Sie sind auf Seite 1von 11

Computers & Fluids 47 (2011) 3343

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

Large-eddy simulations in mixed-ow pumps using an immersed-boundary method


Antonio Posa a, Antonio Lippolis b, Roberto Verzicco c, Elias Balaras d,
a

Dipartimento di Ingegneria Meccanica e Gestionale, Politecnico di Bari, Bari 70125, Italy


Dipartimento di ingegneria dellAmbiente e per lo Sviluppo Sostenibile, Politecnico di Bari, Taranto 74123, Italy
c
Dipartimento di Ingegneria Meccanica, Universit di Roma Tor Vergata, Roma 00133, Italy
d
Fischell Department of Bioengineering, University of Maryland, College Park, MD 20742, USA
b

a r t i c l e

i n f o

Article history:
Received 5 January 2010
Received in revised form 5 October 2010
Accepted 3 February 2011
Available online 18 February 2011
Keywords:
Large-eddy simulation
Finite-difference method
Immersed-boundary method
Turbomachinery ows
Mixed-ow pump

a b s t r a c t
Computations of turbulent and transitional ows in rotating machinery applications are very challenging
due to complexity of the geometry, which usually consists of multiple rotating and stationary parts. The
application of well-established, body-tted methods frequently utilizes overset grids and different reference frames, which have an adverse impact on the overall accuracy and cost-efciency of the method. In
the present work we explore the feasibility of performing computations of such ows using a single reference frame and an immersed-boundary approach. In particular, we report one of the rst large-eddy
simulation in this class of ows, where a structured cylindrical coordinate solver with optimal conservation properties is utilized in conjunction with an immersed-boundary method. To evaluate the accuracy
of the computations the results are compared to the experimental measurements in [1]. Results using the
standard Smagorinsky model and the Filtered Structured Function model are presented. We demonstrate
that the overall approach is well suited for the ow under consideration and the results with the more
advanced subgrid scale model are in good agreement with the experiment. We also briey discuss some
of the features of the instantaneous ow dynamics, to provide a glimpse of the wealth of information that
can be extracted from such computations.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
The development of computational tools to model uid ow
and heat transfer in rotating machinery applications has been at
the forefront of computational mechanics for the past few decades.
The primary challenge that needs to be addressed by all numerical
techniques is the complexity of the geometry, which consists of
multiple rotating and stationary parts. The application of wellestablished, body-tted methods to such problems is not trivial
due to the presence of the moving parts. A frequently adopted
strategy utilizes overset grids and different reference frames to
simulate the ow in the rotor and in the stator respectively (see
for example [14,11]). This approach usually allows for better grid
quality, but the transfer of the solution between reference frames
involves interpolations, which have an adverse impact on the overall accuracy and efciency of the formulation. If a single reference
frame approach is adopted on the other hand, frequent grid deformation/regeneration is necessary, which makes grid quality control problematic [24].

Corresponding author. Tel.: +1 301 405 8268; fax: +1 301 405 9953.
E-mail addresses: antonioposa_81@alice.it (A. Posa), lippolis@poliba.it (A.
Lippolis), roberto.verzicco@uniroma2.it (R. Verzicco), balaras@umd.edu (E. Balaras).
0045-7930/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compuid.2011.02.004

An alternative class of methods that are well suited for applications involving moving boundaries are the so-called immersedboundary type methods, which have been gaining popularity in
the past few years. In such case the requirement for the computational grid to be conform to the body is relaxed and structured
Cartesian or generalized coordinate solvers can be utilized in
highly complex congurations. The presence of a complex boundary is introduced by a discrete forcing function, which is designed
to mimic the effect of the body on the ow. Compared to the
boundary tted strategies above, immersed-boundary methods
greatly simplify the grid generation and the same frame of reference can be used for systems with bodies in relative motion. In
addition, the conservative non-dissipative structured solvers utilized by most immersed-boundary approaches can be ideally coupled to eddy-resolving methods such as direct numerical
simulations (DNS) and large-eddy simulations (LES), thus substantially enhancing the overall predictive capability of the computational tools. Today, immersed-boundary methods have been
utilized in a variety of elds. Early applications were mostly conned to biological ows (see for example [17,18]), but over the
past years a variety of applications in very diverse elds has been
reported. Verzicco and co-workers, for example, analyzed the ow
in a piston-cylinder assembly [6], and an impeller-stirred tank [26],
using a direct-forcing, immersed-boundary approach. More

34

A. Posa et al. / Computers & Fluids 47 (2011) 3343

Nomenclature

DEk
D
hi
hhii

s
mt
x
xs
/

w
~
p
e
S
~
u

e
F2
Ain
i
Ck
Cs
Din
d
Dout
d
Din
i

cyclic unsteadiness
local grid size
phase average operator
passage average operator
subgrid scale stress (SGS) tensor
eddy viscosity
rotational speed of the impeller
specic-speed of the pump
ow coefcient
head coefcient
resolved pressure
resolved strain rate tensor
resolved velocity vector
angle dening the position of the impeller
second-order ltered velocity structure function
area of the impeller at the inow
Kolmogorov constant
Smagorinsky model constant
diameter of the diffuser at the inow
diameter of the diffuser at the outow
diameter of the impeller at the inow

recently applications ranging from complex uid-structure interactions [29,4] to magnetohydrodynamics (MHD) [8] appeared in
the literature. Details on the available immersed-boundary formulations as well as their range of applicability can be found in recent
reviews in [10] and [15].
Turbomachinery applications are a great challenge for immersed-boundary formulations and a stringent test for their
robustness and accuracy. As of today very few applications can
be found in the literature, which usually deal with simplied congurations. You et al. [30,31], for example, performed LES using an
immersed-boundary method and a structured generalized curvilinear coordinate solver, to study the viscous losses associated to
the tip-clearance ow in a linear cascade. To reduce the cost of
the computations a single vane was considered, using periodic
boundary conditions. Their results revealed some interesting
dynamics and the importance of the tip-leakage jet on the vorticity
and turbulence generation. They also demonstrate the importance
of utilizing eddy-resolving methods in such applications, where
experimental measurements are very challenging and classical
Reynolds Averaged Navier Stokes (RANS) models cannot capture
the highly three-dimensional and complex ow physics. As of today, immersed-boundary methods have not been utilized in more
complex turbomachinery ows. The aim of the present work is to
explore the feasibility of utilizing such an approach in the case of
turbopumps, where the complete machine is simulated. In addition, we will investigate the accuracy and applicability of standard
subgrid scale (SGS) models available in the literature. We should
also note, however, that due to cost considerations, the proposed
approach targets primarily laboratory scale pumps. Extension to
industrial scale pumps would require further developments in
approximate wall boundary conditions and sophisticated adaptive
mesh renement methods (see for example [25]).
Currently the majority of turbopump computations are conducted with body-tted meshes utilizing different frames of reference for the stationary and moving parts. These methods usually
employ stable, dissipative, discretizations and are therefore better
suited to RANS modeling strategies. Within this modeling framework, Goto and co-workers [7,23] reported three-dimensional simulations of mixed-ow pumps. They focused on the mechanism of
generation of secondary ow patterns and the formation of the
jet-wake like ow. They also investigated the sensitivity of these

Dout
diameter of the impeller at the outow
i
H
head
N
number of the blades passages
NU
number of phase averages
Nd
number of the diffuser blades
Ni
number of the impeller blades
number of the impeller revolutions
Nr
Q
ow rate
r, h, z
coordinates of the cylindrical system
Re = UL/m Reynolds number based on a reference velocity, U,
length, L, and the kinematic viscosity m.
T
period of the passage of the rotor blades
t
time
t0
initial time instant for the evaluation of the phase-averaged elds
uin
tangential velocity of the tip of the impeller blades at
i
the inow
uout
tangential velocity of the impeller blades at the outow
i

patterns to the tip-clearance. Kaupert et al. [13] investigated the


effects of pressure discontinuities in high-specic-speed radial
pumps, showing reasonable agreement between experiments and
computations. More recent RANS applications include the study
of the ow in a backswept centrifugal impeller by Zhang and coworkers [32,33], the impeller-diffuser interaction in a diffuser
pump stage [20] and various congurations in turbopumps [9,3].
LES of such ows are very limited. Byskov et al. [2] performed
RANS and LES simulations on a six-bladed centrifugal pump impeller. Kato and co-workers [11,12] conducted LES using the Smagorinsky model with van Driest damping near the walls to simulate
a high-specic-speed mixed-ow pump. Pouffary et al. [19] and
Yamanishi et al. [28] used LES to study cavitation phenomena in
centrifugal pumps, with reasonable agreement with experiments.
In the present work we report LES in a mixed-ow pump, where
a structured cylindrical coordinate solver with optimal conservation properties is utilized in conjunction with an immersed-boundary method. Both the Smagorinsky (SM) model [21] and the
Filtered Structured Function (FSF) [5] model were tested. The results are compared to the experimental measurements in [1]. In
the following sections the mathematical formulation and the problem setup are given. Next the results are discussed followed by
some concluding remarks.
2. Mathematical model and numerical method
In the LES approach the large and energetic eddies, which are
resolved directly as in a DNS, are separated from the small scales,
which are modeled, by the application of a ltering operation.
The resulting governing equations for the case of incompressible
ows are as follows:

~0
ru

~
@u
1
~u
~ rp
~ f;
~  r  s r2 u
ru
@t
Re

~ is the velocity vector, p


~ is a
where  denotes the ltered variables, u
modied pressure which also includes the trace of the SGS stress
tensor, s is the SGS stress tensor, and f is a forcing term. t is the time
and Re = UL/m is a Reynolds based on a reference velocity scale, U,
and a length scale, L, which will be dened in the following sections

35

A. Posa et al. / Computers & Fluids 47 (2011) 3343

Fig. 1. (a) Schematic of the ray-tracing procedure used to classify the nodes on the Eulerian grid into  interior,  uid, and h interface ones. Symbol  is a control point and
the dashed lines represent control rays. (b) Linear reconstruction of the solution in the vicinity of the immersed body [6]. Cyan: uid domain; Gray: solid body.

(m is the kinematic viscosity of the uid). Note that we will use


cylindrical coordinates to achieve an efcient distribution of the
grid nodes in the present geometrical conguration.
Two different models are considered for the parametrization of
the SGS stresses: the standard Smagorinsky model [21] and the ltered structure function model [5]. The former, despite its limitations, is a popular and inexpensive eddy viscosity model that is
frequently adopted in rotating machinery applications. The latter
adjusts locally the eddy viscosity based on the available resolved
energy near the cutoff wavenumber, resulting in a more accurate
reproduction of the energy dissipation especially in transitional/
relaminarizing ow congurations such as the one considered in
this study. In the both cases the deviatoric part of the SGS tensor
is modeled using an eddy viscosity model of the form:

s  Trs 2mt eS;

where e
S is the resolved strain rate tensor. In the case of the SM
model the eddy viscosity is dened as:

mSM
C s D2 j e
Sj; and j e
Sj 2 e
Se
S1=2 ;
t

where Cs is the Smagorinsky constant and D is the local lter size.


No effort was made to include wall dumping into the model, which
requires the computation of the distance from the closest wall. In
this application, where the solid surface is never aligned to any
coordinate direction, this is a rather challenging task with a substantial computational overhead. In addition, the denition of such
a distance is ambiguous due to the presence of multiple rotating
and stationary surfaces. For the case of the FSF model the eddy viscosity is dened as follows:
3=2
mFSF
0:0014C k D e
F 2 x; D; t1=2 ;
t

their relation to the immersed body. A ray-tracing procedure is


used for this purpose and all grid points are classied as internal,
uid, and interface points. Internal are the points that are located
within the solid body and interface points are uid points with
at least one neighbor in the solid body. The tracking scheme is
illustrated in Fig. 1a: if C is a control point on the Eulerian grid outside the body, a control ray is a segment which connects C to any
grid point in the neighborhood of the body. For each ray, the intersections with the triangles on the body are found. If the number of
intersections is even, the corresponding grid point is external to
the body, otherwise it is internal. The interface points are central
to the reconstruction of the solution near the solid body. Following
Fadlun et al. [6], a linear reconstruction of the velocity is considered. As shown in Fig. 1b, the desired velocity, Vi, at the interface
points is computed using a linear interpolation along a grid line
(the velocity values Vc and ue, respectively on the body and in
the adjacent external point, are known); then the forcing function
in the Eq. (2) can be found by substituting the interpolated value
for Vi in the discretized momentum equation:

where e
F 2 is a ltered, second-order, velocity structure function (see
[5] for details).
The governing equations are advanced in time using a fractional
step method. The viscous terms in the momentum equation are
treated implicitly with a CrankNicolson scheme and all other
terms are treated with an explicit AdamsBashfort scheme. All spatial derivatives are approximated with second-order nite differences on a staggered grid. The large-band matrix associated with
the solution of the Poisson equation is rst reduced to pentadiagonal problems using FFTs in the azimuthal direction, and then each
problem is solved with a generalized cyclic reduction algorithm
[22]. Details on the overall formulation can be found in [27].
To simulate the ow in a complex conguration, which is immersed in the cylindrical coordinates grid, the direct-forcing approach proposed by Fadlun et al. [6] is utilized. The body is
represented by a series of triangles in stereo-lithography (STL) format, where their density depends on the local curvature of the surface. Initially all points on the Eulerian grid are tagged according to

l12

~l
1
Vl1
u
i
 RHSl2
Dt

where l; l 12 and l + 1 refer to the time level, and the RHS includes the
discrete convective, viscous, and SGS terms in Eq. (2). More details on
the implementation of the above method can be found in [6].
3. Computational setup
We will consider a mixed-ow pump conguration, which has
been studied experimentally by Boccazzi et al. [1]. They reported
Particle Image Velocimetry (PIV) measurements inside the vaned
diffuser at two stations. The geometry of the pump is shown in
Fig. 2a. To better distinguish the different elements each one is represented by a different color: the rotor blades are drawn in red1,
the hub in white and the diffuser blades in blue. Note that the shroud
(green), the diffuser throat and the volute (gray) are partially shown,
for the internal pump elements to be visible. A meridian cut of the
volute geometry is also shown in Fig. 2b. The number of blades for
the rotor and the stator are Ni = 6 and Nd = 7 respectively. The specic-speed, xs, the ow coefcient, /, and the head coefcient, w,
are dened as follows:

xs

x Q
gH

3
4

Q
in
Ain
i ui

w

gH
2 ;
uout
i

where x is the rotational speed of the impeller, Q the ow rate, H


in
out
the head, Ain
are
i is the area of the impeller inlet and ui and ui
1
For interpretation of color in Figs. 13, 58, 12 and 13, the reader is referred to the
web version of this article.

36

A. Posa et al. / Computers & Fluids 47 (2011) 3343

Fig. 2. Geometry of the mixed-ow pump used in the experiments by Boccazzi et al. [1]. (a) Cross cut of the volute; (b) meridian cut of the volute.

Table 1
Main parameters of the mixed-ow pump used in the experiments by Boccazzi et al.
[1].

x = 55.4[rad/s]

xs = 1.08

/opt = 0.314

wopt = 0.443

Dout
i

Ni = 6

Din
d

Dout
d

0:224m

0:233m

0:361m

Din
i 0:154m
Nd = 7

the tangential velocities at the impeller inow and outow respectively. For the inow the speed is referred to the blades tip. All the
geometric and operating parameters are summarized in Table 1,
and are the ones considered in all computations below. Note that
out
Din
are the impeller diameters at the inlet and the outlet,
i and Di

out
while Din
are the corresponding values for the diffuser.
d and Dd
The spanwise dimension of the diffuser is Ls = 44[mm]. The ow
rate corresponding to /opt is Q = 0.025[m3/s]. The Reynolds number,
based on the average inow velocity and the external radius of the
rotor, is equal to Re = 1.5  105.
The pump is fully immersed in the computational domain,
which is discretized with a structured cylindrical coordinates grid.
An example is shown in Fig. 3, where two grid slices in the r  z
and r  h planes are included. The blue portion of the gure represents the part of the structured grid that falls within the uid
domain. The red portion marks the grid nodes that are wasted.
For clarity the volute, the shroud and the inow channel are not

Fig. 3. Pump geometry immersed in the computational mesh. (a) Computational grid in the rz plane; (b) computational grid in the rh plane. Blue color represents the part
of the structured grid that falls within the uid domain. The red color marks the wasted gird nodes. Note that only the impeller and the diffuser blades are shown for clarity.

A. Posa et al. / Computers & Fluids 47 (2011) 3343

hf ih; r; z; u

Fig. 4. Detail of the computational grid in the divergent channel between the
diffuser blades (an r  h plane is shown).

included in this gure. The inow and outow locations are also
indicated. In all computations reported in this study we prescribe
a uniform velocity at the inow plane, while we use a convective
condition at the outow [16]. The convective condition is applied
on the local ow direction, using the averaged velocity on the outow surface as convective speed. The immersed-boundary technique described in the previous section is used to enforce the
boundary conditions on all solid boundaries.
The computational grid consists of 801  350  101 points in
the azimuthal, radial and axial directions respectively (a total of
28 million). It is uniform in the azimuthal and axial directions
and is stretched in the radial one to cluster points in the areas of
high velocity gradients. Since the grid does not conform to the
body, a signicant number of nodes falls outside the useful computational domain, which in the present case is approximately 50%.
This grid, which has a total of 14 million useful nodes, provides
adequate resolution in many critical areas of the computational
box. In the divergent channel between the diffuser blades, for
example, along the direction normal to the blades surface, respectively 40, 45 and 52 computational cells are utilized over the lines
a, b, and c in Fig. 4. Each computation on the above grid requires
approximately 56 CPU hours/revolution on a single 2-way quadcore Opteron 2.1 GHz node with 16 GB of RAM. All eight cores on
the node were utilized using OpenMP. Due to cost considerations,
we could not perform a detailed grid renement study on the full
computational domain, and the above resolution is the maximum
we could achieve with the available computational resources at
the time of the preparation of the manuscript. We did, however,
conduct a grid renement study on a reduced domain, where only
one diffuser passage is considered using periodic boundary conditions. In particular, we compared the results between a coarse grid
with equivalent resolution to the one utilized above in the full domain, and a ne grid with double resolution in the azimuthal and
radial directions. The mean velocities, although are not directly
comparable to the experiment in [1], due to the simplications in
the geometry, were in good agreement and the largest discrepancies (order of 10%) were observed near the walls.
4. Results
In this section we will report a series of computations conducted in the conguration presented above. The primary aim of
the simulations is to establish the accuracy and robustness of the
immersed-boundary methods in rotating machinery applications
by direct comparisons to the experimental results by Boccazzi
et al. [1]. We will also present the results from different SGS models, as well as, an overall view of the instantaneous ow dynamics
in the particular conguration. To facilitate comparisons with the
experiment, for any ow variable f(h, r, z, t), which is a function of
space and time, the following phase average operator is dened:

N
1 X
f h; r; z; t 0 n  1T;
N n1

37

where h, r and z are the azimuthal, the radial and the axial coordinates respectively; u is the angle which denes the position of the
impeller with respect to the stator, for which the phase average is
evaluated; in the present case its value is variable between 0 and
60, since after one blade passage the same impeller-diffuser conguration is repeated; N = NiNr is the total number of blades passages,
i.e., the number of times that the impeller is in the same position
relative to the diffuser (Nr is the number of the impeller revolutions); t0 is the initial time instant and T the period of the passage
of the rotor blades. We can also dene the passage average operator
as follows:

hhf iih; r; z

NU
1 X
hf ih; r; z; u;
NU i1

which is simply the average of the phase-averaged elds, with NU


being the number of phase averages, namely the number of u values for which the phase averages are evaluated. In all computations
reported in this study the phase averages were computed over 5
impeller revolutions, and NU = 12. We found that this sample was
sufcient to obtain smooth rst- and second-order statistics. We
should also note that approximately seven revolutions were required for the effect of the initial conditions to decay and a quasi
periodic state to be reached prior to sampling.
4.1. Comparison to the experiments
First, the head coefcient, w, and the efciency, g, are evaluated
for both computations. For the case of the LES with the SM model
(SM-LES) they are w = 0.44, g = 0.74, and for the FSF model (FSFLES), w = 0.47, g = 0.80. The corresponding experimental values
are w = 0.44 and g = 0.75, respectively. The results from the SMLES are in very good agreement with the experiment (within
1.5%). The FSF-LES slightly over-predicts both w and g by approximately 6%. A possible reason for this behavior is the less dissipative
character of the second model.
In Fig. 5 the phase-averaged velocity magnitude at the diffuser
midspan is shown, for window A indicated in Fig. 6a. These phase
averages correspond to a conguration during the revolution for
which the position of the impeller relative to the diffuser is such
that the trailing edge of a rotor blade at the midspan is in the same
azimuthal position as the leading edge of the stator blade number
4 (see Fig. 6a for the numeration of the diffuser blades and the position of the experimental windows). Results from the LES with the
SM model (SM-LES) and the FSF model (FSF-LES) are compared to
the experiment. It can be seen that all computations capture the
main ow features in this window. The SM-LES, however, overpredicts the velocity in the shear layers developing at the trailing
edge of the impeller and produces higher velocities in the stator
passage between blades 3 and 4. The results of the FSF-LES are in
good agreement with the experiments. Both models, however,
seem to underestimate the wakes of the impeller vanes propagating within the diffuser channels.
A detailed quantitative comparison is shown in Fig. 6, where the
phase-averaged velocity proles at the diffuser midspan in the
direction normal to the pressure side of the blade are shown. In
particular, Fig. 6b shows the tangential velocity at 10% of the chord
of the blade 5 and Fig. 6c and d the tangential velocities at 50% and
90% of the chord of the blade 4, respectively. The exact locations of
the proles, which match the ones in the experimental dataset, are
shown in Fig. 6a. Since these locations are not aligned with the grid
lines, the velocities were interpolated from the surrounding grid
nodes. The spacing on each line was selected to be similar to the

38

A. Posa et al. / Computers & Fluids 47 (2011) 3343

Fig. 5. Phase-averaged velocity magnitude at the diffuser midspan for window A shown in Fig. 6a. (a) PIV measurements [1]; (b) FSF-LES; (c) SM-LES. The velocity magnitude
shown ranges from 0[m/s] (blue) to 4[m/s] (red).

Fig. 6. (a) Experimental measurement windows, blade numbers, and velocity prole locations. Also the following abbreviations are dened: RBSS: rotor blade suction side;
RBPS: rotor blade pressure side; SBPS: stator blade pressure side; SBSS: stator blade suction side. (b) Phase-averaged tangential velocity prole at location a (10% of the chord
of the blade 5); (c) Phase-averaged tangential velocity prole at location b (50% of the chord of the blade 4); (d) Phase-averaged tangential velocity prole at location c (90% of
the chord of the blade 4).  experiment [1], 33 SM-LES,
FSF-LES.

Fig. 7. Phase-averaged velocity magnitude at the diffuser midspan for window B shown in Fig. 6a. (a) PIV measurements [1]; (b) FSF-LES; (c) SM-LES. The velocity magnitude
shown ranges from 0[m/s] (blue) to 4[m/s] (red).

A. Posa et al. / Computers & Fluids 47 (2011) 3343

39

Fig. 8. Phase-averaged tangential velocity proles at the diffuser midspan. (a) location d (10% of the chord of the blade 0); (b) location e (50% of the chord of the blade 0).
experiment [1], 33 SM-LES,
FSF-LES.

Fig. 9. Passage-averaged velocity magnitude (top part) and cyclic unsteadiness (bottom part) at the diffuser midspan for window A shown in Fig. 6a. (a) and (c) PIV
measurements [1]; (b) and (d) FSF-LES.

local grid spacing. In general the agreement with the experiments


at all stations is satisfactory. At station a (Fig. 6b) both computations agree with the experiments and only near the pressure side
of the diffuser blade 4 the velocity is under-predicted, most probably due to the lack of resolution in the thin boundary layer. At station b (Fig. 6c) the differences between the SM model and the
experiments are more profound. This is due to the dissipative character of the SM model near the walls, which results in excessive
dissipation and under-prediction of the wall shear stress. To

conserve mass the velocity increases away from the wall and a local maximum is generated at the edge of the boundary layer. The
FSF is in better agreement with the experiment. At location c near
the trailing edge of blade 4 (Fig. 6d) both computations agree well
with the experiment except for a small area near the wall.
Fig. 7 shows the phase-averaged velocity magnitude at the midspan of the diffuser throat in window B, near the nose of the volute.
The results with the FSF model agree well with the experiment. For
the SM-LES the ow separates much earlier on blade 0 and

40

A. Posa et al. / Computers & Fluids 47 (2011) 3343

generates a large separated region. This has a direct impact on the


angle the ow has as it hits the volute nose, producing a large recirculation area within the volute and between the same volute and
the blade 0. To better quantify the differences between the different computations, also in this case the corresponding phase-averaged tangential velocity proles are shown in Fig. 8. The location
d at 10% of the chord of the blade 0 (Fig. 8a) is before the separation
point, and all numerical results near the pressure side of blade 0
agree well with the experiment. On the suction side of the blade
6, however, the velocity is underestimated, especially with the
SM model. It is worth noting that on the pressure side of the blade
6 the SM-LES predicts slightly negative velocities, indicating that
the ow has already separated. The differences between the SGS
models are more evident at location e at 50% of the chord of the
blade 0 (Fig. 8b). Due to the early separation the SM-LES severely
under-predicts the velocity near the wall. The FSF-LES, on the other
hand, is in good agreement with the experiment.
Due to cost considerations we only computed phase-averaged
statistics for 12 different relative positions between the impeller
and the diffuser (every 5), which also coincide with the corresponding ones from the experiments. Similarly to what has been
observed above, the phase-averaged velocity eld for the case of
FSF-LES agrees well with the experiment for all 12 locations (not
shown here). The SM-LES always results in under-prediction of
the wall shear stress due to its overly dissipative character. In
Fig. 9a and b passage-averaged velocity elds are shown in window A for both the experiment and the FSF-LES. Note that the
impeller blades are not shown, since the passage-averaged eld
is not related to a particular impeller position. The agreement between the computation and the experiment is very good. Another
quantity usually adopted in rotating machinery applications to
characterize the unsteadiness associated to the large scales is the
cyclic unsteadiness, dened as:

1
DEk h; r; z; u hv i2 h; r; z; u  hhv ii2 h; r; z;
2

10

where hvi is the phase-averaged velocity magnitude and hhvii the


passage-averaged velocity magnitude. In Fig. 9c and d the cyclic
unsteadiness is shown for the PIV measurements and the numerical
results from the FSF-LES. The structure of the impeller blade wake
inside the diffuser is well predicted and the local minima on the
pressure side near the leading edge of the stator blade 5 are also
reproduced by the simulation.

4.2. Instantaneous ow dynamics


Having established the accuracy of our computations we will
now examine the instantaneous ow dynamics. Note that we will
only present results from the FSF-LES, which was found to better
agree with the experiment. In Fig. 10 the instantaneous vorticity
magnitude is shown for four consecutive positions of the impeller,
which are 15 apart. Most of the activity, as indicated by the high
vorticity values, is concentrated on the suction side of the impeller
and stator blades. It is also evident that the vortical structures generated on the suction side of the impeller blades and consequently
shed in the wake, interact strongly with the suction side of the stator blades. This is clearly seen in the time sequence shown in
Fig. 10ad: in part (a) of the gure, the tip of the impeller blade
a is close to leading edge of the stator blade 4, generating a disturbance, as manifested by the local increase of the vorticity magnitude in the area; as this impeller blade moves forward (see parts
bd in the gure) its wake impinges on the stator blade and disturbs the boundary layers on its surface, indicating that transition
to turbulence probably occurs through a bypass transition mechanism. On the pressure side of the impeller blades, on the other
hand, the favorable pressure gradient leads to reduced turbulent
activity as shown in Fig. 11, where the impeller-diffuser conguration of Fig. 10a is shown from a different angle. In Fig. 11 it is also
evident that the vorticity magnitude on the pressure side of the diffuser blades is lower when compared to the one on the suction
side. We should note, however, that the effect of the impeller wake

Fig. 10. Instantaneous snapshots of the vorticity magnitude at the midspan of the diffuser for the FSF-LES. Consecutive positions of the impeller blades are shown every 15.

A. Posa et al. / Computers & Fluids 47 (2011) 3343

41

is noticeable also on the pressure side of the stator blades, but in


this case the action of the wake on the boundary layer on the
blades surface is not as profound.
To gain a better insight on the dynamics of the coherent structures which are generated in the rotor and interact with the stator,
the second invariant of the velocity gradient tensor, Q, is utilized.
Fig. 12a shows an instantaneous picture of the ow by means a
Q isosurface, colored by the vorticity magnitude. Four different
groups of coherent vortices can be identied:

Fig. 11. Instantaneous vorticity magnitude at the midspan of the diffuser for the
FSF-LES. Representation of the vorticity values on the pressure side of the rotor and
the stator blades. The same instant in time is shown as in Fig. 10a from a different
angle, in order for the pressure side of the blades to be visible.

(i) The packet of vortices within the area a in the gure, which
is produced near the shroud. In general, the divergent geometry of the shroud generates large adverse pressure gradients
and at the same time the ow turns rapidly by 90, resulting
in complex separation phenomena. The vortices shown in
area a are probably related to the above phenomena and
are not as coherent as the ones in other areas of the machine.

Fig. 12. (a) Isosurfaces of Q = 5  104[1/s2], colored with the instantaneous vorticity magnitude, for the FSF-LES. Green: shroud; Gray: impeller blades, diffuser wall and
volute; Cyan: hub; Blue: diffuser blades; (b) magnication of area b shown in part (a) of the gure (note that Q = 1  105[1/s2] isosurfaces are shown in this case); (c)
instantaneous vorticity magnitude in the section AA shown in part (a) of the gure. DWSS: diffuser wall on the shroud side; DWHS: diffuser wall on the hub side.

42

A. Posa et al. / Computers & Fluids 47 (2011) 3343

Fig. 13. Instantaneous azimuthal (top) and radial (bottom) vorticity elds on the circumferential section at the impeller mean radius from the FSF-LES: (a) 3D view (white:
hub, gray: impeller blades, green: shroud, blue: inow channel); (b) top view of the unrolled plane (RBPS: rotor blade pressure side, RBSS: rotor blade suction side).

(ii) The vortices within the area b in the gure, which is located
at the suction side of the impeller blades. These are much
more coherent, compared to the ones above. These vortices
appear approximately at the same location in all impeller
vanes. In Fig. 12b a closer view of the same area is shown
and one can identify strong co-rotating vortices, which are
aligned with the ow.
(iii) The smaller eddies within the area c in the gure, which are
associated with the wake of the impeller blades. The wake
area is strongly inuenced by the interaction of the vortices
in b and the boundary layer on the impeller blades. The wake
vortices quickly lose their coherence as they move away
from the trailing edge and appear to be the main source of
turbulent uctuations for the ow entering the diffuser
throat.
(iv) Vortical structures inside the diffuser and particularly in the
wake of the diffuser blades (area d in the gure). These form
larger vortices which are shed from the trailing edge, but
they are weaker than the corresponding ones at the rotor
outlet.
To better characterize the coherent structures in the diffuser,
Fig. 12c shows the instantaneous vorticity magnitude at a plane
perpendicular to the ow direction (section AA in Fig. 12a). The
selected position is close to the leading edge on the pressure side
of the blade 2, and to the trailing edge on the suction side of the
blade 1. On the pressure side there are still high vorticity values,
due to the proximity of the wake of the impeller blades. On the
suction side, non-uniformities in the vorticity distribution along
the blade span can be observed, which are probably caused by
the ow coming from the impeller shroud. Fig. 13 shows the
instantaneous azimuthal and radial vorticity on a circumferential
section at the impeller mean radius. Part (a) shows the location
of the section inside the impeller, while in part (b) the plane is unrolled and shown from the top. Referring to part (a) in the gure,
the ow enters the impeller axially, from left to right, and exits
radially. Consistent with the observations above, close to the
shroud (area a) and at the blade suction side (area b), the footprint
of strong vortices can be seen, with the latter being more coherent
as well as having the same vorticity sign (co-rotating). We should

also note that, even if the machine geometry is not symmetric, similar eddy patterns can be identied at all impeller vanes.

5. Summary and conclusions


Turbomachinery applications are a great challenge for
immersed-boundary formulations and a stringent test for their
robustness and accuracy. As of today very few applications of immersed-boundary methods to such ows can be found in the literature and usually deal with simplied congurations (see for
example [30,31]). The aim of the present work is to explore the feasibility of utilizing such an approach in the case of turbopumps,
where the complete machine is simulated. We considered a
mixed-ow pump conguration, which has been studied experimentally by Boccazzi et al. [1]. The pump has been fully immersed
in the computational domain, discretized with a structured grid in
cylindrical coordinates, using approximately 28 million points. The
immersed-boundary technique by Fadlun et al. [6] was used to
enforce the boundary conditions on all solid boundaries.
We tested two different SGS models: the standard Smagorinshy
model [21] and the ltered structure function model [5]. The former was not expected to capture the complex physics in such
ows, due to the lack of proper limiting behavior near walls and
its inability to switch off in areas of laminar ow. Also the use of
a single constant cannot accurately capture the energy dissipation
in the wall layers as well as in the free shear layers in the machine.
Nevertheless it is a popular and inexpensive eddy viscosity model,
which is frequently adopted in rotating machinery applications,
and therefore a closer look at its performance would be useful.
The FSF model, on the other hand, adjusts locally the eddy viscosity
on the basis of the available resolved energy near the cutoff wavenumber, resulting in a more accurate reproduction of the energy
dissipation, especially in transitional/relaminarizing ow congurations, such as the one in this study.
In general, the overall agreement with the experimental results
is good, demonstrating the robustness and feasibility of the approach in rotating machinery applications. Both SGS models capture the main physics of the ow, but as expected the excessive
dissipation produced by the SM model, especially near walls,

A. Posa et al. / Computers & Fluids 47 (2011) 3343

results in the underestimation of the wall stresses and early separation. The results with the FSF model are in very good agreement
with the experiments, indicating that it is a much better choice for
such ows, especially considering that its computational overhead
is not much higher than the one of the SM model. We should also
note, however, that due to cost considerations, a comprehensive
grid renement study was not feasible. Simulations on a reduced,
periodic domain, at comparable and ner grid resolutions, provided mean velocity statistics which differed no more than 10%,
giving us some condence in the results reported here. We are currently working on improving the parallel performance of the solver, which will enable us to do detailed grid renement studies
in the full domain as well.
Finally an outline of the instantaneous ow dynamics is presented. It is evident that the vortical structures generated on the
suction side of the impeller blades, and consequently shed in the
wake, interact strongly with the suction side of the stator blades,
indicating that transition to turbulence on the latter occurs
through a bypass transition mechanism. The inuence of the
impeller wake is noticeable also on the pressure side of the stator
blades, but in this case the effects on the boundary layer on the
blades surface are not as profound.
From a practical perspective, the LES approach is probably
much more valuable for off-design conditions, for which it can offer unique insights into the complex physics of such pumps. The
purpose of the present study, however, is to demonstrate the feasibility/accuracy of the overall simulation strategy, and for this reason we only considered optimal operating conditions, where
comprehensive laboratory data were available. We will consider
off-design and compare with RANS/DES in a future paper.
Acknowledgments
The authors are grateful to Ing. A. Boccazzi and Ing. R. Miorini
for providing their experimental results. Computational resources
were provided by CASPUR (Consorzio interuniversitario per le
Applicazioni di Supercalcolo Per Universit e Ricerca). EB is partially supported by the National Science Foundation (Grant CBET0932613).
References
[1] Boccazzi A, Miorini R, Sala R, Marinoni F. Unsteady ow eld in a radial pump
vaned diffuser. In 8th European conference on turbomachinery (2327 March
2009). Graz, Austria.
[2] Byskov RK, Jacobsen CB, Pedersen N. Flow in a centrifugal pump impeller at
design and off-design conditions-part II: large eddy simulations. J Fluids Eng
2003;125(1):7383.
[3] Cheah KW, Lee TS, Winoto SH, Zhao ZM. Numerical ow simulation in a
centrifugal pump at design and off-design conditions. Int J Rotat Mach
2007;2007:83641-1-8.
[4] de Tullio MD, Cristallo A, Balaras E, Verzicco R. Direct numerical simulation of
the pulsatile ow through an aortic bileaet mechanical heart valve. J Fluid
Mech 2009;622:25990.
[5] Ducros F, Comte P, Lesieur M. Large-eddy simulation of transition to
turbulence in a boundary layer developing spatially over a at plate. J Fluid
Mech 1996;326:136.

43

[6] Fadlun EA, Verzicco R, Orlandi P, Mohd-Yusof J. Combined immersed-boundary


nite-difference methods for three-dimensional complex ow simulations. J
Comput Phys 2000;161(1):3560.
[7] Goto A. Study of internal ows in a mixed-ow pump impeller at various tip
clearances using three-dimensional viscous ow computations. J Turbomach
1992;114(2):37382.
[8] Grigoriadis DGE, Kassinos SC, Votyakov EV. Immersed boundary method for
the mhd ows of liquid metals. J Comput Phys 2009;228(3):90320.
[9] He L, Sato K. Numerical solution of incompressible unsteady ows in
turbomachinery. J Fluids Eng 2001;123(3):6805.
[10] Iaccarino G, Verzicco R. Immersed boundary technique for turbulent ow
simulations. Appl Mech Rev 2003;56(3):33147.
[11] Kato C, Kaiho M, Manabe A. An overset nite-element large-eddy simulation
method with applications to turbomachinery and aeroacoustics. J Appl Mech
2003;70(1):3243.
[12] Kato C, Mukai H, Manabe A. Large-eddy simulation of unsteady ow in a
mixed-ow pump. Int J Rotat Mach 2003;9:34551.
[13] Kaupert KA, Holbein P, Staubli T. A rst analysis of ow eld hysteresis in a
pump impeller. J Fluids Eng 1996;118(4):68591.
[14] Kiris CC, Kwak D, Chan W, Housman JA. High-delity simulations of
unsteady ow through turbopumps and owliners. Comput Fluids
2008;37(5):53646.
[15] Mittal R, Iaccarino G. Immersed boundary methods. Ann Rev Fluid Mech
2005;37(Jan):23961.
[16] Orlanski I. A simple boundary condition for unbounded hyperbolic ows. J
Comput Phys 1976;21(3):25169.
[17] Peskin CS. Flow patterns around heart valves: a numerical method. J Comput
Phys 1972;10(2):25271.
[18] Peskin CS, McQueen D. Modeling prosthetic heart-valves for numericalanalysis of blood-ow in the heart. J Comput Phys 1980;37(1):11332.
[19] Pouffary B, Fortes Patella R, Reboud J-L, Lambert P-A. Numerical simulation of
3D cavitating ows: analysis of cavitation head drop in turbomachinery. J
Fluids Eng 2008;130(6):061301-1061301-10.
[20] Shi F, Tsukamoto H. Numerical study of pressure uctuations caused by
impeller-diffuser interaction in a diffuser pump stage. J Fluids Eng
2001;123(3):46674.
[21] Smagorinsky J. General circulation experiments with the primitive equations.
Monthly Weather Rev 1963;91(3):99164.
[22] Swarztrauber PN. A direct method for the discrete solution of separable elliptic
equations. SIAM J Numer Anal 1974;11(6):113650.
[23] Takemura T, Goto A. Experimental and numerical study of three-dimensional
ows in a mixed-ow pump stage. J Turbomach 1996;118(3):55261.
[24] Tezduyar TE. Finite element methods for ow problems with moving
boundaries and interfaces. Arch Comput Methods Eng 2001;8(2):83130.
[25] Vanella M, Rabenold P, Balaras E. A direct-forcing embedded-boundary
method with adaptive mesh renement for uid-structure interaction
problems. J Comput Phys 2010;229(18):642749.
[26] Verzicco R, Fatica M, Iaccarino G, Orlandi P. Flow in an impeller-stirred tank
using an immersed-boundary method. AIChE J 2004;50(6):110918.
[27] Verzicco R, Orlandi P. A nite-difference scheme for three-dimensional
incompressible ows in cylindrical coordinates. J Comput Phys 1996;123(2):
40214.
[28] Yamanishi N, Fukao S, Qiao X, Kato C, Tsujimoto Y. LES simulation of
backow vortex structure at the inlet of an inducer. J Fluids Eng
2007;129(5):58794.
[29] Yang J, Balaras E. An embedded-boundary formulation for large-eddy
simulation of turbulent ows interacting with moving boundaries. J Comput
Phys 2006;215(1):1240.
[30] You D, Mittal R, Wang M, Moin P. Computational methodology for large-eddy
simulation of tip-clearance ows. AIAA J 2004;42(2):2719.
[31] You D, Wang M, Moin P, Mittal R. Large-eddy simulation analysis of
mechanisms for viscous losses in a turbomachinery tip-clearance ow. J
Fluid Mech 2007;586(Jan):177204.
[32] Zhang M-J, Pomfret MJ, Wong CM. Three-dimensional viscous ow simulation
in a backswept centrifugal impeller at the design point. Comput Fluids
1996;25(5):497507.
[33] Zhang M-J, Pomfret MJ, Wong CM. Performance prediction of a backswept
centrifugal impeller at off-design point conditions. Int J Numer Methods Fluids
1996;23(9):88395.

Das könnte Ihnen auch gefallen