Professional Documents
Culture Documents
1. INTRODUCTION
dynamics that describes the flow of fluids (liquids and gases). It has several
subdisciplines, including aerodynamics (the study of air and other gases in motion)
and hydrodynamics (the study of liquids in motion). Fluid dynamics has a wide
weapon detonation.
measurement and used to solve practical problems. The solution to a fluid dynamics
problem typically involves the calculation of various properties of the fluid, such
as flow velocity, pressure, density, and temperature, as functions of space and time.
applied to gases.
CHAPTER 2
PRELIMINARIES
UL UL
Re = = ,
v
Where, U, L, and are some characteristics values of the velocity, length, density
Definition: Turbulence
condition of the flow. It is the usual state of fluid motion except at low Reynolds
numbers.
Definition: Laminar Flow
Definition: Isotropic
Example: u v w .
Definition: Homogeneous
Example: independent of x .
Smallest scales in the spectrum that form the viscous sub-layer range. In this
range, the energy input from nonlinear interactions and the energy drain from viscous
dissipation are in exact balance. The small scales have high frequency, causing
measure of the rate of divergence of two trajectories that were initially nearby.
Definition: Eddy viscosity
(Reynolds) stresses and mean (large-scale) strain rate, analogous to physical viscosity
Kolmogorov scale is another name for dissipation scales. These scales were
Definition: RANS
essentially the opposite of DNS, namely, nearly all scales of the solution must be
modeled under a RANS formalism; only (time) mean quantities are directly
computed.
Reynolds stresses arise from the averaging procedure used in deriving the
RANS equations. They are the additional unknowns that create the closure problem.
Details of the derivation of the RANS equations will be given in Chap.-2 where
Transition refers to a change from one qualitative state of flow to another, and
turbulent.
1 |()|
= lim lim
0 0 |0 |
The limit 0 0 ensures the validity of the linear approximation at any time.
1
1
(0 ) = lim | ( )|
=0
the attractor (or attractors) of a dynamical system, and there will normally be one set
of exponents associated with each attractor. The choice of starting point may
determine which attractor the system ends up on, if there is more than one. Note:
Hamiltonian systems do not have attractors, so this particular discussion does not
apply to them.) The Lyapunov exponents describe the behaviuor of vectors in the
tangent space of the phase space and are defined from the Jacobian matrix
(
()
) = |
The matrix describes how a small change at the point x0 propagates to the
( ) = ( . ( ))
defines L(x0) a matrix (the conditions for the existence of the limit are given by
the Oseledec theorem). If i(x0) are the eigenvalues L(x0) of , then the Lyapunov
(0 ) = log (0 )
The set of Lyapunov exponents will be the same for almost all starting points
The plane couette flow or simple shear flow is the flow between two parallel
plates one of which is at rest and the other moving with a uniform velocity U in its
In plane poiseullie flow the pressure gradient is non-zero, but the plates are
kept stationary.
CHAPTER 3
1. INTRODUCTION
boundary layers are so different that the precise location of this relatively abrupt
transition can have a profound influence on the overall drag, heat transfer, and
prediction of boundary layer transition is complicated by the fact that it does not
correspond very directly to the onset of instability. Stability analysis for boundary
layers is well developed and very predictive of the behaviour of small disturbances.
analysis is less helpful with the prediction of bypass transition where external free-
stream disturbances bypass the classic instability mechanisms and initiate the
where Rex,tr is the transition Reynolds number based on the local free-stream
( m = 0 for zero pressure gradient), 99 is the local 99% boundary layer thickness,
and T (2/3)k / U is the local free-stream turbulence intensity. This is only one
of many proposed correlations. While they appear relatively simple, they are
quantities, such as free-stream values, distance downstream from the leading edge
general situation are very difficult to formulate. Many CFD codes simply resort to
requiring the user to specify the transition location. Once the transition location has
been determined, there is the further difficulty of determining how the turbulence
model should be prompted to become active. Solutions to this problem range from
specifying large turbulent kinetic energy at the wall where transition is expected to
occur to specifying local large source terms in the turbulence evolution equations
(artificial production). None of these methods of tripping the turbulence model are
particularly reflective of the actual physical transition process, and add further
sort is largely code dependent, so that the same theoretical model can produce
different results in different codes, or even within the same code using different
meshes.
the transition location. This is a very natural approach in by pass transition, since the
low Reynolds number k models have been performed, and [13] gave good reviews
of how various flavours of the k models perform. The overall conclusion is that
none performs well for all the flows considered. More recent work usinga k
intermittency model (Suzen and Huang 2001) has shown more success. However,
2. MOTIVATION
as ease of implementation,it also has a solid theoretical justification that has not been
discussed previously. The derivationof the RANS and accompanying Reynolds stress
+ . () = p + . vu . R (1 a)
. u = 0 (1 b)
2
+ . R = v R (
+
)+ ( + )
2v
( + + )
(2)
whereuis the mean velocity, and Ris the single point, second-order correlation of the
textbooks. The first termin parentheses is the production term. This term does not
require a model if equation (2) issolved. The subsequent terms are pressure strain,
model.
are unclosed, and models must be constructed torepresent the effect of the turbulence
fluctuations on the mean flow. It is these models, notthe equations themselves, that
may, or may not, be able to capture the nonlinear transitionto turbulence. The RST
equation (2) is the basis for many turbulence models (including k/,algebraic
Reynolds stress, and turbulent potential models) and is the starting point for
thisdiscussion.
In a fully turbulent flow all the unclosed terms in the RST equations require
are relatively smalland one might expect the turbulent transport term, which involves
a velocity triple product, tobe very small. Linear stability analysis is akin to assuming
that the turbulent transport term isidentically zero. Furthermore, this term can only
equation (2)) is small in fully developed turbulent boundary layers,and also can only
redistribute energy, so we will assume that like the turbulent transport term itis not a
expectedto be small when the transition process initiates. Eventually, we know that
the turbulence willcreate very strong local gradients so that the dissipation term will
balance the other terms inthe RST equation. However, early in the transition process,
the gradients will be relativelyweak and the turbulence will grow very rapidly,
significant influence on the exact nature of the overshoot(in turbulence intensities and
skin friction) that occurs at the end of transition just before fullydeveloped turbulence
is obtained, but it is unlikely to have much effect on the much more criticaltransition
location.
= 2 , ,
+ , ( ) (3)
,
shows that the pressure has two components, one of which is dependent on the mean
strain term is of thesame order of magnitude as the production term, and tends to
reduce some, but not all, of theproduction. Rapid distortion theory (RDT) can tell us a
great deal about how this term shouldbehave in certain limits. The RDT analysis is
transition. Modelsfor the pressure transport, turbulent transport, and dissipation are
only expected to influencethe very last stages of transition (the overshoot recovery)
and should have little effect on thecritical transition location. Most models for the
rapid pressure strain usually only capturerapid shear correctly because this is the
important case for boundary layer and free-shear flows.Reference [12] have a
such as k/ arevery popular. The k-equation is derived by taking one half of the trace
of equation (2):
2
1
1 (4)
+ . = { ( + )} ( ) + 2
2 2
Two very important changes take place when only the k-equation is solved.
Firstly, pressurestrain disappears entirely. This is because this term only redistributes
energy between variousReynolds stress components, it does not influence the overall
redistribution is important, particularly inthe boundary layer. It takes energy out of the
fluctuations, damping the streamwise fluctuation growth andfeeding the normal and
Inaddition, the production term is no longer exact. The Reynolds stress tensor in this
2
= ( + ) (5)
3
where the eddy viscosity is given by = 2/. More complicated versions of this
algebraic Reynolds stressmodels. They all have the same fundamental problem for
equilibrium with the mean flow. They suppose thata given shear level results in a
given turbulence level. The transition process is as far from anequilibrium situation as
turbulent potentialmodel used in this work [9, 11, 21] represents an effective
inpredicting fully turbulent flows has been discussed in prior publications [2, 10, 16].
In this paper,we focus on its unique abilities to predict transition in boundary layer
transition.
3. MODEL EQUATIONS
The turbulent potential model used in this work is a slight modification of the
model proposedby [9]. The basic motivation and derivation remains the same.
Dk
= ( +
k) + P (6a)
Dt
D
) + (C1 P C2 )
= ( + (6b)
Dt k
= ( + ) + +
1/2
+ (1 1.5 ) ( ) (6c)
D 1/2
= ( + t) + + + + C ( ) k
Dt
(6d)
k 1
P = , t = C , =
1 + 1.5
=
10| 1/2 |
[1 + ]
t 3 6
C = 0.21, Cp1 = 2.0 , Cp2 = , Cp4 = , C = 0.0033
t + 5 5 7
P
= 0.33 + 0.67
P
= 0.33 + 0.5
= 0.33
k2
Ce1 = 1.45, Ce2 = 1.83 0.16 exp (0.25 )
The slow pressure-strain models are
) ( ) k + Cp1 (2 1)
= ( +
k k
(k) (k) P
+ (Cp2 + Cp4 ) [ 25
]
/k (1 + Re ) k
and
= 2( + ) ( ) k Cp1 (1 )
k k
and
P P
= (1 Cp2 ) + Cp2 + C (2 1) 2
k k
Transport terms are modelled by the in the diffusion term. The two terms involving
in the turbulent potential equations (6c) and (6d) are the critical terms for
predictions when is set to zero in fully turbulent flows. However, it has a very
direct effect on transition, with higher values of causing transition to occur earlier.
We have used a value of 0.0033 in all the results presented in the paper, but it should
be observed that the model has the unique ability to explicitly control transition if the
user so desires.
definitions ofsome of the constants. Standard definitions for these constants will also
work quite adequately,these are just our own preferences. What is important to
emphasize is that although we aresolving transport equations this model is
nothing like a /model. In this model thepotentials are defined by the exact
the turbulent potential evolution equations. They are not used todirectly calculate the
Reynolds stress tensor, or the effect of the turbulence on the mean flow.
present model, and arecompared to experimental data. The results presented below
stream turbulence intensity, including verylow level leading to natural transition, the
The mean velocity is initially uniform flow for all casesand the initial values of
are given in table 1. The initial values of the turbulent kinetic energy k are
determinedusing k = 3/2(TuU0 )2 . For the Tu = 0.03, 1.25 and 1.3% cases, the
initial turbulent dissipationrate is calculated from ReT using = k 2 /(ReT ) and the
value of ReT is assumed. The resultsare not very sensitive to reasonable values of ReT .
For the Tu = 3%(T3A) and Tu = 6%(T3B)cases, the data for k are available, and
dk
the initial values of are determined from = U ,where the subscript
dx
represents parameters in the free stream. The values of ReT for thesetwo cases listed in
table1 are just calculated from . The initial potentials and are set as2/3and zero
respectively. All experiments were performed in air so = 1.55 105 was usedin
every case.
1/2 1/7
turbulence flows,C = 0.664 Re and 0.027 respectively, are also plotted
with dashed curves forcomparison. Overall, the ability of the present model to predict
the critical transition location isquite good. However, the natural transition case, Tu =
Table 1.
layer.
( ) (%) Source
intensities. The symbols represent experiment data and the curvesare the present
results. The dashed curves are fully laminar and fully turbulentcorrelations.
[18] to classical predictive theories,such as the e9 rule, which predict a value of2.0
106 [17]. This range of results is commonlyattributed to different noise levels in the
various experiments. We show later that the model canactually predict the entire
The other significant departure of the model predictions from the experiments
is in theovershoot of the friction coefficient that occurs at the very end of the
transition process andin the transition length. Predicting this overshoot and length
the model does show the characteristicovershoot behaviour found in experiments, but
over predicts the extent of this overshoot. Ourcurrent hypothesis is that this
velocity and turbulence variables, and is not due to the model itself. Theboundary
layer equations are based on the premise that streamwise second derivatives are
small.During the latter part of transition, as the flow suddenly becomes turbulent, the
boundary layer
Figure 2.
T3A case.the symbols represent experimental data and thecurves are the predictions.
grows quite dramatically, and the assumption of small streamwise second derivatives
behaviour. The fact that the overshooterror increases as the turbulence intensity
events are also more abrupt and therefore violate the boundarylayer approximations
Two other numerical predictions, both performed by [15], but using their
for the T3A case. The model predictions are not particularly good for this
particular case. In contrast, the model ofSuzen and Huang captured the late stage of
the transition region (including the overshoot) verywell, but the transition onset was
behaviour. In other words, the present model predicts not only a correctposition of
transition onset, but also a fairly accurate length of the transition region.
and shapefactor, H, with experiments and results from other numerical models are
shown in figures-3and 4 respectively for the T3A case. It can be noted that the present
prediction always remainsbetween the other two simulations. The fact that the model
of Suzen and Huang has a slightlater prediction of the onset of transition is also
The streamwise mean velocity profiles calculated using the present model are
compared atfour downstream stations with the experimental T3A data and the
layers are compared with the experimental data of [3] in figure-6. The initial
turbulence levels of these two cases are 3.0% (T3C3) and 6.6% (T3C1) respectively.
For both these flows, the pressure gradientis initially negative (favourable) and then
positive (adverse) in a profile that was designed to roughly approximate the flow over
different cases are roughly equivalent. The initial conditions are listed in table 2 and
are fully determined by the initial experimental data. It is very difficult to use
correlations to predict this variable pressure gradient flow because the pressure
gradient does not correspond closely to any single Falkner-Skan situation. In addition,
the RANS models tested by [13] showed some difficulty with these cases.
The model predicts the transition location in these complex variable pressure
gradientboundary layers quite well. Note that in both cases the initially favourable
pressure gradient has delayed transition compared to the flat plate case. The pressure
gradient data are not provided beyond the range of the experimental data, so we
cannot continue the 3% case beyond the experimental data. Predictions given by [15]
are compared with the present calculation for the T3C1 case again in figure-7.Figure-
8 shows the comparison of Re computed by the present model and the two other
Figure 5.
(a) Re = 1.35 105 ;(b) Re = 2.04 105 ; (c) Re = 2.74 105 ; (d)
Re = 4.19 105 .
Table 2.
( ) (%) Source
models with the experimental data for the T3C1 case. Predictions of the shape factor
profiles using the three models are plotted in figure 9 along with the experimental
data. Again the present model exhibits excellent performance in calculating the shape
It was mentioned previously that acoustic effects (wind tunnel noise) can have
assumed to be largely free of acoustic noise, and therefore gives a high quiet
Figure 7.
were clearly not as successful inthis respect as the later experiments of Well.
(see[9]),and could be a very good measure of the average acoustic forcing. Our initial
natural transition prediction (figure-1) was with relatively high noise levels of =
2/3, and is actually quite close to noisy experiments. If one reduces the initial
scalar potential to a value of 0.12, the on set of the transition will be delayed to
data of [14] Re, = 2.8 106 in table 1 by using = 0.4. All three simulations
are shown inthe C Re curve plotted in figure-10. This demonstrates the ability of
the model to accurately capture the effects of noise on natural transition. The effects
turbulent flow decays to a laminar state. Relaminarization can occur because of strong
stratification, strong rotation,or in the case of channel flow, due to a reduction in the
driving pressure gradient. Figure-11 shows predictions of the skin friction in channel
h
flow for various bulk Reynolds numbers Re = u ,where h is the channel half height
and is the average velocity. The figure also has dashed curves
Figure 10.
The acoustic effects on the natural transition at Tu= 0.03%. The dashed curves are
fully laminar and fully turbulent correlations.
Figure 11.
Predictions of the skin friction in channel flow for various bulk Reynolds numbers.
The dashed curves are fully laminar and fully turbulent correlations.
8
for the laminar exact result C = 3 Re1 and the turbulent correlation of [4]
1
C = (w)/ (2 u20 ), where wis the wall shear stress and u0 is the centre line
improved. However,a theoretical study by [8] has shown that the linear instability for
a channel flow occurs atRe = 1600(Re = 110). It is our observation that this
behaviour of the model (particularly the dissipation and pressurestrain term), and not
variety of boundarylayer flows has been demonstrated. This includes the ability to
accurately calculate the transitional flat plate boundary layerunder favourable and
adverse pressure gradients. The ability of this model to reflect the effect ofnoise on
unlike /, the model does notassume equilibrium between the developing instability
INTRODUCTION
small perturbations to the velocity growing rapidly. Indeed, turbulent flows in closed
domains appear to be chaotic dynamical systems [11]. The result is that the evolution
of the detailed turbulent fluctuations can only be predicted for a finite time into the
future, due to the exponential growth of errors. In a chaotic system, this prediction
which is the average exponential growth rate of typical linear perturbations. The
thus a measure of the strength of the instabilities that underlie the turbulence, and its
inverse defines the time scale over which the turbulence fluctuations can be
meaningfully predicted.
experimentally since the work of Swinney [23], using indirect methods. In numerical
the evolution of linear perturbations. This has been done for weakly turbulent Taylor
Couette flow [21] and very low Reynolds number planar Poiseuille flow [11].
extensively studied both experimentally [3, 4, 14] and using numerical simulations [2,
10, 19, 22]. It is valuable as a model for the small scales of high Reynolds number
turbulence away from walls [6]. It has been speculated that in isotropic turbulence, the
maximum Lyapunov exponent scales with the inverse Kolmogorov time scale [5],
suggesting that the dominant instabilities occur at Kolmogorov length scales as well.
If true, then a study of the maximum Lyapunov exponent and the associated
flows.
This paper focuses on how the maximum Lyapunov exponent and hence the
predictability time horizon scale with Reynolds number and computational domain
that should scale as the inverse Kolmogorov time scale [5] is in agreement with
an estimate from a shell model [1]. However, this scaling has not been directly tested
a direct numerical simulation, one necessarily computes the linear disturbance that is
most unstable (on average). This can be used in the short-time Lyapunov exponent
analysis, as introduced in [21], to characterize the nature of the instabilities. This will
be pursued here for isotropic turbulence. The remainder of this paper includes a
brief review of Lyapunov exponents and how they are computed in numerical
studied here (section III). The results of a scaling study of the Lyapunov exponents
are given in section IV, and a short-time Lyapunov exponent analysis is presented in
of the current study are that 1) solutions evolve toward a stable attractor, and 2)
solution trajectories on the attractor are unstable so that near-by trajectories diverge
Lyapunov exponents, whose characteristics are recalled briefly here. Further details
can be found in [21]. In addition, the use of Lyapunov exponents in the analysis
evolve toward an attracting set in phase space (the attractor); in turbulence this
perturbation u(t0) of the solution at time t0, and its evolution in time. The Lyapunov
1 ()
= lim ( ) (1)
(0 )
exponents, depending on the solution u(t0) and the perturbation u(t0) at the starting
time. However, for almost all u(t0), = 1the largest Lyapunov exponent, and, due to
round-o error and other sources of noise, in practical computations, = 1 for all
u(t0). Furthermore, the Lyapunov spectrum (1> 2> 2> ) does not depend on
u(t0); it is instead a property of the dynamical system. See the review by Eckmann et
limit t
() 1 ()
(()) (()) (2)
() ()
and depend only on the solution at t, and not on the starting conditions
where
long run, growing at the average exponential rate . It is defined by the fact that
itslong-time average growth rate forward in time is and when the evolution is
backward in time the long-time average growth rate is . The short-time Lyapunov
Because and
depend only on the solution at the current time, they can be
used as diagnostics for the instabilities responsible for a system being chaotic. In
particular, when is large, the underlying system is particularly unstable, and at that
when is large, and by analyzing the solution u and the Lyapunov disturbance
at
that time, we can characterize the important instabilities. This is the short-time
In this paper we will be concerned with the scaling of with Reynolds number
and with the chaotic instabilities revealed by short-time Lyapunov exponent analysis.
4.2 SIMULATIONS
Stokes equations which only acts at large scales. The forcing formulation is described
spatial discretization with N modes in each direction, and the vorticity formulation of
Kim et al. [12]. This formulation has the advantage of exactly satisfying the
continuity constraint while eliminating the pressure term. A low-storage explicit third-
perturbation added to the base field. This perturbation satisfies the linearized Navier-
Stokes equations:
1 2
+ ( + ) = + (3)
= 0, (4)
Where ui is the base field and ui is the disturbance field. The disturbance equations
are solved using the same numerical scheme as the Navier-Stokes equations. Note that
the forcing is applied only to the base field and not the perturbation. The
implementation of both the base and disturbance field solvers were verified using the
The goal of the forcing is to inject energy into the large-scale turbulence so
that the isotropic turbulence will be stationary. Forcing is applied to Fourier modes
specified rate of energy injection (forcing power), which, when the system is
stationary, will be the dissipation rate. By specifying the wave number range being
forced, forcing power and viscosity, the integral scale, turbulent kinetic energy and
1 2
+ = + + (5)
= 0 (6)
Following [20], in the Fourier spectral method used here, the Fourier transform of the
() = ||2 ().
(7)
study Lyapunov exponent scaling. Values of are quoted in units in which the
domain size is 2, and averaging times are normalized by eddy turnover time.
Given that ui is a Navier-Stokes solution, fi is guaranteed to be divergence-
free. The coefficient in the above is determined as a function of time so that the
forcing power is the target dissipation rate T. Since the forcing is applied only to a
1
= { ( ||2 () ()) || (8)
||
0
where denotes the complex conjugate, and kfminand kfmax are the bounds on the range
equations, the viscous term has the same structure as fi, so this forcing can be
combined forcing and viscous term is then( )||2 (). In the numerical
solution of the Navier-Stokes equations, this combined term is treated in the same
way as the viscous term would be. Note that fi is just a nonlinear function of ui, so
B. Simulation Cases
Reynolds number and the ratio of the computational domain size L to the integral
scale , eight simulations were performed. These are summarized in table I. To study
the scaling of with Reynolds number, six cases where simulated with the same
forcing wavenumber range and .This resulted in approximately the same integral
scale in each case. The Reynolds number was manipulated by changing the viscosity.
To study the potential variation of with domain size normalized by integral scale,
the domain size was kept fixed at 2 and the integral scale was changed by adjusting
the forced wavenumber range, while keeping the Reynolds number approximately
fixed. In all cases kmax> 1, where kmax is the maximum resolved wavenumber, and
For each case, the simulations were run until the base solution became
statistically steady and then the statistics were gathered by time averaging over a
exponent on the Reynolds number and on the domain size. To address this, the
maximum Lyapunov exponent , the integral scale () and Reynolds number based on
the Taylor micro-scale (Re) are needed, along with their uncertainties. Based on the
= 2 5/(3) [17]. Thus the two statistical quantities that need to be computed
from the DNS are and q2. Both are determined as a time average over averaging
time Tavg (see table I), and the standard deviations of the uncertainty due to finite
averaging time were determined using the technique described by [15]. The values of
, q2 and their standard deviations are given in table II. The standard deviations of the
for each simulation and are specified, there is no uncertainty in their values.
Case
TABLE II.Values of q2 and the maximum Lyapunov exponent , along with the
which the domain size is 2 and = 1, and is normalized by the Kolmogorov time
scale .
Kolmogorov time scale were correct, these data would, within their uncertainty, fall
along a horizontal line. However, this does not appear to be the case. Indeed,
appears to be growing with Re. Also, shown in figure 1 is the dependence of scaled
appear to be consistent with the hypothesis that the Lyapunov exponent does not
given the data and its uncertainties. These scaling relationships serve as the model
for the inference. In Bayesian inference for this problem, the joint probability
(a) the Taylor scale Reynolds number Re and (b) the ratio of the integral scale to the
domain size /L, from the data in table II. The error bars on the data (in blue)
represent one standard deviation. Also shown are the outputs of the models (9) (in
red) calibrated with Bayesian inference, with the dark and light gray shading
(a) (b)
where (d|, ) is the likelihood and (, ) is the prior. The likelihood is the joint
probability density for the observed quantities evaluated for the observed values of
these quantities, as determined by the model with parameters and , and given the
uncertainties in the data. The prior represents our prior knowledge about the
The primary source of uncertainty in such data is statistical sampling error. The
central limit theorem implies that in the limit of large samples, the uncertainty
associated with sampling error is normally distributed with zero mean. Therefore, to
formulate the likelihood, the data are assumed to have Gaussian uncertainty with
standard deviations as reported in table-II. The probability distribution for the ith
2
1 ( 1 1 )
( |, , ) = exp [ ] (11)
2 22
wherexi is the independent variable (Rei or /L, depending on which scaling relation
is being inferred) of the ith observation and i is the standard deviation in associated
with the ith observation. However, there are also uncertainties in the values of the
distribution and standard deviation for the ith observation of xi. In this case, the
1 ( )2
(| ) = exp [ ] (12)
2 22
The conditional distribution of given the parameters and the observed independent
To inform the prior, we consider the range of time scales in the turbulence.
The largest is the eddy turn-over time, which is proportional to q2/, and the smallest
is the Kolmogorov time scale /.The ratio of the turnover to the Kolmogorov
times scales as Re. Therefore, the Lyapunov exponent scaling with the turn-over
time would imply = 1 and scaling with the Kolmogorov scale would imply =0.
However, theoretical arguments suggest that the Lyapunov exponent scales with the
Kolmogorov time scale [5] (=0), and we need to allow for the possibility that this
assessment may be in error in either direction. The bounds on the range of plausible
range was used as a prior for . somewhat arbitrarily, the same range was used for the
prior in the domain size scaling relationship. The parameter is a positive definite
scaling parameter, and so following Jaynes [9], a Jeries distribution ()1/ is used
(, ) = ()().
Given the likelihood and prior described above, and the data in Table II,
Carlo (MCMC) algorithm [8] as implemented in the QUESO library [13, 18]. The
resulting samples were used to characterize the joint posterior distribution of and
for both the Reynolds number and domain size scaling model, as shown in figure-
2and 3. Notice in these figures that the joint distribution has probability mass
FIG. 2. Posterior PDFs for 1 and 1 in the Reynolds number scaling model (9).
Shown are the marginal distributions of both parameters along with contours of their
joint distribution.
changes in can be compensated for by changes in so that the model still fits the
data. The MCMC samples were also used to determine the uncertainty in the model
predictions for the Lyapunov exponent as a function of Reynolds number and domain
size, with the results plotted in figure 1, along with the data. From this, it is clear that
the scaling models as calibrated are consistent with the data and their uncertainty.
relation shows that the most likely values of are between about 1/4 and 1/3, with the
possibility that the value is zero essentially precluded. This is remarkable since it
suggests instability time scales that will become increasingly faster than Kolmogorov
consider is that this fast instability time scale arises as an artifact of the time
discretization of the DNS. However the DNS time step in Kolmogorov units
t/ Re 1/2, so if the Lyapunov exponent were scaling with the DNS time step,
would be 1/2, which is also essentially precluded by the posterior distribution. The
number scaling. This was also verified by running a time refinement study where
was found to be invariant to changing t.As with the time step, interest in the
computational domain size arises because of concern that computational artifacts not
impact our Lyapunov exponent analysis. The posterior distribution of in the domain
size scaling relationship (figure-3) shows that =0 is highly likely, with the most
probable values of ranging from 0.05to 0.05. If there is an eect of the domain
size, the data indicates that it is extremely weak. It therefore appears that the
Lyapunov exponent Reynolds number scaling discussed above and the short-time
size eects.
4.4. SHORT-TIME LYAPUNOV EXPONENT ANALYSIS
As discussed in section II, both the disturbance field (u) used to compute the
Lyapunov exponent and its instantaneous exponential growth rate () depend only on
time Lyapunov exponent analysis, we study and u to learn about the instabilities
First, consider the time evolution of the exponential growth rate , which is
Note that in both cases takes large excursions from the mean, of order 3 times the
mean value. However, the variations in occur on a much shorter time scale and the
large excursions seem to occur more often in the high Reynolds number case. The
time scale on which varies appears to decrease somewhat faster than the
Kolmogorov time scale with increasing Reynolds number, as when plotted against
t/, still varies faster for case-6 (figure-5). At the same Reynolds number (cases-1
computational domain size L/. The fact that the time scale of the instability, as
measured by the Lyapunov exponent, decreases faster than the Kolmogorov time
scale suggests that the instability processes are acting at spatial scales near the
Kolmogorov scale. In this case, a simulation with a larger domain size relative to
intrinsic turbulence length scales would include a larger sample of local unstable
case-1, the relative volume increases by a factor 64, suggesting that the variability of
and the question naturally arises as to what is special about these times. To investigate
this, the spatial distribution of the magnitude of the disturbance energy density is
visualized in figure 6 at three times, just before the beginning of a peak in , a time
half way up that peak and at the peak (tq/ = 9.58, 9.85 and 9.89 in figure 4). Notice
that before the rapid growth of into the peak, the energy in the disturbance field is
broadly distributed across the spatial domain. Half way up the peak, the distribution is
much more spotty, and finally at the peak, the disturbance energy is primarily focused
in a small region, appearing in the lower left corner of figure 6(c). The contour levels
in these images were chosen so that the contours enclose 60% of the disturbance
energy, implying that 60% of the disturbance energy is concentrated in the small
feature in the lower left of figure 6(c). Another indication of the dominance of the
disturbance feature in figure 6(c) is that the contour level needed to enclose 60% of
the energy is about 2500 times the mean disturbance energy density, while in figure
6(a) the contour is only about 15 times the mean. Clearly the growth of the
disturbance field in this concentrated area is responsible for the peak in . However,
the spatially local exponential growth rate of the disturbance energy |u|2 |u|2/t is
not particularly large there, large values of this quantity are distributed broadly across
the spatial domain. It seems, then, that the large peak in is due to a local disturbance
that is able to grow over an extended time until it dominates the disturbance energy,
so that the disturbance is localized in a region of relatively large growth rate. This is
presumably unusual because it requires that the local unstable flow structure
shown in figure-8, the disturbance vorticity is localized on the vortex tubes, with
regions of opposite signed disturbance vorticity to one side or the other of each vortex
tube. This disturbance, when added to the base field would have the effect of
displacing each vortex tube along the line between the positive and negative peaks in
the disturbance vorticity associated with each tube. The instability then appears to be
one associated with slowing (speeding up) the co-rotation of the vortex tubes while
they move away from (toward) each other. Note that the disturbance equations, being
linear and homogeneous, are invariant to a sign change, and so the sign of the vortex
FIG.4. Short-time Lyapunov exponent scaled by . In (d), the time axis is expanded to
zoom in on the peak indicated in (c), and symbols show the times at which the images
The results of the scaling study (section-IV) show definitively that, at least
over the Reynolds number range studied, the Lyapunov exponent does not scale like
the inverse Kolmogorov time scale, as had been previously suggested [1]. Instead,
increases with Reynolds number like Re for in the range from 1/4 to 1/3. Further
note that the analysis of Aurellet al. [1] indicated that a correction for the
intermittence of dissipation would yield < 0, also inconsistent with the current
results.
remarkable, as it would mean that there are instability processes that act on time
scales shorter that Kolmogorov. However, in the highest Reynolds number (Re=210)
simulation performed here, is still only 0.16. it is certainly possible that this
insufficient scale separation between the large scales and the scales at which the
instabilities act, and that the value will reach a plateau at some much higher Reynolds
number. Clearly, this scaling behavior of the maximum Lyapunov exponent is worthy
of further study. The current results suggest that the generally accepted and most
obvious scaling is not correct, and that, unfortunately, turbulent fluctuations are even
instabilities in turbulence act on the smallest eddies. Further, at Re=210, when the
instantaneous disturbance growth rate was the largest (about 3 times the mean), the
disturbance energy was highly localized, suggesting that it was a particular local
instability that was responsible for the rapid growth at that time. However, this was not
due to a particularly large local growth rate, as the logarithmic time derivative of the
spatially local disturbance energy was equally large in regions spread throughout the
importance, except that the underlying structure in the turbulent field was especially
vortices, as in a mixing layer. In this, the short-time Lyapunov analysis pursued here
appears to be a valuable tool for the study of the instabilities underlying turbulence.
FIG. 6 Contour of the magnitude of at three times leading upto the peak as
indicated in figure 4 for Re = 210 (case 6). The contour shown is that at which 60%
of the disturbance energy is enclosed by the contour. To achieve this, the contour
levels are (a) 15, (b) 25 and (c) 2500 times the mean disturbance energy density.
FIG. 7. Contour of the magnitude of the vorticity of the base field for Re = 210 (case
6) at the peak of in the region where the disturbance field is localized (box
highlighted in figure 6c). The contour level is 9.2 times the square root of the mean
enstrophy. The vortex tubes are co-rotating, with the direction of rotation indicated by
FIG. 8.The magnitude of the vorticity (grayscale) and the disturbance vorticity
component normal to the plane (contour lines) in a plane perpendicular to and in the
middle of the vortex tubes shown in figure 7. For the disturbance vorticity, the red and
INTRODUCTION
that occurs within a separation bubble along a flat plate. Boundary layer transition is
important, both economically and ecologically. Since drag increases when a boundary
layer transitions, and because an increase in drag must be overcome by an equal and
opposite production of thrust, flight vehicles and fluid powered devices consume more
fuel and generate greater amounts of greenhouse gas emissions because of these
boundary layer transitions [1]. Furthermore, because transitioning boundary layers are
prone to separate, they can dramatically affect the overall performance of high lift
devices and gas turbines [2, 3]. When a laminar boundary layer separates, transitions
into a turbulent flow and then reattaches downstream, the resulting phenomena is
called a separation bubble [4]. And these separation bubbles produce drag increases
that are greater than those that would otherwise be produced by transitioning but
experimental criteria to determine the size of a separation bubble. Using this work and
others, Mayle [2] developed experimental correlations that would predict the sizes of
separation bubbles. More recently, Haggmark et al. [18] have performed a more
transition [4, 20, 21]. DNS, although computational expensive [22], have produced
some important studies of separation bubbles [4].Spalart and Strelets [23] simulated a
reattaches and then evolves into a normal turbulent boundary layer. Alam and
Sandham [4] performed a simulation of a short separation and Balzer and Fasel [24]
simulated the effects that freestream turbulence has on the length of short separation
bubbles. Rist et al. [21, 25, 26] also studied how disturbances can be used to control
the size of separation bubbles for both short and long separation bubbles.
In this paper we will obtain the locations of separation, reattachment and the
start and end of transition for the flow over separation bubbles. We will verify our
belief that when the separation bubble reattaches, the turbulent boundary layer flow
will be fully developed. We will use a transition model derived from a complex-
lamellar decomposition of the velocity field across transition that exists between a
fully developed laminar and turbulent boundary layer flows. This transition model is
augmented with the experimental correlations of Mayle [2] and Narasimha [27].
(x, y, z) the conservation laws of continuity and linear momentum, and the vorticity
. = 0, (1)
1
= + 2 . (2)
= (. ) + 2 . (3)
= + (u. ) (4)
Complex-Lamellar Decomposition
can be decomposed into a complex-lamellar form in general [28], and the following
= + (5)
Here the velocity field is separated into two components, one potential and the
2 = + (6)
is constructed from the potential , often called either Craigs circulation preserving
This energy dissipation potential also defines the viscous behaviour of the transport of
.
= (7)
2
= 1. (8)
= , (9)
consistency, and contains a source term that either creates or destroys enstrophy
along the particle paths. Given the closed curve c around a material surface s, with
line and surface elements land, the transport of E must satisfy the following
equation:
= 2 (. ), (10)
= 2 [ x]2 , (11)
= 2 [ x]2 , (12)
Augmented Laminar Flow
= + + (13)
previously defined, Eq.(5), has been added to ulam, a baseline laminar flow.
= , (14)
the difference between E, the enstrophy function present in the given flow of
interest, and that which would otherwise be found in a baseline laminar flow, Elam.
= 2 [ x]2 (15)
= 2 )
(2 , (16)
where2 is the square of the magnitude of the vorticity found in the transitioning
2
boundary layer, while is that which would otherwise be found in the baseline
laminar flow. Thus, at the end of the fully developed laminar boundary layer flow,
the location where the boundary layer is assumed to be a fully developed turbulent
[ x ] | = ( ) | , (17)
which is a change in vorticity at the end of transition. It can be shown [29], that for
a 2D boundary layer flow, the following drift function approximation can be made:
= 1, (18)
= 0, (19)
= [2 sin(1 ) + 3 (2 )]. (20)
in regions where v 0, 0 and u 0, or within a boundary layer but away from
The approximate drift function, Eq. (18), when coupled with the change in
1 2 )
[ (2 ] | = ( ) | . (21)
At the ends of this region, we recognize the following boundary conditions:
0
2 2 when =
= {(2 2 ) , (22)
| when =
2 2 2 2 )
( )
= ( | ( ). (23)
( )
=1
which is the change in enstrophy density generated as the flow transitions from
requires that
= 1 (24)
=1
Finally, we can substitute the approximate enstrophy distribution, Eq. (23), into Eq.
| = ( + ) | ( ), (25)
+ 1
=1
an equation that states that the velocity at the end of transition is a function of the
change in enstrophy between the fully developed laminar and turbulent flows, the
sum of the resulting turbulent and otherwise laminar vorticities, at the end of
transition, and the length of the fully developed laminar to fully developed turbulent
region itself, (xB xA). These results are dependent on the series coefficients Cn,
which in turn define the shape of the change in enstrophy density across transition.
5.2. MODEL OF INTERMITTENCY
From the end of the fully developed laminar boundary layer flow, xA, to the
beginning of the fully developed turbulent flow, xB, we can define the scalar factor
as
( )
= , (26)
( )
=1
2 2 2 )
2 = ( | . (27)
It was shown [29] within the measurable extent of the intermittency region,
0.75 identifies the location where the mean flow is approximately 75 percent
turbulent and 25 percent laminar, and = 0.25 identifies its counterpart. The shape
( )2
= 1 [0.412 ], (28)
2
wherext is the equivalent leading edge of the turbulent boundary layer, and the
( )
= = 0.25 = 0.75 (29)
( )
=1
Figure 1 illustrates how the extent of intermittency is small in length compared with
the overall region from xAto xB. As such, we chose to limit the series approximation
to three terms and generate three geometric equations using Mayles [2]
= 4000.7
, (30)
( )
= (31)
3.36
The following geometric equations are used for transition over a separation bubble:
3
( + 154.420.7
)
0.5 = , (32)
=1
( )
3
( + 99.4760.7
)
0.25 = , (33)
=1
( )
3 1
8.257 104 ( + 154.420.7
)
= , (34)
0.7
( )
=1
the Prandtl-von Karman power law velocity profile and the fully developed laminar
boundary layer flow at xA is satisfied using the Pohlhausen laminar velocity profile.
Substituting both the turbulent and laminar velocity profiles, into the
complex-lamellar relationship, Eq. (25), and with this relationship being evaluated
3(2 /2)2
| |
3
(5)2 ( ) (5)3 ( ) 2
4(1 /6)3 3
|
+ 2 [ ( 1) ]
(5)4 ( ) +1
=1
]
1
7
|
=( 4/5
) (35)
0.375 ( ( ))
that exist within our transition model as well as a sketch of the laminar an turbulent
Fig. 1. Transition model over a separation bubble bounded between xA and xB.
5.4. RESULTS
Our transition model, Eqs. (24) and (32) to (34) and section 4, is used to
determine the streamwise locations of separation, reattachment, and the start and
end of intermittency for the transitioning flow across several different separation
bubbles. We will present results that suggest that reattachment occurs when the
turbulent boundary layer becomes fully developed. The system of equations that
2010b, The MathWorks, Inc., Natick, MA) with a convergence tolerance of1012.
Multiple solutions are obtained for each test case, however only the solutions
assumed to be physically relevant, the ones consistent with the DNS simulations,
are presented.
Our first test case will compare the results obtained from our transition
model to those found in the DNS study by Cadieux et al. [36], which is based on the
work of Spalart and Strelets [23]. Since the freestream turbulence level imposed
upon these DNS simulations was assumed to be around Tu = 0.03%, we used this
correlation,
( )
= 163 + [( ) ]. (36)
6.91
where
above the solid surface, at which we will apply our transition model. Thus, for a
where the flow has reattached and becomes fully developed and turbulent, the
resulting boundary layer velocity profile is modeled by the Prandtl velocity profile
[39]. From Cadieux et al. [36] DNS data, we assume the Reynolds number
form
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 1.23 x 105 and the fully developed turbulent boundary layer
of the fully developed laminar boundary layer flow, to the beginning of the
and interpret how the flow changes from a fully developed laminar to a fully
is zero and the laminar boundary layer flow is fully developed and assumed to
negatively, which implies that the boundary layer is now adjusting to the upstream
presence of the separation bubble. Thus while the boundary layer is assumed to
remain laminar, vorticity is being redistributed and the resulting flow is no longer
intermittency, the shape of the change in enstrophy density is assumed to follow the
shape of the universal intermittency function, Eq. (29), which grows from zero to
evolving until finally, at , the turbulent boundary layer flow becomes fully
developed. For the separation bubble modelled by this test case, Fig. 2 indicates that
the end of intermittency occurs at the same streamwise location as the start of the
implies that within this separation bubble the turbulent boundary layer flow
becomes fully developed at the end of transition. Furthermore, the result, that
Table 1.Comparison of the streamwise location of the start of the fully developed
turbulent boundary layer flow with Cadieux et al. [36] reattachment location.
Table 2. Comparison of the separation bubble length between the fully developed
laminar to fully developed turbulent boundary layer flows with Cadieux et al. [36]
The 1.3% difference between and supports the belief that the turbulent
fully developed laminar and turbulent boundary layer flows with the distance
between separation and reattachment from Cadieuxet al. [36] DNS simulation. As
shown in Table 2, the 3.4% difference between these distances implies that the
distance between the fully developed laminar and turbulent boundary layer flows,
separation bubbles, Gaster [19] concluded that for short separation bubbles the
reattachment location. Thus, we can conclude that the separation bubble identified
that the start of the fully developed turbulent boundary layer flow occurs at
reattachment. We compare the results obtained from our transition model to those
found in the Alam and Sandham [4] DNS simulation. From their DNS data, we will
Again, we assume that the fully developed laminar boundary layer that is
profile [38] and the separation parameter D 12. And again, downstream at ,
where the flow has reattached and becomes fully developed and turbulent, the
resulting boundary layer velocity profile is modeled by the Prandtl velocity profile
[39]. From Alam and Sandham [4] DNS data, we assume the Reynolds number at
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 3.31 x104 and the start of the fully developed turbulent boundary
turbulent boundary layer flow with the reattachment location from Alam and
Table 4. Comparison of the separation bubble length between the fully developed
laminar to fully developed turbulent boundary layer flows with Alam and Sandham
In Table 3, we compare the start of the fully developed turbulent boundary layer
For this test case, where the turbulence level was specified as Tu = 0.46%,
we have a 1.2% difference between the start of the fully developed turbulent
boundary layer flow, , and the reattachment location, . This result further
supports the belief that the start of the fully developed turbulent boundary layer
66
In Table 4, we again compare the distance between the fully developed
laminar and turbulent boundary layer flows with the distance between separation
and reattachment from Alam and Sandham [4] DNS simulation. The 1.8%
difference between these distances again supports the belief that the distance
between the fully developed laminar and turbulent boundary layer flows,
In Fig. 3, we show that the change in enstrophy density, Eq. (26), grows in a
similar way to the previous test case, specifically in terms of how the flow develops
from the end of the fully develop laminar boundary layer flow, at , to the start
of the fully developed turbulent boundary layer flow, at . Again, the end of
intermittency, occurs at the same streamwise location as the start of the fully
developed turbulent boundary layer flow, , which implies that the separation
In the next test case, we use our transition model to solve for three different
freestream turbulence levels, Tu=0.05, 0.5 and 2.5% and use the change in
enstrophy density, Eq. (26), to demonstrate how the freestream turbulence level
effects the reattachment location and length of the separation bubbles. We com-pare
our results to those found in Balzer and Fasel [24] DNS simulation.
For each freestream turbulence level, we again assume that at , the fully
[38]. As well, downstream at , where the flow has reattached and becomes
fully developed and turbulent, the resulting boundary layer velocity profile is
67
The first separation bubble is obtained from solving our transition model
following solution:
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 1.75 x 105 and the fully developed turbulent boundary layer flow
68
The second bubble is obtained with a freestream turbulence level of
a Reynolds number at separation, assumed from Balzer and Fasel [24] DNS data, of
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 1.42 x 105 and the start of the fully developed turbulent boundary
From Abu-Ghannam and Shaws [37] correlation, Eq. (36), we obtain a momentum
DNS data from Balzer and Fasel [24], we assume a Reynolds number at separation
69
Fig.4. Change in enstrophy density with associated streamwise locations illustrating
the changes in bubble length with changes in freestream turbulence level.
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 1.32 x 105 and the fully developed turbulent boundary layer flow
We can use the distance from the end of the fully developed laminar to the
of the separation bubble. Thus in Fig. 4, the change in enstrophy density, Eq. (26),
obtained for each freestream turbulence level is plotted. When the freestream
turbulence level is the lowest, at Tu = 0.05%, Fig. 4 shows that a separation bubble
freestream turbulence level is the highest, at Tu = 2.5%, the resulting bubble is the
the freestream turbulence level, the length of the separation bubble will become
shorter. This trend is consistent with the DNS study of Balzer and Fasel [24].
70
In Table 5, when we compare the start of the fully developed turbulent
Balzer and Fasel [24] DNS simulation, the larger differences between these
between the end of the fully developed laminar and the start of the fully developed
turbulent boundary layer flows is compared to the distance between separation and
reattachment from Balzer and Fasel [24] DNS simulation and again greater
differences are not ideal, they are within our expectations, considering the
Table 5. Comparing the streamwise location of the start of the fully developed
turbulent boundary layer flow with the reattachment location from Balzer and Fasel
[24] DNS simulation for three freestream turbulence levels of Tu = 0.05, 0.5 and
2.5%.
Table 6. Comparison of the separation bubble length between the end of the fully
developed laminar to the start of the fully developed turbulent boundary layer flows
with Balzer and Fasel [24] distance from separation to reattachment for the different
71
Distance from Fully Balzer and Fasel [24]
Freestream
Developed Laminar to Distance from separation %
Turbulence
Turbulent Boundary Layer to Reattachment Difference
Level
Flow
0.05% = 5.77 x104 = 4.27 x104 35
0.5% = 4.63 x104 = 3.63 x104 26
2.5% = 2.56 x104 = 3.06 x104 16
In the final test case, we present how changes to a disturbance amplitude can
produce a long separation bubble and a short bubble. We compare the results we
obtain using our transition model with those obtained in the DNS study by Rist
[25].
We can calculate the freestream turbulence level, Tu, that is associated with
2 2
1/2( + )
= . (43)
From Rists [25] DNS study, we obtain a disturbance amplitude of = 106 and
turbulence level of Tu = 0.04%. For each freestream turbulence level, we can again
assume within our transition model that at , the fully developed laminar
boundary layer flow can be modeled by Pohlhausen velocity profile [38] and the
Ghannam and Shaws [37] correlation, Eq. (36) results in the momentum thickness
72
Reynolds number at separation of = = 703. From Rists [25] DNS
| |
= 1.24 x106 . Upon solving the system of equations associated with our
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 1.33 x106 and the fully developed turbulent boundary layer flow
Fig. 5.Change in enstrophy density for long bubble produced with a freestream
73
separation, from Rists [25] DNS data, is assumed to be = 1.29 x106 . Upon
which predicts the effective leading edge of the turbulent boundary layer to be
located at = 1.319 x106 and the start of the fully developed turbulent
In Fig. 5, we plot the change in enstrophy density, Eq. (26), for both
shown as the dashed curve, the change in enstrophy density has similar growth
layer flow, . This is not the case for the lower freestream turbulence level of
implies that for this separation bubble the turbulent boundary layer flow at the end
of intermittency is not fully developed and only becomes fully developed further
downstream.
In his experimental studies, Gaster [19] concluded that for long separation
the end of transition. Thus, we can verify that the turbulent boundary layer flow
Table 7. Comparing the streamwise location of the start of the fully developed
turbulent boundary layer flow with the reattachment location obtained from Rists
74
[25] DNS simulation for freestream turbulence levels of Tu = 0.00045% and
Tu = 0.04%.
Table 8. Comparison of the separation bubble length between the end of the fully
developed laminar to start of the fully developed turbulent boundary layer flows
with Rists [25] distance from separation and reattachment for the different
turbulence levels.
the location of with the reattachment location, identified from Rist [25]
DNS simulation. This comparison, along with the comparison for Tu = 0.04%, is
made in Table 7.
The 1.4 and 2.2% difference for the freestream turbulence levels of
Tu = 0.00045 and 0.04%, respectively, supports the belief that the boundary layer
flow is turbulent and fully developed at reattachment for both short and long
separation bubbles. We can also conclude, for the separation bubbles obtained in
75
We can calculate the distance between the fully developed laminar and
turbulent boundary layer flows to determine the length of the long and short
separation and reattachment from Rists [25] DNS simulation, as shown in Table 8.
76
CONCLUSION
developed laminar to fully developed turbulent flow. This transition model was
used to obtain the locations of separation, reattachment and the starts and ends of
intermittency for several different separation bubbles. The fully developed laminar
boundary layer was modeled by a Pohlhausen velocity profile while its counterpart,
profile. Between these two fully developed boundary layer flows, the transitioning
flow is assumed to separate, reattach and form a separation bubble. Transition over
several different separation bubbles, both long and short, were analyzed and
compared to several different DNS studies. In each of the cases analyzed, the
transition model predicted that the separation bubble reattaches when the turbulent
77
BIBLIO GRAPHY
[2] Are S, Zhang X and Perot J B 2001 Accuracy and conservation properties of
a three-dimensional unstructured staggered mesh scheme for fluid dynamics
Int. J. Rotating Machinery at press
[4] Dean R B 1978 Reynolds number dependence of skin friction and other bulk
flow variables in two-dimensional rectangular duct flow J. Fluids Eng. 100
21523
[5] Dhawan S and Narasimha R 1958 Some properties of boundary layer flow
during the transition from laminar to turbulent motion J. Fluid Mech. 3 418-
36
[7] Keefe L, Moin P and Kim J 1992 The dimension of attractors underlying
periodic turbulent Poiseuille flow J. Fluid Mech. 242 129
78
[10] Perot J B and Taupier J 2000 Modeling three-dimensional boundary layers
using the turbulent potential model AIAA 2000-0914 (Reno, Nevada,
January 2000)
[11] Perot J B and Wang H 1999 Modeling separation and reattachment using the
turbulent potential model Proc. 4th Int. Conf. on Turbulence Modeling and
Measurements (Corsica, France, May 1999)
[16] Tsuei H-S and Perot J B 2000 Turbomachinery predictions using the
turbulent potential model AIAA 2000 0135 (Reno, Nevada, January 2000)
[18] Well C S 1967 Effects of free stream turbulence on boundary layer transition
AIAA J. 5 1724
79
[20] Van Driest E R and Blumer C B 1963 Boundary layer transition, free stream
turbulence, and pressure gradient effects AIAA J. 1 13036
[21] Zhang X and Perot J B 2000 Unsteady flow prediction of turbulent flow
around a triangular cylinder FEDSM 200011172 (Boston, MA, June 2000)
[27] P.A. Durbin and B.P. Reif. Statical theory and modeling for turbulent flows.
John Viley& Sons, 2011.
80