International Journal of Heat and Fluid Flow 54 (2015) 196–210

Contents lists available at ScienceDirect

International Journal of Heat and Fluid Flow

journal homepage: www.elsevier.com/locate/ijhff

Large-Eddy simulation of turbulent pipe flow of power-law fluids

P.S. Gnambode a, P. Orlandi a,b, M. Ould-Rouiss a,⇑, X. Nicolas a
a
Université Paris-Est, Laboratoire Modélisation et Simulation Multi Echelle, MSME, UMR 8208 CNRS, 5 bd Descartes, 77454 Marne-la-Vallée, France
b
Dipartimento di Ingegneria Meccanica e Aerospaziale, Università La Sapienza, Rome, Italy

a r t i c l e i n f o a b s t r a c t

Article history: Fully developed turbulent flows of power-law fluids in a cylindrical stationary pipe are investigated
Received 14 April 2014 numerically by the use of large eddy simulation (LES) for various power law index (0:5 6 n 6 1:4) at dif-
Received in revised form 7 April 2015 ferent Reynolds numbers (4000 6 Res 6 12,000). To validate the present computations, the predictions
Accepted 4 May 2015
are compared to the results reported in the archival literature for laminar and turbulent flows. The LES
predictions agree reasonably favourably with the findings of the literature. The log-region of the mean
axial velocity profile expands with increasing Res and decreasing power-law index n. The predicted fric-
Keywords:
tion factor for n 6 1 at Res ¼ 4000 is slightly overestimated in comparison with Dodge and Metzner cor-
Large eddy simulation (LES)
Power-law fluid
relation, and is better interpolated by Gomes correlation. With increasing n the apparent viscosity
Pipe flow increases close to the wall and decreases for yþ > 30. This implies that the turbulent fluctuations develop
and are more intense further from the wall when n > 1 and closer to the wall when n < 1. The influence
of Res and n on the higher-order statistics (skewness and flatness) is analyzed. Visualizations of the
instantaneous filtered velocity fields exhibit turbulent patterns which develop more as n increases.
Ó 2015 Elsevier Inc. All rights reserved.

1. Introduction turbulent friction and the mean velocity profiles at various gener-
alized Reynolds numbers and different values of the power-law
The turbulent flows of non-Newtonian fluids are of importance index n.
in mechanical and engineering fields. They are encountered in a A more conventional and general k– model was developed by
variety of engineering applications, e.g. drilling hydraulics, sewage Kyoungchul and HongSun (2012) to analyze non-Newtonian fluid
transport, processing of mineral oil and polymer products, blood flows for more complex and various engineering problems. The
flow in arteries, and applications involving relatively high heat modified k– model is based on the standard one with wall and
transfer rates. While the turbulence theory, the mathematical damping functions including the drag reduction phenomenon. In
models and the numerical methods are well-advanced for order to validate their modified k– model, numerical simulations
Newtonian fluids, those for non-Newtonian fluids are not as devel- are performed for shear-thinning fluids, at different values of the
oped. Some attempts have been made to explore the effect of the flow index 0:4 6 n 6 1. The predicted friction factors and mean
power-law index and Reynolds number on the velocity distribution axial velocity profiles agree well with the experimental results of
and turbulence statistics. Computational models for literature (Dodge and Metzner, 1959; Escudier and Presti, 1996;
non-Newtonian fluids can help to bridge the gap in the existing lit- Ptasinski et al., 2001), and this agreement is much better than with
erature, and can contribute to developing the general theories of the standard k– model. The proposed model also agree well with
the turbulent flows of non-Newtonian fluids. Malin’s power law model in the simulation of blood flow. The
Malin (1997) used a modified k– model (a low Reynolds num- authors indicate that the computational time and computer
ber k– model extended to power law fluids) to calculate the fric- resources of the modified k– model are reduced by about one
tional resistance and the velocity profile for fully developed third of those required by the low Reynolds number k– models
laminar and turbulent flows in smooth-walled tubes. A modifica- for power-law fluid (including Malin’s model). The authors point
tion of the viscous damping that improves the predictions for out that their turbulence model better predicts the behaviour of
non-Newtonian fluids is proposed. The presented k– predictions high power-law fluids.
are in fairly good agreement with experimental data for the Direct numeral simulation of turbulent pipe flows of
shear-thinning fluids was carried out by Rudman et al. (2004) for
n = 0.5, 0.69 and 0.75, using a spectral element-Fourier method,
⇑ Corresponding author. Tel.: +33 01 60 95 73 09; fax: +33 01 60 95 72 94.
at a moderate Metzner–Reed Reynolds number (ReMR ’ 3000 and
E-mail address: [email protected] (M. Ould-Rouiss).

http://dx.doi.org/10.1016/j.ijheatfluidflow.2015.05.004
0142-727X/Ó 2015 Elsevier Inc. All rights reserved.
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 197

Nomenclature

Cd coefficient of the dynamic model vr ; vz; vh dimensionless radial, axial and azimuthal velocity com-
D pipe diameter (m) ponents
f mean friction factor, f ¼ 2sw =ðqU 2b Þ yþ distance from the wall in wall units, yþ ¼ ð1  rÞU s =m
f DM Dodge and Metzner (1959) correlation for friction factor z dimensionless coordinate in the axial direction scaled
fG Gomes (1987) correlation for friction factor by the pipe radius
hv 04 i
Fðv 0i Þ flatness factor, Fðv 0i Þ ¼ 02i 2
hv i i
Greek symbols
K consistency index (Pa sn ) gw mean apparent viscosity at the wall (m2 s1 )
Lz length of the computational domain (m) g apparent viscosity, g ¼ K c_ n1
n power law index gd;w dimensionless apparent viscosity at the wall,
qi generic notation for qr ; qh and qz gw =ðqU cL RÞ  1=2
qr ; qh ; qz variables qr ¼ rv r , qh ¼ r v h ; qz ¼ v z c_ shear rate, c_ ¼ 2Sij Sij
Reb Reynolds number based on bulk velocity, Reb ¼ Umb D c_ d;w dimensionless shear rate at the wall, c_ d;w ¼ cw R=U cL
Recr critical Reynolds number between laminar and turbu- h dimensionless angular coordinate in the circumferential
lent flows, Recr ¼ 2100 ð4nþ2Þð5nþ3Þ 2 direction
3ð3nþ1Þ
ReMR Metzner–Reed Reynolds number ReMR ¼ Kð6þ2=nÞ
b 8qU 2n Dn mt turbulent viscosity
n 3
q density (kg=m )
qU 2n
cL R
n
ij ¼ 2mt Sij
Res simulation Reynolds number, Res ¼ K
sij subgrid stress tensor, s
Rew generalized Reynolds number, Rew ¼ qgU s D sw mean averaged fully-established wall shear stress,
w
sw ¼ D4 @p
@z
r dimensionless coordinate in the radial direction scaled
by the pipe radius
R pipe radius (m) Superscripts
Sij strain rate tensor hðÞi statistically averaged
Sðv 0i Þ
hv 03 i
skewness factor, Sðv 0i Þ ¼ i 3 ðÞþ normalized by us or gw
hv 02 i2 
ðÞ filtered variable
i
ui generic notation for the dimensionless velocity compo- ðÞ0 fluctuation component
nents v r ; v h and v z
Ub bulk velocity (m/s) Subscripts
U cL centreline axial velocity. For analytical laminar profile, c centreline
U cL ¼ ð3nþ1ÞU
nþ1
b
L laminar
Uþ mean axial velocity in wall units, U þ ¼ U=U s w wall
Us friction velocity, U s ¼ ðsw =qÞ1=2

4000). A similar DNS study at a higher Metzner–Reed Reynolds index and Reynolds number on the turbulent fields of
number (ReMR ¼ 7500) was conducted by Rudman and Blackburn non-Newtonian fluids.
(2012). In the log-region, the velocity profile was shown to agree There are very few studies employing LES for non-Newtonian flu-
well with the experimental data by Rudman et al. (2001, 2002). ids. To predict the turbulence features in non-Newtonian fluid flows,
The friction factors predicted by DNS were 10–15% higher than Ohta and Miyashita (2014) developed a turbulence model that can
those referred to in earlier research (Dodge and Metzner correla- reproduce the DNS results. They pointed out that constructing a
tions obtained from experiments). The authors reported that this new turbulence model extended for non-Newtonian fluids would
is most likely related to the imperfect fit of the experimental data obviously be difficult, since the new model would have to consider
concerning fluids with power-law rheology. It was shown that, for additional terms in the filtered Navier–Stokes equations (i.e. it can
a given Reynolds number, the flow deviates further from the hardly be expected to construct a turbulence model by introducing
Newtonian profile as the power-law index n decreases, and the turbulence corrections to the additional terms). Therefore, they con-
results suggest that the transition to turbulence is delayed. sidered a different approach: they proposed an extended
Moreover, the shear-thinning or thickening rheologies did not Smagorinsky model with a correction for the filter width of the
result in major changes to the nature of the flow at ReMR ¼ 7500. locally varying viscosity. Ohta and Miyashita (2014) performed
Direct numerical simulation (DNS) and large-eddy simulation DNS and LES of turbulent channel flow, in two non-Newtonian fluids
(LES) are techniques well suited for predicting turbulent with the viscosity described by both the power-law model (n = 0.85
non-Newtonian fluid flows, because a detailed picture of the turbu- and 1.15) and Casson’s model. By performing LES with the
lent structures, profiles of turbulence energy, rms and Reynolds Smagorinsky model as sgs model, extended according to the results
stresses are difficult to obtain experimentally. In DNS, numerically of the DNS, they evaluated the reliability of the extended sgs model.
accurate and complete resolution of all spatial and temporal flow They found that it could more accurately predict the velocity of tur-
scales is required and no turbulence model is used. In LES, an accu- bulent flows of fluids described by both Casson’s model and
rate numerical resolution of a wide range of scales is required and power-law model as compared to the standard Smagorinsky model
only the smallest scales are modelled using a subgrid scale (sgs) (i.e. the results of LES with the extended model agree more with
turbulence model. While DNS is clearly a very useful tool for accu- those obtained by DNS with high resolution). Consequently their
rately simulating the turbulent flows, LES however can yield quan- study showed that the Smagorinsky model of non-Newtonian tur-
titatively accurate predictions at a computational cost which is bulent flows could be universally treated via a spatial scaling of
significantly lower than the corresponding DNS one, since the the locally varying viscosity.
effect of the smallest scales in LES is modelled and the mesh is rel- Thais et al. (2010) proposed an LES approach for viscoelastic
atively coarse. Moreover, when the Reynolds number is significant, turbulent channel flows, based upon a temporal deconvolution
LES provides an effective tool for predicting the effect of the flow method (which was developed for LES of Newtonian channel
198 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

flows) for residual Newtonian stress modelling and secondary reg- filtered governing equations for incompressible non-Newtonian
ularization for unresolved subfilter Newtonian stress. A particular fluids, using the centreline axial velocity (U cL ) of the analytical fully
emphasis is put on the turbulent drag reduction. Excellent agree- developed laminar profile and the pipe radius R as velocity and
ment is obtained between temporal LES and DNS in terms of drag length scales respectively, can be expressed as follows:
reduction prediction. The first normal Reynolds stress component i
@u
is exactly in line (the agreement is even better than observed for ¼0 ð1Þ
@xi
Newtonian flows), while shear is slightly overestimated in the
near-wall region. The subfilter Reynolds stresses are marginally   
i @ u
@u  i uj 
@p 1 @ @u i @ uj @ sij
smaller than observed for a Newtonian flow at a moderate level þ ¼ þ c_ n1 þ  ð2Þ
@t @xj @xi Res @xj d @xj @xi @xj
of the drag reduction, while they are markedly smaller than
observed for a Newtonian flow at a high level of the drag reduction. The Reynolds number of the simulations is defined as
Molla and Paul (2012) carried out LES of a pulsatile qU 2n Rn
Res ¼ cLK . The apparent viscosity g of the fluid is modelled by
transition-to-turbulent non-Newtonian blood flow of arterial steno-
sis, within a 3D-channel, using a new consistent modelling approach a power-law (Ostwald de Waele model) given by g ¼ K c_ n1 , where
for the sgs stress. Their approach used the Cross model for the mod- K is the consistency, n is the power-law index, and c_ is the shear
elling of the molecular viscosity and the advanced dynamic rate. The flow index values 0 < n < 1; n ¼ 1 and n > 1 correspond
non-linear sgs stress model (DNM) of Wang and Bergstrom (2005) to shear thinning (or pseudoplastic), Newtonian, and shear thick-
for the modelling of the sgs stress tensor so that both the molecular ening (or dilatant) cases respectively. The dimensionless shear rate
is defined as: c_d ¼ UcR . The above dimensionless equations are writ-
_
and sgs shear stresses could be represented as non-linear functions cL

of the strain rate tensor. In their research, a high Womersley number ten in cylindrical coordinates and in terms of the variables
(a = 10.5) and a low Reynolds number (Re = 1200), both characteris- qr ¼ rv r ; qh ¼ rv h and qz ¼ v z in the code to avoid the singularity
tics of blood flows in large arteries of humans and animals, were at the axis pipe (r ¼ 0). A mean pressure gradient, at each time sub-
tested. The first and second order flow statistics of the velocity field step, is evaluated to balance the friction losses (the mean pressure
were analyzed. The streamwise velocity at different downstream gradient in the qz equation maintains a constant bulk velocity).
locations agreed with the experimental results of Ahmed and
Giddens (1983). The distributions of the other statistics (wall shear 2.2. Dynamic eddy viscosity model
stress, RMS of velocity fluctuations, Reynolds and sgs stresses) were
plotted but they were not compared with the results from the liter- The subgrid stress tensor sij is related to the of strain rate tensor
ature. The authors concluded that although their preliminary results Sij as
were encouraging, further future studies were necessary to prove
1=2
that their new modelling approach is an effective numerical tool sij ¼ 2mt Sij ¼ 2C d D2 jSjSij ¼ 2C d D2 ½2Sij Sij  Sij ð3Þ
for LES of non-Newtonian fluid flows.
The above literature survey indicates that the few DNS and LES where the turbulent viscosity mt is expressed using an eddy viscos-
devoted to turbulent pipe flows of power law fluids have been per- ity assumption, D is the computational filter, C d is the model con-
h i
@ u
stant, and Sij ¼ 1 @ui þ j .

formed for 0:4 6 n 6 1:4 at ReMR < 7500. The current study deals 2 @xj @xi
with LES of fully developed turbulent pipe flows of power-law flu- Following the model proposed by Germano et al. (1991), with
ids in a cylindrical pipe. The flow index n ranges from 0.5 to 1.4, modifications and extensions provided by Lilly (1992), the coeffi-
and the Reynolds number is up to 12,000. The aim of the current cient of the dynamic model C d is dynamically determined as:
research is to investigate the effects of the power law index, n,
and Reynolds numbers on turbulent power law pipe flows. The 1 hLij Mij i
Cd ¼  ð4Þ
LES at the highest Reynolds number (Res = 12,000) for different 2D2 hMij Mij i
flow index values are new as well as the computation of several
where the tensors Lij and Mij are given as follows
statistics such as:
g
Lij ¼ q ei q
j  q
iq ej ¼ 2C d D2 Mij ð5Þ
– higher-order statistics (skewness, flatness),
– cross-correlation coefficients and kinetic energy, f
D2 ff
– correlations for the peak position of the RMS of the velocity Mij ¼ jSj Sij  jSjg
Sij ð6Þ
fluctuations (with respect to n) and of the kinetic energy. D2
~ larger than the grid filter D which is defined as
Here a test filter D
This study is also intended to examine the effectiveness of the LES
D ¼ ðrDr DhDzÞ1=3 is introduced. The total volume-average box filter
approach for predicting the turbulent pipe flow of power-law fluids.
in cylindrical coordinates is used. The D~ filter width is twice in the
The paper is organized as follows: Section 2 provides the gov-
erning equations and the numerical procedure. The influence of axial and azimuthal directions in comparison to the grid filter D.
the power law index and Reynolds numbers on many turbulent The angle brackets, h i, denote spatial averaging procedure along
statistics (mean velocity profiles, root mean square of fluctuating the homogeneous directions of the flow to make the subgrid-scale
velocities, Reynolds shear stress, high order statistics, friction fac- coefficients well conditioned, and to reduce a large part of the spa-
tor) are discussed in Section 3. The present LES predictions are also tial variations. Negative values of the turbulent viscosity are elimi-
compared to the available results of literature for validation. The nated (i.e. set to zero).
main conclusions are found in Section 4. Note that in the present LES study, the standard dynamic model
of Germano et al. (1991) and Lilly (1992) is used to model the sgs
2. Governing equations and numerical procedure shear stresses. The sgs stress tensor is thus given as sij ¼ 2mt  Sij ,
where the sgs viscosity mt is a linear function of the shear rate
2.1. Governing equations j
Sj(cf. Eq. (3)). However, when the spatial filter is applied to the
Navier–Stokes equations, in non-Newtonian viscous fluid flows
The present study deals with the fully developed turbulent pipe (with spatially varying viscosity characteristics), additional terms
flows of power-law fluids in a cylindrical pipe. The dimensionless are derived (Ohta and Miyashita, 2014) because the stress tensor
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 199

is a non linear function of the strain rate tensor. This means that in streamwise and azimuthal directions. In Eq. (2), the shear rate c_ d
the current LES study with a standard dynamic model, the addi- is calculated explicitly.
tional terms are ignored. Indeed, these terms are smaller than The pipe length in the streamwise direction is equal to Lz ¼ 20R.
the sgs stress: Ohta and Miyashita (2014) plotted the profiles of In the radial direction, the grid distribution is stretched (with an
the sgs stress and of the additional terms that are estimated by fil- hyperbolic tangent function). A large computational time has been
tering the DNS results of Newtonian and power-law fluids. From spent checking the grid and Lz independence of the results, espe-
these profiles, they demonstrated that the additional terms are cially on the axial velocity profiles. Different grids have been tested
much smaller than the sgs stress. Therefore, in their study, they for n ¼ 0.75, 1 and 1.2 and predictions were compared to those of
concentrated on a modification of the sgs model for the viscosity the literature. The grid 65  65  65 was found to provide an accu-
characteristics of non-Newtonian fluid and ignored the additional rate prediction of turbulence statistics (in agreement with the
terms. Moreover, the grid resolution of the present LES approaches available data of the literature) and to give a good compromise
that of the true DNS (LES is performed with high resolution). The between the required CPU-time and accuracy.
coarse DNS and LES with sgs model, both performed at the same In wall bounded flows, the near-wall streaks in the buffer region
grid resolution (653 ), are in satisfactory agreement (see play an important role in the regeneration of turbulent energy. The
Section 3.2). It seems that the grid scale is finer than the dominant maximum production of turbulent kinetic energy occurs in the
scales of the flow so that it suffices to yield the right behaviour of inner layer. Consequently a finer mesh is required in the
the dominant scales (i.e. the dynamic sgs contribution is reduced near-wall region. It is well known that a wall-resolved LES needs
and the LES results approach those of DNS). This provides justifica- a near-wall grid resolution of yþ ’ 1 (which means that the first
tion for why it is reasonable to ignore the additional non linear gridpoint is located within the viscous sublayer) to capture the
terms in the filtered Navier–Stokes equations. dynamically dominant streak-like vortical structures in the viscous
and buffer regions. For present LES, the mesh with 65 nodes in the
radial direction has a grid resolution of yþ1 < 1 for the first grid cell

2.3. Numerical procedures adjacent to the wall (see Table 1 and the following discussion).
That means that the flow simulation is able to capture accurately
The governing equations were discretized on a staggered grid the major part of the eddies that contribute to the momentum
using cylindrical coordinates. The numerical integration was per- transport.
formed by a finite difference scheme, second-order accurate in The two-point correlations of the fluctuating streamwise veloc-
space and time. The time integration and the velocity–pressure ity along the axial and azimuthal directions provide information
coupling are performed by the fractional step method suggested about the flow structures and whether the pipe length is sufficient
by Rai and Moin (1991) and modified by Verzico and Orlandi to capture the largest eddies in the flow. In a previous work
(1996). In a first step, the momentum equations are solved by an (Redjem-Saad, 2008), we shew that the computational domain
incremental factorization method of ADI type to compute a non Lz ¼ 15R is large enough to simulate the largest structures for
solenoidal approximation of the velocity field. This intermediate n ¼ 1 (Newtonian fluid flow) at Re ’ 5000. The computational
velocity field is not locally divergence-free but it is globally domain in the periodic directions was not only large enough to
because periodic boundary conditions are applied in h and z direc- capture the largest eddies of the flow, but also sufficient for the
tions and the wall at r ¼ R is impermeable. In this step, the time scalar field (i.e. to simulate the largest thermal structures). In the
schemes are a third-order Runge–Kutta explicit scheme and a present LES, Lz ¼ 20R was fixed in all simulations. When the length
Crank–Nicolson implicit scheme for the convective and diffusive is increased to Lz ¼ 20R, a convergence to the statistical steady
terms, respectively. In the second step, the non-solenoidal velocity state is reached for n > 0:69. In these cases, the results are obtained
field is projected onto a solenoidal velocity field in order to correct in a reasonably long pipe, rather than spending a large computa-
the pressure and velocity fields and to satisfy the continuity equa- tional time checking the independence of the two-point correla-
tion on each cell. This projection step requires to solve a Poisson tions on Lz . For n 6 0:69, the pipe length is also considered large
equation for a scalar pressure increment /. This Poisson equation enough even though the statistics are less good.
is solved by Fast Fourier Transforms in the two periodic directions. Rudman et al. (2004) performed DNS runs on a domain lengths
This implies that the grids are necessarily uniform in the equal to 4pD for various n. Because their simulation for n ¼ 0:5

Table 1
Parameters of present LES simulations and mean flow quantities for Res ¼ 4000.

Parameters n = 0.5 n = 0.69 n = 0.75 n = 0.8 n = 1.0 n = 1.2 n = 1.4

þ 33.84 38.88 41.14 43.01 50.08 59.85 75.23
Dz
ðr DhÞþ 10.63 12.21 12.85 13.51 15.73 18.80 23.63
Drþ min
0.084 0.059 0.054 0.051 0.040 0.034 0.030
Drþ max 18.57 12.98 12.00 11.27 8.87 7.56 6.69
U b =U cl 0.542 0.545 0.533 0.524 0.494 0.502 0.544
U c =U cl 0.757 0.701 0.687 0.677 0.647 0.662 0.725
U s =U cl 0.027 0.032 0.033 0.034 0.036 0.039 0.045
hc_ d;w i 2.794 6.730 5.902 5.942 5.377 5.366 5.437
hgd;w i  104 1.49 1.38 1.60 1.75 2.50 3.50 4.92
Recr 2464 2280.3 2236.7 2204.2 2100 2024.6 1967.5
ReMR 5302 5033 4870 4754 4449 4512 4731
Rew 10211 8011 7042 6359 4449 3506 2920
Res ¼ U s R=mw 110 126.4 133.72 139.81 162.77 194.52 244.51
f DM  10þ3 5.82 7.32 7.79 8.18 9.58 10.53 11.00
f  103 6.35 8.37 8.85 9.24 10.46 12.54 15.69
f G  103 6.13 7.58 8.06 8.44 9.87 – –
yþ1
0.0342 0.0251 0.0231 0.0217 0.0171 0.0145 0.0129
200 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

Table 2 Indeed, Montreuil (2000) pointed out that three grid resolution’s
Parameters and mean flow quantities for Res ¼ 8000.
levels can be distinguished: LES with Dzþ ’ 35 and ðr DhÞþ < 10 is
Parameters n = 0.8 n=1 n = 1.2 referred to as LES with high resolution; LES with Dzþ ’ 70  80
Dz þ 65.56 78.97 99.37 and ðrDhÞþ ’ 15  20 is referred to as LES with moderate resolu-
ðr DhÞþ 21.17 24.81 31.22 tion; LES with Dzþ ’ 155 and ðr DhÞþ ’ 35  40 is referred to as
Drþ min
0.087 0.063 0.053 LES with low resolution. When employing schemes second-order
Drþ 19.13 14.01 11.65
max accurate in time and space, good results can be derived with
U b =U cL 0.525 0.495 0.529
U c =U cL 0.640 0.617 0.663 Dzþ < 100 and ðr DhÞþ < 12 (Zahrai et al., 1995). Moreover, Zang
U s =U cL 0.029 0.032 0.037 (1991) showed that an accurate LES can be performed using
hc_ d;w i 9.470 7.759 7.931 Dzþ < 80 and ðrDhÞþ < 40, with a minimum of three points within
hgd;w i  104 0.797 1.250 1.891
the viscous sublayer. In the present LES, there are many gridpoints
Recr 2204.2 2100. 2024.6 within the viscous sublayer (more than three points) confirming
ReMR 8485 7928 8368
Rew 12291 7928 6109
that accurate LES are performed in this study.
Res ¼ U s R=mw 219.1 256.7 323.0 To avoid the possible singularity at zero shear rate (g is infinite
f DM  10þ3 7.02 8.29 9.0 when c_ ¼ 0 and n < 1), a cut-off value c_ ¼ 106 is added in the
f  103 7.18 8.23 10.94 code. Below this value, the shear rate is supposed constant and
f G  103 7.15 8.36 – set equal to the cut-off value. However this cut off is never invoked
yþ1
0.0369 0.0270 0.0225 in the present LES runs because c_ is always above the cut off value.
The time is made dimensionless using the pipe radius, R, and
the maximum velocity of the laminar power law profile, U cL . The
was in transitional regime, they also performed a DNS for n ¼ 0:5 calculations have been conducted at the constant Courant,
on a domain length equal to 8pD. However, they observed that Friedrichs and Lewy condition: CFL = 1.7. In this case, the time step
the average flow results of both DNS are very similar and still tran- is computed from the imposed CFL, which is however limited by
sitional for both domain lengths. The main difference is in the axial the value 0:08R=U cL for shear-thickening and Newtonian fluids,
turbulence intensity that is slightly higher with the longest and by 0:04R=U cL or 0:01R=U cL for shear-thinning fluids, in order
domain. The friction factors of the two domains differ by less than to avoid large time discretization errors. The simulation is stopped
2%. It clearly appears that the structures do not fill the domains if the computed time step size becomes very small (which means
(isolated structures persist when the domain length of the simula- that the solution is diverging). This occurs when decreasing flow
tion is increased). The authors concluded that the results on the index n: it was necessary to reduce the time step DT to avoid very
extended domain confirm that this flow is transitional. large (diverging) velocity fields from the dimensionless time step
Table 1 lists the simulation parameters and mean flow quanti- DT = 0.08 (for n > 0:75) to DT = 0.04 (for n ¼ 0:75) and DT = 0.01
ties, at Res ¼ 4000. The first gridpoint, yþ
1 , is also given in Table 1, (for n < 0:75). Hence, in the present LES, the time step varies from
for each simulation. This table shows that due to the 0:01R=U cL to 0:08R=U cL .
non-uniform grid, the first gridpoint, yþ 1 , in all cases is located at The statistics are computed by averaging in the periodic direc-
yþ < 1. The next gridpoints are also within the viscous sublayer: tions and in time. The final data are obtained by ensemble averag-
from a minimum of 21 gridpoints at n ¼ 0:5 to a maximum of 29 ing over the time interval 6750 (i.e. from the dimensionless t ¼ 250
gridpoints at n ¼ 1:4. For Res ¼ 8000 and 12,000, the first gridpoint until t ¼ 7000) for the smallest flow index (n = 0.5 and 0.69). At
yþ þ
1 is also located at y < 1 and there are many gridpoints within higher flow index (n > 0:69), the integrated time over which the
the viscous sublayer: from a minimum of 18 gridpoints to a max- statistics are collected is 9750 (i.e. from the dimensionless time
imum of 25 (see Tables 2 and 3). Thus the simulation run for the t ¼ 250 until t = 10,000).
smallest Reynolds number and flow index (Res ¼ 4000 and
n ¼ 0:5) can be considered as LES with high resolution; for the
3. Results and discussion
other cases, the LES are of moderate resolution except at
Res = 12,000 for n ¼ 1:2 which is rather a low LES resolution.
3.1. Reynolds numbers

First of all, note that the Metzner–Reed Reynolds number, ReMR ,

Table 3 is computed and not imposed in each simulation (the imposed
Parameters and mean flow quantities for Res = 12,000.
Reynolds number is Res ). The relation between the two Reynolds
Parameters n = 0.8 n=1 n = 1.2 numbers is:
Dz þ 92.68 109.65 139.07  n  2n
ðr DhÞþ 29.11 34.44 43.69 Res 1 1 U cL
¼ 3þ ð7Þ
Drþ min
0.127 0.063 0.071 ReMR 8 n Ub
Drþ max 27.87 19.44 15.61
U b =U cl 0.526 0.496 0.530 If the flow is laminar, Eq. (7) reads:
U c =U cl 0.626 0.603 0.652  2  n
U s =U cl 0.027 0.029 0.034 Res 1 3n þ 1 nþ1
hc_ d;w i 12.278 9.774 9.609
¼ ð8Þ
ReMR 8 n þ 1 n
hgd;w i  104 0.504 0.833 1.310
Recr 2204.2 2100. 2024.6 As a consequence, for laminar flows, the ratio Res =ReMR mono-
ReMR 12745 11908 12579 tonically increases from Res =ReMR ¼ 0:601 for n ¼ 0:5 (the most
Rew 19556 11908 8788 shear thinning fluid flow in this study) to Res =ReMR ¼ 1:248 for
Res ¼ U s R=mw 301.2 356.3 452.0
n ¼ 1:4 (the most shear thickening fluid flow), passing through
f DM  10þ3 6.31 7.49 8.22
Res =ReMR ¼ 1 for n ¼ 1 (Newtonian fluid flow). In other words, at
f  103 6.03 7.05 9.52
fixed Res , the values of ReMR decrease when n increases.
f G  103 6.36 7.44 –
On the other hand, from Tables 1–3, one can note that for tur-
yþ1
0.0538 0.0375 0.0301
bulent flows at fixed Res , the values of ReMR decrease when n varies
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 201

from 0.5 to 1, but increase when n increases from 1 to 1.4. This is Shear thinning fluids (n < 1) are generally high polymer solu-
due to the variation of the dimensionless bulk velocity (U b =U cL ) tions or many suspensions. With increasing shear rates, the mole-
with n in Eq. (7): both U b =U cL and the centreline velocity (U c =U cL ) cules are progressively aligned and these fluids become less
vary in a similar way to ReMR in Tables 1–3. viscous with increasing shear rates. The micro-structure of such
Mishra and Tripathi (1971) derived a general criterion to char- materials is even smashed up at higher shear. This results in lower
acterize the transition between laminar and turbulent flows in viscosities, hence the fluid flows more easily. Naturally, smaller the
non-Newtonian fluid flows. This criterion is based on the critical value of n is, more shear thinning the material is.
Reynolds number: Validation of the present LES calculations of the velocity field is
also achieved by comparing the present results to the following
ð4n þ 2Þð5n þ 3Þ results of literature for fully developed turbulent pipe flows:
Recr ¼ 2100 ð9Þ
3ð3n þ 1Þ2
– the Laser Doppler Anemometry (LDA) and Particle Image
Using this criterion, one can calculates Recr for each n. Table 1
Velocimetry (PIV) measurements by Eggels et al. (1994), at
shows that all the simulations are performed for turbulent flows.
Reb ¼ U b D=m ¼ 5300 for n ¼ 1,
– experimental data by Pinho and Whitelaw (1990) for a concen-
3.2. Mean velocity profile tration of a polymer CMC (sodium CarboxyMethyl Cellulose)
with n ¼ 0:9 at Rew ¼ 5070,
The laminar axial velocity profiles, for various flow indexes n at – DNS predictions by Eggels et al. (1994) for n ¼ 1 at Reb ¼ 5300,
Res ¼ 500, are depicted in Fig. 1, along with the analytical – DNS results by Rudman et al. (2004) for n ¼ 0:75 at
distribution: ReMR ¼ 3935.
 r nþ1 
U z 3n þ 1 n
¼ 1 ð10Þ We first present the comparisons with the turbulent profiles
Ub nþ1 R
measured by Eggels et al. (1994) for n ¼ 1 at Reb ¼ 5300 and by
The predictions demonstrate a good agreement with the analyt- Pinho and Whitelaw (1990) for n ¼ 0:9 at Rew ¼ 5070 (see
ical equation. When n < 1, the velocity profiles flatten at the pipe Fig. 2a). The mean axial velocity distributions scaled by the friction
centre because the apparent viscosity g increases where the strain velocity (U þ ¼ U=U s with U s ¼ ðsw =qÞ1=2 ) are plotted on these
rate c_ decreases, and the fluid tends to behave as a solid. Opposite graphs as a function of the distance to the wall (based on the fric-
behaviours are observed for n > 1 or close to the wall where c_ tion velocity and the mean wall viscosity, yþ ¼ qU s ðr  RÞ=gw ). The
increases. Indeed, for dilatant fluids (n > 1), the apparent viscosity relation between the pressure gradient and the wall shear stress
increases with increasing shear rate, i.e. near the wall. A possible gives the friction velocity U s necessary to scale the axial velocity
explanation of the dilatant behaviour is as follows. This behaviour in wall units. The mean wall viscosity gw (gw ¼ K cn1 w ) necessary
is encountered in concentrated suspensions of solid particles. At to scale the distance from the wall (yþ ) is obtained from the simu-
rest, the voidage is at a minimum and the liquid present is only suf- lation results. The axial velocity distribution is slightly underesti-
ficient to fill the void space. At low shear rates, the liquid lubricates mated in the log-region in comparison with this experimental
the motion of each particle past the others thereby minimizing data. This discrepancy may be due to the difference in the
solid–solid friction. Consequently, the resulting stresses are small. Reynolds number values between the present LES and the experi-
At high shear rates, however, the mixture slightly expands (dilates) ments of literature. Indeed, Res ¼ 4000 corresponds to Reb ¼ 4449
so that there is no longer sufficient liquid to fill the increased void for n ¼ 1 (see Table 1), and it corresponds to Rew ¼ 5200 for
space and prevent direct solid–solid contacts and friction. This n ¼ 0:9.
leads to the development of much large shear stresses (which On the other hand, the predicted LES profile for a flow index
result in increased friction) than those seen in a pre-dilated sample n ¼ 0:75 at Res ¼ 4000 indicates a good agreement with the DNS
at low shear rates. This mechanism causes the apparent viscosity result by Rudman et al. (2004) (see Fig. 2b). The influence of the
to rise rapidly with increasing rate of shear (Chhabra, 2010).

Fig. 1. Laminar axial velocity profiles compared to analytical solution. Fig. 2a. Turbulent mean axial velocity profile compared to experiments.
202 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

Fig. 2b. Effect of n on turbulent axial velocity profiles. Fig. 2d. Effect of the sgs model and mesh on the axial velocity profiles, n ¼ 0:8.

Fig. 2c. Effect of Res and n on turbulent axial velocity profiles. Fig. 2e. Effect of the sgs model and mesh on the axial velocity profiles, n ¼ 1:2.

power-law index n on the velocity profile is also illustrated on this

graph. In the viscous sublayer (0 6 yþ 6 5), the normalized velocity The influence of the Reynolds number is shown in Fig. 2c. By fix-
profile agrees with the universal linear law U þ ¼ yþ , for all values ing the power-law index n and increasing Res both an enhance-
of n except for n ¼ 0:5 where the prediction is slightly overesti- ment of the axial velocity values is clearly seen in the log-region,
mated. This means that this flow is in the transitional regime and the log-region is more expanded. Moreover, for a given
and it confirms the observation by Rudman et al. (2004) for Reynolds number, the log-region is more expanded by decreasing
n ¼ 0:5. Indeed, for n = 0.5, the viscosity is higher in the core region n.
and hence the turbulence is not as fully developed here, even Finally, the influence of the grid resolution and the sgs model
though the Reynolds number (RMR ¼ 5302) is slightly higher than are analyzed through Figs. 2d and 2e. LES with and without the
that observed for the other flow index values (see Table 1). This sgs model are compared at the highest Reynolds number
may be due to the difference between the viscosity and Reynolds (Res = 12,000), for two flow indexes (n ¼ 0:8 and n ¼ 1:2), using
number evolutions: the viscosity increases more rapidly than the various meshes, from a very coarse grid (17  30  9) to a finer grid
Reynolds number for more shear-thinning fluid flows, damping (65  65  65). For the coarser grid, the axial velocities without the
the turbulence and inducing a less well-developed turbulent flow. sgs are much smaller than those with the sgs model, especially for
In the log-region (yþ > 30), the mean axial velocity profile for n ¼ 1 n ¼ 0:8, highlighting the improvement due to this model: the sgs
agrees with the log-law U þ ¼ 2:5lnðyþ Þ þ 5:5, and departs from the model provides more damping as expected. For the intermediate
Newtonian behaviour when n is different from unity. This devia- and finer grids, there is no difference between the results with
tion is more pronounced for n < 1 as the flow index n decreases. and without sgs model. It seems that the sgs model has no influ-
Similar trends have been observed by Rudman et al. (2004). ence on the solution and that the finest grid resolution (653 ) is
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 203

potentially good enough for a DNS. Indeed, the role of the sgs
model is reduced when the resolved viscous term (i.e. the gamma
term in Eq. (2)) is greater than the sgs term (i.e. the s
ij term in Eq.
(2)). If the near wall structures are more ordered (when, for
instance, the fluid is more viscous and less turbulent), the sgs
model should be less influential (or less important) because less
energy goes from the resolved to the unresolved scales. Note that
the ratio between the centreline velocities using LES with dynamic
model for the coarser and finer grids is approximately 44/23 for
n ¼ 0:8 and approximately 33/18 for n ¼ 1:2. These ratios denote
a large difference between the centreline velocities, especially for
n ¼ 0:8, due to the large difference between the corresponding fric-
tion velocities: the friction velocity obtained with the coarser grid
is almost twice that calculated with the finer grid. This may be due
to the choice of the grids: LES with resolution less than 653 should
never be performed for any kind of circular pipes (one can never
generates reasonable velocity oscillations in the azimuthal direc-
tion in a domain wide 2p with 32 points). In the simulations with
653 , the grid scale is fine enough to be able to simulate the domi-
nant scales of the flow and its right behaviour, as previously dis-
cussed. Similar conclusions have been reported for Newtonian Fig. 3a. Mean normalized viscosity versus mean normalized shear rate: effect of n.
fluid flows by Feiz et al. (2003).

3.3. Friction factor

The friction factor is defined by f ¼ 2sw =ðqU 2b Þ. For the fully

developed laminar flows simulated at Res ¼ 500 for various n val-
ues, the predicted friction factor value is f ’ 0:0341. This is within
a 2.3% accuracy when compared with the well-known correlation
f ¼ 16=ReMR . On the other hand, for the fully developed turbulent
flows, the friction factor decreases for decreasing n (at a given
Res ) and for increasing Res (at a given n), see Table 1. The present
LES predictions of f for shear-thinning and Newtonian fluids
(n 6 1), at Res ¼ 4000, are slightly overestimated in comparison
to the following Dodge and Metzner (1959) correlation:
a
f DM ¼ ð11Þ
RebMR

where a ¼ 0:0665 þ 0:01175n and b ¼ 0:365  0:177n þ 0:062n2 .

The discrepancies vary between 9% and 14.5% for the cases at
Res ¼ 4000 (see Table 1), between 2.2% and 20.6% at Res ¼ 8000
(see Table 2), and between 4.4% and 15.8% at Res = 12,000 (see
Table 3). A similar observation has been reported by Rudman
et al. (2004): their DNS predictions were 10–15% higher than the Fig. 3b. Behaviour of the mean normalized viscosity.
Dodge and Metzner correlation. For shear-thickening fluids, the dis-
crepancy is enhanced. Gomes (1987) proposed an explicit relation
for pseudo-plastic fluids (n 6 1Þ: The smallest values of the relative shear rate hc_ i=c_ w in Fig. 3a are
located in the pipe centre, while the highest ones are located near
f G ¼ 0:110n0:616 Re0:287 ð12Þ
the pipe wall. For the flow index n ¼ 1, the Newtonian behaviour is
Table 1 compares the LES predicted friction factors to the above find (constant viscosity equal to wall viscosity). At n < 1, this graph
mentioned correlation. This relationship seems to give better (Fig. 3a) clearly displays lower relative viscosity hgi=gw for high
results for Res ¼ 4000: the present friction factor is now 3.7– shear rates. For a given n (n < 1), the relative viscosity increases
10.4% higher than Gomes correlation for n 6 1. There are a number from the pipe wall towards the pipe centre because the fluid tends
of other friction factor correlations (empirical or semi-empirical) to behave like a solid rather than a liquid when approaching the
for pipe flows of power law fluids, in the literature. El-Emam core region of the pipe, due to the lower shear rate in this region.
et al. (2003) tested correlations against published data. They noted This trend is more pronounced as n decreases. For n > 1, the oppo-
that many of these correlations fitted the experimental data of site behaviour is observed. Shear thickening fluids become more
their authors but did not fit that of other authors, indicating a lack viscous with increasing shear rate. Shear thickening behaviour is
of generality. less common and generally arises in fluids that have a highly reg-
ular micro-structure at rest. When the fluid begins to move, the
3.4. Mean normalized viscosity micro-structural components jam against each other, thickening
the fluid thus preventing movement.
The variation of the mean viscosity scaled by the mean wall vis- In the core region, for flows with increasing shear-thinning
cosity, versus the shear rate, for different values of power law behaviour (n < 1), the deviation from the Newtonian behaviour is
index n, at Res ¼ 4000, can give further insights into the flow field. faster than that for flows with increasing shear thickening
204 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

behaviour. The ratio between the viscosity in the centre and the gd , is related to the dimensionless shear rate and the Reynolds
mean wall viscosity is about 1.8, 2.2, 2.8 and 5.1 for the flow index number by hgd i ¼ hc_ n1
d i=Res . Therefore, the Reynolds number dis-
n = 0.8, 0.75, 0.69, and 0.5 respectively, and approximately 0.6 and appears in the relative viscosity hgd i=gd;w ¼ hgi=gw ¼ hc_ n1 i=c_ n1
w
0.4 for n = 1.2 and 1.4 respectively, indicating higher relative vis- However, the plot of the mean relative viscosity hgi=gw versus
cosities (g=gw ) in the core region of the pipe for n < 1 than for the mean shear rate hc_ i clearly exhibits the effect of the Reynolds
n > 1. Note that for all values of n, the mean viscosity remains con- number on hgi=gw , Fig. 3d: the mean relative viscosity hgi=gw is
stant in the viscous sublayer up to yþ ’ 5, denoting a linear veloc- noticeably affected by the Reynolds number for n = 0.8; the influ-
ity distribution in this zone (not shown here). ence of Res (and thus ReMR ) on the viscosity is quite small for
Fig. 3b depicts the behaviour of the dimensionless apparent vis- n = 1.2; with increasing Res , the mean viscosity becomes almost
cosity hgi versus the wall distance yþ , for Res ¼ 4000: near the wall, independent of Res (the LES predictions at Res ¼ 8000 and
the apparent viscosity hgi is constant for a given n up to yþ ’ 5. It Res = 12,000 are very close to each other).
increases with increasing n for yþ < 5, while it decreases with
increasing n after the wall distance yþ ’ 35, for n P 0:69. Once 3.5. Root mean squares of turbulent fluctuations and Reynolds stress
again, the case n ¼ 0:5 deviates from this behaviour probably
because this flow is transitional. In Fig. 4a, the rms of axial, radial and azimuthal velocity fluctu-
When plotting the mean relative viscosity hgi=gw versus the ations and the corresponding experimental data of Eggels et al.
mean relative shear rate hc_ i=c_ w , Fig. 3c, all the distributions col- (1994) are in reasonable agreement for n ¼ 1. In Fig. 4b, for
lapse, as expected. Indeed, the dimensionless apparent viscosity, n ¼ 0:9, the LES predictions are underestimated in comparison to

Fig. 3c. Mean normalized viscosity versus normalized shear rate: effect of n for
Fig. 4a. RMS of velocity fluctuations compared to experiments for n ¼ 1.
different Res .

Fig. 3d. Mean normalized viscosity versus shear rate: effect of Res . Fig. 4b. RMS of velocity fluctuations compared to experiments for n ¼ 0:9.
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 205

the experimental data by Pinho and Whitelaw (1990), particularly

for the axial component near the wall.
The root mean square (rms) of fluctuating axial velocity is plot-
ted in Fig. 5a, at various values of power law index n, for
Res ¼ 4000. The predicted rms for n = 0.75 is in satisfactory agree-
ment with the DNS result of Rudman et al. (2004). The maximum
value of the axial velocity fluctuations increases and its position
moves far away from the wall with increasing n. It moves from
yþ ’ 13; 67 at n = 0.5 to yþ ’ 23; 79 at n = 1.4. The rms of axial
velocity fluctuations, Fig. 5a, is reduced in the vicinity of the wall
when n increases, and is enhanced after reaching the peak location.
This is due to the fact that the apparent viscosity g increases with
increasing n near the wall (in the viscous sublayer), but it decreases
when n increases in the turbulent layer (see Fig. 3c). Thus the tur-
bulent fluctuations develop and are more intense far from the wall
(after the peak) when n > 1 and closer to the wall when n < 1.
The radial and azimuthal velocity fluctuations are smaller than
the axial ones, and reach a maximum further away from the wall,

Fig. 5c. Effect of n on RMS of azimuthal velocity fluctuations.

Fig. 5a. Effect of n on RMS of axial velocity fluctuations.

Fig. 6a. Effect of Res on RMS of axial velocity fluctuations.

Fig. 5b. Effect of n on RMS of radial velocity fluctuations. Fig. 6b. Effect of Res on RMS of radial velocity fluctuations.
206 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

Fig. 6c. Effect of Res on RMS of azimuthal velocity fluctuations. Fig. 8. Cross-correlation coefficient.

Fig. 7a. Effect of n on Reynolds shear stress.

Fig. 9. Peak position of RMS velocity fluctuations and kinetic energy.

Fig. 7b. Effect of Res on Reynolds shear stress. Fig. 10a. Skewness of axial velocity fluctuations: effect of n.
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 207

Figs. 5b and 5c. Both radial and azimuthal turbulence intensities is expected since the mean viscosity decreases towards the pipe
diminish with decreasing n in the turbulent layer (as soon as centre with increasing n. When Res increases, the peak value of
yþ > 20), denoting an augmentation of the mean apparent viscos- kinetic energy is enhanced.
ity in this region. For n < 1, the increase of the mean viscosity Fig. 9 depicts the peak position (yþ ) of the rms of velocity com-
towards the pipe core generates larger and weaker axial vortices, ponents, as well as the peak position of the turbulent kinetic
hence reducing the wall drag. For shear-thickening fluid flows, energy, versus the flow index n, at Res ¼ 4000. All the peak posi-
the mean viscosity is smaller in the pipe core, leading to less dissi- tions shift towards the core region with increasing n. The graph
pation and shorter turbulent structures in this zone. confirms that the maximum of the axial velocity component is
Fig. 6a shows that for a given flow index n, when the Reynolds always closer to the wall pipe than those of the wall-normal and
number increases, the distribution of the fluctuating axial velocity azimuthal velocity components, irrespective of the flow index n.
rms shifts towards the wall, and a significant rise of its peak is The predicted values of the peak position of hu02 1=2
and that of
r i
observed. The influence of n on the peak value of axial velocity fluc- hu02 1=2
can be interpolated by yþ ¼ 82:8n0:29 and yþ ¼ 67:2n0:47
h i
tuations appears to be more pronounced at the lowest Reynolds 1=2
respectively. The peak position of hu02z i can be interpolated by
number (here Res ¼ 4000). At Res = 12,000, the flow index n seems
þ 0:72
to have no effect on the maximum value of axial turbulence inten- the correlation y ¼ 19:0n . Nearly the same n dependency is
sity, while its influence on the peak location is marked. observed for the peak position of kinetic energy: yþ ¼ 20:0n0:69 ,
Similarly, in Figs. 6b and 6c, the distribution of the rms of radial indicating the large contribution of the axial component on the tur-
and azimuthal velocity fluctuations moves towards the wall with bulent kinetic energy. Keeping in mind the viscous sublayer thick-
increasing Reynolds number, since the thickness of the viscous ness (yþ ’ 5) and that of the logarithmic region (yþ P 30), the
sublayer diminishes and hence the turbulence transports higher present predictions show that for 0:5 < n 6 1:4, the maximum of
momentum towards the wall. However, the maximum value of
the radial velocity fluctuations is reduced when Res increases,
whereas the peak of the azimuthal turbulence intensity profile dis-
appears, Fig. 6c, creating two new peaks at Res = 12,000.
The predicted Reynolds shear stress hu0r u0z i=U s agree reasonably
with the experimental data (n ¼ 1) and DNS results (n ¼ 0:75) of
literature, Fig. 7a. The Reynolds shear stress behaves like the axial
turbulence intensity when the flow index varies (see Fig. 5a). Its
distribution is shifted towards the wall with increasing n for a
given Res , and its peak value is noticeably enhanced. Similar beha-
viour is observed when the Reynolds number increases, for a given
n, Fig. 7b. The influence of Res on the Reynolds shear stress is much
more pronounced for yþ < 40 (i.e. before reaching the peak loca-
tion) than for yþ > 40.
Fig. 8 depicts the distribution of the cross-correlation coefficient
of the Reynolds shear stress, Ru0r u0z , with respect to yþ . This graph
exhibits a plateau with almost the same value for all n. The plateau
is more and more extended with decreasing flow index n. This
result means that the axial velocity fluctuations are correlated to
the radial ones along a more extended region when n decreases.
The effect of the power law index n on the turbulence kinetic
energy was also analyzed but is not shown here. The position of Fig. 11a. Flatness of axial velocity fluctuations: effect of n.
the maximum turbulent kinetic energy moves far further from
the wall as n increases, at a given Reynolds number. This behaviour

Fig. 10b. Skewness of axial velocity fluctuations: effect of Res. Fig. 11b. Flatness of radial velocity fluctuations: effect of n.
208 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

the rms of axial velocity as well as the maximum of kinetic energy stretches of small negative values and a few instances of large pos-
are reached in the buffer layer. The peak positions of hu02 1=2
and itive values, with zero time mean, has positive skewness.
r i
hu02 1=2
are mainly located in the log-region, for all values of n. The kurtosis (or flatness) is the fourth moment of v 0 normalized
h i
by the variance. the kurtosis can be used as an indication of the
tails of a PDF. A higher kurtosis indicates that relatively larger
excursions from the mean are more probable. A PDF with longer
3.6. Higher-order statistics tails will have a larger kurtosis than a PDF with more narrow tails.
A time series with most measurements clustered around the mean
A description of a turbulent variable at a given location and has low kurtosis; a time series dominated by intermittent extreme
time is given by the probability density function (PDF) or the events has high kurtosis. For the Gaussian behaviour, the flatness
moments of the variable derived from the PDF. The skewness is value is equal to F ¼ 3.
the third moment of the fluctuation v 0 ¼ v  hv i normalized by hu03 i
the variance. A PDF which is symmetric (a Gaussian PDF) about The profiles of the skewness coefficients Sðu0i Þ ¼ i
3 , at
hu02
i
i2
the mean hv i will have zero skewness. All higher odd moments Res ¼ 4000, have been computed and are presented in Fig. 10a
of such a symmetric PDF will also be identically zero. The skewness and 10b. As the power law index n decreases, the skewness coeffi-
reveals information about the asymmetry of the PDF. A non-zero cient of the three velocity components grows rapidly at the wall,
skewness indicates a skewed or asymmetric PDF, which in turn indicating a more important intermittent (non Gaussian) beha-
means that larger excursions in one direction are more probable viour with decreasing n, especially for the radial velocity fluctua-
than in the other. Positive skewness indicates that the PDF has a tions (not shown here). This indicates that large positive values
longer tail for ðv  hv iÞ > 0 than for ðv  hv iÞ < 0. Hence a positive of the velocity components rather than large negative values pre-
skewness means that variable v 0 is more likely to take on large pos- dominate near the wall. Far from the wall, irrespective of the flow
itive values than large negative values. A time series with long

Fig. 11c. Flatness of azimuthal velocity fluctuations: effect of n.

Fig. 13. Contours of resolved axial velocity field at yþ ’ 15: (a) n ¼ 0:5, (b) n ¼ 0:75,
Fig. 12. Flatness of radial velocity fluctuations, effect of Res . (c) n ¼ 1:0, and (d) n ¼ 1:2.
 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210 209

Fig. 14. Contours of resolved axial velocity field in r–h-plane: (a) n ¼ 0:5, (b) n ¼ 0:75, (c) n ¼ 1:0, and (d) n ¼ 1:2.

index n, the skewness of u0r quickly tends to the Gaussian value the wall are visualized in Fig. 13a–d for Res ¼ 4000. As n increases,
ðSðu0r Þ ¼ 0Þ towards the pipe centre, meaning that positive values the turbulence structures are more random, the axial correlation
of the variable are as probable than negative values. On the con- distance (stretching) is shorter and the streaks are shorter and
trary, the skewnesses Sðu0z Þ, Fig. 10a, and Sðu0h Þ (not shown here) stronger. This is in accordance with increasing turbulence intensi-
go to zero towards the core region more rapidly with increasing ties, from Fig. 13a to d, meaning that the turbulence is more devel-
n. They tend to zero further from the wall with an augmentation oped. For the lowest flow index (n ¼ 0:5), Fig. 13a, the turbulence
of Res , for each given n, Fig. 10b. activity is poor. This flow seems not fully developed, because of
hu04 i the higher viscosity in the core region. However, the mean flow
Profiles of the flatness coefficients Fðu0i Þ ¼ hu02i i2 are sketched in
i statistics (of first and second order) are still accurately predicted.
Figs. 11a–11c, at Res ¼ 4000. Close to the wall, the flatness factors Contours of the resolved axial velocity in r  h plane at
Fðu0z Þ and Fðu0r Þ are significantly dependent on the flow index n. The Res ¼ 4000, Fig. 14a–d, show that the turbulent structures are
rapid increase in the flatness factor near the wall reflects the high essentially located close to the pipe wall for n > 0:5. Further from
intermittency of the velocity fluctuations close to the wall. Such the wall (for yþ > 15), the turbulence activity is attenuated in com-
high values are attributed to the strong sweep events near by the parison to that at yþ ’ 15, especially for n < 1. Indeed, in the core
wall (Xu et al., 1996). The intermittent behaviour is more impor- region, the viscosity is increased for n < 1 (since this zone is a
tant as n decreases, and more pronounced for Fðu0r Þ than for Fðu0z Þ lower shear region) inducing a damping of the turbulent structures
and Fðu0h Þ, Figs. 11a–11c. The more intermittent behaviour close and hence a reduction in radial momentum transfer: the turbulent
to the wall means that the probability of observing large variations structures are weaker and bring less high speed fluid from the core
from the mean axial and radial velocity components close to the region to the wall region. At n ¼ 0:5, when the Reynolds number is
wall is much higher than in the centre of the pipe, especially when higher (Res ¼ 8000, not shown here), the wall streaks are less long
n decreases. Far from the wall, the factor Fðu0r Þ is equal to the than those observed at the same position for Res ¼ 4000, denoting
Gaussian value (F ¼ 3), irrespective of n, whereas Fðu0z Þ and Fðu0h Þ a more developed turbulence.
tend to the Gaussian behaviour for only n P 1. The flatness profiles
extends towards the core region with decreasing Res , for a given n, 4. Conclusion
Fig. 12.
Detailed characteristics of fully developed turbulent pipe flows
3.7. Flow pictures of power law fluids have been numerically explored, using large
eddy simulation with dynamic subgrid-scale model, to investigate
To explore the effects of the flow index on the near-wall struc- the power law index and Reynolds number effects on the velocity
tures, contours of the axial velocity at the distance yþ ’ 15 from field. The Reynolds numbers (Res ¼ 4000, 8000 and 12,000) and
210 P.S. Gnambode et al. / International Journal of Heat and Fluid Flow 54 (2015) 196–210

power law index (0.5 6 n 61.4) have been considered. The main References
new results (in comparison to those in the literature) obtained in
this study are: Ahmed, S.A., Giddens, D.P., 1983. Velocity measurements in steady flow through
axisymmetric stences at moderate Reynolds number. J. Biomechanics 16 (7),
505–516.
– LES at a higher Reynolds number (Res = 12,000) for different Chhabra, Rajendra P., 2010. Non-Newtonian fluids: an introduction. In: Deshpande,
flow indexes n, Abhijit P., Murali Krishnan, J., Sunil Kumar, P.B. (Eds.), Rheology of Complex
Fluids. Springer (Chapter 1).
– analysis of higher-order statistics, Dodge, D.W., Metzner, A.B., 1959. Turbulent flow of non-Newtonian systems. AIChE
– analysis of cross-correlation coefficients and kinetic energy, J. 5, 189–204.
– establishment of correlations for the peak position of the RMS of Eggels, J.G.M., Unger, F., Weiss, M.H., Westerweel, J., Adrian, R.J., Friedrich, R.,
Nieuwstadt, F.T.M., 1994. Fully developed turbulent pipe flow: a comparison
the velocity fluctuations (with respect to n) and for the kinetic
between Direct numerical simulation and experiment. J. Fluid Mech. 268, 175–
energy. 209.
El-Emam, N., Kamel, A.H., El-Shafei, M., El-Batrawy, A., 2003. New equation
The present LES predictions and the available literature results calculates friction factor for turbulent flow of non-Newtonian fluids. Oil Gas J.
101 (36), 74–83.
agree quite well. In the fully developed turbulent flow, the Escudier, M.P., Presti, F., 1996. Pipe flow of thixotropic liquid. J. Non-Newtonian
log-region of the mean velocity profile is more expanded with Fluid Mech. 62, 291–306.
decreasing flow index and increasing Reynolds number. The nor- Feiz, A.A., Ould-Rouiss, M., Lauriat, G., 2003. Large eddy simulation of turbulent flow
in a rotating pipe. Int. J. Heat Fluid Flow 24, 412–420.
malized velocity profile agree with the universal linear law, in Germano, M., Piomelli, U., Cabot, W.H., 1991. A dynamic subgrid-scale eddy
the viscous sublayer, for all values of n except for n ¼ 0:5. In this viscosity model. Phys. Fluids A 3, 1760–1765.
case, the flow is transitional; the delay in the transition to turbu- Gomes, F.J.A.D., 1987. Hydraulic, Power Law Models, Calculation Method using
Rational Polynomial Models. Technical Meeting on Software Developing in
lence is caused by the shear-thinning rheology and a high viscosity Drilling Operations, CAPER/87, Salvador, Brazil.
in the flow core. More generally, the results allow to quantify the Kyoungchul, R.O., HongSun, Ryou, 2012. Development of the modified k–
suppression of turbulence fluctuations with decreasing n in the turbulence model of power-law fluid for engineering applications. Sci. China
55 (1), 276–284.
core region, after the peak position, of the RMS of velocity fluctua- Lilly, D.K., 1992. A proposed modification of the Germano subgrid-scale closure
tions. The peak position of all the rms shifts towards the core method. Phys. Fluids A 4, 633–635.
region with increasing n. Correlations to obtain these peak posi- Malin, M.R., 1997. Turbulent pipe flow of power-law fluids. Int. Commun. Heat Mass
Transfer 24 (7), 977–988.
tions are proposed. At Res ¼ 4000, the maximum of the rms of axial
Mishra, P., Tripathi, G., 1971. Transition from laminar to turbulent flow of purely
1=2
velocity hu02
z i as well as the maximum of kinetic energy are viscous non-Newtonian fluids in tubes. Chem. Eng. Sci. 26, 915–921.
1=2 Molla, M.M., Paul, M.C., 2012. LES of non-Newtonian physiological blood flow in a
reached in the buffer layer, while the peak positions of hu02
r i model of arterial stenosis. Med. Eng. Phys. 34 (8), 1079–1087.
1=2
and hu02h i are mainly located in the log-region, for all values of Montreuil, E., 2000. Simulation Numrique Pour l’arothermique avec des Modles
Sous-maille. PhD Thesis, Universit Pierre et Marie Curie, France.
n. The rapid increase in the flatness factor with decreasing n near
Ohta, T., Miyashita, M., 2014. DNS and LES with an extended Smagorinsky model for
the wall reflects the high intermittency of the velocity fluctuations wall turbulence in non-Newtonian viscous fluids. J. Non-Newton. Fluid Mech.
close to the wall. With increasing flow index and increasing 206, 29–39.
Pinho, F.T., Whitelaw, J.H., 1990. Flow of Non-Newtonian fluids in a pipe. J. Non-
Reynolds number, the axial stretching of the turbulent structures
Newtonian Fluid Mech. 34, 129–144.
is reduced, leading to a better developed turbulence. A reduced Ptasinski, P.K., Nieuwstadt, F.T.M., Van den Brule, B.H.A.A., Hulsen, M.A., 2001.
friction factor for decreasing n (at a given Res ) and for increasing Experiments in turbulent pipe flow with polymer additives at maximum drag
Res (at a given n) is also shown. The apparent viscosity hgi is con- reduction. Flow Turb. Combust. 66, 159–182.
Rai, M.M., Moin, P., 1991. Direct simulations of turbulent flow using finite-
stant for a given n up to yþ ’ 5. It increases with increasing n for difference schemes. J. Comput. Phys. 96, 15–53.
yþ < 5, while it decreases with increasing n after the wall distance Redjem-Saad, L., 2008. Simulation Numrique des Transferts de Chaleur Turbulents
yþ ’ 35, for n P 0:69. par Convection Force dans des Conduites Cylindriques et des Espaces
Annulaires. Phd Thesis, Universit Paris-Est/Marne-la-Valle.
The LES results obtained in this study indicate that LES, with the Rudman, M., Blackburn, H.M., 2012. Turbulence modification in shear-thinning
conventional dynamic model of Germano et al. (1991) and Lilly fluids: preliminary results for power law rheology. In: 18th Australian Fluid
(1992) for evaluating the sgs stresses, using a fine grid, reproduced Mechanics Conference. Launceston, Australia.
Rudman, M., Blackburn, H.M., Graham, L.J.W., Pullum, L., 2001. Weakly turbulent
most the features of the non-Newtonian fluid flows, and thus pipe flow of a power law fluid. In: Proceedings of the 4th A/Asian Fluid
seems capable of predicting non-Newtonian fluid flows and gain- Mechanics Conference, Adelaide, 2001.
ing a comprehensive understanding of the space and time dynam- Rudman, M., Graham, L.J.W., Blackburn, H.M., Pullum, L., 2002. Non-Newtonian
turbulent and transitional pipe flow. Hydrotransport 15 (Banff).
ics of such flows, at a reduced computational cost (without
Rudman, M., Blackburn, H.M., Graham, L.J.W., Pullum, L., 2004. Turbulent pipe flow
resorting to any tuning of the sgs turbulence model). of shear-thinning fluids. J. Non-Newton. Fluid Mech. 118, 33–48.
Furthermore, the present predictions could be useful in developing Thais, L., Tejada-Martinez, A.E., Gatski, T.B., Mompean, G., 2010. Temporal large
eddy simulations of turbulent viscoelastic drag reduction flows. Phys. Fluids 22,
a low-Reynolds number turbulence model for predicting
013103.
non-Newtonian fluid flows. Additional studies could be performed Verzico, R., Orlandi, P., 1996. A finite-difference scheme for three dimensional
in the future to further explore whether a new modelling approach, incompressible flows in cylindrical coordinates. J. Comput. Phys. 123, 402–414.
in which the molecular and sgs stresses would be represented as Wang, Bing-Chen, Bergstrom, Donald J., 2005. A dynamic nonlinear subgrid-scale
stress model. Phys. Fluids 17, 035109.
non-linear functions of the strain rate tensor, will bring any addi- Xu, C., Zhang, Z., den Toonder, J.M.J., Nieuwstadt, F.T.M., 1996. Origin of high
tional improvement to LES of non-Newtonian fluid flows. kurtosis levels in the viscous sublayer. Direct numerical simulation and
experiment. Phys. Fluids 8, 1938–1944.
Zahrai, S., Bark, F.H., Karlsson, R.I., 1995. On anisotropic subgrid modelling. Eur. J.
Acknowledgments Mech. B/Fluids 144, 459–486.
Zang, T., 1991. Numerical simulation of the dynamics of turbulent boundary layers:
The authors would like to thank the reviewers for their useful perspectives of a transition simulator. Philos. Trans. Roy. Soc. London A 336, 95–
102.
comments.