Journal of Chemical Technology and Biotechnology J Chem Technol Biotechnol 78:308–313 (online: 2003)

DOI: 10.1002/jctb.776

Multi-objective optimization of venturi

scrubbers using a three-dimensional model
for collection efficiency†
Gopalan Ravi, Shekar Viswanathan, Santosh K Gupta‡ and Madhumita B Ray*
Department of Chemical and Environmental Engineering, National University of Singapore, 4 Engineering Drive 4, Singapore 117576

Abstract: Multi-objective optimization of a venturi scrubber was carried out using a three-dimensional
model for collection efficiency and non-dominated sorting genetic algorithm (NSGA). Two objective
functions, namely (a) maximization of the overall collection efficiency, and (b) minimization of the
pressure drop were used in this study. Three decision variables including two operating parameters,
viz liquid–gas ratio and gas velocity in the throat, and the nozzle configuration, which takes into
account the three-dimensional nature of the problem, were used in the optimization. Optimal design
curves (non-dominated Pareto sets) and the values of the decision variables corresponding to optimum
conditions on the Pareto set for a pilot-scale scrubber were obtained. The liquid to gas (L/G) ratio,
which is a key decision variable that determines the uniformity of liquid distribution, and a staggered
nozzle configuration can produce uniform liquid distribution in the scrubber. Multiple penetration
using nozzles of two different sizes in a triangular staggered arrangement can reduce liquid loading by
as much as 50%, consequently reducing the pressure drop in the scrubber.
# 2003 Society of Chemical Industry

Keywords: venturi scrubber; collection efficiency; pressure drop; multi-objective optimization; pareto sets;
genetic algorithm

NOTATION Superscript
Cd, Cp Concentration of the droplets and particles l, u Lower and upper limit
(no m
3)
Ed Eddy diffusivity of the droplets (m2 s
1)
I Objective functions for the GA
Lth Length of venture scrubber, m INTRODUCTION
Nc Nozzle configuration Venturi scrubbers have been used extensively since the
Qd Liquid drop source strength (no m
3 -s
1) 1950s as a major gas-cleaning device for the removal of
Qf Amount of liquid flowing as film on the wall fine particulates from industrial exhausts. Although
(no m
3 -s
1) there has been significant progress in developing
R0 Half-width of the venturi throat parallel to models for the collection efficiency and pressure drop,
water injection, mm not much work has been reported on the optimization
t Time (s) of venturi scrubbers. In the pioneering work on
u Decision variable (dimensionless) optimization, Goel and Hollands1 used the model of
Vdx Velocity of the droplets (m s
1) Calvert2 for the determination of the pressure drop
VG,th Gas velocity at the throat (m s)
1 and the collection efficiency and developed optimized
W0 Width of the venturi throat perpendicular to collection efficiency and pressure drop charts. Leith
water injection, mm and Cooper3 proposed an optimization algorithm
based on Calvert’s model, which uses a straight duct
Zo Overall collection efficiency approach, whereas Cooper and Leith4 presented an
(dimensionless) improved approach wherein the scrubber geometry
was considered. All these early approaches attempted
Subscript to optimize the performance of a given scrubber using
i Index of decision variable two important operating variables, viz the ratio, L/G, of

* Correspondence to: Madhumita B Ray, Department of Chemical and Environmental Engineering, National University of Singapore, 4
Engineering Drive 4, Singapore 117576
E-mail: cheraym@nus.edu.sg
†
Paper presented at the Process Innovation and Process Intensification Conference, 8–13 September 2002, Edinburgh, UK
‡
Current address: Dept of Chemical Engineering, IIT Kanpur, India
(Received 28 February 2002; accepted 29 September 2002)

# 2003 Society of Chemical Industry. J Chem Technol Biotechnol 0268–2575/2003/$30.00 308

 Optimization of venturi scrubbers

the liquid-to-gas flow rates, and the gas velocity, VG,th, The model takes into account the jet penetration
in the throat. Several important design variables, such length, the initial momentum of the liquid and the
as the throat dimensions, nozzle diameter, and the non-uniformity in the drop-size distribution. The
arrangement of the liquid injection nozzles, that can model assumes that the movement of the drops in
affect the liquid distribution significantly, were not the axial (x in Fig 1) and lateral (y) directions is due to
considered. In addition, these optimization ap- convection and diffusion, respectively. It also assumes
proaches used simplified models that assumed uni- drops of uniform size, constant film flow, no drop–
form liquid distribution without accounting for film drop interactions, uniform inlet distribution of the
flow. Hence, the above approaches did not provide particles, no interaction between the particles, and that
exact optimized solutions and are scrubber-specific. particulate collection by the droplets is mainly due to
An improved method for the multi-objective optimiza- the inertial impaction of the dust particles onto the
tion of Pease–Anthony type of scrubbers was pre- droplets. An empirical correlation has been used to
sented recently5 using the recent and more detailed estimate the fraction of liquid flowing on the walls.
models developed by Viswanathan and co-workers6,7 The pressure drop, Dp, in the Pease–Anthony
for the collection efficiency and the pressure drop, scrubber is determined using the annular flow model
respectively. However, this study was also somewhat developed by Viswanathan et al. 7 Pressure losses that
limited in scope since it optimized only existing occur due to the acceleration of the gas and of the
(specified) scrubbers and did not incorporate any liquid drops and due to frictional losses are estimated.
design variables. The total pressure drop is calculated by also taking
A detailed optimization study must obtain the best into account the pressure recovery that occurs in the
conditions for both the design and the operating diffuser. The overall collection efficiency and the
variables. In the present study, an important design pressure drop equations are available in Refs 5–7.
variable, viz the optimum nozzle arrangement in a The majority of the collection of particulates occurs
Pease–Anthony scrubber, is included (along with some in the throat because of the presence of a high degree
important operating variables) among the ‘control’ of turbulence in this region caused by large relative
(also called decision) variables used for optimization. velocities between the drops and the particles. A
A three-dimensional model is used to calculate the schematic representation of the throat of the scrubber,
collection efficiency of the scrubber. The optimization sub-divided into a three-dimensional grid, is presented
was conducted using genetic algorithm (GA), a non- in Fig 2. Use of such a three-dimensional approach in
traditional search and optimization method, intro-
duced by Holland.8 GA is known to have several
advantages over conventional optimization techniques
such as objective functions can be multimodal or
discontinuous; information is required only on the
objective function, (and no gradients need be com-
puted); a starting (guess) solution is not needed; the
search is carried out using a population of several
(rather than a single) points simultaneously. In addi-
tion, GA is well suited to handle problems involving
several design or operating decision variables. Non-
dominated sorting genetic algorithm9 (NSGA) is used
to solve a variety of complex problems of industrial
interest involving a vector of several objective func-
tions. In this study, NSGA is used with the two
objective functions, namely, maximizing the overall
collection efficiency and minimizing the pressure drop
in a pilot-scale venturi scrubber.

MODELS FOR COLLECTION EFFICIENCY AND

PRESSURE DROP
The performance of a venturi scrubber (Fig 1)
depends largely on the manner of liquid injection,
including the initial liquid momentum, since this
essentially determines the size-distribution of the
droplets of liquid formed and the liquid-flux distri-
bution. Ananthanarayanan and Viswanathan10 used a
simplified version of the model proposed earlier by
Viswanathan,6 and this model is used in this work to Figure 1. Schematic diagram of the pilot-scale venturi scrubber (all
calculate the droplet flux distribution in the scrubber. dimensions are in cm).

J Chem Technol Biotechnol 78:308–313 (online: 2003) 309

G Ravi et al

the end of the scrubber is then computed as:

R
Cpðx;yÞ dy

o ¼ 1
R ð2Þ
Cpð0;yÞ dy

FORMULATION
The two objective functions chosen in this study, are
the:

(a) maximization of I1, the overall collection effi-

ciency, Zo, and
(b) minimization of I2, the pressure drop, Dp.

Three decision variables are identified and their

bounds were fixed depending on practical considera-
tions. These are: the liquid-to-gas flow ratio, L/G; the
gas velocity at the throat, VG,th; and a (dimensionless)
parameter, Nc, that describes the geometrical place-
Figure 2. Grid lay-out for optimization.
ment of the nozzles at the beginning of the throat.
Figure 3 shows five types of nozzle arrangements, of
the present study enables one to obtain the optimum nine nozzles in each case, in the region, 0 4 y 4 Ro;
nozzle arrangement for venturi scrubbers, a very 0 4 z 4 Wo/2 (there are 17 nozzles on each side of the
important design variable that dictates the flux throat, at the grid points). The central nozzle in Fig 3 is
distribution in this equipment. Due to the symmetry shared by the symmetric other quarter on that side of
in the problem, only a part of the scrubber-throat, the throat. In the first nozzle arrangement, also called
0 4 x 4 Lth; 0 4 y 4 Ro; 0 4 z 4 Wo/2 (see Figs 1 and the single file, nine nozzles are arranged side-by-side in
2), is required for simulation. This reduces the the first row, one at each grid-point. The second
complexity as well as the computational time required configuration is that of a triangular pitch where the
for simulation. The droplets are introduced in the nozzles in the first row are arranged on alternate grid-
scrubber using a total of 34 nozzles placed at the points and those in the second row are staggered with
entrance of the throat (17 on either side of the throat). those in the first. The third type of nozzle configura-
The physical space inside the throat is divided into tion is that of a double file (square). The nozzles in the
cells as shown in Fig 2 (a fixed finite-difference fourth type of nozzle arrangement are arranged in a
Eulerian grid; 9 grid points used in 0 4 z 4 Wo/2, 13 triple file. All the nozzles in these four configurations
grid points used in 0 4 x 4 Lth, 9 grid points used in have the same diameter, viz 2.1 mm, and hence take up
0 4 y 4 Ro) and the Lagrangian mass particles carry the same projected area available for liquid injection.
the fluid from cell to cell by the sum of bulk and
turbulent velocities. In order to evaluate the move-
ment of each mass particle, the bulk velocity, eddy
diffusivity, gas stream drag and the initial liquid
momentum are calculated. The jet penetration length
is determined and subsequently, the flux distribution
at any axial position is obtained by applying a central
difference formula to solve the three dimensional
steady-state continuity equation for the liquid drops as
shown in eqn (1):



@ @ @Cd @ @Cd
ðVdx Cd Þ ¼ Ed þ Ed þ Qd
Qf
@x @y @y @z @z
ð1Þ

This results in a sparse matrix, which is then solved by

the use of appropriate subroutines obtained from the
IMSL library. Particulate matter is introduced in the
scrubber as uniformly distributed dust particles
moving with the same velocity as the gas stream. The
particle distribution at any axial position is then
determined in a way similar to the liquid-flux
distribution. The overall collection efficiency, Zo, at Figure 3. Schematic representation of nozzle configurations studied.

310 J Chem Technol Biotechnol 78:308–313 (online: 2003)

 Optimization of venturi scrubbers

When the nozzles are of the same diameter, the liquid required for optimality and provided considerable
from each of the nozzles penetrates into the same insight into the physical phenomena important under
extent into the scrubber. This phenomenon is termed optimal conditions.
as uniform penetration. The fifth configuration in Fig 3 The values of several of the computational par-
is a staggered arrangement, similar to the triangular file ameters used in this study, as well as those of the other
(the second configuration), but with the nozzle variables and parameters describing the venturi
diameters in the first and second rows being different. scrubber, are listed in Table 1. The results of
The diameters of the nozzles in the first and second optimization are shown in Fig 4 (solid points for the
rows are 1 and 1.5 mm, respectively. Thus, in 3-D case). A plot of Zo versus Dp is shown in Fig 4(a) (a
configuration 5, the penetration lengths are different typical Pareto set, wherein an improvement (increase)
in the first and second rows, and the phenomenon is in Zo is accompanied with a worsening (increase) of
called multiple penetration. Dp). Plots of the three decision variables correspond-
The multi-objective function optimization problem ing to the different points on the Pareto set are shown
studied here is, thus, described mathematically by in Fig 4(b–d). It is observed that the gas velocity at the

 throat, VG,th, varies along the points on the Pareto.
L
Max I1 ðuÞ  I1 ; VG;th ; Nc ¼
o ð3aÞ The values of L/G that provide optimal operating
G conditions are found to be almost constant, varying in

 a narrow range of about 0.8  10
3–1.1  10
3 (Fig
L
Min I2 ðuÞ  I2 ; VG;th ; Nc ¼ Dp ð3bÞ 4(c)), similar to what was found in our previous
G
study.5 The optimum nozzle configuration, Nc (Fig
subject to: 4(d)), is configuration 2 wherein the nozzles are
arranged in a staggered triangular file.
uli 4 ui 4 uui ; i ¼ 1; 2; 3 ð3cÞ A comparison has been made with our previous
study,5 which used a two-dimensional model for the
and model equations in Refs 5-7 ð3dÞ
collection efficiency. Both the studies retain the basic
In this problem, dust of a uniform size (Dp = 5 mm) is characteristics of the Pareto curve but the values are
used in the feed to the venturi scrubber. The optimal slightly different because of the difference in the rigor
value for L/G obtained by Ananthanarayanan and involved in the models used. Higher collection
Viswanathan10 (using only a single decision variable at efficiency at a lower pressure drop in the three-
a time) was 1.2  10
3. Accordingly, a range of dimensional case than that in the two-dimensional
0.3  10
3–1.4  10
3 m3 of liquid/m3 of air has been case may be due to better liquid distribution in the
selected as the bounds for L/G. Bounds (40–120 m scrubber due to the optimal nozzle arrangement
s
1) for the throat gas velocity, VG,th, were decided (Nc = 2) being used here, something that was not
based on industrial practice. All other dimensions of possible in our earlier study. The influence and
the venturi scrubber are the same as used by importance of adding on the new design variable is
Ananthanarayanan and Viswanathan,10 and are given clearly observed. Figure 4 suggests that when all the
in Fig 1. three decision variables are used simultaneously, the
The decision variables that lead to an increase in the Pareto is far more sensitive to VG,th than to L/G, and
overall collection efficiency in a venturi scrubber may that the scatter in the optimal values of L/G in Fig 4(c)
produce a simultaneous undesirable increase in the is essentially a reflection of its relative insensitivity. An
pressure drop. Similarly, decision variables that lead to increase in VG,th at a nearly constant range of L/G leads
a minimization of the pressure drop may lead to lower
collection efficiencies in the scrubber. Thus, the two
objective functions in eqn (3) are conflicting in nature Table 1. Parameters used in the optimization

and would lead, most likely, to a Pareto set of optimal Parameter Value
solutions.
Computational
Maximum number of generations (maxgen) 100
Population size (Np) 100
Probability of crossover (pc) 0.55
RESULTS AND DISCUSSION Probability of mutation (pm) 0.001
The multi-objective optimization problem was solved Random seed 0.87619
on a CRAY J916 computer. The CPU time required to Spreading parameter (s) 0.005
solve this problem was 1.35 s. In solving the problem, Sharing function (a) 2
the domains (bounds) of the several decision variables Grid size (h) (m) 0.004
were varied and the Pareto sets obtained. Thereafter, a Model a
few of the optimal points on the Pareto set were Density of liquid (kg m
3) 993.0
selected (such that they spanned the entire range of the Density of particles (kg m
3) 2500
Pareto set) and simulation results for these cases were Viscosity of gas (Pa s) 1.8  10
5
Viscosity of liquid (Pa s) 1.0  10
3
analysed (results not shown here due to the scope of
the paper). This permitted analysis of the conditions a
References 5–7.

J Chem Technol Biotechnol 78:308–313 (online: 2003) 311

G Ravi et al

Figure 4. Pareto optimal solution showing

pressure drop versus efficiency. (a) DP
versus Z; (b) gas velocity versus Z;
(c) Liquid/Gas ratio versus Z; (d) nozzle
configuration versus Z. *, optimal solution
for 3-D case ~, optimal solution for 2-D
case.

to higher values of Zo because of the formation of a good spreading of the liquid. As the liquid flows down
higher number of smaller droplets. the scrubber, driven by the momentum of the gas flow,
A slightly different case of the multiobjective an even greater spread of liquid droplets occurs due to
optimization problem discussed above, namely, the lateral diffusion.
comparison between uniform and multiple penetra- The optimal L/G ratio in multiple penetration is
tions, is attempted next. In this example, the best located in a narrow range of 0.4  10
3–0.5  10
3,
nozzle configuration from the previous uniform compared with 1.0  10
3 for uniform penetration.
penetration studies (Nc = 2, corresponding to the The lower values of the optimal L/G suggest consider-
triangular file arrangement) is compared with the case able savings in the requirement of the scrubbing liquid.
with Nc = 5, wherein nozzles of two sizes are used. The Such savings are consistent with the recent findings11
optimal Paretos are shown in Fig 5. A comparison of of scrubber optimization using computational fluid
Figs 4 and 5 shows that multiple penetration (con- dynamics (CFD) to model the flow in the scrubber.
figuration No 5 in Fig 3) gives superior solutions
(higher Zo for the same Dp on the Pareto set),
indicating a more uniform liquid distribution in the CONCLUSIONS
scrubber. In the multiple penetration configuration, Multi-objective optimization of a venturi scrubber was
the liquid penetrates to different extents in the first two carried out using a three-dimensional model for the
rows along the axis of the scrubber. As a result, there is collection efficiency with the NSGA algorithm. Simul-

Figure 5. Optimal solutions for multiple

penetration nozzle configuration
(Nc = 5). (a) DP versus Z; (b) gas
velocity versus Z; (c) liquid/gas ratio
versus Z; (d) nozzle configuration
versus Z.

312 J Chem Technol Biotechnol 78:308–313 (online: 2003)

 Optimization of venturi scrubbers

taneous optimization of collection efficiency and REFERENCES

pressure drop in the scrubber was conducted using 1 Goel KC and Hollands KGT, Optimum design of venturi
scrubbers. Atmos Environ 11:837–845 (1977).
three decision variables. Pareto optimal sets were
2 Calvert S, Venturi and other atomizing scrubbers. A IChE J
obtained. The variables include two operating vari- 16:392–396 (1970).
ables (viz L/G and VG,th) and one design variable, 3 Leith D and Cooper DW, Venturi scrubber optimization. Atmos
namely, the nozzle configuration. The ratio of liquid to Environ 14:657–664 (1980).
gas flow (L/G) determines the uniformity of liquid 4 Cooper DW and Leith D, Venturi scrubber optimization
revisited. Aerosol Sci Technol 3:63–70 (1984).
distribution inside the scrubber. An optimal range is
5 Ravi G, Gupta SK, Viswanathan S and Ray MB, Optimization of
obtained for this ratio. VG,th is found to vary along the venturi scrubbers using genetic algorithm. Ind Eng Chem Res
Pareto and contributes to the shape of the Pareto 41:2988–3002 (2002).
design curve. The optimization results (for uniform 6 Viswanathan S, Modeling of venturi scrubber performance. Ind
penetration configurations) indicate that a staggered Eng Chem Res 36:4308–4317 (1997).
7 Viswanathan S, Gnyp AW and St Pierre CC, Annular flow
nozzle configuration, corresponding to Nc = 2, pro-
pressure drop model for Pease–Anthony-type venturi
duces relatively uniform liquid distribution in the scrubbers. A IChE J 31:1947–1958 (1985).
scrubber and leads to improved performance. Multiple 8 Holland JH, Adaptation in Natural and Artificial Systems, Uni-
penetration using nozzles of two different sizes in a versity of Michigan Press, Ann Arbor, MI (1975).
triangular staggered arrangement, Nc = 5, can cause an 9 Srinivas N and Deb K, Multiobjective function optimization
using nondominated sorting genetic algorithms. Evol Comput
almost 50% reduction in the liquid loading, conse-
2:221–248 (1995).
quently reducing the pressure drop in the scrubber. 10 Ananthanarayanan NV and Viswanathan S, Estimating maxi-
The present optimization study, using a three-dimen- mum removal efficiency in venturi scrubbers. A IChE J
sional simulation model, produces the interesting 44:2549–2560 (1998).
information that the nozzle arrangement is an im- 11 Ananthanarayanan NV and Viswanathan S, Effect of nozzle
arrangement on venturi scrubber performance. Ind Eng Chem
portant decision variable in venturi scrubbers. The
Res 38:4889–4900 (1999).
Pareto solutions obtained in this work are essentially
design curves that assist in narrowing down the choices
of a decision-maker.

J Chem Technol Biotechnol 78:308–313 (online: 2003) 313