Computers & Fluids 94 (2014) 161–172

Contents lists available at ScienceDirect

Computers & Fluids

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m p fl u i d

Turbulent flow simulation using large eddy simulation combined

with characteristic-based split scheme
Lei Tan, Baoshan Zhu ⇑, Yuchuan Wang, Shuliang Cao, Kaihong Liang
State Key Laboratory of Hydroscience and Engineering, Department of Thermal Engineering, Tsinghua University, Beijing 100084, China

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

Article history: A numerical method of large eddy simulation (LES) combined with a characteristic-based split scheme
Received 10 July 2012 (CBS) is proposed. The CBS scheme is introduced to discretize the motion equation in the time domain
Received in revised form 10 January 2014 along the characteristic line, and the space domain is discretized by the split algorithm, which calculates
Accepted 28 January 2014
the velocity and pressure separately. Turbulent flow simulations in a lid-driven cubical cavity and two
Available online 11 February 2014
circular section 90° pipes are conducted, and the results are validated by comparison with experimental
data and other direct numerical simulation results. For a circular section 90° pipe, an additional pair of
Keywords:
vortexes that is near the curved section inner side has been observed, and their rotational direction is
Large eddy simulation
Characteristic-based split scheme
the same as that of the main vortex. To the author’s knowledge, this type of four-vortex structure has
Finite element not been previously reported.
Lid-driven flow Ó 2014 Elsevier Ltd. All rights reserved.
Pipe flow

1. Introduction numerical discretization can be as influential as the subgrid scale

Large eddy simulation (LES) directly calculates the large and
energetic vortical structures in turbulent flows, while modeling
the smaller-scale eddies. Therefore, compared to Reynolds-averaged
Navier–Stokes (RANS) models, the advantages of LES are significant.
RANS is effective for steady simulations of fluid flow, but there are
some theoretical imperfections when RANS equations are used to
simulate unsteady flows. Solutions from the RANS equations usually
deteriorate when the flow field of interest involves large-scale
separations.
Nevertheless, generalization of LES methods to industry is still
problematic. Industrial flows are usually physically complex, highly
unsteady, and have large Reynolds numbers. The key challenges
that LES models must meet for overall success in industrial
applications include accurate flow averaging reflecting the true
flow, minimization of discretization errors, performance of the
simulation in a time and cost-effective manner [1]. Although the
number of LES models has increased almost exponentially in recent
years, the original Smagorinsky model is being favored in a large
majority of cases. Due to the intricate and natural coupling between
numerical discretization and LES modeling, the performance of
simulation is solely dependent on the numerical framework. As
noted in the work of Geurts [2] and Chow and Moin [3], low-order
⇑ Corresponding author.
E-mail address: bszhu@mail.tsinghua.edu.cn (B. Zhu).
(SGS) model. The finite volume method is widely used for LES
because it is easy to compose high-order schemes [4].
The finite element method is not used as widely as that of
finite volume in computational fluid dynamics (CFD), although it
is an important numerical procedure in the area of numerical
simulations. Over the past thirty years, some advanced finite
element schemes have been developed in order to solve convection
dominated flow problems, such as the streamline upwind Petrov–
Galerkin method [5], the Taylor–Galerkin method [6], the Galerkin
least square (GLS) method [7], and the characteristic Galerkin (CG)
method [8]. Among these methods, the CG method discretizes the
time domain along the characteristic line, and it is especially effec-
tive at solving convection dominated flow problems. Donea et al. [9]
extended the fractional step method proposed by Chorin [10] into
a finite element context, and the fractional step algorithm was
used as a stabilization technique to restrain spurious pressure
interpolations violating the so-called Ladyzhenskaya–Babuška–
Breezi condition. Zienkiewicz and Codina [11] presented the well-
known characteristic-based split (CBS) algorithm by introducing
Taylor expansion and combining it with a split algorithm into the
CG method. The CBS procedure combines the advantages of the
characteristic method and the split algorithm, and it is effective
and flexible due to many additional improvements that raise
stability and accuracy for incompressible flows in complex geome-
tries. Codina et al. [12,13] compared CBS procedures with other
formulations, such as GLS and SGS, to solve the incompressible
Navier–Stokes equations. The comparisons showed that all these

http://dx.doi.org/10.1016/j.compfluid.2014.01.037
0045-7930/Ó 2014 Elsevier Ltd. All rights reserved.
162 L. Tan et al. / Computers & Fluids 94 (2014) 161–172

formulations are similar, and are stable for the convective term and C 2s ¼ C 2Limit ; if C s > C Limit : ð9Þ
the pressure interpolation. Over more than ten years of develop-
ment, the CBS procedure has been applied to simulate different flow Here, CLimit is defined according to the Courant–Friedrich–Levy
problems, including turbulent incompressible flows [14]. condition
The present work concentrates on the incorporation of a semi-
implicit CBS method to LES. The finite element discrete equations D2
m þ mLimit ¼ ; ð10Þ
for incompressible flows are derived using the framework of LES, Dt
and these equations are applied to solve three dimensional flow
where Dt is the time step of the calculation and D is the grid filter
problems. Numerical simulations of the flows in a lid-driven cubi-
width, which, in the present work, is determined by selecting the
cal cavity and circular section 90° pipes are conducted. The results
minimum of the three coordinate directions, D = min(Dx, Dy, Dz).
are compared with experimental data and direct numerical simu-
lation (DNS) results. The comparisons show that the proposed
scheme can simulate turbulent flows accurately, although the for- 2.2. Time discretization
mulae of second order accuracy both in time and space have been
used. Furthermore, for a circular section 90° pipe, an additional For simplification, the macron symbol is omitted in the follow-
pair of vortexes has been observed, which, to the author’s knowl- ing formulae. Consequently, the left term of Eq. (7) can be written
edge, has been not reported before. in the form of the total derivative dui/dt based on the characteristic
method. In the characteristic method, the characteristic corre-
2. Mathematical model and numerical algorithm sponds to the path line of a particle. An equation for these charac-
teristics can be written as
2.1. LES mathematical algorithm
dxi
¼ ui : ð11Þ
For incompressible flow, by applying the grid filter to the conti- dt
nuity and momentum equations, the following expressions can be Here, xi(i = 1, 2, 3) is the trajectory and ui the characteristic velocity
obtained: of particle i, where the velocity is constant if a linear convection
@ ui equation (negligible diffusion in Eq. (7)) is considered. Therefore,
¼ 0; ð1Þ along the characteristics given by Eq. (11), Eq. (7) can be written as
@xi
     
@ sij dui @p @ @u @u
 i @ðu
@u
þ
i u
j Þ
¼

1 @p
þ
@
m
@ui @ u
þ
j
 ; ð2Þ ¼ þ v0 i þ j : ð12Þ
@t @xj q @xi @xj @xj @xi @xj dt @xi @xj @xj @xi
The discretization of this equation yields
where variables accompanied by a macron (‘‘–’’) here, as elsewhere,
are resolved scale variables after filtration and sij is the subgrid _  
stress. Based on the widely used subgrid eddy viscosity model,
@uni Dt 2 u k @ _ @uni
_
unþ1
i ¼ uni  Dtu j þ uj
the subgrid stress can be written as @xj 2 @xk @xj
 n  n
n _ @Q _ @E
1 þ ð1  hÞDt Q  Dtu j þ ð1  hÞDt En  Dtu j
sij ¼ 2msgs Sij þ dij skk : ð3Þ @xj @xj
3
þ hDtQ nþ1 þ hDtEnþ1 ; ð13Þ
In Eq. (3) above, Sij is the strain rate tensor in resolved scale and msgs
h  i
is the subgrid eddy viscosity coefficient having the following @p @ui @u
where Q denotes  @x and E denotes @
@xj
m0 @x
þ @xj , respectively.
format: i j i

The above parameter h ranges from 0 to 1 corresponding to differ-

0:5
msgs ¼ ðC S DÞ2 ð2Sij Sij Þ ; ð4Þ ent time discretization formulae. In the present work, we take
h  i
@p 0 @ui @uj
where CS is the Smagorinsky coefficient, which can be dynamically h = 1 for the term  @x i
and h = 0.5 for the term @
@x j
m @xj
þ @xi
.
determined by the dynamic Smagorinsky model (DSM) [15]. We therefore obtain
Substituting Eq. (3) into Eq. (2), we obtain    n 
     n @ui
n
@pnþ1 @ 0 @ui
@unj
 i @ðu
@u i u
j Þ @ p  skk @ @ui @ u
j unþ1
i ¼ un
i þ D t u j  þ m þ
þ ¼ þ þ ðm þ mSGS Þ þ : ð5Þ @xj @xi @xj @xj @xi
@t @xj @xi q 3 @xj @xj @xi    n n 
n @u
1 2 n @ @u @ @u j
For concision, the total viscosity (m + msgs) can be written as m0 and þ Dt uk unj i
 m0 i
þ ; ð14Þ
2 @xk @xj @xj @xj @xi
the term ðp=q þ skk =3Þ as p
. Therefore, the governing equations of
LES introducing the subgrid eddy viscosity coefficient for incom- where the third term on the right hand side is the stabilized term.
pressible flow are The semi-implicit formulation given by Eq. (14) in time means that
i
@u an explicit scheme is used to solve the velocity, and an implicit
¼ 0; ð6Þ
@xi scheme is used to solve the pressure.
  
 i @ðu
@u i u
j Þ 
@p @ i @ u
@u j
þ ¼ þ m0 þ : ð7Þ 2.3. Split procedure and space discretization
@t @xj @xi @xj @xj @xi
Because the DSM is employed, CS is not a constant value, and even a The split procedure CBS operator is introduced to calculate the
negative value is possible. A negative value for CS means that the velocity and pressure separately. The general solution process
energy is transferred from small to large scales in the calculated re- consists of three steps: (1) prediction, (2) projection, and (3)
gion, and numerical instability may occur. Therefore, the following correction. Because eight-noded-trilinear hexahedral elements
limitations are applied to CS: are used in this study to discretize the space domain, the basic
Lagrangian interpolation functions are three-order linear. There-
C 2s ¼ 0; if C s < 0; ð8Þ
fore, the accuracy of the space discretization is second order.
 L. Tan et al. / Computers & Fluids 94 (2014) 161–172 163

(1) In the prediction step, the unknown pressure term (@pn+1/ X

M

^ nþ1
@xi) is ignored, and the intermediate velocity u is solved ui ¼ wum uim ¼ wTu ui ; ð16Þ
i
m¼1
as
where M denotes the number of nodes in an element and ui and wTu
   n 
@uni @ @ui @unj denote the velocity vector and the shape function vector for the
U nþ1
i ¼ uni þ Dt unj þ m0 þ
@xj @xj @xj @xi velocity, respectively. Substituting Eq. (16) into Eq. (15) by using
  n
  the standard Galerkin finite element method, we obtain the follow-
1 2 @ @u
þ Dt unk unj i
; ð15Þ ing equations.
2 @xk @xj  
1 Dt n
for a given finite element, ^ nþ1
u i ¼ un
i þ D tM Cu n
i þ Kun
i þ Bu i þ L þ f 1 ð17Þ
2
Z
M¼ wu wTu dX ð18Þ
XðeÞ

Z   @wT
C¼ wu wTu unj u
dX ð19Þ
XðeÞ @xj
Z
@wu @wTu
K¼ m0 dX ð20Þ
XðeÞ @xj @xj
Z
@wu @wTu
B¼ ðwTu unj ÞðwTu unk Þ dX ð21Þ
XðeÞ @xk @xj
Z
@wu @wTu n
L¼ m0 u dX ð22Þ
Fig. 1. Computational domain and coordinate system for cubical cavity. XðeÞ @xj @xi j

Z !
@wTu n
f1 ¼ wu sij þ wu ðwTu unj ÞðwTu unk Þ ui nj dS ð23Þ
CðeÞ @xj

In Eq. (23), nj is the normal vector component on the boundary sur-

face. Solving the assumed equations, the intermediate velocities
^ nþ1
u i for the entire domain can be obtained.
(2) In the projection step, the pressure is calculated by solving
the Poisson equation based on the intermediate velocity
^ nþ1
u i :
 
@ @pnþ1 1 @unþ1
i
¼ ; ð24Þ
@xi @xi Dt @xi
where
X
M
p¼ wpm pm ¼ wTp p: ð25Þ
Fig. 2. Variation of U with time T.
m¼1

Fig. 3. Vmean contours at the Z = 0.5 plane.

164 L. Tan et al. / Computers & Fluids 94 (2014) 161–172

In Eq. (25), p is the pressure vector in an element and wTp denotes the velocity. Eq. (24) shows that pressure is the first order of time
the shape function vector for the pressure. Substituting Eqs. (16) step although it is an explicit function of time.
and (25) into Eq. (24), we obtain the following equations.
1
 3. Numerical simulation and results analyses
Hpnþ1 ¼ ^ nþ1
Gu u i þ f2 ð26Þ
Dt
! 3.1. Flow in a lid-driven cavity
Z T
@wp @wp
H¼ dX ð27Þ
XðeÞ @xi @xi 3.1.1. Description of the problem
Numerical simulation of turbulent flow in a three-dimensional
Z cubical cavity was first conducted. Fig. 1 shows the computational
@wTu
Gu ¼  wp dX ð28Þ domain, where the origin of the coordinate system is located on
XðeÞ @xi
point O. A prescribed velocity u0 is applied to the top surface of
Z the cavity. The Reynolds number is defined as Re = u0H/m = 104,
@wp nþ1
f2 ¼ wp p ni dC ð29Þ where H is the cubical cavity height and m is the kinetic viscosity.
C ðeÞ @xi Dimensions (x, y, z) are transformed to dimensionless forms using
^ nþ1
(3) In the correction step, the intermediate velocity u i is mod-
ified by the pressure obtained in the projection step.
@pnþ1
unþ1
i ¼ unþ1
i  Dt : ð30Þ
@xi
The discretization form is

unþ1
i
^ ni þ DtM 1 Gp pnþ1 ;
¼u ð31Þ
where
Z
@wTp
Gp ¼  wu dX: ð32Þ
XðeÞ @xi
By accomplishing the above three steps, the entire flow field at time
level n + 1 is obtained, and the next time level can be propelled. Be-
cause a single-step finite difference discretization in time is applied,
we select h = 0.5, as shown in Eqs. (14) and (15), since this value
yields second-order accuracy (Crank–Nicolson scheme) in time for Fig. 6. Urms distribution along the vertical central line at the Z = 0.5 plane.

Fig. 4. Umean distribution along the vertical central line at the Z = 0.5 plane. Fig. 7. Vrms distribution along the horizontal central line at the Z = 0.5 plane.

Fig. 5. Vmean distribution along the horizontal central line at the Z = 0.5 plane. Fig. 8. Tao distribution along the vertical central line at the Z = 0.5 plane.
 L. Tan et al. / Computers & Fluids 94 (2014) 161–172
Fig. 9. Tao distribution along the horizontal central line at the Z = 0.5 plane.
Fig. 10. Computational domain and coordinate system for pipe with Rc/D = 2.0.
H, such that (X, Y, Z) = (x/H, y/H, z/H). DNS results [16,18] and exper-
imental results [17] of an equivalent system are used as references
to test the accuracy of the present scheme. We note here that LES
has also been previously carried out for the same flow [19,20].
Fig. 11. Computational grids for pipe with Rc/D = 2.0.
165
In the calculation, the grid density is 100  100  60, and the
number of eight-noded hexahedral elements is 578,259 with finer
elements near walls. A no-slip condition is taken at wall surfaces
and the pressure is determined according to the second boundary
condition @p/@n = 0. Dimensionless time T in this case is defined as
T = tu0/H. The dimensionless time step in this calculation is taken
as DT = 0.004.
3.1.2. Results and analyses
Taking into account the Reynolds hypothesis, the instantaneous
velocities (u, v, w) are separated into their mean values (umean,
vmean, wmean) and fluctuating parts (u0 , v0 , w0 ). All these velocities
can be made dimensionless using the velocity u0, for example,
U = u/u0, Umean = umean/u0.
Fig. 2 shows the variation of the dimensionless velocity U = u/u0
with respect to dimensionless time at a monitoring point
(X, Y, Z) = (0.5, 0.03333, 0.5). Although strong fluctuations occur for
the instantaneous velocity, after T > 400, the mean velocity
Umean = umean/u0 is statistically stable. The statistical average is
determined over dimensionless time from 600 to 800 with an
internal of 0.04.
Fig. 3 shows the contours of the dimensionless mean velocity
Vmean = vmean/u0 on the Z = 0.5 plane. The calculated results by
DNS [16] are also presented. The micro-flow structures near the
left and right bottoms are nearly equivalent for the two calculation
methods.
Figs. 4–9 show comparisons of the present results with
experimental data [17] and DNS results [18] for the mean velocity,
turbulence intensity, and Reynolds stress. The turbulence intensity
pffiffiffiffiffiffiffiffiffiffi
Urms and Vrms are defined as U rms ¼ 10 hu02 i=u0 and V rms ¼
pffiffiffiffiffiffiffiffiffiffi
10 hv 02 i=u0 , respectively, where the hi operator indicates a
statistical average calculation. The dimensionless Reynolds stress
Tao is given by T ao ¼ 500hu0 v 0 i=u20 . Factors of 10 and 500 are used
to normalize the values of turbulence intensity and Reynolds
stress, so that a suitable graphical representation can be obtained.
According to Figs. 4 and 5, the mean velocity of the present LES
method is in good agreement with DNS and experimental results.
As shown in Figs. 6–9, some differences can be observed in the
turbulent intensity and Reynolds stress components, especially in
the peak values; however, similar characteristics are preserved.
166 L. Tan et al. / Computers & Fluids 94 (2014) 161–172
Fig. 12. W contours at the curved section outlet in pipe with Rc/D = 2.0.
Fig. 13. Variation of pressure coefficient CP with time.
3.2. Flow through curved pipes
3.2.1. Calculated pipes
Curved pipes are widely used in engineering, such as fluid
machinery, heat exchangers, and the piping systems in industrial
plants [21]. Flows in these curved pipes are usually turbulent.
Bends with a circular cross section maybe the most widely used
Fig. 14. Instantaneous velocity vectors at curved section outlet.
curved pipes, and a number of experimental measurement studies
[21–24] and numerical simulations [25–27] have been conducted
for these types of curved pipes.
The pipe calculated in the present study is a circular-sectioned
90° bend, and Fig. 10 illustrates the computational domain and
coordinate system of the bend. The bend used in this study has
been experimentally investigated [21]. The central radius Rc of
the curved section is Rc = 208 mm, the pipe diameter is
D = 104 mm, and Rc/D = 2.0. The length of the straight inlet duct
is twice the pipe diameter, and eight times that of the straight out-
let duct. The geometric parameters of the other pipe with Rc/D = 1.0
also used in the calculations are the same as the pipe with
Rc/D = 2.0, except that, in this case, Rc = 104 mm. For the straight
inlet and outlet ducts, cylindrical coordinate systems (r, h, s) are
used to denote the radial, circumferential, and longitudinal coordi-
nates, respectively. The origins of the two cylindrical coordinate
systems are located at points O1 and O2, as shown in Fig. 10.
A natural coordinate system (r, h, u) is built for the curved section,
where u is the curved angle, such that u = 0° at the inlet of the
curved section and u = 90° at the outlet. The value of the circum-
ferential coordinate h is h = 0° at the inner side of the curved duct
and h = 180° at the outer side. A Cartesian coordinate system (x, y, z)
is built for the entire domain, the origin of which is located at point
O in Fig. 10, which is the center of curvature of the curved section.
Velocity profiles at the pipe inlet are specified according to
experimental values [21], and the velocity direction is taken to
be normal to the boundary. Outflow conditions are taken at the
pipe outlet. A no-slip condition is taken at the wall surface and
the pressure is determined according to the second boundary con-
dition @p/@n = 0.
 L. Tan et al. / Computers & Fluids 94 (2014) 161–172 167

The mean velocity Wa at the pipe inlet is given as Wa = 8.7 m/s, The averaged pressure distributions at the pipe inner and outer
and the Reynolds number is Re = WaD/m = 6  104, the same as gi- sides of the two pipes are given in Fig. 16. The pressure variation
ven in Ref. [21]. Dimensionless time T in this case is defined as tendency and the peak point position of the two pipes are nearly
T = tWa/D. The dimensionless time step in this calculation is taken equivalent, but the variation amplitude is different. The highest
as DT = 0.005. pressure coefficient at the pipe outer side for the pipe with
Rc/D = 1.0 is 1.434 times higher than that for the pipe with
Rc/D = 2.0, and the lowest pressure coefficient at the pipe inner side
3.2.2. Results and analyses of the pipe with Rc/D = 1.0 is 2.596 times lower than that of the
The instantaneous velocities (u, v, w) in the coordinate system pipe with Rc/D = 2.0. It is therefore shown that the curvature has
(x, y, z) are separated into their mean values (umean, vmean, wmean) a greater influence on the flow pattern at the pipe inner side than
and fluctuating parts (u0 , v0 , w0 ). All these velocities can be made at the pipe outer side. The highest pressure point on the outer side
dimensionless using the velocity Wa, for example, W = w/Wa. The exists at the curved section with a curved angle of u  70°, while
pressure is expressed using a pressure coefficient C p ¼ ðP  P0 Þ= the lowest pressure point appears at the curved section with a
ð0:5qW 2a Þ, where P0 is the pressure at the center point of the pipe curved angle of u  20°. Those two positions are nearly symmetric
inlet. about the middle position of u = 45°.
Fig. 11 illustrates the meshes at the pipe inlet (s/D = 2) and Fig. 17 shows the mean velocity W along the horizontal central
curved section. A coarse grid of 200,000 and a fine grid of lines of cross sections in the curved section and outlet straight
540,000 are used in the calculation in order to investigate grid duct. The computational results for the pipe with Rc/D = 2.0 con-
effects. Fig. 12 shows the contours of W at the curved section outlet form well with the experimental data [21]. At the curved section
for the pipe with Rc/D = 2.0. Although the calculation results of the inlet, the symmetric flow pattern formed in the straight inlet duct
two grids agree well with the experimental data [21], the following has been completely disrupted. The velocity at the outer side re-
results are all given from calculations using the fine grid of gion decreases, and the main flow moves to the inner side. In the
540,000. front half part of the curved section, the main flow still deflects
Fig. 13 gives the pressure variation with time from T = 400 to 700
at a monitoring point (0, 156.2 mm, 0) for the pipe with
Rc/D = 2.0, and at point (0, 52.2 mm, 0) for the pipe with Rc/D = 1.0.
Turbulent fluctuations are stronger for the pipe with Rc/D = 1.0,
and the amplitude is about 2.5 times larger than that for the pipe
with Rc/D = 2.0. Although there are strong fluctuations for both
pipes, the statistical mean pressure is stable in both cases over the
time interval T = 400–700. Therefore, the statistical averages are
determined over time T = 400–700.
Fig. 14 shows the instantaneous velocity vectors on the cross
section of u = 90° at time T = 700 for the two pipes of different
Rc/D values. Abundant turbulent vortex structures are apparent
in the instantaneous flow field. The interaction of variously scaled
vortexes makes the flow very complex, especially near the pipe
inner side.
Fig. 15 shows the instantaneous velocity vector in the curved
section at the z = 0 plane at time T = 700 for the two pipes of
different Rc/D values. There is a stronger flow separation for the
pipe with Rc/D = 1.0 than that for the pipe with Rc/D = 2.0. This
separation can induce weak flow separation near the pipe outer
side of the curved section. Fig. 16. Pressure coefficient CP at the pipe inner and outer walls along the flow.

Fig. 15. Instantaneous velocity vectors in curved section.

168 L. Tan et al. / Computers & Fluids 94 (2014) 161–172

Fig. 17. Mean velocity W along horizontal central lines of cross sections.

Fig. 18. Mean velocity vectors in curved section.

to the inner side. Then, the main flow begins deflecting to the outer
side until reaching the position s/D = 7 in the straight outlet duct.
Therefore, the curved section can greatly influence the down-
stream velocity in the straight outlet duct. The velocity distribu-
tions at the front half of the curved section and the straight
outlet duct for different curvatures are similar, but different at
the latter half of the curved section, where u = 60° and 90°.
Fig. 18 shows the mean velocity vectors of the curved section.
The flow is not separated in the pipe with Rc/D = 2.0. However,
for the pipe with Rc/D = 1.0, the separation at the curved section in-
ner side at u = 60° is weak and enlarges to a substantial area at
u = 70°, where the velocity distribution profile appears like an
abrupt slope. The velocity in the central region is lower than that
at the pipe inner and outer sides in the straight outlet duct, where
the velocity distribution profile appears like a collapsed surface.
Fig. 19 shows the streamlines and velocity vectors of the sec-
ondary flow in the pipe with Rc/D = 2.0 at different cross sections.
The streamlines are shown in the upper half and the velocity vec-
tors are shown in the lower half at each cross section. A radial pres-
sure gradient already exists before the fluid runs into the curved
section (see Fig. 16), and causes the fluid to move to the inner side
(see Fig. 19(a)). The centrifugal force existing in the curved section
is not equal to the radial pressure, whose direction is counter to the
centrifugal force. Because the centrifugal force is larger than the
radial pressure in the central region of the curved section, the fluid
moves to the curved section outer side. Because the fluid is
 L. Tan et al. / Computers & Fluids 94 (2014) 161–172 169

Fig. 19. Streamlines and velocity vectors of secondary flow in pipe with Rc/D = 2.0.

inversed near the pipe wall where the velocity is small, it then the horizontal central line. This pair of vortexes is called Dean vor-
moves to the curved section inner side. As a result, two vortexes texes. The intensity of the Dean vortexes increases gradually along
with opposite rotational directions are formed above and below the flow, reaches a maximum at the curved section outlet, and then
170 L. Tan et al. / Computers & Fluids 94 (2014) 161–172

Fig. 20. Streamlines and velocity vectors of secondary flow in pipe with Rc/D = 1.0.
 L. Tan et al. / Computers & Fluids 94 (2014) 161–172 171

Fig. 21. Wrms along the horizontal central lines of cross sections in pipe curved section.

decreases slowly. The shapes of the vortexes at different cross sec-

tions along the flow change continuously.
Fig. 20 shows the streamlines and the velocity vectors of the sec-
ondary flow in the pipe with Rc/D = 1.0 at different cross sections. At
the cross section located at s/D = 3, shown in Fig. 20(f), the stream-
lines have crossed the horizontal central line, making this pair of
vortexes asymmetrical. Rutten et al. [25,26] carried out investiga-
tions and determined that separation vortexes were formed at the
pipe inner wall under conditions of large curvature and turbulent
flow. These separation vortexes were then further separated to
form vortex streets, resulting in unsteady flow and Dean vortexes
oscillation. Therefore, the flow in the large curvature pipe with
Rc/D = 1.0 is more complex than that in the pipe with Rc/D = 2.0.
The main vortexes in the secondary flow structure are
symmetrical, but an additional pair of vortexes appears at the left
of the main vortexes, as shown in Fig. 19(b) in the pipe with
Rc/D = 2.0. This pair of vortexes has the same rotational direction Fig. 22. Secondary flow intensity along the flow.
as the main vortexes, but exists in the regions of u from 25° to
35° at a small scale. Flow bifurcations in curved pipes have been Dennis et al. [29] found by numerical computation that double
studied [27,28], consisting of double vortexes, four vortexes, and solutions for laminar flow existed in a circular section pipe. The
even eight vortexes, commonly in a square section pipe. However, additional vortexes observed in this previous study were located
the structure consisting of four vortexes is rare in a circular at the pipe outer side of the curved section, and the rotational
section pipe because there are no corners at the pipe wall. direction was the reverse of that of the main vortexes. The
172 L. Tan et al. / Computers & Fluids 94 (2014) 161–172
additional vortexes in the present study are located at the pipe in-
ner side of the curved section, and the rotational direction is the
same as that of the main vortexes. To the author’s knowledge,
there are no reports related to this phenomenon.
The root mean square (RMS) velocity is used to present the tur-
bulence intensity. Fig. 21 shows the RMS velocity Wrms along the
horizontal central lines of cross sections in the pipe curved section.
In general, the turbulence intensity near the pipe wall is stronger
than that in the pipe central zone for different curvatures and cross
sections. For the two pipes, the turbulence intensity is much differ-
ent at the pipe inner side, while it is nearly equivalent at the pipe
outer side, which indicates a weaker influence of curvature on the
turbulence intensity at the pipe outer side. At the front part of the
curved section (u = 15°) in the two pipes (Fig. 21(a)), the turbu-
lence intensity at the pipe inner side is weaker than that at the pipe
outer side. Then, at u = 30°, the turbulence intensity at the pipe in-
ner side is approximately equal to that at the pipe outer side, and
gradually exceeds that at the pipe outer side at u = 45° and
u = 60°. However, the turbulence intensity at the pipe outer side
remains about the same. The violent variation of turbulence inten-
sity at the pipe inner side is relevant to the complex vortex move-
ment at this side where the additional vortexes appear.
Fig. 22 shows the secondary flow intensity Is for the two pipes,
where Is is defined as
Z p Z D
8 2 2
Is ¼ 2 2
ðU 2mean þ m2mean Þrdrdh:
pD Wa p 0
2
The secondary flow intensity Is in the pipe with Rc/D = 1.0 is much
larger than that in the pipe with Rc/D = 2.0, by a factor of about
three times at the curved section outlet. However, there is a saddle
point at the curved section inlet in the pipe with Rc/D = 2.0, which
may correspond to the occurrence of the additional vortexes, as
shown in Fig. 19(b).
4. Conclusions
A numerical finite element method is proposed to predict tur-
bulent flow based on LES combined with a CBS scheme. The sub-
grid eddy viscosity model is used to calculate the subgrid stress
in LES, and the Smagorinsky coefficient is dynamically determined
by the dynamic Smagorinsky model. The CBS algorithm is intro-
duced to discretize the motion equation in the time domain along
the characteristic line. The space domain is discretized by the split
algorithm, which calculates the velocity and pressure separately.
The simulations of turbulent flow in a lid-driven cubical cavity
and two circular section 90° pipes were conducted using the pro-
posed numerical scheme. Comparisons of the present calculated
results with experimental data and other DNS results show that
the scheme has good accuracy and stability.
Detailed numerical investigations have been conducted for the
circular section 90° pipes. The results show that curvature has sig-
nificant effects on the velocity distributions in the curved sections
as well as on pressure and secondary flow. An additional pair of
vortexes in the pipe with Rc/D = 2.0 has been observed. The vor-
texes occur at the curved section inner side with the same rota-
tional direction as that of the main vortexes.
Acknowledgements
The authors acknowledge financial support from the National
Natural Science Foundation of China (Nos. 51176088, 51179090)
and the National Basic Research Program of China (‘‘973’’ Program)
(No. 2009CB724304) and the Open Research Fund Program of
State key Laboratory of Hydroscience and Engineering (No.
sklhse-2012-E-02).
References
[1] Bouffanais R. Advances and challenges of applied large-eddy simulation.
Comput Fluids 2010;39(5):735–8.
[2] Geurts BJ. An analysis of errors occurring in large eddy simulations of
turbulence. Philos Trans R Soc A 2009;367:2873–83.
[3] Chow FK, Moin P. A further study of numerical errors in large-eddy
simulations. J Comput Phys 2003;184(2):366–80.
[4] Drikakis D, Hahn M, Mosedale A, et al. Large eddy simulation using
high-resolution and high-order methods. Philos Trans R Soc A 2009;367:
2985–97.
[5] Brooks AN, Hughes TJR. Streamline upwind/Petrov–Galerkin formulations for
convection dominated flows with particular emphasis on the incompressible
Navier–Stokes equations. Comput Methods Appl Mech Eng 1982;32(1):
199–259.
[6] Donea J. A Taylor–Galerkin method for convective transport problems. Int J
Numer Methods Eng 1984;20(1):101–19.
[7] Hughes TJR, Franca LP, Hulbert GM. A new finite element formulation for
computational fluid dynamics: VIII. The Galerkin/least-squares method for
advective–diffusive equations. Comput Methods Appl Mech Eng 1989;73(2):
173–89.
[8] Pironneau O, Liou J, Tezduyar T. Characteristic-Galerkin and
Galerkin/least-squares space–time formulations for the advection–diffusion
equation with time-dependent domains. Comput Methods Appl Mech Eng
1992;100(1):117–41.
[9] Donea J, Giuliani S, Laval H, et al. Finite element solution of the unsteady
Navier–Stokes equations by a fractional step method. Comput Methods Appl
Mech Eng 1982;30(1):53–73.
[10] Chorin AJ. Numerical solution of the Navier–Stokes equations. Math Comput
ð33Þ
1968;22(104):745–62.
[11] Zienkiewicz O, Codina R. A general algorithm for compressible and
incompressible flow – Part I. The split, characteristic-based scheme. Int J
Numer Methods Fluids 1995;20(8):869–85.
[12] Codina R, Zienkiewicz O. CBS versus GLS stabilization of the incompressible
Navier–Stokes equations and the role of the time step as stabilization
parameter. Commun Numer Methods Eng 2002;18(2):99–112.
[13] Codina R, Coppola-Owen H, Nithiarasu P, et al. Numerical comparison of CBS
and SGS as stabilization techniques for the incompressible Navier–Stokes
equations. Int J Numer Methods Eng 2006;66(10):1672–89.
[14] Nithiarasu P, Liu CB, Massarotti N. Laminar and turbulent flow calculations
through a model human upper airway using unstructured meshes. Commun
Numer Methods Eng 2007;23(12):1057–69.
[15] Germano M, Piomelli U, Moin P, et al. A dynamic subgrid-scale eddy viscosity
model. Phys Fluids A 1991;3(7):1760–5.
[16] Leriche E. Direct numerical simulation in a lid-driven cubical cavity at high
Reynolds number by a Chebyshev spectral method. J Sci Comput 2006;27(1):
335–45.
[17] Prasad AK, Koseff JR. Reynolds number and end-wall effects on a lid-driven
cavity flow. Phys Fluids A 1989;1(2):208–18.
[18] Leriche E, Gavrilakis S. Direct numerical simulation of the flow in a lid-driven
cubical cavity. Phys Fluids 2000;12(6):1363–76.
[19] Bouffanais R, Deville MO, Leriche E. Large-eddy simulation of the flow in a lid-
driven cubical cavity. Phys Fluids 2007;19:055108.
[20] Habisreutinger MA, Bouffanais R, Leriche E, Deville MO. A coupled
approximate deconvolution and dynamic mixed scale model for large-eddy
simulation. J Comput Phys 2007;224:241–66.
[21] Sudo K, Sumida M, Hibara H. Experimental investigation on turbulent flow in a
circular-sectioned 90-degree bend. Exp Fluids 1998;25(1):42–9.
[22] Azzola J, Humphrey JAC, Iacovides H, et al. Developing turbulent flow in a u-
bend of circular cross-section: measurement and computation. J Fluids Eng
1986;108(2):214–21.
[23] Wendt G, Mickan B, Kramer R, et al. Systematic investigation of pipe flows and
installation effects using laser Doppler anemometry – Part I. Profile
measurements downstream of several pipe configurations and flow
conditioners. Flow Meas Instrum 1996;7(3/4):141–9.
[24] Kalpakli Örlü AR. Turbulent pipe flow downstream a 90° pipe bend with and
without superimposed swirl. Int J Heat Fluid Flow 2013;41:103–11.
[25] Rutten F, Meinke M, Schroder W. Large-eddy simulations of 90 degrees pipe
bend flows. J Turbul 2001;2:1–14.
[26] Rutten F, Schroder W, Meinke M. Large-eddy simulation of low frequency
oscillations of the dean vortices in turbulent pipe bend flows. Phys Fluids
2005;17(3):035107-1–035107-11.
[27] Crawford N, Spence S, Simpson A, et al. A numerical investigation of the flow
structures and losses for turbulent flow in 90° elbow bends. J Proc Mech Eng
(Part E) 2009;223:27–44.
[28] Selmi M, Nandakumar K. Bifurcation study of flow through rotating curved
ducts. Phys Fluids 1999;11(8):2030–43.
[29] Dennis SCR, Ng M. Dual solutions for steady laminar flow through a curved
tube. Q J Mech Appl Math 1982;35(3):305–24.