Professional Documents
Culture Documents
Advanced Turbulence Modelling of Separated Flow in A Diffuser
Advanced Turbulence Modelling of Separated Flow in A Diffuser
Advanced Turbulence Modelling of Separated Flow in A Diffuser
81
© 2000 Kluwer Academic Publishers. Printed in the Netherlands.
D.D. APSLEY
Department of Civil & Structural Engineering, UMIST, Manchester M60 1QD, U.K.
M.A. LESCHZINER
Department of Engineering, Queen Mary and Westfield College, University of London,
London E1 4NS, U.K.
Abstract. The paper describes an investigation into the predictive performance of linear and non-
linear eddy-viscosity models and differential stress-transport closures for separated flow in a nom-
inally two-dimensional, asymmetric diffuser. The test case forms part of a broader collaborative
exercise between academic and industrial partners. It is demonstrated that advanced turbulence mod-
els using strain-dependent coefficients and anisotropy-resolving closure offer tangible advantages
in predictive capability, although the quality of their performance can vary significantly, depend-
ing on the details of closure approximations adopted. Certain features of the flow defy resolution
by any of the closures investigated. In particular, no model resolves correctly the flow near the
diffuser’s inclined wall immediately downstream of the inlet corner, which may reflect the pres-
ence of a “flapping” motion associated with a highly-localised process of unsteady separation and
reattachment.
Key words: turbulence modelling, separated flow, non-linear eddy-viscosity models, differential
stress models.
1. Introduction
Half a century of research in turbulence modelling has spawned numerous closure
models arising from many permutations of intuitive concepts, rational principles
and calibration practices. The largest, oldest and most widely-used class of models
is that combining the eddy-viscosity concept with the linear stress-strain relation-
ship and one or two transport equations for respective surrogates of the turbulent
velocity and length scales. The recognition that this type of model is afflicted with
serious weaknesses, especially in the presence of high curvature, swirl, separation,
stagnation and body forces, has led the turbulence-research community to search
for superior models which are based on a firm physical foundation.
The three principal routes taken over the past few years have been based on
non-linear stress-strain relations [7, 8, 11], transport equations for the stresses
[12, 16, 18, 28] and scalar fluxes and, most recently, on equations for turbulence-
structure parameters [26]. Although there is widespread recognition of the funda-
82 D.D. APSLEY AND M.A. LESCHZINER
mental strengths and superior predictive potential of such models, CFD practition-
ers in industry and vendors of commercial CFD software continue to rely heavily
on linear two-equation models, and have not, on the whole, taken more than tent-
ative steps towards adopting more advanced forms. Indeed, there has been a recent
trend in the aerospace industry (especially in the USA) towards even simpler one-
equation models [2, 27] which rely, intrinsically, on an empirical prescription of a
turbulent length scale. In principle, such models are suited to flows which are close
to equilibrium and in which a single shear stress dominates the linkage between
the mean flow and turbulence.
The reluctance of CFD practitioners to adopt advanced models is rooted in a
number of practical considerations. Greater mathematical and coding complexity,
lower numerical robustness, more stringent grid-quality requirements and more
extensive reliance on boundary conditions for turbulence quantities are some of
the reasons often put forward, and most of these are, indeed, based in reality.
Perhaps the most powerful and potentially damaging argument arises, however,
from the perception that there is no convincing evidence that advanced models per-
form consistently better than carefully-tuned simple models over a broad range of
conditions. This perception is heightened by the outcome of a number of influential
workshops (e.g. [6, 15]) which involved, for prescribed test flows, comparisons of
computational solutions contributed by a number of participants using their own
codes, grids, numerical practices and model implementations. Clearly, within such
loosely-controlled exercises, there is extensive potential for variability in predictive
performance, which may be unrelated to the intrinsic properties of the turbulence
models supposedly being investigated. Thus, occasionally, if not usually, the only
conclusion that can safely be extracted from an extensive exercise of this type is
that it is inconclusive and that an interactive, more carefully-controlled initiative
would be required to remove inexplicable features which had come to light in the
comparisons undertaken. In most circumstances, any follow-up activity is limited
in scope and not very fruitful. The overall damaging consequence is an even greater
level of scepticism and uncertainty than before, and a further strengthening in the
reluctance of practitioners to adopt advanced modelling practices.
The urgent need perceived by industry to gain a much clearer view of the relative
performance of different turbulence models in challenging conditions akin to those
encountered in practice led, in 1997, to the formation of a consortium of industrial
partners (BAE SYSTEMS, Aircraft Research Association, Rolls-Royce, Defence
Evaluation and Research Agency) and academic groups (UMIST, Loughborough
University of Technology) with the goal of undertaking systematic, collaborative,
computational investigations of a selection of well-documented flows spanning
a broad range of flow-physical and geometric features. The consortium, work-
ing within the VoTMATA (Validation of Turbulence Models for Aerospace and
Turbomachinery Applications) project, seeks to provide an objective assessment of
existing turbulence models, in various classes, for flows involving strong curvature,
separation from smooth surfaces, 3-d strain, strong vorticity, impingement and
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 83
shocks. The first three test cases – a 2-d asymmetric diffuser, a transonic axisym-
metric bump and a 3-d wing-body junction – are comparatively simple in terms
of geometry, and are characterised by detailed and accurate experimental data. A
large number of turbulence models have been used to calculate each test case under
strictly controlled numerical conditions.
This paper focuses on the first of the “fundamental” test cases: the asymmetric
plane diffuser of Obi et al. [24]. This flow is of considerable interest to turbulence
modellers, and has featured as one of three test cases in a recent IAHR/ERCOFTAC
Workshop [15]. Besides the original computations of Obi et al. with high-Re k–ε
and differential stress models, it has also been computed by Durbin [9] with the k–
ε–v 2 model which was reported to perform well. The attraction of this test case lies
in its geometric simplicity, its virtually two-dimensional state, the availability of
two sets of closely matching experimental data for mean velocity, turbulence and
surface coefficients, obtained in two separate facilities [4, 24], and the existence
of LES results [10] which further support the validity of the experimental data.
The test case presents a challenging example of separation from a plane surface
in an adverse pressure gradient. The experimental configuration is described in
Section 2.
For geometries where separation is not fixed by sharp corners, determining the
separation line depends on modelling the correct response of the shear stress and,
to some degree also the normal stresses, to deceleration; that is, of predicting the
correct turbulent time scale. Beyond the separation point, the flow loses much of
the direction-constraining influence of the boundary, and all components of the
Reynolds-stress tensor become dynamically significant. At the same time, the rate
of recovery and, hence, the reattachment point, depend on the magnitude of the
shear stress in a (curved) free shear layer. The challenges for turbulence models in
such a flow are, therefore, (i) to predict the correct turbulent time scale in strongly
accelerated wall and free shear layers; and (ii) to predict the correct anisotropy of
the Reynolds-stress tensor.
The turbulence models applied to this test case have been somewhat loosely
categorised into four classes. The minimum degree of complexity is represented
by two-equation, constant-coefficient, linear eddy-viscosity models (EVMs). A
second class of models is formed from the hybrid k–ω/k–ε models of Menter
[23], including the shear-stress transport (SST) scheme. The last two classes
are populated by anisotropy-resolving closures: non-linear eddy-viscosity models
(NLEVMs) and differential stress models (DSMs). Throughout, there are various
choices of turbulence scalars (k–ε, k–ω, k–g) and wall boundary conditions. The
turbulence models and a theoretical assessment of model differences and capab-
ilities are given in Section 3. Section 4 contains a detailed comparison of model
performance against the experimental data. Section 5 attempts to summarise what
we have learnt.
84 D.D. APSLEY AND M.A. LESCHZINER
Figure 2. Finite-volume mesh used for low-Re calculations; (every other grid line omitted for
clarity).
cell height 0.001H ) for low-Re calculations and 292 × 56 control volumes for
high-Re calculations with wall functions. In the latter case, low-Re-grid cells at
less than 5% channel height from each wall were amalgamated, without changing
the interior disposition of control volumes. The near-wall grid sizes were chosen
to obtain a y + value at cell-centres (based on a log-law fit to the inflow velocity
profile) of 1 for low-Re and 25 for high-Re calculations.
Inflow profiles were determined for each turbulence model by a preliminary
computation (with the same y grid) of fully-developed channel flow, with the
same mass flow rate as the experimental data. The remaining boundary conditions
were zero longitudinal gradients for all transport variables at outflow and no-slip
conditions at the diffuser walls.
3. Turbulence Modelling
3.1. L INEAR EDDY- VISCOSITY MODELS
The simplest form of turbulence closure that preserves the symmetries and tensorial
nature of the Reynolds stresses is a linear eddy-viscosity model (EVM), which
assumes a simple proportionality between deviatoric stress and mean strain:
2 1
ui uj − kδij = −2vt Sij − δij Skk , (1)
3 3
where
1 ∂Ui ∂Uj
Sij = + .
2 ∂xj ∂xi
vt = u0 l0 = u20 t0 , (2)
γ A m n Cγ 1 Cγ 2 Cε1 Cε2
where
Cγ 1 − n Cγ 2 − n
Cε1 = , Cε2 = (9)
m m
and (if ε = εabs )
ε 1 1 nε 1 1
Sε = Sγ + ∇ · − vt ∇ε + ∇· − vt ∇k
γm σγ σε mk σγ σk
" 2 2 #
ε vt ∇ε ∇k 2 ∇ε ∇k
+ v+ m +n −m −n . (10)
m σγ ε k ε k
Clearly, all two-equation turbulence models of this type can be interconverted. The
corresponding values of Cε1 and Cε2 deduced from the coefficients of k–ω and k–g
models are given in Table I.
The additional source term consists of parts due to different turbulent Prandtl
numbers and to cross-diffusion. Both stem from the original diffusion model and do
not contain mean-velocity gradients explicitly. In general, it is difficult to establish
the precise effect of the cross-diffusion terms which may be of either sign. More
clear-cut, however, is the effect of the difference Cε2 –Cε1 , particularly as regards
the response to adverse pressure gradients and the prediction of separation: the
smaller the difference between these constants, the greater is ε and hence the less
is the turbulence intensity. Experience shows that the shear stress is extremely sens-
itive to this difference, even though both k–ε and k–ω models satisfy the log-law
in equilibrium flows (for an appropriate choice of σε and σω ). Although the role of
the cross-diffusion terms is not easily analysed, the values of Cε1 and Cε2 in Table I
help to explain why k–ω models (with their standard calibration) are more likely
than k–ε models to predict separation in adverse pressure gradients and, conversely,
why k–ω models tend to over-predict reattachment length in bluff-body separated
flows.
The eddy-turnover time scale is t0 ∼ k/ε. Mean-velocity gradients give rise to
a second (“shear”) time scale, T0 ∼ kSk−1 . Thus, from (1) and (2),
namely that Cµ should vary as T0 /t0 or the reciprocal of the strain rate for large
strains. It is not clear why Menter should have used mean vorticity rather mean
strain, or a combination of the two. There is no difference in boundary layers
(∂U/∂y being the dominant component), but a clear distinction arises in largely
irrotational flow regions, such as near impingement points.
Linear EVMs have a single degree of freedom (the scalar vt ) and, therefore, can
only be justified in flows where a single Reynolds-stress component is dynamically
significant; in fully-attached, near-equilibrium shear flows, for example. However,
by generalising the stress-strain relationship, it is possible to mimic the response
of turbulence to complex strain, whilst still retaining a local, one- or two-equation
turbulence model.
The general stress-strain relationship may be written in non-dimensional form
(with second-rank tensors denoted by bold type, T, contracted products indicated
by juxtaposition as for matrix multiplication, (AB)ij = Aik Bkj , and traces denoted
by {T} = Tkk ):
a = −c1 s
1 2 1 2
+ c2 s − {s }I + c3 (ωs − sω) + c4 ω − {ω }I
2 2
3 3
2
+ c5 ω2 s + sω2 − {ω2 }s − {ωsω}I + c6 (ωs2 − s2 ω)
3
90 D.D. APSLEY AND M.A. LESCHZINER
1 2 2 2 2
+ c7 ω s + s ω − {ω } s − {s }I − {s ω }I
2 2 2 2 2 2
3 3
1
+ c8 s2 ωs − sωs2 − {s2 }(ωs − sω)
2
1 2
+ c9 ωsω − ω sω − {ω }(ωs − sω)
2 2
2
+ c10 (ωs2 ω2 − ω2 s2 s), (19)
where successive lines contain tensor bases of increasing degree in s and ω, and
the coefficients ci are functions of the irreducible invariants {s2 }, {ω2 }, {s3 }, {sω2 }
and {s2 ω2 }. In 2-d, incompressible flow, one finds that
1 2
s2 = (s11
2
+ s12
2
)I2 = {s }I2 ,
2
1
ω2 = −ω12
2
I2 = {ω2 }I2 , (20)
2
where I2 = diag(1, 1, 0), so that, with the (slightly complex) basis chosen here, the
cubic and higher-degree terms vanish. Therefore, a quadratic model is sufficient to
compute a 2-d (incompressible) flow. Note also that, in incompressible flow,
P /ε = −{as} (21)
and the quadratic terms do not contribute to the production of turbulence energy.
As pointed out by a reviewer, all ten basis elements are needed for mathemat-
ical completeness in 3-d flows. However, the stress-strain relationship can generate
normal-stress anisotropy and provide a qualitatively correct response of turbulence
to curvature and swirl if terms up to the cubic are retained [7]. If one adheres to the
principle that pure rotation (s = 0, ω 6 = 0) generates no anisotropy, then c4 = 0.
At cubic level, the number of remaining free coefficients (c1 , c2 , c3 , c5 , c6 ) then
matches the five degrees of freedom of the symmetric, traceless tensor aij . This is
the level marked by the horizontal line in Equation (19).
Different terms in (19) can be identified as yielding particular strain sensitivity.
The first term in (19) corresponds to a linear EVM, with c1 = Cµ fµ . The Craft
et al. [7] and Apsley and Leschziner [1] models include parts proportional to {s2 }s
and {ω2 }s (here regarded as parts of the linear term). In curved shear flow,
∂U U k 2
{s } − {ω } = −2
2 2
, (22)
∂R R ε
where R is the local radius of curvature. The inclusion of these components, there-
fore, yields an important sensitivity to mean streamline curvature. For simple shear,
the linear term makes no contribution to the normal stresses, but the quadratic terms
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 91
give
1
a11 = (c2 + 6c3 − c4 )σ 2 ,
12
1
a22 = (c2 − 6c3 − c4 )σ 2 ,
12
1
a33 = − (c2 − c4 )σ 2 , (23)
6
where σ = (k/ε)(∂U/∂y), and are responsible for normal-stress anisotropy. In 3-d
flow, the cubic terms can be shown to yield a sensitivity to swirl.
Results for three models are reported here: Gatski and Speziale [11], Craft et
al. [7] and Apsley and Leschziner [1]. The first is quadratic and is a high-Re k–ε
model; the last two are cubic, low-Re k–ε closures. The coefficients in the stress-
strain relationships are given in the Appendix. For details on the ε equations the
reader is referred to the original papers.
Differential stress models (DSMs) solve modelled forms of the exact transport
equations:
D ∂
ui uj = dij k + Pij + 8ij − εij . (24)
Dt ∂xk
Whilst the production terms Pij (which transfer energy from the mean flow) and
the advection terms (which incorporate history effects) are exact, the modelling
dilemma is increased over that for the single turbulent scalar k by the need to
model diffusion dij k and dissipation εij for each stress component, as well as the
pressure-strain correlation 8ij , which redistributes energy amongst components.
In addition, numerical stability can be compromised by the lack of a natural dif-
fusivity, although this can be offset by the addition and subtraction of “effective”
eddy viscosities in the momentum equations.
The majority of DSMs use fairly simple models for diffusion, most notably the
generalised gradient-diffusion hypothesis:
k ∂
dij k = vδkl + Cs uk ul ui uj , (25)
ε ∂xl
and incorporate any anisotropy of εij into 8ij , so that
2
εij = δij ε. (26)
3
A notable exception is the model of Jakirlić and Hanjalić [16]. The major differ-
ences between DSMs are the modelling of 8ij and the transport equation for (or
other means of specifying) ε.
92 D.D. APSLEY AND M.A. LESCHZINER
Based on the exact integral expression for 8ij , the pressure-strain correlation is
typically decomposed into “slow”, “rapid” and “wall-reflection” parts:
Whereas the exact expression involves volume and surface integrals, the modelled
form makes use of the local stress and strain tensors. Models examined here can be
written in the form
0 2 2
8 /ε = −C1 a − C1 a − {a }I ,
(1) 2
3
. .
1 1
8 /ε = −C2 P − {P}I
(2)
ε − C3 D − {D}I ε − C4 s,
3 3
2
= c01 s + c11 sa + as − {as}I + c12 (ωa − aω),
3
3 3
8(w)
ij /ε = 8̃kl nk nl δij − 8̃ik nj nk − 8̃j k ni nk , (28)
2 2
where
ε k 3/2 /ε
8̃(w)
ij = C1(w) ui uj + C2(w) 8(2)
ij f, f = ,
k Cl yn
∂Uk ∂Uk
Dij = − ui uk + uj uk . (29)
∂xj ∂xi
With the exception of the Speziale et al. [28] model, all closures tested here use
simple return-to-isotropy (C10 = 0) and isotropisation-of-production (C3 = C4 =
0) approximations, in conjunction with wall-reflection corrections.
Whereas the 8(1) part promotes a return to isotropy, the 8(w)
ij part is designed
to yield the correct anisotropy in the equilibrium layer near walls. The presence
of wall-normal vectors is undesirable, however, and much effort has been made to
eliminate this term, through the use of more complex models for 8(2)ij , and replace
the ni by wall-direction/distance indicators based on the local turbulent length-
scale gradient [5].
Results with four models are reported here: the high-Re models of Gibson and
Launder [12] and Speziale et al. [28] and the low-Re closures of Jakirlić and
Hanjalić [16] and Wilcox [31]. The Speziale et al. closure employs a non-linear
pressure-strain model, with no wall-reflection terms. The Jakirlić and Hanjalić
closure includes anisotropic dissipation and pressure-strain coefficients which are
functions of the stress and dissipation “flatness” factors, A and E, respectively.
The Wilcox multiscale model is ω-based and constructed on the premise that one
can partition the turbulence spectrum into large-scale, energy-bearing eddies and
small-scale, isotropic, dissipative eddies. This construction can be transformed
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 93
Important conclusions about turbulence models can only be drawn after all signific-
ant sources of numerical contamination have been eliminated. Thus, an important
aspect of collaborative testing is to ensure grid-independent and code-independent
solutions.
Computations at UMIST were performed with the in-house finite-volume code
STREAM, employing the SIMPLE pressure-correction algorithm with a cell-
centred, co-located storage arrangement and a second-order, three-point, TVD
advection scheme. The time-independent, incompressible-flow equations were
solved iteratively by standard line-iteration/tri-diagonal matrix solvers. Solution
for the pressure was accelerated by ensuring inflow/outflow mass conservation at
each outer iteration and employing block-correction techniques prior to directional
sweeps in the pressure-correction equation. Comparable pressure-correction meth-
ods were used by Loughborough University and Rolls-Royce. ARA and DERA
used their own SAUNA code, which employs cell-vertex storage, Jameson-style
artificial dissipation and explicit time-marching to steady state. BAE SYSTEMS
used a density-based, Jameson-type, time-dependent, compressible code at very
low Mach number.
Figure 3 shows profiles of streamwise mean velocity and Reynolds shear stress
computed on the primary grid (292 × 96 control volumes) and grids where the
mesh density was halved or doubled in both coordinate directions. Computations
were performed with the Wilcox [30] linear k–ω model, which is observed to give a
moderate degree of separation. A fourth mesh, with reduced near-wall cell height,
was also investigated to examine sensitivity to the numerical implementation of the
ω boundary condition. The results shown confirm that, for low-Re models, the ad-
opted grid was sufficient for numerical accuracy. However, for models employing
wall functions to bridge the viscous sublayer, the minimum y + criterion restricts
the level of grid refinement permissible near boundaries.
Inter-code comparisons on the same grid (Figures 4 and 5) show a moderate
degree of scatter. The figures show mean-velocity profiles computed by different
organisations with linear k–ε [19] and linear k–ω [30] models. Included in the
latter are results from the BAE SYSTEMS k–g model, since, for internal flows,
this is a direct transformation of the Wilcox model. The findings are similar to
those of a recent ERCOFTAC workshop [15]; namely, that the level of agree-
ment between different organisations is substantially better with the k–ω model
than with the Launder–Sharma scheme. This probably reflects the complexity of
the length-scale-determining transport equation. Whilst the Wilcox model makes
no concession to special viscous damping terms, the Launder–Sharma closure
94 D.D. APSLEY AND M.A. LESCHZINER
Figure 3. Grid-dependence tests with Wilcox k–ω model: (a) mean velocity; (b) shear stress.
Figure 6 shows the development of mean-velocity and Reynolds shear stress com-
puted along the diffuser with representative k–ε [19] and k–ω [30] models, and
with the SST model [23]. Comparisons with other high and low-Re k–ε [20, 22]
and k–ω [32] models show the results for this test case to be insensitive to the
near-wall viscous-damping terms, although not to the high-Re source/production
coefficients Cε1 and Cε2 in the ε equation.
The k–ε model fails to predict separation, producing a nearly symmetric mean-
velocity profile across the diffuser. The k–ω model predicts separation at x/H ≈ 5,
but only weak recirculation. The best overall velocity profiles are recorded with
the SST model, but in this case separation occurs far too early (x/H ≈ 2, com-
pared with an experimental value of 7). The Obi et al. data set was not extended
sufficiently far downstream to examine recovery after reattachment; however, com-
parison with the related Buice and Eaton data shows an insufficient rate of recovery,
a typical finding with Reynolds-averaged models applied to separated flow.
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 95
Comparison with experiment in Figure 6 shows that all models predict an an-
omalous retardation of the boundary-layer flow and peak in the shear stress (and,
indirectly, the other stress components) following the first corner. This is not seen
in the LES results [10]. A possible suggestion is that the near-wall shear stress fol-
lowing the corner might be influenced by periodic “flapping” or coherent structures
following minor separation at the corner. This organised motion is not captured
by any Reynolds-averaged Navier–Stokes (RANS) model, but is a very effective
mixer. Because the modelled shear stress is too low at the wall at the first corner,
the boundary layer tends to be closer to separation (stronger inflexion and larger
displacement). This leads to higher velocity gradients away from the wall (due to
excessive displacement – an inviscid feature) and hence to the anomalous turbu-
lence peak. As the flow progresses, this excessive shear stress leads to excessive
velocity-profile flattening, by turbulent mixing, and hence to the reduced back-
flow velocity. This behaviour, which is apparent to a greater or lesser extent in all
model results, is rooted in the models’ inability to capture crucial processes just
downstream of the first corner.
A more detailed analysis of the SST calculation is shown in Figure 7. In Fig-
ure 7a the interpolation factor F1 (Equation (14)) is plotted, with a contour marking
the value 0.5. Since F1 = 0 and F1 = 1 correspond to k–ε and k–ω models, re-
96 D.D. APSLEY AND M.A. LESCHZINER
spectively, it is evident that the model is of k–ω type over a greater part of the flow,
and, in particular, within the important near-wall region. This observation, which is
typical of fully-developed internal flows, but not external flows with a free stream,
explains why calculations with Menter’s baseline model [23] (the hybrid model
without the eddy-viscosity limiter) gives results almost identical to the standard k–
ω model. Figure 7b shows the value of the limiter F2 /aω (Equation (15)), a ratio
of turbulent to shear time scales. The limiter only has an effect when its value is
greater than unity. The figure shows that this occurs in a relatively small region of
the flow: crucially, in the boundary layer just downstream of the first corner, where
it causes separation.
Figure 8 shows derived quantities for eddy-viscosity models: the kinematic
eddy viscosity vt , turbulent kinetic energy k and a “mixing” length l, defined
(independently of the second turbulence variable) by vt = (Cµ1/4 k 1/2 )l. The nor-
malisation is such that l = κyn in the log layer. The plotted models are linear
k–ε, k–ω and SST k–ω models, together with the NLEVM of Craft et al. [7]. Two
features are worthy of note from the inflow profiles (x/H = −11); namely, that
the eddy viscosity is fairly small (the maximum value of vt /v is only about 50)
and that, whilst l/κyn is almost constant across much of the domain, its value is
slightly less than 1.0. The peaks in turbulence energy increase and move towards
the centre of the channel as the flow progresses along the diffuser. Peaks in l/κyn
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 97
Figure 6. Linear eddy-viscosity model calculations of mean velocity and Reynolds shear
stress.
appear relatively close to both walls, being much larger for k–ε than k–ω models.
The length scale predicted by the linear k–ε model increases rapidly along the
diffuser. In contrast, the length-scale ratio for the NLEVM remains limited in a
range roughly 1.0–2.0.
Figure 7. SST model parameters: (a) interpolation factor F1 ; (b) eddy-viscosity limiter
F2 /aω.
in this region causes greater mixing further downstream. The excessive degree of
anisotropy and anomalous peak in u2 are a direct consequence of the large velocity
gradients away from the wall just downstream of the corner.
Whilst the non-linear k–ε models tested clearly show improved predictive
capabilities relative to their linear counterparts, this cannot be attributed to their
non-linear stress-strain relationship per se, since all feature strain-dependent coef-
ficients. Some effort has been made, therefore, to distinguish the effects of a
strain-dependent Cµ from those due to non-linear terms in the Craft et al. model
(Figure 10). Velocity and stress profiles computed with the complete model and
with the successive removal of quadratic and “curvature” terms (({s2 } − {ω2 })s –
regarded as cubic in the original paper) show that, whereas the quadratic terms
are clearly required to predict anisotropy, the major effect on mean-flow dynamics
is due to the linear term and, in this test case, the curvature-related terms are un-
important. The inclusion of quadratic terms increases anisotropy and, thereby, the
gradients ∂u2 /∂x and ∂v 2 /∂y in non-equilibrium shear flow, leading to a slightly
greater tendency to separation.
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 99
Figure 8. Derived quantities for eddy-viscosity models; (a) eddy viscosity; (b) turbulent
kinetic energy; (c) mixing length.
100 D.D. APSLEY AND M.A. LESCHZINER
Figure 10. Relative effect of variable Cµ and non-linear terms in Craft et al. cubic k–ε model.
102 D.D. APSLEY AND M.A. LESCHZINER
Figure 11. Differential stress model calculations of mean velocity and Reynolds stresses.
Figure 11 shows velocity and stress profiles for various differential stress mod-
els: the high-Re models of Gibson and Launder [12] and Speziale et al. [28],
and the low-Re closures of Jakirlić and Hanjalić [16] and Wilcox [31]. There is
a clear distinction between the wall-function calculations, for which the mean-
flow predictions are comparable to those with anisotropic eddy-viscosity closures,
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 103
and the low-Re calculations, which fail to predict separation. In this instance,
despite its manifest incompatibility with non-equilibrium boundary layers, the
wall-function approach, which fixes the turbulent length-scale and anisotropy com-
ponents in the near-wall region, clearly outperforms low-Re schemes calibrated in
near-equilibrium flows.
The relatively poor performance of the Jakirlić–Hanjalić model is surprising,
for it is, in essence, a low-Re extension of the Gibson–Launder model (albeit with
slightly different coefficients). Moreover, the model originators have themselves
reported favourable predictive properties of the model over a wide range of flows,
including separation. Yet, the present results are supported, qualitatively, by cor-
responding weaknesses observed in shock-affected separation [3] with an entirely
different numerical procedure based on a Riemann solver. There are, however,
variants of the Jakirlić–Hanjalić model, distinguished by additional terms in the
ε equation. In particular, Hanjalić et al. [13] use a variant which includes two extra
source terms, intended to improve the performance of the model in non-equilibrium
flows:
Sε0 = Sl + Sn , (30)
where
3/2
εε̃ k
Sl = max[(γ − 1)γ , 0] A,
2 2
γ = ∇ , (31)
k 2.5ε
∂Us ∂Un ε
Sn = −1.16 u2s + u2n . (32)
∂xs ∂xn k
Figure 12. Effect of additional ε source terms in the Jakirlić and Hanjalić differential stress
model.
Figure 13. Effect of varying Cε2 in the SSG differential stress model.
In general, the high-Re DSM predictions of the Reynolds stresses are better
than those with NLEVMs. Whilst both predict comparable u2 /v 2 anisotropy, the
DSMs correctly predict the general increase in turbulence energy along the diffuser.
However, there are abnormal levels of normal stress and an anomalous peak in
Reynolds shear stress apparent at the first measurement station.
Surface pressure and skin-friction coefficients are plotted for representative linear
EVM, NLEVM and DSM in Figure 14.
The pressure coefficient cP (which is only available for the plane wall from
Obi et al.’s data, but for which computations suggest that the distribution is not
significantly different on the opposite wall) reflects the shape of the recirculating
flow region (i.e. the distance between plane surface and separation streamline). The
plot suggests that the eddy-viscosity closures predict the depth of the recirculating
flow fairly well, but that the SSG model (with Cε2 = 1.83) yields too deep a region
of backflow and too great a pressure loss due to mixing along the diffuser.
Obi et al. provide no data for skin friction. However, such data were obtained
by Buice and Eaton to assist in establishing separation and reattachment points. A
106 D.D. APSLEY AND M.A. LESCHZINER
Figure 14. Pressure and skin-friction coefficients: (a) pressure coefficient on plane surface
(experimental data from Obi et al.); (b) skin friction on inclined surface (experimental data
from Buice and Eaton).
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 107
5. Conclusions
This paper conveys a flavour of the difficulties involved in arriving at a reas-
onably secure statement on the relative performance of turbulence models in a
non-trivial, though by no means very complex, flow. In fact, it hides considerably
more than it reveals in terms of the magnitude of the interaction between partners
contributing to this exercise and the numerous iterations needed to illuminate the
sources of inconsistencies and inexplicable differences which characterised early
comparisons. Hence, a first conclusion of the study is that a closely-coordinated
and properly-supported programme, with careful attention to numerical issues and
boundary conditions, is a key to gaining an objective view of the intrinsic predictive
capabilities of alternative closure strategies.
A second conclusion is that no class of models can be said, without qualification
and purely on the basis of agreement with experiment, to be clearly superior or in-
ferior to others. Differential stress models are, of course, fundamentally superior to
eddy-viscosity models in terms of their ability to represent the complex interaction
between different types of strains and all the stress components, but this does not,
in itself, secure superior agreement with flow properties of practical interest in any
but the simplest flows.
In judging the relative merits of models, one issue which poses some uncer-
tainty is the possibility, conjectured on the basis of defects observed consistently
across the entirely range of models investigated, of periodic shear-layer instabilities
provoked by the upstream corner of the diffuser. The existence of such a process
would explain a general failure of all models to resolve the initial development of
the shear layer along the inclined diffuser wall.
108 D.D. APSLEY AND M.A. LESCHZINER
Acknowledgements
This paper would not have emerged without the substantial and sustained efforts
made by the partners of the VoTMATA project, in terms of the provision of com-
putational solutions, technical discussions and financial support. Thanks go, in
particular, to Dr. Alan Gould and Dr Stephen Moir (BAE SYSTEMS), Professor
Jim McGuirk and Dr. R.G.M. Hasan (Loughborough University of Technology),
Dr. Nadji Chioukh (Rolls-Royce), Dr. Nick May (ARA) and Dr. Tony Hutton
(DERA).
The financial support of EPSRC, BAE SYSTEMS, Rolls-Royce and DERA is
gratefully acknowledged. Some of the computations included in the paper have
been performed on the SGI Origin 2000 computer (Fermat) at the CSAR national
facility at Manchester Computing, with resources granted by EPSRC.
(1 + 2ζ 2 )(1 + η4 ) + 23 η2 α3,2
c2,3 = , (33)
(1 + 2ζ 2 )(1 + 2ζ 2 + β3,2 η6 ) 2
1
η = (α3 /α1 )(sij sij )1/2 , ζ = (α2 /α1 )(ωij ωij )1/2,
2
α1 = 0.227, α2 = 0.0424, α3 = 0.0397,
β1 = 7.0, β2 = 6.3, β3 = 4.0. (34)
110 D.D. APSLEY AND M.A. LESCHZINER
0.3[1 − exp(−0.36e0.75s̃ )]
Cµ = ,
1 + 0.35s̃ 3/2
fµ = 1 − exp[−(Rt /90)1/2 − (Rt /400)2 ],
s = (2sij sij )1/2 , ω = (2ωij ωij )1/2 , s̃ = max(s, ω). (36)
References
1. Apsley, D.D. and Leschziner, M.A., A new low-Reynolds-number non-linear two-equation
turbulence model for complex flows. Internat. J. Heat Fluid Flow 19 (1998) 209–222.
2. Baldwin, B.W. and Barth, T.A., One-equation turbulence transport model for high Reynolds
number wall bounded flows. AIAA Paper 91-0610 (1991).
3. Batten, P., Craft, T.J., Leschziner, M.A. and Loyau, H., Reynolds stress transport modelling for
compressible aerodynamic flows. AIAA J. 37 (1999) 785–796.
4. Buice, C.U. and Eaton, J.K., Experimental investigation of flow through an asymmetric plane
diffuser. Report TSD-107, Department of Mechanical Engineering, Stanford University (1997).
ADVANCED TURBULENCE MODELLING OF SEPARATED FLOW IN A DIFFUSER 111
27. Spalart, P.R. and Almaras, S.R., A one-equation turbulence model for aerodynamic flows,
AIAA Paper 92-0439 (1992).
28. Speziale, C.G., Sarkar, S. and Gatski, T.B., Modelling the pressure-strain correlation of
turbulence: An invariant dynamical systems approach, J. Fluid Mech. 227 (1991) 245–272.
29. Speziale, C.G. and Xu, X-H., Towards the development of second-order closure models for
nonequilibrium turbulent flows. Internat. J. Heat Fluid Flow 17 (1996) 238–244.
30. Wilcox, D.C., Reassessment of the scale-determining equation for advanced turbulence models.
AIAA J. 26 (1988) 1299–1310.
31. Wilcox, D.C., Multiscale model for turbulent flows. AIAA J. 26 (1988) 1311–1320.
32. Wilcox, D.C., Simulation of transition with a two-equation turbulence model. AIAA J. 32 (1994)
247–255.