Engineering Applications of Computational Fluid

Mechanics

ISSN: 1994-2060 (Print) 1997-003X (Online) Journal homepage: https://www.tandfonline.com/loi/tcfm20

Multi-objective optimization and comparison of

surrogate models for separation performances
of cyclone separator based on CFD, RSM, GMDH-
neural network, back propagation-ANN and
genetic algorithm

Donggeun Park, Jemyung Cha, Moonjeong Kim & Jeung Sang Go

To cite this article: Donggeun Park, Jemyung Cha, Moonjeong Kim & Jeung Sang Go (2020)
Multi-objective optimization and comparison of surrogate models for separation performances
of cyclone separator based on CFD, RSM, GMDH-neural network, back propagation-ANN and
genetic algorithm, Engineering Applications of Computational Fluid Mechanics, 14:1, 180-201, DOI:
10.1080/19942060.2019.1691054

To link to this article: https://doi.org/10.1080/19942060.2019.1691054

© 2019 The Author(s). Published by Informa Published online: 09 Dec 2019.

UK Limited, trading as Taylor & Francis
Group

Submit your article to this journal Article views: 7

View related articles View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tcfm20
ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS
2019, VOL. 14, NO. 1, 180–201
https://doi.org/10.1080/19942060.2019.1691054

Multi-objective optimization and comparison of surrogate models for separation

performances of cyclone separator based on CFD, RSM, GMDH-neural network,
back propagation-ANN and genetic algorithm
Donggeun Park a , Jemyung Chab , Moonjeong Kimb and Jeung Sang Goa,b
a Department of Advanced Materials and Parts of Transportation Systems, Pusan National University, Busan, Rep. of Korea; b School of
Mechanical Engineering, Pusan National University, Busan, Rep. of Korea

ABSTRACT ARTICLE HISTORY

Pressure drop (p) and collection efficiency (η) are used to evaluate the separation performance Received 16 July 2019
of the cyclone separator. In this study, we conducted comparative study of cyclone models using Accepted 1 November 2019
response surface methodology (RSM), back propagation neural network (BPNN), and group method KEYWORDS
of data handling (GMDH) networks to develop optimal predictive cyclone models. Also, we con- Cyclone separator;
ducted multi-objective optimization for maximizing model and minimizing model using genetic Computational Fluid
algorithm (GA). CFD was performed instead of experimental method to get the estimated values Dynamics (CFD); Response
for modeling of p and η. The validation results of CFD showed 0.5% and 2% errors for p and η, Surface Methodology (RSM);
respectively, compared with the experimental data. Second, design of experiment (DOE) analysis Artificial Neural Network
for 10 cyclone geometrical parameters was executed to obtain the significant geometrical parame- (ANN); Genetic Algorithm
ters. Vortex finder diameter Dx , inlet width a, inlet height b and cone height Hco have a significant (GA); Multi-objective
optimization
effect on η and p. However, interaction effects between the geometrical parameters have small
effects. The cyclone models by RSM, BPNN and GMDH based on 25 CFD training set were devel-
oped. The predictive performance results by the cyclone models were compared by 25 CFD test set.
The GMDH method achieved the best prediction for p (R2 = 99.7, RMSE = 0.102) R2adjusted = 98.99,
RMSE = 0.0119) than the RSM, BPNN cyclone models. Additionally, uncertainty analysis was per-
formed to estimate the quantitative performance of cyclone models. The results show that the
uncertainty width of GMDH models achieved the best prediction (η: ±0.0065, p: ±0.0188). Finally,
GA was applied to optimize the GMDH models simultaneously. GA generated 70 non-dominant solu-
tions. Reproducibility of five optimal points was validated by using CFD. The trade-off optimal point
showed improvement by 24.31%, 18% and 8.79% for p d50 and η, respectively.

1. Introduction is the dimensionless measure of the pressure drop. The

performance parameters are defined as follow;
1.1. Research background
Nout (d)
In many industries, the contaminants are removed using η =1− (1)
Nin (d)
various dust collectors. In comparison with other type
of dust collectors, such as settling chamber, wet scrub- Where Nin is the total number of particles, and Nout is the
bers and electrostatic precipitators etc., gas–solid cyclone total number of particles discharged in the cyclone vortex
separator using a centrifugal force were widely used in finder.
industrial processes due to the advantages of its rela-
P = Pin − Pout (2)
tively low maintenance cost, simple structure and robust-
ness of operating conditions under high temperature and Where Pin is the pressure in the cyclone inlet, and
pressure (Hoffmann & Stein, 2008). Pressure drop ‘p’ Pout is the pressure in the cyclone outlet. The perfor-
(also expressed as Euler number) and collection effi- mance parameters are closely dependent of the geometric
ciency ‘η’ (also known as the separation efficiency and parameters of cyclone. The configuration and nomencla-
cut-off diameter) are used to evaluate the separation per- ture of 10 geometric parameters of cyclone is shown in
formance of the cyclone separator. The cut-off diameter is Figure 1. Ten parameters are diameter of vortex finder
a particle diameter for which η is 50% and Euler number Dx , inlet heighta, inlet width b, cylinder height Hcy , cone

CONTACT Jeung Sang Go [email protected]

© 2019 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 181

and experimental data. Gimbun, Chuah, Choong, and

Fakhru’l-Razi (2005) showed by comparing the CFD
results with the cyclone performances predicted from the
various empirical models.
With the advance of high-performance computers
and CFD technology, CFD has been considered as
another approach to optimize p and η recently. CFD
have advantages of time saving and high accuracy as
well as the ability to calculate physical quantities that
are difficult to be measured experimentally (Elsayed
& Lacor, 2012). Therefore, many CFD studies have
been performed for optimizing p and η depending
on the geometrical parameters. The effects of pres-
sure drop and collection efficiency on the changes of
the cyclone vortex finder shape were investigated using
CFD (De Souza, Salvo, & Martins, 2015; Raoufi, Shams,
Farzaneh, & Ebrahimi, 2008). The CFD simulation of
cyclone flow was performed according to changing the
cyclone inlet shape (Bernardo, Mori, Peres, & Dionísio,
2006; Elsayed & Lacor, 2011; Yang, Sun, & Gao, 2013).
The cyclone performance parameters were investigated
according to changing the cyclone body shape (Brar,
Sharma, & Elsayed, 2015; Hamdy, Bassily, El-Batsh, &
Mekhail, 2017; Safikhani, Akhavan-Behabadi, Shams, &
Rahimyan, 2010). The cyclone flow was investigated by
changing the cyclone dust collector shape (Ganegama
Bogodage & Leung, 2015; Kaya & Karagoz, 2009; Qian,
Zhang, & Zhang, 2006). These studies performed by
changing only one geometrical parameter and others
were fixed. This did not consider the interactive effect
among the geometric parameters. This brings about local
optimization.
Recently, ANN techniques and GA method has been
successfully applied to many industries (Ardabili et al.,
Figure 1. Geometry and nomenclature of 10 geometric parame- 2018; Ebtehaj & Bonakdari, 2016; Fotovatikhah et al.,
ters of cyclone. 2018; Gholami et al., 2018; Moazenzadeh, Mohammadi,
Shamshirband, & Chau, 2018; Najafi, Faizollahzadeh
Ardabili, Shamshirband, Chau, & Rabczuk, 2018; Taherei
height Hco , cone tip diameter Bc , tube height Ht , collec- Ghazvinei et al., 2018). Also, cyclone optimization studies
tor height Hc and collector diameter Dc , vortex finder have been performed by combing Design of Experiment
length S. (DOE) with various modeling methods such as Response
Generally, as p increases, η also increases. However, Surface Methodology (RSM) and Artificial Neural Net-
the optimal cyclone should be designed to have mini- work (ANN) methods. A number of optimization studies
mum p and maximum η. Therefore, many studies have are relatively limited, compared to the experiment and
been reported for several decades to model and optimize simulation studies (Sun & Yoon, 2018).
the performance parameters according to the geometric Elsayed and Lacor (2012) conducted a multi-objective
parameters of cyclone. optimization study for seven geometrical parameters
Early researches developed empirical models for p using non-dominated sorting genetic algorithm (NSGA-
and η according to massive experiments and mathe- II). They obtained the radial basis function neural net-
matical equations (Avci & Karagoz, 2001; Barth, 1956; work (RBFNN) models for Euler number (p) and
Dietz, 1981; Leith, 1990; Muschelknautz & Kram- cut-off diameter (η) based on results of the empirical
brock, 1970; Shephered & Lapple, 2005). The empiri- model. The empirical models can yield the estimated val-
cal model was developed by conservation law’s equation ues faster than using CFD. However it is not suitable
182 D. PARK ET AL.

for cyclonic optimization because the predictive perfor- required for surrogate modeling, as well as to solve
mance of empirical model is less accurate for complex local optimizations that occur when only one design
geometry. variable is changed. Previous work of cyclone using
Sun, Kim, Yang, Kim, and Yoon (2017) conducted CFD overlooks better optimization performance
multi-objective optimization for p and η based on CFD, due to local optimization problems (Elsayed & Lacor,
the desirability function method and response surface 2011; Ganegama Bogodage & Leung, 2015; Raoufi
method (RSM). As optimization results, p and η were et al., 2008).
improved by 20.7% and 24.2%, respectively, compared to (2) Development and Comparison of predictive models
the reference model. Also, Sun and Yoon (2018) added based on RSM, GMDH technique and BPNN tech-
the cyclone surface roughness as a new design variable nique for predicting cyclone performance parameter
and increased the level range of the existing design vari- with respect to the cyclone geometric parameters:
ables and optimized it by using NSGA-II. However, it was Each cyclone performance prediction model is com-
found that cited reference values of p and η are signif- pared quantitatively using AIC, R2, RMSE, and
icantly different at the optimization study using NSGA- R2_adj, and uncertainty analysis is performed on
II. Therefore, a reasonable comparison of optimization the developed models. To the best of our knowl-
results cannot be taken ahead. edge, there are no studies comparing simultaneously
Khalkhali and Safikhani (2012) developed the mod- the predictive performance of p and η models
els of p and η according to the four cyclone geometric obtained by using the various ANN methodologies
parameters based on CFD and GMDH neural network. and the RSM using the same training dataset and test
The developed cyclone models were optimized simul- dataset in many cyclone studies.
taneously using NSGA-II. Multi-objective optimization (3) Multi-objective optimization of the cyclone mod-
was executed based on the NSGA-II. However, the repro- els that have the best predictive performance using
ducibility of the optimized results was not confirmed NSGA-II: The optimized results compare not only
by CFD simulations. This cannot confirm the proper the improvement of the cyclone performance mod-
optimization results. els (p, η) with the reference model, but also ana-
Elsayed and Lacor (2013) optimized the four geomet- lyze from the viewpoint of physical justification of
ric parameters of the Stairmand design cyclone using the optimization result. The limitations of previ-
CFD, RSM and desirability functions. The new design has ous optimization researches lack an analysis of why
achieved high performance. However, the errors between the optimization was made (Elsayed & Lacor, 2013;
RSM results and CFD results for p and η were 17.20% Khalkhali & Safikhani, 2012).
and 20.23%, respectively. In addition, a cyclone perfor-
mance models were developed using RBFNN as an alter-
native to RSM. However, the predicted performance of 1.3. Reference model: Muschelknautz method of
the developed RBFNN model was evaluated quantita- modeling
tively based on the training data set. That approach might
For this optimization study, Muschelknautz method
result in over fitting. A different test set should have been
(MM) of modeling (Muschelknautz & Krambrock, 1970)
used to evaluate the prediction performance of model
was selected as a reference model because it contains the
instead of the training set that used to train the model.
shape of the cyclone dust collector as a design param-
In addition, this study did not verify the physical justifi-
eter, compared to the other cyclone mathematic model,
cation for optimized reasons. It cannot explain why the
and many experimental results (Obermair & Staudinger,
optimized model outperforms than the existing baseline
2001; Obermair, Woisetschläger, & Staudinger, 2003) are
model.
available for CFD validation. The reference model will be
compared to the optimized cyclone model. The 10 geo-
1.2. Study objectives metric parameters and geometric ratio of the MM model
presents in Table 1.
The objectives of present study are summarized in
three:
1.4. Study outline
(1) Analysis of effect of significant geometrical parame-
ters out of 10 geometrical parameters on the cyclone The flow chart relationship in this study is shown in
performance parameters using DOE analysis with Figure 2. This optimization study followed the follow-
CFD results: The significant geometrical parame- ing steps. First, CFD simulations are performed on the
ters assist to reduce the number of DOE data sets cyclone flow using the turbulence models and the DPM.
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 183

Table 1. Geometrical parameters values for Obermair et al. (2003). (D = 0.4 m).
Dx a b Hy Hco Bc Ht Hc Dc S
Factors (x1) (x2) (x3) (x1) (x5) (x6) (x7) (x8) (x9) (x10)
D D D D D D D D D D
Value 0.375 0.25 0.4375 1.25 1.225 0.225 0.74 0.735 0.375 0.45

Figure 2. The flow-chart relationship of present work.

For time and convenience, CFD is used instead of exper- the reproducibility was verified by using CFD with the
imental method to get estimated values for modeling of Reynold stress model. The improvement rates of opti-
p and η. The CFD results are validated by comparison mized results were compared with the reference model
with experiment results of the reference model available quantitatively.
in the literature (Muschelknautz & Krambrock, 1970;
Obermair et al., 2003; Obermair & Staudinger, 2001).
Next, the DOE analysis for the 10 cyclone geometric 2. Numerical simulation of cyclone separator
parameters is performed to get the significant geomet-
2.1. Reynolds averaged Naiver-Stokes (RANS) for
ric parameters. And then, the models based on RSM,
turbulence models
GMDH-NN and BPNN combined with the DOE method
have been developed to predict η and p with respect An appropriate turbulence model should be used to sim-
to the significant geometric parameters. The prediction ulate the turbulence behavior in CFD. The Reynolds
performances of the cyclone performance models are averaged Naiver-Stokes (RANS) turbulence models have
compared to analyze the advantages of each methodol- been successfully applied to predict the turbulence flow
ogy and differences between the models. And then, an in many industrial applications. Many studies have been
optimal predictive model for η and p is selected. conducted on the application of turbulence models to
Finally, the optimal predictive models were simulta- capture the behavior of cyclone flow. Boyan, Ayers, and
neously optimized to get maximum η and minimum p Swithenbank (1982) found the k − ε models cannot pre-
using NSGA-II. The ‘trade-off’ solution between opti- dict the turbulence behavior of cyclone with significant
mal models was selected in optimal Pareto front, and non-equilibrium effects of turbulent transport. Qian et al.
184 D. PARK ET AL.

(2006) showed that the cyclone flow can capture using the coefficient of the spherical particles is calculated as
Reynold stress model (RSM) which assumes a Reynolds
18μ CD Rep
stress tensor with isotropic eddy-viscosity hypothesis. In FD = (10)
addition to the RANS model, LES (Large Eddy Simu- ρp dp2 24
⎧
lation) shows relatively better prediction results about ⎪ 24
⎪
⎪ Rep ≤ 1,
the cyclone than RSM, but LES requires computational ⎪
⎪ Re
⎨
CD = 24(1 + 0.15Rep )
resources several times larger than RSM (Elsayed & 0.687
Lacor, 2013). Therefore, in this study, the RSM was used ⎪
⎪ 1 < Rep ≤ 1000, (11)
⎪
⎪ Rep
to simulate the cyclone flow. The RSM are expressed as ⎪
⎩0.44 1000 < Rep ,
follow:
∂ ∂ where the term CD is drag coefficient of the spherical par-
Rij + ūk Rij = DTij + Pij + Fij + ij − εij (3) ticle, Rep is the particle Reynold number. The turbulence
∂t ∂xk
fluctuations are random functions of space and time. A
DT,ij (turbulent diffusion) discrete random walk (DRW) model is applied to sim-
∂ ulate the instantaneous fluctuation velocity (uk ) in this
=− [ρu i u j u k + ρ(δkj u i + δik u j ) (4)
∂xk study. The uk affects to the particle transport for small
  particles. The values of uk that prevail during the life-
 
∂uj  
∂ui
Pij (stress production) = −ρ u i u k + u iu k time of the turbulent eddy are sampled by assuming that
∂xk ∂xk
they obey a Gaussian probability distribution as (ANSYS
(5)
FLUENT, 2016).
Fij (rotation production)
= −2ρ   + u i u k εjkm ) uk = ξ uk uk (12)
k (u j u m εikm (6)
 
∂u i ∂u j Where, ξ is a normally distributed random number. uk uk
ij (pressure strain) = −p + (7)
∂xj ∂xi is the local RMS value of the velocity fluctuations. The
  characteristic lifetime of the eddy (τe ) is defined either as
∂u i ∂u j a constant given by:
εij (pressure strain) = −2μ (8)
∂xk ∂xk
τe = 2TL (13)
where p and μ are, respectively, pressure, the molecular
where TL is the eddy turn over time given as, τe = 0.3 k/e.
viscosity, and δ is Kronecker delta.
The other option allows for a long–normal random vari-
ation of eddy lifetime that is given by:
2.2. Discrete phase models (DPM)
τe = −TL log(r) (14)
The particle trajectory in cyclone was calculated by
Lagrangian-based discrete phase model (DPM). The where r is a uniform random number between 0 and 1.
DPM assumes that the particle phase does not affect the The option of random calculation of τe yields a more
continuous phase when the volume occupied by the par- realistic description of the correlation function.
ticles in the continuous phase was less than 10% in DPM
(ANSYS FLUENT, 2016). So, DPM can be applied in 2.3. CFD simulation setup
cyclone simulation. Transport momentum equation for
each particle is expressed as: Since cyclone flows have the characteristics of strong
vortex and steep pressure distribution, it is necessary
dup  p − ρ)
g(ρ to accurately predict cyclone flows using powerful and
= FD (uk + uk − u p ) + +F (9) efficient algorithms. Shukla, Shukla, and Ghosh (2011)
dt ρp
conducted a case study to select the optimal algorithm
where the term FD (uk + uk − u p ) is the drag force per that can accurately simulate the cyclone flow. According
unit particle mass, u k is the time-averaged fluid veloc- to the results of the case study, in this study, the veloc-
ity in the k-th direction obtained by solving the RANS ity and pressure was connected by SIMPLE algorithm, for
model, uk is the instantaneous fluctuation velocity in static pressure distribution and the pressure drop, it was
the k-th direction and F is an additional acceleration interporated by the PRESTO method. Advection terms
(force/unit particle mass), ρp , ρ are the particle density were discretized by using the QUICK. The second-order
and the flow density, respectively. The drag force and drag upwind scheme was used for the momentum, turbulence
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 185

Table 2. Boundary condition for the cyclone CFD simulation.

Inlet velocity 12.69 (m/s)
Pressure outlet 101325 (Pa)
Wall No slip condition
Air viscosity 1.82 × 10−5 N × m2
Air density 1.203 kg
Time step size 0.001 s
Iteration number per time step 1500

dissipation rates and turbulence kinetic energy. To reduce

the numerical instability and to improve convergence,
the first-order upwind scheme was used for the Reynolds
stress.
The boundary conditions in this study were identi-
cal to the conditions used in the reference model. The
cyclone inlet was set to uniform velocity and the air flow
rate is 800 m3 /h (12.7 m/s). The cyclone outlet was set to
atmospheric pressure. The cyclone wall was set to no-slip
conditions. For the convergence of the cyclone flow, the
convergence criteria for optimization process are until
there is no fluctuation of pressure drop. The all simula-
tion are conducted as follow, First, the steady-state solu-
tion reached 3,000 iterations for stability of the solution. Figure 3. Mesh for the simulation of the cyclone.
And then, the iteration number per time step and the
time step size were selected as 1500 and 0.001 s, respec-
tively. This creates a physical time of 1.5 s. The boundary 2.5. Comparison between CFD results and reference
conditions for the CFD simulation were listed in Table 2. model
The commercial FVM simulation software ANSYS Flu- 2.5.1. CFD grid independence test
ent 16.2.was used to simulate the cyclone performance As the grid density is increased, the precise solution could
parameters. be obtained. However, as the number of grid increases,
the CPU time also increases. Figure 3 shows grid for
cyclone generated by using cut-cell meshing of ANSYS
2.4. DPM setting
Meshing 16.1. The geometric ratio of cyclone was used
For the discrete phase modeling (DPM) condition, the as the reference model. The near-wall treatment was
collusion condition between the wall and the particles achieved by using scalable wall functions considering the
was selected as elastic conditions. The collection of par- grid refinement with y+ < 11 (ANSYS FLUENT, 2016).
ticles is calculated if the particles reach the dust collector In this study, the grid independence test was con-
bottom. The particles density is 2770 kg/m3 . The number ducted until η and p were converged according to
of injected particles is 104 in the all simulation. The par- increasing the grid edge size. Figure 4 presents that the
ticle size distribution is divided to 10 class based on the results of grid test for η (cf. Figure 4(a)) and p (cf.
Rosin-Rammler theory (ANSYS FLUENT, 2016). Figure 4(b)). It showed that the p and η converged at the
6.5 mm of maximum grid size. Thus, in the all simulation,
n the maximum grid of 6.5 mm was used.
Yd = e−(d/d̄) (15)
2.5.2. Validation of results for velocity distribution
Where the n is the diffusion parameter. The d is the To validate the CFD simulation of cyclone, CFD simula-
mean diameter. For simulation, the d, n is defined to tion of cyclone velocity distribution was compared with
5 μm and 3.5, respectively. Also, the distribution of par- the results of available experimental data. Figure 5 pre-
ticle diameters is set from 1 to 10 μm. The particles are sented the consistency between the experimental data
divided into 10 classes. So, the particles can be mixed using Laser Doppler anemometry (LDA) method (Ober-
sufficiently. The injected 104 particles are tracked to cal- mair et al., 2003) and simulated results for tangential
culate the collection efficiency of the cyclone about all and axial velocities at y = 0.75D and y = 0.368D. As
simulation. shown in Figure 5, The k − ε models cause excessive
186 D. PARK ET AL.

Figure 4. (a) Results of the mesh independence test of the separation efficiency. (b) Result of the mesh independence test of the pressure
drop.
Note: unit % is multiplied by 100.

Figure 5. Comparison of CFD result for distribution of velocity with experiment data. The three rows present comparison against
Moazenzadeh et al. (2018) from left to right: first column – tangential velocity; second column – axial velocity; from top to bottom:
axial location in 0.75D (A-A ) and 0.368D (B-B ), respectively.
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 187

Table 3. Levels and factors for factorial experiment.

Factor Min (−1). (X/D) Ref. (X/D) Max. (X/D)
Dx /D (x1) 0.275 0.375 0.475
a/D (x2) 0.15 0.25 0.35
b/D (x3) 0.3375 0.4375 0.5375
Hcy /D (x4) 0.75 1.25 1.5
Hco /D (x5) 0.5 1.225 1.95
Bc /D (x6) 0.3 0.225 0.25
Ht /D (x7) 0.6 0.74 0.8
Hc /D (x8) 0.635 0.735 0.835
Dc /D (x9) 0 0.375 0.75
S/D (x10) 0.3 0.45 0.6
Cyclone diameter: 0.4 m

(FFD) method and the central composite design (CCD)

method are most commonly used. The DOE method can
Figure 6. Comparison of CFD result for separation efficiency efficiently apply not only to generate train set for deter-
curve with experiment data Moazenzadeh et al. (2018). mining the parameters of model but also to generate
Note: unit % is multiplied by 100. test set for evaluating the predictive performance of the
model.

turbulent viscosities and unrealistic tangential veloci-

ties due to the Boussinesq eddy viscosity assumption. 3.1.1. Fractional factorial design
Therefore, the k − ε models cannot capture the cyclone As the number of design variables increases, the number
flow. Whereas, the Reynolds Stress Model (RSM) can of physical/numerical experiments is increased exponen-
accurately capture the cyclone flow because the RSM tially to generate the DOE matrix. If a lot of design vari-
assumes a Reynolds stress tensor with three-dimensional ables are considered to model ANN and RSM based on
anisotropic property (ANSYS FLUENT, 2016). Further- DOE, it takes lots of time to obtain the physical/ numer-
more, the collection efficiency curve from CFD results ical experiment results. For solving the problem of many
was compared with the experimental results in Figure 6. variables, fractional factorial design (FFD) method is
The CFD results were consistent with those obtained introduced. It is effective to reduce the number of runs by
experimentally. The error between experiment result and compounding high-order interaction effects. It is advis-
CFD results for η and p were 1.1% and 0.5% respec- able to use the FFD to investigate design variables that
tively. Therefore, η and p resulted from CFD showed have a statistically significant effect before implementing
reasonable results by comparing the experiment data. modeling techniques to reduce the experiments size (Sun
This verified CFD simulation can be reasonably used for et al., 2017). Table 3 presents the selected cyclone geo-
various surrogate model method, such as response sur- metric parameters for the FFD with IV resolution and
face methodology and artificial neural network systems. the corresponding range of levels (−1, +1) (Hoffmann &
Stein, 2008). The FFD of IV resolution with 32 CFD runs
was performed. The statistical significance for cyclone
3. Surrogate model development for η and p geometric parameters was analyzed by Pareto chart with
analysis of variance (ANOVA).
3.1. Design of experiment
Figure 7 shows the main effect plot and Pareto chart
Design of Experiment (DOE) is a systematic method, for p (cf. Figure 7(a)) and η (cf. Figure 7(b)) at the
which can be effectively used to investigate the effect α = 0.05 significance level. Higher-order terms are not
of change of design variables on the response(s) and included in the chart because they have very small effects.
to provide reasonable answers about the behavior of Therefore, the higher term has been ignored. The vortex
a system (Mathews, 2005). In particular, the modeling finder diameter Dx (x1), the inlet width a (x2), the inlet
techniques, such as artificial neural networks (ANN) height b (x3) and the cone height H co (x5) have a sig-
and response surface techniques (RSM), are success- nificant effect on η and p. Also, the result of the main
fully combined with DOE methods (Khayet, Cojocaru, & effects for η and p showed the same tendency as Pareto
Essalhi, 2011; Rahimpour, Hatti-Kaul, & Mamo, 2016). chart results. In the case of the x5, it showed a negative
The DOE method can generate a matrix that provides effect on and p and a positive effect on- η. Therefore, the
maximum information by minimizing the combination four geometric parameters (Dx , a,b,H co ) are selected as
of variables. In general, the fractional factorial design design parameter for modeling p and η at next section.
188 D. PARK ET AL.

Figure 7. Analysis of design of experiment for the pressure drop (a) and separation efficiency (b); from top to bottom: Pareto chart, main
effect plot.

The other parameters are set as average values between Where βOLS is a regression coefficient vector, X is the
minimum and maximum values. matrix about the tested four geometrical parameters. The
matrix is generated using DOE method, which includes
all combinations of the four geometrical parameters. Y is
3.2. Response surface methodology (RSM)
the column vector of the experiment/simulation results
RSM uses a strong statistical method based on the least in the DOE matrix. In other words, the main advantage
square method that fits of the estimated values obtained of RSM is that it is simple and fast to develop predictive
from experiment/simulation to the quadratic polynomial models, but the nonlinear nature of the model is relatively
model of RSM. In order to predict the system behav- small compared to ANN.
ior, the quadratic or higher-order polynomial model is A typical DOE matrix for obtaining RSM model is
applied in many industrial applications. For modeling the the central composite design (CCD) method (Mathews,
cyclone performance parameters according to geometric 2005). The CCD enables to estimate the square terms of
parameters, the polynomial equation of RSM is expressed the second-order model efficiently. The CCD consists of
as a general form: a factorial experiment point, a center point and an axial
point. If the number of the design variables is k, the total
4 4 4 4
number of experiments of CCD can be written as
y k = β0 + βi xi + . βij xi xj + βii x2i + ε
i=1 i=1 j=1 i=1
N = 2k + 2k + nc (18)
(16)
where yk is the response surface models (k = 1 for col- where 2k is the factorial experiment points, and 2k is the
lection efficiency and k = 2 for pressure drop), and xi axial points, and nc is the number of iterations at the
and xj are the cyclone geometric parameters. β0 , βi , βij , center point. A value of α depends on the number of
βii are the regression coefficients for each constant, lin- experimental results in the factorial experiment point:
ear, interaction, and square term. ε is the prediction error
of the model. The regression coefficients were obtained α = 2(k−p)/4 (19)
by the ordinary least square (OLS) method. The OLS is
represented as follow, where if k is 0, that is full factorial design. And if k is con-
stants, that is fractional factorial design, α is the rotata-
βOLS = (X T X)−1 X T Y (17) bility. All factors are investigated in five levels (−α, −1,
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 189

0, 1, +α). In this work, the DOE matrix for CCD is gen- as Eq. 23.
erated by using the statistical analysis program (Minitab
17, MINITAB Inc.). n n n
ŷ = w0 + w i xi + wij xi xj
i=1 i=1 j=1
3.3. Group method of data handling (GMDH) type n n n
neural network + wijk xi xj xk + · · · (23)
i=1 j=1 k=1
The GMDH algorithm pertains to computer-based
mathematical modeling techniques. One of the charac-
teristics of GMDH is the fully automatic architecture which is known as a Volterra-Kolmogorov-Gabor poly-
and parametric optimization of the model (Onwubolu, nomial (Onwubolu, 2015). The GMDH-type models of
2015). The GMDH algorithm consists of several lay- p and η will be determined and evaluated in the fol-
ers containing neurons. Each neuron has two inputs (xi , lowing sections. All GMDH calculations were carried out
xj ) and on single output (ŷn ). The output of each neu- using MATLAB (R2018b) code.
ron can be represented through a quadratic polynomial
as Eq.20. 3.4. Artificial neural network technique based on
back propagation neural network (BPNN) algorithm
2 2
ŷn = w0 + w1 xin + w2 xjn + w3 xin xjn + w4 xin + w5 xjn
An artificial neural network (ANN) is a powerful compu-
(20) tational modeling technique for solving the multivariate
regression problems and for modeling of non-linear char-
Where n is the total number of datasets. wi is weigh of acteristic. The prediction accuracy of the ANN is affected
model. The weight of a neuron is obtained by minimizing by the ANN architecture, which consists of an input layer,
the square of the difference between the estimated values hidden layers, and an output layer (Khayet et al., 2011).
(yn ) and the predicted values (ŷn ) using model. The input layer receives the values of input variables and
connects to the hidden layer by multiplying the weight
N (wi ) as Eq. 24 and assigns those to an activation function
e = min (ŷn − yn )2 (21) as Eq. 25. The result of activation function is transferred
n=1 as new input to neurons in the next layer.

∂ N
n=1 (ŷn − yn )2 n
=0 (22) yj(k) = b0 + w i xi (24)
∂wi
i=1
1
That is, the gradient of e is minimized about wi . It called y(k) j_out = (25)
the gradient descent method. Test set is then used to cal- 1 + exp(−y(k) j )
culate the coefficient of determination (R2 ) and Akaike
information criterion (AIC) of the obtained each neuron. Where k is layer number, j is node number. The most
The R2 is the quantitative criteria of predictive perfor- common training algorithm for feed-forward neural net-
mance for the model. The AIC evaluates the overfit risk work is back-propagation (BP) method. ANN training
and overfit risk of the model. If the R2 and AIC of neurons through the BP algorithm is an iterative optimization
in calculating layer are higher than a predefined value, process that adjusts the weights appropriately to mini-
the neurons are transferred as new neurons in the next mize a performance function. A commonly used perfor-
layer. If the best criteria of the current layer are no longer mance function is defined as the root mean square error
higher than the best quantitative criteria of the previous (RMSE) as follow:
layer, the GMDH algorithm will stop. 
In other words, the optimal output neuron is selected 
1 n
automatically. Therefore, the GMDH algorithm can cap- RMSE =  (k)
(yi − yj_out ) (26)
ture not only the mathematical model with nonlinear, n i=1
but also the model with higher-order terms without
instability problems due to the inductive self-organizing Where  yi is i-th prediction value using the obtained
(Onwubolu, 2015). However, the GMDH is difficult to model. According to the BP algorithm, weights and biases
change the network structure such as the layer size, are iteratively updated in the fastest decreasing direction
node size. The final optimal neuron can be represented of the performance function, RMSE. In general, a single
190 D. PARK ET AL.

Table 4. Training parameter for BPNN. Table 5. CFD train set by using central composite design (CCD)
method.
Number of nodes 14 (η model), 8 (p model)
Number of layers 3 (η, p model) Run No. Dx (x1) a (x2) b (x3) Hco (x5) p (Pa) η (* 100%)
Learning rate 0.001 1 0.13 0.08 0.155 0.375 2877.51 0.913
Running algorithm Adam optimizer 2 0.17 0.08 0.155 0.375 1825.94 0.886
Early stopping condition 1E-05 (RMSE) 3 0.13 0.12 0.155 0.375 1728.13 0.869
Number of learning iterations 100,000 4 0.17 0.12 0.155 0.375 1029.31 0.813
5 0.13 0.08 0.195 0.375 2260.91 0.899
6 0.17 0.08 0.195 0.375 1400.14 0.863
iteration of the BP algorithm can be written as: 7 0.13 0.12 0.195 0.375 1419.62 0.843
8 0.17 0.12 0.195 0.375 807.01 0.742
9 0.13 0.08 0.155 0.725 2479.71 0.915
W (k+1) = W (k) − η grad(k) (RMSE) (27) 10 0.17 0.08 0.155 0.725 1595.98 0.906
11 0.13 0.12 0.155 0.725 1536.43 0.877
Where W (k) is a vector of the current weight and bias, 12 0.17 0.12 0.155 0.725 923.58 0.843
13 0.13 0.08 0.195 0.725 1970.14 0.908
grad(k) (RMSE) is the current gradient of the RMSE as 14 0.17 0.08 0.195 0.725 1243.12 0.881
performance function and η is the learning rate. The 15 0.13 0.12 0.195 0.725 1256.01 0.848
16 0.17 0.12 0.195 0.725 733.8787 0.777
update of the algorithm is terminated when the MSE 17 0.11 0.1 0.175 0.55 2407.213 0.896
does not decrease or exceeds a certain number of itera- 18 0.19 0.1 0.175 0.55 884.9296 0.819
19 0.15 0.06 0.175 0.55 2617.483 0.914
tions. In case of BPNN, the number of hidden nodes and 20 0.15 0.14 0.175 0.55 898.3643 0.771
hidden layers of optimal model are not predetermined. 21 0.15 0.1 0.135 0.55 1839.855 0.896
Therefore, it is important to find the optimal number 22 0.15 0.1 0.215 0.55 1116.163 0.841
23 0.15 0.1 0.175 0.2 1602.52 0.846
of node and layer. It called ‘hyperparameter tuning’. In 24 0.15 0.1 0.175 0.9 1257.097 0.881
other words, BPNNs can improve the prediction per- 25 0.15 0.1 0.175 0.55 1403.272 0.869
formance by adjusting the network structure, learning
algorithm, and learning rate. However, not only it takes a
long time to find the optimal learning condition, but also
it takes a long time to converge the model due to iterative Table 6. CFD test set by using full factorial design method (par-
tial).
method. Therefore, in this study, we found the optimal
network structure through greedy search method that Run No. Dx (x1) a (x2) b (x3) Hco (x5) p (Pa) η (* 100%)
increases the variable at regular intervals. For learning 1 0.15 0.1 0.135 0.9 1663.26 0.905405
2 0.11 0.06 0.135 0.55 5581 0.948657
the BPNN model, the learning rate was set to 0.001 and 3 0.11 0.1 0.215 0.55 1984.81 0.871337
the running algorithm was used to the Adam optimizer. 4 0.19 0.1 0.175 0.2 993.661 0.757621
The training of the BPNN model was terminated when 5 0.19 0.06 0.215 0.55 1380.96 0.877083
6 0.11 0.14 0.175 0.2 1904.14 0.835777
the RMSE was lower than 1E-05 or the iterations reached 7 0.19 0.14 0.175 0.9 509.558 0.705693
about 100,000. The learning parameters for BPNN which 8 0.11 0.06 0.175 0.9 3644.77 0.926619
9 0.11 0.14 0.135 0.55 2059.2 0.86531
selected in this work were summarized in Table 4. More 10 0.15 0.14 0.175 0.55 913.551 0.771701
details about mathematical aspects of BP training algo- 11 0.11 0.14 0.215 0.2 1545.63 0.763252
12 0.15 0.06 0.175 0.55 2677.83 0.908115
rithms are described to Ref. (Hagan, Demuth, & Beale, 13 0.15 0.14 0.135 0.9 1075.81 0.842602
1997). All BPNN calculations were carried out using 14 0.11 0.06 0.135 0.2 6806.19 0.950149
PYTHON 3.6. 15 0.15 0.14 0.215 0.9 682.42 0.723664
16 0.19 0.14 0.215 0.2 475.325 0.529653
17 0.19 0.14 0.175 0.55 551.344 0.672337
18 0.15 0.1 0.215 0.9 1024.92 0.846032
3.5. Train set and test set for comparing the 19 0.15 0.06 0.175 0.9 2341.57 0.911294
predictive model of p and η 20 0.11 0.1 0.215 0.9 1716.71 0.86715
21 0.15 0.06 0.135 0.2 4297.99 0.920058
22 0.19 0.14 0.215 0.55 432.684 0.605006
Twenty-five combinations (hereafter, CFD train set) were 23 0.19 0.06 0.215 0.9 1237.09 0.899584
generated using CCD as Table 5. The CFD train set is 24 0.15 0.14 0.215 0.2 824.4 0.687166
used to obtain predictive model for p and η using RSM, 25 0.15 0.14 0.135 0.55 1201.67 0.830576

GMDH and BPNN. Also, 81 combinations were gener-

ated using the full factorial design. Twenty-five out of
81 were randomly selected as CFD test set. The CFD
test set presents in Table 6. The CFD test set is used problem (Onwubolu, 2015). The range of four geomet-
to evaluate the prediction performance of the predictive ric parameters, x1, x2, x3 and x5 used for the modeling
models for p and η. The new combination of test sets methods is the same as Table 3. The other parameters (x4,
not only efficiently grasps the prediction performance x6, x7, x8, x9 and x10) are fixed as the average of their
of the model, but also overcomes the models over train parameter.
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 191

Table 7. Results for fitting of regression coefficient of RSM η terms are less than 0.05. The p-values of all the terms
model and ANOVA results for RSM term. of p and η were similar, except for x2 ∗ x2 term and x1
∗ x term. This indicates that p is sensitively affected by
Terms Regression coefficient of RSM (η) F-Value P-Value 5
β0 0.052 91.86 0 width length of the cyclone inlet. The interactive effect
Linear 293.32 0 between the cone length of cyclone and the outlet diam-
β1 5.41 269.26 0 eter can affect- p. Also, the interactive effects between
β2 8.16 724.9 0
β3 3.24 139.74 0 x1 and x2 have the highest among the interaction term
β4 −0.21 39.4 0 and that the interaction of x3 and x4 has the least effect
Quadratic 5.09 0.017 on p and η. The interactive effects of pair of x1, x2 (cf.
β1 2 −5.76 1.46 0.255 Figure 8(a)) and pair of x3, x5 (cf. Figure 8(b)) are visual-
β2 2 −15.13 10.08 0.01
β3 2 1.12 0.06 0.819 ized in Figure 8. Figure 8(a) showed a steep curve when
β4 2 −0.0262 0.18 0.683 the interactive effect is high. On the other hands, when
Interaction 15.41 0 the interactive effect is low, the curve is the similar as the
β1 *β2 −25.47 40.46 0
β1 *β3 −17.03 18.09 0.002
straight line as shown Figure 8(b).
β1 *β4 1.411 9.5 0.012 The ANOVA showed that the coefficient of determi-
β2 *β3 −19.22 23.04 0.001 nation (R2 ) of the p and the η are 99.83% and 99.21%
β2 *β4 0.518 1.28 0.284
β3 *β4 0.125 0.07 0.79 using CCD results (train set), respectively. The adjusted
y1 = 0.052 + 5.41 x1 + 8.16 x2 + 3.24 x3 − 0.210 x5 − 5.76 x21 − 15.13 x22 coefficient of determination (R2adj ) of the p and the
+1.12 x23 − 0.0262 x25 − 25.47 x1 x2 − 17.03 x1 x3 + 1.411 x1 x5 η considering the degree of freedom of the variables
−19.22 x2 x3 + 0.518 x2 x5 + 0.125 x3 x5
were 99.62% and 98.29% about train set, respectively.
However, the RSM model should be evaluated for accu-
Table 8. Results for fitting of regression coefficient of RSM p rate prediction performance by using a test set that is
model and ANOVA results for RSM term. not used for model parameter estimation rather than
Terms Regression coefficient of RSM p F-Value P-Value train set.
β0 23764 417.83 0 Therefore, the predictive performance of the RSM
Linear 1404.43 0 cyclone models was evaluated based on the CFD test
β1 −98913 2296.44 0 set as shown Figure 9. The R2 values of the p and η
β2 −118886 2636.12 0
β3 −52739 535.58 0 were 91.61% and 94.5%, respectively. In particular, the
β4 −3726 149.58 0 RSM model has relatively poor predictive performance
Quadratic 27.99 0 for collection efficiency data > 92% and pressure drop
β1 2 147586 26.7 0 data > 4000 Pa. Consequently, this indicates that the
β2 2 217493 57.98 0
β3 2 42547 2.22 0.167 RSM cyclone models have relatively low predictability for
β4 2 162 0.19 0.673 the new data.
Interaction 20 0
β1 *β2 168230 49.14 0
β1 *β3 81943 11.66 0.007
β1 *β4 8536 9.69 0.011 3.6.2. Modeling for η and p using GMDH-type
β2 *β3 141231 34.64 0 neural network
β2 *β4 9668 12.43 0.005
β3 *β4 4298 2.46 0.148 The coefficient of GMDH polynomial for η and p is
y2 = 23764 − 98913 x1 − 118886 x2 − 52739 x3 − 3726 x5 + 147586 x21
obtained as Tables 9 and 10, respectively. The adequacy
+217493 x22 + 42547 x23 + 162 x25 + 168230 x1 x2 + 81943 x1 x3 of GMDH model for η and p was evaluated by ana-
+8536 x1 x5 + 141231 x2 x3 + 9668 x2 x5 + 4298 x3 x5 lyzing agreement between the cyclone GMDH output
and the CFD test sets output. Figure 10 (a, b) present
comparison of the prediction results of the performance
3.6. Comparison of prediction performance and
parameters using GMDH with CFD results including
modeling results for η and p using RSM, GMDH and
x-y line. The predictive performance of the obtained
BPNN
the GMDH cyclone models were evaluated quantitatively
3.6.1. Modeling for η and p using RSM based on the correlation coefficient. The correlation coef-
The coefficient of RSM polynomial for η and p is ficients of η and p were 98.9%, 99.7%, respectively.
obtained as Tables 7 and 8, respectively. The RSM cyclone The predictive performance of the obtained the GMDH
models were evaluated by the analysis of variance (here- cyclone models were evaluated quantitatively using the
after, ANOVA). The ANOVA showed that each term of correlation coefficient. As shown Figure 10, the GMDH
the cyclone geometric parameters has significant effect cyclone models predicted effectively the estimated values
on performance parameters when the p-values of each by CFD.
192 D. PARK ET AL.

Figure 8. 3D plot and 2D contour of the response surface about separation efficiency (up) and pressure drop (down); (a) combination
x1 and x2; (b) combination x3 and x5.
Note: all other variables are hold at mean values.

Figure 9. Prediction performance results of cyclone RSM model for (a) separation efficiency and (b) pressure drop.
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 193

Table 9. GMDH η model.

Layer number GMDH model for η
(1) Y1 (1)= 0.877 + 0.107 DX + 2.6872 a + 1.031 D2X –12.18 a2 − 14.17 DX ∗ a
Y2 (1)= 0.784 + 1.751 a + 1.658 b − 9.74 a2–4.23 b2 − 9.73 a ∗ b
Y3 (1)= 0.902 + 0.259 a + 0.069 Hco − 10.75 a2 − 0.035 Hco 2 + 0.115 a ∗ H
co
Y4 (1)= 0.085 + 1.521 DX + 0.199 b − 5.09 D2X − 0.247 b2 –5.44 DX ∗ b
Y1 (2) 2 2
= −10 + 13.02 Y3 (1) + 11.39 Y4 (1) − 1.332Y3(1) − 0.325 Y4(1) − 11.4 Y3 (1) ∗ Y4 (1)
 
(2)
Y2 (2)= −1.11 + 3.97 Y2 (1) − 1.21 Y4 (1) + 1.64 Y22 2 
(1) + 4.61 Y4(1) − 6.81 Y2 (1) ∗ Y4 (1)


Y3 (2)= −3.65 + 3.63 Y1 (1) + 5.57 Y4 (1) + 1.28 Y12(1) − 0.135 Y4 2 − 5.73 Y1
(1) (1) ∗ Y4 
(1)
Y4 (2)= −1.73 + 2.46 Y1 (1) + 2.53 Y2 (1) + 1.764 Y12 (1) + 1.603 Y2 2 − 5.65 Y1
(1) (1) ∗ Y2
(3) Y(η− output) = 0.243 + 1.434 Y1(2) − 1.01 Y2(2) − 3.399 Y12 2 
(2) − 2.404 Y2(2) + 6.14 Y1(2) ∗ Y2(2)


Table 10. GMDH p model.

Layer number GMDH model for p
(1) Y1(1) = 15366 − 90840 DX − 85572 a − 178194 D2X –182057 a2 + 169080 DX ∗ a
Y2(1) = 7020 − 66939 a − 1266 b + 215008 a2 –20596 b2 + 4221 a ∗ b
Y3(1) = 6587–65887 a–1373 Hco + 2118802 a2 + 508 Hco 2 + 2420 a ∗ H
co
Y4(1) = 9641–46907 DX − 24054 b + 100521 D2X + 56256 b2 –22432 DX ∗ b
(2) Y1(2) = 579 − 0.178 Y3(1) − 0.739 Y4(1) + 0.00075 Y32(1) + 0.00022 Y42(1) + 0.00049 Y3(1) ∗ Y4(1)
Y2(2) = 313 − 0.12 Y2(1) − 0.28 Y4(1) + 0.00061 Y22(1) + 0.0001 Y42(1) + 0.00045 Y2(1) ∗ Y4(1)
Y3(2) = 90 + 0.75 Y1(1) + 0.01 Y2(1) − 0.00082 Y12(1) − 0.00053 Y22(1) + 0.00023 Y1(1) ∗ Y2(1)
Y4(2) = −234 + 0.737 Y1(1) + 0.39 Y4(1) − 0.00011 Y12(1) − 0.00018 Y42(1) + 0.00031 Y1(1) ∗ Y4(1)
(3) Y1(3) = −149 + 0.59 Y1(2) + 0.61 Y4(2) + 0.00026 Y12(2) − 0.00011 Y42(2) − 0.00021 Y1(2) ∗ Y4(2)
Y2(3) = −155.7 + 0.55 Y1(2) + 0.65 Y3(2) + 0.00011 Y12(2) − 0.00026 Y32(2) + 0.00091 Y1(2) ∗ Y3(2)
Y3(3) = −3.249 + 0.7478 Y1(2) + 0.2454 Y2(2) − 0.00014 Y12(2) − 0.00017 Y22(2) + 0.00033 Y1(2) ∗ Y2(2)
Y4(3) = −107.88 + 0.655 Y2(2) + 0.4296 Y4(2) − 0.00016 Y22(2) − 0.00017 Y42(2) + 0.00033 Y2(2) ∗ Y4(2)
(4) Y( p− output) = 3.304 + 0.652 Y1(3) + 0.335 Y4(3) − 0.00013 Y12(3) − 0.00021 Y42(3) + 0.00035 Y1(3) ∗ Y4(3)

3.6.3. Modeling for η and p using BPNN So, in this study, the best network architecture has
The prediction performance of the ANN model depends been investigated by grid search method (trial and error
on the number of neurons and the number of layers. method). Based on the grid search method, the root
If the number of datasets is small, larger the number mean square error (RMSE) of each BPNN model for the
of layers, and larger the overfitting that causes problem cyclone performance parameters was measured accord-
in prediction performance of new dataset (Onwubolu, ing to increasing the number of nodes and number of
2015). layers as shown Figure 11. As shown in the Fig. 11(a),

Figure 10. Prediction performance results of GMDH cyclone model for (a) separation efficiency and (b) pressure drop.
194 D. PARK ET AL.

Figure 11. Results of hyperparameter tuning (a) separation efficiency and (b) pressure drop; from up to down: number of node and
number of layer.

Figure 12. Prediction performance results of BPNN cyclone model for (a) separation efficiency and (b) pressure drop.

RMSE of η model is overfitted when the number of nodes and the prediction behavior for the three η models and
is more than 14 and the number of layers is more 3. three p models were analyzed using the CFD test set
Also, Figure 11(b) presents that the RMSE of the as shown Figures 13 and 14, respectively. The histogram
p model is overfitted when the number of nodes is of the error distributions can give better comparison
more than 8 nodes and the number of layers is more 3. for the prediction performance in terms of the possible
Therefore, the RMSE of the BPNN model of p and η error ranges (Onwubolu, 2015). The frequency of error
with optimal structure was 0.12 and 0.16, respectively. between the prediction results by the developed cyclone
Figure 12 present the prediction performance for BPNN models and the estimated results by CFD was normalized
cyclone model based on the CFD test set. It showed for convenient.
the result of the prediction performance in the optimal As shown 13 and 14, the probability density func-
BPNN structure, which is a good agreement, and the R2 tion and the frequency the error were widely distributed
for p and η was 98.1 and 98.5, respectively. in RSM cyclone model for η and p, compared to the
neural network approaches. It showed low predictive per-
3.6.4. Comparison of prediction performance among formance of the RSM cyclone model. On the other hand,
RSM, GMDH and BPNN for p and η the GMDH results and BPNN results have higher pre-
In this study, the predictive model for p and η were diction accuracy because the neural network approaches
obtained using RSM, GMDH and BPNN algorithm. In are considered for high-order nonlinearities unlike RSM
order to compare the prediction accuracy for the pre- approaches. Also, the predictive performance of RSM is
dictive models, the histogram of the error distributions especially poor at data points (3, 10, 22) for collection
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 195

Table 11. Predictive model comparison of η and p parameters

obtained by RSM, GMDH and BPNN.
RSM GMDH ANN
η p η p η p
R2 (%) 94.5 91.6 98.98 99.69 98.5 98.1
RMSE 0.023 0.235 0.012 0.103 0.017 0.106
R2adjusted 93.5 90.22 97.88 99.3 97.44 96.38
AIC 225.24 118.92 186.64 36.8 220.13 72.53

model and the GMDH model was quite similar. How-

ever, training in the GMDH is based on performing linear
algebra which requires short time for convergence rather
than an iterative method of minimizing the error like
the back-propagation algorithm which requires a large
amount of time for convergence (Hagan et al., 1997). The
Figure 13. Comparison analysis of prediction performance test convergence speed of the GMDH model of p and η (41
of separation efficiency model using RSM, GMDH and BPNN based sec) was 15 times lower than that of BPNN model (621
the CFD test set; (a) Error histogram and probability density func- sec).
tion for difference between CFD result and prediction value by
models (b) Error value between CFD result and prediction value Moreover, the GMDH cyclone models are mathemat-
by models. ically expressed as a higher-order polynomial model.
However, in the case of the BPNN cyclone model, the
obtained model is still hidden as the black box model.
The neural structure (the number of layer) for the BPNN
cyclone models were 4. The neural structure for GMDH’s
p, η model were 4, 3, respectively. The BPNN algorithm
needs to select the optimum neural structure by adjust-
ing the hyperparameter, while the GMDH algorithm
can adaptively optimize the structure to avoid overfitting
problem.
Errors between predicted values by each model and
estimated values at each CFD test set are shown in Fig-
ures 13 (b) and 14(b). It shown that the error of GMDH
models is close to zero at all data point, compared to RSM
and BPNN. Also, Table 11 presents the predictive perfor-
mance of the three models using R2 , RMSE, R2adjusted and
AIC based on the CFD test sets. The AIC and R2adjusted
Figure 14. Comparison analysis of prediction performance test evaluate the overfit risk and underfit risk of the model. As
of pressure drop model using RSM, GMDH and BPNN based the shown Table 11, the GMDH cyclone models (p: R2 =
CFD test set; (a) Error histogram and probability density function 99.7, RMSE = 0.102, R2adjusted = 98.4, AIC = 36.794
for difference between CFD result and prediction value by models
(b) Error value between CFD result and prediction value by models and η: R2 = 98.99, RMSE = 0.0119, R2adjusted = 94.45,
AIC = 186.64) showed the best prediction performance
than other models.
efficiency and pressure drop. In case of the data points, Furthermore, in order to quantitatively estimate and
the flow rate is the same, but the inlet area is reduced, compare the uncertainties of the cyclone models, uncer-
and the exit diameter of the cyclone is reduced, which tainty analysis (Gholami et al., 2018) was performed with
increases the inlet velocity, resulting in a strong vor- the 25 CFD test set. To estimate the uncertainties, predic-
tex with a physically sharp pressure gradient inside the tion error (PE), mean prediction error (MPE), standard
cyclone. The GMDH and BPNN models with higher deviation of prediction error (SDPE) and width of uncer-
terms than the RSM show relatively good results in tainty band (WUB) and prediction error interval of 95%
order to predict this sudden increase in pressure gradi- (PEI) are introduced. The value of MPE indicates an
ent. When comparing the BPNN results and the GMDH overestimated model for positive numbers and an under-
results, the variance of the error histogram of the BPNN estimated model for negative numbers. The WUB and
196 D. PARK ET AL.

Table 12. Uncertainty analysis results for cyclone performance cyclone models for p and η was reasonably selected as
models of GMDH, BPNN and RSM. the optimal predictive model in this paper. The GMDH
# Number cyclone models for p and η are used in next section for
Models of samples MPE SDPE WUB 95% PEI the Pareto multi-objective optimization.
GMDH for η 25 0.023 0.0218 ±0.0065 (0.016 0.029)
BPNN for η 25 0.029 0.0288 ±0.0085 (0.021 0.037)
RSM for η 25 0.039 0.0478 ±0.0132 (0.025 0.523) 4. Multi-objective optimization for η and Δp by
GMDH for p 25 0.019 0.0865 ±0.0188 (0.002 0.038)
BPNN for p 25 0.057 0.208 ±0.0464 (0.011 0.103) using genetic algorithm
RSM for p 25 0.1105 0.416 ±0.0955 (0.015 0.206)
4.1. Genetic algorithm
GA (Genetic Algorithms) are one of the optimization
PEI can be expressed by the SDPE.
methods to solve local optimization problem. The GA
n n inspired by the natural evolution. The population evolves
MPE = ej = (Pi − Ti ) (28) towards the optimal form under the natural conditions.
i=1 i=1 The GA repeatedly updates the population of individual
 solutions. In each step of the GA process, the GA pro-
n
i=1 (ej
− MPE)
SDPE = (29) duces initial random population and evaluates fitness for
n−1 the individual. The best individuals are selected as the
PEI = ±1.96 ∗ SDPE (30) parent, and then the GA produces the offspring by apply-
ing crossover and mutation for the next generation. The
Where, Pi is the predicted value, Ti is the esti- GA is terminated when it reaches to the certain criterion.
mated value. The uncertainty analysis results for the all However, in the many multi-optimization problems,
cyclone models are shown in Table 12. The WUB of the objective functions conflicted with each other. There-
GMDH models showed the lowest values (η = ±0.0065, fore, it is hard to obtain the best optimized points to
p = ±0.0188) compared to the RSM and the BPNN optimize all objective functions simultaneously. For solv-
results. Also, it was seen that the model of smallest MPE ing the problem, Den, Agrawal, Pratap, and Meyarivan
and 95% PEI are the GMDH model. Furthermore, the (2000) proposed the non-dominant genetic algorithm-
width of uncertainty band between the estimated values II (NSGA-II). The non-dominated solutions set (also
by CFD and the predicted values by the three models expressed as Pareto optimal front) provide the optimal
were visualized in Figure 15. The upper line and lower solution by determining the point among solution set.
line of 95% PEI for η and p predicted by GMDH are For more details about the GA, see the Ref (Elsayed &
best distributed in the exact line. Therefore, the GMDH Lacor, 2013).

Figure 15. Comparison analysis of qualitative scatter plot about width of uncertainty band of (a) Separation efficiency model and (b)
Pressure drop model by three proposed method (BPNN, GMDH and RSM).
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 197

Table 13. Genetic algorithm setup for multi-objective optimiza- used for the GA optimization. The constraints of geomet-
tion. rical parameters used in the GA optimization are same as
Population type Double vector range of parameters at Table 3.
Fitness scaling Rank Figure 16 presents the results of 70 Pareto optimal
Selection operation Roulette method front for η and p. The η is expressed as negative for the
Crowding distance fraction 0.35
Crossover fraction 0.8 optimization purpose. The five design points (A,B,C,D,E)
Crossover operation Intermediate crossover with the default from 70 Pareto optimal front were determined to decide
value of 1.0
Mutation operation The constraint dependent default the optimal p and η. Point A have the high pressure
Maximum number of generations 1000 drop compared to other point, it may require the high
Population size 120 power to operate the manufacturing system. Point E has
the low collection efficiency compared to other point. It
cannot function properly as a cyclone. Point C was cho-
sen to optimization cycle. The separation efficiency and
pressure drop of Point C are 91.1% and 98.5 Pa, respec-
tively, it has good performance compared the reference
data. Also, Table 14 shows the reproducibility results. The
reproducibility results show that GMDH cyclone models
have high predictive performance.
Finally, Table 15 shows the improvement rate of p
and η of the optimized cyclone and the dimension ratio
of the optimized cyclone. The η, d50 and p of optimized
cyclones improved 8.79%, 18%, and 24.31% respectively
over the performance of the reference cyclone. Figure 17
presents the comparison of the collection efficiency curve
between the reference cyclone and the optimization
cyclone. It presented that the collection efficiency for par-
Figure 16. Pareto front for multi-objective optimization for ticles less than 1 micron was improved by about 20%.
cyclone performance parameters.
Figure 18 presents the particle behavior in the optimized
cyclone. Therefore, the optimized cyclone demonstrates a
reasonable approach in the GMDH-type neural network
4.2. Results of multi-objective optimization for p
combined DOE and NSGA-II optimization processes in
and η using NSGA-II
the present work.
For multi-objective optimization for p and η, the non-
dominated solutions set was obtained by NSGA-II using
4.3. Results of numerical simulation comparison
MATLAB 2018a. Table 13 presents a setting condition for
between optimized cyclone and reference model
multi-objective optimization. The conditions in NSGA-
II were set up to search the entire design area. The fitness In order to analyze why the optimized cyclone model
functions from GMDH-cyclone model for p and η were could achieve better performance than the reference

Table 14. Validation of Pareto optimal front for the cyclone performance.
GA results CFD results Error (%)
Dx (x1) a (x2) b (x3) Hco (x4) η (%) p (Pa) η (%) p (Pa) η p
Point A 0.059 0.135 0.185 0.9 0.97 2072 0.93 2144 4.3 3.35
Point B 0.085 0.135 0.185 0.9 0.93 1351 0.91 1349 2.15 0.15
Point C 0.105 0.1385 0.187 0.87 0.91 984 0.88 970 3.29 1.42
Point D 0.116 0.175 0.189 0.8 0.79 725 0.78 689 1.26 4.96
Point E 0.126 0.213 0.192 0.24 0.61 618 0.58 563 4.91 8.89

Table 15. Comparison of results for cyclone performance parameters between optimized
cyclones and reference model.
Dx (x1) a (x2) b (x3) Hco (x4) η (%) p (Pa) Cut-off size (µm)
Reference model 0.15 0.1 0.175 0.49 0.83 1300 1.83
Optimized model 0.105 0.1385 0.187 0.87 0.91 984 1.5
Improvement (%) 30 −38.5 −6.85 −77.55 8.79 24.31 18.03
198 D. PARK ET AL.

cyclone, the separation efficiency of optimization cyclone

outperform.
The contour of pressure drops of the optimized
cyclone and reference cyclone presents in Figure 19(b),
respectively. The maximum pressure of the optimized
cyclone was about 200 lowers than the reference cyclone.
This high pressure drop requires the high power to oper-
ate the manufacturing system and affects exposure of the
particles in the cyclone called the phenomenon of ‘lip
leakage’. This resulted in an improvement in the cyclone
separation performance.
Figure 19(c) presents the turbulent kinetic energy
(TKE) of the reference cyclone and the optimized
cyclone. As shown in the Figure 19(e), the TKE was
observed relatively high near the cyclone vortex finder
Figure 17. Comparison of the separation efficiency curve
between the optimized cyclone and the reference model. of the reference model. As the TKE increases, the col-
note: unit % is multiplied by 100. lection efficiency decreases. The maximum TKE of the
optimized cyclone decreased by 25%, compared to the
reference cyclone.

model, the flow pattern of optimized cyclone was com-

pared with the reference model by using CFD. The
5. Conclusion
secondary circulation in cyclone decreases the cyclone
separation performance due to particle re-entrainment This study performed a comparative study among RSM,
phenomenon. Figure 19(a) presents the comparison for BPNN and GMDH algorithm for predictive models of
velocity vector field of the optimized cyclone and ref- p and η, and multi-objective cyclone optimization was
erence cyclone. The secondary circulation occurred at successfully performed by using NSGA-II. In this study,
the reference cyclone. However, the optimized cyclone MM model was selected as the reference model, and
showed that the area of secondary circulation was the optimization objectives were defined to minimize
decreased as shown Figure 19(a), compared to refer- p and maximize η. After the DOE analysis for the
ence model. Therefore, when compared with reference cyclone geometric parameters using the validated CFD,

Figure 18. The behavior when the particles of different sizes (1–10 μm)are injected into the optimized cyclone.
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 199

optimal design points were obtained by GA. The five

optimal point were selected to validate accuracy of GA
results using CFD. The reproducibility results showed
less than 10% error and proved the reliability of the GA
optimization process. The optimized cyclone (Point C)
provided a more reasonable improvement than the MM
model. The optimized cyclone reduced the pressure drop
by 24.31% (from 1300 to 984 Pa), the cut-off diameter by
18% (from 1.83 μm to 1.5 μm) and increased the over-
all efficiency by 8.79% (from 83% to 91%). In addition,
the optimized reason was demonstrated by analysis of
the flow part of the cyclone (flow field, pressure drop,
turbulence kinetic energy).
The results presented can help with selection of appro-
priate prediction modeling techniques about the cyclone,
and the obtained 70 Pareto front design points can
assist designers and decision makers to optimize cyclone
design. However, the optimization results need to be ver-
ified using experimental methods in addition to CFD.
Therefore, future studies will further demonstrate the
feasibility of the research by fabricating the cyclone based
on optimized cyclone dimensions and conducting exper-
iments on separation performance and pressure drop.

Disclosure statement
No potential conflict of interest was reported by the authors.

ORCID
Donggeun Park http://orcid.org/0000-0002-6524-2858

References
ANSYS FLUENT, 16.1. (2016). User’s and theory guide. Canons-
burg, PA, USA: ANSYS, Inc.
Ardabili, S. F., Najafi, B., Shamshirband, S., Bidgoli, B. M., Deo,
R. C., & Chau, K. W. (2018). Computational intelligence
approach for modeling hydrogen production: A review.
Engineering Applications of Computational Fluid Mechanics,
12, 438–458. doi:10.1080/19942060.2018.1452296
Avci, A., & Karagoz, I. (2001). Theoretical investigation of
pressure losses in cyclone separators. International Com-
Figure 19. Results of flow pattern analysis; from left to right: Ref- munications in Heat and Mass Transfer, 28, 107–117.
erence cyclone and optimized cyclone; (a) comparison for velocity doi:10.1016/S0735-1933(01)00218-4
vector field. (b) comparison for pressure drop. (c) comparison for Barth, W. (1956). Design and layout of the cyclone separator
turbulent kinetic energy. on the basis of new investigations. Brennstow-Wäerme-Kraft
(BWK), 8(4), 1–9.
Bernardo, S., Mori, M., Peres, A. P., & Dionísio, R. P. (2006).
3-D computational fluid dynamics for gas and gas-particle
the predictive models for p and η were obtained based flows in a cyclone with different inlet section angles. Powder
on RSM, GMDH technique and BPNN technique. As Technology, 162, 190–200. doi:10.1016/j.powtec.2005.11.007
a result of comparing the obtained models, the GMDH Boyan, F., Ayers, W. H., & Swithenbank, J. (1982). A Funda-
mental mathematical modeling approach to cyclone design.
cyclone models predicted the p and η most efficiently.
Transactions of the Institution of Chemical Engineers, 60,
The R2 of η and p was 98.9%, 99.7%, respectively. 222–230.
Therefore, the GMDH cyclone models were used in the Brar, L. S., Sharma, R. P., & Elsayed, K. (2015). The effect
GA as the fitness functions. A total of 70 non-dominated of the cyclone length on the performance of Stairmand
200 D. PARK ET AL.

high-efficiency cyclone. Powder Technology, 286, 668–677. Kaya, F., & Karagoz, I. (2009). Numerical investigation of
doi:10.1016/j.powtec.2015.09.003 performance characteristics of a cyclone prolonged with
Den, K., Agrawal, S., Pratap, A., & Meyarivan, T. (2000). a dip leg. Chemical Engineering Journal, 151, 39–45.
A fast elitist non-dominated sorting genetic algorithm for doi:10.1016/j.cej.2009.01.040
multi-objective optimization: NSGA-II. International Con- Khalkhali, A., & Safikhani, H. (2012). Pareto based multi-
ference on Parallel Problem Solving from Nature, 849–858. objective optimization of a cyclone vortex finder using CFD,
doi:10.1007/3-540-45356-3_83 GMDH type neural networks and genetic algorithms. Engi-
De Souza, F. J., Salvo, R. D. V., & Martins, D. D. M. (2015). neering Optimization, 44, 105–118. doi:10.1080/0305215X.
Effects of the gas outlet duct length and shape on the per- 2011.564619
formance of cyclone separators. Separation and Purification Khayet, M., Cojocaru, C., & Essalhi, M. (2011). Artificial neu-
Technology, 142, 90–100. doi:10.1016/j.seppur.2014.12.008 ral network modeling and response surface methodology
Dietz, P. W. (1981). Collection efficiency of cyclone separators. of desalination by reverse osmosis. Journal of Membrane
AIChE Journal, 27, 888–892. doi:10.1002/aic.690270603 Science, 368, 202–214. doi:10.1016/j.memsci.2010.11.030
Ebtehaj, I., & Bonakdari, H. (2016). Assessment of evolution- Leith, D. (1990). The logistic function and cyclone fractional
ary algorithms in predicting non-deposition sediment trans- efficiency. Aerosol Science and Technology, 12, 598–606.
port. Urban Water Journal, 13, 499–510. doi:10.1080/15730 doi:10.1080/02786829008959373
62X.2014.994003 Mathews, P. (2005). Design of experiments with MINITAB. Mil-
Elsayed, K., & Lacor, C. (2011). The effect of cyclone inlet waukee: ASQ Quality Press.
dimensions on the flow pattern and performance. Applied Moazenzadeh, R., Mohammadi, B., Shamshirband, S., & Chau,
Mathematical Modelling, 35, 1952–1968. doi:10.1016/j.apm. K. W. (2018). Coupling a firefly algorithm with support vec-
2010.11.007 tor regression to predict evaporation in northern Iran. Engi-
Elsayed, K., & Lacor, C. (2012). Modeling and Pareto optimiza- neering Applications of Computational Fluid Mechanics, 12,
tion of gas cyclone separator performance using RBF type 584–597. doi:10.1080/19942060.2018.1482476
artificial neural networks and genetic algorithms. Powder Muschelknautz, E., & Krambrock, W. (1970). Aerodynamis-
Technology, 217, 84–99. doi:10.1016/j.powtec.2011.10.015 cheBeiwerte des Zyklonabscheidersaufgrundneuer und
Elsayed, K., & Lacor, C. (2013). CFD modeling and multi- verbesserterMessungen. Chemie Ingenieur Technik - CIT, 42,
objective optimization of cyclone geometry using desirabil- 247–255. doi:10.1002/cite.330420503
ity function, artificial neural networks and genetic algo- Najafi, B., Faizollahzadeh Ardabili, S., Shamshirband, S., Chau,
rithms. Applied Mathematical Modelling, 37, 5680–5704. K. W., & Rabczuk, T. (2018). Application of anns, anfis
doi:10.1016/j.apm.2012.11.010 and rsm to estimating and optimizing the parameters that
Fotovatikhah, F., Herrera, M., Shamshirband, S., Chau, K. W., affect the yield and cost of biodiesel production. Engineering
Ardabili, S. F., & Piran, M. J. (2018). Survey of computa- Applications of Computational Fluid Mechanics, 12, 611–624.
tional intelligence as basis to big flood management: Chal- doi:10.1080/19942060.2018.1502688
lenges, research directions and future work. Engineering Obermair, S., & Staudinger, G. (2001). The dust outlet of a
Applications of Computational Fluid Mechanics, 12, 411–437. gas cyclone and its effects on collection efficiency. Chemical
doi:10.1080/19942060.2018.1448896 Engineering & Technology, 24, 1259–1263. doi:10.1002/1521-
Ganegama Bogodage, S., & Leung, A. Y. T. (2015). CFD sim- 4125(200112)24:12 < 1259::AID-CEAT1259 > 3.0.CO;2-O
ulation of cyclone separators to reduce air pollution. Pow- Obermair, S., Woisetschläger, J., & Staudinger, G. (2003). Inves-
der Technology, 286, 488–506. doi:10.1016/j.powtec.2015. tigation of the flow pattern in different dust outlet geome-
08.023 tries of a gas cyclone by laser Doppler anemometry. Pow-
Gholami, A., Bonakdari, H., Ebtehaj, I., Mohammadian, M., der Technology, 138, 239–251. doi:10.1016/j.powtec.2003.
Gharabaghi, B., & Khodashenas, S. R. (2018). Uncertainty 09.009
analysis of intelligent model of hybrid genetic algorithm and Onwubolu, G. (2015). GMDH methodology and implementa-
particle swarm optimization with ANFIS to predict thresh- tion in MATLAB. London: Imperial College Press.
old bank profile shape based on digital laser approach sens- Qian, F., Zhang, J., & Zhang, M. (2006). Effects of the pro-
ing. Measurement: Journal of the International Measurement longed vertical tube on the separation performance of
Confederation, 121, 294–303. doi:10.1016/j.measurement. a cyclone. Journal of Hazardous Materials, 136, 822–829.
2018.02.070 doi:10.1016/j.jhazmat.2006.01.028
Gimbun, J., Chuah, T. G., Choong, T. S. Y., & Fakhru’l-Razi, Rahimpour, F., Hatti-Kaul, R., & Mamo, G. (2016). Response
A. (2005). A CFD study on the prediction of cyclone col- surface methodology and artificial neural network mod-
lection efficiency. International Journal of Computational elling of an aqueous two-phase system for purification of a
Methods in Engineering Science and Mechanics, 6, 161–168. recombinant alkaline active xylanase. process Biochemistry,
doi:10.1080/15502280590923649 51, 452–462. doi:10.1016/j.procbio.2015.12.018
Hagan, M. T., Demuth, H. B., & Beale, M. (1997).Neural net- Raoufi, A., Shams, M., Farzaneh, M., & Ebrahimi, R. (2008).
work design. Boston: PWS Publishing Co. Numerical simulation and optimization of fluid flow in
Hamdy, O., Bassily, M. A., El-Batsh, H. M., & Mekhail, T. A. cyclone vortex finder. Chemical Engineering and Processing:
(2017). Numerical study of the effect of changing the cyclone Process Intensification, 47, 128–137. doi:10.1016/j.cep.2007.
cone length on the gas flow field. Applied Mathematical 08.004
Modelling, 46, 81–97. doi:10.1016/j.apm.2017.01.069 Safikhani, H., Akhavan-Behabadi, M. A., Shams, M., &
Hoffmann, A. C., & Stein, L. E. (2008).Gas cyclones and Swirl Rahimyan, M. H. (2010). Numerical simulation of flow field
Tubes: Principle, design and operation (2nd ed.). Berlin: in three types of standard cyclone separators. Advanced Pow-
Springer. der Technology, 21, 435–442. doi:10.1016/j.apt.2010.01.002
 ENGINEERING APPLICATIONS OF COMPUTATIONAL FLUID MECHANICS 201

Shephered, C. B., & Lapple, C. E. (2005). Flow pattern and pres- computational fluid dynamics. Powder Technology, 325,
sure drop in cyclone dust collectors. Industrial & Engineering 347–360. doi:10.1016/j.powtec.2017.11.012
Chemistry Research, 31, 972–984. doi:10.1021/ie50356a012 Taherei Ghazvinei, P., Hassanpour Darvishi, H., Mosavi,
Shukla, S. K., Shukla, P., & Ghosh, P. (2011). Evaluation A., Bin Wan Yusof, K., Alizamir, M., Shamshirband,
of numerical schemes using different simulation methods S., & Chau, K. W. (2018). Sugarcane growth prediction
for the continuous phase modeling of cyclone separators. based on meteorological parameters using extreme learning
Advanced Powder Technology, 22, 209–219. doi:10.1016/ machine and artificial neural network. Engineering Appli-
j.apt.2010.11.009 cations of Computational Fluid Mechanics, 12, 738–749.
Sun, X., Kim, S., Yang, S. D., Kim, H. S., & Yoon, J. Y. doi:10.1080/19942060.2018.1526119
(2017). Multi-objective optimization of a Stairmand cyclone Yang, J., Sun, G., & Gao, C. (2013). Effect of the inlet dimensions
separator using response surface methodology and com- on the maximum-efficiency cyclone height. Separation and
putational fluid dynamics. Powder Technology, 320, 51–65. Purification Technology, 105, 15–23. doi:10.1016/j.seppur.
doi:10.1016/j.powtec.2017.06.065 2012.12.020
Sun, X., & Yoon, J. Y. (2018). Multi-objective optimization
of a gas cyclone separator using genetic algorithm and