerical computations. The results of
boundary layer and pressure distribution measurements are shown in Figure 3. The
impinging shock causes a moderate separation zone which is marked in the skin friction
coefficient distribution by the two solid symbols "S" and "R" for separation and reat-
tachment position. The static pressure increase up to the inviscid pressure level occurs
over an enlarged distance in this case. The pressure drop at the rear of the plate is
caused by the expansion fan built up at the trailing edge of the shock generator. In Fig-
ure 4 a sketch of the flow field obtained from the measurements is compared with the
corresponding shadowgraph. The boundary layer development, the incoming shock and
the separation and reattachment shocks were clearly discerned in the measurements.
FIGURE 3: Impinging shock/boundary layer interaction at β = 10° .
Finally, the strength of the impinging shock was increased by setting the shock gen-
erator to an incidence of ß = 14˚. The impinging shock, thus generated, causes a more
severe separation. A small plateau appears over the separated region before it rises to
the theoretical inviscid level. After reattachment the skin friction coefficient is drasti-
cally increased. The flow in the separated region is severely disturbed by the impinging
shock.
5. Computational results
The computations of the DLR experiment described in the previous section give a clear
picture of the differences in prediction using different classes of turbulence modeling
FIGURE 4: Impinging shock/boundary layer interaction at β = 10° .
for high speed separated flows and will be used for illustration of the project results.
The conclusions drawn from the other test cases were in line with these results.
5.1. The DLR experiment of shock wave-boundary layer interaction. The

three shock generator plate angles β = 6° , 10° and 14° were computed using differ-
ent turbulence models. The emphasis was put on the 10° case, where moderate separa-
tion occurs. The computational grid used in this study consists of 80x80 cells,
providing in most flow regions well resolved results. The grid points are clustered
around the incoming shock, in the boundary layer and around the shock wave-boundary
layer interaction, with the use of different grids for the varying strength of the incoming
shock.
5.1.1. Comparison of results using different levels of turbulence modeling. In order to

illustrate the improvement of flow prediction with more complex turbulence models,
computations were performed with a range of models, from an algebraic mixing length
model to a DRSM. The results in terms of skin friction are shown in Figure 5. The
Baldwin-Lomax (1978) mixing length model (FFA_BL in the figure legend), was
designed for attached boundary layers and gives for this case a much too large separa-
tion length with a separation onset somewhat downstream the experimental one. The
two-equation Wilcox (1994) k – ω model (TUM_k-w) underestimates the separation
length, in a way typical for two-equation eddy viscosity models without further correc-
tions for separated flows. The separation lengths predicted with the Reynolds stress
models represented by the recommended EARSM, built on the k – ω model,
(TUM_EARSM_k-w) and the SSG (Speziale, et.al. 1991) DRSM (TUM_SSG_k-w)
are almost equal. The predicted skin friction peaks (maximum and minimum) are
somewhat higher using DRSM compared to EARSM.
0.010
FFA_BL
TUM_k−w
TUM_EARSM_k−w
TUM_SSG_k−w
Exp_prof
0.008 Exp_VD
Exp_HBC
S+R
0.006
Cf [−]
0.004
0.002
0.000
−0.002
0.30 0.35 0.40 0.45
x [m]
FIGURE 5: Different level of modeling. Impinging shock at Mach 5, β = 10° ,

Experiment by Krogmann & Schülein (1996).
FFA_BL
TUM_k−w
TUM_EARSM_k−w
40 TUM_SSG_k−w
Expr
30
p_wall [kPa]
20
10
0
0.30 0.35 0.40 0.45
x [m]

Figure 6 shows the computed wall pressure for the same case. The underprediction
of the separation length by the eddy-viscosity model gives a too steep pressure rise. The
slope of the pressure rise computed with the Reynolds stress models better matches the
experimental data. We can se a small difference between the algebraic and differential
Reynolds stress models in the beginning of the separation zone, where the DRSM
shows a somewhat steeper pressure rise compared to the EARSM and also that there is
a tendency of a ‘pressure plateau’ in the DRSM results. This can be interpreted as the
DRSM predicting a thicker separation bubble than EARSM. However, no pressure pla-
teau could be seen in the experiment for this case. The computed pressure rise is shifted
somewhat downstream and the experimental pressure level downstream of the interac-
tion is not completely reached due to approximate inflow conditions used for the com-
putations.
The computed velocity profiles are compared to the measured values at different
cross sections in Figure 7. Sections 2 to 4 are located upstream of the interaction while
the sections 7 to 10 are downstream. The position of the measured velocity profiles cor-
responds to the position of the skin friction measurements where sections 4 and 7 to 10
are seen in Figure 5. There are only minor differences among the computed velocity
profiles before the interaction and a consistent difference between the computed and
measured velocity profiles. This difference is also consistent with the underpredicted
skin friction coefficient before the interaction in Figure 5. The reason is that the inflow
profile used in the computations at x = 0.2 m could not be perfectly matched to the
experimental profile. After the interaction, stations 7 to 10, the differences among the
computed velocity profiles are much larger and it is quite clear that the computed pro-
files using the two different Reynolds stress models are closest to the experiments.
There are only minor differences between the algebraic and differential Reynolds stress
models. Overall the recommended EARSM model gives substantially improved predic-
tions of this case compared to standard two-equation models.
5.1.2. Computations using different EARSMs. In order to investigate the difference

between the EARSM formulation developed in the current project and previously pub-
lished versions, the new EARSM was compared to the Shih, Zhu & Lumley (1992)
(SZL) model, both built on the k – ω two-equation model. The SZL model was modi-
fied with respect to its near wall treatment. The computed skin friction is shown in Fig-
ure 8. The modified Shih et.al. model gives results similar to those using the k – ω
model upstream of the separation while it is closer to the new EARSM downstream of
the separation. The modified Shih et.al. model is, however, did not predict the correct
separation length.
In order to select the best two-equation basis for the EARSM, results using different
two-equation models were compared (see Figure 9). Here we can see that all models
give the correct separation length besides the EARSM based on the Patel & Chen k – ε
model (DAA_EARSM_k-e) where the flow reattaches too far downstream. The behav-
iour of the different models downstream of the reattachment point is similar to that of
the same models with the original eddy-viscosity hypothesis. The overpredicted skin
section 2 section 3 section 4 section 7 section 8 section 9 section 10

6.0
FFA_BL
TUM_k−w
TUM_EARSM_k−w
TUM_SSG_k−w
Expr
5.0
4.0
y [mm]
3.0
2.0
1.0
0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.5 1.0
U/Uo
0.010
TUM_k−w
TUM_SZL_k−w
TUM_EARSM_k−w
Exp_prof
Exp_VD
0.008 Exp_HBC
S+R
0.006
Cf [−]
0.004
0.002
0.000
−0.002
0.30 0.35 0.40 0.45
x [m]
FIGURE 8: Different EARSM. Impinging shock at Mach 5, β = 10° , Experi-

ment by Krogmann & Schülein (1996).
friction using the Chien k – ε model basis (FFA_EARSM_k-e) and the underpredicted
skin friction with the Patel & Chen k – ε model are present also in the original two-
equation formulations. The overall best results are obtained using the k – ω model as
the basis. This conclusion was also drawn from computing the other test cases and that
version was chosen as the recommended model.
There are noticable differences between the two different EARSM computations
based on the k – ω model (FFA_EARSM_k-w and TUM_EARSM_k-w) in Figure 10.
0.010
FFA_EARSM_k−w
TUM_EARSM_k−w
FFA_EARSM_k−e
DAA_EARSM_k−e
Exp_prof
0.008 Exp_VD
Exp_HBC
S+R
Cf [−] 0.006
0.004
0.002
0.000
−0.002
0.30 0.35 0.40 0.45
x [m]
FIGURE 9: Newly developed EARSM using different two-equation model

basis. Impinging shock at Mach 5, β = 10° , Experiment by Krogmann &
Schülein (1996).
These are similar to the differences between the two computations with the same model
based on the eddy-viscosity assumption. The differences are probably due to differ-
ences in the FFA and TUM implementations of the wall boundary conditions and also
in the definition of the effective eddy viscosity which is used in the Modeling of the tur-
bulent diffusion of k and ω . This demonstrates that near wall and boundary condition
formulations are quite important and merit further study.
5.1.3. Shock strength. The major effort has been in computing the 10° case where the
results are presented above but also the 6° and 14° cases have been computed. The
stronger shock in the 14° case induces a much larger separation compared to the 10°
case, see Figure 10. The different turbulence models show the same characteristic devi-
ations. It is worth noting that the proposed EARSM based on the k – ω model
(FFA_EARSM_k-w) predicts the separation length and skin friction behaviour after
reattachment correctly. The good result at 10° was not a coincidence.
6. Discussion and recommendations

Our main conclusions from the literature survey, development and computations with
regards to the requirements on turbulence modelling for high speed separated flows
may be summarised
• Two-equation eddy viscosity models do not include the relevant physics involved in
separated flows without ad-hoc modifications. This class of models did not give
accurate predictions for the chosen test cases.
0.010
FFA_EARSM_k−w
DAA_EARSM_k−e
FFA_k−w
FFA_k−e
FFA_BL
0.008 Exp_prof
Exp_VD
Exp_HBC
S+R
Cf [−] 0.006
0.004
0.002
0.000
−0.002
0.30 0.35 0.40 0.45
x [m]
FIGURE 10: Computed skin friction. Impinging shock at Mach 5, β = 14° ,

• Differential Reynolds stress models (DRSMs) are more general in the sense that the
intercomponent transfer in the Reynolds stress tensor can be considered. The
numerical treatment and computational effort in complex flow situations is a major
disadvantage of these models.
• By developing a suitable explicit algebraic Reynolds stress model to replace the
eddy viscosity hypothesis of the standard two-equation models, we were able to
obtain results which showed the correct trends for the chosen test cases, closer to
experimental values. In a limited comparative study the EARSM produced results
closer to DRSM results than two-equation results.
The three different two-equation models with the original eddy viscosity hypothesis
give quite similar results in zero pressure gradient flows but different results in flows
with nonzero pressure gradient. These differences are inherited when the two-equation
models are used together with the EARSM. One of the main short-comings of the
Chien k – ε model is the large overprediction of Cf after reattachment. The k – ω
model has a more regular behaviour in the separation region and the ω equation per-
forms better in pressure gradient flows than the Chien ε equation. For these reasons the
k – ω model was selected as the recommended basis.
The new EARSM anisotropy model has been found to be well suited for integration
to the wall. It includes a near wall treatment ensuring realizability for the individual
stress components which can be well predicted by introducing one simple wall damp-
ing function of ‘van Driest’ type. When the proposed EARSM is used together with an
existing two-equation eddy viscosity model, it was found by Wallin et.al. (1997) that
the additive constant B in the log-law is slightly overpredicted for both the Chien k – ε
and Wilcox low Reynolds number k – ω models for zero pressure gradient flows. The
near wall corrections within the different length-scale determining equations must thus
be modified in order to give the correct near wall behaviour. The near wall behaviour is
improved if the effect of the near wall modification in the original ε or ω equation is
reduced which indicates that the inclusion of the proposed EARSM reduces the need
for ad hoc modifications in the length-scale determining equation.
There are deficiences in the prediction of the presented models which are not clearly
understood. One is the lag in prediction of the pressure rise in cases like the DLR
experiment, also known from two-equation models. One must also keep in mind that
the EARSM with its algebraic anisotropy model is not able to predict strongly nonequi-
librium turbulent situations. In conclusion, however, it seems that the model class gives
improvements in engineering prediction of a large set of important complex flows.
The new recommended model is one of the first in a new class of models, and not a
final step. A thorough testing of the model in different parameter regimes must be made
in order to assess the range of validity for the recommended model. Moreover, a
number of new possible extensions are now within reach with the proposed model class
as the platform for further developments.
The numerical behaviour of the new recommended model has been found to be sim-
ilar to the basic two-equation model used. The numerical efficiency and robustness
could be further improved by analysing the numerical behaviour of the basic two-equa-
tion model together with the proposed EARSM.
One of the results from this study is that the length-scale equation has a major
impact on the prediction capability of the EARSM model. The recommendation of
using the k – ω model as the basis gives limitations known from the model in stand-
ard form. As already mentioned, the choice of the dependent variable has an impact on
the near wall treatment and wall boundary conditions as well as far-field boundary con-
ditions and far-field behaviour. A new lengthscale determining equation should be
developed with the use of the proposed EARSM to avoid the risk of inheriting terms
that are needed to balance the errors introduced by the eddy viscosity hypothesis.
Our understanding of the features and behaviour of truly compressible turbulent
flows is quite limited. General models to describe compressibility effects due to turbu-
lent fluctuations are not yet available. A critical review in this topic was performed in
this project which pointed out limitations and lack of generality of the models currently
available. More basic research in this area is needed to increase the understanding of
this phenomena through numerical simulations of compressible turbulence and novel
experiments, see also Friedrich and Bertolotti (1997). One interesting observation made
in the review is that intrinsic compressibility is affected by the Reynolds stress anisot-
ropy. One necessary condition to be able to catch this behaviour is the possibility to cor-
rectly represent the individual Reynolds stresses. This makes the EARSM class of
models an interesting platform for future compressibility correction models.
The experimental test cases available for the validation of high speed separated flows
were reviewed as part of the project. A substantial set of generic cases relevant for high
speed air intake and control surface flows are available in different databases. A gap in
the available set was identified, which could be filled by the DLR experiment per-
formed in the current project. Closer scrutiny of the existing high speed test cases with
separated flows reveal uncertanties which often limits the usefulness of the data for val-
idation of turbulence models.
• Flow definition
- External, in particular inflow, conditions are often uncertain. Either because they
are not measured carefully enough or disturbed by transition effects with the tur-
bulence in a nonequilibrium state. An obvious need for turbulence model valida-
tion is that the inflow turbulence data has been measured, in particular turbulent
kinetic energy profiles. There is also a difficulty in keeping external conditions,
such as pressure gradients or flow homogeneity under full control.
- In situations with stronger interactions the simple flow geometry assumed in the
generic geometry of e.g. a two-dimensional test case is strongly modified by
three-dimensional flow.
• The well proven measurement techniques available at high speed are limited. The
main tool is pitot pressure measurement, calibrated to generate velocity profiles in a
way which becomes questionable near separation regions. Alternative techniques
such as the optical measurement technique, used to determine skin friction directly
in the DLR experiment, may give useful additional data.
The data from the DLR experiment was obtained with great care. There were some
differenences between experiments and boundary layer computations of the inflow con-
ditions, which were not fully resolved during the project. The deviations from the
expected conditions were however limited and systematic and we may conclude from
the work that this is a very valuable contribution to the database of high speed valida-
tion test cases.
As seen in the current project also numerical experiments, using Direct Numerical
Simulation have a useful role in the development of advanced turbulence models. In
particular the near wall behaviour and the anisotropy of the Reynolds stress tensor are
difficult to obtain in any other way.
Future experimental work should be developed with the requirements for validation
and calibration of Reynolds stress turbulence models in mind. There is a need for fur-
ther refined experiments, where the emphasis is put on obtaining a well defined flow sit-
uation also with regard to turbulent data, in creating experimental series with variation
in interaction strength parameters and in measuring turbulent data.
The DLR experiment should be complemented with quantitative measurements in
the separated zone and if possible also by obtaining turbulent data. New experimental
techniques, such as the optical interferometry measurement of skin friction should be
applied through the interaction region.
We have concluded that the main problems in the predicition of high speed wall
bounded turbulent separated flows more lie in the formulation of the basic turbulence
model, than in specific high speed effects. Also, the more advanced Reynolds stress tur-
bulence models of the class developed here require more turbulent data, under different
conditions (pressure gradients etc) for their calibration and validation. It is thus reason-
able to use experimental test cases at lower speed where there are more experimental
techniques available, in particular hot wire and LDV techniques, giving turbulence data.
Still there is a severe problem both having high enough Reynolds numbers to obtain
equilibrium turbulent inflow conditions and at the same time measuring with high reso-
lution in the boundary layers and separated flow regions.
We also recommend that a strong effort is put on generating high quality DNS data
for the calibration of Reynolds stress turbulence models. This will probably be the most
reasonable way of getting enough detail about all the Reynolds stress components.
DNS computations should be done in simple geometries and e.g. with variation in
streamwise pressure gradients to allow the calibration of models suitable for separated
flows.
Acknowledgements
The current project was funded by ESA/ESTEC within the TRP programme under con-
tract no. 11018/94/NL/FG/(SC). The support of ESA is gratefully acknowledged.
References
[1] Blaisdell, G.A., Mansour, N.N., Reynolds W.C. 1991: Numerical simulations of
compressible homogeneous turbulence. Dept. of Mechanical Engineering, Stanford
University, Report No. TF-50.
[2] Friedrich, R. and Bertolotti, F. 1997 ”Compressibility effects due to turbulent fluctu-
ations”, To appear in Applied Scientific Research.
[3] Gatski, T. B., Speziale, C. G., “On explicit algebraic stress models for complex tur-
bulent flows”, J. Fluid Mech., vol. 254, 1993.
[4] Girimaji, S.S., 1995, “Fully-Explicit and Self-Consistent Algebraic Reynolds Stress
Model”, ICASE Report No. 95-82.
[5] Johansson, A. V., Wallin, S., 1996, “A new explicit algebraic Reynolds stress
model”, Proc. Sixth European Turbulence Conference, Lausanne, P. Monkewitz (ed).
[6] Krogmann, P. & Schülein E., 1996, ”Documentation of Two-dimensional Impinging
Shock/Turbulent boundary Layer Interaction Flow”, DLR IB 223- 96 A 49.
[7] Launder, B. E., & Spalding, D. B., 1974 ”The Numerical Computation of turbulent
Flows”, Computer Methods in Applied Mechanics and Engineering, 3, 269-289.
[8] Launder, B. E., Reece, G. J., Rodi, W., 1975, “Progress in the development of a
Reynolds-stress turbulence closure”, J. Fluid Mech., vol. 41.
[9] Lindblad, I. A. A., Wallin S., Courty J.-C., Stanewsky, E., Friedrich, R. & Johans-
son, A. V., 1995, ’Critical Review of Prediction Methods for Compressible Turbulent
Attached and Separated Flows’, FFA TN 1995-43.
[10] Patel, V. C., Chen, H. C., 1991, “Near wall turbulence models for complex flows
including separation”, AIAA J., Vol. 29, No. 6.
[11] Pope, S.B., 1975, “A more general effective-viscosity hypothesis”, J. Fluid Mech.,
vol. 72, part 2.
[12] Rizzi, A., Eliasson, P., Lindblad I., Hirsch, C., Lacor C. & Häuser, J., 1993, “The
engineering of multiblock/multigrid software for Navier-Stokes flows on structured
meshes”, Comp. Fluids., Vol. 22, pp 341-367.
[13] Rodi, P.E., Dolling, D.S., 1992, “An Experimental/ Computational Study of Sharp
Fin Induced Shock Wave/Turbulent Boundary Layer Interactions at Mach 5: Experi-
mental Results”, AIAA-92-0749.
[14] Rodi, W., 1972, ”The prediction of free turbulent boundary layers by use of a two
equation model of turbulence”, Ph.D. thesis, University of London.
[15] Rodi, W., 1976, “A new algebraic relation for calculating the Reynolds stresses”,
Z. angew. Math. Mech. 56, T219-221.
[16] Sarkar, S., Erlebacher, G., Hussaini, M.Y., Kreiss H.O. 1991: The analysis and
modelling of dilatational terms in compressible turbulence. J. Fluid Mech., 227,
pp.473-493.
[17] Sarkar, S. 1995: The stabilizing effect of compressibility in turbulent shear flow. J.
Fluid Mech., 282, pp. 163-186.
[18] Schmatz, M. A., 1989, ”NSFLEX - An Implicit Relaxation Method for the Navier-
Stokes Equations for a Wide Range of Mach Numbers”, MBB Internal report MBB/
FE122/S/PUB/366.
[19] Schmatz, M.A., Höld, R.K., Monnoyer, FF., Mundt, Ch., Wanni, K.M. 1991,
”Numerical Methods for Aerodynamic Design II”. Space Course Aachen 1991.
[20] Shih, T. H., Zhu, J., Lumley, J. L., “A Realizable Reynolds Stress Algebraic Equa-
tion Model”, NASA TM 105993, ICOMP-92-27, CMOTT-92-14, 1992.
[21] Taulbee, D. B., “An improved algebraic Reynolds stress model and corresponding
nonlinear stress model”, Phys. Fluids A 4 (11), Nov., 1992.
[22] Wallin, S., Krogmann, P., Stanewsky, E. Shülein, E., Lechner, R., Ravachol, M. &
Courty, J.-C., 1996, “Preparatory Activities for Theoretical and Experimental Tasks in
Turbulence Modeling of Hypersonic Air Intake Flows”, FFA TN 1996-46.
[23] Wallin, S. & Johansson, A. V., 1996, A New Explicit Algebraic Reynolds Stress
Turbulence Model Including an Improved Near-wall Treatment”, Proc. 6th Int. Symp.
Flow Mod. and Turb. Meas., Tall. Fl., Sept 8-10, 1996, Ed. Chen, C.-J., Shih, C.,
Lienau, J. & Kung, R.
[24] Wallin, S.,Lechner, R. Ravachol, M., 1997, “Validation of Improved Turbulence
Model”, FFA TN 1997-14.
[25] Wallin, S. & Krogmann, P., 1997, “Validation of Improved Methodology”, FFA
TN 1997-15.
[26] Wallin, S, Johansson, A. V., 1997, “A complete explicit algebraic Reynolds stress
turbulence model for incompressible and compressible flows”,FFA TN 1997-51.
[27] Wilcox, D. C., 1991, ”Progress in Hypersonic Modeling”, AIAA paper 91-1785.
[28] Wilcox, D. C., 1993, ”Turbulence Modeling for CFD, DCW Industries Inc., La
Canada, USA.
[29] Wilcox, D. C., 1994, ”Simulation of Transition with a Two-equation Turbulence
Model, AIAA J. 32 no. 2, 247-255.
[30] Zeman, O., 1990, ”Dilatational Dissipation: The Concept and Application in Mod-
eling Compressible Mixing Layers”, Phys. of Fluids A, 2, 178-188.
Paper 3
3
EVOLUTION OF AN ISOLATED TURBULENT TRAILING
VORTEX
Stefan Wallin
Sharath S. Girimaji
Institute for Computer Applications in Science and Engineering (ICASE)
NASA Langley Research Center, Hampton, VA 23681
Abstract. The long time development of an isolated wing tip turbulent

vortex has been studied by Reynolds averaged Navier-Stokes computations.
The vortex is assumed to extend to infinity and to be axisymmetric and
homogeneous in the axial direction, and the axial velocity is assumed to
be negligible. The validity of different turbulence models, ranging from
Reynolds stress transport models to eddy-viscosity K-ε models, has been
assessed and qualitative comparisons with field measurements are made.
Reynolds stress transport models correctly predict strong suppression of the
turbulence in the rotation-dominated vortex core and a reasonable decay
rate of the vortex. Outside of the core, the different Reynolds stress models
differ significantly. The standard eddy-viscosity K-ε model is insensitive
to rotation and, thus, overpredicts the vortex decay rate. Computations
using explicit algebraic Reynolds stress models that are strictly based on
Reynolds stress transport models show that the algebraic form compares
reasonably well with the full transport form. It is, however, important that
the curved flow algebraic assumption is invoked when deriving the explicit
algebraic model.
1. Introduction
The long time development of wing tip vortices is of major interest in determining
the safe distance between airplanes at takeoff and landing. The vortices persist
for long times because the turbulence in the vortex core is strongly suppressed
by the rotation. The vortex decays from the outer parts leaving the core more
or less unaffected for long times, and thus, the radius of the maximum swirl
velocity does not change much in time.
The decay mechanism for aircraft trailing vortices evolving in atmospheric
turbulence is complex and involves several different processes. It is believed
(Proctor [1], Sarpkaya and Daly [2], Sarpkaya [3], and others) that the decay
mechanism is influenced by ambient atmospheric turbulence and stratification,
three-dimensional vortex pair dynamic instabilities, and ground effects, as well as
self-generated turbulence in the vortex. Each of these effects may be important
at different conditions and should not be excluded in a complete analysis.
119
120 EVOLUTION OF AN ISOLATED TURBULENT TRAILING VORTEX
Coherent Doppler laser radar (LIDAR) measurements of trailing vortices of

aircraft at takeoff and landing have been analyzed by Proctor [1], and the over-
all conclusion is that there exist strong correlations between the vortex decay
rates and ambient conditions. That suggests that the self-generated turbulence
within the vortex is of secondary importance, especially during daytime when
the atmospheric dissipation rates are high. The vortex dissipation mechanisms
are complex, are not fully understood, and involve many different aspects that
will not be addressed within this study. In a typical wake vortex, the Reynolds
number is of the order of 107 , which means that molecular diffusion is of lesser
importance compared to turbulent diffusion.
Measurements of the near-field vortex rollup process [4] [5] show that the
vortex is initially fully turbulent. The source of turbulence is primarily the
highly unstable shear layer from the wing trailing edge that winds up around
the core. The winding of the vortex sheet generates a wakelike deficit in the
axial velocity. That is another strong source of turbulence, shown by Ragab and
Sreedhar [6] using large-eddy simulations. The turbulence in the vortex core is,
however, strongly suppressed. It has been observed, experimentally as well as
numerically, that the Reynolds shear stresses and turbulence production tend to
zero approaching the vortex core. The reason for that is that the flow in the
core is close to solid body rotation, which is strongly stabilizing. Because the
Reynolds number is large, the core expands very little in time.
Further downstream, the winding vortex sheet and the axial velocity deficit
have been diffused by the turbulence, and any strong sources of turbulence do
not exist any more. In this study, we will focus on this later stage of the vortex
development. Because physical or numerical experimental data are rare for far-
field vortices, we must rely on field observations and measurements. From field
measurements the following features of trailing vortices have been established:
1) The core persists for a reasonable long period of time as the decay begins at
the outer edges of the vortex and propagates inward. 2) there is no discernible
overshoot in the circulation profile.
In a previous study by Zeman [7] of an isolated single vortex, it is reported
that two-equation eddy-viscosity models, such as the K-ε model, 1) strongly
overpredict the decay rate of the vortex and 2) predict massive overshoots in
circulation. The results obtained using a Reynolds stress transport (RST) model
were in fact almost laminar. The turbulence shear stress uv was much lower than
the molecular stress all over the vortex. The observed vortex decay is, however,
much faster than laminar decay for the high Reynolds number vortices from a
typical aircraft.
In an accompanying ongoing study by Wallin and Girimaji, single finite and
infinite vortices using more advanced RST models are studied. Their results are
more in line with observed trailing vortex behavior as they capture all of the
qualitative aspects quite accurately. The objective of this study is to determine
STEFAN WALLIN AND SHARATH S. GIRIMAJI 121
if a two-equation explicit algebraic Reynolds stress model (EARSM) is indeed

capable of capturing the observed features of the trailing vortex. This study
is motivated by our ultimate goal of developing a three-dimensional unsteady
Reynolds averaged Navier-Stokes (RANS) capability for detailed investigation
of trailing vortices. Clearly, for such an effort a two-equation EARSM is more
viable than the more expensive RST models.
We will later show the importance of the pressure-strain model in this highly
rotating flow and that models based on the Speziale–Sarkar–Gatski (SSG) [8]
model give results in line with observations whereas models based on the Launder–
Reece–Rodi (LRR) [9] model give lower turbulence levels and, thus, a vortex
decay rate slower than observed. This will explain the laminar solution that
Zeman [7] obtained with the so-called isotropization of production model, which
is close to the LRR model.
We will also show that explicit algebraic Reynolds stress models are capable
of predicting the qualitative behavior if 1) the model is fully self-consistent, 2)
the model is based on SSG, and 3) the weak equilibrium assumption is made in
an appropriate curvilinear coordinate system (see Girimaji [10]).
2. Modeling assumptions
We will consider the wing-tip vortex in the farwake after any three-dimensional
and time-dependent disturbances have vanished. The axial velocity component
and the axial derivatives are assumed to be negligible. The vortex may, thus, be
considered as axisymmetric, homogeneous in the axial direction, and developing
in time [11]. It is, therefore, natural to express the flow in the vortex in a
cylindrical coordinate system where êi = [r̂, θ̂, ẑ].
We define the spatial differentiation operator as
∂
∇ ≡ êi (1)
∂xi
where

∂ ∂ 1 ∂ ∂
≡ , , (2)
∂xi ∂r r ∂θ ∂z
in a cylindrical coordinate system. In the problem considered, ∂/∂z ≡ 0 because
the axial direction is homogeneous. The problem is also homogeneous in the
azimuthal direction, but the θ̂ derivative may not be excluded from the nabla
operator because the θ̂ derivative of the unit vectors r̂ and θ̂ are nonzero. Let
us define the derivative of the unit vectors as follows:
∂êi
= Γijk êk (3)
∂xj
where Γijk is all zero except Γ122 = 1/r and Γ221 = −1/r. By the use of this
definition, one may expand the gradient operating on any tensor of any rank
simply by the use of the chain rule, that is, the gradient of the velocity expands
to

∂ ∂Uj
∇U = êi (êj Uj ) = êi êj + Γkij Uk . (4)
∂xi ∂xi
By the use of the preceding assumptions, the Reynolds averaged velocity may
be written as U = θ̂V (r), and the Reynolds averaged momentum equation in
the θ̂ direction reads

∂V 1 ∂ ∂ V
= 2 r3 ν − r2 uv (5)
∂t r ∂r ∂r r
To close the momentum equation one needs an expression for uv, which is the
r̂θ̂ component of the Reynolds stresses, τ = r̂r̂u2 + θ̂ θ̂v 2 + ẑẑw2 + (r̂θ̂ + θ̂r̂)uv.
2.1. Reynolds stress transport modeling. Strong rotational and streamline

curvature effects are present in the vortex. The stabilizing effects are in fact
so strong that any preexisting turbulence in the core of the vortex is almost
completely suppressed. It is, thus, believed that RST closures are needed in
order to capture the turbulence physics properly. The transport equation for
the Reynolds stress tensor in an inertial frame of reference may be written in
symbolic form as

D τij 2
− Dij = Pij + Πij − ε eij + δij (6)
Dt 3
The molecular and turbulent diffusion Dij the pressure-strain term Πij , and the
dissipation rate anisotropy eij need to be modeled because they contain higher-
order correlations. The production term is, however, explicit and reads
Pij = − (τik Skj + Sik τkj ) + (τik Ωkj − Ωik τkj ) (7)
where the mean flow strain and rotation rate tensors are
1 1 ∂V
V
T
S = (∇U) + ∇U = − r̂θ̂ + θ̂r̂
2 2 ∂r r

1 T 1 ∂V V
Ω = (∇U) − ∇U = + θ̂r̂ − r̂θ̂ (8)

2 2 ∂r r
The standard form of the transport equation for the dissipation rate ε is
adopted. For the vortex considered, the advection of ε, as well as the advection
of any scalar quantity, is zero, and the equation for ε reads
∂ε ε
− D(ε) = (Cε1 P − Cε2 ε) (9)
∂t K
where the production P ≡ Pii /2 and D(ε) is the molecular and turbulent diffusion
of ε. Cε1 = 1.45 and Cε2 = 1.90 is used.
The assumptions made for the vortex are that both the ẑ and θ̂ directions are
homogeneous. Moreover, the r̂ component of the mean velocity vanishes. The
advection of the Reynolds stresses then only contain algebraic (nondifferential)
Table 1. The values of the C coefficients for different linear pressure-

strain models
C10 C11 C2 C3 C4
Linearized SSG (L-SSG) 3.4 1.8 0.36 1.25 0.40
Original LRR 3.0 0 0.8 1.75 1.31
Re-calibrated LRR (LRR) 3.6 0 0.8 2 1.11
terms that originate from the θ̂ derivatives of the unit vectors. The material
derivative in cylindrical coordinates, thus, becomes
D τij ∂τij (r) (r)

= − τik Ωkj − Ωik τki (10)

Dt ∂t
where
V
Ω(r) = θ̂r̂ − r̂θ̂ (11)

r
that is, the advection does not contain any spatial derivatives. With V /r = ω R ,
this term is actually half of the Coriolis term that appears in the Reynolds stress
equation written in a rotating frame of reference, where ω R is the (constant)
rotation rate of the system. The other half of the Coriolis term originates from
a similar transformation of the production term.
The modeling of the pressure-strain term is an important issue, especially
in rotation-dominated flows. Models that are linear in the Reynolds stress
anisotropy have been widely used due to the simplicity and general good be-
havior in a number of benchmark flows of engineering interest. Recently more
attention has been placed on non-linear models, for example, the SSG model [8],
which has been found to be superior to linear models in many aspects. In this
study, we will focus on the performance of EARSMs where only linear pressure-
strain models may be considered.
The pressure-strain and dissipation rate anisotropy are usually lumped to-
gether into the general linear form, for example see Girimaji [12],

Πij 1 1P K
− eij = − 0
C1 + C1 aij + C2 Sij
ε 2 ε ε

C3 K 2 C4 K
+ aik Skj + Sik akj − akl Slk δij − (aik Ωkj − Ωik akj ) (12)
2 ε 3 2 ε
where the Reynolds stress anisotropy aij is defined as
τij 2
aij ≡ − δij (13)
K 3
and K ≡ τii /2 is the turbulent kinetic energy. The classical LRR model [9]
and the SSG model [8] linearized around equilibrium homogeneous shear may be
included in this general form where the C coefficients are given in Table 1 (for
example, see Gatski and Speziale [13] for linearization of the SSG model).
The original calibration of the LRR model have not been optimal in all situ-
ations, and more recent calibrations of the LRR model imply that the Rotta co-
efficient c1 ≈ 1.8 and the coefficient in the rapid pressure strain model c2 ≈ 5/9,
for example, see Taulbee [14], Lumley [15], Shabbir and Shih [16], and Wallin
and Johansson [17] [18]. All LRR computations in this study were made using
the recalibrated set of coefficients. The SSG model, also in the linearized form,
is known to be superior to the LRR model in rotation-dominated flows, for ex-
ample, see Gatski and Speziale [13], where the SSG model, the differential as
well as its algebraic form, performs well for homogeneous rotating shear flows.
In this study we will investigate the differences between these models for the
vortex flow.
The term that remains to be modeled in the Reynolds stress transport equa-
tion (6) is the turbulent transport term. The molecular and turbulent diffusion
of the Reynolds stresses, Dij , is

êi êj Dij = ∇ · ν∇τ − T (14)
where T is the turbulent flux of Reynolds stresses. In the present computation

we will adopt the Daly and Harlow [19] model that reads
K
T = −cs τ · ∇τ (15)
ε
where cs = 0.25. With the assumptions we have made for the vortex, the diffu-
sion becomes

1 ∂ K ∂τij (ex)
Dij = r ν + cs u 2 + Dij (16)
r ∂r ε ∂r
(ex)
where the extra diffusion Dij consists of all terms that emanate from the
nonzero θ̂ derivatives of the unit vectors. The nonzero components of the extra
diffusion read

1 ∂ K 2
K uv ∂ 2
v2 uv
Duv =
(ex)
cs uv u − v 2 + cs u −v −4
2 − 4ν 2
r ∂r ε ε r ∂r r r

1 ∂ K
Duu
(ex)
= −Dvv (ex)
= −2cs uv 2
r ∂r ε


v 2 u 2 − v2
 − 2ν u − v
K uv ∂uv 2 2
− 2cs  + (17)
ε r ∂r r2 r2
Finally, the Daly and Harlow model for the dissipation reads

1 ∂ K ∂ε
D(ε) = r ν + cε u 2 (18)
r ∂r ε ∂r
where cε = 0.15.
2.2. Algebraic Reynolds stress modeling. An alternative of writing the

system of equations in terms of the Reynolds stresses is to reformulate the equa-
tions in terms of the Reynolds stress anisotropy and the turbulent kinetic energy.
In flows where the anisotropy varies slowly in time and space, the transport equa-
tion for the Reynolds stress anisotropy tensor reduces to an implicit algebraic
relation. Also in many inhomogeneous flows of engineering interest, the flow is
steady and the advection and diffusion of the Reynolds stress anisotropy may
be neglected [20] [21]. This is equivalent to assuming that the advection and
diffusion of the individual Reynolds stresses scale with those of the turbulent
kinetic energy.
The advection of the Reynolds stress tensor is, however, dependent on the
coordinate system and would, thus, lead to an algebraic relation and a model
that is dependent of the choice of coordinate system. This could be of major
importance in flows with strong streamline curvature, such as strongly rotating
flows. In circular flows it is, therefore, most appropriate to neglect the advection
of the anisotropy following a streamline, which is similar to neglecting the rate of
change of the anisotropy following a fluid particle. For the vortex flow considered
in this study, this is fulfilled exactly by including the algebraic terms in (10).
In a more general flow it is possible to define a Galilean invariant curvilinear
coordinate system based on the acceleration vector [10].
By the use of the general linear pressure-strain model in (12), the transport
equation for the Reynolds stress anisotropy tensor reads

K ∂aij (a) L1 P K
− Dij = − L01 + 1 aij + 2L2 Sij
ε ∂t 2 ε ε

K 2 K
+L3 aik Skj + Sik akj − akl Slk δij − L4 aik Ω∗kj − Ω∗ik akj (19)
ε 3 ε
The advection of the Reynolds stress anisotropy is, for the vortex, exactly in-
cluded in the modified rotation rate tensor
Ω(r)
Ω∗ = Ω − (20)
L4
where Ω(r) is given by (11). The L coefficients are related to the C coefficients
in (12) through
C10 C2 2 C3 C4
L01 = − 1, L11 = C11 + 2, L2 = − , L3 = − 1, L4 = −1
2 2 3 2 2
(21)
The transport equation for the turbulent kinetic energy is obtained by taking
half of the trace of the equation for the Reynolds stress tensor
∂K
− D(K) = P − ε (22)
∂t
where the diffusion of turbulent kinetic energy is

1 1 ∂ K ∂K
D(K) = Dii = r ν + cs u 2 (23)
2 r ∂r ε ∂r
Neglecting the time dependency and the molecular and turbulent diffusion
(a)
of the Reynolds stress anisotropy Dij , the left-hand side of (19), leads to an
algebraic relation

L11 P K
0 = − L1 + 0
aij + 2L2 Sij
2 ε ε

K 2 K
+ L3 aik Skj + Sik akj − akl Slk δij − L4 aik Ω∗kj − Ω∗ik akj (24)
ε 3 ε
Also in the algebraic equation, advection of the Reynolds stress anisotropy is

exactly included through the modified rotation rate tensor Ω∗ (20). Thus, the
importance of the advection may be investigated by excluding the Ω correction.
2.3. Explicit algebraic Reynolds stress modeling. Using representation

theory, Pope [22] proposed that the implicit relation (24) may be solved so that
an explicit relation for the anisotropy is obtained. The scalar nonlinearity in
P/ε, because P = −Kaij Sji , may be expressed as a cubic polynomial that has
an explicit solution [12] [23]. That will result in a fully explicit and self-consistent
algebraic relation for the Reynolds stress anisotropy tensor that may be written
as
2 2
K K 2 K
aij = β1 Sij + β2 Sik Skj − IIS δij + β4 Sik Ω∗kj − Ω∗ik Skj
ε ε 3 ε
(25)
where
A1 N A1 A2 A1
β1 = − β2 = 2 β4 = − (26)
Q Q Q
and
2 2
K 2 K
Q = N2 − 2 IIΩ∗ − A22 IIS (27)
ε 3 ε
The invariants IIS ≡ Sik Skj and IIΩ∗ ≡ Ω∗ik Ω∗kj , and N ≡ A3 + A4 P/ε. The
solution to the cubic equation for P/ε or N reads

1/3
1/3
A3 + P1 + P2 + P1 − P2 , P2 ≥ 0
3
N= A3 2 1/6 1 P1 (28)
+ 2 P1 − P2 cos arccos 2 , P2 < 0
3 3 P1 − P2
where
2 2
A3 2 A1 A4 2 2 K 2 K
P1 = + − A2 IIS − IIΩ∗ A3
27 6 9 ε 3 ε
2 2 3
A3 2 A1 A4 2 2 K 2 K
P2 = P1 −
2
+ + A2 IIS + IIΩ∗ (29)
9 3 9 ε 3 ε
When computing the third root, one need to make sure that the real root is
returned even for negative arguments, and the arccos function should return an
angle between 0 and π. The A coefficients are related to the L coefficients in
(19) through
L2 L3 L01 1 L11
A1 = 2 , A2 = , A3 = − , A4 = − (30)
L4 L4 L4 2 L4
Thus, the Reynolds stress anisotropy is fully explicitly expressed in terms
of the mean flow gradient and the turbulent time scale (K/ε). This replaces
the transport equation for the anisotropy tensor and is used together with the
transport equations for the turbulent kinetic energy (22) and the dissipation
rate (9). The Reynolds stresses are directly related to the anisotropy through
definition (13). The Daly and Harlow modeling of the transport of K and ε, (23)
and (18), is also kept at this level of modeling.
The formal reduction of the full Reynolds stress transport model results in a
two-equation turbulence model where the standard eddy-viscosity assumption is
replaced with a more complete constitutive relation. This can be compared to
the eddy-viscosity assumption that relates the Reynolds stresses linearly to the
mean flow strain rate tensor
K
aij = −2Cµ Sij (31)
ε
where the coefficient Cµ = 0.09. Even the first term in (25) gives a much better
description of the anisotropy since the effective Cµeff = −β1 /2 is not a constant
but is dependent on the local flow state. Moreover, the eddy-viscosity assumption
does not give realistic values of the normal Reynolds stress components, and thus,
the Daly and Harlow diffusion model could not be used. The diffusion terms in
the K and ε equations, for the standard eddy-viscosity K-ε model, are thus
modeled using a simple gradient diffusion model that reads

1 ∂ Cµ K 2 ∂K
D (K)
= r ν+
r ∂r σK ε ∂r
2

1 ∂ Cµ K ∂ε
D(ε) = r ν+ (32)
r ∂r σε ε ∂r
where the standard values σK = 1.0 and σε = 1.3 are used.

3. Numerical methods and boundary conditions

The system of equations is integrated in time using a first-order implicit method
(backward Euler) from a prescribed initial condition. For every time step, the
system of equations is iterated until convergence by reducing the residual 6
decades. Second-order central differences on a staggered grid is used for the
spatial discretization.
Typically, 1000 time steps for two-equation computations and 2000 time steps
for full Reynolds stress computations were used in the computations. Decreasing
the time step did not influence the result here. 100 computational points in the
radial direction were used, 20 points with an equidistant distribution within
r = 2Rc and the other points smoothly stretched towards the outer boundary.
Doubling the number of grid points alters the computed circulation by less than
1%.
To keep the symmetry for the velocity components in the cylindrical coordi-
nate system, the tangential velocity as well as the Reynolds shear stress compo-
nent must be zero. Moreover, the normal Reynolds stress components as well as
K and ε should have zero spatial derivatives. The boundary conditions at r = 0
are
∂u2 ∂v 2 ∂w2 ∂K ∂ε
V = = = = = = uv = 0 (33)
∂r ∂r ∂r ∂r ∂r
The outer edge of the computational domain is situated far away from the
vortex. The boundary condition for the tangential velocity is obtained by as-
suming that the flow is inviscid, and thus, the vortex circulation Γ approaches a
constant. The boundary condition for the turbulence quantities are set in such
a way that an initial freestream turbulence level would properly decay in time.
This requires, however, that the computational domain is large enough so that
the production and diffusion at the outer boundary, Rmax , are negligible. In this
study Rmax ≈ 100 core radii. The boundary conditions at r = Rmax are
∂ ∂u2 ∂v 2 ∂w2 ∂K ∂ε ∂uv
(rV ) = = = = = = = 0. (34)
∂r ∂r ∂r ∂r ∂r ∂r ∂r
The system of equation and the boundary conditions are specified in a differen-
tial form using the symbolic manipulation tool MAPLE. FORTRAN subroutines
that compute the discrete system matrix are then automatically generated with
the aid of routines written in MAPLE. The sources of programming errors are
then reduced to the much more limited number of lines of the MAPLE code
compared to the final Fortran code.
4. The test case and computational results

In the test case adopted for this study we will eliminate all effects of ambient
influence, such as atmospheric turbulence, stratification, or the other vortex, to
study the isolated effect of vortex decay subjected to self-generated turbulence.
1.2
Lidar
Lidar
Proctor
Lamb
0.8
Gamma
0.4
0.0
-1.0 0.0 1.0 2.0
log(r)
Figure 1. Different initial circulation profiles, Proctor and Lamb,

compared to measured profiles at t = 5s.
1.2
Lidar
Lidar
Initial
0.8
Gamma
0.4
0.0
-1.0 0.0 1.0 2.0
log(r)
Figure 2. Measured circulation profiles at t = 55s compared to the

initial profile (Proctor).
The outer edge boundary condition is, thus, put far away, and the vortex is
initiated assuming that it extends to infinity. LIDAR measurements of an aircraft
trailing vortex at the Memphis airport, case 1252 [24] [25], will be used for a
qualitative comparison with the computations. The case is characterized by
low ambient turbulence during nighttime and slow vortex decay rate. Despite
the low turbulence levels, it is still believed that the ambient conditions are of
importance for the vortex decay rate. Comparisons with the observed vortex
decay rate are, thus, not fully relevant because these effects are not present in
the computations. Nevertheless, the measured circulation profile at t = 5s will
be used as initial conditions for the computations, and the computational results
will be compared to profiles measured at t = 55s, shown in Figures 1 and 2.
4.1. Initial conditions. The vortex circulation, defined as Γ ≡ 2πrV , in-
creases from zero at the core center and levels up to a fairly constant level, Γ0 ,
outside of the vortex core. The vortex strength Γ0 is typically around 350m2 /s
and the core radius Rc , defined as the position of the peak tangential velocity,
is typically around 1.8m. That gives a Reynolds number based on Γ0 of about
20 · 106 .
The problem is made nondimensional based on Γ0 = 350m2 /s and Rc = 1.8m.

That means that all lengths are normalized by Rc , time by Rc2 /Γ0 , velocity by
Γ0 /Rc , turbulent kinetic energy and Reynolds stresses by Γ20 /Rc2 , and dissipation
rate by Γ30 /Rc4 .
Laminar vortices where the viscosity is constant are self-similar in time with
the normalized vortex circulation given by

Γ = 1 − exp −αr2 (35)
This is referred to as a Lamb vortex (see Saffman [26], p. 253). The coefficient α
is usually around 1.26, which gives a maximal normalized velocity Vmax ≈ 0.11.
Measurements of aircraft wing-tip trailing vortices over airfields show that the
typical circulation distribution is somewhat different from a Lamb vortex. By
considering a large set of data from vortex measurements, Proctor [1] proposed
an alternative form that better matches the data
3/4
1
1.4 1 − exp −10 1 − exp −1.2527r2 ,r < 1
B
Γ= r
3/4 (36)
1 − exp −10 ,r ≥ 1
B
where B is the (normalized) span width of the airplane that generates the vortex.
This gives the normalized maximal velocity Vmax ≈ 0.11. The initial circulation
profiles, Lamb and Proctor, are compared to measured data in Figure 1. The
vortex is initially prescribed according to (36) using the normalized span width
B = 17.7. The initial profile does not perfectly match the data, but the initial
core radius was fit to a large number of measurements and the data in Figure 1
are only one set of measurements.
The initialization of the turbulence is, however, not obvious because tur-
bulence field measurements within the vortex are rare. The measured atmo-
spheric turbulence could, however, act as a lower limit for the initial vortex
turbulence. For the low turbulence case considered (Dallas 1997 Wake Vortex
Data Set CD-ROM, personal communication with Proctor, Flight Dynamics and
Control Division, NASA Langley Research Center), the turbulent kinetic energy
K ∼ 0.02m2 /s2 and the dissipation rate ε ∼ 10−5 m2 /s3 . That corresponds to
a normalized turbulence level of the order of K ∼ 10−7 and ε ∼ 10−12 . The
integral length scale is of the order of 102 (core radius) which may be unrealis-
tically high for the self-generated turbulence in the vortex. The dissipation rate
is, however, universal over all turbulent scales, and by assuming a length scale of
the vortex self-generated turbulence in the order of unity, one obtains K ∼ 10−8
and ε ∼ 10−12 . At these low initial levels the growth rate is so slow that the
flow remains nearly laminar in the RANS computations.
The rollup of a wing-tip vortex takes place in the very near-wake region, and
a fully developed and nearly symmetric vortex is formed within the near wake.
In the experiment by Chow et al. [4], rollup happened within less than one chord
length from the trailing edge of the wing, whereas Devenport et al. [5] observed
a much more extended rollup of the near wake. The vortex is fully turbulent,
although the turbulence shear stress in the core is strongly suppressed. The tur-
bulence level at the most downstream measured section is of the order of 10% of
the axial freestream velocity, and the maximum swirl velocity was of the same
order as the axial freestream velocity. By assuming that the turbulence length
scale is of the order of the core radius, one obtains the normalized K = 10−4 and
ε = 10−6 . Using that as initial condition, one found that the turbulence levels
quickly decrease to some asymptotic level that is about one order of magnitude
lower, shown in Figure 3 (a). During that process, the high-intensity turbu-
lence diffuses the tangential velocity before the decay rate settles down to some
asymptotic level. The transient behavior of the circulation is seen in Figure 4
(a). By the setting of the initial turbulence level to the order of that asymptotic
state, the initial transient is eliminated and the computed time history becomes
somewhat independent of small initial variations and the exact form of the initial
profiles.
The turbulence is initially assumed homogeneous, or uniform, with the nor-
malized K0 = 10−5 and ε0 = 10−8 , which gives a freestream turbulence level
of 2% of Vmax and the turbulent macro lengthscale Λ = K 3/2 /ε ≈ 3.2. In the
differential Reynolds stress computations, the normal Reynolds stress anisotropy
is prescribed to be zero, and the normalized Reynolds stress shear component
uv = K0 V (r). Figure 4 (b) shows a typical time evolution of a computation
initiated with these conditions.
The vortex flow was integrated for a time period of 50s, which corresponds
to 600 vortex turnover times defined as Rc /Vmax . The computed results may
then be compared to the measured data at t = 55s. Sarpkaya and Daly [2]
introduced an alternative timescale based on the downsweep of the vortex pair
and the initial vortex separation t∗ = V0 t/B, where V0 = Γ0 /(2πB). Note that
55s corresponds to t∗ = 3.0.
4.2. Reynolds stress transport models. The flow in the core of the vortex
approaches solid body rotation, and thus, the production of the turbulence is
suppressed. In the outer part of the vortex, the flow gradually changes from
rotation-dominated to completely strain dominated. It is well known that stan-
dard eddy-viscosity two-equation models, such as the K-ε model, are unable
to describe the turbulence in rotation-dominated flows satisfactorily. Also the
response of different RST models are known to be substantially different in
rotation-dominated flows.
The results using the different RST models are shown in Figure 5. The SSG
model is known to perform better than the LRR model in rotation-dominated
flows, and the results show that the linear SSG compares well with the mea-
surements in this case whereas the LRR model predicts a decay rate slower than
observed. Moreover, no overshoot in circulation could be seen, which is in line
0.12E-03 0.12E-03
0 0
(a) 1 (b) 1
2 2
3 3
4 4
5 5
0.80E-04 6 0.80E-04 6
K 7 K 7
8 8
9 9
10 10
0.40E-04 0.40E-04
0.00E+00 0.00E+00
-1.0 0.0 1.0 2.0 -1.0 0.0 1.0 2.0
log(r) log(r)
Figure 3. The evolution of the computed turbulent kinetic energy

subjected to different initial turbulent levels using EARSM based on
SSG. (a): High levels where K = 10−4 and ε = 10−6 . (b): Nominal
levels where K = 10−5 and ε = 10−8 . The spacing between each line
is 5s.
1.2 1.2
0 0
(a) 1 (b) 1
2 2
3 3
4 4
5 5
0.8 6 0.8 6
Gamma 7 Gamma 7
8 8
9 9
10 10
0.4 0.4
0.0 0.0
-1.0 0.0 1.0 2.0 -1.0 0.0 1.0 2.0
log(r) log(r)
Figure 4. The evolution of the computed circulation subjected to

different initial turbulent levels using EARSM based on SSG. (a):
High levels where K = 10−4 and ε = 10−6 . (b): Nominal levels
where K = 10−5 and ε = 10−8 . The spacing between each line is 5s.
with observations. A further discussion about a possible overshoot is given in

section 4.5.
Figure 5 also shows the Reynolds shear stress for the different RST models.
The models correctly predict the strong damping of the turbulence in the vortex
core, but in the outer part of the vortex the two models predict very different
turbulence shear stresses. The low uv-values predicted using LRR are consistent
with the slow vortex decay rate. Figure 6 shows the turbulent kinetic energy,
which, opposite to uv, do not approach zero in the vortex core. This has also
been observed by Devenport et al. [5] experimentally and was explained as
influence from the surrounding turbulence.
1.2 0.3E-05
Initial L-SSG
(a) L-SSG (b) LRR
LRR
0.8 0.2E-05
Gamma uv
0.4 0.1E-05
0.0 0.0E+00
-1.0 0.0 1.0 2.0 -1.0 0.0 1.0 2.0
log(r) log(r)
Figure 5. The computed vortex circulation (a) and shear stress (b)
profiles at t = 55s using different RST models. The flow is initiated
using the Proctor profile.
0.6E-04
Initial
L-SSG RST
LRR RST
SSG EARSM
LRR EARSM
0.4E-04
K
0.2E-04
0.0E+00
-1.0 0.0 1.0 2.0
log(r)
Figure 6. The computed turbulent kinetic energy profiles at t = 55s

using different models. The flow is initiated using the Proctor profile.
4.3. Algebraic Reynolds stress models. Reynolds stress transport models

are still not widely used in engineering methods, and so an interesting question
is, is it possible to obtain computational results similar to the full SSG by means
of much simpler two-equation models?
Self-consistent EARSMs are the most consistent approximations of the basic
RST models at a two-equation level. In Figure 7, computational results using
EARSM based on both the linearized SSG and the LRR are shown. The EARSM
level of modeling follows the behavior of the full Reynolds stress models they are
based on. The EARSM based on the LRR predicts a low decay rate of the vortex
whereas the EARSM based on the linearized SSG predicts a vortex decay rate
in line with observations. For comparison, the result from using the standard
K-ε eddy-viscosity model is also shown in Figure 7. As expected, the vortex
decay rate is very much overpredicted, and there is a massive overshoot in the
circulation profile.
An important difference between the differential and algebraic form of the

linearized SSG is the prediction of the Reynolds shear stress and the turbulent
kinetic energy (Figure 6). In the outer parts of the vortex the EARSM predicts
a much lower uv compared to the full RST. However, this has just a small
implication on the mean swirl velocity as long as the shear stress in and near the
vortex core is reasonably predicted.
In the EARSM computations, the weak equilibrium assumption is made in
a Galilean invariant curvilinear coordinate system. The importance of that is
investigated by computing one case, EARSM based on SSG, where the weak
equilibrium assumption is made in an inertial system. That is equivalent of
letting Ω∗ = Ω in (24). The results are labeled (SSG no rot) in Figure 7 and
show a vortex decay rate that is complete in error. This result demonstrates the
importance of the correct treatment of the advection term in algebraic modeling
of turbulent strongly curved flows. Neglecting this effect makes EARSM as bad
as standard eddy-viscosity models.
The EARSM approach taken for this study is that the explicit solution is
fully self-consistent. That means that the implicit algebraic relation is exactly
solved without further assumptions. The importance of this approach is assessed
by computing the case using the EARSM based on the linearized SSG model,
but without considering the solution for P/ε or N in (28), but rather to use
the equilibrium value in homogeneous shear. This is similar to the Gatski and
Speziale model [13], but in this study we do not consider the regularization of the
denominator to avoid singular behavior that is a part of the Gatski and Speziale
model, and thus, the only difference between the models is the consistency con-
dition. This model is labeled (G&S) in Figure 8. The G&S model gives a similar
vortex decay rate compared to the SSG, but the circulation profile is distorted
just outside of the core, and the decay rate is somewhat lower in the outermost
regions of the core. It is, however, possible that the regularization would cure
this behavior, but that is not considered in this study.
The turbulent diffusion of K and ε in the EARSM computations are deter-
mined using the Daly and Harlow model, which relates the diffusivity to the u
fluctuations, equations (23) and (18). The importance of the diffusion modeling
is determined by using the standard eddy-viscosity gradient diffusion model of
the diffusion (32) as a comparison. To obtain somewhat realistic values on the
diffusivity, the Cµ coefficient is not taken as a constant, but is related to β1 in
(26), Cµ = −β1 /2. Figure 8 shows this comparison for the EARSM based on
the linearized SSG model, where (SSG) and (SSG nut) labels the computations
using the Daly and Harlow, and eddy-viscosity modeling of the diffusion, respec-
tively. There are only minor differences between the two modeling approaches
indicating that the turbulent diffusion modeling is not a critical issue.
The LRR and SSG models gave very different behavior for this flow, and one
important difference between the two models is the C2 coefficient before the Sij
1.2 0.3E-05
Initial SSG
(a) SSG (b) LRR
LRR SSG no rot
SSG no rot k-eps
k-eps
0.8 0.2E-05
Gamma uv
0.4 0.1E-05
0.0 0.0E+00
-1.0 0.0 1.0 2.0 -1.0 0.0 1.0 2.0
log(r) log(r)
Figure 7. The computed vortex circulation (a) and shear stress

(b) profiles at t = 55s using different two-equation models, differ-
ent EARSM’s and the standard eddy-viscosity K-ε model (k-eps).
The flow is initiated using the Proctor profile.
term in (12). In the linearized SSG model, the equilibrium value C2 = 0.36 is
used whereas C2 = 0.8 in the LRR model, which corresponds to rapid distortion
theory. The algebraic form of the LRR model (using the newly re-calibrated
coefficients) is especially attractive because the complexity of the algebra in
three-dimensional mean flows is very reduced [14] [23]. It is possible to keep
exactly the same algebraic form by changing the L coefficients in (21) so that
the A coefficients in (30) remain unchanged and the C2 coefficient takes the
value of 0.36 rather than 0.8. That leads to the alternative set of C coefficients
where C10 = 4.92, C11 = 1.65, C2 = 0.36, C3 = 2 and C4 = 0.378. Because
the A coefficients remain unchanged compared to the LRR model, the basic
EARSM remains unchanged, and the simple algebraic form is preserved. The
only difference enters through the modified rotation rate tensor to account for
the advection. The computational result from using this model, (mix) in Figure
8, is much closer to the SSG model than the LRR model. Thus, it is reasonable
to believe that this mechanism is the most important difference between the
SSG and LRR models for this flow and that the decay rate of the vortex is much
controlled by the C2 coefficient. This mixed model is, however, not yet another
proposal of a linear pressure-strain model, but only constructed to assess the
differences between the SSG and LRR models and the importance of the C2
coefficient.
4.4. Variation of the initial circulation profile. The observations in the
preceding sections are made using an initial circulation profile that corresponds
to observed trailing vortices. That initial circulation profile deviates somewhat
from the Lamb vortex that has been used in previous studies of trailing vortices,
for example, see Zeman [7]. The two different initial profiles are shown in Figure
1. The computations with this alternative (Lamb) initial profile follows the
previous computations, see Figure 9. The SSG-based models, in differential
1.2 0.3E-05
Initial SSG
(a) SSG (b) SSG nut
SSG nut mix
mix G&S
G&S
0.8 0.2E-05
Gamma uv
0.4 0.1E-05
0.0 0.0E+00
-1.0 0.0 1.0 2.0 -1.0 0.0 1.0 2.0
log(r) log(r)
Figure 8. The computed vortex circulation (a) and shear stress (b)
profiles at t = 55s using different EARSM’s based on the linearized
SSG model. The flow is initiated using the Proctor profile.
1.2
Initial
L-SSG RST
LRR RST
SSG EARSM
laminar*
0.8
Gamma
0.4
0.0
-1.0 0.0 1.0 2.0
log(r)
Figure 9. The computed vortex circulation profiles at t = 55s us-

ing different models. The laminar results was obtained at a lower
Reynolds number, Re = 104 . The flow is initiated using the Lamb
profile.
form as well as in algebraic form, give reasonable results whereas the LRR-based
models strongly underpredict the vortex decay rate. Also here one may observe
that the algebraic modeling assumption gives results comparable to the full RST
models.
Also shown in Figure 9 are the laminar results where the Reynolds number
has been decreased from 20 · 106 to 104 ; otherwise there would have been almost
no difference between the initial and final profiles. The laminar result shows
a complete different nature where the vortex decays mostly from the core and
outwards giving an increased vortex radius while leaving the exterior part of the
vortex unaffected. The laminar decay rate is also much slower than the turbulent
decay rate.
4.5. Comments on the obtained results. There are some questions con-
cerning the obtained results that need to be further discussed. The vortices
studied here are linearly stable according to Rayleigh’s [27] stability criterion.
The model computations show an initial growth of turbulence that saturates
and starts to decrease slowly (Figure 3b). The level at which the turbulence
saturates is, however, fairly low, K ∼ 10−5 (normalized by Γ0 and Rc ). This
could be compared to the large-eddy simulations by Ragab and Sreedhar [6],
where the computations were interrupted when K ∼ 10−4 , or the measurements
by Devenport et al. [5], where the maximal level at the final streamwise position
was K ∼ 10−4 .
If the initial turbulence level in the model computations is prescribed much
higher than the saturation level, no initial growth could be seen. The turbu-
lence rapidly decreases to about the saturation level (Figure 3a). If the initial
turbulence level is prescribed to be very low, the initial growth is very weak, but
still positive, and the saturation level will not be reached within the time period
investigated here.
The same principal behavior could be observed using models based on SSG, as
well as LRR, but the saturation level is much lower when using the LRR model.
The question is if this behavior, predicted by the models, is physically reason-
able or not. We know that the vortex is linearly stable (at least if we ignore that
the mean flow is not stationary), and thus, we do not have exponential growth of
any of the eigenmodes. Also if the eigenmodes are orthogonal, we would expect
an overall exponential decay of any disturbances. However, the eigenmodes are
not orthogonal, and thus, there exists a mechanism for disturbances to grow.
The perhaps most classical example of this is plane Couette flow that is linearly
stable for all Reynolds numbers [28] but transient growth nevertheless could lead
to turbulence.
Schmid et al. [29] studied the eigenvalue sensitivity for the Batchelor vortex,
for example, which is linearly unstable due to an axial velocity deficit. Large
sensitivity is closely connected to the possibility of transient growth. For this
case they found strong transient growth that was not related to the most unsta-
ble mode but rather due to large sensitivities of stable or weakly unstable modes.
From this, it is not unrealistic to expect similar behavior for other types of vor-
tices, even if they are linearly stable, and that an alternative growth mechanism
may exist.
To our knowledge it is not clear whether the Lamb type of vortex could sustain
turbulence or not. Numerical experiments (for example, large-eddy simulation
by Ragab and Sreedhar [6]) did not show any evidence of sustained turbulence for
the linearly stable case. However, the saturated turbulence level could be quite
low without any large structures, which is typical for cases with a highly unstable
mode, and may be overlooked if not explicitly searched for. These computations
were interrupted when K ∼ 10−4 , which is one order of magnitude larger than
the saturated level predicted by our SSG model computations. Moreover, the
stable case computed by Ragab and Sreedhar shows an initial behavior that
could be interpreted as transient growth.
An other question concerns a possible overshoot in the circulation profile.
The angular momentum must be conserved, and Govindaraju and Saffman [11]
showed that a vortex that entrains the ambient at a rate greater than that due to
molecular diffusion must develop a circulation overshoot. However, no overshoot
could be seen in our computational results, except for the strongly diffusive
eddy-viscosity K-ε model. In the analysis by Govindaraju and Saffman, the
Reynolds shear stress uv was assumed to decrease with increasing radius at a
rate faster than r−2 . That means that an effective eddy viscosity must decrease
with increasing radius. In our computations we have initiated the turbulence
with a uniform K and ε field, which gives an effective eddy viscosity that far
exceeds the molecular viscosity also at the outer boundary of the computational
domain. The angular momentum or circulation will, thus, be transported out
of the computational domain without developing any overshoot of circulation
within the domain.
The assumption of an initially uniform turbulent field is not unreasonable
when studying vortices in atmospheric conditions because the atmospheric tur-
bulence level may act as a lower limit on the eddy viscosity at increasing radius.
This could actually be the reason for the lack of experimentally observed circu-
lation overshoots.
5. Conclusions
The turbulent vortex flow investigated is strongly affected by rotation, and even
at the Reynolds stress transport modeling level, there are important differences
between different models. All models based on the Reynolds stress transport
equations, differential as well as algebraic, predict the strong suppression of
the turbulence within the vortex core, consistent with observations. However,
models based on the LRR pressure-strain model predict that the turbulence
is almost completely suppressed also in the exterior part of the vortex, and
the predicted vortex decay rate is much lower than observed. In that sense,
models based on SSG perform more in line with observations. Computations
show a fully turbulent vortex outside of the stabilized vortex core, and the vortex
decay rate is of the same order as observed decay rates during low turbulent
ambient conditions. The computations were made by assuming a vortex that
extends to infinity, and thus, any real influence from ambient conditions are not
included. The comparison with measured data is therefore not fully relevant
because strong correlations between ambient conditions and vortex decay rates
has been observed and direct comparisons should be avoided. For the idealized
case of an isolated vortex, it is not obvious which of the SSG or LRR models
give the most physically reasonable results, nor if the isolated vortex is capable
References 139
of sustaining turbulence at all. However, the particular observation used in this

study was one of the slowest decaying ones, and thus, it would be somewhat
surprising if the idealized case would substantially deviate from the observed
one.
EARSMs are attractive due to reduced complexity and computational effort
compared to full RST models. The computational results using EARSM closely
follow the results obtained with the full RST models. It is, however, important
that the algebraic assumption is made in a streamline based coordinate system.
Neglecting this effect makes the model as bad as the eddy-viscosity K-ε model,
which strongly overpredicts the vortex decay rate. Moreover, it is important that
the EARSM is based on a self-consistent approach where the implicit algebraic
relation is exactly solved without further assumptions. The modeling of the
diffusion terms in the K-ε equations is not very important in this flow. The
anisotropy diffusivity approach by Daly and Harlow [19] gives results comparable
to the standard eddy-diffusivity approach as long as the eddy diffusivity is based
on an effective Cµ determined from the EARSM relations rather than a universal
constant (usually 0.09).
Acknowledgments
The authors gratefully acknowledge funding for this study from the Aeronautical
Research Institute of Sweden (FFA) and from the Flight Dynamics and Control
Division at NASA Langley Research Center. This study was performed at the
Institute for Computer Applications in Science and Engineering, NASA Langley
Research Center, during a visit by the first author. The authors would like to
thank F. Proctor for introducing us to this problem and sharing with us his
experience in this field.
References
[1] Proctor, F.H. 1998 The NASA-Langley Wake Vortex Modeling Effort
in Support of an Operational Aircraft Spacing System. Invited paper, 35th
Aerospace Sciences Meeting & Exhibit, 12-15 January, Reno, NV AIAA Paper
No. 98-0589, 19 pp.
[2] Sarpkaya, T. & Daly J.J. 1987 Effect of Ambient Turbulence on Trailing
Vortices. J. Aircraft 24 (6) 399–404.
[3] Sarpkaya, T. 1998 Decay of Wake Vortices of Large Aircraft. AIAA J. 36
(9) 1671–1679.
[4] Chow, J.S., Zilliac, G.G. & Bradshaw, P. 1997 Mean and Turbulence
Measurements in the Near Field of a Wingtip Vortex. AIAA J. 35 (10) 1561–
1567.
[5] Devenport, W.J., Rife, M.C., Liapis, S.I. & Follin, G.J. 1996 The
structure and development of a wing-tip vortex. J. Fluid Mech. 312, 67–106.
140 References
[6] Ragab, S. & Sreedhar, M. 1995 Numerical simulation of vortices with

axial velocity deficits. Phys. Fluids 7 (3) 549–558.
[7] Zeman O. 1994 The persistence of trailing vortices: A modeling study. Phys.
Fluids 7 (1) 135–143.
[8] Speziale, C.G., Sarkar, S. & Gatski, T.B. 1991 Modelling the pressure-
strain correlation of turbulence: an invariant dynamical systems approach.
J. Fluid Mech. 227, 245–272.
[9] Launder, B.E., Reece, G.J. & Rodi, W. 1975 Progress in the develop-
ment of a Reynolds-stress turbulence closure. J. Fluid Mech. 41, 537–566.
[10] Girimaji, S.S. 1997 A Galilean invariant explicit algebraic Reynolds stress
model for turbulent curved flows Phys. Fluids. 9 (4) 1067–1077.
[11] Govindaraju, S.P. & Saffman, P.G. 1971 Flow in a Turbulent Trailing
Vortex. Phys. Fluids 14 (10) 2074–2080.
[12] Girimaji, S.S. 1996 Fully explicit and self-consistent algebraic Reynolds
stress model Theor. and Comp. Fluid Dyn. 8 (6) 387–402.
[13] Gatski, T.B & Speziale C.G. 1993 On explicit algebraic stress models
for complex turbulent flows. J. Fluid Mech. 254, 59–78.
[14] Taulbee, D.B. 1992 An improved algebraic Reynolds stress model and
corresponding nonlinear stress model. Phys. Fluids A 4 (11) 2555–2561.
[15] Lumley, J.L. 1978 Computational modeling of turbulent flows. Adv. Appl.
Mech. 18 123–176, Academic Press, New York.
[16] Shabbir, A. & Shih, T.H. 1992 Critical Assessment of Reynolds Stress
Turbulence Models Using Homogeneous Flows NASA TM 105954, ICOMP-92-
24, CMOTT-92-12.
[17] Wallin, S. & Johansson, A.V. 1996 A new explicit algebraic Reynolds
stress turbulence model including an improved near-wall treatment. In Proc.
Flow Modeling and Turbulence Measurements VI, Tallahassee FL, C.-J. Chen,
C. Shih, J. Lienau & R.J. Kung (eds), 399–406.
[18] Wallin, S. & Johansson, A.V. 1999 An explicit algebraic Reynolds stress
model for incompressible and compressible turbulent flows. To appear in J.
Fluid Mech. (Also FFA TN 1997-51).
[19] Daly, B.J. & Harlow, F.H. 1970 Transport equations in turbulence. Int.
Phys. Fluids 13 (11) 2634–2649.
a two-equation model of turbulence. Ph.D. thesis, Imperial College of Science
and Technology, Dept. of Mechanical Engineering; HTS/72/24.
stresses. Z. angew. Math. Mech. 56 (3) T219–221.
[22] Pope, S.B. 1975 A more general effective-viscosity hypothesis J. Fluid
Mech. 72, 331–340.
[23] Johansson, A.V. & Wallin, S. 1996 A new explicit algebraic Reynolds
stress model. In Proc. Sixth European Turbulence Conference, Lausanne, July
References 141
1996, Eds. S. Gavrilakis, L. Machiels & P.A. Monkewitz, Kluwer, 31–34.

[24] Campbell, S.D., Dasey, T.J., Freehart, R.E., Heinrichs, R.M.,
Matthews, M.P. & Perras, G.H. 1996 Wake Vortex Field Measurement
Program at Memphis, TN. 34th Aerospace Sciences Meeting & Exhibit, Reno,
NV, AIAA Paper No. 96-0399, January 1996.
[25] Campbell, S.D., Dasey, T.J., Freehart, R.E., Heinrichs, R.M.,
Matthews, M.P., Perras, G.H. & Rowe, G.S. 1997 Wake Vortex Field
Measurement Program at Memphis, TN Data Guide. Project Report: NASA/L-
2, January 1997.
[26] Saffman, P.G. 1992 Vortex Dynamics. Cambridge University Press, New
York.
[27] Rayleigh, Lord 1916 On the dynamics of revolving fluids. Proc. R. Soc.
A 93 148–154.
[28] Herron, I.H. 1991 Observations on the role of vorticity in the stability
theory of wall bounded flows. Stud. Appl. Math. 85 (3) 269–286.
[29] Schmid, P.J., Henningson, D.S., Khorrami, M.R. & Mujeeb, R.M.
1993 A study of eigenvalue sensitivity for hydrodynamic stability operators.
Theoret. Comput. Fluid Dyn. 4 (5) 227–240.
Paper 4
4
A COMPUTATIONAL STUDY OF UNSTEADY TURBULENT
BUFFET AERODYNAMICS
Dieqian Wang, Stefan Wallin, Martin Berggren and Peter Eliasson
Abstract. The periodic self-excited turbulent flow around an 18-percent

thick circular-arc airfoil at transonic speeds has been computed by solving
the unsteady Reynolds-averaged Navier-Stokes equations. Three different
turbulence models were used: an algebraic model, an eddy-viscosity K − ω
model, and an explicit algebraic Reynolds stress model (EARSM). Grid and
time-step studies were performed to assess the numerical accuracy. The
EARSM is a fully self-consistent approximation of a Reynolds stress trans-
port model in the weak-equilibrium limit and behaves reasonable well also
in non-equilibrium flows. Moreover, the model has the correct near-wall
asymptotic behavior for all individual Reynolds stresses with a near-wall
damping function formulated without use of the wall skin friction (or y + ).
Computational results indicate that the EARSM predicts the unsteadiness
due to shock and boundary-layer separation better than the two-equation
model based on the Boussinesq hypothesis and the algebraic turbulence
model. The EARSM predicts the frequency of self-exited unsteady tran-
sonic flow to high accuracy and the motion of the shock agrees well with
the experimental data.
1. Introduction
Flows around blunt bodies, airfoils near and in stall, flows around turbine blades
and shock induced separation on airfoils are some examples of unsteady turbulent
flows of engineering interest. The unsteadiness may be forced or self-induced due
to flow instabilities, and interaction with the structure (aeroelasticity) may be
present in the examples above.
In this study we focus on buffet aerodynamics. Transonic flow over an airfoil
may result in a periodic motion of the shock over the surface of the airfoil. The
unsteadiness is driven by the interaction between the shock, boundary layer and
the vortex shedding in the wake. The reduced frequency is usually less than one
which means that the time scale of the wall-bounded turbulence is much smaller
than the periodic unsteadiness which motivates a quasi-steady approach.
The 18% bicircular-arc airfoil has been excessively studied experimentally and
numerically with respect to the buffet behavior and was chosen as the valida-
tion case in this study. Previous computational studies using Reynolds averaged
Navier-Stokes methods were performed by Levy[2], Steger[3], Seegmiller et al.[4],
Edwards and Thoms[5]. All employed algebraic turbulence models. In all of these
studies the self-excited oscillations were reproduced, but the reduced frequencies
145
146 UNSTEADY TURBULENT BUFFET AERODYNAMICS
were predicted about 20 percent lower than in the experiment of McDevitt[6].

The same case was also computed in the context of the ETMA[7] research pro-
gram where the partners employed algebraic or two-equation turbulence models.
The general conclusion from ETMA is that the principal flow characteristics are
predictable, even for the simple algebraic turbulence models, since an impor-
tant part of the complex physical process is contained in the periodic motion,
which is resolved by the phase-averaging approach. But also here, the predicted
frequencies are lower than in the experiment.
In the paper of Rumsey[8], results with the algebraic Baldwin–Lomax[9] (B-
L) and the one-equation Spalart–Allmaras[10] (S-A) turbulence models. The
S-A turbulence model calculations predict the experimental reduced frequency
of the oscillations as well as the Mach number range within which hysteresis
takes place. But the motion of the shock toward and ahead of the mid chord is
not predicted by the S-A model.
In this study, which is a continuation of earlier work[1], we will identify the
general requirements on a turbulence model for unsteady turbulent flows and
make a choice of suitable models to test. The numerical method used for accu-
rate time integration will be described and the accuracy assessed before we will
present computational results on the bicircular arc airfoil.
2. Modelling of unsteady periodic turbulence

Modelling of unsteady turbulence very much resembles the modelling of statis-
tically steady turbulence. The main difference is the definition of the averaging
operation. In steady turbulence an instantaneous flow variable φ is decomposed
into a mean part φ and a fluctuating part φ , that is φ = φ + φ . The mean
is typically defined as the time or ensemble average. Averaging the momentum
equation results in the Reynolds averaged Navier-Stokes (RANS) equation for
the mean flow quantities. The Reynolds stress tensor ui uj is an additional stress
term that appear in the RANS equations and needs to be modelled in terms of
the mean flow quantities.
The mean flow is considered as unsteady when the time scale of the mean
flow unsteadiness is much larger than the characteristic integral time scale of
the turbulence. In that case the turbulence energy spectrum is well separated
from the unsteadiness. The turbulence may, thus, be modelled while the mean
flow unsteadiness is left to be resolved in the unsteady RANS solution. This is
known as the quasi-steady approach where a standard RANS turbulence model,
with the time derivatives included, may be used. No additional modification of
the turbulence model is in principal needed due to the unsteadiness.
If there is no clear separation between the turbulence scales and the unsteadi-
ness, then standard RANS turbulence models cannot be used. This class of prob-
lems needs to be tackled using large eddy simulations (LES) where subgrid-scale
models are used to represent the unresolved stresses. The Reynolds averaged
D. WANG, S. WALLIN, M. BERGGREN AND P. ELIASSON 147
and the LES-filtered Navier-Stokes equations are in principal similar but the
unknown correlation ui uj is modelled completely different in RANS and LES.
When the unsteady turbulent flow is, by nature or forced to be, periodic, the
mean is conveniently defined as a phase average, see e.g. Jin & Braza[13]. The
flow variables are then decomposed into three parts, φ = φ + φ” + φ , where φ
is the time-averaged value, φ” is the periodic component and φ the turbulent
fluctuations. The phase-averaged variable is then defined as φ ≡ φ + φ”.
In the quasi-steady approach RANS turbulence models will be used without
any explicit modifications due to the unsteadiness. However, the unsteadiness of
the mean flow poses some additional requirements on the turbulence model, or
will stress existing requirements on models for steady flows even further. This
will be discussed in the following.
2.1. Requirements on RANS models for unsteady turbulence. Fan &

Lakshminarayana[11] and Mankbadi & Mobark[12] consider unsteady boundary
layers where either the wall or the free stream fluctuates. The flow is character-
ized by a near-wall flow that fluctuates in phase with the wall and an outer region
with a substantial phase shift. An instantaneous log-law does not in general exist
and near-wall formulations based on the log-law and the equilibrium assumption
are not appropriate, according to Fan & Lakshminarayana. This disqualifies al-
gebraic eddy-viscosity models and two-equation models with log-law boundary
conditions or with near-wall damping functions based on the wall friction (such
as y + ). Moreover, it is believed that the correct near-wall representation of the
Reynolds stresses is important in representing unsteady near-wall turbulence.
When the frequency of the unsteadiness increases and approaches the tur-
bulence characteristic frequency, the turbulence becomes more directly affected
by the fluctuating mean flow. That means that non-equilibrium effects become
more important. The most extreme kind of unsteady turbulence is rapid dis-
torted turbulence where the turbulence is far from the equilibrium state. Turbu-
lence models that are well behaved in non-equilibrium turbulence may thus be
supposed to be suitable for unsteady turbulent flows. This has also been briefly
discussed by Mankbadi & Mobark[12].
One application of unsteady turbulent flows is the unsteady flow around tran-
sonic airfoils. The shock may induce boundary layer separation that may be un-
steady in some parameter regimes. In such flows it is crucial that the turbulence
model is able to predict separated flows accurately. This aspect is discussed by
Jin & Braza[13] who show that reasonable predictions are obtained if the model
is able to predict the boundary layer separation.
The following general requirements on a two-equation model for unsteady
wall-bounded flows can thus be identified: (i) No y + or log-law dependency, (ii)
correct near-wall asymptotic behavior, (iii) good prediction of non-equilibrium
turbulence and (iv) good prediction of boundary layer separation.
2.2. Choice of turbulence model. The self-consistent explicit algebraic Rey-

nolds stress model (EARSM) by Wallin & Johansson[14] based on K −ω matches
well the requirements above and has been found to be able to predict flow cases
where standard eddy-viscosity models fail. This model has been shown to give
results that in many aspects are similar to those of full Reynolds stress transport
models but for a computational effort that is comparable with standard two-
equation models; an issue of major importance, especially in three-dimensional
unsteady computations.
The EARSM is solved together with a low-Reynolds number K − ω model,
and has a near-wall damping function that gives the correct near-wall behavior
for all individual Reynolds stresses. The √ near-wall damping in the EARSM
formulation is expressed in terms of Rey ≡ Ky/ν while the K − ω model has
damping functions in terms of the turbulent Reynolds number, Ret ≡ K 2 /νε.
The proposed model is fully self-consistent, which means that the model rep-
resents an exact solution of the Reynolds stress transport model in the weak-
equilibrium limit where the advection and diffusion of the Reynolds stress anisot-
ropies are neglected. This gives a model that has the correct asymptotic behavior
for large strain rates. Wallin & Johansson[14] show that the model behaves rea-
sonable well even for the highly non-equilibrium homogeneous rapid sheared flow
where the initial SK/ε = 50.
The prediction of boundary layer separation is important for capturing the
unsteady flow around transonic airfoil. The eddy-viscosity assumption fails to
correctly predict the production of the turbulent kinetic energy which is over-
estimated in strong adverse pressure gradients and near separation or reattach-
ment points. Improvements can be obtained by an empirically based limiter on
the eddy viscosity or production, as in the Menter SST K − ω model[19]. A
similar effect is obtained from the proposed EARSM which in a natural way
decreases the turbulent shear stress in such flows. Transonic highly loaded wings
and supersonic impinging shocks were computed by Wallin & Johansson[14] and
they found that the shock position, separation length and skin friction was con-
siderably improved compared to standard eddy-viscosity models. The reduced
frequency for the case adopted in this study is, however, not very high and thus
accurate prediction of the shock position and boundary layer separation is the
most important requirement here.
To assess the results from EARSM together with the low-Reynolds number
K − ω model, the Wilcox eddy-viscosity low-Reynolds number K − ω model[20]
and the Baldwin & Lomax algebraic model[9] are applied for the study, and the
results are compared with experiment[6] and other computational results.
3. Numerical method
The 2D/3D structured multi-block, cell-centered finite-volume code Euranus[15]
is used for the study. The code solves the Euler or Navier–Stokes equations
for steady as well as unsteady flow problems, and has numerous options con-
cerning numerical schemes, convergence acceleration, and options for reacting
gas in equilibrium or non equilibrium. Algebraic as well as several two-equation
turbulence models are implemented in the code.
Euranus can be used for time-accurate computations with explicit Runge–
Kutta time marching or implicit time integration with pseudo-time relaxation[16,
17]. In this study, an implicit second-order backward-difference scheme for time
advancement is employed. This scheme yields a nonlinear system of equations
at each time step which is resolved by pseudo-time marching combined with
multigrid and local time stepping. The spatial discretization are treated with a
second-order upwind scheme with symmetric TVD and van Leer limiters[18].
4. Computational Results
We consider the 18% bicircular-arc airfoil in free space at M∞ = 0.76, α = 0
and Re = 11×106 . Experimental data[6] as well as previous computations are
available for comparison.
The baseline grid, generated by FFANET[21], is of C-type with 193 × 65
points and with 129 points on the airfoil. The first layer over the airfoil surface
is located at about unit y + . The outer boundary is about 25 chords away. A
close-up of the grid near the airfoil is shown in figure 1. The grid is symmetric
around the symmetric airfoil. A finer grid with 385 × 129 grid points is also used
for assessing grid convergence. The solution on the baseline grid was found to
be somewhat grid dependent.
4.1. Convergence Investigation. The reason for introducing fully implicit

temporal discretizations is that these are expected to allow significantly larger
time steps than the ones dictated by explicit scheme’s stability conditions. The
time step can then be chosen solely according to the time scale of the physical
problem. The use of dual time stepping introduces, however, the complication
of deciding how many multigrid cycles that are needed in the subiterations each
time step. It is reasonable to expect a trade-off; longer time steps advances the
computation faster but will need more iterations each time step. We study these
issues below, concentrating on following questions.
• How small time steps are needed for this problem?
• How many multigrid iterations are needed in the subiterations for each
time step?
• Does it pay off to increase the time step?
For all convergence investigations results reported here, we used the 193 × 65
mesh and the Wallin & Johansson EARSM[14].
4.1.1. Effects of changing time step and convergence criterion. For two different
time steps, corresponding to 165 and 350 time steps per period, respectively,
figure 2 shows the lift coefficient Cl as a function of reduced time t∗ = t/(c/U∞ );
Figure 1. Close up of the baseline computational grid.
0.5
350 time steps per period

Cl
0
165 time steps per period
−0.5
0 10 20 30 40 50 60 70 80
t*
Figure 2. Effect of changing the time step size on the lift coefficient.
c is the airfoil chord and U∞ the free-stream speed. Note that the longer time step
lowers the buffet frequency somewhat and introduces a low-frequency modulation
of the amplitude. Thus, at least 350 time steps per period seems necessary for
an accurate prediction of the lift.
Figure 3 shows the effect on Cl versus t∗ when relaxing the convergence cri-
terion for the subiterations; instead of requiring one order-of-magnitude residual
reduction, we prescribe a small, fixed number (10) of iterations. The latter leads
to a residual reduction which is much less than one order of magnitude and a
0.5
Cl
0 1 order reduction(~60 subiter. )
<<1 order reduction(10 subiter.)
−0.5
0 2 4 6 8 10 12 14 16
t*
Figure 3. Effect of the convergence criterion, one order and less

than one order reduction on the lift coefficient.
0.5
1 0rder reduction(~60 subiter. )

2 order reduction(500 subiter. )
0.25
Cl
−0.25
−0.5
0 2 4 6 8
t*
Figure 4. Effect of the convergence criterion, one order and two

order reduction, on the lift coefficient.
reduced frequency for the buffet around k = 0.481, whereas the stricter conver-
gence criterion yields k = 0.493. The residual is here defined as the maximum of
the density residuals in the domain and the reduction of the residual is measured
within one time step. Figure 4 shows the effect on Cl versus t∗ when sharpen-
ing the convergence criterion for the subiterations from one order-of-magnitude
residual reduction to two. The reduced frequency using the sharpened conver-
gence criterion is still around k = 0.49. We conclude that one order-of-magnitude
residual reduction seems to be a sufficient for this case. The time step in figures 3
and 4 corresponds to about 350 time steps per period.
Regarding the last of the three questions above, we performed several runs
with different time steps and recorded the number of subiterations needed to
achieve the same residual reduction, one order of magnitude. Table 1 shows the
Table 1. Effect of changing the time step size while keeping one
order of magnitude residual reduction.
No. of ∆t per period 50 100 165 350

No. of iteration/∆t ∼ 260 ∼ 200 ∼ 120 ∼ 60
Iterations/period (103 ) ∼ 13 ∼ 20 ∼ 20 ∼ 21
result. We make the somewhat surprising observation that it does not pay off
to reduce the number of time step from 350 to 165 per period; the number of
subiterations doubles approximately. A small improvement in efficiency can be
noted for the larger time steps in table 1. However, these longer time steps yield
an inaccurate solution; see the remarks above on the time-step selection. Fixing
the number of subiterations to 500, we are able to achieve more than two orders-
of-magnitude residual reduction when using 350 time steps per period, whereas
less then two orders-of-magnitude is achieved when using 165 time steps per
period.
4.1.2. The main conclusion from the convergence investigation.

• The time step should be small enough to get a good solution. About 350
time steps per period seems appropriate for the grid and model used.
• Larger time step does not save time when using the same convergence
criterion (keeping a reasonable resolution in time).
• Using one or two order of magnitude residual reduction yield very similar
solutions and, thus, one order of magnitude reduction was found to be
sufficient in this case.
• Using a small (10) number of iterations per time step yields smaller reduced
frequency and slightly wavy Cl curves.
Note that the temporal resolution that is needed, 350 time steps per period,
is much higher than the smooth behavior of the Cl -versus-time curve would
suggest. Even if our interest only lies in a slowly-varying integrated quantity
such as Cl , this suggests that it seems important to capture accurately the time
history of the flow on a faster time scale. This differs from our experience with
unsteady Euler calculation of forced airfoil pitching at transonic speeds [17, 22],
where accurate results were obtained with quite large time steps, say 24 time
steps per period.
4.2. Turbulence modelling effects. To assess the performance of the Wallin–

Johansson EARSM model described above, we compare with results obtained
using the Wilcox K − ω [20] and the Baldwin–Lomax [9] model, with other com-
putations from the literature, and with experiments. Free-stream conditions are
used as initial conditions in the time accurate calculations. All of the computa-
tions are performed using the baseline grid 193 × 65 and ∼ 350 time steps per
period (but ∼ 490 time steps per period when using the Baldwin–Lomax model).
Table 2. Computed reduced frequency compared to experiment by McDevitt[6].
model EARSM K −ω B − L Exp.

k 0.493 0.441 0.461 ∼ 0.49
0.5
EARSM
k−omega
0.25
Cl
−0.25
−0.5
0 2 4 6 8 10
t*
Figure 5. Computed lift coefficient using EARSM and K − ω.
0.5
EARSM
B−L
0.25
Cl
−0.25
−0.5
0 2 4 6 8
t*
Figure 6. Computed lift coefficient using EARSM and Baldwin-Lomax.
4.2.1. Reduced frequency and the Cl evolution. Table 2 compares the experimen-
tal reduced frequency [6] with the ones obtained using the different turbulence
models. Rumsey [8] obtained k = 0.492 using the Spalart–Allmaras model and
a 165 × 65 mesh.
The lift coefficients as a function of reduced time for the different turbulence
models are given in figures 5 and 6.
We are thus able to predict accurately the reduced frequency only when using
the Wallin–Johansson EARSM. Both the Baldwin–Lomax and the Wilcox K − ω
model yield too low frequency. It should also be noted that the turbulence models
1.4
EARSM 193*65
k−omega 193*65
1.2 exp.
0.8
Xshock/c 0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
t/T
Figure 7. Computed shock position using EARSM and K − ω com-

pared to experiment[6].
1.4
EARSM 193*65
B_L 193*65
1.2 exp.
0.8
Xshock/c
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
t/T
Figure 8. Computed shock position using EARSM and Baldwin-

Lomax compared to experiment[6].
differ with respect to how long time it take to initiate the buffeting from the
initial, free-stream state. The time it takes to reach a periodic state corresponds
to about 0.5, 3, and 7.5 periods for the Baldwin–Lomax, the EARSM, and the
Wilcox K − ω models, respectively. The Wilcox K − ω model appears to give
too high turbulence levels for this application, damping the natural unsteadiness
and thus both the frequency and amplitude are underpredicted.
4.2.2. Instantaneous shock position. Figures 7 8 and 9 depicts the shock posi-
tion versus time in the experiments [6], in our calculations using the different
turbulence models, and the results from Rumsey [8] using the Spalart–Allmaras
turbulence model. In these pictures, the shock position at the airfoil is nondi-
mensionalized by the chord c, and the time by the period of oscillation T . In the
1.4
EARSM 193*65
Rumsey et al. 185*65
1.2 exp.
0.8
Xshock/c 0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
t/T
Figure 9. Computed shock position using EARSM compared to

computations by Rumsey[8] using Spalart-Allmaras and experi-
ment[6].
experiment, the time t/T = 0 is taken to be the time when the shock is at mid
chord. However, this is not a convenient definition for the computational results,
since some models do not predict motion of the shock forward of mid chord. In
the computational results of Rumsey[8], t/T = 0 is the time when the shock is
closest to mid chord. In our calculations, t/T = 0 is defined as the time when
Cl = 0 and ∂Cl/∂t > 0, since there are several time instants when the shock is
positioned at mid chord in our calculations.
The best prediction of the shock position is obtained when using the EARSM.
The Baldwin–Lomax model predicts the shock position well, whereas the Wilcox
K − ω model gives a poor prediction.
4.2.3. Instantaneous pressure distribution. Figure 10 shows instantaneous pres-

sure distributions on the upper airfoil surface at t/T = 0, 0.1, . . . , 0.9. Com-
putations using the EARSM on two different grid levels are compared to mea-
surements. Also here we define time t/T = 0 as the instant when Cl = 0 and Cl
is increasing. With respect to pressure distribution, the computational results
agree rather well with the experimental results upstream of the shock but the
pressure recovery in the separated region downstream of the shock is not well
predicted. Note that the weak secondary shock appearing in the experiment at
times t/T = 0.2 and 0.3 is not present in the computational results. The most
upstream computed shock position is at x/c = 0.42 while the shock is present
further upstream in the experiment, which is also clear from Figures 7, 8 and 9.
4.3. Grid convergence. The influence from grid refinement in studied from
computational results using the EARSM on two different grids; 385 × 129 and
193 × 65. Figure 11 shows the lift and drag coefficients, Cl and Cd , versus
reduced time. The reduced frequency for the finer and baseline grids differ
t/T=0. t/T=0.1 t/T=0.2 t/T=0.3

−1.6 −1.6 −1.6 −1.6
385*129 385*129
193*65 193*65
exp. exp.
−0.8 −0.8 −0.8 −0.8

Cp
Cp
0 0 0 0
0.8 0.8 0.8 0.8

0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1
x/c x/c x/c x/c
t/T=0.4 t/T=0.5 t/T=0.6 t/T=0.7

−1.6 −1.6 −1.6 −1.6
385*129 385*129
193*65 193*65
exp. exp.
−0.8 −0.8 −0.8 −0.8

Cp
Cp
0 0 0 0
0.8 0.8 0.8 0.8

0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1
x/c x/c x/c x/c
t/T=0.8 t/T=0.9
−1.6 −1.6
385*129
193*65
exp.
−0.8 −0.8
Cp
0 0
0.8 0.8
0 0.5 1 0 0.5 1
x/c x/c
Figure 10. Computed pressure distribution at different times using

EARSM on different grids compared to experiment[6].
only in the third digit; k = 0.497 for the finer grid and k = 0.493 for the
baseline grid. However, the difference in the amplitude of the lift oscillations is
large. Figure 12 shows that the prediction of the temporal evolution of the shock
position improves for the finer mesh. The amplitude of Cl is strongly correlated
to the time variation of the shock position and thus it may be reasonable to
believe that also the Cl amplitude is more accurate predicted on the finer grid.
0.5
385*129 Cl
193*65 Cd
385*129 Cd
193*65 Cd
0.25
Cl & Cd
0
−0.25
−0.5
0 2 4 6 8
t*
Figure 11. Computed lift and drag coefficients using EARSM on

different grids.
1.4
EARSM 385*129
EARSM 193*65
1.2 exp.
0.8
Xshock/c
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
t/T
Figure 12. Computed shock position using EARSM on different grids.
5. Conclusions
The implicit time integration procedure used in this study was found to be robust
and efficient. The overall computational time was not very sensitive to the size of
the time step within the range studied, 50–350 time steps per cycle. The reason
for this was that the number of pseudo-time relaxation steps could be reduced as
the size of the time step decreased, keeping the same reduction of the residual.
Small time steps were needed for accuracy reasons and at least 350 time steps
per cycle was recommended which is much higher than the smooth behavior of
the Cl -versus-time curve would suggest. Even if our interest only lies in a slowly-
varying integrated quantity such as Cl , this suggests that it seems important to
capture accurately the time history of the flow on a faster time scale.
The solution was, however, not very sensitive to the number of subiterations
as long as the residual was reduced at least one order of magnitude. Grid con-
vergence could not be demonstrated completely, The differences between the
baseline, 193x65, grid and a fine, 385x129, grid was small with respect to the
frequency, whereas the differences in shock position and the amplitude of the lift
oscillations were larger.
Three different turbulence models were studied and the influence of and
differences between the models were significant. The more physically com-
plete model, the explicit algebraic Reynolds stress model (EARSM) proposed by
Wallin & Johansson[14], significantly improved the computational results over
the Baldwin-Lomax[9] and Wilcox K − ω[20] models. Both the shock position
and the frequency of the periodic motion were predicted to a high accuracy using
the EARSM. The Baldwin-Lomax model gave a reasonable shock position but
underpredicted the frequency. The K − ω model failed in both these respects.
Compared with the experimental pressure distribution, EARSM still can not
correctly predict the flow recovery downstream of the shock induced separation,
especially when the shock is positioned ahead of mid chord. The inability to cor-
rectly predict the flow recovery downstream of separation seems to be a general
problem with turbulence models.
Four general requirements on a turbulence model for unsteady mean flows has
been identified: (i) No y + or log-law dependency, (ii) correct near-wall asymp-
totic behavior, (iii) good prediction of non-equilibrium turbulence, and (iv) good
prediction of boundary layer separation. The Wallin & Johansson EARSM was
chosen for this study since that model reasonable well fulfills these requirements.
It is, however, reasonable to believe that accurate prediction of boundary layer
separation is the most important quality for a turbulence model in this case.
Acknowledgment
This research was performed as a part of the UNSI project (Unsteady Vis-
cous Flow in the Context of Fluid–Structure Interaction), a collaboration be-
tween Alenia, BAe, CASA, Dasa–M (Coordinator), Dassault, DERA, DLR, FFA,
IMFT, IRPHE, NUMECA, ONERA, Saab, TUB, and UMIST. The project is
funded by the European Commission, DG XII, Brite/EuRam, under the IMT
initiative (Project Ref: BRPR-CT97-0583). Partial funding was also provided by
NUTEK, the Swedish National Board for Industrial and Technical Development.
References
[1] Grønland, T. A., Eliasson, P., and Nordström, J., “Accuracy of Unsteady
Transonic Flow Computations,” ICAS-98, 1998.
[2] Levy, Jr., L., “Experimental and Computational Steady and Unsteady Tran-
sonic Flow about a Thick Airfoil,” AIAA Journal, Vol. 16, No. 6, 1978, pp.
564–572.
References 159
[3] Steger, J., “Implicit Finite-Difference Simulation of Flow about Arbitrary

Two-Dimensional Geometries,” AIAA Journal, Vol. 16, No. 7, 1978, pp.679-
686.
[4] Seegmiller, H. L., Marvin, J. G., and Levy L. L., “Steady and Unsteady
Transonic Flow,” AIAA Journal, Vol. 16, No. 12, 1978, pp.1461-1471.
[5] Edwards, J., and Thomas, J., “Computational Methods for Unsteady Tran-
sonic Flows,” Unsteady Transonic Aerodynamics, AIAA Progress in Astronau-
tics and Aeronautics, Nixon, D. (ed.), Vol. 120, 1989, pp.211-261.
[6] McDevitt, J., “Supercritical Flow about a Thick Circular-Arc Airfoil,” NASA
TM-78549, 1979.
[7] “Unsteady airfoil flow.” In Computation and Comparison of Efficient Turbu-
lence Models for Aeronautics – European Research Project ETMA, pp. 535–572.
(ed. by A. Dervieux, M Braza & J.-P. Dussauge) Vieweg 1998. Notes on Nu-
merical Fluid Mechanics, Vol. 65.
[8] Rumsey, C. L., Sanetrik, M. D., Biedron, R. T., Melson, N. D. and Parlette,
E. B., “Efficiency and Accuracy of Time-Accurate Turbulent Navier-Stokes
Computations,” AIAA-95-1835, 1995.
[9] Baldwin, B., and Lomax, H., “Thin Layer Approximation and Algebraic
Model for Separated Turbulent Flow,” AIAA-78-257, 1978.
[10] Spalart, P., and Allmaras, S., “A One-Equation Turbulence Model for Aero-
dynamic Flows,” AIAA-92-0439, 1992.
[11] Fan, S., and Lakshminarayana, B., “Low-Reynolds-Number k − ε Model for
Unsteady Turbulent Boundary-Layer Flows,” AIAA Journal, Vol. 31, No. 10,
1993, pp. 1777-1784.
[12] Mankbadi, R. R., and Mobark, A., “Quasi-steady Turbulence Modeling of
Unsteady Flows,” Int. J. Heat and Fluid Flow, Vol. 12, 1991, pp. 122-129.
[13] Jin, G., and Braza, M., “Two-Equation Turbulence Model for Unsteady
Separated Flows around Airfoils,” AIAA Journal, Vol. 32, 1994, pp. 2316-2320.
[14] Wallin, S., and Johansson A. V., “An explicit algebraic Reynolds stress
model of incompressible and compressible flows,” Accepted for publication in J.
Fluid Mech. 1999.
[15] Rizzi, A., Eliasson, P., Lindblad, I., , Hirsch, C., Lacor, C., and Häuser, S.,
“The Engineering of Multiblock/Multigrid Software for Navier-Stokes Flows on
Structured Meshes,” Computers & Fluids, vol. 22, 1993, pp. 341-367.
[16] Jameson, A., “Time Dependent Calculations using Multigrid, with Appli-
cations to Unsteady Flows past Airfoils and Wings,” AIAA-91-1596, 1991.
[17] Eliasson, P., and Nordström, J., “The Development of an Unsteady Solver
for Moving Meshes,” FFA TN 1995-39, 1995.
[18] Lacor, C., Zhu, Z.W., and Hirsch, C., “A New Family of Limiters Within
the Multigrid/Multiblock Navier-Stokes code Euranus,” AIAA/DGLR 5th Int.
Aerospace Planes and Hypersonics Conf., Munich. 1993.
160 References
[19] Menter, F. R., “Zonal Two-Equation k − ω models for Aerodynamic Flow,”

AIAA-93-2906, 1993.
[20] Wilcox, D.C., “Simulation of transition with a two-equation turbulence
model,” AIAA Journal Vol. 32, 1994, pp. 247–255.
[21] Tysell, L, and Hedman, S. G., “Towards a general three-dimensional grid
generation system,” ICAS Paper 88-4.7.4. 1988.
[22] Berggren, M., “Geometric conservation for structured moving meshes,”
Technical Report TN 1998-28, FFA, the Aearonautical Research Institute of
Sweden, P. O. Box 11021, S-161 11 Bromma, Sweden, 1998.
Paper 5
5
DERIVATION AND INVESTIGATION OF A NEW EXPLICIT
ALGEBRAIC MODEL FOR THE PASSIVE SCALAR FLUX
Petra M. Wikström
Stefan Wallin
Arne V. Johansson
Abstract. An algebraic relation for the scalar flux, in terms of mean flow
quantities, is formed by applying an equilibrium condition in the trans-
port equations for the normalized scalar flux. This modelling approach is
analogous to explicit algebraic Reynolds stress modelling (EARSM) for the
Reynolds stress anisotropies. The assumption of negligible advection and
diffusion of the normalized passive scalar flux gives, in general, an implicit,
nonlinear set of algebraic equations. A method to solve this implicit re-
lation in a fully explicit form is proposed, where the nonlinearity in the
scalar-production-to-dissipation ratio is considered and solved. The non-
linearity, in the algebraic equations for the normalized scalar fluxes, may be
eliminated directly by using a nonlinear term in the model of the pressure
scalar-gradient correlation and the destruction and thus results in a much
simpler model for both two- and three-dimensional mean flows. The perfor-
mance of the present model is investigated in three different flow situations.
These are homogeneous shear flow with an imposed mean scalar gradient,
turbulent channel flow and the flow field downstream a heated cylinder.
The direct numerical simulation data are used to analyse the passive scalar
flux in the homogeneous shear and channel flow cases and experimental
data are used in the case of the heated cylinder wake. Sets of parameter
values giving very good predictions in all three cases are found.
1. Introduction
Proper modelling of the passive scalar flux is important in many engineering
applications. The passive scalar quantity may, for example, be temperature,
species concentrations in combustion flows, or pollutant in atmospheric or ocean
flows.
In analogy with the eddy viscosity concept for the Reynolds stresses, the
passive scalar flux, ui θ, is commonly modelled through a simple gradient diffusion
assumption,
νt ∂Θ
ui θ = − , (1)
Prt ∂xi
163
164 A NEW EXPLICIT ALGEBRAIC MODEL FOR THE PASSIVE SCALAR FLUX
where ui and θ represents the fluctuating parts of the velocity and scalar, respec-
tively. Here Θ is the mean scalar value, and νt is the turbulent (eddy) viscos-
ity. In a zero equation model the turbulent Prandtl number, Prt (or turbulent
Schmidt number) is assumed to be constant. In models without the assumption
of a constant turbulent Prandtl number, the scalar variance, θ2 , and the dissi-
pation rate of half the scalar variance, εθ , are usually needed to be solved for
in analogy with K − ε modelling. Still the eddy diffusivity approach, using a
scalar eddy diffusivity, is unable to predict realistic values of all components of
ui θ, since it predicts the scalar flux to be aligned with the mean scalar gradient.
A tensor eddy diffusivity proportional to the Reynolds stresses was introduced
by Daly & Harlow [1] giving the following model for the scalar fluxes,
∂Θ
ui θ = −cθ τθ ui uj , (2)
∂xj
where cθ is a model parameter and τθ is an appropriate time scale. In the case
that the mean scalar gradient exists only in the y direction, the ratio of the scalar
fluxes, uθ/vθ, is then given by the Reynolds stress ratio uv/v 2 . In the near-wall
region of wall-bounded shear flows θ correlates more strongly with u than with v
and (2) underpredicts the scalar flux ratio, see for example Kim & Moin [2] and
Abe & Suga [3]. The following model was therefore proposed by Abe et al. [3],
uk uj ∂Θ
ui θ = −cθ τθ ui uk , (3)
K ∂xj
which instead gives a scalar flux ratio approximately proportional to u2 /uv in

the near-wall region when the mean scalar gradient exists only in the y direction.
For this quadratic form much better predictions were obtained than with those
of (2), in wall bounded shear flows, see [3].
In transport-equation modelling, of the passive scalar flux, the transport equa-
tions for, Kθ ≡ θ2 /2, its destruction rate, εθ , and the passive scalar flux, ui θ, are
modelled. The former two (Kθ , εθ ) are only needed for the construction of the
decay time scale Kθ /εθ for the passive scalar. Information about this timescale
may be particularly important in situations when it differs significantly from
K/ε. The set of transport equations is solved together with transport equations
for the Reynolds stresses and the dissipation rate of the turbulent kinetic energy.
This level of modelling is supposed to capture most of the relevant phenomena
involved in engineering flows, but it is not very attractive in complicated ge-
ometries since it leads to 12 transport equations for the turbulence quantities in
three-dimensional mean flows.
There is a renewed interest in algebraic models which are obtained from the
transport equations using some equilibrium assumption. The most common
approach is the weak equilibrium assumption, where the advection and diffusion
√
of the normalized scalar flux ui θ/ KKθ is neglected, rather than the scalar
flux itself, see e.g. Adumitroaie et al. [4], Girimaji & Balachandar [5], Abe
PETRA M. WIKSTRÖM, STEFAN WALLIN AND ARNE V. JOHANSSON 165
et al. [6], Shih & Lumley [7] and Shih [8]. This is equivalent to that used
to formulate algebraic Reynolds stress models (ARSMs). A different approach
is taken by Shabany & Durbin [9] where the advection and diffusion of the
normalized dispersion tensor are neglected.
The equilibrium assumption for the normalized scalar flux results in an al-
gebraic relation for ui θ that is implicit and nonlinear in the scalar flux. The
nonlinearity forms a major obstacle for deriving a relation that is fully explicit
and self-consistent, i.e. an expression that fulfills the basic implicit algebraic
equation. Adumitroaie et al. [4] leave this nonlinearity to be implicitly solved
as a part of the solution procedure. Explicit forms of the models are, however,
attractive since this leads to decreased numerical problems and computational
efforts. Abe et al. [6] obtain a fully explicit expression by modelling the nonlin-
ear part in terms of known quantities and thus self-consistency is not fulfilled.
The model proposed by Girimaji & Balachandar [5] is both fully explicit and
self-consistent but is restricted to a special case, Rayleigh–Bernard convection
with vanishing mean flow gradients.
In the present paper a new explicit algebraic scalar flux model is presented
which is fully explicit and self-consistent. The approach is similar to that of
Girimaji & Balachandar [5] but extended to be valid for general flows. The
present modelling approach for the scalar fluxes, ui θ, is thus analogous to that
of an self-consistent explicit algebraic Reynolds stress model (EARSM) for the
Reynolds stresses, see Johansson & Wallin [10] and Girimaji [11]. An implicit
algebraic equation for ui θ is constructed as an algebraic approximation of the
modelled transport equation for the scalar fluxes. The implicit system of equa-
tions is inverted and the scalar production to dissipation ratio is determined to
obtain an explicit formulation. Both the dynamical time-scale, K/ε, and the
scalar time-scale, Kθ /εθ , are present in the formulation. For a complete model
of the scalar fluxes, models of the transport equations for K, Kθ , ε and εθ are
therefore to be included. This explicit algebraic scalar flux model (EASFM) for
ui θ is suitably used together with an EARSM for ui uj , such as that of Wallin &
Johansson [12] and [13].
In section 2 the concept of second-order moment closures for the scalar fluxes
is considered as well as the equilibrium assumption needed to obtain algebraic
forms. The solution of the algebraic relation for the normalized scalar fluxes,
derived in section 2, is obtained in section 3. A method to solve this implicit
relation in a fully explicit form is proposed where the nonlinearity in the scalar
production to dissipation ratio is considered and solved. The solution is exact
in two-dimensional mean flows. In three-dimensional mean flows the tensorial
form is exact, but an approximation for the scalar production to dissipation
ratio needs to be introduced for a general set of model parameters. However, the
nonlinearity, in the algebraic equations for the normalized scalar fluxes, may be
eliminated directly by using a nonlinear term in the model of the pressure scalar-
gradient correlation and the destruction. This special case, which results in a
much simpler model for both two- and three-dimensional mean flows, is proposed
as a particularly interesting and practically useful model. In section 4 a set of test
cases is described and in section 5 data from these are compared to the present
EASFM predictions. It is also investigated how appropriate the equilibrium
assumption is for each of the test cases. Some conclusions and summarizing
comments are given in section 6.
2. Scalar flux modelling

Transport-equation models for the Reynolds-stress tensor and passive scalar flux
vector contains substantially more of the underlying physics than standard two-
equation models based on, e.g., K − ε and Kθ − εθ , respectively. A considerable
part of this improvement can be retained also on algebraic approximations of
the Reynolds-stress and scalar flux transport equations. The modelling level for
the scalar fluxes should not be higher than that of the Reynolds stresses, since
the scalar transport predictions heavily rely on the velocity field description.
We first describe a second-order modelling approach for the passive scalar
fluxes analogous to Reynolds stress transport models for ui uj . The Reynolds
averaged transport equation for the mean scalar, Θ, for incompressible flows
reads

DΘ ∂ ∂Θ
= α − uj θ , (4)
Dt ∂xj ∂xj
where D/Dt(≡ ∂/∂t + Uj ∂/∂xj ) implies the substantial derivative and α is the
molecular diffusivity. The scalar flux vector, ui θ, appears as an unknown quan-
tity, which originates from the averaging of the nonlinear term in the transport
equation of the total scalar field. This is analogous to the Reynolds averaged
Navier-Stokes (RANS) equations. Also in analogy with the transport equations
for the Reynolds stresses, a transport equation of the scalar-flux vector can be
written in symbolic form as
Dui θ
− Di = Pθi + Πθi − εθi , (5)
Dt
where the production term is given by
∂Θ ∂Ui
Pθi = −ui uj − uj θ . (6)
∂xj ∂xj
The molecular and turbulent diffusion, Di , the pressure scalar-gradient correla-
tion, Πθi , and the destruction rate tensor, εθi , contain unknown and higher-order
correlations, and the system of equations is not closed. A scalar flux transport
(SFT) model is obtained by modelling these unknowns in terms of the gradi-
ents of the mean velocity and scalar fields, ∂Ui /∂xj and ∂Θ/∂xj , the Reynolds
stresses, ui uj , and the passive scalar fluxes, ui θ. Also, the turbulent kinetic en-
ergy K ≡ ui ui /2, the half-scalar variance Kθ ≡ θ2 /2, and their dissipation rates
ε and εθ , respectively, may be involved in this modelling.
An equivalent alternative to (5) is to form a transport equation for the nor-
malized scalar flux,
ui θ
ξi ≡ √ , (7)
KKθ
which reads

Dξi (ξ) 1 Pθ − ε θ PK − ε Pθi − εθi + Πθi
− Di = − ξi + + √ . (8)
Dt 2 Kθ K KKθ
This corresponds to the transport equation for the Reynolds stress anisotropy
tensor, defined as
ui uj 2
aij ≡ − δij , (9)
K 3
The transport equation for aij together with the equation for the turbulent ki-
netic energy represents an alternative to formulating an equation for the Reynolds
stress tensor.
The K and Kθ transport equations are given by
DK
− D(K) = PK − ε
Dt
DKθ
− D(θ) = Pθ − εθ , (10)
Dt
where the production terms are
∂Ui ∂Θ
PK = −ui uj Pθ = −ui θ . (11)
∂xj ∂xi
(ξ)
The term Di in (8) is the molecular and turbulent diffusion of ξi and reads

(ξ) Di 1 D(K) D(θ)
Di = √ − + ξi . (12)
KKθ 2 K Kθ
To close the system of equations for the normalized scalar flux, transport
equations for ε and εθ are needed. Moreover, the Reynolds stresses, ui uj need
to be modelled in a way such that the individual anisotropy components are
accounted for. This disqualifies eddy-viscosity models.
The modelled transport equation for the normalized scalar flux vector, ξi ,
may be written in the following symbolic form,
Trans (ξi ) = fiξ (amn , Smn , Ωmn , Θm , r) , (13)
where the strain- and rotation-rate tensors, normalized with the turbulent time
scale, are

1 K ∂Ui ∂Uj 1 K ∂Ui ∂Uj
Sij ≡ + Ωij ≡ − . (14)
2 ε ∂xj ∂xi 2 ε ∂xj ∂xi
The last argument in fiξ is the ratio of the scalar to dynamical time-scales,
Kθ /εθ
r≡ . (15)
K/ε
Transport terms, advection and diffusion, are included in Trans (ξi ) and the
redistribution terms are lumped into fiξ . In the modelling this is taken to be
a function of the stress anisotropy, Smn , Ωmn and the normalized mean scalar
gradient defined as

K K ∂Θ
Θi ≡ . (16)
ε Kθ ∂xi
2.1. The equilibrium assumption for the scalar flux. In nearly homoge-
neous steady flows the advection and diffusion of the nondimensional scalar flux
may be neglected, see e.g. Adumitroaie et al. [4], Abe et al. [6] and Girimaji
et al. [5]. This is in many engineering flows a reasonable approximation es-
pecially if the driving forces (the velocity and scalar gradients) are large. By
using the equilibrium assumption, for which the left-hand side of (8) is zero, the
following implicit algebraic equation system is obtained:

1 Pθ − ε θ PK − ε Pθi − εθi + Πθi
ξi + = √ . (17)
2 Kθ K KKθ
This is in direct analogy with the equilibrium assumption for the Reynolds
stresses where the advection and diffusion of the Reynolds stress anisotropy is ne-
glected, introduced by Rodi (1972, 1976). Equation (17) may also be rearranged
using (15) to

1 Pθ PK 1 Kθ
ξi −1+r −1 = (Pθi + Πθi − εθi ) (18)
2 εθ ε εθ K
The modelling needed is of the form

K
Πθi − εθi = εθ Fi (amn , ξm , Smn , Ωmn , Θm , r) , (19)
Kθ
where Fi is a nondimensional function of the six nondimensional arguments.
The equation (18) together with the modelling expression (19) can be seen as
a system of equations for the nondimensional scalar flux, ξi . Since the scalar
flux is included in the expression for the production of the scalar variance, Pθ ,
according to (11), the algebraic equation system has a scalar nonlinearity in ui θ
(or, equivalently, ξi ) which needs to be considered in a solution that is fully self-
consistent, i.e., that the ratio of scalar production and dissipation, Pθ /εθ , used
in the lhs of (17) is identical to that of (11).
A similar nonlinearity arises in the algebraic Reynolds stress equations that
needs to be considered for a fully self-consistent explicit algebraic Reynolds stress
model, see Johansson & Wallin [10] and Girimaji [11]. The self-consistent ap-
proach for the Reynolds stresses results in a polynomial equation for the produc-
tion to dissipation ratio, PK /ε, and is crucial in flows far from equilibrium and
ensures, e.g., that the solution has the correct asymptotic behaviour for large
strain rates.
The equation system (18) may in principal be solved by forming the most
general tensorial form for the scalar flux as a function of the nondimensional
mean scalar gradient, the normalized mean flow strain and rotation rate tensors,
as well as the Reynolds stress anisotropy, aij , i.e,
ξi = ξi (amn , Smn , Ωmn , Θm , r) . (20)
The next step would then be to solve for the coefficients of the linearly indepen-
dent vector groups. By using an explicit algebraic Reynolds stress model the
anisotropy may be written in terms of the strain and rotation rate tensors, Sij
and Ωij , and thus the scalar flux is then of the following form:
ξi = ξi (Smn , Ωmn , Θm , r) . (21)
Since the exact transport equation of ui θ only involves linear terms of Θm , it is

natural to restrict the relation to
ξi = −Dij (Smn , Ωmn , r) Θj . (22)
The dispersion tensor, Dij , also introduced by Shabany & Durbin [9], has no
symmetry properties and is not, in general, traceless, so the general tensor form of
Dij in terms of Sij and Ωij consists of 17 linearly independent tensor groups. Shih
& Lumley [7] and Shih [8] include also an 18th, sixth-order term, which actually
may be expressed in the other terms. Due to the complexity this approach is of
minor practical use in three-dimensional mean flows. In section 3 a solution to
the implicit algebraic ξi -relation given by (17)–(19) in a fully explicit form will
instead be derived.
An alternative to neglecting the advection and diffusion of the nondimensional
scalar flux was proposed by Shabany & Durbin [9]. They form a transport
equation for the dispersion tensor. The alternative equilibrium assumption is
then to neglect the advection and diffusion of the nondimensional dispersion
tensor. The two different approaches are similar in flows where the direction of
the mean scalar gradient varies only slowly in the mean flow direction. In cases
where the direction of the mean scalar gradient varies rapidly in the mean flow
direction, the nondimensional dispersion tensor may be nearly constant while the
direction of the nondimensional scalar flux varies with the mean scalar gradient.
Neglecting the advection and diffusion of the nondimensional dispersion tensor
may thus be more appropriate than neglecting the advection and diffusion of the
nondimensional scalar flux. The algebraic equation associated with the Shabany
& Durbin [9] equilibrium assumption is, however, not as conveniently solved in
three-dimensional flows since the general form with the 17 linearly independent
tensor groups cannot be avoided in that approach.
2.2. Modelling Πθi and εθi . The dissipation vector, εθi , and the pressure
scalar-gradient correlation, Πθi , need to be modelled in terms of the dependent
variables, according to (19). A model form that has been widely used, see, e.g,
Shabany & Durbin [9] and also Launder [16], can be written as
ε ∂Ui ∂Uj ∂Θ
Πθi − εθi = −cθ1 ui θ + cθ2 uj θ + cθ3 uj θ + cθ4 ui uj . (23)
K ∂xj ∂xi ∂xj
This is the most general linear form, conserving the superposition principle for
passive scalars. A time scale consisting of both the dynamical timescale, K/ε,
and the scalar timescale, Kθ /εθ , could be used, though, instead of K/ε in the
cθ1 term. More generally, one could allow the four model parameters to depend
on r. This would lead to a violation of the superposition principle for passive
scalars, but an analogous violation is made, e.g., when modelling the pressure-
strain rate tensor. A model for the pressure-strain rate tensor that is nonlinear
in the Reynolds stress tensor violates the superposition principle of Reynolds
stress spectrum tensors.
From a formal solution of the Poisson equation for the rapid pressure a linear
model for Πrθi ≡ (pr /ρ)(∂θ/∂xj ) may be derived. This model contains only
rapid terms involving mean velocity gradients, that is, the second and third
terms of (23), (see, e.g., Shih [8]). In (23) also a fourth term is added, which
is a composite model of dissipative and redistributive terms in the scalar flux
equation. To capture effects of mean scalar gradients, an alternative to the cθ4
term would be to include the nonlinear term −cθ5 (1/Kθ )uk θ(∂Θ/∂xk )ui θ, which

would give the following model, − cθ1 + cθ5 (K/εKθ )uk θ(∂Θ/∂xk ) (ε/k)ui θ, for
the sum of slow pressure scalar-gradient correlation and the destruction. (Also
this term gives a violation of the superposition principle). The mean scalar
gradient has been included in a nonlinear term in a model of the slow term, in
a similar manner by Craft & Launder [17]. The following model for Πθi − εθi ,

K ∂Θ ε ∂Ui
Πθi − εθi = − cθ1 + cθ5 uk θ ui θ + cθ2 uj θ
εKθ ∂xk K ∂xj
∂Uj ∂Θ
+ cθ3 uj θ + cθ4 ui uj , (24)
∂xi ∂xj
will here be used, where effects of the mean scalar gradients are to be captured
by the cθ4 or the cθ5 terms.
3. Solution of the implicit algebraic equation for ξi

The algebraic scalar-flux equation system (17) and (24) can be solved by direct
inversion, which gives the scalar fluxes on an explicit form. Adumitroaie et al.
[4] have proposed a solution method which is quite simple and compact also
for three-dimensional flows. That model is, however, not fully explicit since
the production to dissipation ratios, PK /ε and Pθ /εθ are left implicit. We will
follow the solution method proposed by Adumitroaie et al. [4] to obtain a formal
solution of the scalar fluxes in general three-dimensional mean flows, where Pθ /εθ
is left as an unknown.
The formal solution is then used to formulate an equation for Pθ /εθ . This
equation has a closed solution for general two-dimensional mean flows in which
the solution becomes fully explicit and self-consistent. Moreover, this solution
is a reasonable approximation for three-dimensional mean flows. This approach
follows the ideas proposed by Girimaji et al. [5] for the special case of Rayleigh–
Bernard convection where the mean flow shear vanishes. Here, these methods
will be extended to general mean flow fields. The associated difficulties will be
discussed and a fully explicit model will be proposed. It will be shown that
a drastic algebraic simplification, and full self-consistency, is obtained for the
parameter value choice of cθ5 = 1/2.
The nondimensional system of equations for the scalar flux (17) and (24) is
given by

2
Nθ ξi = −(1 − cθ4 ) aij + δij Θj − [(1 − cθ2 − cθ3 ) Sij + (1 − cθ2 + cθ3 ) Ωij ]ξj ,
3
(25)
where

1 Pθ 1 1
Nθ = G + − cθ5 = G − − cθ5 ξl Θl
r εθ 2 2

1 1 Pk
G = 2cθ1 − 1 − + .
2 r ε
For cθ5 = 1/2 the equation system (25) is nonlinear due to the term ξl Θl in Nθ
that multiplies ξi . Equation (25) may be rewritten as

2
Aij ξj = −c 4 aij + δij Θj , (26)
3
where the matrix Aij is given by
Aij = Nθ δij + cS Sij + cΩ Ωij , (27)
and
cS = 1 − cθ2 − cθ3 cΩ = 1 − cθ2 + cθ3 c4 = 1 − cθ4 . (28)
For cθ5 = 1/2 the formal solution of the system (26), where Nθ is not yet
determined reads

−1 2
ξi = −c4 Aij ajk + δjk Θk . (29)
3
It is obtained by using the inverse of the matrix Aij . One interesting observation
is that the influences from the mean velocity gradient through A−1 ij , the Reynolds
stress anisotropy, and the mean scalar gradient are tensorially separated.
3.1. Solution for three-dimensional flows. In the following, bold face de-
notes a second rank tensor, e.g. A ≡ Aij . The matrix A is conveniently inverted
analytically, as proposed by Adumitroaie et al. [4], with the aid of the Cayley–
Hamilton theorem. Any general matrix satisfies its own characteristic polynomial
equation, which for a 3×3 matrix reads
1
A3 − tr{A}A2 + tr{A}2 − tr{A2 } A − det(A)I = 0, (30)
2

where det(A) is the determinant of A, tr{} denotes the trace, A2 ij ≡ Aik Akj ,
and I is the identity matrix. By multiplying (30) with A−1 the inverse of A is
obtained as

2 tr{A} − tr{A } I − tr{A}A + A
1 2 2 2
−1
A = . (31)
det(A)
The inverse of A exists only if the determinant of A is nonzero. Using the
definition (27), the inverse of A may be written in terms of S and Ω as
2 1
−1 Nθ − 2 Q1 I − Nθ (cS S + cΩ Ω) + (cS S + cΩ Ω)2
A = , (32)
Nθ3 − 12 Nθ Q1 + 12 Q2
where
2 3
Q1 ≡ c2S II S + c2Ω II Ω Q2 ≡ c III S + 2cS c2Ω IV (33)
3 S
and the invariants of the mean flow gradients are
II S ≡ tr{S2 } II Ω ≡ tr{Ω2 } III S ≡ tr{S3 } IV ≡ tr{SΩ2 }. (34)
For arbitrary cθ5 , the quantity Nθ depends on the scalar-variance production

to dissipation ratio, Pθ /εθ , and may be determined by using some iteration pro-
cedure. The relation is, however, strongly nonlinear and the iteration procedure
is thus not guaranteed to converge towards the physical, correct root. Moreover,
the iteration procedure, for Nθ might influence the convergence rate, to a steady
state solution, and it is preferable to have a fully explicit relation.
A self-consistent solution is obtained if Nθ may be determined without making
any assumptions about the scalar-variance production to dissipation ratio, Pθ /εθ .
A self-consistent and explicit model is thus obtained by using the formal solution
for the scalar flux, (29), in the definition of Nθ to obtain an explicit relation for
Nθ . This is similar to the model proposed by Girimaji et al. [5] for the special
case of Rayleigh–Bernard convection.
Following these ideas, a polynomial equation for Nθ is obtained by inserting
(32) into the definition (26), that is

1 1 2
Nθ = G − − cθ5 ξl Θl = G + − cθ5 c4 A−1
ij a jk + δ jk Θk Θi . (35)
2 2 3
The equation for Nθ is in general a fourth-order polynomial equation, of which

the solution may be found in any mathematical textbook,
2Nθ4 − 2GNθ3 − (Q1 + R1 ) Nθ2 + (GQ1 + R2 + Q2 ) Nθ
1
−R3 − GQ2 + Q1 R1 = 0 (36)
2
where

1 2
R1 =2 − cθ5 c4 tr a + I Θ2 ,
2 3

1 2
R2 =2 − cθ5 c4 tr (cS S + cΩ Ω) a + I Θ2 , (37)
2 3

1 2 2
R3 =2 − cθ5 c4 tr (cS S + cΩ Ω) a + I Θ2
2 3

and Θ2 ij ≡ Θi Θj . The problem of determining the physical, correct root is,
however, not considered here. For two-dimensional mean flows the equation for
Nθ is cubic and has a reasonably simple closed solution. This solution may be
used as a first approximation for Nθ in three-dimensional mean flows.
3.2. Solution for two-dimensional flows. The tensor A and the anisotropy
tensor a are 3x3 matrices in general three-dimensional mean flows as well as in
two-dimensional (2D) mean flows. In the latter case, when the third direction
is the homogeneous one, the components Aα3 = A3α = aα3 = a3α = Θ3 = 0 for
α = 3. (The aα3 components may be nonzero in 2D mean flows only if the initial
conditions of aα3 are nonzero). The A33 and a33 components are nonzero, but
they do not influence any of the ξi components, which is obvious by inspecting
(26). The tensor A may thus be considered as a 2x2 matrix, denoted A2D ,
during the inversion of A. The inverse of the matrix A2D in two-dimensional
mean flows is somewhat simpler than for three-dimensional mean flows and reads
2D −1 2Nθ I2D − A2D
A = , (38)
det(A2D )
where the determinant is given by
1
det(A2D ) = Nθ2 − Q1 . (39)
2
The 3x3 form of the inverse of A defined by (27) may now be written in terms
of S and Ω as

−1 Nθ I − (cS S + cΩ Ω) Nθ 1
A = − − I − I2D . (40)
det(A2D ) det(A2D ) Nθ
The last term in (40) contributes only to the A−1
33 term and may be dropped in
the expression for the scalar flux according to the discussion earlier.
The equation for Nθ is obtained from the relation (35) or from the equation
for Nθ in three-dimensional flows, (36), by considering that Q2 = 0 and R3 =
Q1 R1 /2 in two-dimensional flows. The equation for Nθ in two-dimensional flows

is then given by
2Nθ3 − 2GNθ2 − (Q1 + R1 )Nθ + GQ1 + R2 = 0. (41)
This equation has three roots and at least one of them is real.
The physical, correct root is identified by looking at the special case when the
mean scalar gradient, Θi , is zero. A local, or algebraic, model should then predict
a scalar flux, ξi = 0, which can be seen from (26). Θi = 0 gives R1 = R2 = 0
and the equation for Nθ can be written as
det(A2D ) (Nθ − G) = 0 (42)
with the aid of (39). This equation has three roots. The root Nθ = G is the
physical, correct root since the other two roots, det(A2D ) = 0, implies that a
solution to (26) does not exist. It is also easily seen in (26) that Nθ = G is the
physical, correct root for zero mean scalar gradients. The solution to the cubic
equation is then formulated so that the correct root is obtained for the special
case of vanishing scalar gradient. The solution for general two-dimensional mean
flows reads
G
Nθ = +
 3 √ √ 1/3 √ √ 1/3

 sgn Pθ1 + Pθ2 |Pθ1+ Pθ2 | +sgn Pθ1 − Pθ2 |Pθ1 − Pθ2 | , Pθ2 ≥ 0


 2 1/6
2 Pθ1 − Pθ2 cos 13 arccos √ P2 θ1 , Pθ2 < 0, Pθ1 ≥ 0
Pθ1 −Pθ2


 2

 2 Pθ1 − Pθ2 1/6 cos − 13 arccos √ −P 2
θ1
+ π
3 , Pθ2 < 0, Pθ1 < 0
Pθ1 −Pθ2
(43)
where
G3 1 1 1
Pθ1 = − GQ1 + GR1 − R2
27 6 12 4
2 3
G 1 1
Pθ2 = Pθ1 −
2
+ Q1 + R1 . (44)
9 6 6
For arbitrary parameter values the solution is not continuous in Pθ1 and Pθ2
since there is a discontinuity when Pθ1 < 0 and Pθ2 changes sign. A parameter
choice leading to this discontinuity should be avoided. One possibility is to use
the parameter choice cθ5 = 1/2, for which this problem is eliminated since then
Nθ = G is the only solution for any value of the mean scalar gradient.
3.2.1. The determinant of A in 2D flows and realizability constraints. The ex-

plicit relation (29) becomes singular if the determinant of A is zero. In that case
no solution of the implicit relation (17) exists and it is reasonable to suspect
that there would be numerical problems associated with the solution of the cor-
responding transport model. If the determinant of A is zero for some parameter
regime, the solution is valid only outside of that regime. The determinant of A
may be written as
1 1 2
det(A2D ) = Nθ2 − Q1 = Nθ2 − c II S + c2Ω II Ω . (45)
2 2 S
The obvious problem is that the determinant consists of both positive and neg-
ative terms and it is not possible from this form to judge if the determinant
has the possibility to become zero. Since Nθ is strongly coupled to the param-
eters II S and II Ω it is not obvious whether the determinant may become zero
for some set of parameters. One possibility would be to choose the coefficient
cS = 0, which gives a strictly positive determinant since II Ω < 0 by definition.
cS , however, can not be neglected for arbitrary choices of model constants (cθ2
and cθ3 ).
Let us study the cubic equation (41) which may be rewritten as
2det(A2D ) (Nθ − G) = R1 Nθ − R2 , (46)
where det(A2D ) is given by (45). The only possibility for the determinant to
become zero is if the rhs term, R1 Nθ − R2 , is zero. Since the physical correct
root then is Nθ = G, it is thus sufficient to investigate whether the determinant
can become zero for the special case Nθ = G. In that case the determinant is
2
r + 1 PK
4det(A ) = 4G − 2Q1 = 2cθ1 −
2D 2
+ − 2c2S II S − 2c2Ω II Ω . (47)
r ε
The determinant is strictly positive if

1 1
cθ1 > 1+ (48)
2 r
and
2
1 r + 1 PK II Ω
c2S < 2cθ1 − + − c2Ω . (49)
2II S r ε II S
PK /ε becomes negative only in rare conditions. This behaviour is not caught by
algebraic Reynolds stress models and thus not considered here. The limitation
on the cθ1 coefficient could always be violated for sufficiently small time-scale
ratios, r. The only possibility to avoid this is by letting cθ1 depend on r such that
cθ1 is nearly constant for reasonable flows but that the asymptotic behaviour is
cθ1 ∼ 1/r for small time-scale ratios, r.
A suitable form is
r+1
cθ1 = cθ1 , (50)
r
where cθ1 is a constant. The limitation (48) then gives cθ1 > 1/2. For all
practical cases that approximately fulfill the equilibrium condition assumed here,
the time-scale ratio should be roughly constant and of order unity. Inclusion of
this time-scale dependence, though, could give better model predictions in flows
with moderate deviations from equilibrium. This will be illustrated later for the
case of a heated cylinder wake.
The limitation on the cS coefficient may not be that severe since all terms on
the rhs. of (49) are positive. For large strain rates, II S , the term (PK /ε)2 /II S
on the rhs. should asymptotically approach a positive constant when using an
anisotropy model with a correct asymptotic behaviour. The limitation on the
constant cS depends on the particular choice of the other model parameters as
well as on the particular choice of the Reynolds stress anisotropy model. For
instance, the explicit algebraic Reynolds stress model by Wallin & Johansson
[10], [12] and [13] does have the correct asymptotic behaviour for large strain
rates. For II Ω = 0 and II S → ∞ that model gives the following limitation:

4
cS < ≈ 0.52. (51)
15
In parallel mean flows II S = −II Ω , and (49) is then automatically satisfied if,
e.g., we choose cS = cΩ (and satisfy (48)). It is, however, possible to construct
extreme flow cases for which cS = 0 is the only possible choice that avoids the
singularity. One could thus also here consider some weak flow-field dependency
in cS .
The limitations obtained for the cθ1 and cS coefficients should also be consid-
ered when using an implicit algebraic model or a transport model for the scalar
flux to prevent odd numerical behaviour.
3.3. A Daly & Harlow-type model. The parameter choice cθ2 = 1 and
cθ3 = 0 implies that both cS and cΩ are zero. This special choice of parameters
gives a drastically simplified model. The last term in equation (25) then vanishes
and the model reduces to

1 − cθ4 2
ξi = − aij + δij Θj . (52)
Nθ 3
This simplified model is very similar to the model of Daly & Harlow [1], given in
(2). For both models the dispersion tensor, Dij , is proportional to the Reynolds-
stress tensor.
The equation for Nθ is in this case given by
2Nθ (Nθ − G) = R1 (53)
with the physical solution

1 1 2 1
Nθ = G + G + R1 (54)
2 4 2
valid for arbitrary three-dimensional mean flows. From (29) it is evident that
the restriction c4 > 0, that is cθ4 < 1, has to be made. Then (54) has only real
roots. For R1 = 0 the solution Nθ = G is the physical one, see (26) and (37).
Hence, no singularity will occur as long as the condition

1 1 1
cθ1 > 1+ ⇔ cθ1 > (55)
2 r 2
is satisfied.
There is actually some theoretical support for a value of cθ2 close to unity and
a value of cθ3 close to zero (see section 2.2). Basic mathematical constraints (see
e.g. Shih [8]) give cθ2 = 4/5 and cθ3 = −1/5. In the present case a fourth term is
also added in (24), which is a composite model of dissipative and redistributive
terms in the scalar flux equation.
3.4. A new simple, fully explicit model. Models where the dispersion ten-
sor is proportional to the Reynolds-stress tensor, such as the simplified model in
the previous section, have not proven to be successful. It was pointed out in the
introduction that such models fail in e.g. the near-wall region of wall bounded
flows.
The general solution could also be cumbersome due to the complex algebra
even i two-dimensional mean flows, and it is not obvious how to treat fully
three-dimensional mean flows.
For the special choice
1
cθ5 = (56)
2
the simple solution
Nθ = G, (57)
valid for arbitrary three-dimensional mean flows is obtained. Hence, no equation
for Nθ needs to be solved, and a fully explicit, self-consistent model
2 1 2
G − 2 Q1 I − G (cS S + cΩ Ω) + (cS S + cΩ Ω)
A−1 = (58)
G3 − 12 GQ1 + 12 Q2
to be inserted in (29), is automatically obtained. The simplicity is here retained
for any choice of the model parameters cθ1 -cθ3 , and thus the dispersion tensor is
not necessary directly proportional to the Reynolds-stress tensor. The calibra-
tion of the model parameters will be considered later. A restriction is also here
that they should be chosen such that the denominator of (58) is guaranteed to
be nonzero.
4. The test cases

4.1. Homogeneous shear flow. We will here use data from a direct numeri-
cal simulation (DNS) of homogeneous shear flow with mean scalar gradients in
each of the three orthogonal directions made by Rogers et al. [18]. The simu-
lations were performed with 128×128×128 spectral modes and a mesh spacing
of (∆x, ∆y, ∆z)=(0.07792, 0.03896, 0.03896). In the comparisons, in section 5,
data from the simulations with a Prandtl number 0.7 and a turbulence Reynolds
number, 4K 2 /νε, of 847 are used. The shear rate, S, will be used to nondimen-
sionalize the time.
4.2. Turbulent channel flow. A DNS of a turbulent channel flow with a
passive scalar and a Prandtl number of 0.71 has been performed by Wikström
& Johansson [19]. The Reynolds number based on the centerline mean velocity
and the channel half-width, δ, is 5000, and the Reynolds number based on the
wall friction velocity, uτ , and the channel half-width is 265. The simulation
code uses spectral methods, with Fourier representation in the streamwise (x)
and spanwise (z) directions and Chebyshev polynomials in the wall-normal (y)
direction. The computational domain is 12.56δ, 2δ, and 5.5δ in the streamwise,
wall-normal, and spanwise directions, respectively, and the number of spectral
modes is 256×193×192. This gives a resolution in the x-, y and z directions of
13.0, 2.7(on average), and 7.6 wall units, respectively. The boundary conditions
used for the scalar field, with δ = 1, are
Θ (x, −1, z, t) = −1, Θ (x, 1, z, t) = 1 (59)
where Θ is the dimensionless total scalar field, Θ + θ. The scalar value, e.g.,
temperature, at each wall is thus kept constant (i.e., the scalar fluctuations on
the walls are zero) with a higher temperature of the upper wall. This boundary
condition, which represents a case where the passive scalar is introduced at the
upper wall (y = 1) and removed from the lower wall (y = −1), results in an
antisymmetric mean scalar profile in the channel. The DNS data presented in
the following consist of averages of 24 statistically independent fields.
4.3. The heated cylinder wake. Measurements in the self-similar region of a
heated cylinder wake has been made by Wikström et al. [20]. The experiments
were performed in the MTL wind tunnel at KTH, Stockholm, which has a 7.0 m
long test section of 1.2×0.8 m2 cross section and a free stream turbulence level
less than 0.05%. The diameter of the wake-generating cylinder was 6.4 mm and
all the measurements were made at a velocity, U0 , of 10.1 m/s giving a maximum
mean velocity deficit of 0.5 m/s at x/d = 400. The present Reynolds number,
U0 d/ν = 4300, is about three times higher than that of Browne & Antonia
[21]. The cylinder was electrically heated, giving a maximum mean temperature
excess, Θs , of 0.8 ◦ C above the ambient air temperature at x/d = 400.
Simultaneous measurements of velocity and temperature statistics were made
using a three-wire probe configuration consisting of an X-probe for velocity mea-
surements and a single cold wire for temperature detection, located 0.5 mm in
front of the X-wire mid-point. The hot wires had a length of 0.5 mm and a
diameter of 2.5 µm. The corresponding dimensions for the cold wire were 1.0
mm and 0.63 µm. Voltages from the constant temperature and constant current
circuits were filtered at 5 kHz and sampled at 10 kHz.
Cross-stream derivatives of measured quantities were obtained by using cubic-
spline smoothing. Derivatives in the streamwise direction, needed to determine
Table 1. The different combinations of the parameter values in the

model of Πθi − εθi given by (24).
model cθ1 cθ2 cθ3 cθ4 cθ5

(a) 3.2 0 0 0 0
(b) 3.2 0.5 0 0 0
(c) 3.2 1 0 0 0
(d) 3.2 0.5 0.5 0 0
(e) 2.5 0 0 0.35 0
(ar ) 1.2 (r + 1) /r 0 0 0 0
(br ) 1.2 (r + 1) /r 0.5 0 0 0
(er ) 1.0 (r + 1) /r 0 0 0.35 0
(aWWJ ) 1.6 (r + 1) /r 0 0 0 0.5
advective terms, were obtained by assuming self-similarity. It was demonstrated

that this assumption was valid at x/d = 400. The thermal dissipation, εθ , was
determined from the transport equation of Kθ and ε was determined from the
transport of K by neglecting pressure diffusion.
The results presented below are obtained from data at x/d = 400, where the
turbulent Reynolds number, 4K 2 /νε, is about 3200 at the location of maximum
production.
5. Model comparison
In the following, model predictions with the present explicit algebraic scalar
flux model (EASFM), for different choices of the set of parameter values, are
compared to data for the three different flow situations described in the previous
section. In this comparison the Reynolds stresses are taken directly from DNS
or experimental data and not obtained from any model. The different model
parameters used are presented in Table 1. Model (a) represents a model for
Πθi − εθi aligned with the scalar flux vector and model (b) is that of Launder
[16]. Model (c) is the Daly–Harlow-type model, with cθ1 =3.2 and cθ4 =0 as in the
two previous cases, and in model (d) cS = 0, i.e., singularities are guaranteed not
to occur for any flow situation. For model (e) cθ1 =2.5, cθ2 =cθ3 =0, cθ4 =0.35 and
the coefficients of this model are tuned for a good performance in the channel
flow case.
In the first four models the cθ1 parameter value is the same (that of Launder
[16]) to be able to compare the effects of the cθ2 and cθ3 terms. In models (ar ),
(br ), (er ) and (aWWJ ) the time-scale dependence on the cθ1 -term, discussed in
section 3.2.1, is taken into account. Apart from this modification, models (ar ),
(br ), and (er ) are equivalent to models (a), (b), and (e) respectively.
In the last case, model (aWWJ ), the mean scalar-gradient correction of the cθ1
term is included. For cθ5 = 1/2, the nonlinear terms in (25) vanishes. This very
attractive parameter choice is adopted here and cθ2 –cθ4 are set to zero.
Note that for all of the last four models, cθ1 > 1/2, hence, satisfying the
condition (48), necessary to avoid singular behaviour for small time-scale ratios.
Also, comparisons with the models of Rogers et al. [22] and Abe et al. [6] are
made. The model of Rogers et al. [22] is given by
ε ∂Θ ∂Ui
CD ui θ = −ui uj − uj θ , (60)
2K ∂xj ∂xj
where
0.25 −2.08
130 12.5 4K 2
CD = 18 1 + 1+ , ReT = (61)
PrReT ReT 0.48 εν
is an empirically determined function calibrated against the homogeneous shear
flow data by Rogers et al [18], [22]. This model is an implicit algebraic ap-
proximation of the ui θ transport equation. The right-hand side of (60) is the
production term Pθi .
5.1. Homogeneous shear flow. In Table 2 the models for Πθi − εθi are com-
pared to the DNS data of homogeneous shear flow, described in the previous
section. In Case 1 the mean scalar gradient is imposed in the streamwise (x)
direction and in Case 2 it is imposed in the cross-stream (y) direction. The best
predictions are given by models (a), (ar ), and (aWWJ ). Models (e) and (er ) also
give reasonably good predictions. Models (b) and (br ), for which cθ2 = 0.5,
somewhat overpredict the amplitude of the Πθ1 − εθ1 -values, and for the Daly–
Harlow-type model, (c), for which cθ2 = 1, this tendency is even worse. Model
(d) gives severe overpredictions of the Πθ2 − εθ2 -values.
In the simulation of Rogers et al. [18] the normalized scalar flux becomes
approximately constant for large simulation times. The equilibrium assumption
is thus quite appropriate in this flow case. In Table 3 the predictions of the
scalar flux ratio uθ/vθ are given. Good predictions are given by models (a),
(ar ), (aWWJ ), (e), (er ), and the model of Rogers et al. [22].
Models (b) and (br ) somewhat underpredict the amplitude of the scalar-flux
ratios and for the Daly–Harlow-type model, (c), this is more accentuated. Model
(d) gives predictions that both qualitatively and quantitatively deviate severely
from the DNS data. The model of Abe et al. [6] gives a quite large overprediction
of the amplitude of the flux-ratio in Case 1.
Only the models that behave reasonable well in homogeneous shear flow will
be kept in the following. In Table 4 the predictions of models (a), (e), (ar ),
(er ), (aWWJ ), and the models of Rogers et al. [22], and Abe et al. [6], for
the individual flux components are given. In Case 3 the mean scalar gradient is
imposed in the spanwise (z) direction. The simple model (aWWJ ) gives a very
good compromise for both components of the scalar fluxes in all three cases. The
Table 2. The predictions of the components of Πθi − εθi of models

(a)-(aWWJ ) compared to the DNS data of Rogers et al. [18], case
C128U, Pr = 0.7, at St=12.
Case 1 Case 1 Case 2 Case 2

Πθ1 − εθ1 Πθ2 − εθ2 Πθ1 − εθ1 Πθ2 − εθ2
DNS data 33.1 −6.90 −15.9 7.14
(a) 39.1 −7.31 −15.3 5.84
(b) 45.6 −7.31 −20.4 5.84
(c) 52.0 −7.31 −25.6 5.84
(d) 45.6 −41.4 −20.4 19.1
(e) 30.6 −5.71 −11.9 4.57
(ar ) 42.0 −7.85 −15.0 5.74
(br ) 48.5 −7.85 −20.2 5.74
(er ) 45.6 −9.52 −15.6 7.87
(aWWJ ) 38.6 −7.21 −14.7 5.62
Table 3. The predictions of the scalar flux ratio of models (a)-

(aWWJ ), the model of Rogers et al. [22], and the model of Abe et al.
[6], compared to the DNS data of Rogers et al. [18], case C128U,
Pr = 0.7, at St=12.
uθ/vθ, Case 1 uθ/vθ, Case 2

DNS data −5.37 −2.57
(a) −4.96 −2.33
(b) −4.27 −1.65
(c) −3.53 −0.97
(d) 2.30 −51.6
(e) −5.35 −2.73
(ar ) −4.90 −2.35
(br ) −4.24 −1.66
(er ) −5.21 −2.68
(aWWJ ) −4.92 −2.45
Rogers et al. −4.89 −2.33
Abe et al. −6.92 −2.17
model predictions of (aWWJ ) and those of the model by Rogers et al. [22] attain
approximately the same values in all cases.
The model of Abe et al. [6] underpredicts the amplitudes of all components in
all cases, whereas the predictions for vθ in Case 2 and wθ in Case 3 are somewhat
better. In the investigation of the model performance by Abe et al. [6] only
information of vθ in the cross-stream direction (i.e., two-direction) was needed.
Table 4. The predictions of the scalar fluxes of models (a), (b), (e),
(ar )-(aWWJ ), the model of Rogers et al. [22], and the model of Abe
et al. [6], compared to the DNS data of Rogers et al. [18], case C128U
at St=12.
Case 1 Case 2 Case 3

uθ vθ uθ vθ wθ
DNS data −2.41 0.45 0.94 −0.36 −0.67
(a) −2.13 0.43 0.99 −0.42 −0.41
(b) −1.90 0.44 0.67 −0.42 −0.41
(e) −1.90 0.36 0.97 −0.35 −0.35
(ar ) −2.02 0.41 1.01 −0.43 −0.37
(br ) −1.80 0.42 0.71 −0.43 −0.37
(er ) −1.71 0.33 0.92 −0.34 −0.30
(aWWJ ) −2.05 0.42 1.13 −0.46 −0.70
Rogers et al. −2.00 0.41 0.98 −0.42 −0.71
Abe et al. −1.12 0.16 0.53 −0.24 −0.42
0.3
1 (a) (b)
0.2
0.5
0.1
0 0
−0.1
−0.5
−0.2
−1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

y y
Figure 1. Validity of the equilibrium assumption, in the chan-
nel flow. (a): ξ1 : - - -, −ξ1 (Pθ /(2Kθ ) + PK /(2K)); - · -
√ √
, ξ1 (εθ /(2Kθ ) + ε/(2K)); ·, Pθ1 / KKθ ; —, (Πθ1 − εθ1 ) / KKθ ;
◦, the sum. (b): ξ2 : - - -, −ξ2 (Pθ /(2Kθ ) + PK /(2K)); - · -,
√ √
ξ2 (εθ /(2Kθ ) + ε/(2K)); ·, Pθ2 / KKθ ; —, (Πθ2 − εθ2 ) / KKθ ; ◦,
the sum
5.2. The turbulent channel flow. In Fig. 1 the validity of the equilibrium as-
sumption, in the present channel flow, is investigated. According to this assump-
(ξ)
tion the sum of all the terms on the right hand side of (8), that is, Dξi /Dt − Di ,
should vanish, i.e., be negligible in comparison with characteristic magnitudes of
individual terms on the right-hand side. Since the advective terms, Dξi /Dt, are
zero in the channel flow, the sum of all the terms on the right-hand side of (8)
(ξ)
equals −Di . For the ξ1 -component the equilibrium assumption is appropriate
0.25 0.15
(a) (b)
0.2 0.1
0.15 0.05
0.1 0
0.05 −0.05
0 −0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y
Figure 2. Model predictions of Πθi − εθi in the channel flow, using
the parameter combinations given in Table 1. (a): —, (a); - - -, (b);
- · -, (e); ∗, Πθ1 −εθ1 from DNS. (b): —, (a), (b); - · -, (e); ◦, Πθ2 −εθ2
from DNS.
0.25 0.15
(a) (b)
0.2 0.1
0.15 0.05
0.1 0
0.05 −0.05
0 −0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y
Figure 3. Model predictions of Πθi − εθi in the channel flow, using
the parameter combinations given in Table 1. (a): —, (ar ); - - -, (br );
- · -, (er ); · · · , (aWWJ ); ∗, Πθ1 − εθ1 from DNS. (b): —, (ar ), (br );
- · -, (er ); · · · , (aWWJ ); ◦, Πθ2 − εθ2 from DNS.
except near the wall. This is also the case for the ξ2 -component except in the
(ξ)
center of the channel where D2 may be said to be non-negligible.
In Figs. 2 and 3 the two components of Πθi − εθi are shown together with the
model predictions of (24) using the different parameter choices given in Table 1
except for (c) and (d). All terms are normalized by α (∂U/∂y)wall (∂Θ/∂y)wall ,
where α is the molecular diffusivity. Let us first consider models (a) and (b),
where the cθ1 parameter is taken to be 3.2. Since the cθ2 parameter only affects
the Πθ1 − εθ1 - component, in the present case, these two models give the same
predictions of the Πθ2 − εθ2 -component. Of these models, i.e., cθ1 = 3.2, the
aligned model, (a), gives the best agreement and the discrepancy in the Πθ1 −
εθ1 -component prediction increases as cθ2 increases. Models (a) and (e) give
(a) (b)
6 6
4 4
2 2
0 0
−2 −2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y
Figure 4. Model predictions of the scalar flux components, in the
channel flow, using the parameter combinations given in Table 1. (a):
—, (a); - - -, (b); - · -, (e). (b): —, (ar ); - - -, (br ); - · -, (er ); · · · ,
(aWWJ ). ∗, −uθ; ◦, vθ from DNS.
−2
0 0.2 0.4 0.6 0.8 1
y
Figure 5. Comparison of model (aWWJ ) and the model of Rogers

et al. [22], in the channel flow. —, Rogers et al.; · · · , (aWWJ ); ∗, −uθ
from DNS; ◦, vθ from DNS.
the best agreement for both components, but also the simple model (aWWJ )
gives encouragingly good results. All models slightly overpredict the Πθ2 − εθ2 -
component in the center of the channel.
It is evident that the model, given by (24), is unable to capture the behaviour
of Πθi − εθi near the wall. Since it is linear (without the cθ5 -term) and thereby
not realizable, there is a need for damping functions, in analogy with, e.g., the
EARSM by Wallin & Johansson [12], [13]. It could also be considered to add
wall-reflection terms to the model to obtain better predictions of Πθi − εθi near
the wall. From Fig. 3 it is seen that inclusion of the time-scale dependency in
the cθ1 -term slightly improves the prediction abilities of (24). The difference,
though, is quite small since the timescale-ratio variation is small in the present
channel flow.
Figure 4 shows the model predictions of the scalar fluxes, using the model
parameters given in Table 1 except for (c) and (d). The terms are normalized by
α (∂Θ/∂y)wall . Increasing the cθ2 parameter results in a decrease of the model
prediction of the amplitude of the uθ-component. An increase of cθ1 or cθ4
results in a decrease of the predictions of the amplitudes of both the scalar-flux
components.
Models (a), (ar ) and (aWWJ ) capture the behaviour of the scalar-flux com-
ponents quite well. This was also the case in the homogeneous shear flow. A
perfect prediction of Πθ2 − εθ2 for |y| ≤ 0.2 results in an underprediction of vθ in
(ξ)
this region. This is due to the fact that the equilibrium assumption, D2 =0, is
not quite fulfilled here. All the models discussed here give a much more realistic
description of the scalar fluxes than a standard eddy diffusivity model, which
predicts a zero uθ-component throughout the channel.
In Fig. 5 a comparison of the prediction of model (aWWJ ) and that of Rogers
et al. [22] is made. By decreasing the cθ1 value in model (aWWJ ) (and also
in model (ar )), the prediction of the scalar fluxes are very similar to that of
the model of Rogers et al. [22]. The cθ1 parameter is here tuned to give the
best performance in the homogeneous shear-flow case and with that value the
predictions of the scalar fluxes, in the channel-flow case, are very good in the
region where the equilibrium assumption is approximately satisfied.
5.3. The heated cylinder wake. In Fig. 6 the validity of the equilibrium
assumption, in the present wake flow, is investigated. For the ξ2 -component
the equilibrium assumption is quite appropriate except near the free-stream.
This is also the case for the ξ1 -component except in the center of the wake where
(ξ)
Dξ1 /Dt−D1 is of importance. The relative magnitude of the terms neglected in
the equilibrium assumption are somewhat larger in this case than in the channel
case. Applying the equilibrium assumption is thus slightly more appropriate in
the channel case. For η > 2 there is a large deviation from equilibrium, but since
the scalar fluxes approach zero for η > 2 this is actually of minor importance.
In Fig. 7 the two components of Πθi − εθi are shown together with the model
predictions of (24) using the parameter choices given in Table1 except for (c) and
(d). All terms are normalized by Us2 Θs /l, where Us and Θs are the maximum
mean velocity defect and mean temperature excess respectively and l is the mean
velocity defect half-width. Since the cθ2 parameter only affects the Πθ1 − εθ1 -
component, in the present case, models (a) and (b) give the same prediction
of the Πθ2 − εθ2 -component. The aligned model, (a), and model (e) give a
severe underprediction of the Πθ1 − εθ1 -component. The Launder model, (b)
gives an improvement compared to these. The importance of the rapid terms
when modelling Πθi − εθi in the heated cylinder wake has been investigated by
Wikström et al. 1998 [20].
15 15
(a) 10 (b)
10
5
5
0
0 −5
−10
−5
−15
−10
−20
−15 −25
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
η η
Figure 6. Validity of the equilibrium assumption, in the cylin-
der wake. (a): ξ1 : - - -, −ξ1 (Pθ /(2Kθ ) + PK /(2K)); - · -
√ √
, ξ1 (εθ /(2Kθ ) + ε/(2K)); ·, Pθ1 / KKθ ; —, (Πθ1 − εθ1 ) / KKθ ;
◦, the sum. (b): ξ2 : - - -, −ξ2 (Pθ /(2Kθ ) + PK /(2K)); - · -,
√ √
ξ2 (εθ /(2Kθ ) + ε/(2K)); ·, Pθ2 / KKθ ; —, (Πθ2 − εθ2 ) / KKθ ; ◦,
the sum
0.06 0.06
(a) (b)
0.04 0.04
0.02
0.02
0
0
−0.02
−0.02
−0.04
−0.04
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
η η
Figure 7. Model predictions of the components of Πθi − εθi , in the
cylinder wake, using the parameter combinations given in Table 1.
(a): —, (a); - - -, (b); - · -, (e). (b): —, (ar ); - - -, (br ); - · -, (er );
· · · , (aWWJ ). ∗, Πθ1 − εθ1 ; ◦, Πθ2 − εθ2 .
From Fig. 7, it is seen that inclusion of the time scale dependency in the
cθ1 -term improves the prediction abilities of (24), especially for the Πθ1 − εθ1 -
component. The effects of this inclusion is here larger than in the channel flow
case, since the time-scale-ratio variation is larger in the wake flow.
Figure 8 shows the model predictions of the scalar fluxes, using the model
parameters given in Table 1 except for (c) and (d). The terms are normalized
by Us Θs . The aligned model, (a), and model (e) give a severe overprediction of
the amplitude of the uθ-component, whereas they gave good predictions in both
the homogeneous shear-flow and the channel cases. One interesting observation
is that models (a) and (e) give approximately the same predictions, of the scalar
0.06 0.06
(a) (b)
0.04 0.04
0.02 0.02
0 0
−0.02 −0.02
−0.04 −0.04
−0.06 −0.06
−0.08 −0.08
−0.1 −0.1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
η η
Figure 8. Model predictions of the scalar flux components, in the
cylinder wake, using the parameter combinations given in Table 1.
(a): —, (a); - - -, (b); - · -, (e). (b): —, (ar ); - - -, (br ); - · -, (er );
· · · , (aWWJ ). ∗, uθ; ◦, vθ.
0.06
0.04
0.02
−0.02
−0.04
−0.06
−0.08
−0.1
0 0.5 1 1.5 2 2.5
η
Figure 9. Comparison of model (aWWJ ) and the model of Rogers

et al. [22], in the cylinder wake. —, Rogers et al.; · · · , (aWWJ ); ∗,
uθ; ◦, vθ.
fluxes, in both the channel flow and the cylinder wake. Increasing the cθ2 param-
eter results in a decrease of the model prediction of the amplitude of uθ. From
Fig. 8 it is also seen that inclusion of the time scale dependency in the cθ1 -term
significantly improves the prediction abilities of (24). The effect of this inclu-
sion on the scalar fluxes (and the sum of pressure scalar-gradient correlation and
destruction) is much larger here than in the channel-flow case. These types of
models all give a reasonably good agreement with the experimental data. Model
(br ), for which cθ2 = 0.5, gives an excellent agreement here. This was not the
case for the homogeneous shear flow and the channel flow.
Models (ar ) and (aWWJ ), though, give a very good agreement in all three test
cases. Since no equation for the scalar quantity Nθ has to be solved when using
the parameter choice cθ5 = 1/2, model (aWWJ ) appears particularly attractive
and can be recommended for a wide class of situations.
By using a model with the time-scale-ratio dependence of models (ar )-(aWWJ ),
numerical problems associated with extreme values of the time-scale ratio are
also avoided. As in the previous case, all models in Table 1 give a more realistic
description of the scalar fluxes than a standard eddy diffusivity model, which,
e.g., predicts a zero value of the uθ-component throughout the wake. The con-
sequences of applying the equilibrium assumption, in the present case, may be
seen in Figs. 7 and 8. Model (e) gives an excellent prediction of Πθ2 − εθ2 for
|η| ≤ 1.5, whereas vθ is overpredicted in this region. This is due to the fact that
(ξ)
the equilibrium assumption, Dξ1 /Dt−D1 =0, is not quite fulfilled in this region.
Since all mean scalar gradients are zero at the center line, a zero prediction of uθ
is obtained. To be able to give a nonzero prediction at the center line, nonlocal
effects have to be taken into account and the equilibrium assumption should not
be applied. In Fig. 6 it is seen that the transport terms are important in the
region near the center line.
In Fig. 9 a comparison of the prediction of model (aWWJ ) and that of Rogers
et al. [22] is made. As in the homogeneous shear flow, and in the channel flow,
the model predictions are very similar for each component of the scalar flux.
5.4. Comparison with the model of Rogers et al. When comparing the
(aWWJ ) model with that of Rogers et al. [22] there are several things that should
be pointed out.
First, the model of Rogers et al. is an implicit model, which for large Reynolds
numbers becomes a Daly & Harlow type of model and thus in this limit will
underpredict the scalar flux ratio. See Abe et al. [3] and model (c) in Table 3.
This model involves an empirically determined parameter, CD , a function of
the turbulence Reynolds number and the turbulence Peclet number, which is
calibrated against homogeneous shear flow DNS data. Model (aWWJ ) is, on
the other hand, an explicit model which involves a function of the time-scale
ratio, r. This function is obtained directly from the modeled equations to avoid
singularities.
Second, and perhaps the most important issue, is that the model of Rogers
et al., which is formed by considering homogeneous shear flows, and neglecting
the diffusion terms in the transport equation of the dimensional scalar flux. Also
the time evolution of the scalar flux, ∂ui θ/∂t, is assumed to be proportional to
the flux itself. This last term is then combined with the model of Πθi − εθi ,
and together they are modelled by Πθi − εθi − ∂ui θ/∂t = −CD (ε/2K)ui θ. For
general flows exactly the same model formulation is obtained by applying the
equilibrium assumption in the transport equation for the nondimensional scalar
√
flux, (8), adding and subtracting ∂ui θ/∂t / KKθ to the right-hand side of,
(8), using the model Πθi − εθi − ∂ui θ/∂t = −CD (ε/2K)ui θ, and finally assuming

∂ui θ 1 Pθ − ε θ PK − ε
− / KKθ = − ξi + . (62)
∂t 2 Kθ K
For statistically stationary (nonhomogeneous) flows exactly the same model for-
mulation is thus obtained by assuming

1 Pθ − ε θ PK − ε
− ξi + = 0, (63)
2 Kθ K
and using the model Πθi − εθi = −CD (ε/2K)ui θ, since ∂ui θ/∂t = 0. In the
present non homogeneous test cases, where diffusion terms are present, i.e, the
channel flow and the cylinder wake, (63) is approximately satisfied, which may
be seen in Figs.1 and6. This explains why there are no major differences between
the model predictions of the model of Rogers et al., (ar ), and (aWWJ ), which all
are based on models for Πθi − εθi that are aligned with the scalar flux, whereas
models (ar ) and (aWWJ ) do not include the assumption (63). (In the case of
(aWWJ ) the mean scalar gradient is invoked in the model of the slow term, see
(24)). Model (aWWJ ) is obtained by neglecting the advection and diffusion terms
in the transport equation of the nondimensional scalar flux and keeping the first
term on the right-hand side of (8), that is, not making the assumption (63), and
would in general do a much better job in nonhomogeneous flows where (63) is
not valid (and the equilibrium assumption is appropriate).
6. Conclusions
An algebraic relation for the scalar flux in terms of mean flow quantities has
been formed by applying an equilibrium assumption in the transport equation
for the scalar flux. The resulting set of algebraic equations is implicit and non-
linear unless the model parameter cθ5 = 1/2 is used in the model of the sum of
the pressure scalar-gradient correlation and the destruction. A solution to the
general implicit relation for the normalized scalar flux in a fully explicit form
is proposed. The solution is exact in two-dimensional mean flows. In three-
dimensional mean flows the tensorial form is exact but an approximation for the
scalar production to dissipation ratio is introduced.
The Daly–Harlow-type model, cθ2 = 1 and cθ3 = 0, can readily be transformed
into a formulation of a model for the dispersion tensor, Dij (see e.g. (22)), which
inserted into the implicit algebraic equation for ξi , yields
1 − cθ4 ui uj
Dij = , (64)
Nθ k
where Nθ is given by (54). Hence, for this simplified model we have a dispersion
tensor (in two-dimensional as well as in three-dimensional mean flows) that is
proportional to the Reynolds stress tensor. The dependency of the mean flow
gradients here enters through the scalar Nθ . For this model the distribution of
the relative dispersion is proportional to the relative turbulence intensities in
the different directions. This is analogous to the Daly & Harlow [1] model for
turbulent diffusion of the Reynolds stresses. As a comparison one may note that
the simplest eddy-diffusivity approach implicitly assumes an isotropic dispersion
tensor.
From the general model formulation, it is seen that for model parameters
that deviate from cθ2 = 1 and cθ3 = 0, a dispersion tensor with principal axes
different from those of the Reynolds stress tensor is obtained.
With the parameter choice cθ5 = 1/2, the equation system (25) for the scalar
fluxes becomes explicit without any need for solving an equation for Nθ . This is
the case for both two- and three-dimensional mean flows. The parameter choice
cθ5 = 1/2 is thus very attractive. The present EASFM is then given by

−1 2
ξi = −(1 − cθ4 ) Aij ajk + δjk Θk , (65)
3
where the inverse of the matrix A is given by
2 1 2
G − 2 Q1 I − G (cS S + cΩ Ω) + (cS S + cΩ Ω)
A−1 = (66)
G3 − 12 GQ1 + 12 Q2
Predictions of the present EASFM have been compared to data for three different
types of flow situations. These are homogeneous shear flow, turbulent channel
flow, and the flow behind a heated cylinder. With the Daly–Harlow-type model,
i.e., cθ2 = 1 and cθ3 = 0, an underprediction of the uθ-component compared
to the vθ-component is obtained in all cases. Using model (aWWJ ) in Table 1
(given by (65) and (66) with cS = cΩ = 1 and cθ4 = 0), for which cθ5 = 1/2,
very good predictions are obtained for all test cases. Since no equation for the
scalar quantity Nθ has to be solved when using the parameter choice cθ5 = 1/2,
model (aWWJ ) is strongly recommended.
By using a model with the time-scale-ratio dependence of models (ar )-(aWWJ ),
numerical problems associated with extreme values of the time-scale ratio are also
avoided. In the wake flow where the largest variations of the time-scale ratio are
present, compared to the other two cases, the inclusion of this time-scale depen-
dence is shown to be of significant importance. This time-scale ratio dependence
in the cθ1 -parameter is necessary to obtain good predictions in the heated cylin-
der wake, using the model parameters that give good model predictions in the
present homogeneous shear and channel-flow cases.
In a comparison of the prediction of model (aWWJ ) and that of Rogers et al.
[22] it is found that the model predictions are very similar for each component in
all test cases of the scalar flux. Model (aWWJ ), which is obtained by neglecting
the advection and diffusion terms in the transport equation of the nondimen-
sional scalar flux and keeping the first term on the right-hand side of (8), that
is, not making the assumption (63), would in general do a much better job in
nonhomogeneous flows where (the equilibrium assumption is appropriate and)
(63) not is valid.
References 191
An explicit algebraic scalar flux model as proposed here ((65) and (66)) should
suitably be used together with an appropriate EARSM for the Reynolds stresses.
To make the model complete model equations for θ2 and εθ are needed to be
able to determine the time-scale ratio r.
References
[1] Daly, B. J. & Harlow, F. H. 1970 Transport equations in turbulence. Phys.
Fluids 13, 2634–2649.
[2] Kim, J., & Moin, P. 1989 Transport of passive scalars in a turbulent channel
flow. in proc. Turbulent Shear Flows 6, Toulouse Sept. 7-9 1987, Springer,
85–96
[3] Abe, K., & Suga, K. 1998 Large eddy simulation of passive scalar fields
under several strain conditions. in proc. Turbulent Heat Transfer II, Manch-
ester May 31 – June 5, 2, 8-15–8-30.
[4] Adumitroaie, V., Taulbee, D. B. & Givi, P. 1997 Explicit Algebraic Scalar-
Flux Models for Turbulent Reacting Flows. A. i. ch. e. J. 43, 1935–2147.
[5] Girimaji, S. S. & Balachandar, S. 1998 Analysis and modeling of boyancy-
generated turbulence using numerical data. Int. J. Heat Mass Transfer 41,
915–929.
[6] Abe, K., Kondoh, T. & Nagano, Y. 1996 A two-equation heat transfer model
reflecting second-moment closures for wall and free turbulent flows. Int. J.
Heat and Fluid Flow 17, 228–237.
[7] Shih, T. H. & Lumley, J. L. 1993 Remarks on Turbulent Constitutive Re-
lations. Mathl. Comput. Modelling 18(2), 9–16.
[8] Shih, T. H. 1996 Constitutive relations and realizability of single-point tur-
bulence closures. Turbulence and Transition Modelling, M. Hallbäck, D.S.
Henningson, A.V. Johansson & P.H. Alfredsson (eds), Vol 2, Kluwer.
[9] Shabany, Y. & Durbin, P. A. 1997 Explicit Algebraic Scalar Flux Approxi-
mation. AIAA J. 35, 985–989.
[10] Johansson, A. V. & Wallin, S. 1996 A new explicit algebraic Reynolds stress
model. Proc. Sixth European Turbulence Conference, Lausanne, July 1996,
Ed. P. Monkewitz, 31–34.
[11] Girimaji, S. S. 1995 Fully explicit and self-consistent algebraic Reynolds
stress model. Theor. Comp. Fluid Dyn. 8, 387.
[12] Wallin, S. & Johansson, A. V. 1996 A complete explicit algebraic Reynolds
stress turbulence model including an improved near-wall treatment. Proc.
Flow Modeling and Turbulence Measurements VI, Tallahassee FL, C.-J.
Chen, C. Shih, J. Lienau & R.J. Kung (eds), 399–406, Balkema, Rotterdam.
[13] Wallin, S. & Johansson, A. V. 2000. An explicit algebraic Reynolds stress
model for incompressible and compressible turbulent flows. To appear in J.
Fluid Mech. 403, 89–132.
192 References
a two equation model of turbulence. Ph.D. thesis, University of London.
stresses. Z. angew. Math. Mech. 56, T219–221.
[16] Launder, B. E. 1978 Heat and mass transport. In Turbulence, P. Bradshaw
(ed.), Springer, Berlin, 231–287.
[17] Craft, T. J., Launder, B. E. 1996 Recent developments in second-moment
closure for boyancy-affected flows. Dynamics of Atmospheres and Oceans
23, 99–114.
[18] Rogers, M.M., Moin, P. & Reynolds, W.C. 1986 The structure and modeling
of the hydrodynamic and passive scalar fields in homogeneous turbulent
shear flow. Dept. Mech. Engng. Rep. TF-25, Stanford University, Stanford,
California.
[19] Wikström, P. M. and Johansson, A. V. 1998 DNS and scalar-flux transport
modelling in a turbulent channel flow. in proc. Turbulent Heat Transfer II,
May 31-June 5, Manchester, 1998.
[20] Wikström, P. M., Hallbäck, M., and Johansson, A. V. 1998 Measurements
and heat-flux transport modelling in a heated cylinder wake. Intl. J. Heat
Fluid Flow 19, 556–562.
[21] Browne, L. W. B., & Antonia, R. A. 1986 Reynolds shear stress and heat
flux measurements in a cylinder wake. Phys. Fluids A. 29, 709–713.
[22] Rogers, M.M., Mansour, P. & Reynolds, W.C. 1989 An algebraic model for
the turbulent flux of a passive scalar. J. Fluid Mech. 203, 77–101.
Paper 6
6
AN EFFICIENT EXPLICIT ALGEBRAIC REYNOLDS STRESS k-ω
MODEL (EARSM) FOR AERONAUTICAL APPLICATIONS
Stefan Wallin
Abstract. The self-consistent explicit algebraic Reynolds stress model (EARSM)

proposed by Wallin & Johansson (1997) with the k-ω wo-equation transport
model as platform was the starting point for some further developments. The orig-
inal near-wall damping function was reformulated in terms of Re y ≡ k y ⁄ ν and
an additional alternative without any damping functions whatsoever was intro-
duced. The EARSM approach clearly improves the shock position for transonic
wings without significal increase in computational cost. The EARSM may be
expressed as an extension to an eddy-viscosity model and thus no major problems
should be expected in implementing EARSM into a flow solver with existing two-
equation eddy-viscosity models.
1. Introduction
Standard two-equation models based on the eddy-viscosity assumption are still domi-
nating in the context of industrial flow computations. In flows with strong effects of
streamline curvature, adverse pressure gradients, flow separation or system rotation,
such models fail to give accurate predictions. Turbulence models based on the transport
equations for the individual Reynolds stresses have the natural potential for dealing
with, e.g., the associated complex dynamics of intercomponent transfer. The eddy-vis-
cosity hypothesis may in this context be said to be replaced by transport equations for
the individual Reynolds stress anisotropies. As yet, there are non-trivial numerical
aspects of flow computations with such models in complex flow situations. This repre-
sents an active area of research. In parallel with such efforts there has been a considera-
ble renewed interest in various forms of explicit algebraic approximations of the
anisotropy transport equations, see e.g. Pope (1975), Gatski & Speziale (1993), Gir-
imaji (1995), Johansson & Wallin (1996) and Wallin & Johansson (1997).
The model proposed by Wallin & Johansson has been proven to give important
improvements over standard two-equation models based on the eddy-viscosity assump-
tion in flows with strong effects of streamline curvature, adverse pressure gradients,
flow separation or system rotation. The near-wall formulation in the model is, however,
based on y + which may cause problems in some situations.
In the present work the near-wall formulation in the model proposed by Wallin &
Johansson will be revised in terms of the wall distance which replaces the y + -depend-
ency. Moreover, a simplified version is proposed without near-wall damping functions
whatsoever which may be attractive for computation of complicated 3D flows.
195
196 AN EFFICIENT k-ω EARSM FOR AERONAUTICAL APPLICATIONS
2. Description of the models

The turbulence models that will be described here are variations of the explicit alge-
braic Reynolds stress model (EARSM) described in detail by Wallin & Johansson
(1997). That model is based on a recalibrated Launder, Reece & Rodi (1975) model
which is a differential Reynolds stress model. A self-consistent and fully explicit alge-
braic relation for the Reynolds stresses in terms of the mean flow field is then obtained
by applying the so-called ‘equilibrium assumption’, Rodi (1972 & 1976), where the
advection and diffusion of the Reynolds stress anisotropy is neglected.
The model is based on transport equations for the turbulent kinetic energy k and the
specific dissipation rate ω and we are recommending two options: (i) EARSM based
on the Wilcox (1988) standard k – ω model denoted ‘WJ_skw’ and (ii) EARSM based
on the Wilcox (1994) low-Reynolds number k – ω model denoted ‘WJ_kw’. There is
actually a third option that is an eddy-viscosity model obtained from further simplifica-
tion of (i). That is (iii) Linear EARSM based on the Wilcox (1988) standard k – ω
model denoted ‘LWJ_kw’.
A well known problem with these options is that the k – ω model is unphysically
sensitive to the free-stream turbulence level. That has recently been resolved by Kok
(1999) by adding a cross-diffusion term in the ω -equation. An interesting, but not thor-
oughly tested alternative is thus EARSM based on Kok k – ω , alternative (iv).
Which of these options that should be recommended depends on the application. The
following may be used as a guide:
(i) EARSM based on standard k – ω is more simple to implement and needs no
wall-distance information. The turbulent kinetic energy and the Reynolds stress profiles
are qualitatively wrong in the near-wall region, y + < 100 , but the velocity profiles and
skin friction are correct. This alternative is thus most suitable as an engineering tool for
complicated 3D flows where detailed information of the near-wall turbulence is not
needed.
(ii) EARSM based on low-Reynolds number k – ω is somewhat more complex and
needs the distance to the closest wall. The model is, however, independent of the local
wall skin friction or y + . The influence by the near-wall damping functions are limited
to the very near-wall region and are practically zero for y + > 100 . For this reason, it
should not be an important issue how multiple walls and corners are treated. The turbu-
lent kinetic energy and the Reynolds stress profiles are well predicted also in the near-
wall region. This alternative should thus be used if detailed information of the near-wall
turbulence is needed.
(iii) Linear EARSM based on standard k – ω model is only motivated from the
somewhat simplified implementation. The model is actually an eddy-viscosity model
and thus only the routine that computes the eddy viscosity needs to be modified. Some
of the prediction capabilities of the full EARSM, (i) or (ii), is lost but this alternative is
still better than the more ad-hoc standard eddy-viscosity models with constant C µ .
S. WALLIN 197
(iv) EARSM based on Kok k – ω is similar to alternative (i) regarding implementa-

tion and prediction capabilities since the only difference is a cross-diffusion term added
to the ω -equation. The advantage with this alternative is that the unphysical free-
stream dependency in the standard k – ω is eliminated. However, this model is not
thoroughly tested yet.
These three models will be described with special emphasis on the extension to an
EARSM.
2.1. EARSM based on standard k-ω
2.1.1. Transport equations. The transport equations for k and ω read
∂  µ T ∂k
µ + ------
D
------ ( ρk ) = P – ρε +
Dt ∂x j  σk  ∂ x j
(1)
ω ∂  µ T ∂ω
µ + -------
D
------ ( ρω ) = α ---- P – βρω 2 +
Dt k ∂x j  σ ω ∂ x j
where
D ∂ ∂
------ ≡ + U j . (2)
Dt ∂ t ∂x j
The production of the turbulent kinetic energy is defined as
∂U i
P ≡ – ρu i u j (3)
∂x j
and the dissipation rate
ε ≡ β * ωk . (4)
α = 5 ⁄ 9 , β = 3 ⁄ 40 , β*
= 9 ⁄ 100 , σ k = 2.0 and σ ω = 2.0 are the model con-
stants.
In the original eddy-viscosity model the Reynolds stresses are obtained from
2 1  ∂U i ∂U j 2 ∂U k 
ρu i u j = --- ρkδ ij – 2µ T S ij* where S * ij = ---  + – --- δ  (5)
3 2  ∂ x j ∂ x i 3 ∂ x k ij
and the turbulent viscosity is given by

ρk
µ T = ------ . (6)
ω
2.1.2. Reynolds stresses. The Reynolds stresses ρu i u j expressed by EARSM may be

(ex)
written in terms of an effective turbulent viscosity µ T and an extra anisotropy a ij
2 (ex)
ρu i u j = --- ρkδ ij – 2µ T S ij* + ρka ij . (7)
3
The effective turbulent viscosity then reads

1
µ T = – --- ( β 1 + IIΩ β 6 )ρkτ (8)
2
= β 3  Ω 2 – --- IIΩ I + β 4 ( SΩ – ΩS )
(ex) 1
a
 3 
. (9)
+ β 6  SΩ 2 + Ω 2 S – IIΩ S – --- IV I + β 9 ( ΩSΩ 2 – Ω 2 SΩ )
2
 3 
The invariants are
IIS = tr { S 2 } , IIΩ = tr { Ω 2 } and IV = tr { SΩ 2 } . (10)
Here a , S and Ω denotes second rank tensors. tr{ } denotes the trace and I (or
δ ij ) is the identity matrix. The inner product of two matrices is defined as
( SS ) ij ≡ ( S 2 ) ij ≡ S ik S kj . The normalized mead strain- and rotation rate tensors are
defined as
S ij = τS * ij and Ω ij = τΩ * ij (11)
where τ is the turbulent time-scale. The dimensional strain- and rotation rate tensors
are
1  ∂U i ∂U j 2 ∂U k  1  ∂U i ∂U j
S * ij = ---  + – --- δ ij and Ω * ij = ---  – . (12)
2  ∂ x j ∂ xi 3 ∂ xk  2  ∂ x j ∂ xi 
In the near-wall region, viscous effects become important and the turbulent time-
scale in (11) need to be limited by the Kolmogoroff time scale as proposed by Durbin
(1993)
µ
τ = max  --, C τ ------
k
(13)
ε ρε
where C τ = 6.0 is a model constant.
The β -coefficients in (8) and (9) are given by
N ( 2N 2 – 7IIΩ ) 12N – 1 IV
β 1 = – ------------------------------------ β 3 = – ----------------------
Q Q
(14)
2 ( N 2 – 2IIΩ ) 6N 6
β 4 = – ------------------------------- β 6 = – ------- β 9 = ----
Q Q Q
with the non-singular denominator
5
Q = --- ( N 2 – 2IIΩ ) ( 2N 2 – IIΩ ) . (15)
6
N is given by
S. WALLIN 199
 C 1'
 ------- + ( P 1 + P 2 ) 1 / 3 + ( P 1 – P 2 ) 1 / 3 , P2 ≥ 0
 3
N =  (16)
 C 1' 1  P1  
 ------- + 2 ( P 12 – P 2 ) 1 / 6 cos  --- arccos  -----------------------  , P 2 < 0
 3 3  P2 – P  1 2
where
C 1' 2 9 C 1' 2 9 3
P 1 =  ---------- + ------ IIS – --- IIΩ C 1' , P 2 = P 12 –  ---------- + ------ IIS + --- IIΩ
2 2
(17)
 27 20 3   9 10 3 
and
9
C 1' = --- ( c 1 – 1 ) . (18)
4
When computing the third root one need to make sure that the real root is returned even
for negative arguments and the arccos function should return an angle between 0 and
π . c 1 = 1.8 is a model constant.
In 2D mean flows the β -coefficients are
6 N 6 1
β 1 = – --- ----------------------
- β 4 = – --- ----------------------
-
5 N 2 – 2IIΩ 5 N 2 – 2IIΩ (19)
β3 = 0 β6 = 0 β9 = 0
and the relation (9) simplifies significantly.
2.2. EARSM based on low-Re k-ω. The transport equation for k and ω are given
by (1) to (3) and the dissipation rate ε is given by (4) as for the standard k – ω model.
The differences enter through the model coefficients that in the case of low-Reynolds
number k – ω are functions of the turbulent Reynolds number Re T .
These functions read:
9 5 ⁄ 18 + ( Re T ⁄ R β )
4
β * = --------- ----------------------------------------------
-
100 1 + ( Re ⁄ R ) 4
T β
5 α 0 + Re T ⁄ R ω
α = --- ---------------------------------- ⋅ ( α * ) – 1 (20)
9 1 + Re T ⁄ R ω
β ⁄ 3 + Re T ⁄ R k
α * = ------------------------------------
1 + Re T ⁄ R k
where the turbulent Reynolds number is defined as
ρk
Re T = ------- . (21)
ωµ
R β = 8 , R ω = 2.7 , R k = 6 , α 0 = 1 ⁄ 10 , and β = 3 ⁄ 40 are the model constants.
2.2.1. Reynolds stresses. The Reynolds stresses ρu i u j is also here written in terms of
(ex)
the effective turbulent viscosity µ T and the extra anisotropy a ij as in (7) but here a
near-wall damping function needs to be introduced. The effective turbulent viscosity
then reads
1
µ T = – --- f 1 ( β 1 + IIΩ β 6 )ρkτ (22)
2
3B 2 – 4
-  S 2 – --- IIS I
(ex) 1
a = ( 1 – f 12 ) ------------------------------
eq  3 
max(IIS, IIS )
 B2 
+ f 12 β 3  Ω 2 – --- IIΩ I +  f 12 β 4 – ( 1 – f 1 ) ---------------------------------
1
- ( SΩ – ΩS ) (23)
 3   eq
2max(IIS, IIS )
+ f 1 β 6  SΩ 2 + Ω 2 S – IIΩ S – --- IV I + f 12 β 9 ( ΩSΩ 2 – Ω 2 SΩ )

2
 3 
where
eq 405c 12
IIS = ----------------------------- (24)
216c 1 – 160
and B 2 = 1.8 is an additional model constant.
The near-wall damping function f 1 in (22) and (23) was original defined in terms of
y + by Wallin & Johansson (1997) but has later been redefined in terms of Re y . The
near-wall damping is needed when EARSM is used together with low-Reynolds
number two-equation models with correct near-wall asymptotic behaviour.
f 1 = 1 – exp ( – C yA Re y – C yB Re y2 ) (25)
where
ρ ky
Re y ≡ ------------- . (26)
µ
C yA = 0.092 and C yB = 1.2 ⋅ 10 – 4 are model constants. The β -coefficients in (22)
and (23) are defined as previous in (14) to (17) or in (19) for 2D mean flows.
2.3. Linear EARSM based on standard k-ω. An alternative to EARSM based on

standard k – ω that is even more simple to implement, since it actually is an eddy-vis-
cosity model, is here described. The approach is to keep only the eddy-viscosity part of
(ex)
the relation (7) which means that a ij = 0 . Some of the physics covered of the full
EARSM is naturally missing in this approach and it cannot be motivated from
increased robustness or decreased computational cost. However, as an eddy-viscosity
model this approach is based on a more sound basis compared to standard eddy-viscos-
S. WALLIN 201
ity models with constant C µ . The only motivation for this approach is that only the
routine that computes the eddy-viscosity needs to be modified.
2.4. EARSM based on Kok k-ω. The dependenc on free-stream ω -values in the
k – ω model has been resolved by Kok (1999) by adding a cross-diffusion term in the
ω -equation and modifying the diffusion coefficients.
This model is identical to EARSM based on standard k – ω (see section 2.1) except
a few details. The transport equation for ω reads
ω ∂  µ T ∂ω ρ ∂k ∂ω 
µ + ------- + σ d ---- max 
D
------ ( ρω ) = α ---- P – βρω 2 + ,0 (27)
Dt k ∂x j  σ ω ∂ x j ω ∂ x j∂ x j 
where the diffusion coefficients are σ ω = 2.0 and σ d = 0.5 . The diffusion coefficient
in the k -equation (1) is modified to σ k = 1.5 .
3. Implementation
3.1. Formulation in terms of an effective viscosity. The Reynolds stresses

expressed by EARSM is written in terms of an effective turbulent viscosity µ T and an
(ex)
extra anisotropy a ij . The reason for this is that many CFD solvers already have dif-
ferent eddy-viscosity two-equation models implemented and that many routines may be
reused when the model is formulated in terms of an effective turbulent viscosity µ T .
Our experience is that it is sufficient only to consider the effective turbulent viscosity
µ T when determining the stability limit for explicit methods or the system Jacobian for
(ex)
implicit methods. The extra terms associated with the extra anisotropy a ij may be
treated as a fully decoupled source term without loss of performance of the solution
method.
3.2. Terms to be modified. 3.2.1. Production of the turbulent kinetic energy. The
production of the turbulent kinetic energy is defined as
∂U i
P ≡ – ρu i u j (28)
∂x j
and needs no further modelling since the Reynolds stresses are explicitly known. In an
eddy-viscosity two-equation model, this term is modelled as
∂U i
=  2µ T S ij* – --- ρkδ ij
(EVM) 2
P   ∂x j
. (29)
3
Many important physical aspects are lost through this assumption, like effects of
streamline curvature and local and global rotation and this term is thus the most impor-
tant to modify. By using the effective turbulent viscosity µ T in (29) and by adding the
contribution from the extra anisotropy the original definition of the production of the
turbulent kinetic energy is preserved
(EVM) (ex) ∂U i
P = P – ρka ij . (30)
∂x j
3.2.2. The turbulent transport of momentum. The second most important term to mod-
ify is the turbulent transport of momentum in the momentum equation. This term is also
exactly defined in terms of the Reynolds stresses and reads
∂
Ti ≡ ( – ρu i u j ) . (31)
∂x j
In an eddy-viscosity two-equation model, this term is modelled as
∂ 
2µ S * – --- ρkδ ij .
(EVM) 2
Ti = (32)
∂ x j  T ij 3 
Also for this term many important physical aspects are lost through this assumption,
like turbulence induced secondary motions. By using the effective turbulent viscosity
µ T in (29) and by adding the contribution from the extra anisotropy the original defini-
tion of turbulent transport of momentum is preserved
(EVM) ∂ (ex)
Ti = Ti – ( ρka ij ) . (33)
∂x j
3.2.3. Turbulent transport terms. There are also a number of other turbulent transport
terms, but these terms needs to be modelled even in terms of the Reynolds stresses. In
eddy-viscosity two-equation models, these terms are most commonly modelled using
gradient diffusion where the diffusion coefficients are related to the turbulent viscosity.
Also here the effective turbulent viscosity should be used, but no further modifications
of these terms are necessary.
It is, however, possible to adopt some more elaborate modelling in terms of the Rey-
nolds stresses also for these terms, but in that case one must use the near-wall correct
EARSM including near wall damping functions.
3.3. Wall boundary conditions. The physical wall boundary conditions on k and
ω are
∂k 6µ
k = 0, = 0 and ω → ------------2 (34)
∂y βρy
where y is the wall distance in the wall-normal direction. In a cell centred FVM, the
boundary condition should result in a correct wall flux. That is obtained by setting
S. WALLIN 203
k 0 = – k 1 for the convective fluxes and k 0 = k 1 for the viscous fluxes. Index 0 indi-
cates the dummy cell value while index 1 indicates the first interior cell.
The boundary condition on ω is ω 0 = 2ω wall – ω 1 where the wall value is given
by
6µ
ω wall = C ω-w -----------------
- (35)
βρy 2
1⁄2
where y is the wall normal distance to the first cell centre (for a cell centred
1⁄2
scheme) and C ω-w is a constant. Menter (1994) originally chosen that constant to 10
but later Hellsten (1998) suggested a value of 1.5 which minimizes the influence of the
near-wall grid spacing. The approach of Hellsten was adopted here, thus C ω-w = 1.5 .
3.4. General numerical discretization. The numerical treatment of standard two-

equation turbulence models described by Eliasson (1999), concerning e.g. multigrid,
positiveness, second order upwind scheme, was adopted also for the EARSM approach.
4. Shock - turbulence interaction

A shock may cause problems in computations using standard eddy-viscosity two-equa-
tion models. The problem is that the turbulence may grow at the shock and the growth
increases typically with increasing grid density. This problem do not occur in the
boundary layer and if EARSM is used. This phenomena is here analysed.
Let us consider a model problem, a one-dimensional moving shock. The shock
velocity is V , the velocity jump over the shock ∆V and the shock thickness ∆x . The
thickness of the shock is in high Reynolds number flows a function of the numerical
resolution of the shock.
Following the shock, the equation for the turbulent kinetic energy becomes, if the
diffusion and dissipation are neglected
∂
V (ρK ) = P (36)
∂x
and by integrating once
P ∆x
∆ ( ρK ) = ∫ ---V- dx ≈ P -----
V
-. (37)
We have here assumed that the production may be considered as constant over the
shock.
The production is
∂U i ∂U i 2 ∂U ∆V
= – ρK  a ij + --- δ ij = – ρK  a 11 + ---
2
P = – ρu i u j 
∼ ρK -------- (38)
∂x j 3  ∂x j  3 ∂ x ∆x
since the anisotropy is limited to around unity. The integrated K -equation then
becomes
∆V
∆ ( ρK ) ∼ ρK -------- . (39)
V
The turbulence growth over the shock is thus independent of the shock thickness, i.e.
the resolution of the shock.
In an eddy-viscosity model the production is modelled as
∂U i ∆V 2
=  2µ T S ij* – --- ρK δ ij ∼ µ T  --------
(EVM) 2
P   ∂x j  ∆x 
(40)
3
and the integrated K -equation becomes
( ∆V ) 2
∆ ( ρK ) ∼ µ T --------------- (41)
V∆x
and the turbulence growth over the shock increases with decreasing shock thickness.
For high Reynolds numbers, the turbulence may grow almost unbounded and thus gen-
erates unphysical high turbulence levels downstream of the shock. In fact also in sub-
sonic flows, any small disturbance in the velocity field or grid may generate unphysical
growth of free stream turbulence.
To avoid that effect in eddy-viscosity models the production may be limited. One
often used form is to limit the production by the dissipation rate P < Cρε where C is
some sufficiently high number. The problem with this form is that in rapidly growing
turbulence the production rate is, and should be, independent of the dissipation rate. A
more appropriate form is obtained by considering the relation (38)
∂U i
P = – ρK  a ij + --- δ ij
2
≤ ρK a ij S ij* < ρK S ij* = ρK S ij* S *ji (42)
3 ∂x j
which could serve as an upper limit for P . This limit is related to the realizability con-
straint that the norm of the anisotropy should be limited. The limit could also be
expressed in terms of the eddy-viscosity production as
(EVM)
(lim) P
P = ρK ----------------- (43)
µT
(EVM) (lim)
and the limited production is obtained by P = min ( P ,P ) which is easily
(EVM) (lim) (EVM)
implemented. P <P for P < ε ⁄ ( C µ f µ ) which implies that the limit
would not be active for P ⁄ ε < 10 . The limitor should thus not interfere with the predic-
tion capability of the model.
This problem does not occur for the proposed EARSM since the production is here
not modelled using the eddy-viscosity hypothesis. Even the linear EARSM described
previous does not have this deficiency. The reason for this is that the effective turbulent
viscosity, or the effective C µ , is a function of the strain rate such that
C µ → 1 ⁄ S ij* S *ji for large strain rates.
S. WALLIN 205
5. Test cases
5.1. RAE2822: 2D transonic wing profile. The two-dimensional transonic wing

profile RAE2822, case10 illustrates the behaviour of the proposed model. The wing is
highly loaded and a boundary layer separation occurs at the shock impingement. Figure
1 shows the wall pressure and skin friction coefficients, C p and C f . Standard eddy-
viscosity models, like the Wilcox (1988) standard k – ω model denoted ‘std_kw’, pre-
dict a shock position downstream of the observed one. The proposed EARSM approach
clearly improves the shock position and the results are very much in line with differen-
tial Reynolds stress computations by Hellström et al. (1994). For this case there are no
major difference in computational results between EARSM based on standard or low-
Reynolds number k – ω , ‘WJ_skw’ or ‘WJ_kw’. The eddy-viscosity version
‘LWJ_kw’ gives, however, a shock position downstream of the EARSM’s. Also the
computational results using the Menter (1994) SST model, ‘SST_kw’, is shown in the
figure. That model gives a somewhat better shock position. The best prediction of the
shock position is, however, obtained with EARSM based on the Kok k – ω model. Fig-
ure 3 shows the velocity profiles at different x -positions.
There are still some deviations from the experimental data and there are other com-
putations published on this case that have better agreements. The reason for this is that
there exists a number of different geometries for this case, the measured or design
geometry with or without an additional camber correction. For this case the measured
geometry with the camber correction was used.
The RMS ( ρ ) -residual convergence history is shown in figure 2. The eddy-viscosity
k – ω model has the best convergence rate but the different EARSM’s have similar
convergence rates. The EARSM approach increases the computational time with less
that 10% and the different computations were obtained by using the same numerical
settings (CFL, multigrid levels, upwind discretization, et.c.). In flows without separa-
tion the convergence curves are even closer to each other.
5.2. LANN: 3D transonic wing. A further example for illustrating the model in
three-dimensional flows is the transonic LANN-wing, measured by Horsten et al.
(1983). Figure 4 shows the pressure coefficient distribution at different spanwise posi-
tions and figure 5 shows the surface pressure distribution as well as the computed skin
friction lines. Also here the EARSM approach (‘WJ_skw’) improves the predicted
shock position compared to the Wilcox (1988) standard k – ω model (‘std_kw’).
-1.50
std_kw
SST_kw
WJ_skw
-1.00 WJ_kw
LWJ_skw
WJ_Kkw
expr
-0.50
Cp
0.00
0.50
1.00
0.00 0.20 0.40 0.60 0.80 1.00
x/c
0.0100
std_kw
SST_kw
WJ_skw
WJ_kw
0.0060 LWJ_skw
WJ_Kkw
expr
zero
Cf
0.0020
-0.0020
-0.0060
0.00 0.20 0.40 0.60 0.80 1.00
x/c
FIGURE 1: Predicted C p and C f distributions compared to measurements for

6
the RAE2822 wing profile at Ma = 0.754 , α = 2.57 and Re = 6.2 ⋅ 10
(Case 10). The standard eddy-viscosity k – ω model (std_kw) and Menter
SST k – ω (SST_kw) compared to the EARSM based on standard k – ω
(WJ_skw), EARSM based on low-Reynolds number k – ω (WJ_kw),
EARSM based on Kok k – ω (WJ_Kkw marked with small squares) and the
Linear EARSM based on standard k – ω (LWJ_kw). Experimental data from
Cook et al. (1979).
S. WALLIN 207
0.0
std_kw
WJ_skw
WJ_kw
LWJ_skw
-2.0
rho_rms
-4.0
-6.0
0. 2000. 4000. 6000.
iter
FIGURE 2: ρ -residual convergence history for the RAE2822 wing profile. The
standard eddy-viscosity k – ω model (std_kw) compared to the EARSM
based on standard k – ω (WJ_skw), EARSM based on low-Reynolds number
k – ω (WJ_kw) and the Linear EARSM based on standard k – ω (LWJ_kw).
0.03 0.03
std_kw
SST_kw
WJ_skw
x/c=0.404 x/c=0.650 WJ_kw
LWJ_skw
WJ_Kkw
0.02 0.02 expr
y/c y/c
0.01 0.01
0.00 0.00
0.0 0.4 0.8 1.2 -0.2 0.2 0.6 1.0
U/Uf U/Uf
0.03 0.045
std_kw
SST_kw
WJ_skw
x/c=0.750 x/c=0.900 WJ_kw
LWJ_skw
WJ_Kkw
0.02 0.030 expr
y/c y/c
0.01 0.015
0.00 0.000
0.0 0.4 0.8 1.2 0.0 0.4 0.8 1.2
U/Uf U/Uf
0.08 0.08
std_kw
SST_kw
x/c=1.000 x/c=1.025 WJ_skw
WJ_kw
LWJ_skw
WJ_Kkw
0.04 0.04 expr
y/c y/c
0.00 0.00
-0.04 -0.04
0.0 0.4 0.8 1.2 0.0 0.4 0.8 1.2
U/Uf U/Uf
FIGURE 3: Predicted velocity profiles at different x -positions compared to

measurements for the RAE2822 wing profile at Ma = 0.754 , α = 2.57 and
6
Re = 6.2 ⋅ 10 (Case 10). The standard eddy-viscosity k – ω model
(std_kw) and Menter SST k – ω (SST_kw) compared to the EARSM based
on standard k – ω (WJ_skw), EARSM based on low-Reynolds number k – ω
(WJ_kw), EARSM based on Kok k – ω (WJ_Kkw marked with small squa-
res) and the Linear EARSM based on standard k – ω (LWJ_kw). Experimen-
tal data from Cook et al. (1979).
S. WALLIN 209
0.200 0.325
0.475 0.650
0.825 0.950
FIGURE 4: Predicted C p distribution at different spanwise positions compared

to measurements for the LANN wing at Mach = 0.82 , α = 2.60 and
6
Re = 7.16 ⋅ 10 . The standard eddy-viscosity k – ω model ( ) compa-
red to the EARSM based on standard k – ω ( ). Experimental data from
Horsten et al. (1983).
k–ω
EARSM
FIGURE 5: Predicted skin friction lines for the LANN wing at Mach = 0.82 ,
6
α = 2.60 and Re = 7.16 ⋅ 10 . The standard eddy-viscosity k – ω model
(top) compared to the EARSM based on standard k – ω (bottom).
6. Concluding remarks
The EARSM approach was shown to improve the shock position over standard eddy-
viscosity two-equation models, for highly loaded transonic wings, without significal
increased computational cost. The EARSM without damping functions with the stand-
ard k – ω model as the platform is recommended for this kind of problems. The addi-
tional accuracy in the near-wall region, that the EARSM with damping functions
provides, is probably not needed in the cases studied here.
The Kok k – ω model is an interesting alternative since the erroneously free-stream
dependency in standard k – ω models have been removed. The EARSM based on that
S. WALLIN 211
model also shows the best predictions for the RAE2822 case. This combination
(EARSM + Kok k – ω ) should, thus, be further investigated and tested, especially con-
cerning the free-stream dependency.
The Menter SST model predicts the shock position well for the RAE2822 and could
also be recommended for these kind of flow problems. This model is, however, an
eddy-viscosity model with an empirical limitation on the turbulent viscosity and thus
lacks the generality of the more fundamental EARSM. That is obvious in rotational
dominated flows where the EARSM has the correct qualitative behaviour while eddy-
viscosity models and in some cases also incomplete quadratic EARSMs fail, reported
by Wallin & Johansson (1997).
Acknowledgment
This work has been carried out within the the AVTAC Project (Advanced Viscous
Flow Simulation Tools for Complete Civil Transport Aircraft). AVTAC is a collabora-
tion between British Aerospace, DASA, CASA, Dassault Aviation, SAAB, Alenia,
DLR, ONERA, CIRA, FFA and NLR. The project is managed by British Aerospace
and is funded by the CEC under the IMT initiative (Project Ref: BRPR CT97-0555)
References
[1] Cook, P.H., MacDonald, M.A. & Firmin, M.C.P. 1979 Aerofoil 2822 - Pressure dis-
tributions, boundary layer and wake measurements. AGARD AR 138.
[2] Durbin, P.A. 1993 Application of a near-wall turbulence model to boundary layers
and heat transfer. Int. J. Heat and Fluid Flow 14, 316–323.
[3] Eliasson, P. 1999 A robust and positive scheme for viscous, compressible steady
state solutions with two-equation turbulence models. FFA TN 1999-81.
[4] Gatski, T.B. & Speziale, C.G. 1993 On explicit algebraic stress models for complex
turbulent flows. J. Fluid Mech. 254, 59–78.
[5] Girimaji, S.S. 1995 Fully-explicit and self-consistent algebraic Reynolds stress
model. ICASE Report No. 95-82.
[6] Hellsten, A. 1998 On the solid-wall boundary condition of ω in the k – ω -type tur-
bulence models. Helsinki University of Technology, Report No B-50, Series B.
[7] Hellström, T., Davidson, L. & Rizzi A. 1994 Reynolds stress transport modelling of
transonic flow around the RAE2822 airfoil. AIAA 94-0309. 32nd Aerospace Sciences
Meeting, Reno, Jan. 1994.
[8] Horsten, J.J., Den Boer, R.G. & Zwaan, R.J. 1983 Unsteady transonic pressure
measurements on a semi-span wind tunnel model of a transport-type supercritical
wing (LANN model). AFWAL-TR-83-3039.
[9] Johansson, A.V. & Wallin, S. 1996 A new explicit algebraic Reynolds stress model.
Proc. Sixth European Turbulence Conference, Lausanne, P. Monkewitz (ed). 31–34.
[10] Kok, J.C. 1999 Resolving the dependence on free-stream values for the k – ω tur-
bulence model. Submitted to AIAA J.
[11] Launder, B.E. Reece, G.J. & Rodi, W. 1975 Progress in the development of a Rey-
nolds-stress turbulence closure. J. Fluid Mech. 41, 537–566.
[12] Menter, F.R. 1994 Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA J. 32, 1598–1604.
[13] Pope, S.B. 1975, A more general effective-viscosity hypothesis. J. Fluid Mech. 72
331–340.
[14] Rodi, W. 1972 The prediction of free turbulent boundary layers by use of a two
equation model of turbulence. Ph.D. thesis, University of London.
[15] Rodi, W. 1976 A new algebraic relation for calculating the Reynolds stresses. Z.
angew. Math. Mech. 56, T219–221.
[16] Wallin, S & Johansson, A. V. 1996, A new explicit algebraic Reynolds stress tur-
bulence model including an improved near-wall treatment. Proc. Flow Modeling and
Turbulence Measurements VI, Tallahassee, Chen, Shih, Lienau & Kung (eds).
[17] Wallin, S. & Johansson, A.V. 1997 A complete explicit algebraic Reynolds stress
model for incompressible and compressible turbulent flows. FFA TN 1997-51 (submit-
ted for publication). Also to appear in J. Fluid Mech, 403, 89-132, (2000).
[18] Wilcox, D.C. 1988 Reassessment of the scale-determining equation for advanced
turbulence models. AIAA J. 26, 1299–1310.
[19] Wilcox, D.C. 1994 Simulation of transition with a two-equation turbulence model.
AIAA J. 32, 247–255.
Paper 7
7
A ROBUST AND POSITIVE SCHEME FOR VISCOUS, COMPRES-
SIBLE STEADY STATE SOLUTIONS WITH TWO-EQUATION TURBU-
LENCE MODELS
Peter Eliasson and Stefan Wallin
Abstract. An efficient and robust method has been designed for the simulation of
the compressible Navier-Stokes equations with two-equations turbulence models.
The compressible Navier-Stokes equations are integrated explicitly to steady state
and a novel approach is used to maintain the positivity of the turbulent quantities.
The novel approach is based on a conservative estimate of the characteristics of
the turbulent equations leading to a restriction of the local time step determined
by the residual. As steady state is approached the residual and the restriction of
the time step is reduced and hence the asymptotic rate of convergence is not
affected. Multigrid is used to accelerate the convergence and a similar approach is
used when adding the corrections to guarantee a positive scheme. A higher order
restriction operator improves the robustness and the performance of the scheme. It
is demonstrated that the higher order restriction can improve the rate of conver-
gence, especially for ‘bad’ grids with large stretchings and locally large variations
in grid size. Numerical results are presented for a model problem and on a
RAE2822 airfoil.
1. Introduction
The design of efficient and robust numerical algorithms is essential for the simulation
of compressible turbulent flows, especially for engineering-type applications. The main
objective of this paper is to demonstrate new approaches to gain increased robustness in
the iteration of an explicit Navier-Stokes solver coupled with a two-equation turbulence
model and accelerated with a multigrid method.
It is well known that positivity of the discrete turbulent variables at any time during
iteration is crucial. Unphysical negative values rapidly lead to instability and positivity
must therefore be ensured. Various methods exist to achieve this. Ad hoc methods like
enforcing the positivity by taking the absolute values of the turbulent quantities often
prevents the convergence or cause divergence of the iterative process.
A more classical approach for avoiding the generation of spurious oscillations is to
impose TVD-like conditions [10]. Explicit TVD schemes for transport equations have
been considered by Harten [8] and implicit schemes by Yee et al. [16]. When transport
equations for turbulent variables are present additional constraints need to be made to
the source term. Jongen & Marx have developed TVD-like criteria for a general advec-
tion-diffusion equation with a source term. Both explicit and implicit time integration
are considered. They consider separately the convective, viscous and source term of the
equation and give a positivity criterion for the advection term that is a weaker require-
215
216 ROBUST AND POSITIVE SCHEME FOR STEADY STATE 2-EQUATION TURBULENCE MODELS
ment of a TVD criterion. The central viscous term can not increase the total variation.
The negative part of the source term is treated implicitly and the positive part explicitly.
A criterion is given on how the linearization of the negative terms must be chosen to
maintain positivity. The criterion for the source term often restricts the local time step
more than does the stability criterion. In addition, Jongen & Marx consider only the
update of the variables in the time integration. They give no criteria of how to maintain
positivity of the variables in a multigrid procedure where the variables are updated from
prolongated coarse grid corrections.
In this paper an alternative, novel approach is suggested to guarantee positive turbu-
lent variables. The approach is based on an estimate of the spectral radius of the com-
plete turbulent equations and produces an underrelaxation of the local time step based
on the residual. The underrelaxation is only active in those regions where the residual is
large compared to the positive dependent variable and where the sign of the residual is
such that the variable is decreased. The underrelaxation does not affect the asymptotic
rate of convergence since the relaxation is only active initially when the solution is far
from being converged. This results in a method that is not required to fulfil the TVD
criterion for the convection, nor from the positivity requirements setup for the source
term to have a positive solution in the iterative time stepping. The only requirement is
that a stable time step is chosen.
An explicit Runge-Kutta finite-volume method is used to solve the compressible,
Reynolds averaged Navier-Stokes equation. Local time steps are used based on a stabil-
ity analysis of the convective and viscous terms. A point-implicit treatment of the tur-
bulent source terms is chosen to guarantee stability in the iteration to steady state of the
turbulence.
The convergence is further accelerated by using FAS multigrid. To ensure positive
variables when the prolongated coarse grid corrections are added, a similar procedure
as in the time stepping is introduced. An additional parameter is introduced to scale the
underrelaxation, the size of this parameter is investigated numerically.
A higher order restriction operator is used which is basically the transpose of a linear
prolongation operator. For some reason, to the authors knowledge this restriction oper-
ator has not been used before for CFD calculations. It is demonstrated in this paper that
a higher order operator is as important as a higher order (linear) prolongation operator
used on a common basis and that a lower order operator may prevent convergence. It is
shown that especially when the grid is ‘bad’ with local large variations and when the
grid is stretched with high cell aspect ratios the combination of the higher order transfer
operators are beneficial.
Numerical results are presented for a model problem and for a 2D RAE2822 airfoil
(case 10). The computations are all started from free stream, with or without multigrid.
Although the novel approach is not restricted to a specific turbulence model, results are
presented fro the k−ε equations only with the near wall treatment by Chien [2]. The
focus will not be on the predicted results, only robustness and the rate of convergence
are considered.
P. ELIASSON AND S. WALLIN 217
2. Construction of positive scheme

Finite-volume discretization schemes for a general transport equation are considered
integrated to a steady state. The purpose is to investigate the stability and positivity
requirements in the time integration. A positive discretization of the advection may be
obtained by a TVD like criterion, see e.g. Jongen & Marx [10] where it also shown that
the diffusion does not increase the total variation. A positivity requirement of the
source term is also given leading to a restriction of the local time step. This requirement
is usually more restrictive than the requirement for stability. Below an alternative
approach is suggested in which the stability requirement only is required. The positivity
is ensured by a an underrelaxation of the time step in the initial stage of the computa-
tion.
2.1. 1-D model problem. To investigate the positivity requirements a simple model
problem is considered. A scalar, linear semi-discrete model problem with diffusion and
a source term is investigated
dq i ν
= R ( q ) = --------2- ( q i + 1 – 2q i + q i – 1 ) + aq i (1)
dt ∆x
where q represents a positive, transported quantity (e.g. k or ε), ν represents a positive
viscosity, ∆x the step length and a is a constant multiplying the source term. The
advection term is left out simply to reduce the complexity of the algebra.
a > 0 leads to an exponential growth without the presence of the diffusion. The
source term may be positive locally in regions where there is a turbulent growth. Posi-
tive source terms do not restrict the local time step, they are integrated explicitly and
will eventually be balanced by the negative source terms or the diffusion when the tur-
bulence stops to grow. Positive source terms do no possess a threat to turn q negative
since q is growing, therefore only negative source terms a < 0 are considered which
may push q towards zero.
The diffusion term is integrated explicitly, as in the CFD calculation. Without the
source term explicit Euler forward integration gives the following stability bound on
the time step:
n+1 n 2
qi – qi ν n n n ∆x
------------------------ = --------2- ( q i + 1 – 2q i + q i – 1 ) ⇒ ∆t v ≤ --------- (2)
∆t v ∆x 2ν
which is easily verified by a von Neumann analysis. With the negative source term
present and explicit integration leads to
n+1 n
qi – qi ν n n n n 1 ∆t v
------------------------ = --------2- ( q i + 1 – 2q i + q i – 1 ) + aq i ⇒ ∆t ≤ ------------------ = -------------------- (3)
∆t ∆x 2ν a a∆t
--------- – --- 1 – ----------v-
2 2
∆x 2
for a stable time step provided ∆t v is chosen as its maximum. The source term restricts
the time step by an additional term in the denominator. To guarantee positivity in the
integration (3) the following constraint must be made to the time step:
1 ∆t v
Positivity: ∆t ≤ ------------------ = -------------------- (4)
2ν 1 – a∆t v
--------- – a
2
∆x
When the source term is dominating the positivity requirement restricts the time step
by a factor of two.
Jongen & Marx [10] propose to treat negative source terms point implicitly leading to
n+1 n
qi – qi ν n n n n+1
------------------------ = --------2- ( q i + 1 – 2q i + q i – 1 ) + aq i ⇒
∆t ∆x (5)
n+1 n
 ----- qi – qi ν
– a ( q i
1 n+1 n n n n n
– q i ) = ------------------------ = --------2- ( q i + 1 – 2q i + q i – 1 ) + aq i
 ∆t  *
∆t ∆x
which increases the stability limit for the time step ∆t :
1
∆t ≤ ------------------- (6)
2ν a
--------- + ---
2 2
∆x
However, the time step used in the time integration is the efficient time step
*
∆t = ∆t ⁄ ( 1 – a∆t ) which is restricted by the same stability and positivity criteria as
the explicit scheme, (3) and (4). In either case of explicit or point implicit time integra-
tion, the time step has to be restricted due to the source term. In the case of this scalar,
one-stage integration problem explicit and point-implicit time integration become iden-
tical.
It is not attractive to restrict the time step due to a positivity criterion more restrictive
than the stability criterion since it will decrease the rate of convergence. Ideally, the
positivity requirement should be used initially only when the residual is large and may
cause the solution to turn negative. When the residual becomes smaller the stability
limit should determine the size of the time step. Below such an approach is presented.
If the residual is assumed to be negative and large and the residual is dominated by
the source term the model problem may be approximated as:
dq i
= R ( q ) ≈ aq i (7)
dt
Hence the size of the source term may be estimated as
R(q)
a ≈ ----------- (8)
q
According to the expression for the time step in the model problem (3) a positive
time step ∆t p may be obtained as
∆t
∆t p ≤ -------------------------------------------------- (9)
R(q)
1 – ∆t min  -----------, 0
 q 
which ensures that the time step is determined by the stability requirement only as the
residual becomes small compared to the solution. ∆t is then determined from reasons
of stability only, the additional term in the denominator in (9) guarantees positivity. For
a point implicit approach, the time step may be determined from the explicit terms
(advection and diffusion) which for the example above would lead to
∆t v
∆t p ≤ ----------------------------------------------------- (10)
R(q) a
1 – ∆t v min  -----------, ---
 q 2
where ∆t v is given in (2). Note that the time step is only reduced when the residual is
negative, i.e. when the solution is decreasing. When it increases the time step is deter-
mined from stability reasons only and the expressions (9) and (10) are identical. Note
that the expressions (9) and (10) are not identical for negative residuals. (9) is more
restrictive for large negative residuals. The expressions (9) and (10) are chosen because
of their similar appearance, other choices of time steps are possible that guarantees pos-
itivity.
The one-stage integration may be extended to m-stage Runge-Kutta schemes. E.g.
the 2-stage point-implicit scheme:
1 n  ν n n n 1
q i = q i + α 1 ∆t  --------2- ( q i + 1 – 2q i + q i – 1 ) + aq i 
 ∆x 
(11)
n+1 1  ν 1 1 1 n + 1
qi = qi + ∆t  --------2- ( q i + 1 – 2q i + q i – 1 ) + aq i 
 ∆x 
where α 1 is a Runge-Kutta coefficient. This scheme will, as the one-stage scheme, be

* *
integrated with the efficient time step ∆t = ∆t ⁄ ( 1 – a∆t ) and ∆t will be subject to
the same stability and positivity requirement as the corresponding explicit time integra-
tion with time step ∆t . The stability requirement becomes for the explicit scheme:
1 ⁄ α1
∆t ≤ ------------------ (12)
4ν
--------- – a
2
∆x
The positivity requirement may be found for each stage in the scheme to be
2
Positivity: ∆t ≤ ---------------------- (13)
4ν
--------- – 2a
2
∆x
provided α ≤ 1 . For a dominating source term the positivity requirement is certainly

more restrictive. For vanishing source terms ( a → 0 ) Shu [13] gives the necessary CFL-
conditions for arbitrary m-stage Runge-Kutta to be TVD.
For a general m-stage Runge-Kutta scheme a CFL criterion the condition (9) may be
written
∆t
∆t p = -------------------------------------------------------------- (14)
∆t R(q)
1 – -------------- min  2 -----------, 0
CFL v  q 
for an explicit integration of the viscous and source term where ∆t is a stable time step
based on the spectral radius of the viscous and source term
CFL v
∆t = ------------------ (15)
4ν
--------- – a
2
∆x
and where CFL v is the absolute value of the intersection of the stability region with the
negative real axis in the complex plane.
For a point implicit approach the equation (10) for a general m-stage Runge-Kutta
scheme may be written
∆t v
∆t p = -------------------------------------------------------------- (16)
∆t v R(q)
1 – -------------- min  2 -----------, a
CFL v  q 
where the time step is based on the spectral radius of the viscous term only:
CFL v
∆t v = -------------- (17)
4ν
---------
2
∆x
The equations (14) and (16) are valid for explicit and point implicit time integration
respectively and guarantee a positive and stable time integration for an m-stage Runge-
Kutta scheme applied to the model problem (1) provided the Runge-Kutta coefficients
0 ≤ α1 ≤ i ≤ m ≤ 1 .
2.2. The k-ε equations. The positive time steps for the model problem above may
easily be extended to two-equation turbulence models As an example the k-ε equations
are used, but any two-equation turbulence model may be used. The equations for the
Chien k-ε is:
µ µk
------ ρk = ∇ ⋅  µ + -----T- ∇k + P – ρε – 2 ------
D
Dt  σk  2
y
(18)
1 +
µ ε µε 2
– --- y
------ ρε = ∇ ⋅  µ + -----T- ∇ε + ( C ε1 P – C ε2 f 2 ρε ) -- – 2 ------ e
D
Dt  σε  k 2
y
where
2
2 2 k
P = S µ T – --- ρkD, µ T = C µ f µ ρ ----- (19)
3 ε
The negative source terms are treated point implicitly and must be accounted for so
-
that a stable time step is obtained. The complete Jacobian of the negative source S
terms may be used
 µk  2µ
 – ρε – 2 -----2-  – --------- –1
 y  ρy
2
S =
-  , ∂ (S -) =
 1 +  ∂q 1 + (20)
– --- y 2 – --- y
 ε µε 2  ε ε µε 2
 ( – C ε2 f 2 ρε ) -k- – 2 -----2- e  C ε2 f 2 ----2- – 2C ε2 f 2 -- – 2 -----2- e
k
 y  k y
-
or an estimate of the spectral radius ρ of it may be used. In the computations pre-
-
sented in this paper ρ is used and approximated as
ε µ ε µ – y + ⁄ 2
ρ = min  – 2 -- – 2 --------
-
-, – 1.5C ε2 f 2 -- – 2 --------- e (21)
 k 2 k 2 
y ρ y ρ
As will be shown using the spectral radius or the complete Jacobian has a small
influence on the rate of convergence.
The time step used in the computations is
∆t
∆t* = --------------------------------------------------------------------------------------- (22)
∆t ( ρk ) R ( ρε ) -
1 – ----------- min  2 ---------------, 2 ---------------, ρ 
R
CFL  ρk ρε 
where ∆t is determined from a stability analysis of the convective and diffusive terms
of the mean flow [14]. Hence ∆t is used to update the mean flow, ∆t* is used for the
turbulence. A common time step instead of two separate time steps was chosen for the
two turbulent equations since they are closely coupled.
3. Multigrid
To accelerate the convergence of the explicit Runge-Kutta time integration with local
time steps, FAS multigrid is used. The multigrid is applied to both the mean flow as
well as to the turbulence. Below the positive update in the time integration is extended
to the update of the corrections in multigrid. A higher order restriction operator is also
presented.
3.1. Positivity. The dependent variables are updated in the Runge-Kutta time integra-
tion but also in the multigrid procedure where a prolongated correction is added to the
variables, see e.g. Hackbush [7]. Positivity of the turbulence is then not only of concern
in the time integration but also when the multigrid corrections are added.
The corrections are added as:
n+1 n
q = q + ∆q (23)
where ∆q is the correction. There is nothing in the update of the corrections (23) that
prevents the turbulent variables to turn negative.
By considering the corrections as residuals the approach from the time integration
may be extended. The following modified expression for adding the corrections ensures
positivity:
n+1 n ∆q
q = q + ----------------------------------------- (24)
∆q 
1 – βmin  ------ -, 0
 n 
q
Note that the correction is only reduced when it is negative and that the reduction
becomes small as soon as the correction becomes small compared to the solution. An
additional parameter β > 1 has been introduced to keep the turbulence sufficiently far
away from zero. In the computations presented below it will be shown that the value of
β has very small influence on the rate of convergence provided it is chosen big enough
to prevent initial divergence and blow up. β should for most cases be chosen β ≥ 10
for reasons of robustness, in the computations β = 100 is usually used.
For the k-ε equations the following expression is used:
n+1 n ∆q
q = q + --------------------------------------------------- (25)
∆k ∆ε 
1 – βmin  ------ , ------, 0
 n n 
k ε
where q is either k or ε and ∆q the correction ∆k or ∆ε respectively. Note that k
and ε are equally restricted not to destroy their close coupling.
3.2. A higher order restriction operator. In multigrid the residuals of a fine grid
is restricted to coarser grids to form a forcing function on the right hand side [7]. A
semi-discrete equation discretized by a finite-volume method is denoted
dq l
Vl = R ( ql ) + d l (26)
dt
where V l is the volume of a cell and where subscript l ≥ 1 denotes the grid level,
l = 1 being the finest. d l is the forcing function restricted from finer levels, d 1 = 0 .
For more details see e.g. Rizzi et al. [14].
A restriction operator commonly used is the transpose of the simplest prolongation
operator, the injection. These restriction and prolongation operators are illustrated in
Figure 1 in two space dimensions. Note that the restriction is for the residuals only, to
restrict the unknowns a factor of 1/4 multiply fine grid unknowns. This is a result of the
finite-volume discretization.
Restriction Prolongation (injection)

1 1 1
4 1
1 1
FIGURE 1: Prolongation operator of lowest order (injection) in two dimen-

sions, restriction as its transpose.
The injection prolongation is usually not accurate enough so a linear prolongation

operator may be used instead. The linear prolongation may be used to form a higher
order restriction by taking the transpose of the prolongation, Figure 2.
1 3 3 1 Restriction Prolongation (bilinear)
1 3 9 9 3
_
16 3 9 9 3 9 3
------ ------
16 16
1 3 3 1
3 1
------ ------
16 16
FIGURE 2: Linear prolongation in two dimensions and restriction as its trans-

pose.
The higher order restriction needs values outside the boundaries, it needs boundary
conditions on the residual. At solid walls the boundary condition
R0 = – R1 (27)
is used for all residuals. At external boundaries the same boundary condition is used
although the boundary condition on these boundaries have negligible impact on the rate
of convergence. For more details about the multigrid operators as well as analysis for
model problem, see Eliasson [3] and Eliasson et al. [4].
Note that the higher order restriction should not be used for the variables. Although
it should be possible to apply any consistent operator in theory, in practice the large var-
iations and non-linearities of the turbulent variables often cause problem for this wider
operator and hence the lower order, compact operator is used for the variables.
The extension of the operators to three space dimensions is obtained from the trans-
pose of the tri-linear prolongation operator. The coarse grid cell centre forcing function
F i, j, k may in three dimensions be obtained as
j 2j+1 1 3 3 1 3 9 9 3 3 9 9 3 1 3 3 1
2j 3 9 9 3 9 27 27 9 9 27 27 9 3 9 9 3
1 2j-1 3 9 9 3 9 27 27 9 9 27 27 9 3 9 9 3
- 2j-2 1 3 3 1 3 9 9 3 3 9 9 3 1 3 3 1
64
2i-2 2i-1 2i 2i+1
i
k=2k-2 k=2k-1 k=2k k=2k+1
FIGURE 3: Higher order restriction operator in three dimensions. Coarse grid
cell centre forcing function F i, j, k as a result.
It will be demonstrated that in some cases the lower order operators prevent conver-
gence.
4. Solution technique
4.1. Time integration. The equations are integrated to steady state using an explicit
5-stage Runge-Kutta scheme with point-implicit treatment of the turbulent negative
source terms described above. The Runge-Kutta coefficients are:
α 1 = 0.0814, α 2 = 0.1906, α 3 = 0.342, α 4 = 0.574, α 5 = 1 (28)
The viscous terms and the source terms are calculated in the first stage only. This
scheme provides optimum damping for both central and upwind schemes [3], the CFL
number in the computations is CFL=1.5.
4.2. Spatial discretization. The mean flow convective terms as well as the turbu-
lent convective terms are discretized using a second order accurate symmetric TVD
scheme [11]. Second order accuracy is obtained by using a van Leer limiter for the mean
flow applied to the characteristics and a minmod limiter for the turbulence.
An entropy fix is used to prevent unphysical solution, the eigenvalues of the mean
flow and turbulence are not allowed to be less than 5% of the spectral radius u + c .
4.3. Coarse grid simplification. On coarser grids the mean flow and the turbulent
convective terms are reduced to first order accuracy. The entropy fix is increased to
20% and the turbulent production is set to zero, i.e. it is assumed to be constant from
the fine grid.
5. Numerical Results
5.1. Model equation. To investigate the time discretization and the different time
step restrictions a model equation is used [15]. The model problem varies in time only
and is basically the k-ε equations without convection and near-wall damping:
Pk – ε
dq
= R(q) = V (q) + S(q) = – α k – 1 + ε (29)
dt ε–1 ( C ε1 P k – C ε2 ε ) --
k
T
where q = [ k, ε ] and V ( q ) represents schematically the viscous terms where α is
2 2
positive and related to µ ⁄ ∆x . S ( q ) is the source term, P k = C m k ⁄ ε is the turbu-
lent production and C ε1 = 1.45, C ε2 = 1.9 .
The spectral radius to the negative terms of S ( q ) may be found from the Jacobian
- –ε 0 –1
∂S ∂
J (q) = = 2 = ε
2
ε (30)
∂q ∂ q – C ε----- C ε2 ----2- – 2C ε2 --
ε2 k k
k
to be
- ε
ρ = – ( C ε2 + C ε2 ( C ε2 – 1 ) ) -- (31)
k
One stage explicit Euler time integration is made with three different limitations to
the time step. In the first approach a strictly positive time step is chosen according to
Equation (13):
CF L v
∆t = ------------------ (32)
-
α – 2ρ
The second time step used is a time step based on stability requirements only
CF L v
∆t = -------------- (33)
-
α–ρ
and may therefore initially turn k or ε negative. An ‘ad hoc’ approach is then used by
setting the negative value to 1% of the initial value.
In the third approach the new way of restricting the time step is used in accordance
with Equation (14)
CF L v
∆t = --------------------------------------------------------------------------------- (34)
-
α – ρ – min ( R ( k ) ⁄ k, R ( ε ) ⁄ ε, 0 )
or, equivalently, with the notation in Equation (14)
∆t
∆t p = ------------------------------------------------------------------------------------ (35)
∆t
1 – -------------- min ( R ( k ) ⁄ k, R ( ε ) ⁄ ε, 0 )
CF L v
where ∆t is defined in equation (33).
In Figure 4 and Figure 5 two computations with different values of the parameters
are visualized. The results are compared against an exact solution in time. Since a
steady state problem is used the focus is the number of iterations required to reach the
asymptotic steady state values. The computed results are therefore plotted using the
same time step although different time steps (32)-(34) were used in the computations.
As can be seen, for both cases the new proposed time integration reaches the asymp-
totic value fastest.
Other ‘ad hoc’ resetting of negative values can result in both worse and better behav-
iour than demonstrated below. The example demonstrates the danger with that proce-
dure and the increased robustness with the new proposed time step.
a=1.0 dt=0.9 k0=1 e0=10 Cm=1

3.00
exact
pos,spec
stab,spec
new
2.00
1.00
0.00
0.0 6.7 13.3 20.0
t
a=1.0 dt=0.9 k0=1 e0=10 Cm=1

3.00
exact
pos,spec
stab,spec
new
2.00
eps
1.00
0.00
0.0 6.7 13.3 20.0
t
FIGURE 4: Model example with k 0 = 1, ε 0 = 10, α = 1 . Time evolution of

k (top) and ε (bottom). ___ exact solution, positive time step, O stable
time step with ‘ad hoc’ resetting, ∆ new proposed time step.
a=0.01 dt=0.9 k0=0.01 e0=1 Cm=0.01

0.0300
exact
pos,spec
stab,spec
new
0.0200
0.0100
0.0000
0.0 6.7 13.3 20.0
t
a=0.01 dt=0.9 k0=0.01 e0=1 Cm=0.01

0.0300
exact
pos,spec
stab,spec
new
0.0200
eps
0.0100
0.0000
0.0 6.7 13.3 20.0
t
FIGURE 5: Model example with k 0 = 0.01, ε 0 = 10, α = 0.01 . Time evo-

lution of k (top) and ε (bottom). ___ exact solution, positive time step,
O stable time step with ‘ad hoc’ resetting, ∆ new proposed time step.
5.2. RAE2822 airfoil, Case10. A demanding test case is the transonic flow over a
RAE2822 airfoil. This case has been subject for several earlier investigations in Euroval
[5] and ECARP [6], the particular case investigated here is denoted Case 10 and involves
shock boundary layer separation. In this paper only aspects concerning robustness and
rate of convergence are considered although the case is very interesting concerning the
prediction of shock location, leading edge suction peak, skin friction etc.
The flow conditions are

6
M = 0.754, α = 2.57°, Re = 6.2 ×10 . (36)
The trailing edge of the airfoil is sharp and hence a C-type of grid was used with a
size of 257 × 65 nodes, 65 nodes from the wall to the outer boundary located about 10
chords away. A close up of the grid can be seen Figure 6. The distance to the second
–5 –5
layer of nodes from the wall varies from 0.3 ×10 at the leading edge to 1 ×10 at the
trailing edge. Transition is specified at 3% of the chord on both upper and lower side.
FIGURE 6: C-type of grid around the RAE2822 airfoil used in the calculation
with 257 × 65 nodes.
Unless otherwise stated 3 levels of multigrid with V-cycles is used in the computa-
tions as well as a higher order prolongation operator, higher order restriction operator,
the restriction damping factor β = 100 from Equation (25), everything initialized to
free stream values and a CFL number of CFL = 1.5 . No smoothing of the multigrid
corrections is applied. The rate of convergence is presented for the density and turbu-
lent kinetic energy residuals.
Without the new proposed time step (22) and correction relaxation (25) the computa-
tions starting from free stream values diverge immediately. The option to the new
approach would be to drastically reduce the CFL number initially, finding a better ini-
tial solution, leave out multigrid initially etc. All of these option imply additional time
and cost to find a converged solution which makes the proposed improvements supe-
rior. It is difficult though, to quantify the benefits from it since the options involve dif-
ferent procedures in getting a solution.
The rate of converge using 1-4 multigrid levels can be seen in Figure 7.
0.0
1 grid
2 grids
3 grids
4 grids
-2.5
Res
-5.0
-7.5
0. 2000. 4000. 6000.
Iter
0.0
1 grid
2 grids
3 grids
4 grids
-2.0
Res
-4.0
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 7: Rate of convergence using 1-4 multigrid levels. Top: density resi-
dual, bottom: k residual.
As can be seen the speed up is considerable. The residual of the turbulent kinetic
energy drops to a constant level from which there is no further decrease.
The ambition has not been to obtain the best possible convergence but to study delta
effects on the rate of convergence. The rate of convergence can be further improved by
optimizing CFL numbers, using residual smoothing, increasing numerical diffusion etc.
Although the density residual is reduced more than 7 orders of magnitude it is not
required for engineering accuracy, e.g. the integrated forces converge to its steady state
values within 0.1% in about 1500 iterations using 4 grid levels, see Figure 8
0.810
1 grid
2 grids
3 grids
4 grids
0.780
Cl
0.750
0.720
0. 2000. 4000. 6000.
Iter
0.0400
1 grid
2 grids
3 grids
4 grids
0.0350
Cd
0.0300
0.0250
0. 2000. 4000. 6000.
Iter
FIGURE 8: Rate of convergence using 1-4 multigrid levels. Top: lift coefficient
Cl, bottom: drag coefficient, Cd
The choice of the parameter β in the update of the multigrid corrections in Equation
(25) is investigated in Figure 9 using 4 grid levels. The rate of convergence is practi-
cally unchanged for different β provided β is chosen large enough. Divergence is

obtained for β = 2 which is obviously a too low value for this case.
In Figure 10 different restriction operators are used. With a lower order restriction
operator the residuals show a somewhat oscillatory behaviour using 3 grid levels
although convergence is obtained. With 4 grid levels there is no longer convergence
using 4 grid levels.
0.0
2
10
100
1000
-2.5
Res
-5.0
-7.5
0. 2000. 4000. 6000.
Iter
0.0
2
10
100
1000
-2.0
Res
-4.0
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 9: Rate of convergence using 4 grids and variation of

β = 2, 10, 100, 1000 . Top: density residual, bottom: k residual.
0.0
3, low
3, high
4, low
4, high
-2.5
Res
-5.0
-7.5
0. 2000. 4000. 6000.
Iter
0.0
3, low
3, high
4, low
4, high
-2.0
Res
-4.0
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 10: Rate of convergence using 3 and 4 multigrid levels with lower and
higher order restriction. Top: density residual, bottom: k residual.
The influence of using the full Jacobian in the determination for the local time step
instead of the spectral radius can be seen in Figure 11. The difference is very small
which justifies the use of the scalar spectral radius.
0.0
-2.5
Res
-5.0
-7.5
0. 2000. 4000. 6000.
Iter
0.0
-2.0
Res
-4.0
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 11: Rate of convergence with 3 multigrid levels using spectral radius
(straight line) and full Jacobian (dashed line). Top: density residual, bottom: k
residual.
In general multigrid is often less robust applied to the turbulent equations due to
their stiffness. In many industrial codes multigrid is therefore only applied to the mean
flow equations. In Figure 12 the difference between using multigrid and single grid for
the turbulence is plotted. As can be seen the rate of convergence for the density is not
affected much whereas the kinetic turbulent kinetic energy converges much slower.
This also indicates that the evolution of the mean flow and the turbulence is fairly
decoupled.
0.0
Orig.
1 grid tur
-2.5
Res
-5.0
-7.5
0. 2000. 4000. 6000.
Iter
0.0
Orig.
1 grid tur
-2.0
Res
-4.0
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 12: Rate of convergence with 4 multigrid levels. Comparison with

multigrid applied to all equations (solid line) and multigrid applied only to the
mean flow equations (dashed line). β = 1000 . Top: density residual, bottom:
k residual.
5.3. Multigrid applied to ‘bad grids’. In practical application for complex geom-
etry in three dimensions it not possible to avoid ‘bad grids’ with large stretchings, grid
singularity, high cell skewness etc. These distortions often lead to a degradation in the
quality of the solution and in the rate of convergence. In this section the influence of
different multigrid transfer operators on grids with large and sudden stretchings to the
rate of convergence is carried out.
To simulate a complex 3D grid with large and sudden stretchings, a less complex
problem was chosen. Every third point has been removed in the 2D grid of the
RAE2822 original grid of 257 × 65 points. Points are removed in the I-direction, in the
J-direction and in both directions. The grids can be seen in Figure 13. The grid sizes
and the sizes of the coarser grids in multigrid can be seen in Table 1. Note that for the
grid with points removed in both directions only 3 multigrid levels are possible and
hence used. Also note that removing points in one directions implies that the grid is
made coarser in only one direction between grid levels 3 and 4.
The flow conditions and the numerical conditions are the same as in the previous
Section 5.2 .
FIGURE 13: .Upper left: original grid. Upper right: grid with every third point
removed in I. Lower left: grid with every third point removed in J. Lower right:
grid with every third point removed in I and J.
TABLE 1: Size of the original grids and the grids with every third point
removed. Sizes of the coarser grids used in multigrid are displayed as well
(Grid 2). J is the direction normal to the airfoil to the outer boundary.
Original Removal in I Removal in J Removal in I,J

Grid Size Grid Size Grid Size Grid Size
1 257×65 1 173×65 1 257×45 1 173×45
2 129×33 2 87×33 2 129×23 2 87×23
3 65×17 3 44×17 3 65×12 3 44×12
4 33×9 4 44×9 4 33×12
The rate of convergence using the lower and higher order multigrid operators in Fig-
ure 1 and Figure 2 for the four grids in Figure 13 can be seen in Figure 14- Figure 17.
It is evident that the combination of higher order prolongation and restriction opera-
tors give the best convergence as well as convergence curves with least oscillations.
Lower order restriction or prolongation may prevent convergence in some cases. The
benefit from the higher order operators is most evident when removing points in the I-
direction and hence increasing the local cell aspect ratio. The combination of lower
order restriction and prolongation operators give divergence and can usually not be
used for problems with more than two multigrid levels.
Finally the rate of converge is plotted in Figure 18 using three multigrid levels for
the four different grids with the higher order restriction and prolongation operators. The
rate of convergence is improved when removing points normal to the wall (J) which
reduces the cell aspect ratio. By removing in the I-direction the convergence becomes
slower with higher cell aspect ratios. This is typical for explicit methods.
6. Summary and conclusions

A new way of keeping the turbulent dependent variables positive in the time integration
is proposed and evaluated. The local time step in the steady state time integration is
restricted by the local residual in such a way that the solution cannot turn negative. The
restriction is only initial and the asymptotic rate of convergence is not effected. A simi-
lar procedure is used in the update of the variables from the multigrid corrections. The
benefits of the proposed method is demonstrated on a model example and on the
RAE2822 airfoil at transonic speed.
A higher order restriction operator is demonstrated to give improved rate of conver-
gence in some cases. Especially when the grid has large local variations and the cell
aspect ratio is high it is shown that only with the combination of linear prolongation
and higher order restriction convergence is obtained.
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.5
Res
Lower restr.
-5.0
Lower prol.
Higher restr&prol
-7.5
0. 2000. 4000. 6000.
Iter
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.0
Res
Lower restr.
-4.0
Higher restr&prol
-6.0 Lower prol.
0. 2000. 4000. 6000.
Iter
FIGURE 14: Rate of convergence, original grid with 4 multigrid levels. Top:
density residual, bottom: k residual.
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.5
Lower prol.
Res
Lower restr.
-5.0
Higher restr.&prol.
-7.5
0. 2000. 4000. 6000.
Iter
0.0
R=1,P=1
R=0,P=1
R=1,P=0
Lower prol.
-2.0
Res
Lower restr.
-4.0
Higher restr.&prol.
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 15: Rate of convergence with points removed in I and 4 multigrid

levels. Top: density residual, bottom: k residual.
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.5
Res
-5.0 Lower prol.

Lower restr.
Higher restr.&prol.
-7.5
0. 2000. 4000. 6000.
Iter
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.0
Res
Lower restr.
-4.0
Lower prol.
Higher restr&prol
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 16: Rate of convergence with points removed in J, 4 multigrid levels.

Top: density residual, bottom: k residual.
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.5
res
Lower prol.
-5.0
Higher restr.&prol
Lower restr.
-7.5
0. 2000. 4000. 6000.
Iter
0.0
R=1,P=1
R=0,P=1
R=1,P=0
-2.0
Lower prol.
Res
-4.0
Higher restr.&prol.
-6.0 Lower restr.

0. 2000. 4000. 6000.
Iter
FIGURE 17: Rate of convergence with points removed in I and J, 3 multigrid

levels. Top: density residual, bottom: k residual.
0.0
Orig.
I-rem
J-rem
IJ-rem
-2.5
res
-5.0
I,J removed
I removed
Original
J removed
-7.5
0. 2000. 4000. 6000.
Iter
0.0
Orig.
I-rem
J-rem
IJ-rem
-2.0
res
-4.0
Original
I removed
I,J removed
J removed
-6.0
0. 2000. 4000. 6000.
Iter
FIGURE 18: .Rate of convergence for the four grids with higher order transfer
operators using 3 multigrid levels. Top: density residual, bottom: k residual.
Acknowledgment
This work has been carried out within the the AVTAC Project (Advanced Viscous
Flow Simulation Tools for Complete Civil Transport Aircraft). AVTAC is a collabora-
tion between British Aerospace, DASA, CASA, Dassault Aviation, SAAB, Alenia,
DLR, ONERA, CIRA, FFA and NLR. The project is managed by British Aerospace
and is funded by the CEC under the IMT initiative (Project Ref: BRPR CT97-0555)
References
[1] Almeida, G.P., Durao, D.F. and Heitor, M. V. (1993) “Wake Flows behind Two-
dimensional Model Hills”, Experimental and Thermal Fluid Science, vol. 7, pp. 87-
101.
[2] Chien, K.Y. (1982) “Prediction of Channel and Boundary Layer Flows with a Low
Reynolds Number turbulence Model”, AIAA Journal, Vol. 20, No. 1.
[3] Eliasson, P. (1993) “Dissipation Mechanisms and Multigrid Solutions in a Multi-
block Solver for Compressible Flow”, PhD thesis, TRITA-NA-R9314, ISSN-0348-
2952, Stockholm.
[4] Eliasson, P. and Engquist, B. (1995) “The Effects of Dissipation and Coarse Grid
Resolution for Multigrid in Flow Problems”, 7th Copper Mountain Conf. on Multigrid
Methods, NASA Conf. Publication 3339, Colorado.
[5] Haase, W., Bradsma, F., Elsholz, E., Leschziner, M. and Schwamborn, D. (1993)
“EUROVAL - An European Initiative on Validation of CFD Codes”, Notes on Numeri-
cal Fluid Mechanics, Vol. 42, Vieweg Verlag.
[6] Haase, W., Chaput, E., Elsholz, E., Leschziner, M.A. and Mueller U.R. (1997)
“ECARP - European Computational Aerodynamics Research Project: Validation of
CFD Codes and Assessment of Turbulence Models”, Notes on Numerical Fluid
Mechanics, Vol. 58, Vieweg Verlag.
[7] Hackbush, W. (1985) “Multi-Grid Methods and Applications”, Springer-Verlag.
[8] Harten, A. (1983) “High Resolution Schemes for Hyperbolic Conservation Laws”,
Journal of Computational Physics, vol. 49, pp. 357-393.
[9] Hundsdorfer, W., Koren, B. van Loon, M. and Verwer, J. (1995) “A Positive Finite-
Difference Advection Scheme”, Journal of Computational Physics, vol. 117, pp. 35-
46.
[10] Jongen, T. and Marx, Y.P. (1997) “Design of an Unconditionally Stable, Positive
Scheme for the Κ−ε and Two-Layer Turbulence Models”, Computers & Fluids, Vol.
26, No. 5, pp. 469-487.
[11] Lacor, C., Zhu, Z.W. and Hirsch, C. (1993) “A New Family of Limiters Within the
Multigrid/Multiblock Navier-Stokes code Euranus”, AIAA/DGLR 5th Int. Aerospace
Planes and Hypersonics Conf., Munich.
[12] Lien, F.S. and Leschziner, M.A. (1993) “Approximation of Turbulence Convection
in Complex Flows with a TVD-MUSCL Scheme”, Proc. 5th Int. Symp. on Refined
Flow Modelling and Turbulence Measurements, Paris, Sep. 7-10, 1993.
[13] Shu, C. (1988) “Total-Variation-Diminishing Time Discretizations”, SIAM J. Sci.
Stat. Comput., Vol. 9, No. 6.
[14] Rizzi, A., Eliasson, P., Lindblad, I., Hirsch, C., Lacor, C. and Haeuser, J. (1993)
“The Engineering of Multiblock \ Multigrid Software for Navier-Stokes Flows on
Structured Meshes”, Computers and Fluids, Vol. 22, pp. 341-367
[15] Wallin, S. (1996) “Numerical Discretization of Two-Equation turbulence Models
in Euranus”, FFAP-B-010, Internal report.
[16] Yee, H.C., Warming, R.F. and Harten, A. (1985) “Implicit Total Variation Dimin-
ishing (TVD) schemes for Steady-State Calculations”, Journal of Computational
Physics, vol. 57, pp. 327-360.

PHD 2000 Stefan Wallin

Uploaded by

Copyright:

Available Formats

PHD 2000 Stefan Wallin

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PHD 2000 Stefan Wallin

Uploaded by

Copyright:

Available Formats

Engineering turbulence modelling for CFD with a focus

on explicit algebraic Reynolds stress models

Akademisk avhandling som med tillstånd av Kungliga Tekniska Högskolan i

Descriptors: Turbulence model, Reynolds stress, passive scalar ﬂux, explicit

Paper 1. Wallin, S. & Johansson, A. V. (2000) An explicit algebraic

Paper 2. Lindblad, I.A.A., Wallin, S., Johansson, A.V., Friedrich, R.,

Paper 3. Wallin, S. & Girimaji, S.S. (2000) Evolution of an isolated turbu-

Paper 4. Wallin, S., Wang, D., Berggren, M. & Eliasson, P. (2000) A

Paper 5. Wikstrm, P.M., Wallin, S. & Johansson, A.V. (2000) Deriva-

Paper 6. Wallin, S. (1999) An eﬃcient explicit algebraic Reynolds stress k–ω

Chapter 2. Basic concepts in turbulence modelling 3

Chapter 3. The proposed modelling approach 16

Chapter 4. Conclusions and outlook 33

Basic concepts in turbulence modelling

2.1. Governing equations and Reynolds averaging

In practically all engineering cases, the turbulence needs to be considered

2.2. Eddy-viscosity models

2.2.2. Two-equation models. The major deﬁciency in algebraic and one-

P = 2νT Sij Sji . (9)

Finally the eddy viscosity is related to K and Z, for instance through

2.3. Reynolds stress models

improvements compared to eddy-viscosity two-equation models in which this

2.3.1. Diﬀerential Reynolds stress models (DRSM). In diﬀerential

for example, see Girimaji[13]. τ = K/ε is the turbulent timescale. Classical

Table 2.1 The values of the C coeﬃcients for diﬀerent quasi-linear

An alternative to using the Reynolds stress transport equations (5) is to re-

2.3.2. Algebraic Reynolds stress models (ARSM). In ﬂows where

This system is implicit in the Reynold stress anisotropy. It is also nonlinear

2.3.3. Explicit algebraic Reynolds stress models (EARSM). The

and may, in principal, be solved directly. However, by the inspection of (22)

ten tensorially independent groups

2.3.4. Nonlinear eddy-viscosity models (NL-EVM). A diﬀerent fam-

The β coeﬃcients should, therefore, be functions of the strain rate so that

2.4. Scalar ﬂux models

vector may symbolically be written as

The proposed modelling approach

Engineering problems are typically characterized by complexity in that dif-

3.1. The weak equilibrium assumption

-0.2 hom. shear

3.1.1. The consistency condition. Since the weak equilibrium assump-

Figure 3.2 Time evolution of the turbulent kinetic energy in rapidly

3.1.2. Slowly distorted turbulence. In slowly distorted turbulence where

3.1.3. Streamline curvature eﬀects. Turbulent ﬂows over curved sur-

D aij ∂aij ε (r) (r)

= + G (aij ) − aik Ωkj − Ωik akj (34)

where G contains spatial derivatives evaluated in a curvilinear coordinate system,

Ω(r) = θ̂r̂ − r̂θ̂ (35)

and the spatial derivative of the anisotropy components becomes zero G = 0.

local rotation vector may be obtained as

Table 3.1 The values of the A-coeﬃcients for diﬀerent quasi-linear

Figure 3.3 Computed evolution of the turbulent kinetic energy K in

Rotating homogeneous shear ﬂow is another example in which the A0 -coeﬃ-

Figure 3.4 The Reynolds stresses in channel ﬂow. Comparison of

3.2. Near-wall treatment

Figure 3.5 The velocity proﬁle for an adverse pressure gradi-

3.3. Numerical behaviour and some applications

3.3.1. Eﬀective eddy viscosity. The Reynolds stresses expressed by the

3.3.2. Positivity. It is well known that positivity of the turbulence vari-