Computers & Fluids 46 (2011) 212217

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

Computational uid dynamic analysis and design optimization of jet pumps

J. Fan a,1, J. Eves a,, H.M. Thompson a, V.V. Toropov a,b, N. Kapur a, D. Copley c, A. Mincher c
a
School of Mechanical Engineering, University of Leeds, UK
b
School of Civil Engineering, University of Leeds, UK
c
Parker-Hannin, Racor Filter Division Europe, Dewsbury, UK

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

Article history: Jet pumps have a wide variety of applications and are commonly used in thermal power plants and refrig-
Received 30 April 2010 eration systems. An initial jet-pump design was developed using an analytical approach and its efciency
Received in revised form 21 October 2010 was improved using an efcient and accurate computational uid dynamics model of the compressible
Accepted 31 October 2010
turbulent ow in the pump, whose predictions agreed well with corresponding experimental data. Para-
Available online 29 November 2010
metric studies were performed to determine the inuence of the pumps geometry on its performance
and the high delity CFD solutions were used to build surrogate models of the pumps behavior using
Keywords:
the moving least squares method. Global optimization was carried out using the surrogates. This
CFD
Optimization
approach resulted in pump efciency increasing from 29% to 33% and enabled the energy requirements
Jet pump of the pump to be reduced by over 20%.
Ejector pump 2010 Elsevier Ltd. All rights reserved.

1. Introduction
Supersonic jet pumps were originally developed in the 19th
century to maintain vacuum pressures in the condensers of steam
engines and are simple devices that can pump and compress a ow
without any moving parts [1]. A typical jet pump consists of two,
counter-facing, straight-sided cones separated by a parallel throat
section (Fig. 1). Within this section the primary ow, in the form of
a high velocity jet issuing from the nozzle, entrains (and mixes
with) the secondary ow, thereby generating the pumping behav-
ior. The second cone is divergent and acts to raise the static pres-
sure so that it equals that downstream of the pump, thus
minimizing exit losses. Research into jet pumps has seen a resur-
gence in recent years, since they offer an environmentally-friendly
component of refrigeration and air conditioning systems that can
be activated by low grade heat release from renewable energy
sources such as solar energy. A wide variety of jet pump applica-
tions exist, ranging from single-phase gas ows to two-phase ows
which employ liquid as the primary uid and gas as a secondary
uid. Examples of the latter arise in steam-driven jet pumps used
in thermal power plants and in refrigeration systems where nucle-
ation and growth of condensing droplets can be inuential [2].
Corresponding author. Tel.: +44 113 343 6360, mobile: +44 7717 327 202.
E-mail addresses: menjfa@leeds.ac.uk (J. Fan), j.s.v.eves@leeds.ac.uk (J. Eves),
h.m.thompson@leeds.ac.uk (H.M. Thompson), v.v.toropov@leeds.ac.uk (V.V. Toro-
pov), n.kapur@leeds.ac.uk (N. Kapur), dcopley@parker.com (D. Copley), adrian.
mincher@parker.com (A. Mincher).
1
Permanent address: School of Astronautics, Northwestern Polytechnical Univer-
sity, Xian, China.
This paper is concerned with the analysis and design of super-
sonic gas ows in jet pumps, and several mathematical models
have now been developed to analyze and design such ow sys-
tems. The majority are one-dimensional thermodynamic models
which obtain state and operating parameters along the jet pump.
Eames [1], for example, developed a useful analytical method for
jet-pump design where a constant rate of momentum change is
prescribed within the diffuser in order to prevent the occurrence
of thermodynamic shocks. However the complexity of the ow in
supersonic jet pumps, involving compressible ow, shock interac-
tions and turbulent mixing of two streams, has led to increasing
reliance on computational uid dynamics (CFD) as a design tool
for improving jet pump efciencies [3].
Previous CFD studies have highlighted the important role of
good grid resolution and an appropriate choice of turbulence mod-
el in any accurate CFD analysis of supersonic ow in jet pumps.
There is currently some debate about the best choice of turbulence
model for such ows, with the accuracy of a range of models, such
as the RNG ke and SST kx, having been analyzed [4]. Very recent
comparisons between CFD and experiments have shown that the
standard ke model is capable of predicting accurately the impor-
tant global performance indicators, such as the entrainment and
pressure lift ratios, and that the main difference between turbu-
lence models lies in their prediction of local ow structure [2].
One of the key design issues limiting the wider exploitation of
jet pump technology is the dramatic reduction of efciency and
pressure lift ratios that occur due to the onset of thermodynamic
shocks within the nozzle. Although the effects of various aspects
of the jet-pump geometry have been considered previously, there
is still some disagreement over important issues such as the

0045-7930/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compuid.2010.10.024
 J. Fan et al. / Computers & Fluids 46 (2011) 212217
Fig. 1. A typical jet-pump design.
optimal nozzle position, pump dimensions and their dependence
on operating conditions [5]. In this paper we combine high delity
CFD analysis and experimentation within a formal optimization
framework in order to improve jet pump efciency, in terms of
pressure lift and entrainment ratios, by optimizing jet pump geo-
metrical and operating parameters.
2. Problem formulation
The compressible ow within the jet pump is modeled using
the NavierStokes equations, based on the conservation of mass,
momentum and energy [6]. Flow within the jet pump is turbulent
and following several previous studies, e.g. [2], in the present work
the standard ke turbulence model [7] is adopted. In order to quan-
tify a jet pumps performance, the efciency can be dened as:
Powerout V_ ent pent
Efficiency
Powerin V_ pri ppri
where V_ is the volumetric ow rate and p is the static pressure of
the primary (pri) and entrained (ent) ows, respectively. This paper
focuses on the optimization of the efciency of practical jet-pump
designs.
2.1. CFD validation
An initial baseline jet-pump design (Fig. 2) was obtained using
the analytical method developed by Eames [1] where a constant
rate of momentum change is prescribed within the diffuser in or-
der to prevent the occurrence of thermodynamic shocks. This is
represented by the following equation:
dM 0 dc
mg 1 Rm b 2
dx dx
where M0 is the momentum of the ow, x is the axial position along
the diffuser, mg is the mass ow rate through the primary ow inlet,
Rm is the entrainment ratio (ms/mg), c is the velocity at x and b is the
constant rate of change of momentum. Inlet conditions for the
Fig. 2. Axisymmetric model of jet pump with geometrical variables.
213
diffuser are calculated by assuming uniform static pressure
throughout the entrainment region and an (arbitrary) low velocity
at the outlet. Pressure, temperature and Mach number distributions
can then be calculated based on one-dimensional compressible ow
theory from which a diffuser radius prole can be determined. This
should produce designs which avoid high pressure gradients and re-
duce pressure losses within the pump.
More detailed analysis of jet-pump behavior was based on a
series of higher delity CFD solutions of the compressible Na-
vierStokes equations governing the axisymmetric motion of the
gas ows within the jet-pump geometry using the uent CFD pack-
age [8]. Flow proles were compared with those from the analyti-
cal model.
Results from the initial CFD study showed the ow within the
pump to be signicantly different from the analytical calculations
(Fig. 3). This can be explained by the one-dimensional nature of
the analytical method, where it is assumed that the ow entering
the diffuser has a uniform Mach number. However as shown in
Fig. 4, this is not the case, rather the ow entering the diffuser con-
sists of a central supersonic jet surrounded by low speed entrained
ow. This also means that the effects of thermodynamic shocks
within the supersonic jet are not captured within the analytical
framework. Consequently, improvements in efciency throughout
the optimization of the geometry should be possible using a higher
delity approach than that available from the one-dimensional
model.
1 Due to the computational expense of each CFD solution, the rel-
ative merits of two techniques for solving the compressible ow
equations, namely the pressure based NavierStokes (PBNS) and
density based NavierStokes (DBNS) solvers, were compared.
Although the mesh independent solutions from either method
were found to be indistinguishable from one another the PBNS sol-
ver was found to require signicantly higher mesh densities in or-
der to resolve detailed ow features such as the chain of shocks
that occur in the supersonic jet ow (Fig. 4).
However, the overall solution time required by the algorithms
on the grid densities needed to achieve mesh independence are
much shorter for the PBNS approach than for the DBNS solver, as
shown in Table 1, and for this reason the CFD simulations used
in the optimization study carried out here were based on the PBNS
approach. A study of turbulence modeling was also carried out in
which the ke model variants and the Reynolds Stress Model
(RSM) were applied to the CFD model of the jet pump (Fig. 5). All
ke models were found to produce similar results. The RSM
generated slightly better predictions of vacuum pressure when
compared against experiments carried out on a prototype jet pump
at Parker-Hannin, but not enough to justify the additional
214 J. Fan et al. / Computers & Fluids 46 (2011) 212217

Fig. 3. Comparison between analytical and CFD predictions of Mach number along the jet pumps central axis.

Fig. 4. Contours of Mach number in jet pump showing thermodynamic shocks.

Table 1 between CFD results and experimental data was found to be gen-
Comparison of cell count required for mesh independency and CPU time for pressure erally good when operating at lower entrained ow rate of up
based and density based solution algorithms.
to 200 L/min the difference between theory and experiment is very
Cell CPU Vacuum pressure at Vacuum pressure at small. However at higher ow rates where the pump approaches
count hours 150 L/min (Pa) 250 L/min (Pa) its maximum possible performance, the accuracy of predictions
PBNS 71,000 6.5 11,794 9686 starts to decrease. At these critical operating conditions where
DBNS 47,000 8 11,778 9718 the pump is close to stalling its performance becomes highly
dependent on the ratio of the nozzle radius to diffuser radius. Inac-
curacies in the CFD predictions in these conditions can be ex-
computational cost. Therefore, the standard ke model was be used plained by small defects in the experimental model due to the
in all future studies due to its 10% reduction in computational cost rapid prototyping process. Some aspects of the prototype pump
with no loss in accuracy compared to the other two equation tur- geometry may vary from the CAD model by up to 3% and surface
bulence models. roughness is unavoidable during the rapid prototyping production
The CFD model resulting from these preliminary studies was process.
validated by comparison with experimental measurements carried In summary, the CFD model was shown to be capable of captur-
out for a range of nozzle pressures and entrained ow rates on a ing important ow features and making accurate predictions of
prototype jet-pump design at Parker-Hannin (Fig. 6). Agreement performance for all but the most extreme operating conditions.
 J. Fan et al. / Computers & Fluids 46 (2011) 212217 215

Fig. 5. The effect of turbulence models on CFD calculations of vacuum pressure achieved by a jet pump at various entrained ow rates.

Fig. 6. Comparison between CFD model and experimental data at three nozzle pressures.

2.2. Parametric study ential, offering up to 300% improvements in vacuum pressure

The jet-pump geometry (Fig. 2) was parameterized in terms of
eight key design variables. The diffuser prole is determined using
a hyperbolic tangent function in order to reduce the required num-
ber of variables.
rx A1 tanhA2 x  A3 A4
where r is the diffuser radius and x is the normalized axial coordi-
nate. A1 and A4 determine the inlet and outlet radii, and A2 and A3
describe the shape of the diffuser, allowing a wide range of designs
to be produced using only four design variables. A fth variable is
used for the diffuser length and a further two variables describe
the nozzle size. The eighth variable denes the length of the
entrainment zone.
CFD analysis was performed using a baseline mesh and geome-
try changes were made using an automated mesh morphing algo-
rithm. Flow variables from the CFD analysis of the baseline
geometry were then imported to the new geometry and used to
initialize the new CFD solution. A parametric study was carried
out for each design variable in order to identify those with the
greatest inuence on pump performance. The inlet radius of the
diffuser and the nozzle radius were found to be particularly inu-
(Fig. 7). However decreasing the diffuser radius or increasing the
nozzle radius too far caused the pump to stall as the entire ow
entering the diffuser is supersonic and the diffuser accelerates
the ow, leading to a sudden drop in efciency. Small changes in
operating conditions or manufacturing defects in pumps with
these highly optimized designs could result in a high degree of per-
3 formance variability. Neither the nozzle length nor the diffuser
length were found to be inuential, therefore the full optimization
study was carried out using only the remaining six design
variables.
2.3. Optimization
A surrogate modeling approach was used for the optimization
study since this is well suited to applications with small numbers
of design variables and expensive tness evaluations, as is the case
with the CFD simulations used here. An Optimal Latin Hypercube
(OLH) method was used for the design of experiments (DOE). This
was achieved using a multi-objective permutation genetic algo-
rithm [9], where the uniformity of build points, validation points
and combined DOEs can be optimized. The combined DOE con-
sisted of a 70 build point DOE and a 30 validation point DOE. High
216 J. Fan et al. / Computers & Fluids 46 (2011) 212217

Fig. 7. Inuence of diffuser inlet radius on vacuum pressure, showing potential performance improvements against current design.

delity CFD simulations were carried out at all design points using Table 2
the same approach as adopted in Section 2.2, with a 200 L/min Pump performance at stages of the design process.
boundary condition dened at the entrained ow inlet. The resul- Responses Vacuum Primary air Efciency (%)
tant performance data is shown in Fig. 8. from pressure (kPa) usage (m3/s)
Response surfaces of vacuum pressure and mass ow rate Original design CFD 14.61 8.28E4 29.4
through the primary ow inlet were built using two different mov- Best design from CFD 13.49 6.97E4 32.3
ing least squares (MLS) approaches available in Altairs HyperStudy DOE
software package [10]. Third order base polynomials were con- Optimized design MLS 13.33 6.52E4 34.1
after GA
structed at each design point in the DOE study by minimizing Optimized design MLS 13.25 6.53E4 33.9
the weighted sum of the squared errors, as described by (4). The after SQP
weighting factor, w, was calculated using the Gaussian decay func- CFD validation of CFD 12.90 6.49E4 32.9
tion in (5). Optimum
X
min wi kf xi  fi k2 4
i
The rst approach used the validation points to optimize the
closeness-of-t parameter in the MLS model, and the second ap-
wi exphr 2i 5
proach used a high closeness-of-t parameter to create an approx-
where f is the function value predicted by the response surface, fi imation that interpolated the model building points. The primary
are the function values observed from the CFD model, r is the dis- air usage response was shown to be easily predictable and almost
tance between design point i and the point at which the polynomial entirely dependent on only the nozzle radius. The vacuum pressure
is being constructed, and h is a closeness-of-t parameter. response was found to be more complex. Both MLS surfaces gave

Fig. 8. Vacuum pressure and ow rates for the 100 design points. Vertical line shows pressure constraint.
 J. Fan et al. / Computers & Fluids 46 (2011) 212217
Fig. 9. CFD predictions of pump efciency for original and optimized designs at a range of entrained ow rates.
equally good agreement with building and combined DOEs how-
ever there were differences in agreement with the validation
points (R2 values of 0.8934 and 0.8538 for optimized and high
closeness-of-t approaches respectively). Because of this the opti-
mized closeness-of-t approach was chosen for use in the optimi-
zation study.
The optimization problem was formulated as follows:
minimize : primary flow rate
subject to : vacuum pressure > 13:25 kPa
The required vacuum pressure of the pump was 12 kPa, how-
ever the use of a conservative lower limit on this response ensured
that designs avoided the critical region where pump efciency
drops. The equality constraint on entrained ow rate was satised
by dening it as a boundary condition in the CFD model. A genetic
algorithm (GA) was used to nd a global minimum with tness
evaluations carried out using the surrogate models. Sequential
Quadratic Programming (SQP) was then employed to ne tune
the design. The response surfaces were rened with an additional
50 CFD evaluations in a small proportion of the design space, cen-
tered around the optimum, with the allowable range of each design
variable reduced by a factor of 10. Optimization using a GA and
SQP was then repeated. The GA produced a design which, the sur-
rogate model predicted would reduce boost air usage, and there-
fore the power requirements of the pump, by 21%. The use of
SQP starting from this design point made no further improvements
(Table 2).
A CFD study was made of the optimized design and showed
good agreement with the surrogate models with a vacuum pres-
sure of 12.90 kPa and primary air usage of 6.49E4 m3/s at 200 L/
min, giving an efciency of 32.9%. Improvements were also made
at lower entrained ow rates when compared to the original de-
sign (see Fig. 9).
3. Conclusions
Jet pumps are important components of a wide variety of indus-
trial applications. High delity CFD models of the compressible
217
ow in jet pumps can be developed which agree well with exper-
imental data, offering signicant improvements over analytical
methods. The latters inability to represent important ow physics,
such as shock formation in the nozzle, preclude the use of a multi-
delity optimization strategy. Appropriate combinations of solu-
tion algorithms and turbulence model provide the computational
efciency needed to embed CFD modeling within a formal optimi-
zation framework, using surrogate models and a genetic algorithm,
which enable jet pump performance to be improved signicantly.
The CFD optimization study carried out here has enabled substan-
tial improvements in jet-pump design to be achieved: pump ef-
ciency has increased from 29% to 33% and, more signicantly, the
energy requirements of the pump have been reduced by over 20%.
Acknowledgement
The authors wish to thank the European Science Foundation for
funding the attendance of the ICFD 2010 Conference on Numerical
Methods for Fluid Dynamics, where this paper was rst presented.
References
[1] Eames IW. A new prescription for the design of supersonic jet-pumps: the
constant rate of momentum change method. Appl Therm Eng 2002;22:12131.
[2] Hemidi A, Henry F, Leclaire S, Seynhaeve J-M, Bartosiewicz Y. CFD analysis of a
supersonic air ejector: part 1: experimental validation of single-phase and
two-phase operation. Appl Therm Eng 2009;29:152331.
[3] He S, Li Y, Wang RZ. Progress of mathematical modeling of ejectors. Renew
Sustain Energy Rev 2009;13:176080.
[4] Bartosiewicz Y, Aidoun Z, Mercadier Y. Numerical assessment of ejector
operation for refrigeration applications based on CFD. Appl Therm Eng
2006;26(56):60412.
[5] Chunnamond K, Aphornatana A. Ejectors: applications in refrigeration
technology. Renew Sustain Energy Rev 2004;8:12955.
[6] Versteeg HK, Malalasekera W. An introduction to computational uid
dynamics: the nite volume method. 1st ed. Pearson Prentice Hall; 1995.
[7] Launder BE, Spalding DB. The numerical computation of turbulent ows.
Comput Methods Appl Mech Eng 1974;3:26989.
[8] Fluent 6.3. <http://www.ansys.com/products/uid-dynamics/uent>.
[9] Narayanan A, Toropov V, Wood AS, Campean IF. Simultaneous model building
and validation with uniform designs of experiments. Eng Optim
2007;39:497512.
[10] HyperWorks 10. <http://www.altairhyperworks.co.uk>.