Sunteți pe pagina 1din 28

Pseudospectral simulation of turbulent viscoelastic channel

ow
Antony N. Beris
*
, Costas D. Dimitropoulos
Department of Chemical Engineering, University of Delaware, Newark, DE 19716, USA
Received 14 December 1998
Abstract
The methodology and validation of direct numerical simulations of viscoelastic turbulent channel ow are presented here. Using
dierential constitutive models derived from kinetic and network theories, numerical simulations have demonstrated drag reduction for
various values of the parameters, under conditions where there is a substantial increase in the extensional viscosity compared to the
shear viscosity (Sureshkumar, Beris, Handler, Direct numerical simulation of turbulent channel ow of a polymer solution, Phys.
Fluids 9 (1997) 743755 and Dimitropoulos, Sureshkumar, Beris, Direct numerical simulation of viscoelastic turbulent channel ow
exhibiting drag reduction: eect of the variation of rheological parameters, J. Non-Newtonian Fluid Mech. 79 (1998) 433468). In this
work, new results pertaining to the Reynolds stress and the pressure are presented, and the convergence of the pseudospectral algo-
rithm utilized in the simulations, as well as its parallel implementation, are discussed in detail. It is shown that the lack of mesh re-
nement, or the use of a larger value for the articial stress diusivity used to stabilize the conformation tensor evolution equations,
introduce small quantitative errors which qualitatively have the eect of lowering the drag reduction capability of the simulated uid.
However, an insucient size of the periodic computational domain can also introduce errors in certain cases, which albeit usually
small, can qualitatively alter various features of the solution. 1999 Elsevier Science S.A. All rights reserved.
1. Introduction
One of the most spectacular successes of numerical simulations of Newtonian (viscous) uid ow is
without doubt the capability to produce reliable data under fully developed turbulent conditions [35].
Under these conditions the ow is time-dependent and fully three-dimensional, exhibiting chaotic behavior
(i.e., showing extreme sensitivity on the initial conditions). But in this case, chaotic does not mean random
and there are structures and correlations developed which make the modeling of turbulent ow using
probabilistic descriptions still an open challenge [68]. Thus, direct numerical simulations (i.e., where all the
pertinent length and time scales are resolved) provide the only theoretical means, so far, to simulate tur-
bulent ows from rst principles, with results that are in quantitative agreement with experiments [911].
Advances in computation have helped direct numerical simulations to become the method of choice in
order to generate detailed velocity and pressure data for turbulent ows, which can be used in model re-
duction [12]. In addition to isotropic turbulence [11,13], several sets of data have now become available for
turbulent Newtonian ow in a channel [9], in a tube [10], with a (planar) free surface [14,15], over riblets
[1618], etc. Moreover, data from direct numerical simulations have recently become available for turbulent
viscoelastic channel ow [1,2]. These data sets, which are generated from rst principles for the rst time,
showed that drag reduction can occur if the ow and material parameters have such values so that the
extensional viscosity of the polymer solution can become signicantly larger than the shear viscosity. In
www.elsevier.com/locate/cma
Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
*
Corresponding author. Tel.: 001 302 831 8018; fax: 001 302 831 1048; e-mail: beris@che.udel.edu
0045-7825/99/$ - see front matter 1999 Elsevier Science S.A. All rights reserved.
PII: S 0 0 4 5 - 7 8 2 5 ( 9 9 ) 0 0 1 7 4 - 7
addition, the computational viscoelastic data showed changes from Newtonian turbulence when drag re-
duction was observed, that agreed qualitatively with the experimental observations, which are currently
available only at higher Reynolds numbers [1922]. Moreover, these changes, involving the velocity uc-
tuations, streaks and vorticity uctuations were found to be increasing with an increase in the drag re-
duction [2], as observed experimentally [22].
On the other hand, we need to acknowledge that the presence of multiple time and length scales (with a
range increasing rapidly with the Reynolds number) places severe restrictions on the turbulent ow condi-
tions that can be reliably simulated with xed computational resources, and therefore the Reynolds numbers
utilized are currently limited to relatively small values. These limitations become more enhanced when vis-
coelasticity is involved, due to the higher cost to solve the viscoelastic ow equations for a given mesh res-
olution, as opposed to the Newtonian ones. Not only does the former involve six more equations than the
latter (the components of the stress or the conformation tensor), but they also typically have much smaller
operational window in terms of the size of the time-step increment for stability [1]. In addition, viscoelastic
simulations require the use of numerical diusion (in the stress, or conformation tensor evolution equations
[23]) and they tend to exhibit much stronger spatial correlations (as seen, for example, in the formation of
much more well developed velocity streaks [22]). Thus, we believe that in viscoelastic turbulent direct nu-
merical simulations it is especially important to carefully evaluate the needed resolution as well as to monitor
the sensitivity of the solution to numerical parameters. This task becomes even more important when the use
of over-stabilizing schemes for example, the use of articial (numerical) diusivity makes possible for
smooth solutions to be obtained with under-resolved meshes [2427]. Once the validity of a solution is
demonstrated, then the results can be utilized in a properly scaled fashion (wall scaling for turbulence cal-
culations [28]) to infer conclusions for substantially higher Reynolds numbers with high condence.
The motivation behind the study of turbulent viscoelastic ow lies in the observation that viscoelasticity
(as introduced to the ow through the presence of even minute quantities of polymer additives of the order
of ppm) can result in substantial drag reduction under turbulent ow conditions [19]. However, although the
earliest experimental observations of polymer-induced drag reduction were made more than 50 years ago [29]
and despite a wealth of experimental observations collected since then [19,3032], our predictive capabilities
of the phenomenon are still quite poor, primarily because of the complexity of the ow. Of course, thanks to
the experimental observations, we already have substantial evidence regarding the primary factors behind the
drag reduction. In particular, the primary mechanism has already been proposed in the 1960s [19,33]. Briey,
the increased extensional viscosity due to the viscoelasticity is considered responsible in reducing the fre-
quency of the wall eddies and increasing their size, which makes them less eective in supporting the tur-
bulence energetically. The primary action of viscoelasticity (at least for dilute polymer solutions) is thus
centered in the boundary between the viscous and inertial sublayers where eddies are formed. This has been
supported both experimentally [20] and recently numerically through direct numerical simulations [1], using
the FENE-P dierential viscoelastic model [34]. Also, additional numerical calculations have helped to
elucidate the role of some of the other rheological properties (in addition to the extensional viscosity char-
acteristics), on the phenomenon of drag reduction [2]. The data collected oer a new window of opportunity
to understand polymer-induced drag reduction although clearly a lot more data are still needed in order to
allow us to construct quantitative, predictive, models.
The objective of the present work is to present an analysis of the numerical method utilized so far within
our group for the direct numerical simulation of viscoelastic ows [1,2]. For that, we have used the same
pseudospectral numerical technique as developed in [23] based on the successful spectral methods that have
been developed for direct numerical simulations of Newtonian turbulent ows [35,9,12]. However, the
viscoelastic simulations also involve additional evolution equations governing the chain conformation
statistics which necessitate special numerical integration to guarantee numerically stable results [23]. In this
work we examine the inuence of several numerical parameters on the results.
The rest of the paper is organized as follows. In Section 2 we describe the governing equations. The
numerical algorithm is presented in Section 3. Section 4 consists of the numerical implementation of
the algorithm with emphasis on its adaptation for multiprocessor (parallel) computers and a description of
the simulation conditions. Section 5 contains some representative results and a discussion of convergence
issues related to the inuence of mesh resolution, the magnitude of the articial diusivity and the
dimensions of the periodic computational cell. Finally, in Section 6 we oer our concluding remarks.
366 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
2. Description of the governing equations
We solve for the full three-dimensional and time-dependent ow and internal structure elds of an in-
compressible viscoelastic ow within a channel. Thus, the governing equations consist of a set of ten partial
dierential equations built from the three components of the momentum conservation equation, suitably
modied to accommodate the extra stress due to viscoelasticity, the divergence-free condition and the six
independent components of a constitutive relationship expressing the evolution in time of the internal
structure second-order tensor parameter, c, upon which the viscoelastic extra stress depends.
All equations are expressed in an orthogonal Cartesian coordinate system appropriate to this problem.
In particular, the x coordinate is parallel to the channel walls and in the (streamwise) direction along which
we impose externally a constant pressure gradient, DP=DL, y is along the shear direction perpendicular to
the channel walls and z is along the neutral (spanwise) direction parallel to the channel walls and per-
pendicular to the streamwise direction, x.
As is customary in turbulent calculations [35,9,28], the proper scaling for the velocity (wall scaling) is
specied by the friction velocity, U
s
, which is dened as U
s
=

s
w
=q
p
in terms of the wall shear stress, s
w
,
and the mass density of the polymer solution, q. However, because of the shear thinning, i.e., the fact that in
polymer solutions the shear viscosity is not constant but (typically) decreases with increasing shear rate,
special care needs to be exercised in order to properly dene the wall scaling for the length, L
v
. The mo-
tivation behind the proper length scaling is to preserve the linear law of the wall
U

= y

; (1)
which characterizes the dependence of the average streamwise velocity component (under stationary
conditions) on the distance from the wall within the viscous boundary layer adjacent to the wall, when the
quantities scaled with the wall units (indicated by the superscript + in Eq. (1)) are used (U

is scaled with
respect to U
s
and y

is scaled with respect to L


v
). In order to achieve this we had to scale the y coordinate
through the use of an effective kinematic viscosity computed from the simulation data to account for
possible shear thinning model effects [1,2].
The eective kinematic viscosity m was computed from the simulation time- and space-averaged data for
the average velocity along the x-direction, U, using the relationship
m =
bg
0
dU
dy
y=0
(1 b)T
xy;y=0
q
dU
dy
y=0
; (2)
where m
0
denotes the zero-shear kinematic viscosity. However, this makes the friction Reynolds number,
dened as the dimensionless distance of the centerline from the wall using wall units, uid-dependent. So, in
the simulations we rather use the centerline distance from the wall (half-width) h as our length scale and we
only rescale the results with respect to the wall unit of length a posteriori. Moreover, for convenience, we
keep the zero shear friction Reynolds number, Re
s
0
=

s
w
q
_
h=g
0
, where g
0
is the zero shear rate viscosity,
constant in our simulations. We chose to use the zero shear friction Reynolds number instead of the actual
friction Reynolds number since specication of the later requires a collection of turbulent data in order to
dene the effective shear viscosity. The drawback of that selection is that the data obtained for uids of
different shear thinning correspond to different friction Reynolds numbers and they are not directly
comparable. However, as it turns out, the differences in friction Reynolds numbers are small if the sim-
ulations involve dilute polymer solutions, since they only exhibit small shear thinning. Indeed, the shear
viscosity of a dilute polymer solution can at maximum decrease to the proportion of the solvent viscosity to
the total (at high shear rates when the polymer contribution has effectively decayed to zero), a parameter
symbolized here as b, which typically ranges from values of 0.91. For such small shear thinning the scaling
provided by the wall units makes the scaled quantities highly comparable and the need for additional
simulations at different zero shear friction Reynolds numbers is avoided. The quantity that varies the most
is the mean Reynolds number, but even for that one, the largest difference was calculated to be at most
2.5% obtained by shifting its values through the Blasius friction law (Re
m
Re
8=7
s
) [2].
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 367
Finally, correspondingly to the half-width length scale, we take h=U
s
as our time scale. Utilization of
these scales and g
0
U
s
=h for the stress, leads to the following dimensionless equations for the conservation of
momentum and mass
o~v
ot
= ~v ~ x \~ p
b
Re
s
0
\
2
~v
1 b ( )
Re
s
0
\
~
T e
x
; (3)
\ ~v = 0; (4)
where ~v, ~ x, ~ p and
~
T are the dimensionless velocity, vorticity, pressure and the viscoelastic contribution to
the total stress, respectively. The pressure is scaled by the wall shear stress s
w
. The last term in Eq. (3)
represents the constant, mean pressure drop per unit length across the channel, which in the dimensionless
units used here is represented by one. The parameter b is the ratio of the solvent g
s
( ) to the total zero-shear
rate solution viscosity, g
0
. Finally, note that the viscoelastic extra-stress tensor T is made dimensionless
using g
p0
U
s
=h = g
p0
=g
0

g
0
U
s
=h = 1 b ( )g
0
U
s
=h, where g
p0
(= g
0
g
s
) is the polymer contribution to the
total zero-shear rate solution viscosity.
Eqs. (3) and (4) are not complete without a specication for the viscoelastic extra stress, T. This is ac-
complished in this work by providing two equations. With the rst equation, the viscoelastic extra-stress is
obtained explicitly in terms of a macroscopic measure of the polymer deformation. This measure is taken to
be the second-order conformation tensor which attempts to model the second moment of the orientation
distribution function for the end-to-end chain distance. Therefore, the second equation is naturally an
evolution equation for the conformation tensor. Although for the constitutive models used here (the FENE-
P and the Giesekus models) one could have eliminated the conformation tensor by solving the rst equation
for it and then substituting it in the second, leading to a single equation for the extra-stress tensor, we have
opted to use the two equations explicitly in the numerical simulations for two reasons. First, in order to
allow future use, with minimum code changes, of other models, for which the conformation tensor back-
substitution is not possible [35,36]. Second, in order to allow for an explicit representation of the confor-
mation tensor. Not only can the conformation tensor be used to provide insight on the chain extension and
orientation during the ow, but it also needs to be monitored to control the approximation error in the
simulation. Indeed, as demonstrated in a previous investigation [23] the approximation error during the
simulation typically rst manifests as the loss of positive deniteness of the conformation tensor. Such an
event is not only aecting the physical interpretation of the results (since a non-positive denite second
moment is aphysical) but rst and foremost the stability of the numerical scheme. In fact, when the con-
formation tensor is not positive denite, it has been shown that for a variety of models [35,37] the evolution
equation for the conformation tensor (or the resulting evolution equation for the extra stress, should the
conformation tensor be eliminated) ceases to be evolutionary. When this happens, Hadamard instabilities
develop which quickly cause the calculations to fail [38]. In order to guard against such an event, an articial
diusivity is added to the original evolution equations for the conformation tensor. This additional dif-
fusivity has been shown in the past to be eective in stabilizing the calculations of ow transients without
adversely aecting the accuracy of the results, provided its magnitude is kept low [23]. Some limited data
that address the eect of the diusivity in the calculations are also available [1,2]. In this work the eect of
the diusivity is further examined by looking at additional data and for longer integration times.
The FENE-P model is a constitutive equation that has been developed especially for dilute polymer
solutions using the Peterlin (pre-averaging) approximation in the kinetic theory of Finitely Extensible
Nonlinear Elastic (FENE) dumbbells [39,40]. In addition to a polymer viscosity g
p
and a relaxation time k,
this model also possesses an additional characteristic parameter L which is proportional to the maximum
extension of the polymer chain. Therefore, it controls the maximum increase of the extensional viscosity
observed at high extensional rates and the Trouton ratio (i.e., the ratio of the extensional to the zero-shear
rate polymer viscosity) increases from 3 (the Newtonian value) to 2L
2
at high extensional rates [39,40]. In
addition, we have also used the Giesekus model which has been developed for concentrated polymer
solutions and melts based on a network theory and using an anisotropic polymer mobility [35,41,42]. The
Giesekus model also possesses an extra parameter a in addition to the polymer relaxation and viscosity. For
a = 0 the upper convected Maxwell model is recovered, which suffers from the prediction of an unbounded
368 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
extensional viscosity for a nite extensional rate. However, as a increases from 0 (physical values are ob-
tained for 0 6a 60:5) the model starts shear thinning and at the same time its extensional viscosity is
bounded. The predicted Trouton ratio increases with extensional rate, approaching 2=a at high extensional
rates. The Giesekus model also introduces (for non-zero a) a non-zero second normal stress difference W
2
in
simple shear ows, which is shear thinning (as the rst normal stress and the viscosity). Its ratio with the
rst normal stress coefcient reaches an asymptotic value approximately equal to a=2 at high shear rates.
The investigation of the role of the second normal stress in drag reduction was the subject of a previous
publication, and was actually one of the reasons for introducing the Giesekus model in the viscoelastic
turbulent ow simulations [2]. This was motivated from the role that W
2
plays in the stability of simple
laminar viscoelastic ows like the TaylorCouette, where negative W
2
values have been found to stabilize
the ow [4345]. In addition, W
2
has been proposed to play a critical role for turbulent ows [46]. However,
results obtained so far with the Giesekus model showed only secondary differences from those obtained
with the FENE-P model and for such parameter values that there is a match in the maximum attained
extensional viscosity at high extensional rates [2]. Nevertheless, the Giesekus model appears to offer some
important quantitative differences (for example, the results for the average velocity show a different slope in
the log-linear regime as well as a different intercept) that seem to follow experimental trends associated with
concentrated polymer systems (as opposed to dilute), for which the use of the Giesekus model should
therefore be preferred from that of the FENE-P model. Of course, we also need to mention here that both
models, Giesekus and FENE-P, involve only one characteristic relaxation time. Even though this may be
adequate in order to qualitatively capture the observed phenomena, a spectrum of relaxation modes needs
to be utilized for a quantitative analysis. This can be implemented in a straightforward fashion by simply
carrying out the calculations outlined below for each mode separately. Following the indicated procedure,
this will increase the workload and storage requirements proportionally, at most, to the number of utilized
modes. Such calculations are now quite common place for laminar ows [47], however, in our turbulence
calculations, we have initially focused, for simplicity, only at models with one relaxation mode.
We now proceed to describe the equations used that correspond to the aforementioned models. In the
FENE-P model, the viscoelastic extra-stress tensor,
~
T, is a non-linear function of the conformation tensor ~c.
More specically,
~
T =
~
f (~r)~c 1
We
h
; We
h
=
kU
s
h
=
kU
2
s
m
0
m
0
U
s
h
=
We
s
Re
s
0
; (5)
where We
s
, the Weissenberg number, is a dimensionless relaxation time dened as the product of the
polymer relaxation time and a characteristic shear rate We
s
= kU
2
s
=m
0

based on its turbulence scale. In the
above equation, the conformation tensor is made dimensionless with k
B
T=H, where k
B
is the Boltzmann
constant, T the absolute temperature and H the spring constant of the FENE-P dumbbell. The function
~
f ~ r ( ) is the Peterlin function [48], dened as
~
f (~ r) =
L
2
3
L
2
~r
2
; (6)
where ~ r is the root mean square chain extension, i.e., ~r
2
= trace ~c ( ), and L is the maximum chain extensi-
bility. Note that in Eq. (6) there is a small correcting factor (3) in the numerator, added for consistency, as
discussed in [35]. The equilibrium value of L is

3
_
.
The corresponding evolution equation for the conformation tensor ~c is given by
o~c
ot
(~v \)~c [~c (\~v) (\~v)
,
~c[
j
U
s
h
\
2
~c =
~
f (~r)~c 1
We
h
: (7)
As mentioned before, this equation contains, in addition to the classical terms on the left-hand side that
constitute the upper convective time derivative, an articial diusive term for the stress, added to ensure the
stability of the numerical integration of Eq. (7). In earlier work [23], we have shown that if the magnitude of
this term is kept reasonably small (j=(U
s
h) ~ Dh
2
=Dt, where Dh is a characteristic mesh size and Dt the
time-step), convergence is established as Dh ~ Dt 0. We should note here that the addition of this term
gives a streamwiseupwinding (SU) character in the numerical method, similar to that proposed by Crochet
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 369
[24], but in contrast to that work we have found that due to the chaotic nature of the turbulent ow it is not
necessary to utilize a non-isotropic numerical diusivity along the streamline direction. This is fortunate,
because a non-isotropic diusion term would have not allowed the use of fast Poisson solvers for the
implicit time-integration of this term. Finally, the extra boundary conditions for the conformation tensor
that are now needed once the articial diusivity is added to the equations, are generated by integrating in
time the original evolution equation (without articial diusivity) at the solid boundaries [23].
The introduction of the Giesekus model changes the equations for the viscoelastic extra stress and the
evolution equation for the conformation tensor to
~
T =
~c 1
We
h
(8)
and
o~c
ot
~v (\~c) [~c (\~v) (\~v)
,
~c[
j
U
s
h
\
2
~c =
[1 a(~c 1)[ (~c 1)
We
h
; (9)
where the parameter a is related to the extensibility of the polymer chains. Note the presence of the stress
diusivity term, introduced here for stability reasons, exactly as in the FENE-P model.
Eqs. (3), (4) and either Eqs. (5)(7) or Eqs. (8) and (9), along with the no-slip boundary conditions for
the velocity fully describe the turbulent viscoelastic channel ow we are simulating by solving them nu-
merically.
As far as the notation to be used in the subsequent sections is concerned, average values (in space or
time) of a ow variable, say ~q (~q can be ~v, ~c,
~
T, ~ p) will be denoted by capitals (Q), whereas uctuations will
be denoted by lowercase letters (q). Notice that the ow we are examining is on the average a shear ow
V = U(y)e
x
, where e
x
is the unit vector along the streamwise (x) direction.
3. Numerical algorithm
In the viscoelastic, turbulent channel ow simulations we used a pseudospectral numerical algorithm
that was originally developed by Beris and Sureshkumar [38], based on a time-splitting semi-implicit
spectral technique similar to that used for the direct numerical simulation of turbulent Newtonian channel
ow [35]. The computational domain has the shape of a rectilinear parallelepiped of (base) dimensions
10h 2h 5h in the streamwise (x), shear (y) and spanwise (z) directions, respectively, where h is the
channel half-width. Periodicity conditions were applied along the streamwise and spanwise directions, with
their respective sizes, 10h and 5h selected based on earlier work on Newtonian turbulent ow simulations
under similar ow conditions [50] that have indicated that those sizes are adequate in order to capture the
streaky structures and the elongated vortical structures, developed during the turbulent ow [1]. However,
as the structural features of the ow increase signicantly in size with increasing drag reduction [2], larger
periodic cells are also used in this work in order to test the effect of the periodic conditions on the ow
simulations. The basic mesh resolution used in our simulations was 64 65 64 along x; y; z, respectively,
following the suggestions of earlier work under similar ow conditions that had demonstrated this reso-
lution sufcient to obtain sustained turbulence both in Newtonian and viscoelastic simulations [1,2,50]. To
check the convergence of the results with respect to various parameters, selected runs were also performed
at ner resolutions. In the two periodic directions, x; z, Fourier representations were used, whereas in the
non-homogeneous shear direction, y, an expansion in terms of orthogonal Chebyshev polynomials was
employed.
The integration of the governing equations in time is based on a time-splitting inuence matrix method
similar to those used in Newtonian ow calculations [35]. A second order AdamsBashforth method was
used for the explicit update of the non-linear terms and a second order AdamsMoulton scheme for the
implicit update of the linear terms. Then, it can be shown that the velocity ~v at the (n 1)th time step can be
written as
370 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
~v
n1
~v
n
=
Dt
2
[3(~v ~ x)
n
(~v ~ x)
n1
[ (1 b)
Dt
2Re
s
0
[\
~
T
n1
\
~
T
n
[ Dt\~ p
n1=2
b
Dt
2Re
s
0
[\
2
~v
n1
\
2
~v
n
[ e
x
; (10)
where Dt is the time step size and ~ p is calculated from the divergence of Eq. (11), which is simply a Poisson
equation. The boundary conditions for the pressure Poisson equation are constructed using the inuence
matrix method [23,38,49]. Briey, the boundary conditions for the solution of the Poisson equation for p
are evaluated imposing the requirement that \ ~v = 0 on the boundaries. Taking the Fourier transform of
that equation along the two periodic directions leads to a decoupled set of equation pairs, two equations for
each combination of Fourier modes, i and k, involving the ith and kth Fourier transforms of the velocity
evaluated at the two wall boundaries. To use those equations to evaluate the yet unknown boundary
conditions for the normal derivatives of the pressure, we take advantage of the linearity to express the
solution for the velocity as the linear superposition of a known term calculated based on the inhomoge-
neous right-hand-sides in the bulk equations, but with homogeneous (zero) boundary conditions for the
velocity components and the normal derivatives of the pressure, plus a series involving velocity ``Green's
functions'' weighted by the unknown pressure boundary conditions. The Green's functions are calculated
based on the homogeneous (zero right-hand side) part of the governing equations and boundary conditions
for the pressure which involve a normal pressure of unity only for one of the Fourier components at one of
the boundaries and zero everywhere else. Since the various Fourier modes are decoupled, the end result is
that each pair of equations representing a given Fourier transform of the divergence-free condition at the
two solid boundaries only involves two of the Green's functions weighted by the corresponding Fourier
Fig. 1. The parallelization scheme used. The data are distributed in sets of planes amongst the processors (whose number is a power of
2) and are split along the x direction in spectral space and along the z direction in physical space. This minimizes the use of com-
munication in the XYZFFT routine used to transform the data from physical to spectral space and vice versa.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 371
transforms of the normal derivatives of the pressure at the two boundaries, which therefore can be solved
easily thereafter see [23,38,49] for more details.
The above set of equations leads to an update of the velocity components which is obtained in three
stages. First, the contribution to the velocity from the inertial terms and the viscoelastic extra stress is
calculated. The contribution of the inertial terms since it involves an explicit time-integration can be cal-
culated directly from the velocity data obtained at the two previous time steps. The contribution of the
viscoelastic extra stress is evaluated from the conformation tensor using Eq. (5) or Eq. (8) after it is cal-
culated at the new time step, following the procedure outlined in the next paragraph. The second step
involves evaluation of the pressure which is performed by solving a Poisson equation constructed by taking
the divergence of Eq. (11), using the fact that the divergence of the velocity at the new time step is zero.
Finally, the contribution of the viscous terms is evaluated in the third stage of the calculations by solving a
set of Helmholtz equations, one for each of the velocity components.
Note that as mentioned above, the calculations in stages two and three are repeated twice, rst in order
to obtain the values of the Green's functions of the velocities (this set of calculations is done only once in
the beginning of each run and the results are stored for subsequent use) and second in order to get the
values of the velocities corresponding to zero normal pressure gradient boundary conditions. The actual
normal pressure gradient values are then subsequently evaluated by solving several sets of 2 2 linear
equations (one for each Fourier transform in the periodic x, z directions) as explained above, following the
inuence matrix technique [23,38,49].
The conformation tensor components at the new time step are evaluated in a similar multistage fashion,
by expressing the discretized equations for ~c as
Fig. 2. The scalability of the parallel implementation of our algorithm in the SGI/CRAY Origin 2000. For dedicated use, we see ideal
performance for 32 processors and a small deviation for larger numbers, where external routers and Craylink cables are utilized for
interprocessor communication. For time-shared use, the speedup was not linearly scalable for more than eight processors and dete-
riorates with size, practically limiting simulations to 16 processors for adequate performance gain.
372 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
~c
n1
= ~c
n

Dt
2
(3
~
F
n

~
F
n1
)
jDt
2U
s
h
[(\
2
~c)
n
(\
2
~c)
n1
[; (11)
where
~
F = (~v \)~c ~c \~v \~v
,
~c
~
U. The term
~
U is dened by the right-hand side of Eqs. (7) and (9),
for the FENE-P and Giesekus models, respectively.
Eq. (11) allows us to evaluate ~c
n1
in two stages. In the rst stage, we update the components of the
conformation tensor to account for all the terms except the diusive one, generating ~c
n1=2
:
~c
n1=2
= ~c
n

Dt
2
(3
~
F
n

~
F
n1
): (12)
In the second stage, the diusive term is accounted for implicitly, by solving the Helmholtz equation
\
2
~c
n1

2U
s
h
jDt
~c
n1
= \
2
~c
n

2U
s
h
jDt
~c
n1=2
: (13)
The boundary conditions used here to solve Eq. (13) is given by
~c
n1
= ~c
n1=2
for y = 1: (14)
As mentioned before, the updating procedure for ~c needs to precede the evaluation of velocity at the
(n 1)th step following the three-stage procedure outlined above. Except for the addition of the viscoelastic
extra stress contribution, the procedure is a standard one for spectral time-dependent simulations, rst
Fig. 3. The Reynolds shear stress for Newtonian, FENE-P (L = 10, b = 0:9) and FENE-P (L=30, b = 0:98874). In both viscoelastic
cases j=(U
s
h) = 0:01 and We
s
= 50. The purely viscous and viscoelastic contributions to the total shear stress are also plotted along
with the total stress for both viscoelastic ows. The mesh size is 64 65 64.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 373
introduced for simulating Newtonian turbulence [35]. A more detailed description can be found in section
4:1 of Beris and Sureshkumar [38].
4. Numerical implementation and simulation conditions
The need for ne three-dimensional spatial discretization and small time-step size for the time inte-
gration leads to signicant computational requirements that make the use of a parallel supercomputer
necessary. For the simulations presented in this paper a SGI/CRAY Origin 2000 parallel supercomputer
was used. In order to implement eciently the numerical algorithm in a parallel environment, we employed
the Message Passage Interface (MPI) in order to provide reliable interprocessor communication. The key
for a successful parallelization was the minimization of the interprocessor communication and the very
good (almost 100%) load balancing, made possible by exploiting the structure of the computations within
our numerical algorithm.
Briey, exploiting the regular structure of the three-dimensional data sets manipulated in our work, we
assign to each processor the same computational tasks, to be performed however on a dierent subset of the
data. Provided that the number of processors used, np, divides exactly the number of half modes along the x
direction, nxh, and the number of modes in the z direction, nz, respectively, we can assign to each processor
kxh = nxh=np or kz = nz=np planes of data, depending on whether the data arrays hold information in the
physical or spectral space (see Fig. 1). Then, all the data manipulations are performed locally on the data
sets within each processor in exactly the same fashion as in a single processor machine. This ensures
simplicity of programming, no need for additional redundant data manipulations and almost 100% load
balancing. The only overload that comes from the parallelization happens during the Fast Fourier
Fig. 4. The Reynolds shear stress for Newtonian, FENE-P (L = 30) and Giesekus a = 0:01). In both viscoelastic cases b = 0:9,
j=(U
s
h) = 0:01 and We
s
= 50. The purely viscous and viscoelastic contributions to the total shear stress are also plotted along with the
total stress for both viscoelastic ows. The mesh size is 64 65 64.
374 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
Transforms (FFTs), when the data need to be re-shufed between all the processors. This is where we make
use of the MPI routines in order to perform the data transfer, implemented within a three-dimensional
FourierChebyshev transform routine (XYZFFT). During the transfers, since all the data along the y-
direction are always kept together, we can transfer the data in blocks along that direction. Given the or-
dered distribution of data along the z and x directions, the number of global transpose operations is
minimized, the necessary operations for communication are quite straightforward to implement and the
interprocessor data transfers can be performed quite efciently and with a small penalty on the overall code
performance. All the interprocessor communications take place within the three-dimensional XYZFFT
routine. This is also the place where most of the computations are done, growing slightly more than linear
with the number of unknowns, O(N log
2
N). The algorithm's performance is therefore dictated by the
performance of the XYZFFT. Noting that the skew-symmetric form is used for the non-linear terms, we
have a requirement to perform 21 XYZFFT calls per time step for the Newtonian case, 81 for the FENE-P
model and 63 for the Giesekus model. Thus, viscoelastic turbulence calculations are more demanding than
Newtonian by a factor that depends on the complexity of implementation of the constitutive equation for
the stress. However, almost linear scalability with the number of unknowns is retained and the challenge is
to implement the algorithm in a parallel environment in such a way that it also scales linearly with the
number of processors over a wide range. Our implementation satises both requirements and the resulting
algorithm is capable of being used for performing efciently a wide range of large scale computations to
study various cases of viscoelastic turbulent channel ow.
Specically, an estimate of the parallelization performance achieved can be obtained by following the
speedup realized as we increase the number of processors, while we keep the size of the problem the same.
This is indicated in Fig. 2 where it can be seen that for the size of the problem investigated (128 65 128)
Fig. 5. The average pressure distribution between the channel walls for Newtonian ow and ow with the FENE-P model for L = 10
and L = 30. In both viscoelastic cases b = 0:9, j=(U
s
h) = 0:01, We
s
= 50. The mesh size is 64 65 64.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 375
and when the simulations are performed in dedicated mode, a linear scaling was obtained up to 32 pro-
cessors and a slightly sub-linear up to 60 processors. Here it is worthwhile to note that for 16 processors we
have super-linear performance, probably due to the better use of the cache. In fact, if our reference was not
the performance for 8 processors, but that for 1 processor (we did not generate data for smaller numbers of
processors for practical reasons), it is highly possible that we would have seen similar super-linear behavior
for a wider range of processor numbers, especially for smaller mesh sizes. In addition, the large runs for 60
processors were performed for a 120 65 120 mesh, due to the constraints imposed by our parallelization
scheme. We did not use 64 processors because the runs were performed on a dedicated machine with 64
processors and it is best to leave some resources free for the operating system in order to maximize per-
formance in machines such as the SGI/CRAY Origin 2000. The small decrease in performance for more
than 32 processors is due primarily to the use, due to the machine architecture, of external routers and
Craylink cables for interprocessor communication in these congurations. However, note that for time-
shared mode of operation, where the system is shared with other users, we saw considerably degraded
parallelization performance as one can see in Fig. 2. This has led us to limit all simulations in time-shared
mode to 16 processors, since this size seems to be the limit where one can achieve high enough performance
gain to justify the larger allocation of additional processors.
Somewhat better performance can possibly be achieved using the native shared-memory (SHMEM)
routines. However, their usage requires a dedicated use of the multiprocessor system for robust perfor-
mance, and even then there were instances of unreliable behavior given the fact that in the SGI/CRAY
Origin 2000 they are not implemented with hardware barriers like in the CRAY T3E. It was decided that it
was not practical to sacrice the portability of MPI in favor of an elaborate scheme to ensure proper
Fig. 6. The root mean square pressure uctuations between the channel walls for Newtonian ow and ow with the FENE-P model
for L = 10 and L = 30. In both viscoelastic cases b = 0:9, j=(U
s
h) = 0:01, We
s
= 50. The mesh size is 64 65 64.
376 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
communication, which will inevitably have required an additional communication load for data validation
and thus decrease or annul any potential performance gain.
The dimensionless time-step used in these simulations was of the order of 10
4
dened using the channel
half-width as a length-scale (h=U
s
), which is equivalent to 0.025 using turbulence scales (m
0
=U
2
s
) and typi-
cally 5 times smaller than that used in Newtonian turbulent channel ow simulations. As initial conditions
for the simulations, we have used data corresponding to either Newtonian or viscoelastic ows which had
been simulated previously [1,2]. All these ows correspond to a friction Reynolds number Re
s
0
= 125, for
which the mean ow Reynolds number based on the channel half-width Re
m
(= Uh=m
0
) in the Newtonian
case is approximately 1840. This value for Re
s
0
is close to the maximum we can have in order to resolve the
dissipative time scales for the specied computational resolution. At this Re
s
0
, it has been shown that one
can obtain fully sustainable turbulent ow [50]. Drag reduction has been observed experimentally at higher
Reynolds numbers (Re
m
P8900) [20], so in this work we are compensating by using more elastic uids
through a high value of the relative viscoelastic contribution to the stress, corresponding to 1 b = 0:1, this
parameter being physically proportional to the polymer concentration. The Weissenberg numbers (We
s
)
used in this work is 50, high enough to show an appreciable viscoelastic eect. For the FENE-P model we
simulated ows at values of the parameter L equal to 10 and 30, whereas for the Giesekus model the
parameter a was equal to 10
2
. In both cases, these values were chosen to investigate the effect of increasing
the extensional viscosity of the system. In addition, in order to investigate separately the effect of the
molecular weight, roughly proportional to L
2
, versus that of the concentration, we also performed simu-
lations for b = 0:99874, for high chain extensibility (FENE-P model; L = 30) and Weissenberg number
(We
s
= 50), the parameter b chosen here so that the maximum extensional solution viscosity is the same as
before for b = 0:9, L = 10. In all these cases the stress diffusivity j=(U
s
h) was taken to be 10
2
in accor-
dance to what is discussed in [23,1]. The mesh size was 64 65 64 for most simulations. The effect of the
numerical diffusivity j=(U
s
h) was studied by performing also simulations with this parameter equal to
2 10
2
. However, we also studied the effect of mesh-renement by doubling the mesh in all directions
(128 129 128). Finally, we also studied the effect of the size of the computational domain by repeating
Fig. 7. The mean ow velocity as a function of the distance from the wall for the FENE-P model with L = 10, b = 0:9, j=(U
s
h) = 0:01,
We
s
= 50 and two dierent meshes, 128 129 128 and 64 65 64. Also shown is the mean ow velocity for Newtonian ow with
64 65 65 mesh and the respective semi-logarithmic asymptotes in the logarithmic layer.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 377
the simulations for the FENE-P model with b = 0:9, L = 30 and j=(U
s
h) = 10
2
with a 80 65 80 mesh
and then increasing the size of the streamwise and spanwise directions by a factor of 1:5 and using a mesh
equal to 120 65 120 in order to preserve the same mesh resolution.
5. Results and discussion
5.1. New results of drag reduction-induced changes in the ow structure
In previous work, Sureshkumar et al. [1] presented data from simulations with the FENE-P model for
various Weissenberg numbers and for b = 0:9 and L = 2; 10. There it was shown that drag reduction sets in
at the higher value of L = 10 after a critical Weissenberg number, We
cr
, 12:5 < We
cr
< 25. The drag re-
duction was found to be 15% (using Dean's correlation [51]) for the highest We
s
used in those calculations,
We
s
= 50 (based on estimates for the same ow rate the simulations always involved the same pressure
drop and monitor the ow). More recently, Dimitropoulos et al. [2] have shown similar drag reduction,
obtained with a dierent combination of the parameters of FENE-P model, namely b = 0:98874 and
L = 30. These parameter values were chosen to correspond to the same limiting value for the extensional
viscosity of the solution, obtained however with a less concentrated solution (the concentration being
proportional to 1 b) and a higher molecular weight polymer (the molecular weight increasing as L in-
creases). In addition, similar values for the drag reduction were obtained with the Giesekus model as well.
Actually, the percentage of drag reduction was slightly higher, around 19%, at We
s
=50) using a value for
the parameter a equal to 0:01, which matches its limiting extensional viscosity to that of the FENE-P,
L = 10 and b = 0:9 [2]. In contrast, the same authors have seen dramatic increases in drag reduction, up to
44% for the same We
s
=50, using an even higher value for L with the FENE-P model, L = 30, corre-
sponding to a substantially higher extensional viscosity (remember that the limiting value of the extensional
viscosity at high extensional rates increases proportionally to L
2
[40]) [2].
Fig. 8. The root mean square velocity uctuations as a function of the distance from the wall for Newtonian ow with 64 65 64
mesh and the FENE-P model for L = 10, b = 0:9, j=(U
s
h) = 0:01, We
s
= 50 and two dierent meshes, 128 129 128 and
64 65 64.
378 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
The above observations provide signicant evidence about the primary role played by the extensional
viscosity in determining the extend of drag reduction. Detailed data on the mean ow Reynolds number,
average velocity and the root-mean-square (RMS) statistics of the uctuating velocities and vorticity are to
be found in the above mentioned works [1,2]. Here we would like to provide another angle not covered in
those investigations, namely a comparison on the average (along the x and z directions and in time)
contributions to the shear stress of the Reynolds stress, the viscous and the viscoelastic extra stress as a
function of the distance from the wall of the channel and a comparison on the average and RMS uctu-
ations for the pressure. These effects are discussed in the following subsections.
5.1.1. Averaged stress contributions to the shear stress
In Fig. 3 a comparison is made between the predictions for the average stress contributions at We
s
=50
of two dierent versions of the FENE-P model used in the previous investigations, one corresponding to
L = 10; b = 0:9 and the other to L = 30; b = 0:99874. As Fig. 3 shows, the Reynolds stress contributions are
very similar and signicantly lower to those obtained in the simulation using a Newtonian uid (indicated
with a dashed line in Fig. 3). Not surprisingly, the results of the more dilute system (FENE-P with
b = 0:99874) are closer to the Newtonian values, in fact they are almost identical to the Newtonian values
as we approach the centerline of the channel. The fact that we still see the same drag reduction with the two
FENE-P models, is additional evidence to the key role that the boundary region between the viscous and
inertial sublayers plays (around y=h ~ 0:2 in the gure), where most of the changes from the Newtonian
data are observed. As far as the other data are concerned, the observed dierences between the two FENE-
P models can be explained given the fact that the b = 0:99874 version has a higher solvent viscosity and a
smaller polymer viscosity contribution (for shear dominated ows). Note that the polymer contributions of
the two models are almost identical as we approach the centerline where the character of the (time-de-
pendent and uctuating) ow is extensional, given the fact that the b = 0:99874 case has a higher value for
the molecular extensibility (L = 30), chosen so that it exactly compensates for the lower polymer
Fig. 9. The root mean square vorticity uctuations as a function of the distance from the wall for Newtonian ow with 64 65 64
mesh and the FENE-P model for L = 10, b = 0:9, j=(U
s
h) = 0:01, We
s
= 50 and two dierent meshes, 128 129 128 and
64 65 64.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 379
concentration at high extensional rates. Thus, this matching of the viscoelastic stress as the centerline of the
channel is approached, is one more independent indication of the extensional character of the ow kine-
matics in that region. Finally note that as far as the total shear stress is concerned, both models should have
given a straight line decreasing from one at the wall surface to zero at the centerline. The observed devi-
ations (especially obvious for the b = 0:9 case) need to be rationalized in terms of an insucient time
averaging of the data-typically around 7 large eddy turnover times, or around 1000 turbulent time units
used here. Unfortunately this insucient time averaging will also be observed in several of these runs, since
given the fact that we wanted to examine several sets of parameters it was not computationally feasible to
do large averages. Should more quantitative results be necessary, more extended averages need to be
performed, and in that respect, monitoring the average shear stress can provide an internal self-consistency
test for the adequacy of the time integration (however, other tests need also be taken into account).
In Fig. 4 we show a comparison between the predictions made, at the same We
s
, ( =50), and for the same
solvent viscosity, b = 0:9, by two dierent models, the Giesekus model with a = 0:01 and the FENE-P
model with L = 30. In this case, since the two models predict very dierent extensional and shear behavior
(the Giesekus model shear-thins more and extensional-thickens less than the FENE-P, L = 30) we see quite
dierent predictions. As Fig. 4 shows, the Reynolds stress contributions are very dierent, the observed
extent of the decrease from the Newtonian values (indicated with a dashed line in Fig. 4) being proportional
to the observed drag reduction (19% for the Giesekus and 44% for the FENE-P, L = 30), which also follows
closely the extensional viscosity of the two models, that of the Giesekus model being about nine times less
than that of the FENE-P L = 30 in the limit of high extension. This is also consistent with the important
role played by the extensional viscosity in this problem. As far as the other data are concerned, the observed
dierences in the viscoelastic stress are explained by the fact that the FENE-P L = 30 model shows less
shear thinning and has a higher extensional viscosity than the Giesekus a = 0:01. This also explains why the
viscoelastic stress obtained with the former is always larger than that obtained with the later for all values
of y from the wall, where the ow is shear-dominated, to the centerline where it has a highly extensional
character. Interestingly enough, we also see a signicantly higher viscous contribution for the FENE-P than
Fig. 10. The trace of the conformation tensor as a function of the distance from the wall for the FENE-P model with L = 10, b = 0:9,
j=(U
s
h) = 0:01, We
s
= 50 and two dierent meshes, 128 129 128 and 64 65 64.
380 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
for the Giesekus model, despite the fact that for both models the solvent viscosity contribution is the same
(b = 0:9). This needs to be attributed solely to the changes in the kinematics introduced by the drag re-
duction, especially the large decrease in the Reynolds contribution, which is picked up by an increased
viscous contribution, higher than the corresponding one for even a purely Newtonian uid which is close to
that reported by the Giesekus model (not shown in Fig. 4). This demonstrates the nonlinear character of the
drag reduction, where most of the changes need to be accounted for in the dynamics of the eddies, well
hidden in the average data shown in Fig. 4. Finally, as in Fig. 3, the observed deviations from the straight
line for the total shear stress contributions (especially obvious for the FENE-P case) need to be rationalized
in terms of an insufcient time averaging of the data.
5.1.2. Eect of drag reduction on the pressure
The pressure eld can be back-calculated from the velocity eld in a post-processing step by solving a
Poisson equation similar to the one used in stage 2 of the time-integration algorithm (see for example [52]).
In this case, the pressure boundary conditions at the wall are of Neumann type, with the values for the
normal derivatives supplied by evaluating the y-component of the momentum equation at the wall
boundaries. Note that due to the relatively small friction Reynolds number used in this work, it is im-
portant to consider non-homogeneous Neumann boundary conditions for the pressure, as of course we
have also done consistently in advancing the solution in time, using the inuence matrix procedure ex-
plained in Section 3. The effect of the non-homogeneous boundary conditions for the pressure was studied
before, for a Newtonian uid, by Kim [52] who showed that for a friction Reynolds number of 180 their
contribution to the RMS value of the pressure uctuations was of the order from 10% to 20%. Since they
scale inversely with a friction Reynolds number, they are therefore anticipated to have an even greater
weight in the simulations presented in this work which involve a lower friction Reynolds number.
Figs. 5 and 6, show the average and RMS uctuation values, respectively, for the pressure for a FENE-P
viscoelastic uid at We
s
= 50, b = 0:9 and two values of the chain extensibility parameter, L = 10 and
L = 30. For comparison purposes, the data corresponding to a Newtonian simulation for the same zero
shear friction Reynolds number (125) are also shown. First, an enhanced asymmetry of the viscoelastic
results is observed as compared to those of the Newtonian simulations with respect to the channel cen-
terline that it is attributed to the shorter time over which the data were integrated. Clearly, for more
quantitative predictions we need additional data in time. Nevertheless, even with the existing data, a clear
pattern emerges, according to which as the extensibility of the solution increases (and correspondingly the
drag reduction eect), both the average value of the pressure and the RMS pressure uctuations get de-
pressed in magnitude and vary less along the shear direction in the channel. Parenthetically, evaluation of
the pressure values was found to be one of the most dicult computational tasks, sensitively dependent on
the accuracy of the satisfaction of the divergence-free condition. We therefore anticipate the reported values
to be sensitive to the convergence of the overall solution, which is the primary reason why we have un-
dertaken a more systematic study of the dependence of the solution to various numerical parameters before
proceeding to a more quantitative (and longer in time) analysis of the viscoelastic turbulent ow problem.
5.2. Convergence study
The rest of our results in this work focus on an investigation of the inuence of various numerical
parameters on the data in an eort to ascertain the limits of accuracy of the numerical calculations and its
rate of convergence. We have already seen in the previous section a signicant eect of the time window
used in the time averaging. Other issues involve the mesh resolution, the magnitude of the articial dif-
fusivity entering the conformation tensor evolution equation, the boundary conditions for the pressure and
the size of the periodic computational domain. These are addressed in the subsections that follow.
5.2.1. Eect of the mesh size
In viscoelastic turbulent ow simulations, as in the Newtonian turbulent simulations, an important
component is the adequacy of the mesh size, i.e., whether it is ne enough to capture all the important
length scales of the ow. For Newtonian turbulent channel ow simulations at Re
s
= 125 we know from
previous work that a mesh size similar to the one used in this work (64 65 64) is adequate [50].
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 381
However, with viscoelastic turbulent simulations we had performed previously only a limited mesh re-
nement study comparing the results between two mesh sizes, 64 65 64 and 128 129 128 [1]. A
more extensive study is undertaken in this work, also using the same mesh sizes and the FENE-P model for
parameter values We
s
= 50, L = 10 and b = 0:9. In this case, we kept the numerical diusivity constant,
j=(U
s
h) = 0:01, with the inuence of variations of its value examined separately below. Selective results of
this mesh renement study are shown in Figs. 711 where we compared results obtained with the two
above-mentioned mesh sizes. For comparison purposes, results obtained for a purely viscous (Newtonian)
uid with the smaller mesh are also shown.
In general, Figs. 711 indicate qualitatively similar results corresponding to the two dierent mesh sizes.
Quantitatively the dierences observed between the two mesh sizes are small, signicantly smaller to those
observed between the viscoelastic and the Newtonian data. In addition, the results corresponding to the
ner mesh demonstrate higher drag reduction, as seen, for example, in the higher average velocity values
shown in Fig. 7. The higher drag reduction is consistently accompanied by higher RMS velocity uctuation
values in the streamwise direction and smaller RMS velocity uctuation values in the other two directions,
as seen in Fig. 8, and smaller RMS vorticity uctuations, as seen in Fig. 9. Moreover, although the average
molecular extension, which can be assumed to be related to the trace of the conformation tensor, c, appears
unchanged by the mesh renement, as shown in Fig. 10, the RMS values for c
xx
, which are more indicative
of the transient deformation that the chains are subject to, appear again to increase in the ner mesh, as
shown in Fig. 11. This behavior is again consistent to an increased extensional viscosity (which is expected
to accompany higher molecular deformation) and therefore the increased drag reduction observed with the
ner mesh calculations. These results are consistent with the ndings of the previous mesh renement study
[1]. Finally, note that again there is an issue of adequate time integration, especially associated with the ne
mesh study, where the ``wiggles'' in the data have to be attributed. However, as there were no substantial
dierences between the two mesh results, and given the high cost of the computations with the ne mesh, no
further time integration was pursued with the ne mesh in this study.
Fig. 11. The root mean square uctuations of the xx component of the conformation tensor as a function of the distance from the wall
for the FENE-P model with L = 10, b = 0:9, j=(U
s
h) = 0:01, We
s
= 50 and two different meshes, 128 129 128 and 64 65 64.
382 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
5.2.2. Eect of the value of the numerical diusivity
The need to stabilize the calculation of the conformation tensor in the viscoelastic turbulent calculations
was amply shown in previous work [23,38]. The addition of an articial diusion term alongside the
convective terms in the evolution equation for the conformation tensor was shown to be eective in sta-
bilizing the calculations [1,2,23]. However, the eect of the articial diusivity on the accuracy of the re-
ported results was only examined in detail for sample calculations of the eigensolutions in the channel
Poiseuille ow [23] and only examined jointly with the eect of mesh renement in turbulent simulations [1].
In this work, we systematically compare the results obtained with two dierent articial diusivity values,
j=(U
s
h) = 0:01 and j=(U
s
h) = 0:02, with the FENE-P model using a 64 65 64 mesh and parameter
values We
s
= 50, L = 10 and b = 0:9. Our new results expand the brief discussion of the eect of the
numerical diusivity in [2] and reinforce our previous conclusions [1,2,23]. The comparisons are shown in
Figs. 1216.
As seen in those gures, the results obtained with the two dierent numerical diusivity values are very
close. In particular, as far as the average velocity proles are concerned (and therefore also the associated
predictions for the drag reduction) they are indistinguishable, at least within the resolution shown in
Fig. 12. As far as the RMS velocity uctuations are concerned shown in Fig. 13, there are small but sys-
tematic deviations where the results obtained with the smallest diusivity appear to show a more drag-
reducing eect accompanied by higher RMS values for the streamwise velocity uctuations and lower for
the components in the other two directions. Moreover, although no signicant changes appear in the av-
erage molecular extension as represented by the average value of the trace of the conformation tensor c
shown in Fig. 14, there is a signicant enhancement of the RMS uctuations of the c
xx
component of the
conformation tensor observed with the lower diusivity runs, as seen in Fig. 15, indicating a higher de-
formation of the polymer chains allowed during the transient dynamics. In addition, there are some small
but systematic dierences in the average contributions to the shear stress, as seen in Fig. 16, again pointing
out to the results obtained with the smaller articial diusivity as corresponding to a more enhanced drag
reduction. These trends can be justied by the fact that the presence of the numerical diusivity is
Fig. 12. The mean ow velocity as a function of the distance from the wall for the FENE-P model with L = 10, b = 0:9 and We
s
= 50,
for j=(U
s
h) = 0:01 and j=(U
s
h) = 0:02, as well as for the Newtonian ow. Also shown are the respective semi-logarithmic asymptotes
in the logarithmic layer. The mesh size is 64 65 64.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 383
Fig. 13. The root mean square velocity uctuations as a function of the distance from the wall for the FENE-P model with L = 10,
b = 0:9 and We
s
= 50, for j=(U
s
h) = 0:01 and j=(U
s
h) = 0:02, as well as for the Newtonian ow. The mesh size is 64 65 64.
Fig. 14. The trace of the conformation tensor as a function of the distance from the wall for the FENE-P model with L = 10, b = 0:9
and We
s
= 50, for j=(U
s
h) = 0:01 and j=(U
s
h) = 0:02. The mesh size is 64 65 64.
384 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
Fig. 16. The Reynolds shear stress for Newtonian and FENE-P (L = 10, b = 0:9, We
s
= 50, 64 65 64 mesh) with j=(U
s
h) = 0:01
and j=(U
s
h) = 0:02. The purely viscous and viscoelastic contributions to the total shear stress are also plotted along with the total
stress for both viscoelastic ows. The mesh size is 64 65 64.
Fig. 15. The root mean square uctuations of the xx component of the conformation tensor as a function of the distance from the wall
for the FENE-P model with L = 10, b = 0:9 and We
s
= 50, for j=(U
s
h) = 0:01 and j=(U
s
h) = 0:02. The mesh size is 64 65 64.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 385
anticipated to have as a side-eect to dampen the conformation changes and thus the extensional viscosity
and the resulting drag reduction. However, it does appear from the comparisons shown that the eect of
numerical diusivity for the values employed in this work is small. This is encouraging because it increases
our condence that the articial diusivity, while necessary in order to ensure numerical stability, does not
lower signicantly the accuracy of the results.
5.2.3. Eect of the dimensions of the periodic computational domain
One of the key predictions made in the earlier numerical simulations of the viscoelastic ow was a
signicant increase in the thickness and spacing of the streamwise streaky structures, observed in snapshots
of the instantaneous streamwise velocity uctuations at an xz plane surface situated at about the location
of the boundary between the viscous and inertial sublayers, y

= 15 [1,2]. These changes have been also


observed experimentally under conditions where drag reduction takes place [22]. Those changes have been
correlated to corresponding changes in the streamwise vorticity uctuations, thus providing further evi-
dence on the proposed mechanism for drag reduction through an increase in size of the turbulence feeding
wall eddies [19,33]. Moreover, the degree of drag reduction has been experimentally correlated to the in-
crease of the streak spacing [22], which was found to be in good agreement with the results of the numerical
simulations [1,2].
However, as the range of the velocity correlations increases, exemplied with the increase of the thickness
and spacing of the streaky structures, there is a concern for the adequacy of the size of the periodic com-
putational cell, which in all the previous calculations was 10h 2h 5h along the x, y and z directions,
respectively. Indeed, a typical case is shown in Fig. 17, for the FENE-P, L = 30, b = 0:9 and We
s
= 50,
calculated with an enhanced mesh of 80 65 80. There, one can see that only four streaks can t along the
spanwise, z, direction within the computational domain while, at the same time, strong correlations appear in
the streamwise, x, direction as well. Thus, although a domain of that size seems to be adequate for the
simulations of turbulent Newtonian ow [50], given the signicant increase of the correlation length along
both the periodic directions observed with strongly drag reducing turbulent viscoelastic ows, there is a
concern that the short periodicity imposed by the small periodic computational domain may introduce
systematic errors in the calculations. To test this hypothesis, a series of runs was performed using the same
FENE-P model, L = 30, b = 0:9 and We
s
= 50, but a 50% larger computational domain in each one of the
periodic directions, i.e., of a size 15h 2h 7:5h along the x, y and z directions, respectively. To remove the
possibility that the effects seen are due to differences in resolution, we have also increased the mesh resolution
proportionally, i.e., we have used in these calculations a mesh size of 120 65 120.
The increased size of the computational domain had a clear eect in the spacing and form of the streaky
structures, which now appear of a much more variable form and of increased spacing, as shown in Fig. 18.
Note that in order to facilitate comparison Fig. 18 is presented at the same scale as Fig. 17. Comparing the
two gures one can clearly see large qualitative changes. The question then arises whether those changes
seen in an arbitrary snapshot also translate to quantitative changes in the average and RMS values for
various ow statistics. A comparison of the results obtained by two computational domains with dierent
size is oered in Figs. 1923.
As Fig. 19 shows, the ow simulated in the larger computational domain shows signicantly higher
drag reduction. This is consistent with the fact that the streaky structures corresponding to the larger
computational domain appear to have a higher spacing, thus corresponding to larger eddy structures.
These structures are indicative of higher drag reduction, since they are correlated with the mechanism
according to which the wall-eddies are responsible for feeding, and thus sustaining, the turbulence with energy
[19,33]. On the other hand, it is important to note that this increased drag reducing eect of the larger com-
putational domain is very dierent from a ``true'' enhanced drag eect through, for example, modeling.
Whereas, as we increase drag reduction through, for example, increasing the Weissenberg number, we have
also seen an increase of the intensity of the streamwise velocity uctuations and a decrease of the spanwise and
shearwise ones, as well as a decrease of the vorticity uctuations and an increase in the uctuations of the
molecular deformation, here we either do not see these eects or we even see the opposite, as Figs. 2023
indicate. We believe that this is because what we see here is the relief of an articially imposed tight periodicity
constraint which was not letting the physical, model-induced structures to develop freely. Thus, we feel that
some of the previous results, specically those for the FENE-P model with L = 30 and b = 0:9 that were
386 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
Fig. 18. Instantaneous snapshots of the streamwise uctuating velocity in the xz plane at y

~ 15 for the FENE-P model (L = 30,


We
s
= 50, b = 0:9 and j=(U
s
h) = 0:01). The channel size is 15h 2h 7:5h, whereas the mesh size is 120 65 120. The dark regions
represent the low speed uid, whereas the light regions the high speed uid. The effect of drag reduction is a little more intense than in
Fig. 17.
Fig. 17. Instantaneous snapshots of the streamwise uctuating velocity in the xz plane at y

~ 15 for the FENE-P model (L = 30,


We
s
= 50, b = 0:9 and j=(U
s
h) = 0:01). The channel size is 10h 2h 5h, whereas the mesh size is 80 65 80. The dark regions
represent the low speed uid, whereas the light regions the high speed uid.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 387
Fig. 19. The mean ow velocity as a function of the distance from the wall for the FENE-P model (L = 30, b = 0:9, We
s
= 50,
j=(U
s
h) = 0:01) as well as for the Newtonian ow. One viscoelastic ow case corresponds to a channel with dimensions
15h 2h 7:5h and 120 65 120 mesh, whereas the other case and the Newtonian ow correspond to 10h 2h 5h and
80 65 80 mesh. Also shown are the respective semi-logarithmic asymptotes in the logarithmic layer.
Fig. 20. The root mean square velocity uctuations as a function of the distance from the wall for the FENE-P model (L = 30, b = 0:9,
We
s
= 50, j=(U
s
h) = 0:01) as well as for the Newtonian ow. One viscoelastic ow case corresponds to a channel with dimensions
15h 2h 7:5h and 120 65 120 mesh, whereas the other case and the Newtonian ow correspond to 10h 2h 5h and
80 65 80 mesh.
388 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
Fig. 21. The root mean square vorticity uctuations as a function of the distance from the wall for the FENE-P model (L = 30,
b = 0:9, We
s
= 50, j=(U
s
h) = 0:01) as well as for the Newtonian ow. One viscoelastic ow case corresponds to a channel with di-
mensions 15h 2h 7:5h and 120 65 120 mesh, whereas the other case and the Newtonian ow correspond to 10h 2h 5h and
80 65 80 mesh.
Fig. 22. The trace of the conformation tensor as a function of the distance from the wall for the FENE-P model (L = 30, b = 0:9,
We
s
= 50, j=(U
s
h) = 0:01). One viscoelastic ow case corresponds to a channel with dimensions 15h 2h 7:5h and 120 65 120
mesh, whereas the other case and the Newtonian ow correspond to 10h 2h 5h and 80 65 80 mesh.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 389
obtained with the smaller computational domain, contain small systematic errors that fortunately do not alter
the conclusions that had been reached [2]. The observed dierences as the computational domain size is in-
creasedare solely due to the ``restructuring'' of the turbulent oweld in the larger domain. However, as it was
mentioned before, the errors are in general small and the relative trends when one increases, for example, the
extensional viscosity are preservedandare the same as before, whenone examines a domain of a givensize. It is
expected that after a critical size, the periodicity conditions will not aect the turbulent ow eld signicantly
and one will not observe dierences. But, as more drag reducing ows are simulated, we need to evaluate
carefully the appropriate domain size for the calculations, in order not to bias the simulation results. Addi-
tional work is currently under way in order to further test yet larger computational domain and mesh sizes.
6. Conclusions
We have presented an algorithm for direct numerical simulation of turbulent channel ow of a visco-
elastic dilute polymer solution. The algorithm is implemented on a parallel supercomputer and exhibits
almost linear scalability with the number of unknowns and linear scalability with the number of processors
over a wide range. The viscoelastic contribution to the stress was obtained using two dierent and inde-
pendently evaluated rheological models with no adjustable parameters, the FENE-P and Giesekus models,
for the polymer chains. In a consistent fashion to previous work by Sureshkumar et al. [1] and Dim-
itropoulos et al. [2], we have observed a reduction in the turbulent drag for a relatively wide range of
parameters when the ratio of extensional to shear viscosity is suciently high and the Weissenberg number
permits a fast enough response of the polymer macromolecule. For example, the Reynolds stress and the
pressure statistics are consistent with what one would expect from the results for other quantities with
increasing drag reduction.
Mesh renement studies have shown small quantitative changes, with the drag reduction increasing
slightly with the mesh renement accompanied by a consistent increased drag-reduction behavior from
Fig. 23. The root mean square uctuations of the xx component of the conformation tensor as a function of the distance from the wall
for the FENE-P model (L = 30, b = 0:9, We
s
= 50, j=(U
s
h) = 0:01). One viscoelastic ow case corresponds to a channel with di-
mensions 15h 2h 7:5h and 120 65 120 mesh, whereas the other case and the Newtonian ow correspond to 10h 2h 5h and
80 65 80 mesh.
390 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392
various other quantities such as the mean ow velocity, the RMS velocity and vorticity uctuations, as well
as the trace of the conformation tensor. The RMS values of the xx component of the conformation tensor
reinforced our hypothesis of the critical role of polymer extension on the phenomenon of drag reduction.
The effect of the numerical diffusivity was also studied, where it is evident that even though there is a
damping effect as the diffusivity is increased, the results do not become signicantly different and they are
qualitatively similar to a slightly less drag-reducing ow. Thus, we are quite condent that the numerical
simulations do capture the essential physics of the phenomenon.
Finally, we also undertook a preliminary study of the eect of the size of the computational domain on
the results. For the old 10h 2h 5h domain, a mesh of 80 65 80 gave identical results with the
previous 64 65 64 mesh. However, when one increases both the domain size and the mesh size by a
factor of 1.5 (so that the mesh resolution remains the same), one can see that this has a signicant qual-
itative eect on the structure of the turbulence ow eld, leading to larger sizes for the streamwise streaky
structures and slightly larger drag reduction. In fact, the increase in the domain apparently removes a size
constraint that existed previously which appears to have also aected the results quantitatively. Therefore,
the importance of using a computational domain of an adequate size in order to eventually be able to
achieve quantitative comparisons is evident from this work. In particular, the computational domain needs
to be larger than the largest coherent structure generated within the turbulent ow eld, otherwise there is a
risk to have adverse eects on the accuracy of the solution from the periodic boundary conditions in the
streamwise and spanwise directions. This seems to be an important issue for viscoelastic turbulent ow
simulations, where in contrast to Newtonian simulations, the proper computational domain size depends
not only on the Reynolds number, but also on the rheological properties of the uid which is examined.
Acknowledgements
The authors would like to acknowledge the nancial support provided by ONR, Grant No. N00014-98-
1-0008 (CDD and ANB) and NSF, Grant No. CTS-9114508 (ANB). We are grateful to the National
Center for Supercomputing Applications (NCSA) for providing the computational resources used in this
research. We will also like to acknowledge Mark A. Straka for his help in porting our code to the SGI/
CRAY Origin 2000 system at NCSA and Dr. Danesh Tafti for his support during our transition to NCSA.
References
[1] R. Sureshkumar, A.N. Beris, R.A. Handler, Direct numerical simulation of turbulent channel ow of a polymer solution, Phys.
Fluids 9 (1997) 743755.
[2] C.D. Dimitropoulos, R. Sureshkumar, A.N. Beris, Direct numerical simulation of viscoelastic turbulent channel ow exhibiting
drag reduction: eect of the variation of rheological parameters, J. Non-Newtonian Fluid Mech. 79 (1998) 433468.
[3] S.A. Orszag, L.C. Kells, Transition to turbulence in plane Poiseuille and plane Couette ow, J. Fluid Mech. 96 (1980) 159205.
[4] P. Moin, J. Kim, On the numerical solution of time-dependent viscous incompressible uid ows involving solid boundaries,
J. Comput. Phys. 35 (1980) 381392.
[5] J. Kim, P. Moin, Application of a fractional-step method to incompressible NavierStokes equations, J. Comput. Phys. 59 (1985)
308323.
[6] W.D. McComb, The Physics of Fluid Turbulence, Oxford University Press, Oxford, 1990.
[7] U. Frisch, Turbulence, Cambridge University Press, Cambridge, 1995.
[8] T.B. Gatski, M.Y. Hussaini, J.L. Lumley, Simulation and Modeling of Turbulent Flows, Oxford University Press, New York,
1996.
[9] J. Kim, P. Moin, R. Moser, Turbulence statistics in fully-developed channel ow at low Reynolds number, J. Fluid Mech. 177
(1987) 133166.
[10] J.G.M. Eggels, F. Unger, M.H. Weiss, J. Westerweel, R.J. Adrian, R. Friedrich, F.T.M. Nieuwstadt, Fully-developed turbulent
pipe ow. A comparison between direct numerical simulation and experiment, J. Fluid Mech. 268 (1994) 175209.
[11] S.M. de Bruyn Kops, J.J. Riley, Direct numerical simulation of laboratory experiments in isotropic turbulence, Phys. Fluids 10
(1998) 21252127.
[12] P. Moin, K. Mahesh, Direct numerical simulation: A tool in turbulence research, Ann. Rev. Fluid Mech. 30 (1998) 539578.
[13] M. Nelkin, S. Chen, The scaling of pressure in isotropic turbulence, Phys. Fluids 10 (1998) 21192121.
[14] R.A. Handler, T.F. Swean, R.I. Leighton, J.D. Swearingen, Length scales and the energy balance for turbulence near a free
surface, AIAA J. 31 (1993) 19982007.
A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392 391
[15] D.T. Walker, R.I. Leighton, L.O. GarzaRios, Shear-free turbulence near a at free surface, J. Fluid Mech. 320 (1996) 1951.
[16] D.C. Chu, G.Em. Karniadakis, A direct numerical simulation of laminar and turbulent ow over riblet-mounted surfaces, J. Fluid
Mech. 250 (1993) 142.
[17] H. Choi, P. Moin, J. Kim, Direct numerical simulation of turbulent ow over riblets, J. Fluid Mech. 255 (1993) 503539.
[18] D. Goldstein, R. Handler, L. Sirovich, Direct numerical simulation of turbulent ow over a modeled riblet covered surface,
J. Fluid Mech. 302 (1995) 333376.
[19] J.L. Lumley, Drag reduction by additives, Ann. Rev. Fluid Mech. 1 (1969) 367384.
[20] T.S. Luchik, W.G. Tiederman, Turbulent structure in low concentration drag-reducing channel ows, J. Fluid Mech. 190 (1988)
241263.
[21] W.W. Wilmarth, T. Wei, C.O. Lee, Laser anemometer measurements of Reynolds stress in a turbulent channel ow with drag
reducing polymer additives, Phys. Fluids 30 (1987) 933935.
[22] W.G. Tiederman, The Eect of dilute polymer solutions on viscous drag and turbulence structure, in A. Gyr (Ed.), Structure of
Turbulence and Drag Reduction, Springer-Verlag, 1990 187200.
[23] R. Sureshkumar, A.N. Beris, Eect of articial stress diusivity on the stability of numerical calculations and the ow dynamics of
time-dependent viscoelastic ows, J. Non-Newtonian Fluid Mech. 60 (1995) 5380.
[24] J.M. Marchal, M.J. Crochet, A new mixed nite element method for calculating viscoelastic ow, J. Non-Newtonian Fluid Mech.
26 (1987) 77114.
[25] M.J. Crochet, V. Legat, The consistent streamline-upwind/PetrovGalerkin method for viscoelastic ow revisited, J. Non-
Newtonian Fluid Mech. 42 (1992) 283299.
[26] F. Debae, V. Legat, M.J. Crochet, Practical evaluation of four mixed nite element methods for viscoelastic ow, J. Rheol. 38
(1994) 421441.
[27] J. Rosenberg, R. Keunings, Numerical integration of dierential viscoelastic models, J. Non-Newtonian Fluid Mech. 39 (1991)
269290.
[28] M.T. Landahl, E. MolloChristensen, Turbulence and random processes in uid mechanics, 2nd ed, Cambridge University Press,
New York, 1992.
[29] B.A. Toms, Some observations on the ow of linear polymer solutions through straight tubes at large Reynolds numbers.
Proceedings of the International Congress on Rheology, Vol. 2 (North Holland, Amsterdam, 1949) 135-141.
[30] J.L. Lumley, Drag reduction in turbulent ow by polymer additives, J. Polymer Sci. Macromol. Rev. 7 (1973) 263290.
[31] P.S. Virk, Drag reduction fundamentals, AIChE J. 21 (1975) 625656.
[32] N.S. Berman, Drag reduction by polymers, Ann. Rev. Fluid Mech. 10 (1978) 4764.
[33] A.B. Metzner, M.G. Park, Turbulent ow characteristics of viscoelastic uids, J. Fluid Mech. 20 (1964) 291303.
[34] R.B. Bird, C.F. Curtiss, O. Armstrong, O. Hassager, Dynamics of Polymeric Fluids, vol. 2, 2nd ed., John Wiley & Sons, New
York, 1987.
[35] A.N. Beris, B.J. Edwards, Thermodynamics of Flowing Systems with Internal Microstructure, Oxford University Press, New
York, 1994.
[36] R. Sureshkumar, A.N. Beris, Uniformly valid approximations for the conformational integrals resulting from Gaussian closure in
the Hookean dumbbell model with internal viscosity, J. Rheol. 39 (1995) 13611384.
[37] D.D. Joseph, Fluid Dynamics of Viscoelastic Liquids, Springer-Verlag, New York, 1990.
[38] A.N. Beris, R. Sureshkumar, Simulation of time-dependent viscoelastic channel ow at high Reynolds numbers, Chemical
Engineering Science 51 (1996) 14511471.
[39] R.I. Tanner, Stresses in dilute solutions of bead-nonlinear-spring macromolecules. II. Unsteady ows and approximate
constitutive relations, Trans. Soc. Rheol. 19 (1975) 3765.
[40] L.E. Wedgewood, R.B. Bird, From molecular models to the solution of ow problems, Ind. Eng. Chem. Res. 23 (1988) 13131320.
[41] H. Giesekus, A simple constitutive equation for polymer uids based on the concept of deformation-dependent tensorial mobility,
J. Non-Newtonian Fluid Mech. 11 (1982) 69109.
[42] R.B. Bird, R.C. Armstrong, O. Hassager, Dynamics of Polymeric Fluids, vol. 1, 2nd ed., John Wiley & Sons, New York, 1987.
[43] R.F. Ginn, M.M. Denn, Rotational stability in viscoelastic liquids, AIChE J. 15 (1969) 450454.
[44] H. Giesekus, P.K. Bhatangar, On the stability of viscoelastic uid ow. IV. Overstability in plane Couette ow, Rheol. Acta 10
266-274.
[45] A.N. Beris, M. Avgousti, A. Souvaliotis, Spectral calculations of viscoelastic ows: evaluation of the Giesekus constitutive
equation in model ow problems, J. Non-Newtonian Fluid Mech. 44 (1992) 197228.
[46] M. Renardy, On the mechanism of drag reduction, J. Non-Newtonian Fluid Mech. 59 (1995) 93101.
[47] A. Oztekin, R.A. Brown, G.H. McKinley, Quantitative prediction of the viscoelastic cone-and-plate ow of a Boger uid using a
multimode Giesekus model, J. Non-Newtonian Fluid Mech. 54 (1994) 351377.
[48] A. Peterlin, Streaming birefringence of soft linear macromolecules with nite chain length, Polymer 2 (1961) 257264.
[49] T.N. Phillips, I.M. Soliman, Inuence matrix technique for the numerical spectral simulation of viscous incompressible ows,
Num. Meth. PDE 7 (1991) 924.
[50] R.A. Handler, E. Levich, L. Sirovich, Drag reduction in turbulent channel ow by phase randomization, Phys. Fluids A 5 (1993)
686694.
[51] R. Dean, Reynolds number dependence of skin friction and other bulk ow quantities in two-dimensional rectangular duct ow,
Trans. ASME I: J. Fluids Eng. 100 (1978) 215223.
[52] J. Kim, On the structure of pressure uctuations in simulated turbulent channel ow, J. Fluid Mech. 205 (1989) 421451.
392 A.N. Beris, C.D. Dimitropoulos / Comput. Methods Appl. Mech. Engrg. 180 (1999) 365392

S-ar putea să vă placă și