PHD Thesis Khairy Elsayed

FACULTY OF ENGINEERING Department of Mechanical Engineering
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies
Thesis submitted in fulllment of the requirements for the award of the degree of Doctor in de Ingenieurswetenschappen (Doctor in Engineering) by
Khairy Elsayed
Brussels, October 2011 Advisor: Prof. Dr. Ir. Chris Lacor
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies
by
Khairy Elsayed
Submitted to the Department of Mechanical Engineering, in partial fulllment of the requirements for the degree of Doctor in Engineering
Vrije Universiteit Brussel October 2011
Advisor: Prof. Dr. Ir. Chris Lacor
Analysis and Optimization of Cyclone Separators Geometry Using RANS and LES Methodologies Khairy Elsayed Department of Mechanical Engineering, Vrije Universiteit Brussel Pleinlaan 2, B-1050 Brussels, Belgium Thesis submitted in partial fulllment of the requirements for the academic degree of Doctor in Engineering
Promoter: Prof. dr. ir. Chris Lacor Jury: Prof. dr. ir. Johan Deconinck, voorzitter Prof. dr. ir. Rik Pintelon, vice-voorzitter Prof. dr. ir. Gunther Steenackers, secretaris Prof. dr. ir. Gert Desmet Prof. dr. ir. Harry van den Akker (Delft University of Technology, Netherlands) Prof. dr. ir. Herman Deconinck (Von Karman Institute, Belgium)
2011 Khairy Elsayed 2011 Uitgeverij University Press Leegstraat 15 B-9060 Zelzate Tel +32 9 342 72 25 E-mail: info@universitypress.be www.universitypress.be Vrije Universiteit Brussel Faculteit Ingenieurswetenschappen Pleinlaan 2 1050 Brussel Contact: +32 (0)2 629 39 10 http://www.vub.ac.be/IR secr-dtw@ir.vub.ac.be ISBN 978-94-9069-594-1 All rights reserved. No parts of this book may be reproduced or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without the prior written permission of the author.
Abstract
The gas-solids cyclone separator is an industrial equipment that has been widely used for more than a century. Due to its industrial relevance, a large number of experimental, theoretical and computational studies have been reported in the literature aimed at understanding and predicting the performance of cyclones in terms of pressure loss and collection efciency (cut-off diameter). The currently used mathematical models for the prediction of cyclone performance, however, exhibit limited accuracy and generality. Moreover, the cyclone performance can be calculated using the articial neural networks approach. An alternative approach is to simulate the gas-particle ow eld in a cyclone by computational uid dynamics (CFD). As a result of the recent progress of computational power and numerical techniques, CFD has been widely applied to industrial ow problems. The cyclone performance parameters are governed by many operational parameters (e.g., the gas ow rate and temperature) and geometrical parameters. This study focuses only on the effect of the geometrical parameters on the ow eld pattern and performance of the tangential inlet cyclone separators using three different approaches, the most robust mathematical models, articial neural networks and CFD approaches. The study was limited to reverse-ow gas-solids cyclone separators operating at low solids loading. The objective of this study is four-fold. First, to determine the most signicant factors affecting the cyclone performance based on the previous studies and statistical analysis of data using response surface methodology. Second, to study each (signicant) parameter separately to obtain more details about its effect on the ow eld pattern and the performance. Third, to obtain the most efcient cyclone design for minimum pressure drop (using the most robust mathematical models, articial neural networks and CFD dataset). Finally, to obtain the most efcient cyclone design for best performance (minimum pressure drop and minimum cut-off diameter) using multi-objective optimization techniques with two different optimization techniques (both the Nelder-Mead with desirability function and the genetic algorithms (NSGA-II)). The response surface methodology has been performed using dataset obtained from the Muschelknautz method of modeling (MM) to determine the most signicant parameters. Four geometrical factors have signicant effects on the cyclone performance viz., the vortex nder diameter, the ini
let width, the inlet height and the cyclone total height. There are strong interactions between the effect of inlet dimensions and the vortex nder diameter on the cyclone performance. The same investigation has been repeated using articial neural network approach based on the experimental pressure drop. A radial basis neural network (RBFNN) is developed and employed to model the pressure drop for cyclone separators. The neural network has been trained and tested by experimental data available in literature. The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone pressure drop. The analysis indicates the signicant effect of the vortex nder diameter and the vortex nder length, the inlet width and the total height. Furthermore, Four mathematical models (Muschelknautz method MM, Stairmand, Ramachandran and Shepherd and Lapple) have been tested against the experimental values. The residual error (the difference between the experimental value and the model value) of the MM model is the lowest. The numerical simulations of cyclone ow were carried out by solving the unsteady-state, three-dimensional Reynolds averaged Navier-Stokes (RANS) equations combined with a closure model for the turbulent stresses and the large eddy simulation approach. The modeling of the cyclonic ow by computational uid dynamics (CFD) simulation has been reported before in the literature. Using the experimental data available in literature, a generic assessment was carried out on a number of turbulence closure models. Only the Reynolds stress turbulence model (RSM) and large eddy simulation (LES) captured the cyclone ow eld best compared to the experimental measurements. The CFD model also predicted the collection efciency, where the particulate phase was treated in a Lagrangian framework by tracking a large number of particles of different size classes through the computational domain. The stochastic nature of the particle motion due to the uid turbulence was taken into account by a particle dispersion model. There was a reasonable agreement between the calculated and measured cut-off diameter for both the RSM and LES results. The effect of the cone tip-diameter on the ow eld and performance of cyclone separators was investigated because of the discrepancies and uncertainties in the literature about its inuence. Three cyclones with different cone tip diameters were studied using large eddy simulation (LES). The ow eld pattern has been simulated and analyzed with the aid of velocity components and static pressure contour plots. The obtained results demonstrate that the cone tip-diameter has an insignicant effect on the collection efciency (the cut-off diameter) and the pressure drop. The simii
ulation results agree well with the published experimental results and the mathematical models trend. The effect of the cyclone inlet dimensions on the performance and the ow eld pattern has been investigated computationally using the Reynolds stress turbulence model (RSM) for ve cyclone separators. The maximum tangential velocity in the cyclone decreases with increasing the cyclone inlet dimensions. Increasing the cyclone inlet dimensions decreases the pressure drop. The cyclone cut-off diameter increases with increasing cyclone inlet dimension. Consequently, the cyclone overall efciency decreases due to weakness of the vortex strength. The effect of changing the inlet width b is more signicant than the inlet height a, especially for the cut-off diameter. The optimum ratio of inlet width to inlet height b/a is from 0.5 to 0.7. The effect of the vortex nder dimensions (both the diameter and length) on the performance and ow eld pattern has been investigated computationally using the large eddy simulation (LES) for nine cyclone separators. The maximum tangential velocity in the cyclone decreases with increasing the vortex nder diameter. Whereas, a negligible change is noticed with increasing the vortex nder length. Increasing the vortex nder length makes a small change in both the static pressure, axial and tangential velocity proles. However, decreasing the vortex nder diameter gradually changes the axial velocity prole from the inverted W to the inverted V prole. Decreasing the cyclone vortex nder diameter increases the maximum tangential velocity. The maximum tangential velocity approaches asymptotically 1.589 times the inlet velocity when decreasing the vortex nder diameter. The Euler number (dimensionless pressure drop) decreases with increasing the vortex nder diameter. Increasing the vortex nder length slightly increases the Euler number. The Stokes number increases with increasing the vortex nder diameter and slightly increases as the vortex nder length is increased. The effect of the cyclone height (both the barrel and cone) on the performance and ow eld pattern has been investigated computationally for six cyclone separators. The maximum tangential velocity in the cyclone decreases with increasing the cyclone (barrel or cone) height. Increasing the barrel height, makes a small change in the axial velocity, whereas increasing the cone height changes it considerably. Increasing the cyclone (barrel or cone) height decreases both the pressure drop and the cut-off diameter. The changes in the performance beyond h/D = 1.8 are small at constant cone height, whereas the performance improvement stops after hc /D = 4.0 (Ht /D = 5.5) at constant barrel height where h is the barrel height, hc is the cone height, Ht is the total cyclone height and D is the barrel diameter. iii
The effect of changing the cone height on the ow pattern and performance is more signicant than that of the barrel height. The CFD model was used to predict the pressure drop and the collection efciency of a range of cyclone geometries based on Stairmands highefciency design. These predictions were used to obtain an algebraic equation that relates the performance of a cyclone to its design and a limited set of dimensionless quantities (Euler number and Stokes number). This approach towards predicting cyclone performance by varying many geometrical parameters has not been reported before. To obtain new optimized cyclone separators, several optimization studies have been conducted in this thesis. Both the response surface methodology (RSM) and the radial basis function neural network (RBFNN) have been used as meta-models. Three different sources of data have been used to t the second order polynomial in case of RSM and for training the RBFNN. These data come from analytical models, experimental measurements and CFD simulations. Two optimization techniques have been used to optimize the cyclone geometry for minimum pressure drop, namely, the NelderMead and the genetic algorithms techniques. To handle the bi-objective optimization problem (both the pressure drop and the cut-off diameter), two approaches have been applied, the desirability function and NSGA-II techniques. All the new optimized cyclones obtained either for single objective or for bi-objective problems exhibit better performance than the Stairmand design. Moreover, a new correlation between the Stokes number and the Euler number is obtained. The new correlation can be used to estimate the Stokes number if the Euler number is known.
iv
Acknowledgments
First and foremost, I would like to gratefully acknowledge the enthusiastic supervision of my promoter Prof. Chris Lacor who gave me the opportunity to do a PhD under his guidance. I particularity thank him for our weekly technical discussions, which had a major inuence on this thesis. I am indebted to him for showing great condence in me and always pushing me to achieve greater heights, as well as for granting me sufcient freedom to pursue my own ideas. I can say for sure that the past years at VUB have been the most productive days of my learning. I also thank all the members of my thesis committee: Prof. Johan Deconinck, Prof. Rik Pintelon, Prof. Gunther Steenackers, Prof. Gert Desmet, Prof. Harry van den Akker and Prof. Herman Deconinck, whose constructive criticism and valuable suggestions improved the quality of this dissertation. I warmly thank the IT support of our system administrator Alain Wery. His support is invaluable for the research at our department. I greatly appreciate him for his good mood and everlasting patience through the perpetual stream of requests and computer problems coming towards him. I am yet to meet someone who is so patient and always ready to help others. Thank you very much, Alain! The support of our secretary Jenny Dhaes started even before I arrived in Belgium. She was there for help, starting from lling down my admission forms in Dutch, to organizing my PhD defense. Thanks a lot Jenny. A word of thanks should also go to Birgit Buys and my Egyptian colleague Mahmoud El-kafafy who helped me in printing the draft version. I am pleased to acknowledge my colleagues, Ghader Ghorbaniasl. The many discussions on mathematics and physics I have had with Ghader were always fruitful. Santhosh Jayaraju and Kris van den Abeele gave v
me the template that was used for this thesis, and in doing so, saved a lot of much needed time for me. I also enjoyed the scientic discussion with them. In this regard, I should mention Willem Deconinck as well. I taught the students a basic techniques in computer simulation course with him. He made it fun to do so with his pleasant mood and sense of humor, even though I was under the pressure of writing my thesis at that time. I would like to thank Willem once more for proofreading of some part of my thesis. At my rst days at VUB during which we were ofce mates, I have shared many laughs and a lot of joy with Mahdi Zakyani. I am pleased to acknowledge my present and former colleagues Dean Vucinic, Matteo Parsani, Patryk Widera, Xiadong Wang, Vivek Agantori, Florian Krause and Dinesh Kumar. We have nice discussions and good fun at the coffee corner. In addition, I am sincerely thankful to Prof. Momtaz F. Sedrak, Prof. Ahmed F. Helal, Prof. Mohammed M. Abdelrahman, Prof. Mohammed Fatouh and Prof. Samira Elshereef who played a major role in my scientic career. I have learned a lot from them during my Master and Bachelor studies. I consider them as good examples for Egyptian professors. All my thanks are given to the Egyptian community at Belgium for advices, support and continuous encouragement. Special thanks are given to Omar Ellabban, Sameh Sorror, Romany Abskharon, Ehab Khatab and Wael Mohammed for the help and advice they gave me during my stay here at Brussel, especially at my rst days at Belgium. I cannot forget to give all thanks to the spirit of my late parents who I am indebted with all my life. Lastly, and most importantly, my utmost gratitude is reserved to my dear wife and my two sons Omar and Ahmed for their patience and encouragement.
vi
Jury Members
President Prof. dr. ir. Johan Deconinck Vrije Universiteit Brussel
Vice-President
Prof. dr. ir. Rik Pintelon Vrije Universiteit Brussel
Secretary
Prof. dr. ir. Gunther Steenackers Vrije Universiteit Brussel
Internal Member
Prof. dr. ir. Gert Desmet Vrije Universiteit Brussel
External members
Prof. dr. ir. Harry van den Akker Delft University of Technology
Prof. dr. ir. Herman Deconinck Von Karman Institute
Promoter
Prof. dr. ir. Chris Lacor Vrije Universiteit Brussel
vii
viii
Contents
1 Introduction 1.1 Overview of dust collectors . . . . . . . . . . . . . . 1.2 Cyclone separators: types and principals . . . . . 1.2.1 Advantages and disadvantages of cyclones 1.2.2 Applications . . . . . . . . . . . . . . . . . . 1.2.3 Principals of cyclonic separation . . . . . . 1.2.4 Factors affecting the cyclone performance . 1.3 Motivation of this work . . . . . . . . . . . . . . . . 1.4 Outline of the thesis . . . . . . . . . . . . . . . . . 2 Literature Review 2.1 Classication of study approaches . 2.2 Mathematical models . . . . . . . . . 2.3 Experimental methods . . . . . . . . 2.4 Computational uid dynamics (CFD) 2.5 Discrepancy in the previous studies 2.5.1 The cone tip diameter . . . . . 2.5.2 The dust outlet geometry . . . 2.5.3 The inlet dimensions . . . . . 2.5.4 The vortex nder dimensions 2.5.5 The cyclone heights . . . . . . 2.5.6 Previous optimization studies 2.6 Summary and research plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 2 5 6 7 10 11 13 15 15 16 17 17 18 18 19 19 20 21 21 23 25 25 28 29 31 32 37
3 Governing Equations 3.1 Turbulence . . . . . . . . . . . . . . . . . . . . . . 3.2 The governing equations for the gas phase . . . 3.2.1 Reynolds averaged Navier Stokes (RANS) 3.2.2 Reynolds stress model (RSM) . . . . . . . 3.2.3 Large eddy simulation (LES) . . . . . . . 3.3 Discrete phase modeling . . . . . . . . . . . . . . ix
3.3.1 Governing equations for the particles . . . . . . . . . 3.3.2 Modeling the particle phase . . . . . . . . . . . . . . . 3.3.3 Stochastic trajectory approach . . . . . . . . . . . . . 4 Sensitivity Analysis of Geometrical Parameters 4.1 Sensitivity analysis . . . . . . . . . . . . . . . . . 4.1.1 Response surface methodology (RSM) . . 4.1.2 Design of experiment (DOE) . . . . . . . . 4.1.3 Analysis of response surfaces . . . . . . . 4.1.4 Conclusions . . . . . . . . . . . . . . . . . 4.2 The cone-tip diameter . . . . . . . . . . . . . . . . 4.2.1 Numerical simulation . . . . . . . . . . . 4.2.2 Results and discussion . . . . . . . . . . . 4.2.3 The ow pattern in the three cyclones . . 4.2.4 Comparison with mathematical models . 4.2.5 Conclusions . . . . . . . . . . . . . . . . . 4.3 The dust outlet geometry . . . . . . . . . . . . . . 4.3.1 Numerical simulation . . . . . . . . . . . 4.3.2 Results . . . . . . . . . . . . . . . . . . . . 4.3.3 Conclusions . . . . . . . . . . . . . . . . . 4.4 Closure . . . . . . . . . . . . . . . . . . . . . . . . 5 The Vortex Finder Dimensions 5.1 Introduction . . . . . . . . . . . . . . . . . . 5.2 Numerical settings . . . . . . . . . . . . . . 5.2.1 Conguration of the tested cyclones 5.2.2 Solver settings . . . . . . . . . . . . . 5.2.3 Boundary conditions . . . . . . . . . 5.2.4 Grid independency study . . . . . . . 5.3 Results and discussions . . . . . . . . . . . 5.3.1 The axial variation . . . . . . . . . . 5.3.2 The ow pattern . . . . . . . . . . . . 5.3.3 The radial variation . . . . . . . . . . 5.3.4 The cyclone performance . . . . . . . 5.4 Conclusions . . . . . . . . . . . . . . . . . . 6 The Inlet Dimensions 6.1 Introduction . . . . . . . . . . . . . . . . 6.2 Numerical settings . . . . . . . . . . . . 6.2.1 Conguration of the ve cyclones 6.2.2 Boundary conditions . . . . . . . 6.2.3 Selection of the time step . . . . . 6.2.4 CFD grid . . . . . . . . . . . . . . x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37 40 41 45 45 46 47 48 49 52 54 63 64 70 71 72 73 76 87 88
89 . 89 . 92 . 92 . 92 . 93 . 93 . 95 . 95 . 96 . 99 . 102 . 106 . . . . . . 109 109 111 111 111 111 112
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 The axial variation of the ow properties 6.3.2 The ow pattern . . . . . . . . . . . . . . . 6.3.3 The cyclone performance . . . . . . . . . . 6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . 7 The Cyclone Height 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . 7.2 Numerical settings . . . . . . . . . . . . . . . . . 7.2.1 Conguration of the tested cyclones . . . 7.2.2 Boundary conditions . . . . . . . . . . . . 7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 The axial variation . . . . . . . . . . . . . 7.3.2 The radial variation . . . . . . . . . . . . . 7.3.3 The ow pattern . . . . . . . . . . . . . . . 7.3.4 The performance . . . . . . . . . . . . . . 7.3.5 The cone height versus the barrel height 7.4 Conclusions . . . . . . . . . . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
113 113 115 120 124 127 127 128 128 128 130 130 131 135 137 141 146 147 147 149 150 154 158 158 164 169 179 185 186 187 194 202 204 215 215 220 231 235 239
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
8 Optimization 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Single-objective using MM model . . . . . . . . . . . . . 8.2.1 CFD comparison between the two designs . . . . 8.2.2 Conclusions . . . . . . . . . . . . . . . . . . . . . 8.3 Single-objective using RBFNN . . . . . . . . . . . . . . 8.3.1 Radial basis function neural networks (RBFNN) 8.3.2 Evaluation of different mathematical models . . 8.3.3 Design of experiment (DOE) . . . . . . . . . . . . 8.3.4 CFD Comparison between the two designs . . . 8.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . 8.4 Multi-objective optimization using GA . . . . . . . . . . 8.4.1 Articial neural network (ANN) approach . . . . 8.4.2 Single objective optimization . . . . . . . . . . . 8.4.3 Optimal cyclone design for best performance . . 8.4.4 Conclusions . . . . . . . . . . . . . . . . . . . . . 8.5 Multi-objective optimization using CFD data . . . . . . 8.5.1 Design variables and approaches . . . . . . . . . 8.5.2 The desirability function . . . . . . . . . . . . . . 8.5.3 Articial neural network (ANN) approach . . . . 8.5.4 Optimization Using Genetic Algorithms . . . . . 8.5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . xi
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
9 Conclusions and Future Directions 9.1 Conclusions . . . . . . . . . . . . . . . . . . . . 9.1.1 The most signicant geometrical factors 9.1.2 The impact of geometry . . . . . . . . . 9.1.3 Optimization . . . . . . . . . . . . . . . . 9.1.4 Multi-objective optimization . . . . . . . 9.2 Future Directions . . . . . . . . . . . . . . . . . A Mathematical models A.1 General assumptions . . . . . . . . . . . . . . . A.2 Barth model . . . . . . . . . . . . . . . . . . . . A.3 The Muschelknautz method of modeling (MM) A.4 Stairmand model for pressure drop . . . . . . . A.5 Purely empirical models for pressure drop . . . A.6 Iozia and Leith model for the cut-off diameter A.7 Rietema model for cut-off diameter . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
243 243 243 244 245 246 248 251 251 252 256 259 260 261 262 263 263 266 268 268 272 281
B Optimization Techniques B.1 Nelder-Mead . . . . . . . . . . . . . . . . . . . . . . . B.2 Genetic algorithms (GA) . . . . . . . . . . . . . . . . B.2.1 Description of the genetic algorithm process B.2.2 Genetic operators . . . . . . . . . . . . . . . . B.3 Multi-objective optimization . . . . . . . . . . . . . . Bibliography
xii
Nomenclature
Upper-Case Roman Ai AR Bc CD Cin D Dx Eu Fdrag Frx HCS Ht K K Li Le Ln P P Pr Qin R R2 Rb Rij Rep ReR Rm Rx inlet cross sectional area [m] total inside area of the cyclone [m2 ] cyclone cone-tip diameter [m] particle drag coefcient inlet dust concentration, [kg/m3 ] cyclone barrel diameter [m] cyclone vortex nder diameter [m] Euler number [-] drag force [N] Froude number (inertia force / gravitational force) [-] length of the control surface [m] cyclone total height [m] uctuating kinetic energy [m2 /s2 ] vortex nder entrance factor distance between the inlet section and cyclone center [m] distance between the outlet section and the barrel top [m] cyclone natural length [m] uctuating kinetic energy production [m2 /s3 ] mean pressure [N/m2 ] Prandtl number [-] gas volume ow rate [m3 /s] cyclone radius [m] coefcient of multiple determination dust outlet radius [m] Reynolds stress tensor [m2 /s2 ] Reynolds number based on the relative particle velocity [-] cyclone body Reynolds number [-] geometric mean radius [m] vortex nder radius [m] xiii
S Sg Stk50 V
vortex nder length [m] geometrical swirl number [-] Stokes number at the cut-off diameter [-] cyclone volume [m3 ]
Lower-Case Roman a b d50 dp f fair fr fsm gi h hc k ks m m p p t tres ui ui u i upi vx vzw x x50 xi cyclone inlet height [m] acceleration [m/s2 ] cyclone inlet width [m] cut-off diameter [m] particle diameter [m] total friction factor [-] gas friction factor [-] friction factor due to wall roughness [-] friction factor for smooth wall [-] acceleration due to gravity in i direction [m/s2 ] barrel height [m] cone height [m] turbulent kinetic energy [m2 /s2 ] wall relative roughness of the cyclone wall [m] mass [kg] dust mass ow rate [kg/s] static pressure [N/m2 ] ow physical time [s] ow average residence time [s] ow velocity component in i direction [m/s] mean velocity [m/s] uctuating velocity component in i direction [m/s] particle velocity in i direction [m/s] mean gas velocity through the vortex nder [m/s] wall axial velocity [m/s] particle diameter [ m] cut-off diameter [m] position [m]
Upper-Case Greek P Pbody Px lter width [m] pressure drop in the cyclone [N/m2 ] pressure drop in the cyclone body [N/m2 ] pressure drop in the vortex nder [N/m2 ]
xiv
Lower-Case Greek ij t t bulk p str ij ij gas moment-of-momentum ratio at inlet [-] ratio of inlet width to the cyclone radius [-] Kronecker delta [-] turbulence dissipation rate [m2 /s3 ] viscous dissipation [m2 /s3 ] dynamic viscosity [kg/(m s)] turbulent (eddy) viscosity [kg/(m s)] kinematic viscosity ( = /) [m2 /s] turbulent (eddy) kinematic viscosity (t = t /) [m2 /s] spatial separation [m] gas density [kg/m3 ] bulk density of the solid [kg/m3 ] particle density [kg/m3 ] bulk density of the strand layer at the walls [kg/m3 ] viscous stress tensor [N/m2 ] subgrid scale stress tensor [N/m2 ]
Subscripts
CS
g
in
p
w
at the control surface gas at the inlet surface particle properties angular (tangential component) close to the wall
Abbreviations ANOVA ANN CFD CFL DNS DOE DPM GA Analysis Of Variance Articial Neural Networks Computational Fluid Dynamics Courant-Friedrichs-Lewy number Direct Numerical Simulation Design Of Experiment Discrete Phase Modeling Genetic Algorithm
xv
GCI GEC LDA LES MM PIV RANS RBFNN RNG RSM RSM SGS
Grid Convergence Index Grade Efciency Curve Laser Doppler Anemometry Large Eddy Simulation Muschelknautz Method of modeling Particle Image Velocimetry Reynolds Averaged Navier-Stokes Radial Basis Function Neural Network Renormalization Group Response Surface Methodology Reynolds Stress turbulence Model Subgrid Scale Model
xvi
Chapter 1 Introduction
1.1 Overview of dust collectors
There are four principal types of industrial dust collectors [77] namely, inertial separators, fabric collectors, wet scrubbers and electrostatic precipitators. The selection of one type depends mainly on the particle size as shown in Fig. 1.1. The classication of different dust collectors is shown in Fig. 1.2. The inertial separators separate dust from gas streams using a combination of forces, such as centrifugal, gravitational, and inertial. These forces move the dust to an area where the forces exerted by the gas stream are minimal. The separated dust is moved by gravity into a hopper, where it is temporarily stored. The three primary types of inertial separators are settling chambers, bafe chambers, and centrifugal collectors (e.g., cyclone separator). A settling chamber consists of a large box installed in the ductwork. The sudden expansion at the chamber reduces the speed of the dust-lled airstream and heavier particles settle down. Settling chambers are simple in design and can be manufactured from almost any material. However, they are seldom used as primary dust collectors because of their large space requirements and low efciency. A practical use is as precleaners for more efcient collectors. Bafe chambers use a xed bafe plate that causes the conveying gas stream to make a sudden change of direction. Large-diameter particles do not follow the gas stream but continue into a dead air space and settle. Bafe chambers are used as precleaners for more efcient collectors. Fabric collectors are commonly known as baghouses. Fabric collectors use ltration to separate dust particulates from dusty gases. They are one 1
Chapter 1. Introduction
Settling chambers
Cyclone separator
Liquid scrubbers
Filters Electrostatic precipitators
10-3
10-2
10-1
100 101 Particle size (micron)
102
103
Figure 1.1: Suitable methods for removing particles from a gas stream [128]
of the most efcient types of dust collectors available and can achieve a collection efciency of more than 99% for very ne particulates. Dust collectors that use liquid are commonly known as wet scrubbers. In these systems, the scrubbing liquid (usually water) comes into contact with a gas stream containing dust particles. The greater the contact of the gas and liquid streams, the higher the dust removal efciency. The electrostatic Precipitators use electrostatic forces to separate dust particles from exhaust gases. A number of high-voltage, direct-current discharge electrodes are placed between grounded collecting electrodes. The contaminated gases ow through the passage formed by the discharge and collecting electrodes. The airborne particles receive a negative charge as they pass through the ionized eld between the electrodes. These charged particles are then attracted to a grounded or positively charged electrode and adhere to it [77].
1.2 Cyclone separators: types and principals

The gas cyclones belong to the type of centrifugal separators. A gas cyclone is a stationary mechanical device that utilizes centrifugal force to separate solid or liquid particles from a carrier gas. The ow enters near the top through the tangential inlet, which gives rise to an axially descending spiral of gas and a centrifugal force eld that causes the incoming particles to concentrate along, and spiral down, the inner walls of the cyclone separator. The collected particulates are allowed to exit out an underow 2
1.2. Cyclone separators: types and principals
Figure 1.2: Classication of dust collectors
Figure 1.3: Typical cyclone separator
pipe while the gas phase reverses its axial direction of ow and exits out through the vortex nder (gas outlet tube) [77]. Figure 1.3 shows a typical cyclone separator. The cyclone separator is one of the most efcient and robust dust separators. Its robustness results from lack of moving parts and the ability to withstand harsh operating environments. Moreover, cyclones are well suited for high pressure and temperature applications. Centrifugal collectors use cyclonic action to separate dust particles from the gas stream. In a typical cyclone, the dust gas stream enters tangentially forcing the ow into a spiral movement. The centrifugal force created by the circular ow throws the dust particles toward the wall of the cyclone. After striking the wall, the particles fall into a hopper located underneath. The most common types of centrifugal, or inertial, collectors in use today are single-cyclone separators and multiple-cyclone separators (multiclone). Single-cyclone separators create a dual vortex to separate the dust from the gas. The main vortex spirals downward and carries most of the heavier particles. The inner vortex, created near the bottom of the 4

cyclone, spirals upward and carries ner dust particles. Multiple-cyclone separators consist of a number of small-diameter cyclones, operating in parallel and having a common gas inlet and outlet. Multiclones operate on the same principle as cyclonescreating a main downward vortex and an ascending inner vortex.
1.2.1 Advantages and disadvantages of cyclones

Compared with the other separation devices, the cyclone separators advantages are [77]: the collected product remains dry and, normally useful. low capital investment and maintenance costs in most applications. very compact in most applications. can be used under extreme processing conditions, in particular at high temperatures and pressures and with chemically aggressive feeds. no moving parts. very robust. can be constructed from most any material suitable for the intended service including plate steel, casting metals, alloys, aluminum, plastics, ceramics, etc. can be equipped with erosion or corrosion resistant or particle repelling type liners, such as Teon. Internal surfaces may be electro polished to help combat fouling. can be fabricated from plate metal or, in the case of smaller units, cast in molds. can, in some processes, handle sticky or tacky solids with proper liquid irrigation. can separate either solids or liquid particulates; sometimes both in combination with proper design.
Some disadvantages of cyclones are [77]: low efciency for particle sizes below their cut-off diameter when operated under low solids-loading conditions. usually higher pressure loss than other separator types, including bag lters and low pressure drop scrubbers. subject to erosive wear and fouling if solids being processed are abrasive or sticky. can operate below expectations if not designed and operated properly. Although this problem, as well as the erosion and fouling problem mentioned above, is not unique to cyclones. 5
Figure 1.4: Applications of cyclone separator
1.2.2
Applications
Due to the mentioned advantages, cyclones have found application in virtually every industry where there is a need to remove particles from a gas stream. Figure 1.4 presents some examples of cyclones industrial applications with wide range of sizes, locations and applications. Today, cyclone separators are found in: ship unloading installations power stations spray dryers uidized bed and reactor riser systems (such as catalytic crackers and cockers) synthetic detergent production units food processing plants crushing, separation, grinding and calcining operations in the mineral and chemical industries fossil and wood-waste red combustion units (normally upstream of 6
a wet scrubber, electrostatic precipitator or fabric lter) vacuum cleaning machines dust sampling equipment Cyclones have also been used to classify solids on the basis of their characteristic such as their mass, density, size, or shape. Because of their simple construction and high reliability, cyclones are also used very effectively to separate two-phase gas-liquid mixtures, such as the entrained droplets exiting a venturi scrubber or other types of scrubber. Other examples include the removal of water droplets from steam generators and coolers and oil-mist from the discharge of air compressors. Likewise, they have been widely applied in process machinery to remove entrained oil and hydrocarbon droplets generated from spraying, injection, distillation, or most any process that results in the production of entrained droplets or a two-phase mixture. They have even been used as inlet devices to prevent foaming in gravity separation drums [77].
1.2.3 Principals of cyclonic separation

In centrifugal devices, the dust-laden gas is initially brought into a swirling motion. The dust particles are slung outward to the wall, and transported downward to the dust outlet by the downwardly directed gas ow near the wall. A sketch of a standard reverse-ow, cylinder-on-cone cyclone with a tangential, slot-type inlet is shown in Fig. 1.5. For the standard, reverse-ow cyclone, (with a so-called slot type of entry) the swirling motion is brought about by designing the inlet in such a manner that it forces the gas to enter the unit on a tangent to the inner body wall. The inlet is normally of rectangular cross section. As the gas swirls, it moves axially downwards in the outer part of the separation space. In the conical part of the cyclone, the gas is slowly forced into the inner region of the cyclone, where the axial movement is upwardly directed. This ow pattern is often referred to as a double vortex: an outer vortex with a downwardly directed axial ow and an inner one with an upwardly directed ow. The gas exits the cyclone through the so-called vortex nder, which extends downward from the center of the roof. This outlet pipe goes by many different names, with vortex tube and dip-tube being the most common, aside from the vortex nder [77]. The particles in the inlet gas are slung outwards to the wall in the centrifugal eld, and are transported to the dust exit by the downwardly directed gas ow near the wall. Below more details of the ow pattern in the separation space will be given. The geometry of a cyclone with a slot type inlet is determined by the following eight dimensions as shown in Fig. 1.5: 7
Dx S h D b Ht a
hc
Bc
Figure 1.5: Sketches of a reverse-ow, cylinder-on-cone cyclone with a tangential inlet. The geometrical notation is indicated in the right sketch
1. 2. 3. 4. 5. 6. 7.
the body diameter (barrel diameter) D the total height of the cyclone (from roof to dust exit) Ht the vortex nder diameter Dx the vortex nder length (from the roof of the separation space) S the inlet height a the inlet width b the height of the conical section hc or the height of the cylindrical section h 8. the cone-tip diameter (dust exit diameter) Bc
1.2.3.1 Real vortex ow Swirling ow, or vortex ow, occurs in different types of equipment, such as cyclones, hydrocyclones, spray dryers and vortex burners [77]. Two basic types of swirling ows can be distinguished: 1. forced vortex ow, which is a swirling ow with the same tangential velocity distribution as a rotating solid body 2. free vortex ow, which is the way a frictionless uid would swirl. 8
Tangential velocity
Fo r
ced
Rankine vortex
vo r
tex
Free vort ex
Distance from center
Figure 1.6: The tangential velocity distribution in a real vortex [77]
The tangential velocity in such a swirl is such that the moment-ofmomentum of uid elements is the same at all radii. The tangential velocity distribution in a real swirling ow is intermediate between these two extremes. Now imagine rst that the swirling uid has an innite viscosity (behaves like a solid body). Hence, no shearing motion exists between uid layers at different radii. In this case, the uid elements at all radial positions are forced to have the same angular velocity which equals v /r where v is the tangential velocity. This is the forced vortex ow or solid-body rotation: v = r (1.1)
In the other extreme, if the swirling uid has no viscosity, the motion of a given uid element is not inuenced by the neighboring elements at smaller and larger radii. If in such a uid, we bring an element to a smaller radius, its tangential velocity will increase, since its moment-ofmomentum (mv r) will be conserved. Such a vortex is called a free or frictionless vortex. In such a ow, we have rv = C , with C a constant, so that: C (1.2) r This is the second basic swirl ow. A real swirling ow normally has a core of near solid-body rotation surrounded by a region of near loss-free rotation as sketched in Fig. 1.6. This is called a Rankine vortex. The ow and pressure distribution within cyclones is more easily understood if we make clear the relation between static and dynamic pressures; p and 1/2v 2 , respectively, with the density. The well-known Bernoulli v = 9
equation for steady ow of a frictionless, constant density uid, which can be derived from the Navier-Stokes equations, states that: 1 p + gh + v 2 = constant along a streamline 2 (1.3)
In this equation, we recognize the static and dynamic pressures (the latter is often called the velocity head) as the rst and third terms on the left-hand side. They have been divided by the uid density. This equation shows that static and dynamic pressures can be interchanged in the ow eld [77]. In areas where the velocity is high, the static pressure will be low and vice versa. It is especially important to appreciate this interdependence between static and dynamic pressure when dealing with swirling ows [77]. The left-hand side of Eq. 1.3 is sometimes called Bernoullis trinomial. The second term is unimportant relative to the two others when discussing gas cyclones, since the uid density is relatively low, and height differences not very large. In an actual ow situation, the uid is not frictionless. Frictional dissipation of mechanical energy will therefore cause Bernoullis trinomial to decrease in the ow direction, i.e. the trinomial is no longer constant, but decreases along a streamline. Frictionless ow is, nevertheless, a reasonably good approximation in the outer part of the swirl in a cyclone; Bernoullis trinomial does not change very much there [77].
1.2.4
Factors affecting the cyclone performance
Figure 1.7 indicates the possible factors affecting the cyclone performance and ow pattern. These factors can be sub-classied as follows: 1. Cyclone dimensions Cyclone diameter Inlet height Inlet width Vortex nder diameter Vortex nder length Cylinder height Cyclone total height Cone tip diameter
2. Particle properties Density Shape Diameter and distribution 10
1.3. Motivation of this work
Mass loading 3. Gas properties Velocity Density Viscosity Temperature Pressure
4. Other factors Wall roughness Shape of vortex nder Eccentricity of vortex nder
1.3 Motivation of this work

In spite of the fact that the use of cyclone separators is common in many industrial applications, an accurate prediction tool for their behaviors is still not available. The challenge of this work is therefore, a detailed study of the ow phenomena in cyclones and the design of an optimum cyclone separator (minimum pressure drop and maximum collection efciency). The cyclone performance parameters are governed by many operational parameters (e.g. the gas ow rate and temperature) and geometrical parameters. This study focuses only on the effect of the geometrical parameters on the ow eld pattern and on the performance of the tangential inlet cyclone separators. For that the specic goals of the work are the following: Determine the most signicant factors affecting the cyclone performance based on previous studies and statistical analysis of data. Study each (signicant) parameter separately to obtain more details on its effect on the ow eld pattern. To obtain the most efcient cyclone design for minimum pressure drop (using the most robust mathematical models, articial neural networks and CFD data sets). To obtain the most efcient cyclone design for best performance (minimum pressure drop and minimum cut-off diameter) using multiobjective optimization techniques. 11
12
Figure 1.7: Cause and effect plot for cyclone separator
1.4. Outline of the thesis

The study is limited to reverse ow gas cyclone separators operating at low mass loading.
1.4 Outline of the thesis

The thesis is organized in ten chapters. After an introduction and overview given in the previous sections, chapter 2 deals with the different study approaches of cyclone separators. In addition, an overview is presented of work reported in the literature on the application of CFD in cyclone modeling. Moreover, the discrepancy of the results from the previous studies has been discussed in details. Furthermore, a summary of the previous optimization studies is given. Chapter 3 presents the governing equations for the carrier (gas) and the discrete phase (solids) in detail. The application of response surface methodology and design of experiment statistical techniques to estimate the most signicant geometrical parameters are given in chapter 4. Furthermore, chapter 4 deals with the uncertainty of the signicant effect of cone tip diameter and the necessity of including the dust outlet geometry in the simulation domain. Chapter 5 presents the study of the effect of the vortex nder dimensions, whereas chapter 6 deals with the effect of the inlet dimensions. The effects of both the cone height and the barrel height on the ow eld and performance are discussed in chapter 7. The new optimized cyclones for minimum pressure drop and best performance using different techniques are analyzed in chapter 8. The used dataset are collected from different sources; mathematical models calculations, articial neural network models and CFD simulations data. Chapter 9 summarizes the main conclusions and some future directions. The thesis contains two appendices. Appendix A, deals with the details of eight mathematical models used to predict the cyclone performance parameters. The details of the two optimization techniques used in this thesis are given in appendix B.
13
14
Chapter 2 Literature Review

The most important parameter that affects the cyclone performance and ow pattern is the cyclone geometry. For reversed ow cyclones, there are seven geometrical parameters, viz. the inlet height a, the inlet width b, the vortex nder diameter (gas outlet tube diameter) Dx , the vortex nder length S , the cylindrical part height h, the cyclone total height Ht , and the cone-tip diameter Bc . All these dimensions are expressed in terms of barrel diameter D as shown in Fig. 2.1. The two performance indicators used are the pressure drop and the particle separation (collection) efciency. The latter is normally expressed as a grade efciency curve a graph of the collection efciency against the particle diameter. For low mass loading cyclone separators, the cut-off diameter x50 is usually given instead of grade efciency curves.
2.1 Classication of study approaches

There is a widespread literature on the effect of cyclone geometry on performance, using one or more of the four main approaches of study, which are: 1. Analytical methods (mathematical models), which can be classied into [194]: (a) theoretical and semi-empirical models (b) statistical models 2. Experimental measurements 3. Computational uid dynamics (CFD) simulations 4. Articial neural networks (ANN) approach Recently, optimization studies based on data available from one of the main four approaches have been performed. Also articial neural net15
Chapter 2. Literature Review
Figure 2.1: The cyclone separator dimensions
works become a tool to study the effect of cyclone geometry on performance.
2.2 Mathematical models

The theoretical models were developed by many researchers e.g., Shepherd and Lapple [157], Alexander [1], First [58], Stairmand [166], Barth [9], Avci and Karagoz [5], Zhao [192], Karagoz and Avci [90] and Chen and Shi [22]. These models were derived from physical descriptions and mathematical equations. They require a very detailed understanding of gas ow pattern and energy dissipation mechanisms in cyclones. In addition, due to using different assumptions and simplifying conditions, different theoretical or semi-empirical models can lead to signicant differences between predicted and measured results. Predictions by some models are sometimes twice the experimental values [172]. Since the rst application of aerocyclones in 1886 [3], theories for the estimation of both particle collection efciency and pressure drop of cyclone have been developed by many contributors using different methods with various simplifying assumptions. During the past 50 years, interest in particle collection and pressure theories has steadily increased [196]. The most widely used mathematical models for the cut-off diameter and pressure drop estimation are: Barth model [9] The Muschelknautz method of modeling (MM) [29, 77, 114116, 174, 175] 16
2.3. Experimental methods

Stairmand model [165] Shepherd and Lapple model [157] Casal and Martinez-Bent model [21] Ramachandran model [139] Iozia and Leith model [84] Rietema model [142] The interested reader can refer to appendix A for more details about these eight models.
2.3 Experimental methods

There are numerous experimental investigations performed on the cyclone separators. The majority of these studies used either laser doppler anemometry (LDA) or particle image velocimetry (PIV) to obtain the ow eld pattern. Some of the studies only measured the pressure drop and collection efciency without any details of the ow elds. For instance, Dirgo and Leith [43] measured the collection efciency and pressure drop for the Stairmand high efciency cyclone at different ow rates. Hoekstra et al. [75] measured the mean and uctuating velocity components for gas cyxD clones with different geometric swirl numbers (SG = D 4ab ) by means of the laser doppler anemometry technique. The experimental data shows the strong effect of the geometric swirl number on the mean ow characteristics, in particular with respect to vortex core size and the magnitude of the maximum tangential velocity. It is shown that the forced vortex region of the ow is dominated by the so-called precessing vortex core. Hoffmann et al. [76] investigated the effect of the cyclone length on the separation efciency and the pressure drop experimentally and theoretically by varying the length of the cylindrical segment of a cylinder-oncone cyclone. They found for cyclone lengths from 2.65 to 6.15 cyclone diameters, a marked improvement in cyclone performance with increasing length up to 5.5 cyclone diameters; beyond this length the separation efciency was dramatically reduced. For the interested reader, other experimental results on cyclones can be found in [36, 43, 66, 7476, 93, 102, 105, 121124, 127, 138, 152, 163, 184].
2.4 Computational uid dynamics (CFD)

Boysan et al. [14] presented the rst CFD investigation in the eld of cyclone separators. From that time, the CFD technique becomes a widely used approach for the ow simulation and performance estimation for cyclone separators. For example, Grifths and Boysan [68] computationally 17

investigated three cyclone samplers. They reported that the CFD predicted pressure drops are in excellent agreement with the measured data. The CFD modeling approach is also able to predict the features of the cyclone ow eld in great details, which providing a better understanding of the uid dynamics in cyclone separators [68]. Consequently, CFD approach is a reliable and relatively inexpensive method of examining the effects of a number of design changes. Moreover, this makes the CFD methods represent a cost-effective route for geometry optimization in comparison with the experimental approach. Another example, Gimbun et al. [64] successfully applied CFD to predict and to evaluate the effects of temperature and inlet velocity on the pressure drop of gas cyclones [194]. The successful application of CFD technique in different studies in cyclone separators has been reported by many researchers [e.g., 6, 11, 23, 49, 50, 52, 62 64, 146, 147, 186, 198]. Nevertheless, CFD is still more expansive in comparison with the mathematical models approach. The main reasons behind the cost of the CFD approach with respect to the mathematical methods are: 1. In essence, the CFD process requires expert intervention by an expert researcher at every stage (mesh generation, solver settings and post processing). 2. The license cost of the grid generator, solver and post processor. 3. The running cost especially for unsteady simulations which need also parallel processing. 4. CFD results always need (i) validation with experimental results (ii) perform the same simulation on different grids to be sure that the obtained results are grid independent.
2.5 Discrepancy in the previous studies

In this section, the effect of the geometrical parameters on the performance in terms of the two indicators; the pressure drop and the separation efciency (cut-off diameter) will be discussed briey based on the available literatures. More details will be presented in the subsequent chapters.
2.5.1
The cone tip diameter
Very little information is available on the effects of changing the cone bottom (tip) diameter, which determines the cone shape if other cyclone dimensions are xed [184]. Regarding this effect, discrepancies and uncertainties exist in the literature. Bryant et al. [17] observed that if the vortex touched the cone wall, particle re-entrainment occurred and efciency decreased, so collection efciency will be lower for cyclones with a small 18
2.5. Discrepancy in the previous studies

cone opening (cone tip diameter). While according to Xiang et al. [184], a cone is not an essential part for cyclone operation, although it serves the practical purpose of delivering collected particles to the central discharge point. However, Zhu and Lee [200] stated that the cone provides greater tangential velocities near the bottom for removing smaller particles.
2.5.2 The dust outlet geometry

Conventional cyclones always have a dustbin attached to the cone to collect the separated solid particles. When a gas ow stream enters the dustbin (closed at bottom), some of the ow will return into the cone and distribute some of the separated particles. This phenomena called re-entrainment and it will affect the separation efciency of the cyclone [138]. There has been little work concerning the dust outlet geometries [e.g., 40, 47, 78, 123]. Regarding this inuence, discrepancies and uncertainties exist in the literature. Xiang and Lee [186] reported that the dustbin connected to the cyclone should be incorporated in the ow domain as it affects the results obtained. On the other hand, numerous studies were performed without dustbin [e.g., 159, 178] with good matching with experimental results. Obermair et al. [123] performed cyclone tests with ve different dust outlet geometries to nd the inuence of the dust outlet geometry on the separation process. They showed that separation efciency can be improved signicantly by changing the dust outlet geometry, and they reported that further research is needed to clarify precise effects of dust outlet geometry. The effect of a dipleg (a vertical tube between the cyclone and the dustbin) was posed and investigated by several researchers [e.g., 78, 92].
2.5.3 The inlet dimensions

The effects of the inlet dimensions on the cyclone performance (pressure drop and cut-off diameter) have been reported in many articles. Casal and Martinez-Benet [21] proposed the following empirical formula for the dimensionless pressure drop (Euler number), Eu = 11.3 ab Dx
2
+ 2.33
(2.1)
implying proportionality with the square of the inlet area. On the other hand, Ramachandran et al. [139] proposed, Eu = 20 ab 2 Dx 19
S D H h Bc DD D 1/3
(2.2)

i.e., a linear relation with the inlet area. Iozia and Leith [84, 85] presented a correlation to estimate the cut-off diameter x50 and found proportionality to (ab)0.61 . The importance of the inlet dimensions becomes clear after the study of the natural length (or vortex length) by several researchers, e.g., Alexander [1]. The cyclone has two spiral motions, outer and inner. In the reverse ow cyclone, the outer vortex weakens and changes its direction at a certain axial distance Ln from the vortex nder [29]. This distance is usually called the turning length, natural length or vortex length of the cyclone. The inlet area is one of the relevant parameters inuencing the natural length. Alexander [1] found that Ln decreased proportionally to the inlet area but the opposite trend has been also reported [29]. Numerous studies have been performed for the effect of geometrical parameters on the ow pattern and performance [e.g., 15, 62, 102, 140, 184] while the effect of cyclone inlet dimensions remained largely unexplored. The articles investigating the effect of cyclone geometry report only briey on the effect of inlet section dimensions without sufcient details about the effects on the ow pattern and velocity proles. The new trend is to study the multi-inlet cyclone [e.g., 103, 187, 195].
2.5.4
The vortex nder dimensions
The vortex nder size is an especially important dimension, which signicantly affects the cyclone performance as its size plays a critical role in dening the ow eld inside the cyclone, including the pattern of the outer and inner spiral ows. Saltzman and Hochstrasser [151] studied the design and performance of miniature cyclones for repairable aerosol sampling, each with a different combination of three cyclone cone lengths and three gas outlet diameters. Iozia and Leith [84] optimized the cyclone design parameters, including the gas outlet diameter, to improve the cyclone performance using their optimization program. Kim and Lee [95] described how the ratio of the diameters of cyclone body D and the vortex nder Dx affected the collection efciency and pressure drop of cyclones, and proposed an energy-effective cyclone design. Moore and Mcfarland [111] also tested cyclones, with six different vortex nders, and concluded that the variation in the gas outlet diameter under the constraint of a constant cyclone Reynolds number produced a change in the aerodynamic particle cut-off diameter. Recently, Hoekstra [74] investigated the effect of gas outlet diameter on the velocity prole using 2-D axisymmetric simulations. Lim et al. [102] examined experimentally the effect of the vortex nder shape on the collection efciency at different ow rates but without any explanation on its effect of the ow eld pattern and velocity proles. Raou et al. [140] duplicated numerically the same study of Lim et al. 20
2.5. Discrepancy in the previous studies

[102] with limited details about the effect of the gas outlet diameter on the ow eld pattern and velocity prole.
2.5.5 The cyclone heights

Limited literatures are available for the effect of cyclone height. Zhu and Lee [200] have conducted detailed experiments on cyclones of different height and found that, the cyclone height can inuence considerably the separation efciency of the cyclones. However, they did not provide any information about the ow pattern or even explanation for the efciency results. Hoffmann et al. [76] investigated the effect of the cyclone length on the separation efciency and the pressure drop experimentally and theoretically. The cyclone performance improves with increasing length up to 5.5 cyclone diameters beyond this length the separation efciency was dramatically reduced. However, they did not present any contour plot or velocity prole to assist the explanation for the effect of cyclone height on performance. Recently, Xiang and Lee [186] have repeated the same study of Zhu and Lee [200] for the effect of cyclone height computationally via steady three-dimensional simulation using Reynolds stress turbulence model (RSM). They found that the tangential velocity decreases with increasing cyclone height, which is responsible for the lower separation efciency observed in long cyclones. The explanation of this behavior was not adequate. Moreover, no particle tracking study was presented.
2.5.6 Previous optimization studies

Due to the wide range of industrial applications of the cyclone separator, it was a matter of study for decades. However, the optimization studies on it is quite limited in literature. Moreover, many of these studies are not coherent studies. Ravi et al. [141] carried out a multi-objective optimization of a set of N identical reverse-ow cyclone separators in parallel by using the non-dominated sorting genetic algorithm (NSGA). Two objective functions were used: the maximization of the overall collection efciency and the minimization of the pressure drop. Non-dominated Pareto optimal solutions were obtained for an industrial problem in which 165 m3 /s of air was treated. In addition, optimal values of several decision variables, such as the number of cyclones and eight geometrical parameters of the cyclone, are obtained. Their study shows that the barrel diameter, the vortex nder diameter, and the number of cyclones used in parallel, are the important decision variables inuencing the optimal solutions. Moreover, their study illustrates the applicability of NSGA in solving multi-objective optimization problems involving gas-solid separations. The main drawbacks of 21

their study are: (1) They used the model of Shepherd and Lapple [157] for predicting the dimensionless pressure drop (Euler number). In the Shepherd and Lapple model, the Euler number depends on only three factors 2 (Eu = 16ab/Dx ) and they used it to optimize the seven geometrical parameters. (2) The barrel diameter, number of parallel cyclones and the gas velocity have been included into the optimization design space. Consequently, it is not devoted to the geometrical ratio. (3) They used many side constraints on the geometrical values (0.4 a/D S/D, 0.15 b/D (1 Dx /D)/2 if 0.5 Dx /D 0.6) these constraints prevent searching for the global optimization geometrical ratios for the seven geometrical parameters. (4) No table for the non-dominated Pareto front points are presented from which the designer can select certain geometrical ratio set (optimal solution). Swamee et al. [172] investigated the optimum values of the number of cyclones to be used in parallel, the diameter of cyclone barrel D and exit pipe Dx , when a specied ow rate of gas is to be separated from solid particles, and the cut diameter is already specied. They used Stairmand model for calculation of pressure drop and Gerrard and Liddle formula for the cut-off diameter [172] which is not a widely used model. Instead of handling two objective functions, they blended the two objective into a single objective problem which is not the suitable method to considering two conicting objectives (the pressure drop and cut-off diameter). Sakhani et al. [148] performed a multi-objective optimization of cyclone separators. First, they simulated many cyclones to obtain the pressure drop and the cut-off diameter and used articial neural network approach to obtain the objective function values. Finally, a multi-objective genetic algorithms are used for Pareto based optimization of cyclone separators considering two conicting objectives. However, the design variables are only four (instead of seven): the barrel height, the cone height, the vortex nder diameter and length. So they ignored the effect of inlet dimensions, which has been acknowledged by other researchers as signicant geometrical parameters for the cyclone ow eld and performance (cf. Elsayed and Lacor [50, 5254]). Moreover, they did not explain why they selected these particular parameters. Furthermore, they applied four side constraints on the four tested variables, which prevent searching for the global optimization. Pishbin and Moghiman [130] applied genetic algorithm for optimum cyclone design. They studied the seven geometrical parameters. The data used for optimization was obtained from 2-D axisymmetric simulations. However, the ow in the cyclone separator is 3-D unsteady. Instead of using multi-objective genetic algorithm (e.g., non-dominated sorting genetic algorithm II (NSGA-II) [34]) they used the weighted-sum genetic algorithm. In this technique, a weighting factor is assigned for each objective 22
2.6. Summary and research plan

function based on the user preference. The main shortages of the Pishbin and Moghiman [130] study are: (1) How to select the weighting factor, in scientic and engineering problems, it is a non-trivial task to nd the one solution of interest to the decision maker [26]. The decision makers weight (no matter how dened) could be greater than necessary as more acceptable solutions are missed. Optimizing mostly prot could lead to poor quality or reliability, not a good compromise [26]. The weighted-sum genetic algorithm usually does not nd all Pareto front points of interest. But this approach is a simple approach for handling multi-objective optimization problem. Another simple but better result can be obtained using the desirability function approach [54, 126]. (2) No table for the nondominated Pareto front points is presented from which the designer can select a certain geometrical ratio. Sakhani et al. [149] carried out a multi-objective optimization using the genetic algorithm technique to obtain the best vortex nder dimension (diameter and length) and shape (convergent and divergent). Four design variables have been investigated; the vortex nder diameter, angle, upperpart length and lower-part length of the vortex nder. They applied neural networks to obtain a meta-model for the pressure drop and collection efciency from CFD dataset. The main shortages of the Sakhani et al. [149] study are: (1) They used dimensional values instead of dimensionless, and applied side constraints, which prevent the optimization procedure from obtaining global optimization. (2) The selection of only the vortex nder dimension as the design variables and neglecting the interaction with the vortex nder diameter with the other dimensions, especially the inlet dimensions [52, 53].
2.6 Summary and research plan

After studying the existing literature on cyclones, the following conclusion can be drawn: The separation mechanism inside cyclone separators is not well understood yet, and needs more investigations. Nearly all published articles have no systematic and complete study for the effect of geometrical parameters on the ow eld and performance. In more details: the geometry parameters are not given as dimensionless numbers; the effect of a certain parameter is obtained with no knowledge about the effect of others or possible interactions; the results of different articles are sometimes in contradiction. In some cases, there are a lot of results about the effect of a certain parameter on the performance but on different dimensions (and not 23

on the same dimensionless ratio) and also sometimes at different operating conditions Some parameters have less interest compared with others like the effect of vortex nder shape and number of inlet sections. The particle tracking in almost all computational investigations are not well reported (or missing), and need more investigations Some articles give only the ow pattern results without a study of the effect of any parameter on the performance, or with limited discussion. From the remarks above, it is clear that more investigation is still needed about the ow patterns inside cyclone separators and how the geometrical parameters affect the performance. This guided us to construct the following research plan: 1. Undertake a systematic study for the effect of seven geometrical parameters to obtain the most signicant factors on the cyclone performance. This step can be subdivided into: (a) Application of design of experiment to obtain a table of runs, the performance parameters will be estimated using the most robust mathematical models. (b) Application of the response surface methodology to determine the most signicant factors and any possible interactions. 2. Once, the most signicant factors are selected, CFD simulations for each particular parameter will be performed to investigate in detail the effect of these parameters on the ow eld pattern and performance (pressure drop and cut-off diameter). 3. Optimization of the cyclone geometry, rst single-objective optimization for minimum pressure drop and then for best performance (multiobjective optimization for minimum pressure drop and minimum cutoff diameter). The required data for optimization can be obtained using mathematical models, articial neural networks data (based on experimental measurements available in literature) or CFD simulations.
24
Chapter 3 Governing Equations

3.1 Turbulence
Hinze [73] described turbulence as follows: Turbulence uid motion is an irregular condition of ow in which various quantities show a random variation with time and space coordinates so that statistically distinct average values can be discerned. Wilcox [183] explains that turbulence consists of a continuous spectrum of scales that vary from smallest to largest over several orders of magnitude. The idea of a series turbulent eddies is often used. He also stated that: A turbulent eddy can be thought of as a local swirling motion whose characteristic dimension is the local turbulence scale. These eddies overlap in space and the large ones carry the smaller ones. The conversion of energy in a turbulent ow follows a cascading process where the kinetic energy is transferred from the larger eddies to the smaller ones.
The energy cascade

The concept of the energy cascade has been introduced in 1922 by Richardson [132]. The idea is that the kinetic energy enters the turbulence at the largest scales of motion. This energy is then transferred by inviscid processes to smaller and smaller scales until the smallest scales. The energy is dissipated by viscous processes [35], Fig. 3.1. In 1941, Kolmogorov identied the smallest scales of turbulence to be those that now bear his name [35, 132]. In this concept, the turbulence can be considered to be composed of eddies of different sizes that overlap in space. The largest eddies are characterized by the length scale l which is comparable to the ow scale 25
Chapter 3. Governing Equations
Figure 3.1: Schematic representation of the energy cascade [32, 35]. is the Kolmogorov length scale, lDI is the length scale dividing the dissipation and the inertial subrange, lEI is the length scale dividing the energy-containing range and the inertial subrange, l0 is the length scale and L is the characteristic length [35].
(the large scales are of the order of the ow geometry), and a characteristic velocity u which is on the order of the turbulence intensity. If l and u are the length and velocity scales of the largest eddy, the time scale is derived as, l (3.1) u The large energy containing eddies give away their kinetic energy to slightly smaller-scale eddies with which the large scales interact. The process of kinetic energy transfer continues in a similar fashion until the smallest scale eddies are reached, where the frictional forces become so large that the kinetic energy is converted into internal energy. This process of energy transfer and dissipation is referred to as the energy cascade process. The scales at which the dissipation () takes place are the smallest scales, and are also referred to as the Kolmogorov scales. They can be estimated from the large-scale properties as follows, = = u2 u3 = l 26 (3.2)
3.1. Turbulence
Figure 3.2: Energy spectrum for a turbulent ow [183]
Since the processes of dissipation in the smallest scales are due to viscous forces, the properties of the smallest eddies can be estimated using the ow kinematic viscosity ( ) and the dissipation () itself. The length, velocity and time scales are given by:
l =
1/4
(3.3)
1/4
u = () =
(3.4) (3.5)
1/2
The turbulent length scale l is related to the wave number as = 2/l. The energy spectrum E () for a turbulent ow is as shown in Fig. 3.2. From dimensional analysis, the Kolmogorov -5/3 law characterizes the inertial subrange which is given by, E () = C 2/3 5/3 C is the Kolmogorov constant. 27 (3.6)
3.2 The governing equations for the gas phase

All uid motions (laminar or turbulent) are governed by a set of dynamical equations namely the continuity, momentum and the energy equation (Navier-Stokes equations),
(ui ) = 0 + t xi p ij (ui uj ) = + (ui ) + t xj xi xj (Hui ) = (E ) + t xi (ji uj qi ) xi
(3.7) (3.8) (3.9)
ui (x, t) represents the i-th component of the uid velocity at a point in space x and time t. p(x, t) is the static pressure. ij (x, t) are the viscous stresses. (x, t) is the uid density. E and H are the total energy and total enthalpy per unit mass. qi in Eq. 3.9 is the heat ux which is proportional to the temperature gradient. T qi = (3.10) xi where is the thermal conductivity. The Mach numbers associated with air ow in cyclone separators are very nominal, which allows the ow to be treated as incompressible. Furthermore, the air behaves as a Newtonian uid, in which case the viscous stresses are related to the incompressible uid motion using a property of uid, viscosity [86]. 1 ij = 2 sij skk ij 3 sij is the instantaneous strain rate tensor given by, 28 (3.11)
3.2. The governing equations for the gas phase
sij =
1 2
ui uj + xj xi
(3.12)
For incompressible ows, Eqs. 3.7 and 3.8 are simplied to the following form,
uj xj ui ui + uj t xj
= =
0 1 p 2 ui + xi xj xj
(3.13) (3.14)
In this thesis, the temperature effects are ignored and hence Eq. 3.9 is uncoupled from the continuity and momentum equations. The four main numerical procedures for solving the Navier-Stokes equations are the direct numerical simulation (DNS), the large eddy simulation (LES), the detached eddy simulation (DES) and the reynolds averaged navier stokes (RANS) approach. The most accurate approach is DNS where the whole range of spatial and temporal scales of turbulence are resolved. Since all the spatial scales, from the smallest dissipative Kolmogorov scales (l ) up to the energy containing integral length scales (l), are needed to be resolved by the computational mesh, the number of points required in one direction is of the order, N= l l (3.15)
The number of points required for a resolved DNS in three dimensions can be estimated as, 9/4 3 l ul = Re9/4 (3.16) N= l The number of grid points required for fully resolved DNS is enormously large, especially for high Reynolds number ows, and hence DNS is restricted to relatively low Reynolds number ows. DNS is generally used as a research tool for analyzing the mechanics of turbulence, such as turbulence production, energy cascade, energy dissipation, noise production, drag reduction, etc [86].
3.2.1 Reynolds averaged Navier Stokes (RANS)

When the ow is turbulent, it is convenient to analyze the ow in two parts, a mean (time-averaged) component and a uctuating component 29

[86], Ui P Tij = = = U i + ui P +p T ij + ij

Overline is a shorthand for the time average and in case of RANS, Ui Ui and ui =0. The above technique of decomposing is referred to as Reynolds Decomposition. Inserting this decomposition into the instantaneous equations and time averaging results in the Reynolds averaged Navier-Stokes equations (RANS).
U j xj U i U i + Uj t xj

= =
0 1 P 2U i ui uj + xi xj xj xj
(3.17)
(3.18)
ui uj in the last term of Eq. 3.18 represents the correlation between uctuating velocities and is called the Reynolds stress tensor. All the effects of turbulent uid motion on the mean ow are lumped into this single term by the process of averaging [86]. This will enable great savings in terms of computational requirements. On the other hand, the process of averaging generates six new unknown variables. Now, in total there are ten unknowns (3-velocity, 1-pressure, 6-Reynolds stresses) and only four equations (1-continuity, 3 components of momentum equation). Hence, we need six equations to close this problem. This is referred to as the Closure problem. Based on the way we close the Reynolds stress tensor, there are two main categories, namely the eddy viscosity models and the Reynolds stress model. The Reynolds stress tensor resulting from time averaging of the NavierStokes equations is closed by replacing it with an eddy viscosity multiplied by velocity gradients. This is referred to as the Boussinesq assumption. U j U i + xj xi
u i u j = t
(3.19)
where t is the turbulent (eddy) kinematic viscosity. In order to make Eq. 3.19 valid upon contraction because of Eq. 3.17, it should be rewritten as, 30
u i u j = t
U i U j + xj xi
2 + ij k 3
(3.20)
where ij is the Kronecker delta, ij = 1 if i = j and ij = 0 if i = j . k is the turbulent kinetic energy given by, k= 1 uu 2 i i (3.21)
The eddy viscosity is treated as a scalar quantity and is determined using a turbulent velocity scale v and a length scale l, based on the dimensional analysis.
t vl
(3.22)
There are different types of eddy viscosity models (EVM) based on the way we close the eddy viscosity. Algebraic or zero equation EVMs normally use a geometric relation to compute the eddy viscosity. In one equation EVMs, one turbulence quantity is solved and a second turbulent quantity is obtained from algebraic expression. These two quantities are used to describe the eddy viscosity. In two equation EVM models the two turbulent quantities are solved to describe the eddy viscosity. The interested reader can refer to [24, 177, 183] for more details. In the Reynolds stress models (RSM), an equation is solved for each Reynolds stress component as well as one length scale determining equation. However, RSMs are computationally much more demanding when compared to EVMs.
3.2.2 Reynolds stress model (RSM)

RSM is regarded as the most appropriate RANS turbulence model for cyclone ows [178]. Equation 3.18 can be written as [147], Ui Ui 1 P 2 Ui = + Rij + Uj t xj xi xj xj xj (3.23)
where Rij = u i uj is the Reynolds stress tensor. The RSM turbulence model provides differential transport equations for evaluation of the turbulence
31

stress components (Eq. 3.24). Uj Ui t Rij = Rij Rik + Rjk Rij + Uk k t xk xk xk xk xk 2 2 2 Rij ij K C2 Pij ij P ij C1 K 3 3 3 (3.24) where the turbulence production terms Pij are dened as [147]: Pij = Rik Uj 1 Ui , P = Pij + Rjk xk xk 2 (3.25)
With P being the uctuating kinetic energy production. t is the turbulent (eddy) kinematic viscosity; and k = 1, C1 = 1.8, C2 = 0.6 are empirical constants The transport equation for the turbulence dissipation rate, , is given as [98]: = + Uj t xj xj In Eq. (3.26), K = + t Ui 2 C 1 Rij C 2 xj K xj K (3.26)
1 u u is the uctuating kinetic energy, and is the 2 i i turbulence dissipation rate. The values of constants are = 1.3, C 1 = 1.44 and C 2 = 1.92.
3.2.3
Large eddy simulation (LES)
The large eddy simulation technique was developed based on an implication from Kolmogorovs theory of self-similarity that the large eddies of the ow are dependent on the geometry while the smaller scales are more universal [86]. Hence, the big three-dimensional eddies which are dictated by the geometry and boundary conditions of the ow involved are directly calculated (resolved) whereas the small eddies which tend to be more isotropic are modeled. Large eddy simulation (LES) treats the large eddies more exactly than the small ones. In LES, the large scales in space are calculated directly as illustrated in Fig. 3.3. The LES technique is based on a separation between large and small scales. A grid size rst has to be determined. Those scales that are of a characteristic size greater than the grid size are called large or resolved scales, and others are called small or subgrid scales. The subgrid scales are included by way of a model called the subgrid model [154]. An 32

elaborate explanation on LES can be found in several text books such as [132, 150, 183].
Figure 3.3: Decomposition of the energy spectrum [154]
3.2.3.1 LES equations In case of RANS, the instantaneous continuity and momentum equations (Eq. 3.7 and 3.8) are time averaged to obtain a steady form of the averaged equations (Eq. 3.17 and 3.18). In case of LES, instead of time-averaging, the instantaneous time-dependent equations are ltered. Filtering is a method that separates the resolvable scales from the subgrid scales. Filtering can be performed in either wave number space or the physical space. The lter cut-off should lie somewhere in the inertial range of the spectrum (Fig. 3.2). In nite volume methods, box lters are always used because the nite volume discretization itself implicitly provides the ltering operation. One of the earliest volume average box lters was given by Deardorff [33].
(X, t) = =
1 3
x0.5x x0.5x
y 0.5y y 0.5y
z 0.5z z 0.5z
(, t)ddd
(3.27) (3.28)
+ s
In the above equation, denotes the resolvable scale ltered variable and s denotes the sub-grid scale uctuation. is the lter width given by = (xy z )1/3 . Leonard [100] dened a generalized lter as a convolution integral which is given by,
(X, t) =
G(X ; ) (, t) d3 33
(3.29)

G is the lter function that determines the scale of resolved eddies. The lter function is normalized by requiring that,
G(X ; ) d3 = 1
(3.30)
The lter function in terms of the volume average box lter (Eq. 3.27) can be written as, 1/3 , |x | < x/2 0, otherwise
G(X ; ) =
Finally, the decomposition of the ow into a ltered part and a sub-grid part looks like, Ui P Tij = = = Ui + u s i P + ps
s Tij + ij
The hat operator . in the above equations represents the ltering operation as opposed to the time-averaging in case of RANS. Moreover, contrary to RANS, where the average of uctuations is zero, in LES, Ui = Ui [150] and us = 0. Further details on the ltering methods can be found in [32, 97, 132, 183]. Inserting the above decomposition into the instantaneous equations results in the following ltered Navier-Stokes equations,
Uj xj Ui Ui + Uj t xj
= =
0
sgs ij 1 P 2 Ui + xi xj xj xj
(3.31) (3.32)
sgs where ij are the sub-grid scale stresses.
3.2.3.2 SGS modeling From the energy cascade, explained in the beginning of this chapter, it is apparent that the energy transfer occurs from the bigger scales to the 34

smaller scales. Hence, the main purpose of an SGS model is to represent the energy sink [86]. The representation of the energy cascade is an average process. However, locally and instantaneously the transfer of energy can be much larger or much smaller than the average. Moreover, there is also the phenomenon of energy backscatter in the opposite direction [129]. Ideally speaking, SGS models should actually account for all these phenomena. However, if the grid scale is much ner than the dominant scales of the ow, even a crude SGS model will result in good predictions of the behavior of the dominant scales [86].
sgs The sub-grid scale stresses ij in Eq. 3.32 are given by,
sgs ij = Ui Uj Ui Uj
(3.33)
By using the denition of ltering as given by Eq. 3.28 we can further work sgs out ij as,
sgs ij s Ui Uj (Ui + us i )(Uj + uj ) s s s us i uj + (Ui uj Uj ui ) + Ui Uj Ui Uj Reynolds Crossterm Leonard
= =
(3.34) (3.35)
sgs ij
Leonard [100] shows that the Leonard stresses can signicantly drain energy from the resolvable scales and they can be directly computed. On the other hand, Wilcox [183] mentions that Leonard stresses are of the same order of magnitude as the truncation error when a nite-difference scheme of second-order accuracy is used, and thus it is implicitly represented. The cross-term stresses are dispersive in nature and largely account for the backscatter effects. Modeling them with a purely dissipative model such as Smagorinsky would be in conict because of its dispersive nature [97]. In many applications, it is assumed that the Leonard and cross-term stresses can be neglected, and only the Reynolds stresses remain to be modeled. It is the same case in the present work [86]. The interested reader is referred to Sagaut [150] for the detailed review of various SGS models available in literature. Smagorinsky model One of the simplest SGS model is the Smagorinsky model [161]. The unknown subgrid-scale stresses are modeled employing the Boussinesq assumption as in the case of RANS. The subgrid-scale stress are related to 35

the eddy viscosity as follows, 1 ij kk ij = t 3 The eddy-viscosity is modeled as, t = L 2 s 2Sij Sij (3.37) Ui Uj + xj xi (3.36)
where Ls is the length-scale for the sub-grid scale and is given by Cs V 1/3 , where V is the computational cell volume. It is interesting to note that the length scale is now the lter width rather than the distance to the closest wall as in RANS. Cs is a constant which is taken to be 0.17. The only disadvantage of the Smagorinsky model is the constant Cs , which is not really a constant, but is ow dependent. It is found to vary between 0.065 [110] and 0.3 [89]. In the dynamic version, which was rst proposed by Germano et al. [60], Cs is dynamically computed based on the information provided by the resolved scales of motion. The specication of Ls as Cs V 1/3 is not justiable in the viscous wall region as it incorrectly leads to a nonzero turbulent shear-stress at the wall. In order to rectify this, Moin and Kim [110] use a Van Driest damping function to specify the length scale as, Ls = Cs V 1/3 1 exp y+ A+ (3.38)
where y + = u d/ is the non-dimensional distance from wall, u is the wall shear stress velocity, d is the distance to the nearest wall and A=25 is the Van Driest constant. The above-described SGS model is a standard version as dened in Smagorinsky [161]. The LES simulations in the present thesis are performed employing the Fluent ow solver. The Smagorinsky model implemented in Fluent deviates slightly from the standard version in the following ways [86], The length-scale for the sub-grid scale is computed as min(d, Cs V 1/3 ). is the von Karman constant (typically a value of 0.41 is used), d is the distance to the closest wall. d is indeed one of the rst mixing length models in the literature to handle the turbulent viscosity and was proposed by Prandtl [133]. Van Driest damping is basically an improved version of Prandtls mixing length model. Both the Prandtl and the Van Driest model are algebraic and from the zero-equation models category. 36
3.3. Discrete phase modeling
The constant Cs in Fluent is taken to be 0.1 instead of 0.17 as was originally proposed. The value of 0.17 for Cs was originally derived for homogeneous isotropic turbulence in the inertial subrange. However, this value was found to cause excessive damping of large-scale uctuations in transitional ows near solid boundaries, and has to be reduced in such regions [59]. A Cs value of around 0.1 has been found to yield the best results for a wide range of ows, and is the default value in Fluent. Dynamic Smagorinsky-Lilly model Germano et al. [60] and subsequently Lilly [101] conceived a procedure in which the Smagorinsky model constant Cs is dynamically computed based on the information provided by the resolved scales of motion [59]. The dynamic procedure thus obviates the need for users to specify the model constant Cs in advance. The Smagorinsky model constant is dynamically computed instead of given as an input to the solver, but clipped to zero or 0.23 if the calculated model constant is outside this range to avoid numerical instabilities [59]. The second advantage of the dynamic SmagorinskyLilly model over the Smagorinsky model is the treatment near the wall. In the dynamic Smagorinsky-Lilly model, a damping function for the eddy viscosity near the wall is not required, since the model constant goes to zero in the laminar region just near the wall [55, 113].
3.3 Discrete phase modeling

3.3.1 Governing equations for the particles
Based partly on the physical properties of dust particles and partly on the mathematical modeling effort required, there are certain reasonable assumptions made to describe the particles transport in a uid medium [86]. The major simplifying assumptions are as follows, The particles are assumed to be spherical. The ratio of particle to uid density is very large: The density of the dust particle is much higher when compared to the uid medium which is air. Drag force is the dominant force: This is a direct result of the previous assumption. Since the density of the particles are much higher than the density of the uid medium, several forces such as the lift force, Basset force and buoyancy force can be readily discarded as they have negligible effect on the particles transport [86]. 37
p/e
Particles enhance production Negligable effect on turbulence Particles enhance dissipation
102
0
10
10-2 10-4 10-7

One-Way Coupling
Fluid
10-5
Two-Way Coupling
10-3
10-1
Four-Way Coupling
p
Particles
Particles Fluid
Particles
Fluid
Particles
Dilute suspension
Dense suspension
Figure 3.4: Map for particle-turbulence modulation [44]. p is the ratio of particles volume to the volume occupied by particles and uids. p is the particle response time, p = p d2 /(18), where p is the particle density, d is the particle diameter, is the uid viscosity. e is the turnover time of large eddy (time scale= l/u) [44].
One-way coupling: The phenomenon of mutual mass, momentum and energy transfer between the phases is termed as coupling. Elghobashi [44] proposed a map of regimes of interactions between particles and uid turbulence as shown in Fig. 3.4. For values of dispersed-phase volume fraction less than 106 , particles have negligible effects on turbulence and this is termed as one-way coupling. The volume fraction of dust particles we are dealing with in the present thesis is much less than 106 and hence one-way coupling is assumed. In the second regime which lies between 106 103 , the existence of particles can augment the turbulence if the ratio of the particle response time to the turnover time of a large eddy is greater than unity, or can attenuate turbulence if the ratio is less than unity. This interaction is called two-way coupling. In the third regime where the volume fractions are greater than 103 , in addition to two-way coupling between particles and turbulence, particle collisions take place and hence this regime is termed as four-way coupling.
38
Incorporating all the above assumptions, the Lagrangian equations governing the particle motion can be written as [86]: dxp dt dup dt
= =
up Fd (u up ) + gx (p ) p
(3.39) (3.40)
xp is the particle position, gx is the gravitational force, and p are the density of the uid and the particle respectively. Generally, the particle moves with a different velocity than the uid at any given point. The difference in uid velocity (u) and the particle velocity (up ), termed as the slip velocity (u up ), leads to an unbalanced pressure distribution as well as viscous stresses on the particle surface which yields a resulting force called drag force. In Eq. 3.40, the term Fd (u up ) is the drag force per unit particle mass. Fd is given by [86]: Fd = 1 Cd Rep p 24 (3.41)
where p is the particle relaxation time given by, p = Laws of drag coefcient The drag coefcient Cd is a function of particle Reynolds number (Rep ). Various experimentally based empirical correlations for the drag coefcient based on Rep are available in the literature. The Reynolds number of the particle is dened as: Rep = dp |u u p | (3.43) p d2 p 18 (3.42)
In Fluent, the drag coefcient for spherical particles is calculated by using the correlations developed by Morsi and Alexander [112]. It is given by, Cd = a1 + a2 a3 + Rep Re2 p (3.44)
where a1 , a2 and a3 are constants that apply to smooth spherical particles in a stipulated range of Rep as given in Table 3.1. 39

Table 3.1: Drag coefcient parameters used for Morsi and Alexander [112]
Rep < 0 .1 0 .1 < 1 .0 1 < 10.0 10.0 < 100.0 100.0 < 1000.0 1000.0 < 5000.0 5000.0 < 10000.0 10000.0 < 50000.0
a1 0 3.69 1.222 0.6167 0.3644 0.357 0.46 0.5191
a2 24 22.73 29.1667 46.5 98.33 148.62 -490.546 -1662.5
a3 0 0.0903 -3.8889 -116.67 -2778 -4.75 57.87 5.4167
3.3.2
Modeling the particle phase
Coming to the fundamental mathematical modeling of two-phase ow, the two most widely used approaches are the Eulerian continuum approach and the Lagrangian trajectory approach. Eulerian continuum approach In an Eulerian approach, the particles are treated as a second uid which behaves like a continuum and the equations are developed for average properties of the particles. For example, the particle velocity is the average velocity over an averaging volume. This approach is most suitable when one requires a macroscopic eld description of dispersed phase properties such as pressure, mass ux, concentration, velocity and temperature. Eulerian approach is more suitable for simulating large-scale particle ow processes. However, this approach requires sophisticated modeling in order to describe the key effects and phenomena found in industrial processes [31, 86]. Lagrangian trajectory approach A Lagrangian approach is useful when the particle phase is so diluted that the description of particle behavior by continuum models is not feasible. The motion of a particle is expressed by ordinary differential equations in Lagrangian coordinates and are directly integrated to obtain individual tracks of particles [86]. To solve the Lagrangian-equation for a particular moving particle, the dynamic behavior of the gas phase (generally obtained by an Eulerian approach) and other particles surrounding this moving particle should be pre-determined. Since the particle velocity and the corresponding particle trajectory are calculated for each particle, this approach is more suitable to obtain the discrete nature of motion of particles. However, to obtain statistical averages with reasonable accuracy, a 40

large number of particles will have to be tracked. An advantage of using the Lagrangian approach is the ability to vary easily the physical properties associated with individual particles such as diameter, density, etc. Moreover, local physical phenomena related to the particle ow behavior can be easily probed. Hence, the Lagrangian models can also be used for validation, testing and development of continuum models [31]. The Lagrangian approach is classied into two types namely, deterministic trajectory methods and Stochastic trajectory methods based on the effect of turbulence. In a deterministic method, all the turbulent transport processes of the particle phase are neglected where as the stochastic method takes into account the effect of uid turbulence on the particle motion by considering instantaneous uid velocity in the formulation of the equation of particle motion. In the present thesis, the dust particles are modeled with a stochastic Lagrangian approach [86].
3.3.3 Stochastic trajectory approach

One of the most frequently used models is the eddy interaction model (EIM) rst introduced by Hutchinson et al. [83] and further developed by Gosman and Ioannides [67]. The instantaneous motion of particles governed by Equations 3.39 and 3.40 can be written in a general form as given below [86]. dx dt dup dt
= =
up 1 (u up ) + g p
(3.45) (3.46)
The instantaneous uid velocity u in the above equation is represented as the sum of the mean and uctuating velocity, u=U +u Assuming isotropic turbulence, we have, u 2 = v 2 = w 2 = 2 k 3 (3.48)
(3.47)
where k is the turbulent kinetic energy. Furthermore, it is assumed that the local velocity uctuations of the uid phase obey a Gaussian probability density distribution. Most stochastic models in practical use are de41

rived from the formulation of Gosman and Ioannides [67], which is given by, u =
2 k 3
(3.49)
where is a random number drawn from a normal probability distribution with zero mean and unit standard deviation. The minimal random number generator of Park and Miller with Bays-Durham shufe [135] is implemented [86]. The random number generator returns a uniform random derivative with zero mean and unit standard deviation. The chosen uctuation is referred to a turbulent eddy whose size (length scale) and life-time (time scale) is known. Sommerfeld et al. [164] proposed the following relations for eddy parameters, k 2 k 3
te le where ct was taken to be 0.3.
= =
ct te
(3.50) (3.51)
Figure 3.5 shows a 2-D schematic representation of an eddy inside a rectangular domain. At any given particle position (xp , yp ), the eddy parameters are rst evaluated based on the local uid kinetic energy and dissipation rate. The particle position (xp , yp ) is assumed to be located at the center of this hypothetical eddy. It is accepted that each eddy has its own uctuation u , which remains constant until the particle leaves this eddy. The particle leaving an eddy is based on a certain interaction time of the particle with the eddy. Once this interaction time is reached while time integration of particle equations, the particle is assumed to have left the present eddy. Now, based on the new position of the particle, new eddy pa rameters are calculated and a new uctuation u is assigned to this eddy. This procedure may be repeated for as many interaction times as required for the particle to traverse the required distance. If a statistically significant number of particles are tracked in this way, the ensemble averaged behavior should represent the turbulent dispersion induced by the prevailing uid eld [67]. The interaction time is the minimum of two time scales, one being a typical turbulent eddy lifetime and the other the crossing-time of the particle in the eddy [67]. tint = min(te , tc ) 42 (3.52)
Figure 3.5: 2-D illustration of a particle within an eddy [86]
The crossing-time is dened as, tc = p ln 1 le p |u up | (3.53)
where p is the particle relaxation time, le the eddy length scale and |u up | the magnitude of slip velocity. In circumstances where le /(p |u up |) > 1, Eq. 3.53 has no solution. This can be interpreted as the particle trapped by an eddy, in which case tint = te [67]. The mentioned eddy interaction model is needed only for RANS simulation to take into account the effect of turbulence on the particle. In LES simulations, the effect of the resolved velocity uctuations on the particles is accounted for and there is no need for an eddy interaction model like in RANS [86]. In this thesis, the effect of the subgrid scale velocity uctuations on the particle dispersion is assumed negligible and hence not modeled. Figure 3.6 represents a ow chart demonstrating the steps involved in tracking one injected particle.
43
Start with the location of one particle injected from the inlet surface at a certain point
Compute the surrounding cells of the current control volume where the particle lies
Interpolate the ow variables at the particle position
Determine the eddy parameters (and uctuating velocity using interpolated ow variables and the random number generator in case of RANS)
Compute the forces acting on the particle
Determine the integration time-step based on the cell size & eddy parameters
Perform a time-step integration to obtain updated particle position and velocity
No
Is the particle still in the current cell?
No
Is the current cell a boundary cell?
Check the distance between particle position & the nearest boundary cell (wall or outlet)
No
Is the distance to wall/outlet <= particle radius?
Particle is stuck on wall (trapped) or reached outlet (escaped)
Figure 3.6: Flow chart demonstrating the steps involved in tracking one injected particle [86]
44
Chapter 4 Sensitivity Analysis of Geometrical Parameters

4.1 Sensitivity analysis
The cyclone separator performance and the ow eld are affected mainly by the cyclone geometry where there are seven geometrical parameters, namely, the inlet section height a and width b, the vortex nder diameter Dx and length S , the barrel height h, the cyclone total height Ht and the cone-tip diameter Bc . all of these parameters are always expressed as a ratio of cyclone diameter D, as shown in Fig. 4.1 and Table 4.1. In this study, only the effect of geometry was taken into account. Nevertheless, what about the effect of ow rate on the performance. Overcamp and Scarlett [125, p 369, Fig. 6] studied the effect of changing Reynolds number on the cut-off diameter (Stokes number) and found that for Reynolds number values beyond 1E4, the effect of increasing the Reynolds number on the cut-off diameter is very limited. Furthermore, Karagoz and Avci [90, p 863, Fig. 7] studied the effect of increasing the Reynolds number on the pressure drop and found that beyond Reynolds number of 2E4 any increase in the Reynolds number has nearly no effect on the pressure drop. The values of the Reynolds number for all tested cases in this thesis have Reynolds number higher than 2E4. Consequently, the effect of ow rate
Table 4.1: The Stairmand high-efciency design
a/D 0.5
b/D 0.2
Dx /D 0.5
Ht /D 4.0 45
h/D 1.5
S/D 0.5
Bc /D 0.375
Chapter 4. Sensitivity Analysis of Geometrical Parameters
Figure 4.1: The cyclone separator dimensions
can be safely neglected.
4.1.1
Response surface methodology (RSM)
The usual method of optimizing any experimental set-up is to adjust one parameter at a time, keeping all others constant, until the optimum working conditions are found. Adjusting one parameter at a time is necessarily time consuming, and may not reveal all interactions between the parameters. In order to fully describe the response and interactions of any complex system a multivariate parametric study must be conducted [30]. Since there are seven geometrical parameters to be investigated, the best technique is to perform this study using the response surface methodology (RSM). RSM is a powerful statistical analysis technique which is well suited to model complex multivariate processes, in applications where a response is inuenced by several variables, and the objective is to optimize this response. Box and Wilson rst introduced the theory of RSM in 1951 [13]. RSM today is the most commonly used method of process optimization. Using RSM one may model and predict the effect of individual experimental parameters on a dened response output, as well as locating any interactions between the experimental parameters which otherwise may have been overlooked. RSM has been employed extensively in the eld of engineering and manufacturing, where many parameters are involved in the process [70, 106, 118, 167170]. 46
4.1. Sensitivity analysis
In order to conduct a RSM analysis, one must rst design the experiment, identify the experimental parameters to adjust, and dene the process response to be optimized. Once the experiment has been conducted and the recorded data tabulated, the RSM analysis software models the data and attempts to t a second-order polynomial to this data [30]. The generalized second-order polynomial model used in the response surface analysis was as follows:
7 7
Y = 0 +
i=1
i X i +
i=1
ii Xi2 +
i<j
ij Xi Xj
(4.1)
where 0 , i , ii , and ij are the regression coefcients for intercept, linear, quadratic and interaction terms, respectively. Xi and Xj are the independent variables, and Y is the response variable (Euler number).
4.1.2 Design of experiment (DOE)

The statistical analysis is performed through three main steps. Firstly, construct a table of runs with combinations of values of the independent variables via the commercial statistical software Statgraphics centurion XV by giving the minimum and maximum values of the seven geometrical factors under investigation as input. Secondly, perform the runs by estimating the pressure drop (Euler number) using the MM model (cf. Sec. A.3, page 256). Thirdly, ll in the values of pressure drop in the Statgraphics worksheet and obtain the response surface equation with main effect plot, interaction plots, Pareto chart and response surface plots beside the optimum settings for the new cyclone design. Table 4.2 depicts the parameters ranges selected for the seven geometrical parameters. The study was planned using BoxBehnken design, with 64 combinations. A signicant level of P < 0.05 (95% condence) was used in all tests. Analysis of variance (ANOVA) was followed by an F-test of the individual factors and interactions.
Table 4.2: The values of the independent variables Variables Inlet height, a/D =X1 Inlet width, b/D =X2 Vortex nder diameter, Dx /D =X3 Total cyclone height, Ht /D =X4 Cylinder height, h/D =X5 Vortex nder length, S/D =X6 cone-tip diameter, Bc /D =X7 minimum 0.4 0.14 0.2 3.0 1.0 0.4 0.2 center 0.55 0.27 0.475 5.0 1.5 1.2 0.3 maximum 0.7 0.4 0.75 7.0 2.0 2.0 0.4
47
Fitting the model

The analysis of variance (ANOVA) shows that the resultant quadratic polynomial models adequately represented the input data with the coefcient of multiple determination R2 being 0.92848. This indicates that the obtained quadratic polynomial model was adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (see Table 4.3). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables. Based on the ANOVA results presented in Table 4.3, the variable with the largest effect on the pressure drop (Euler number) was the linear term of vortex nder diameter, followed by the linear term of inlet width and inlet height (P < 0.05); the other four linear terms (barrel height, vortex nder length, cyclone total height and cone-tip diameter) did not show a significant effect (P > 0.05). The quadratic term of vortex nder diameter also had a signicant effect (P < 0.05) on the pressure drop; however, the effect of the other six quadratic terms was insignicant (P > 0.05). Furthermore, the interaction between the inlet dimensions and vortex nder diameters (P < 0.05) also had a signicant effect on the pressure drop, while the effect of the remaining terms was insignicant (P > 0.05).
4.1.3
Analysis of response surfaces
For visualization of the calculated factor, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve is proportional to the size of the effect, and the direction of the curve species a positive or negative inuence of the effect [61](Fig.4.2(a)). Based on the main effect plot, the most signicant factor on the Euler number are (1) the vortex nder diameter, with a second-order curve with a wide range of inverse relation and a narrow range of direct relation, (2) direct relation with inlet dimensions, (3) inverse relation with cyclone total height and insignicant effects for the other factors. Pareto charts were used to summarize graphically and display the relative importance of each parameter with respect to the Euler number. The Pareto chart shows all the linear and second-order effects of the parameters within the model and estimates the signicance of each with respect to maximizing the Euler number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto chart judges, whether each effect is statistically signicant within the generated 48
response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 4.2(b)) and ANOVA table (Table 4.3) there are four signicant parameters (six terms in the ANOVA table ) at a 95% condence level: the negative linear vortex nder diameter; the linear inlet width; the linear total cyclone height; a second-order vortex nder diameter; negative interaction between vortex nder diameter and inlet dimensions. These are the major terms in a polynomial t to the data. Therefore, the pareto chart is a perfect supplementation to the main effects plot. To visualize the effect of the independent variables on the dependent ones, surface response of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range while holding the other factors at their central values [189]. Thus, Fig. 4.2(c) was generated by varying the inlet height and the inlet width while holding the other ve factors xed at their central value. The trend of the curve is linear, with more signicant effect for inlet width, with no interaction between the inlet height and width. The response surface plots given by Figs. 4.2(d), 4.2(e) and 4.2(f) show that there are interactions between both inlet width and inlet height with the vortex nder diameter. The effect of cyclone total height is less signicant with respect to the vortex nder diameter, but its effect is higher than that of the vortex nder length, the barrel height and the cone-tip diameter.
4.1.4 Conclusions
Mathematical modeling (the Muschelknautz method of modeling (MM)) has been used to understand the effect of the cyclone geometrical parameters on the cyclone performance. The most signicant geometrical parameters are: 1. 2. 3. 4. the vortex nder diameter the inlet section width the inlet section height the cyclone total height.
The effect of both the barrel height and the vortex nder length on the cyclone separator performance are small in comparison with these most signicant geometrical parameters. There are strong interactions between the effects of inlet dimensions and the vortex nder diameter on the cyclone performance. This study conrms the insignicant effect of the cone-tip diameter on the cyclone performance. However, the discrepancy exists in literature for this 49
Table 4.3: Analysis of variance and the regression coefcients

Variable 0 Linear 1 2 3 4 5 6 7 quadratic 11 22 33 44 55 66 77 Interaction 12 13 14 15 16 17 23 24 25 26 27 34 35 36 37 45 46 47 56 57 67 R2 Regression coefcient -43.0742 178.176 372.26 -161.452 -1.55344 8.5875 -7.23112 19.5663 1.08238 -12.2111 403.419 -0.223597 -2.67108 1.81257 -62.1739 91.0488 -355.892 0.459314 -3.27883 2.19997 26.2787 -720.685 1.03571 -2.53478 4.2616 -5.28466 5.2034 2.77536 0.985086 32.579 -0.0452174 0.345301 -1.5016 -0.422227 3.82354 -6.40134 0.92848 8.11 19.79 232.04 0.6 0 0.1 0 0 0 107.8 0.09 0.05 0.15 0.04 0.22 14.75 0 0 0 0.01 42.42 0 0 0.01 0 0.51 0.01 0 0.05 0 0.02 0.01 0 0 0.02 0.0075 0.0001 0.0000 0.446 0.9691 0.757 0.9537 0.9931 0.9446 0.0000 0.7641 0.8223 0.6994 0.8364 0.6427 0.0005 0.9726 0.9514 0.9465 0.9191 0.0000 0.9467 0.9675 0.9112 0.9862 0.4799 0.9249 0.9568 0.8221 0.9911 0.8902 0.9404 0.9667 0.9622 0.8945 F-Ratio P-Value
Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
50
(a) Main effect plot
(b) Pareto chart
(c) X1 versus X2
(d) X1 versus X3
(e) X2 versus X3
(f) X3 versus X4
Figure 4.2: Analysis of design of experiment (cf. Table 4.2)
51

geometrical factor motivates us to study this individual parameter in more details using large eddy simulation methodology as given in Sec. 4.2. Before proceed to apply CFD technique in the optimization of the cyclone geometry and to study the effect of each signicant factor factor, a question appears. Should the dust outlet geometry be included in the simulation domain? The answer of this question is given in Sec. 4.3.
4.2 The cone-tip diameter

Until now, a considerable number of investigations has been performed either on small sampling cyclones or larger industrial cyclone separators, for example; Buttner [19], Iozia and Leith [84], Kim and Lee [95], Zhu et al. [199], Elsayed and Lacor [50, 52, 53] and Sakhani et al. [147]. In these studies, almost all of the cyclone dimensions listed in Table 4.4, were varied and the changes in cyclone performance characteristics brought about by these variations were studied. However, very little information is available on the effect of changing the cone bottom (tip) diameter (which determines the cone shape if other cyclone dimensions are xed [184]) on the ow pattern and performance. Regarding this effect, discrepancies and uncertainties exist in the literature. Bryant et al. [17] observed that if the vortex touched the cone wall, particle re-entrainment occurred and efciency decreased, so the collection efciency will be lower for cyclones with a small cone opening (cone-tip diameter). However, according to Stern et al. [171] (cited in Xiang et al. [184]), a cone is not an essential part for cyclone operation, whereas it serves the practical purpose of delivering collected particles to the central discharge point. However, Zhu and Lee [200] stated that the cone provides greater tangential velocities near the bottom for removing smaller particles. Furthermore, the sensitivity analysis presented in Sec. 4.1 indicates the insignicant effect of the cone-tip diameter on the cyclone performance. However, the understanding and knowledge of the ow eld inside a cyclone has been developed rapidly over the last few years, the exact mechanisms of removing particles are still not fully understood. Therefore, most existing cyclone theories are based on simplied models or depend upon empirical correlations [23]. Xiang et al. [184] carried out experiments with cyclones of different cone dimensions and evaluated a few models, namely Barth [9], Leith and Licht [99] and Iozia and Leith [85]. All these models could simulate correctly the trend of Xiangs experimental data. However, the quantitative agreement was not satisfactory. CFD has a great potential to predict the ow eld characteristics and particle trajectories as well as the pressure drop inside the cyclone [68]. Chuah et al.[23] carried out a numerical investigation on the same cyclone dimensions used by Xiang et 52
4.2. The cone-tip diameter
al. [184] with the commercial nite volume code Fluent. Using different turbulence models they proved that Fluent with Reynolds stress turbulence model (RSM) predicts well the cyclone collection efciency and the pressure drop. The CFD simulation results from Chuah et al. [23] agree well with Xiangs experimental results in that cyclones with a smaller cone diameter result in a slightly higher collection efciency compared to cyclones with a bigger cone-tip diameter (only if the cone-tip diameter is not smaller than the gas exit tube diameter). Moreover, the change in the pressure drop will not be signicant when the cone size is varied. Both Xiang and Chuah did not give any results about the effect of the cone-tip diameter on the ow eld inside the cyclone separator, except some plots for axial and tangential velocity proles at two stations in the ow eld for Chuah et al. [23]. Xiang and Lee [185] computationally investigated the effect of the cone-tip diameter on the ow eld using the Reynolds stress turbulence model. They did not present any contour plots for either the static pressure, tangential and axial velocity. However, the comparisons between the tangential and axial velocity proles at different sections indicating no valuable difference between the three cyclones [185, Fig. 8, p. 216 and Fig. 9, p. 217 ], they mentioned that the cone-tip diameter has a signicant effect on the ow eld. No particle tracking study has been performed in the study of Xiang and Lee [185]. Currently a better understanding of the ow eld inside cyclone separators is an important concern, especially with the application of large eddy simulation (LES). The present study was undertaken in an effort to carry out a numerical study on the effect of the cone-tip diameter on the ow eld and the cyclone performance using LES available in Fluent commercial nite volume solver.
Table 4.4: The geometrical dimensions of the three cyclones
Dimension Body diameter, D Gas outlet diameter, Dx Inlet height, a Inlet width, b Cyclone height, Ht Cylinder height, h Gas outlet duct length, S Cone-tip diameter, Bc Cyclone I Cyclone II Cyclone III Length (mm) 31 15.5 12.5 5 77 31 15.5 19.4 15.5 11.6 Dimension ratio (dimension/D) 1 0.5 0.4 0.16 2.5 1 0.5 0.625 0.5 0.375
The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = 0.75D from the cyclone center (cf. Fig. 4.3).
53
Dx a h D
Le S Li b
Ht
Bc
Figure 4.3: Schematic diagram for the cyclone geometry and coordinate denition
4.2.1
Numerical simulation
4.2.1.1 Conguration of the three cyclones The cyclones used in this study had a reversed ow tangential inlet. The geometry and dimensions are shown in Fig. 8.1 and Table 4.4. Three cyclones with different cone-tip diameters are used viz., Bc /D= 0.625, 0.5 and 0.375. The three cyclones are identical to those used by both Xiang et al. [184, 186] and Chuah et al. [23]. Four plotting sections are used to investigate the effect of the cone-tip diameter Bc on the velocity proles as given by Table 4.5. 4.2.1.2 Selection of the turbulence model (RANS versus LES) For the turbulent ow in cyclones, the key to the success of CFD lies with the accurate description of the turbulent behavior of the ow [68]. To model the swirling turbulent ow in a cyclone separator, there are different turbulence models available in Fluent. These range from the standard k model to the more complicated Reynolds stress model (RSM)
Table 4.5: The position of different plotting sections
Section z (mm) z/D S1 5 0.16 S2 15 0.48 S3 30 0.97 S4 50 1.61
54

and large eddy simulation (LES) methodology as an alternative for RANS models. The standard k , RNG k and Realizable k models were not optimized for strongly swirling ows found in cyclones [23]. The Reynolds stress turbulence model (RSM) requires the solution of transport equations for each of the Reynolds stress components and yields an accurate prediction on swirl ow pattern, axial velocity, tangential velocity and pressure drop on cyclone simulation [159]. Large eddy simulation (LES) has been widely accepted as a promising numerical tool for solving the large-scale unsteady behavior of complex turbulent ows. Encouraging results have been reported in recent literature and demonstrate the ability of LES to capture the swirling ow instability and the energy containing coherent motion of such highly swirling ows [41]. LES methodology has been used in many articles to study the highly swirling ow in cyclone separators, [e.g., 38, 39, 48, 155, 156, 159, 191]. It will be used in this study to reveal the effect of changing the cone-tip diameter on the turbulent ow in the cyclone separator. The simulation will start with the Reynolds stress turbulence model for ow initialization, then the large eddy simulation methodology will be applied. 4.2.1.3 Solver settings Selection of the discretization schemes The choice of the discretization schemes has a tremendous inuence on the simulation results and the Fluent solver offers many different schemes for pressure-velocity coupling, pressure, momentum, kinetic energy, rate of kinetic energy dissipation discretization [59]. Both Kaya and Karagoz [91] and Shukla et al. [158] investigated the performance of different discretization schemes in the steady and unsteady simulation of cyclone separators. The schemes used in this study are given in the following paragraphs together with an explanation of the reasons behind their selection. Kaya and Karagoz [91] reported the advantages of SIMPLEC (semi implicit method for pressure linked equations consistent) scheme for pressurevelocity coupling in terms of convergence. For the pressure discretization they stated that only the PRESTO (pressure staggering option) pressure interpolation scheme only can predict precisely the mean velocity proles static pressure distribution and the pressure drop in the cyclone separator with good agreement with the experimental values. This scheme is also recommended by the Fluent manual [59] for highly swirling ows. For momentum, the QUICK (quadratic upwind differencing) scheme has been recommended by both Kaya and Karagoz [91] and Fluent manual [59] for the ow in cyclone separators. For the discretization of kinetic 55

energy and its dissipation rate equation, the second-order upwind scheme has been used [158]. The rst-order upwind scheme has been used for the discretization of the Reynolds stress equations [91, 158]. For the LES simulations the bounded central difference scheme is the default and the recommended convection scheme by the Fluent manual [59]. The time step The unsteady simulation started with the Reynolds stress model for initialization with a time step of 1E-4s using implicit coupled solution algorithm. For the LES simulation, the time step is 1E-5 s. The selected time step results in an average inlet Courant number of 2.3 for the three cyclones. But as the solver is a segregated implicit solver, there is no stability criterion that needs to be met in determining the time step (and consequently the Courant number) [59, 137]. However, to model transient phenomena properly the Fluent manual [59] suggested using a time step of at least one order of magnitude smaller than the smallest time constant in the system. In the cyclone separator studies, the average residence time (cyclone volume/ gas volume ow rate) is widely used to estimate the time step [23, 43, 52]. In this study, the cyclone volumes (calculated by the Fluent solver) are 5.045387E-5 m3 , 4.738813E-5 m3 and 4.468306E-5 m3 for cyclones I-III respectively. For a ow rate of 30E-3 m3 /s, the corresponding average residence time values are 0.00168s, 0.00158s and 0.001489s for cyclones I-III respectively, i.e., the used time step is just a small fraction of the average residence time. This conrms that the used time step can reveal the transient phenomena properly. However, the interest in this study lies in the simulation of averaged scalars and vectors (the average velocities and pressures in order to estimate the cyclone performance). To resolve the high-frequency phenomena in the time domain may be a time step smaller by two orders of magnitude (tiny fraction) than the average residence time would be required which is not the case in this study. Furthermore, to verify that the choice for the time step was appropriate after the calculation is complete; Fluent manual [59] suggests to check the maximum value of Courant number at the most sensitive transient regions of the domain (in this study, it is the central region) should not exceed a value of 20-40 [59]. For the three cyclones the maximum values of Courant number are 9.78, 12.1, 16.92 for cyclones I-III respectively. This veries again that the choice of the time step was proper. Convergence criteria With regard to the convergence criteria, two aspects should be considered. Firstly, the scaled residuals should be below 1E-5 (The default convergence 56

criterion of Fluent is that scaled residuals of all equations fall below 1E3). Secondly, some representative quantities such as velocity and pressure should be monitored until they are constant [137]. Although the present simulations were converged at about (t=1.5-1.6 s), they were only terminated at t=2s to get more accurate time averaged values. After achieving constant tangential velocity with time at a certain point in the middle of the cyclone domain, Fluent begins the data sampling for time statistics for the whole domain (the velocity components and the static pressure) each time step for sufciently long time (t=1.5s until t=2.0s). From this step, the following time averaged values are available; the mean and the root mean square values of the static pressure p, the velocity magnitude v, the x-velocity vx , the y-velocity vy and the z-velocity vz . The mean z-velocity vz is identical to the mean axial velocity vaxial . The mean tangential velocity v can be obtained using Fluent custom eld function [cf. 59] according to Eq. (4.2). = v x cos + v y sin v (4.2)
x is the time averaged x-velocity, v y is the time averaged y-velocity where v and is the angular coordinate. Boundary conditions and other settings Velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall (no-slip) boundary condition at all other boundaries. The air inlet velocity Uin equals 8 m/s, corresponding to air inlet volume ow rate Qin =30 l/s, air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s, leading to a Reynolds number of 1.18E4 based on the cyclone diameter and the area averaged inlet velocity. The turbulence intensity I equals 5% and the turbulence characteristic length equals 0.07 times the inlet width [74]. At the cyclone inlet, the Reynolds stress specic method in Fluent solver is the Reynolds stress components. The diagonal components of the Reynolds 3 2 , I Uin stress tensor (normal stresses) are assigned to 2kin /3, kin = 2 where kin is the kinetic energy at the inlet [59, 81]. The shear stresses (non-diagonal components) at the inlet are set to zero. To take into account the stochastic component of the turbulent ow at the inlet for the LES simulation, articial perturbations have been generated using the spectral synthesizer method available in the Fluent solver [59, 82, 162], where the uctuation velocity components are computed by synthesizing a divergence-free velocity-vector eld from the summation of 100 Fourier harmonics [59]. The uctuations are added to the mean inlet velocity. The reason for introducing these articial perturbations instead of selecting the no-perturbation option in the Fluent solver, is that the unpertur57

bated at turbulent proles at inlet generates unrealistic turbulent eddies [59]. For the near-wall treatment, the enhanced wall function [59] has been used in the RSM simulation. For subgrid scale model, the dynamic Smagorinsky-Lilly model [59, 60, 101] has been used. The Smagorinsky model constant is dynamically computed instead of given as an input to the solver, but clipped to zero or 0.23 if the calculated model constant is outside this range to avoid numerical instabilities [59]. The second advantage of the dynamic Smagorinsky-Lilly model over the Smagorinsky model is the treatment near the wall. In the dynamic Smagorinsky-Lilly model, a damping function for the eddy viscosity near the wall is not required, since the model constant goes to zero in the laminar region just near the wall [113]. The boundary condition at the gas outlet is the outow boundary condition provided by Fluent [59], where all transport variables have a zero normal gradient. This boundary condition is valid for fully developed ow, and this is the reason why the vortex nder is extended 1/2 cyclone diameter above the top of the cyclone in the present study to allow the exit ow to be a fully developed ow [185]. The effect of this distance on the ow eld has been the subject of different investigations [e.g., 153, 178], where this distance was varied between zero and four cyclone diameters. Wang et al. [178] investigated the effect of outow length on the ow eld and velocity prole and suggested to have the gas outlet boundary condition at a distance longer than the cyclone radius. The grid independency study The grid independence study has been performed for the three tested cyclones. Three levels of grid for each cyclone have been tested, to be sure that the obtained results are grid independent. The hexahedral computational grids were generated using GAMBIT grid generator and the simulations were performed using Fluent 6.3.26 commercial nite volume solver on a 8 nodes CPU Opteron 64 Linux cluster. The computational results of the three grid types are presented in Table 4.6. As seen the maximum difference between the results obtained from the ne and medium meshes is 1% for the calculation of the cut-off diameter and the Euler number which is in the range of experimental error [52, 137]. It has been observed that even medium grids provide a sufcient grid independency. However, for excluding any uncertainty, computations have been performed using the ne grid, where the total number of grid points was not that critical with respect to the computation overhead [10]. Consequently, the used grid produces grid independent results (the author only checked the mean values, so for future studies with unsteady phe58

Table 4.6: The details of the grid independence study for cyclones I-IIIa
Cyclone Coarse Medium Fine % differenceb % differencec
a
N 632153 861077 1021616
I Eu 2.48 2.405 2.39 3.76 0.63
x50 1.396 1.355 1.35 3.4 0.37
N 513021 863852 1025778
II Eu 2.36 2.28 2.27 3.96 0.44
x50 1.335 1.257 1.25 6.8 0.56
N 513991 712576 1027982
III Eu 2.869 2.712 2.687 6.77 0.93
x50 1.24 1.21 1.2 3.33 0.83
N is the number of hexahedral cells, Eu is the Euler number (dimensionless pressure drop = pressure drop / average kinetic energy at inlet) and x50 is the cut-off diameter; the particle diameter that will produce 50% collection efciency (cf. Sec. 3.3). b The percentage absolute difference between the coarse and ne grid values for Euler number and cut-off diameter. c The percentage absolute difference between the medium and ne grid values for Euler number and cut-off diameter.
nomena like vortex-core precession (cf. Ref. [38]), the effect of the grid on the Strouhal number associated with the simulated vortex-core precession [38] should be included in the grid independency study probably requiring ner grids, but this is not part of the present study). Moreover, to evaluate accurately the numerical uncertainties in the computational results (especially because of the large difference between the results obtained on the coarse and the ne mesh which is about 7%), the concept of grid convergence index (GCI) was adopted using three grid levels per cyclone.
Grid convergence index (GCI) Roache [143145] suggested a quantitative measure for the grid convergence; the grid convergence index (GCI). The GCI can be computed using two levels of grid; however, three levels are recommended in order to estimate accurately the order of convergence and check that the solution is within the asymptotic range of convergence [160]. For a consistent numerical analysis the discretized equations will approach the solution of the actual equations as the grid resolution approaches zero [160]. The appropriate level of grid resolution is a signicant issue in numerical investigations. It is a function of many variables including the ow condition, type of analysis, geometry and many other variables. The GCI is based upon a grid renement error estimator derived from the theory of the generalized Richardson extrapolation [160]. The GCI is a measure of how far the computed value is away from the value of the asymptotic numerical value. Consequently, it indicates how much the solution would change with a further renement of the grid. A small value of GCI indicates that the computation is within the asymptotic range. 59

The GCI on the ne grid is dened as: GCIne = Fs || (rp 1) (4.3)
where Fs is a factor of safety. Fs = 3 for comparison of two grids and 1.25 for comparison over three grids or more. For the coarse grid: GCIcoarse = Fs ||rp (rp 1) (4.4)
is a relative error measure of the key variable f between the coarse and ne solutions, = f2 f1 f1 (4.5)
where f2 is the coarse-grid numerical solution obtained with grid spacing h2 . f1 is the ne-grid numerical solution obtained with grid spacing h1 . r is the grid renement ratio (r = h2/h1 > 1). For complicated geometries r is replaced by the ratio of the number of control volumes in the ne and coarse mesh [107] which is the case in this study, N1 N2
1 D
r12 =
(4.6)
where D = 2 and 3 for two-dimensional and three-dimensional geometries respectively [143, pp. 410]. N1 and N2 are the number of control volumes in the ne and coarse mesh respectively. p is the order of the discretization method. p equals two if the second order discretization is used for all terms in space [107] (However, Slater [160] stated that if all discretization in space was of second-order, p will be less than 2. The difference is due to grid stretching, grid quality, non linearity in the solution, presence of shocks, turbulence modeling and perhaps other factors). For the grid renement study, three meshes have been used with N1 , N2 and N3 cells for the ne, medium and coarse three-dimensional mesh respectively.
N2 1 r12 = N , r23 = N , e12 = f2 f1 , e23 = f3 f2 , where ei,i+1 = N2 3 fi+1 fi is the difference in the key variable f resulting from the use of different grids. If r12 = r23 then,
1 3 1 3
p = ln
e23 e12 60
/ln(r)
(4.7)

If r12 = r23 which is the case in this study, Roache [145] proposed to solve Eq. 4.8 e12 e23 p = r12 p p (r23 1) (r12 1) (4.8)
Equation 4.8 is transcendental in p. Using the iterative technique with relaxation factor introduced in Roache [144, 145] p = + (1 ) where = ln( ) ln(r12 ) (4.9)
p 1)e23 (r12 , = 0.5 and is the previous iteration of p. The p (r23 1)e12
author suggest to use = ln will stop if GCIne 23 = Now one can calculate, 12 =
1.25|23 | . p 1) (r23 | p p |
e23 e12
/ln(r12 ) as a rst guess. The iteration 23 =

f3 f2 f2 ,
< 1E 5.
f2 f1 f1 ,
GCIne 12 =
ne GCIne 12 should be smaller than GCI23
1.25|12 | p 1) (r12
and
To check if the solution is in the asymptotic range, 1 (cf. Eq. 4.10) =

p r12 GCIne 12
GCIne 23
(4.10)
The Richardson extrapolation can be used to obtain the value of f when the grid spacing h vanishes (h 0) [2, 143].
p fexact = f1 + (f1 f2 )/ (r12 1)
(4.11)
Table 4.7 presents the grid convergency calculations using GCI method and three grid levels for cyclones I -III. The following conclusions have been obtained from the GCI analysis: The results are in the asymptotic range for the three cyclones, because the obtained values for are close to unity. The ratio R is less than unity this means monotonic convergence [2] (Ali et al. [2] classied the possible convergence conditions into three groups, namely (1) monotonic convergence; 0 < R < 1 (2) oscillatory convergence; R < 0 (3) divergence; R > 1.) . There is a reduction in the GCI value for the successive grid renene ments (GCIne 12 < GCI23 ) for the two variables (Eu and x50 ). This indicates that the dependency of the numerical results on the cell size has been reduced. Moreover, a grid independent solution has been achieved. Further renement of the grid will not give much change 61

in the simulation results. For the two variables (Eu and x50 ), the extrapolated value is only slightly lower than the nest grid solution. Therefore, the solution has converged with the renement from the coarser grid to the ner grid [2]. Figure 4.4 presents a qualitative proof that the obtained results are in the asymptotic range, i.e., the obtained results are mesh independent.
Table 4.7: Grid convergency calculations using GCI method and three grid levels for cyclones I - III
i 0c 1 Eu I x50 2 3 0 1 2 3 0 1 Eu II x50 2 3 0 1 2 3 0 1 Eu III x50 2 3 0 1 2 3
a b c
Ni 1021616 861077 632153 1021616 861077 632153 1025778 863852 513021 1025778 863852 513021 1027982 712576 513991 1027982 712576 513991
fi 2.3750 2.3900 2.4050
ri,i+1
ei,i+1
i,i+1
GCIne i,i+1 % 0.7849
Ra
1.0586 1.1085 2.4800 1.3471 1.3500 1.0586 1.3550 1.1085 1.3960 2.2523 2.2700 1.0589 2.2800 1.1897 2.3600 1.2413 1.2500 1.0589 1.2570 1.1897 1.3350 2.6835 2.6870 1.1299 2.7120 1.1150 2.8690 1.1962 1.2000 1.1299 1.2100 1.1150 1.2400
0.0150 0.0750
0.0063 0.0312
0.2013 1.5597
1.0063
0.0050 0.0410
0.0037 0.0303
0.2649 0.1224 0.7252 1.0037
0.0100 0.0800
0.0044 0.0351
0.9754 0.1256 1.5194 1.0044
0.0070 0.0780
0.0056 0.0621
0.8740 0.0902 1.5652 1.0056
0.0250 0.1570
0.0093 0.0579
0.1619 0.1607 1.3127 1.0093
0.0100 0.0300
0.0083 0.0248
0.3913 0.3361 1.4211 1.0083
R=12 /23 . p GCI12 /GCI23 . = r12 The value at zero grid space (h 0). i=1, 2 and 3 denote the calculations at the ne, medium and coarse mesh respectively.
62

1.5 3
Cyclone I (Eu) Cyclone I (X50) Cyclone II (Eu) Cyclone II (X50) Cyclone III (Eu) Cyclone III (X50)
1.45
2.8
1.4
Cut-off diameter
Euler number
2.6
1.35
1.3 2.4 1.25 2.2 1.2
5E-07
1E-06
1.5E-06
1.15
(h --> 0)
-1
Figure 4.4: Qualitative representation of the grid independency study. The Euler number and the cut-off diameter for each cyclone at the three grid levels. N 1 is the reciprocal of the number of cells, h 0 means the value at zero grid size (cf. Table 4.7). To obtain a smooth curve; the spline curve tting has been applied in Tecplot post-processing software.
4.2.2 Results and discussion Validation of results

The obtained numerical results are compared with the LDA velocity measurements of Hoekstra [74] measured using laser doppler anemometry (LDA) system. Figure 4.5 shows the comparisons between the LES simulation and the measured axial and tangential velocity proles at axial station Z=94.25 cm from the cyclone bottom [74]. The LES simulation predicts a similar trend as observed experimentally. The non exact matching between experimental and LES simulation has been reported in some other literatures [e.g., 38]. Considering the complexity of the turbulent swirling ow in the cyclones, the agreement between the simulations and measurements is considered to be quite acceptable. Another comparison between the current LES results and the Reynolds stress turbulence model (RSM) results of Xiang and Lee [186] for cyclone III is given in Fig. 4.6 which indicates LES can also depict the main ow features of cyclonic ow as the Reynolds stress turbulence model can do. 63

2 1.8 0.6
Tangential velocity/ Inlet velocity
1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 LDA Simulation -0.5 0 0.5 LDA Simulation -0.5 0 0.5
Axial velocity/ Inlet velocity Radial position / Cyclone radius
0.4
0.2
-0.2
-0.4
Radial position / Cyclone radius
Figure 4.5: Comparison of the time averaged axial and tangential velocity between LDA measurements [74] and the LES simulations at section Z=94.25 cm from the cyclone bottom (cf. Hoekstra [74] for more details about the used cyclone geometry in this validation).
4.2.3
The ow pattern in the three cyclones
The pressure eld Figure 4.7 shows the time-averaged static pressure contours plots. In the three cyclones, the static pressure decreases radially from the wall to the center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, while the gradient in the axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low-pressure zone at the cyclone center (twisted cylinder). The static pressure contour plots for the three cyclones are almost the same. The velocity eld Based on the contours plots of the time averaged tangential velocity (Fig. 4.7) and the radial proles at sections S1, S2, S3 and S4 shown in Fig. 4.8 (cf. Table 4.5), the following comments can be drawn. The maximum tangential velocity equals around 1.25 times the average inlet velocity and occurs in the annulus cylindrical part. The tangential velocity distribution for the three cyclones are nearly identical at the corresponding sections. The tangential velocity prole at any section is composed from two regions, inner and outer. In the inner region the ow rotates approximately like a solid body rotation (forced vortex), where the tangential velocity increases with radius. After reaching its peak the tangential velocity decreases with radius in the outer part of the prole (free vortex). This prole is so-called 64

1.25 Present study Xiang and Lee (2005) Dimensionless tangential velocity Dimensionless axial velocity 1 0.5 Present study Xiang and Lee (2005)
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance
0.75
-0.5 -1
-0.75
0.75
Figure 4.6: Comparison of the time averaged tangential and axial velocity between RSM results of Xiang and Lee [186] and the current LES results at Z/D=1.29 for Cyclone III (Bc /D = 0.375). Note: the dimensionless radial distance = the distance / the cyclone radius, the dimensionless velocity = the velocity / inlet velocity.
Rankine type vortex which include a quasi-forced vortex in the central region and a quasi free vortex in the outer region. The radial proles given in Fig. 4.8 represent the time averaged tangential velocity in the lower part of the cyclone. The tangential velocity distributions at the bottom sections show good axis-symmetrical distribution. The tangential velocity proles for the three cyclones are identical in the inner region, where the maximum tangential velocity nearly equals the inlet velocity and occurs at a position 0.25 - 0.45 of the cyclone radius as given in Table 4.8. The axial velocity contours (Fig. 4.7) indicate the existence of two ow streams. Downward ow directed to the cyclone bottom (negative axial velocity), and upward ow directed to the vortex nder exit. The axial velocity plots for the three cyclones are nearly identical to the corresponding sections in the conical part. The axial velocity equals zero at the walls and maximum close to the position of maximum tangential velocity. The axial velocity proles shown in Fig. 4.8 exhibit a severe asymmetrical feature.
Table 4.8: Comparison between the maximum tangential velocity value and its position at different sections.
Section Cyclone / v v in max x/R I 1.08 0.3 S1 II 0.97 0.28 III 1.02 0.26 I 1.02 0.3 S2 II 0.95 0.33 III 1.05 0.33 I 0.96 0.32 S3 II 0.89 0.37 III 1.04 0.41 I 0.94 0.35 S4 II 0.88 0.43 III 1.035 0.425
The ratio between the maximum tangential velocity and the area average inlet velocity. The dimensionless distance between the centerline and the point of maximum velocity, R is the cyclone radius.
65
Comparison of the velocity proles in the three cyclones However, from the previous discussions it is clear that, the effect of the cone-tip diameter on the ow eld in the conical section is insignicant, in comparison with other geometrical parameters such as the vortex nder diameter. Nevertheless, in this section a comparison between the axial and tangential velocity proles at four sections (Table 4.5) will be analyzed as presented in Fig. 4.9. The tangential velocity proles in the forced vortex region are nearly identical in the three cyclones at each sections. The tangential velocity in the free vortex region increases as the cone-tip diameter is reduced. The tangential velocity proles for the three cyclones are almost the same. The axial velocity prole has the shape of an inverted W for all cyclones. The highest axial velocity occurs at 0.25 - 0.5 of the cyclone radius down the vortex nder until the cyclone bottom, and between 0.25 and 0.5 of the cyclone radius in the annulus and through the vortex nder. No considerable difference exists in the axial velocity proles for the three cyclones. Since the axial velocity proles is almost the same for the three cyclones, the average residence time of particles is nearly the same. Furthermore, the position of the highest axial velocity moves inward in the conical part as the cone-tip diameter reduced. From the previous analysis, the region of downward ow is nearly the same, for the three cyclones, while the tangential velocity slightly increases as the cone-tip diameter reduced, so the particles will experience a higher tangential velocity for cyclone III than in other cyclones for the same time (as the region of downward axial velocity is nearly equal). This results in a slightly higher collection efciency. This is consistent with the measured results reported by Xiang et al. [184] and simulation by Xiang and Lee [186] and Chuah et al. [23]. The change of the cone-tip diameter affects the ow eld in the cyclone separator but this change is so limited, i.e., the reduction of cone-tip diameter enhances the collection efciency but with a small percentage, as the ow eld pattern is so closed for the three cyclones. The DPM results In order to calculate the effect of the cone-tip diameter on the cut-off diameter, 104 particles were injected from the inlet surface with particle velocity equals the gas inlet velocity. The particle density is 860 kg/m3 and the maximum number of time steps for each injection was 9E5 steps. The DPM analysis results for the three cyclones are shown in Table 4.9 and Fig. 4.10. It is found that the cut-off diameter decreases slightly with decreasing the 66
Figure 4.7: The contours plots for the time averaged ow variables at Y=0 . From top to bottom: the static pressure N/m2 , the tangential velocity and axial velocity m/s. From left to right cyclone I, II and III respectively.
67

1.25 0.5
Dimensionless tangential velocity
S1 S2 S3 S4
S1 S2 S3 S4
Dimensionless axial velocity

-0.75 -0.5 -0.25 0 0.25 0.5 Dimensionless distance 0.75 1
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.5 -1
-0.75
0.75
1.25
S1 S2 S3 S4
0.5
S1 S2 S3 S4

0.25
0.75
0.5
-0.25
0.25
0 -1
-0.5 -1
-0.75
0.75
1.25
S1 S2 S3 S4
0.5
S1 S2 S3 S4

0.25
0.75
0.5
-0.25
0.25
0 -1
-0.5 -1
-0.75
0.75
Figure 4.8: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) for each cyclone. From top to bottom: Cyclone I, II and III respectively.
68

1.25 Cyclone I Cyclone II Cyclone III
0.5
Cyclone I Cyclone II Cyclone III
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
-0.5 -0.25 0 0.25 0.5 Dimensionless distance Cyclone I Cyclone II Cyclone III
0.75
-0.5 -1
-0.75
0.75
1.25
0.5
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
0.75
-0.5 -1
-0.75
0.75
1.25
0.5
0.25
0.75
0.5
-0.25
0.25
0 -1
-0.75
0.75
-0.5 -1
-0.75
0.75
1.25
0.5
0.25
0.75
0.5
0.25
0 -1
-0.75
0.75
-0.25 -1
-0.75
0.75
Figure 4.9: Comparison between the radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0). From top to bottom: section S4 - S1 respectively.
69
cone-tip diameter while the pressure drop is increasing slightly. Consequently, the effect of the cone-tip diameter on the cyclone performance is insignicant. The trend of changing the cut-off diameter with the cone-tip diameter given by Chuah et al. [23] (Qin = 60l/min) supports the conclusion of the slightly decrease of the cut-off diameter by decreasing the cone-tip diameter (insignicant effect), Table 4.9.
4.2.4
Comparison with mathematical models
Table 4.10 presents a comparison between the Euler number (dimensionless pressure drop) and the cut-off diameter obtained from CFD, experimental investigation [184] and seven mathematical models, viz. (the Barth model [9], the Muschelknautz method of modeling (MM) [116, 174], the Stairmand model [165], the Casal and Martnez-Benet model [21], the Shepherd and Lapple model [157], the Iozia and Leith model [85] and the Ritema model [142] (cf. Sec. 2.2) The Euler numbers obtained from the models of Shepherd and Lapple, Casal and Martnez-Benet are constant, because these models do not include the effect of the cone-tip diameter Bc in their formulas. The three other models (Barth, MM and Stairmand) indicate less effect on both the Euler number and the cut-off diameter by changing the cone-tip diameter. The models of Iozia and Leith in addition to that of Rietma indicate no change in the cut-off diameter with changing the cone-tip diameter. The results of mathematical models and the experimental investigation support the CFD results that the cone-tip diameter has an insignicant effect on the cyclone separator performance.
Table 4.9: The cut-off diameter and pressure drop for the three cyclones Cyclone Bc /D Cut-off diameter [m] Cut-off diameter [m] (Chuah et al. [23] ) Pressure drop [N/m2 ] Euler number Eu
I 0.625 1.35 1.65 76.5 2.39
II 0.5 1.25 1.45 72.7 2.27
III 0.375 1.2 1.1 86 2.687
Qin = 60 l/min.
70

5 Euler number [-] (Exp.) Euler number [-] (LES) Cut-off diameter [micron] (LES) 1.6
Cut-off diameter [micron]
Ritema 1.4 1.4 1.4
1.8
Euler number [-]
3 1.4 2
1.2 1
0 0.3
0.35
0.4
0.45
0.5 Bc /D
0.55
0.6
0.65
1 0.7
Figure 4.10: The effect of cone-tip diameter on the pressure drop (Euler number) and the cut-off diameter (with spline curve tting to get a smooth curve).
4.2.5 Conclusions
Large eddy simulation has been used to study the effect of the cone-tip diameter on the cyclone ow eld and performance. Three cyclones with different values of Bc /D viz. 0.625, 0.5 and 0.375 (at constant vortex nder diameter Dx /D = 0.5) have been investigated. The following conclusion can be drawn. The cone-tip diameter has an insignicant effect on the ow pattern and performance. As the cone-tip diameter decreases, the maximum tangential velocity increases slightly, while its position is almost the same.
Table 4.10: The cyclone performance parameters using CFD, Experimental [184] and different seven mathematical models
Euler number Eu [-] Cyclone I II III I II III CFD 2.39 2.27 2.687 1.35 1.25 1.2 Barth 6.94 7.40 7.43 1.22 1.22 1.28 MM 4.88 4.95 4.95 2.13 2.089 2.089 Stairmand 6.68 6.69 6.69 Sphered 4.1 4.1 4.1 Casal 4.07 4.07 4.07 Iozia 1.44 1.44 1.44 Exp. 2.8 2.8 3.25 3.01 2.60 2.36
Cut-off diameter x50 [m]
The mathematical model used for estimation of the pressure drop only. The mathematical model used for estimation of the cut-off diameter only. Different particle density.
71

The ow pattern and performance parameters of the three cyclones are almost the same. Decreasing the cone-tip diameter increases the pressure drop slightly. The reverse trend is obtained for the cut-off diameter. Seven mathematical models used for estimation of the effect of conetip diameter on the cyclone performance, and all of them support the CFD results for the insignicance of varying the cone-tip diameter on the cyclone performance. However, the main nding of the current study is the insignicant effect of the cone-tip diameter on the cyclone performance, in comparison to the other geometrical parameters like the vortex nder diameter Dx or the inlet dimensions [cf. 50, 52], the cone is an essential part for cyclone operation. If the cone would be removed from the cyclone, it will be only cylindrical part and that one will have another geometry (not the conventional cyclone). Moreover, the particles are collected over the cone wall and then moved to the cyclone bottom. Consequently, a very low collection efciency will be expected for a cyclone separator without a cone.
4.3 The dust outlet geometry

Conventional cyclones always have a dustbin attached to the cone to collect the separated solid particles. When a gas ow stream enters the dustbin (closed at bottom), some of the ow will return the cone and carry some of the separated particles. This phenomena is called re-entrainment and it will affect the separation efciency of the cyclone [138]. However, many works have been carried out to investigate the inuence of different geometrical parameters such as cyclone length, inlet and outlet pipe geometries on the performance of cyclones [15, 62, 102, 140, 184], there has been little work concerning the dust outlet geometries [40, 47, 78, 123]. Regarding this inuence, discrepancies and uncertainties exist in the literature. Xiang and Lee [186] reported that the dustbin connected to the cyclone should be incorporated in the ow domain as it affects the results. On the other hand, numerous studies were performed without dustbin [e.g., 159, 178] with good matching with experimental results. Obermair et al. [123] performed cyclone tests with ve different dust outlet geometries to nd the inuence of the dust outlet geometry on the separation process. They showed that separation efciency can be improved signicantly by changing the dust outlet geometry, and they reported that further research is needed to clarify precise effects of dust outlet geometry. 72
4.3. The dust outlet geometry

The effect of a dipleg (a vertical tube between the cyclone and the dustbin) was posed and investigated by several researchers [78, 92]. The previous studies on the effect of dust outlet geometry on the cyclonic ow can be classied into the following categories: 1. Comparison between cyclones with dustbin and that with dustbin plus dipleg [78, 123, 138]. Obermair et al. [123] investigated experimentally the effect of different dust outlet geometries on the ow pattern. Nevertheless, the given ow pattern was limited to the cyclone bottom. The effects of the dust outlet geometry on the velocity proles are not given. 2. Comparison between the cyclone with and without dipleg (prolonged cyclones). Kaya and Karagoz [92] numerically investigated the ow characteristics and particle collection efciencies of conventional (without dipleg) and prolonged cyclones. 3. Comparison between cyclone with and without dustbin [47]. Elsayed and Lacor [47] numerically investigated two cyclones with and without dustbin. They reported a negligible effect of the dustbin on the performance. They reported further research is needed with the cyclone with dipleg and dust hopper (dustbin plus dipleg). All studies above do not contain enough information about the effect of the dust outlet geometry on the ow pattern, velocity proles and separation performance. Furthermore, there is no clear comparison between the four cases: separation space only (cylinder on cone), cyclone with cylindrical shaped dustbin, cyclone with vertical tube (dipleg) and cyclone with dustbin plus dipleg. The present numerical investigation aims at examining the inuence of the dust outlet geometry on the ow pattern and the cyclone performance using the Reynolds stress turbulence model. This serves two purposes: (1) to demonstrate whether simulation of the separation space only without including the dustbin or the dipleg is sufcient to estimate the cyclone performance. (2) to give information for the designer about the effect of the conguration under the separation space (the dustbin or the dipleg) on the ow pattern and performance.
4.3.1 Numerical simulation

4.3.1.1 Congurations of the four tested cyclones The numerical simulations were performed on four cyclones. Cyclone I has only the cylinder on cone shape (separation space only), cyclone II has a dustbin, cyclone III has a dipleg while cyclone IV has a dust hopper (dustbin plus dipleg). Figure 4.11 and Table 4.11 give the cyclones dimensions. 73
Table 4.12 gives more details for the used cyclones, including the number of cells, cyclone volume and the ow residence time for each cyclone. Nine sections are used to plot the velocity proles as shown in Table 4.13.
(a) I
(b) II
(c) III
(d) IV
(e) The surface mesh
Figure 4.11: Schematic diagrams and surface meshes for the four tested cyclone separators
4.3.1.2 Solver settings Based on the study of Kaya and Karagoz [91] for the best selection of numerical schemes to be used with RSM model, the following discretization schemes have been used. The PRESTO scheme has been used for the pres74

Table 4.11: The geometrical dimensions of the four tested cyclones Dimension Body diameter, D Inlet height, a Inlet width, b Gas outlet diameter, Dx Gas outlet duct length, S Cone-tip diameter, Bc Cylinder height, h Cyclone height, Ht Length (m) 0.205 0.105 0.041 0.105 0.105 0.076875 0.3075 0.82 Dimension ratio (Dimension/D) 1 0.5 0.2 0.5 0.5 0.375 1.5 4
The outlet section is above the cyclone surface by Le = 0.5D. The inlet section located at a distance Li = 0.75D from the cyclone center, the height of the dustbin and the dipleg, LD = 2D.
sure interpolation, the SIMPLEC algorithm for pressure velocity coupling, the QUICK scheme for momentum equations, the second-order upwind for the turbulent kinetic energy and the rst-order upwind discretization scheme for the Reynolds stresses [52], cf. Sec. 4.2.1.3 for more details. Chuah et al. [23] stated that the time step should be selected as a tiny fraction of the residence time tres . From Table 4.12, the value of tres varies between 0.237 and 0.399 s. Therefore a time step of 1E-4 is an acceptable value for the current simulations [52]. The simulations have been performed using FLUENT 6.3.26. 4.3.1.3 CFD grid and boundary conditions The mesh sensitivity study has been performed for the four tested cyclones with three levels for each cyclone, to be sure that the obtained results are grid independent. For example, three different meshes with respectively 130596, 260230 and 478980 cells have been used for cyclone I. The computational results on the three grids are presented in Table 4.14. As the maximum difference between the results is less than 5%, so the grid template 130596 produces the grid independent results [52, 146]. It has been observed that even 130596 grid provides a sufcient grid indepenTable 4.12: The details of the four tested cyclones
Cyclone Number of cellsa Cyclone volume x102 [m3 ] tres [s] b
a b
I 260230 1.99 0.237
II 593125 3.35 0.399
III 322286 2.18 0.259
IV 441062 2.76 0.328
The total number of hexahedral cells after the mesh sensitivity study The average residence time, tres = V /Qin where V is the cyclone volume and Qin is the gas ow rate.
75

Table 4.13: The position of different plotting sections
Section S1 S2 S3 S4 S5 S6 z`/D 2.75 2.5 2.25 2 1.75 1.5 z` is measured from the top of the inlet section (cf. Fig. 4.11(a)). S7 1.25 S8 1.0 S9 0.75
dency. However, for excluding any uncertainty, computations have been performed using the 260230 cells grid, where the total number of grid points was not that critical with respect to the computation overhead [10]. Figure 4.11(e) shows the surface mesh of the four cyclones. The hexahedral computational grids were generated using the GAMBIT grid generator. The boundary condition at the inlet section is the velocity inlet. An outow boundary condition is used at the outlet. The no-slip (wall) boundary condition is used at the other boundaries [52]. The air volume ow rate Qin =0.08405 m3 /s for all cyclones, air density 1.225 kg/m3 and dynamic viscosity of 178.940E-6 Pa s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [74].
DPM settings
A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 m at a velocity equals to the gas velocity.
4.3.2
Results
Validation of the numerical model

In order to validate the obtained results, it is necessary to compare the prediction with experimental data. The comparison performed with the measurements of Hoekstra [74] of the Stairmand cyclone using Laser doppler anomemetry (LDA). The present simulation are compared with the measured axial and tangential velocity proles at an axial station located at 94.25 cm from the cyclone bottom (Dx /D = 0.5) as shown in Fig. 4.12 (cf.
Table 4.14: The details of the grid independency study for cyclone I
Total number of cells 130596 260230 478980 % difference
Static pressure drop N/m2 955.51 960.25 961.12 0.584
Cut-off diameter m 1.48 1.5 1.51 1.987
The percentage difference between the coarsest and nest grid
76

2
LDA RSM
1.75 0.5 1.5

Tangintial velocity / inlet velocity
1.25 1 0.75 0.5 0.25 0 -1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
Axial velocity / inlet velocity

LDA RSM
0.25
-0.25
-1
-0.75
-0.5
-0.25
0.25
0.5
0.75
Radial position /cyclone radius
Radial position /cyclone radius
Figure 4.12: Comparison of the time averaged tangential and axial velocity between the LDA measurements, Hoekstra [74] and the current Reynolds stress turbulence model (RSM) results at 94.25 cm from the cyclone bottom.
Hoekstra [74] for more details about the used cyclone in this validation). The RSM simulation matches the experimental velocity prole with underestimation of the maximum tangential velocity, and overestimation of the axial velocity at the central region. Considering the complexity of the turbulent swirling ow in the cyclones, the agreement between the simulations and measurements is considered to be quite acceptable. The cyclone pressure drop is calculated as the pressure difference between the inlet and the average pressure across the vortex nder exit [74]. The experimental pressure drop of the cyclone can be calculated by the difference between the static pressures at the inlet and outlet [137]. A comparison of the pressure drop, the cut-off diameter (at particle density of 2740 kg/m3 ) obtained from the experimental data [74], CFD prediction is shown in Table 4.15. Table 4.15 indicates a very small deviations from the experimental values in both the calculated pressure drop and cut-off diameter. As the errors are less than 4%, so it is in the same magnitude as the experimental error [137]. The above comparison results show that the numerical model employed in this study can be used to analyze the gas ow eld and performance of the cyclone separator.
Table 4.15: Validation of the computational pressure drop and cut-off diameter Experimental [74] CFD % error Static pressure drop [N/m2 ] 300 309 3 Cut-off diameter [micron] 1 0.965 3.5
77
The dominant velocity component of the gas ow in cyclones is the tangential velocity, which results in the centrifugal force for particle separation [186]. The axial velocity is responsible more than the gravity for the transport of particles to the collection devices [29, 104]. These velocity components will be discussed in details in order to investigate the effect of the dust outlet geometry on the ow properties. Moreover, the pressure distribution in the swirling ow in these four cyclone separators will be discussed in details.
The axial variation of the ow properties

Figures 4.13 and 4.14 present the radial proles of the time-averaged tangential and axial velocity and static pressure at nine axial stations. As expected, the tangential velocity proles exhibit the so-called Rankine vortex, which consists of two parts, an outer free vortex and an inner solid rotation in the center (Fig. 4.13). The tangential velocity distribution in the inner region is rather similar in different sections for the same cyclone. In the outer region, due to the sharp drop in velocity magnitude in the near wall region, the distribution is different but the maximum tangential velocity is similar in all sections. Generally, the tangential velocity distribution varies only slightly with axial positions for the same cyclone, which is also reported in other articles [e.g., 66, 127, 163, 186]. This means that, if the tangential velocity increases at one section of the cyclone, it will increase at all other sections. The axial velocity proles at nine different stations are shown in Fig. 4.13. Two types of axial velocity proles are observed. Cyclone I, III and IV show an inverted W prole. Only cyclone II (with dustbin) has an inverted V prole. The reason for the two different axial velocity proles can be explained by the change in the ow eld pattern caused by the dustbin. The radial proles of the time averaged static pressure are given in Fig. 4.14 for the four cyclones. Like for the tangential velocity, the axial variations of static pressure are very small for the same cyclone. An exception is the cyclone I which shows some variations in the central part. Figure 4.15 compares the static pressure proles for the four cyclones at sections S7S9 (located at the cylindrical part of the cyclone, which is the most effective part of the cyclone in the separation process, the location of the highest area average tangential velocity). The plots at sections S7-S9 are also representative for the other sections, as the axial variations in the ow variables are small. From the comparison between the radial proles of the four cyclones, the minimum pressure at the cyclone center is almost the same for all cyclones. The static pressure radial proles of cyclones I, III and IV 78
30
30
30
30
25
Tangential velocity (m/s) Tangential velocity (m/s)
25
Tangential velocity (m/s)
25
25
20
20
20
20
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
S1 S2 S3 S4 S5 S6 S7 S8 S9
10
10
10
10
0 -1
-0.5
0.5
0 -1
-0.5
0.5
0 -1
-0.5
0.5
0 -1
-0.5
0.5
Radial position/ Cyclone radius
Axial velocity (m/s)
79
20
15
10
S1 S2 S3 S4 S5 S6 S7 S8 S9
20
15
10
S1 S2 S3 S4 S5 S6 S7 S8 S9
20
S1 S2 S3 S4 S5 S6 S7 S8 S9
20
S1 S2 S3 S4 S5 S6 S7 S8 S9
15
15
10
10
-5 -1
-0.5
0.5
-5 -1
-0.5
0.5
-5 -1
-0.5
0.5
-5 -1
-0.5
0.5
Figure 4.13: The radial prole for the time averaged tangential and axial velocity at different sections.

almost coincide. Cyclone II (with dustbin) depicts fewer gradients in the radial direction. It has the lowest maximum pressure value. Referring to Fig. 4.15, there is no difference between the radial pressure proles at the three sections for the same cyclone.
1400 1200 1000 800 600 400 200 0 -200 -1
S1 S2 S3 S4 S5 S6 S7 S8 S9
1200
1000
Static pressure (N/m )
2
800
600
S1 S2 S3 S4 S5 S6 S7 S8 S9
400
200
0 0 0.5 1 -1
-0.5
-0.5
0.5
(a) Cyclone I
(b) Cyclone II (dustbin)
1400 1200 1000 800 600 400 200 0 -200 -1

S1 S2 S3 S4 S5 S6 S7 S8 S9
1400 1200
1000 800 600 400 200 0

S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.5
0.5
-1
-0.5
0.5
(c) Cyclone III (dipleg)
(d) Cyclone IV (dustbin plus dipleg)
Figure 4.14: The radial proles for the time-averaged static pressure at different sections for the four cyclones.
The velocity proles

Figure 4.16 compares the tangential and axial velocity proles at sections S7-S9. The variation of the velocity proles (both the axial and the tan80

gential) from cyclone to cyclone is mainly located at the central region for both the tangential and axial velocity. Since the effect of the dust outlet geometry on the tangential velocity (centrifugal force) is minor, it is expected that the collection efciency (cut-off diameter) of the four cyclones will be comparable.
1600 1400
With dust bin Without dust bin With dipleg With dipleg & dust bin
2
1600 1400 1200 1000 800 600 400 200 0 -1
1200 1000 800 600 400 200 0 -1
-0.5
0.5
-0.5
0.5
(a) The four cyclones at section S7
(b) The four cyclones at section S8
1600 1400
1200 1000 800 600 400 200 0 -1
-0.5
0.5
(c) The four cyclones at section S9
Figure 4.15: The radial prole for the time-averaged static pressure at different sections for the four cyclones.
The axial velocity in the inner vortex is either reported as an inverted V 81
or inverted W-shaped prole,i.e., with a maximum (V-shaped) or a dip (Wshaped) at the symmetry axis [29]. Hoekstra et al. [75] stated that the shape of the axial velocity prole is affected by the cyclone geometry. They referred the dip in the inverted W prole to the loss of swirl in the vortex nder (the friction force of the vortex nder wall attenuates the swirling ow), which results in an adverse pressure gradient at the centerline [78]. Hence, uid with fewer swirls is drawn back from the exit pipe into the cyclone. This core ow prevails throughout the entire separation space of the cyclone in spite of the attenuation of swirl in the conical part of the cyclone [78]. This explains the reason behind the inverted W-shaped prole exhibited by cyclones I (cylinder on cone), cyclone III (dipleg) and cyclone IV (dustbin plus dipleg). However, why cyclone II (with dustbin) exhibit the inverted V-shaped prole? In cyclone II, the dustbin has an equal diameter to the cyclone barrel, which means sudden expansion to the downward ow and sudden contraction to the upward ow (the ow inside the dustbin also has two streams due to the absence of a vortex stabilizer [75] which can prevent the re-entrainment of the collected particles). The upward gas ow (directed from the end of dustbin) has a higher kinetic energy especially at the cone-tip diameter (sudden contraction) which can overcome the adverse pressure gradient at the centerline (caused by the swirl attenuation in the vortex nder) and results in the inverted V axial velocity prole. One more question may appear now, why cyclones III (dipleg) and IV (dustbin plus dipleg) did not exhibit the inverted V-shaped prole? The reason can be referred to the diameter of the dustbin directly connected to the cyclone. For cyclones III and IV there is no change in the ow area at the connection, consequently no ow acceleration happens. The effect of dustbin dimensions (diameter, height) still need more investigations. The author believe, if a cone is inserted at the entrance of the cone-tip inside the dustbin of cyclone II, the axial velocity will become inverted W-shaped. However, the dimensions and location of this cone still need more investigations (cf. Obermair et al. [123] for more details).
The ow pattern
Regarding the effect of neglecting the dust outlet geometry (dustbin or dipleg or dustbin plus dipleg) in the simulating domain on the ow eld pattern, the following comments can be drawn (Fig. 4.17). 1. From the comparison between the static pressure contour plots of each cyclone and that of cyclone I, the highest value of static pressure is obtained in case of cyclone III (dipleg). The lowest value is given by cyclone II (dustbin). The highest value of the static pressure 82
35
30
25
20
25
15
20
10
15
10
-5
0 -1
-0.5
0.5
-10 -1
-0.5
0.5
35
30
25
20
25
15
20
10
15
10
-5
0 -1
-0.5
0.5
-10 -1
-0.5
0.5
35
30
25
20
25
15
20
10
15
10
-5
0 -1
-0.5
0.5
-10 -1
-0.5
0.5
Figure 4.16: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at the inlet region (sections S7, S8 and S9). From top to bottom: section S7-S9.
83
in cyclone IV (dustbin plus dipleg) is in between that for cyclones III and II. This indicates a slight underestimation of the pressure drop by neglecting the effect of dipleg or dustbin plus dipleg (cyclone III, IV versus cyclone I). Furthermore, a slight overestimation of the pressure drop is observed by neglecting the effect of the dustbin (cyclone II versus cyclone I). 2. The tangential velocity pattern is very similar in the four cyclones (Rankine prole). The highest value is almost the same for the four cyclones. Consequently, the collection efciency (cut-off diameter) of the four cyclones will be almost the same. Moreover, the contour plots for the tangential velocity of the four cyclones are quite similar in the main separation space (cylinder and cone). 3. The axial velocity patterns for cyclones I, III and IV have the shape of an inverted W prole while that of cyclone II has an inverted V prole, indicating different ow behavior. Close to the cone bottom there are different ow patterns as a consequence of different dust outlet geometry. Figure 4.18 shows a qualitative view of the complex ow in the four tested cyclones with the streamtraces plots of the time averaged velocities colored by the time-average axial velocity. The swirling, downward ow at the outer region of the cyclone is clearly visible. Near the bottom of the cyclone, it is diverted into an upward ow near the cyclone center. In cyclone II (with dustbin) and IV (with dustbin and dipleg), the ow behavior in the dustbin is quite different because of the different length of the dustbin.
The performance
Figure 4.19 presents the grade efciency curves (GECs) for the four tested cyclones. As expected, the frictional efciencies of all the cyclones are seen to increase with the increase in particle size. The shapes of the grade collection efciency curves of all models have a so-called S shape [195]. It is clear from Fig. 4.19 that the effect of the cyclone dustbin or dipleg, on the cut-off diameter (particle diameter of 50% collection efciency) is small. Neglecting the effect of dust outlet geometry in the cyclone simulation slightly overestimates the cut-off diameter (when compare it with cyclone I). A deeper look to the GECs indicates the variation of the collection efciency for particles diameters larger than 1.5 m, with a higher efciency for cyclone II and the lower efciency for cyclone I. This behavior is due to the increase in the separation space and the change in the maximum tangential velocity. For particles with diameters less than 0.8 m, 84
Figure 4.17: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From left to right: the static pressure (N/m2 ), the tangential and the axial velocity (m/s).
85
Figure 4.18: The streamtraces plots for the time averaged ow variables, colored by the average axial velocity (m/s).
86

the collection efciency tends to zero except in case of cyclone IV which depict a non zero collection efciency. In order to estimate the effect of dust outlet geometry on the performance parameters, the pressure drop and cut-off diameter have been calculated. Table 4.16 shows the effect of dust outlet geometry absence from the simulation domain on the pressure drop and cut-off diameter. From the comparison between the estimated pressure drop of cyclone I and the other three cyclones, the following comments are obtained. A slight underestimation is obtained when omitting the dipleg or dustbin plus dipleg from the simulation space. A slight overestimation result from omitting the dustbin. The pressure drop values given in Table 4.16 support the results obtained from the analysis of the static pressure contour plots.
Table 4.16: The effect of dust outlet geometry on the cyclone performance Cyclone Pressure drop (N/m2 ) Cut-off diameter (m) I 960 1.5 II (dustbin) 890 1.0 III (dipleg) 1017 1.25 IV (dustbin plus dipleg) 1008 1.2
1 Without dustbin With dustbin With dipleg With dustbin plus dipleg 0.8
Collection efficiency
0.6
0.4
0.2
4 5
Particle diameter [micron]
Figure 4.19: The grade efciency curves for the four cyclones.
4.3.3 Conclusions
Four cyclones have been simulated using the Reynolds stress model (RSM), to study the effect of the dust outlet geometry on the cyclone separator performance, ow pattern and velocity proles. The following conclusions have been obtained. 87

The maximum tangential velocity in the four cyclones is very similar. No radial acceleration occurs in the cyclone space (the maximum tangential velocity is nearly constant throughout the cyclone). The cyclone without dustbin slightly overestimates both the pressure drop and cut-off diameter. So the simulations with and without dustbin will produce nearly the same performance parameters. The cyclone without dipleg slightly underestimates the pressure drop and overestimates the cut-off diameter. Consequently, the simulations with and without dipleg will produce nearly the same performance parameters. The axial velocity patterns obtained by the four cyclones are different. If the main target of the CFD investigation is the performance parameters, one can safely simulate only the main separation space (cylinder on cone). However, if the aim is to investigate the ow eld pattern, the dust outlet geometry should be included in the simulation domain.
4.4 Closure
From the previous sections, it becomes clear that the most signicant factors are four (the vortex nder diameter, the inlet width, the inlet height and the total cyclone height (cone height)). Consequently, any minor modication in these particular factors will result in a considerable change in the cyclone performance. The effect of the cone-tip diameter on the cyclone performance is insignicant. Since the main target of this thesis is to optimize the cyclone performance, it is accepted to exclude the dust hopper (dustbin or dipleg) from the simulation domain and use the obtained CFD data to obtain the optimized cyclone design.
88
Chapter 5 The Vortex Finder Dimensions

5.1 Introduction
The vortex nder size is an especially important dimension, which signicantly affects the cyclone performance as its size plays a critical role in dening the ow eld inside the cyclone, including the pattern of the outer and inner spiral ows. Numerous studies have been performed for the effect of geometrical parameters on the ow pattern and performance [15, 52, 62, 102, 140, 184] whereas only limited number of studies have been devoted to the effect of the vortex nder dimensions. Iozia and Leith [84] optimized the cyclone design parameters, including the vortex nder diameter, to improve the cyclone performance using an optimization code. Kim and Lee [95] described how the ratio of the diameters of cyclone body, and the vortex nder diameter Dx /D affected the collection efciency and pressure drop of cyclones, and proposed an energyeffective cyclone design. Moore and Mcfarland [111] also tested cyclones, with six different vortex nders, and concluded that the variation in the gas outlet diameter under the constraint of a constant cyclone Reynolds number produced a change in the aerodynamic particle cut-off diameter. Bakari and Hamdullahpur [7] investigated experimentally the effects of both the inlet gas velocity, the cyclone inlet width, the vortex nder length and the vortex nder diameter on the cyclone performance. They reported that the vortex nder length and the vortex nder diameter have a strong effect on the cyclone performance parameters. Moreover, the results indicated that the vortex nder length has a direct effect on the cyclone performance. The longer vortex nder minimizes the short circuiting of incoming gases, preventing the dusty gas from owing directly from the inlet to the outlet. Zhu and Lee [200] carried out a set of experimental 89
Chapter 5. The Vortex Finder Dimensions
(a) The cyclone geometry
(b) The surface mesh for cyclone S1 (D5)
Figure 5.1: Schematic diagram for the cyclone separator
investigations on the particle collection efciency of small cyclones operating at high-ow rates. Special emphasis was given to the effects of the barrel height and the vortex nder length on the particle collection efciency. The length ratios of the barrel height h and vortex nder length S to the cyclone body diameter D were varied from 0.75 to 4.5 and from 0.5 to 1.5, respectively. Pressure drop decreased substantially either as the barrel height h became longer or as the vortex nder length S became shorter. It was also found that the difference between the cyclone barrel height h and the vortex nder length S affects the particle collection characteristics signicantly. The optimum performance will be obtained if (h S )/D = 1.0.
The performance of a cyclone, with different vortex nders, was evaluated by Lim et al. [102] to examine the effect of the vortex nder shape on the characteristics of the collection efciency. Four cylinder-shaped and six cone-shaped vortex nders were designed and employed to compare the collection efciencies of the cyclone, at ow rates of 30 and 50 l/min. The cylinder-shaped-vortex nders had different diameters and the coneshaped vortex nders had different cone lengths. The result indicates that two cone-shaped vortex nders, with different diameters, had the collection efciencies between those of the cylinder-shaped vortex nders with the same diameter, and that a smaller pressure drop per ow rate unit could be achieved for the cone-shaped design, but the cone length did not affect the collection efciency and pressure drop of the cyclone. Raou et al. [140] duplicated numerically the same study of Lim et al. [102] with limited details about the effect of the gas outlet diameter on the ow eld 90
5.1. Introduction
pattern and velocity prole. You-hai et al. [188] simulated the three-dimensional gas-phase ow eld in the cyclone separator with different vortex nder diameters. The results show that when the diameter of the vortex nder decreases, the downward ow decreases and the tangential velocities of the whole cyclone separator increase, but at the cost of the pressure drop increase. Horvath et al. [79] refereed the reason behind the two classes of axial velocity in cyclone separator (class V and class W) to the inuence of the vortex nder diameter. The pressure drop and collection efciency of a swirl tube with different vortex nder geometries were studied numerically by Jian and You-hai [87]. The gas ow elds were simulated by the Reynolds stress model (RSM) and the stochastic tracking approach in discrete phase model (DPM). The results indicate that the decrease of the vortex nder diameter leads to higher tangential velocity, which helps to improve the separation efciency. The back ow can be observed in converging coned-shaped vortex nder, meanwhile the diverging coned-shape vortex nder can make the ow move smoothly to reduce pressure drop. Ficici et al. [57] performed an experimental study using three cylindershaped vortex nders with diameters of 80, 120 and 160 mm. They investigated the effects of gas inlet velocity, the vortex nder diameter Dx and length S on the cyclone performance. They reported a linear relationship between the length of the vortex nder and the pressure loss. Khalkhalia and Sakhania [94] performed a multi-objective optimization of a cyclone vortex nder shape using CFD simulations data set. Two meta-models based on the evolved group method of data handling (GMDH) type neural networks are used as tness functions for Pareto-based optimization. In summary, all previous studies reported the signicant effect of the vortex nder dimensions on the cyclone performance and ow pattern. Nevertheless, the previous studies are not coherent, and did not present sufcient details about the effect of these two geometrical parameters on the pressure drop and the cut-off diameter. Moreover, detailed studies about the effect of the vortex nder dimensions on the velocity proles are scarce in the literature. The present study is intended to computationally investigate the effect of increasing the vortex nder diameter Dx and length S on the pressure drop and cut-off diameter and to obtain more details about the ow eld pattern and velocity proles using the large eddy simulation (LES) methodology. 91
5.2 Numerical settings

5.2.1 Conguration of the tested cyclones
The numerical simulations were performed on ve cyclones with different vortex nder diameters (at constant S ) and ve cyclones with different vortex nder length (at constant Dx ). Figure 5.1 and Table 5.1 give the cyclones geometrical dimensions.
5.2.2
Solver settings
The simulations started with unsteady simulation using the Reynolds stress turbulence model with a time step of 1E-4 for initialization of the ow eld. Afterwards, the turbulence model switched to the large eddy simulation with a time step of 1E-5s using implicit coupled solution algorithm. The selected time step results in an average inlet Courant number of 0.0288 for the tested cyclones. The cyclones volumes and the corresponding residence times for the tested cyclones are given in Table 5.2. The minimum value of residence time is 0.0812s i.e., the used time step is just a small fraction of the average residence time. This conrms that the used time step can reveal the transient phenomena properly. For the tested cyclones the maximum values of the courant number are less than 0.2. This veries again that the choice of the time step was proper (cf. Sec. 4.2.1.3, page 56 for more details).
Table 5.1: The geometrical dimensions of the tested cyclonesa
Dimension Inlet height, a Inlet width, b Barrel height, h Total cyclone height, Ht Cone tip-diameter, Bc Vortex nder diameter, Dx S/D = 0.5 Cycloneb dimension/D 0.375 0.2625 1.5c 4.0 0.375 0.30 0.35 0.40 0.45 0.50 0.5 0.625 0.875 1.0 1.0 (h S )/D Dx /S
Vortex nder length, S
Dx /D = 0.5
D1 D2 D3 D4 D5 S1 S2 S3 S4 S5
1.0 1.0 1.0 1.0 1.0 1.0 0.875 0.625 0.5 1.0c
0.6 0.7 0.8 0.9 1.0 1.0 0.8 0.625 0.5 1.0
a Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = D from the cyclone center, cf. Fig. 5.1(a). b c
Cyclone D5 and S1 are identical. cyclone S6 has different barrel height h/D = 2.0 .
92
5.2. Numerical settings

Table 5.2: The details of the tested cyclones
Cyclone Number of cellsa Cyclone volume x 105 [m3 ] tres b [s] D1 601482 6.765 0.0812 D2 620088 6.802 0.0816 D3 657841 6.845 0.0821 D4 666336 6.894 0.0827 D5 =S1 714029 6.949 0.0834 S2 795321 6.948677 0.0834 S3 798143 6.947 0.0834 S4 798256 6.946 0.0834 S5 888612 8.116 0.0974
a The total number of hexahedral cells after the grid independency study, cf. Sec. 5.2.4 b The residence time t res = V/Qin where V is the cyclone volume and Qin is the gas ow rate (50 l/min).
5.2.3 Boundary conditions

Velocity inlet boundary condition is applied at inlet, outow at the gas outlet and wall (no-slip) boundary condition at all other boundaries. The air inlet volume ow rate Qin =50 l/min, air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s. The turbulence intensity I equals 5% and the turbulence characteristic length equals 0.07 times the inlet width [52, 55, 74]. All other settings are identical to that given in Sec. 4.2.1.3.
5.2.4 Grid independency study

The grid independence study has been performed for the tested cyclones. Three levels of grid for each cyclone have been tested, to be sure that the obtained results are grid independent. The hexahedral computational grids were generated using GAMBIT grid generator and the simulations were performed using Fluent 6.3.26 commercial nite volume solver on an eight nodes CPU Opteron 64 Linux cluster. To evaluate accurately the numerical uncertainties in the computational results, the concept of grid convergence index (GCI) was adopted using three grid levels per cyclone. Grid convergence index (GCI) The grid convergence index (GCI) proposed by Roache [143145] was employed to test the grid independence of the simulations. The GCI are comTable 5.3: The details of the grid independence study for cyclone S1 (D5)
Number of cells 714029 1174029 1793459 % differenceb % differencec Eua 3.375 3.18 3.145 -6.8148 -1.1006 Cut-off diameter 1.985 1.436 1.318 -33.602 -8.2173
a 2 Euler number is the dimensionless pressure drop Eu = P/( 1 2 Vin ) where P is the static pressure drop, is the gas density, Vin is the gas inlet velocity. b The percentage difference between the coarsest and nest grid. c The percentage difference between the ne and nest grid.
93
puted using three levels of grids in order to estimate accurately the order of convergence and check that the solution is within the asymptotic range of convergence [160]. The GCI is based upon a grid renement error estimator derived from the theory of the generalized Richardson extrapolation [55, 160]. The GCI is a measure of how far the computed value is away from the value of the asymptotic numerical value. Consequently, it indicates how much the solution would change with a further renement of the grid. A small value of GCI indicates that the computation is within the asymptotic range (cf. Sec. 4.2.1.3 for more details). Table 5.4 presents the grid convergency calculations using GCI method for three grid levels for cyclone S1 as an example for the tested cyclones. The following conclusions have been obtained from the GCI analysis [55]: The results are in the asymptotic range, because the obtained values for are close to unity. The ratio R is less than unity this means monotonic convergence [2]. There is a reduction in the GCI value for the successive grid renene ments (GCIne 12 < GCI23 ) for the two variables (Eu and x50 ). This indicates that the dependency of the numerical results on the cell size has been reduced and a grid independent solution has been achieved. Further renement of the grid will not give much change in the simulation results. For the two variables (Eu and x50 ), the extrapolated value is only slightly lower than the nest grid solution. Therefore, the solution has converged with the renement from the coarser grid to the ner grid [2]. Figure 5.2 presents a qualitative proof that the obtained results are in the asymptotic range.
Table 5.4: Grid convergency calculations using GCI method and three grid levels for cyclone S1
i 0c 1 Eu 2 3 0 1 x50 2 3 Ni 1793459 1174029 714029 1793459 1174029 714029 fi 3.1336 3.1450 3.1800 1.1803 3.3750 1.2703 1.3180 1.1517 1.4360 1.1803 1.9850 0.5490 0.3823 14.4222 0.1180 0.0895 4.5223 0.2342 1.0895 0.1950 0.0613 1.8228 ri,i+1 ei,i+1 i,i+1 GCIne i,i+1 % Ra b
1.1517
0.0350
0.0111
0.4520 0.1815 1.0111
a R= / . 12 23 b = rp GCI 12 /GCI23 . 12 c The value at zero grid space (h 0). i=1, 2 and 3 denote the calculations at the ne, medium and coarse mesh respectively.
94
5.3. Results and discussions

3.4
Euler number Cut-off diameter
2.2
2 3.3
Euler number
1.8 3.2 1.6
3.1
1.4
1.2 3 0 5E-07 1E-06
(h --> 0)
N -1
Figure 5.2: Qualitative representation of the grid independency study. The Euler number and the cut-off diameter for cyclone S1 at the three grid levels. N 1 is the reciprocal of the number of cells, h 0 means the value at zero grid size (cf. Table 5.4). To obtain a smooth curve; the spline curve tting has been applied in Tecplot post-processing software. Table 5.5: The position of different plotting sections Section z`/Da
a
S1 2.75
S2 2.5
S3 2.25
S4 2
S5 1.75
S6 1.5
Cut-off diameter
S7 1.25
S8 1.0
S9 0.75
z`measured from the inlet section top (cf. Fig. 5.1(a)).
5.3 Results and discussions

The ow velocity can be decomposed into three components. The tangential and the axial velocity components are the major velocity components in comparison with the radial velocity component. Xiang and Lee [186] stated that the tangential velocity is the dominant gas velocity in gas cyclones, which results in the centrifugal force for particle separation. The axial component is responsible for the two ow streams (downward and upward).
5.3.1 The axial variation

Figures 5.3 and 5.4 present the radial proles of the time-averaged tangential and axial velocity and static pressure at nine axial stations (cf. Table 5.5). As expected, the tangential velocity proles exhibit the so-called Rankine vortex, which consists of two parts, an outer free vortex and an inner solid rotation in the center. The tangential velocity distribution in the inner region is rather similar at different sections for the same cyclone. 95
In the outer region, due to the sharp drop in the velocity magnitude in the near wall region, the distribution is different but the maximum tangential velocity is similar at all sections for the same cyclone (S1-S5). The maximum tangential velocity increases with decreasing the vortex nder diameter. The cyclones S1 to S5 show the inverted W prole but cyclones D1 to D3 show the inverted V prole. The radial proles of the time averaged static pressure are given in Fig. 5.3. Like for the tangential velocity, the axial variations of static pressure are very small for the same cyclone. The maximum value of the static pressure decreases when the vortex nder diameter is increased for cyclones D1 to D5. However, the maximum value of the static pressure slightly increases when the vortex nder length is decreased for cyclones S1 to S4. Cyclone S5 (Dx /D = 0.5, S/D = 0.5, h/D = 2.0) differs than cyclone S4 in only the barrel height (h/D = 1.5). From Fig. 5.4, a slight difference in the maximum tangential velocity between the two cyclones, which predict close values for the collection efciency. Moreover, the axial velocity proles for the two cyclones are close but less central dip in the axial velocity is exhibited in cyclone S5. The distinct difference between cyclones S4 and S5 is that in the maximum static pressure, where cyclone S5 indicates a reduction in the maximum static pressure. This indicates that the effect of changing the barrel height is more signicant than that of the vortex nder length. The sharp changes in the radial proles appear in Fig. 5.4 are due to crossing the vortex nder.
5.3.2
The ow pattern
Figure 5.5 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones D1-D5. The time-averaged static pressure decreases radially from the wall to the center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, whereas the gradient in axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low-pressure zone at the cyclone center (twisted cylinder). Two vortical motions are exist one moving down (outer vortex) and the other moving up (inner vortex). The highest value of the static pressure decreases with increasing the vortex nder diameter. Consequently, a smaller pressure drop can be expected when increasing the vortex nder diameter. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value decreases with increasing the vortex nder diameter. Consequently, a better collection efciency can be expected when decreasing the vortex nder 96
D1
16 16 12 12
D2
16 12
D3
14 12 10
D4
12 10
D5
S1 S2 S3 S4 S5 S6 S7 S8 S9
-0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-4 -0.015
-0.015
-4 -0.015
0 -0.015
0 -0.015
15
12
10
10
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
7.5
S1 S2 S3 S4 S5 S6 S7 S8 S9
2.5
0 -2 -2 -2.5
Static pressure (N/m2)
97
-5 -0.015
-0.01
-0.005 0 0.005 Radial position (m)
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
-4 -0.015
-0.01
0.01
0.015
-5 -0.015
-0.01
0.01
0.015
500
400 320
280
200
240 150 400 300 240 200 100
160
300
200
160
120
50
200
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
80
S1 S2 S3 S4 S5 S6 S7 S8 S9
80
40
-40 -0.015
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
0 -0.015
-0.015
-0.015
-100 -0.015
Figure 5.3: The radial prole for the time-averaged tangential and axial velocity and static pressure at different sections for cyclones D1 - D5.
S1
12 12 10 10
S2
12 10
S3
12 10
S4
12 10
S5
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
0 -0.015
0 -0.015
0 -0.015
0 -0.015
0 -0.015
10
10
10
10
10
7.5
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
2.5
-2 -2.5 -4 -5 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015 -5 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015 -5 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
-2
-4 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
100
100
100
100
98
200
200
200
200
200
150
150
150
150
150
100
50
50
50
50
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-100 -0.015
-100 -0.015
-100 -0.015
-100 -0.015
-100 -0.015
Figure 5.4: The radial prole for the time-averaged tangential and axial velocity and static pressure at different sections for cyclones S1 - S5.
diameter. Figure 5.6 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones S1-S5. The highest value of the static pressure slightly decreases with increasing the vortex nder length. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value is very closed for cyclones S1 to S4 but there is a small difference between cyclones S4 and S5. The axial velocity patterns for the ve cyclones (S1 to S5) have the shape of an inverted W prole. Cyclone S5 exhibit a less central dip in the axial velocity.
5.3.3 The radial variation

The tangential and axial velocity proles at section S6 (as a representative for the other sections, because the axial variations in the ow variables are small) for the tested cyclones are compared in Fig. 5.7. From the comparison between the radial proles of the ve D cyclones, the minimum pressure at the cyclone center is almost the same for cyclones (D3 - D5). The slope of the static pressure radial prole becomes atter with increasing the vortex nder diameter. Contrarily, the variation of the static pressure for cyclones S1 to S5 is very small. Decreasing the cyclone vortex nder diameter, increases the maximum tangential velocity. The maximum tangential velocity approaches asymptotically 1.589 times the inlet velocity when decreasing the vortex nder diameter. The effect of increasing the vortex nder length on the maximum tangential velocity is limited and small reduction in the maximum tangential velocity by increasing the barrel height for the same vortex nder dimensions (cyclone S5). The variation of axial velocity with changing the vortex nder length is limited close to the wall, especially in the outer part. In the central region, the change in axial velocity prole is more pronounced with the dip in axial velocity decreasing with increasing the vortex nder length. Cyclone S5 exhibit the smallest dip among cyclones S1 to S5. The variation of axial velocity in the ve D cyclones is limited close to the wall with a big change in the central part. The axial velocity prole gradually changes from the inverted W to be inverted V with decreasing vortex nder diameter. Again, limited effect on the axial velocity prole (inverted W) for the vortex nder length is recorded. Moreover, increasing the vortex nder length has only small inuence on the central part. Furthermore, The inuence of the barrel height is more pronounced than that of the vortex nder length. Generally, the effect of changing the vortex nder length 99
D1
D2
D3
D4
D5
100
Figure 5.5: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section for cyclones D1 - D5. From top to bottom: the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone D1 - D5.
S1
S2
S3
S4
S5
101
Figure 5.6: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section for cyclones S1 - S5. From top to bottom: the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone S1 - S5.
on the axial velocity prole is very small in comparison with the vortex nder diameter. The shape of the axial velocity prole is affected by the cyclone geometry [75]. Hoekstra et al. [75] referred the dip in the inverted W prole to the loss of swirl in the vortex nder (the friction force of the vortex nder wall attenuates the swirling ow), which results in an adverse pressure gradient at the centerline [78]. Hence, uid with fewer swirls is drawn back from the exit pipe into the cyclone. This core ow prevails throughout the entire separation space of the cyclone in spite of the attenuation of swirl in the conical part of the cyclone [78]. This explains the reason behind the inverted W-shaped prole exhibited by cyclones S1 to S5. However, why decreasing the vortex nder diameter gradually change the axial velocity prole from inverted W to exhibit the inverted V-shaped prole? The answer is, when the vortex nder diameter decreases, the swirl in the cyclone increases (as is clear from the tangential velocity proles for cyclones D1 to D5). Consequently, the ow can overcome the adverse pressure gradient and exhibit the inverted V prole for cyclone D1.
5.3.4
The cyclone performance
In order to estimate the effect of the vortex nder dimensions on the performance parameters, the Euler number (the dimensionless pressure drop) have been calculated. A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 micron. Figure 5.8 shows a sharp decrease of the Euler number (dimensionless pressure drop) with increasing the vortex nder diameter Dx and a small increase with increasing the vortex nder length S . This behavior can be explained as follows. The pressure drop in the cyclone is composed of three main contributions [52]: (1) the pressure drop at the inlet section. (2) the pressure drop in the cyclone body due to swirling motion and due to wall friction, this contribution may increase with increasing the cyclone barrel height (cyclone S5) or with increasing the vortex nder length (cyclones S1 - S4) as the wall friction will increase due to friction with a larger wall surface, or decreases as the vortex strength will decrease because the maximum tangential velocity decreases. (3) the main contribution to the cyclone pressure drop is the energy loss in the vortex nder, which mainly depends on the maximum tangential velocity in the cyclone. As is clear from Fig. 5.7 the maximum tangential velocity decreases with increasing the vortex nder diameter. As the inlet section is the same in the all tested 102
500
200
400
150
300
100
200
50
100
D x/D=0.30 D x/D=0.35 D x/D=0.40 D x/D=0.45 D x/D=0.50

-50
S/D=0.5 S/D=0.625 S/D=0.875 S/D=1.0 (h/D=1.5) S/D=1.0 (h/D=2.0)

-0.015
-100 -0.015
12 14 10 12 8
10

S/D=0.5 S/D=0.625 S/D=0.875 S/D=1.0 (h/D=1.5) S/D=1.0 (h/D=2.0)

-2 -0.015
-2 -0.015
2.5
-2.5 0
S/D=0.5 S/D=0.625 S/D=0.875 S/D=1.0 (h/D=1.5) S/D=1.0 (h/D=2.0)

-2 -0.015
-0.01
0.01
0.015
-5 -0.015
Figure 5.7: Comparison between the radial proles for the time averaged static pressure, tangential and axial velocity at section S6.
103

cyclones, the pressure drop in the inlet section does not vary with increasing the vortex nder dimensions. The sharp decrease of the Euler number between the D cyclones with increasing the Dx is mainly due to the decrease in the pressure drop as a result of the decrease in the maximum tangential velocity. The Euler number in cyclones (S1 - S4) increases as the vortex nder length increases. Since, the maximum tangential velocity in the four cyclones is almost the same. Consequently, this trend is due to the small increase in the energy loss at the wall surface with increasing S . Cyclone S exhibits a sudden (small) drop in the Euler number. There are two competing contributions: increase of the pressure drop due to friction (the barrel height in cyclone S5 is larger than that for cyclone S4) and decrease of the pressure drop due to the small reduction in the vortex strength (maximum tangential velocity, cf. Fig. 5.7).
0.6
12 11 10 9
0.8
Euler number (Dx) Stk50 x 10 3 (Dx) Euler number (S) Stk50 x 10 3 (S)
2.5
Euler number
8 7 6 5
1.5
0.5 4 3 0
0.3
0.35
0.4
0.45
0.5
Dx
Figure 5.8: The variation of the Euler number and the Stokes number with the vortex nder dimensions.
The trend of increasing the Stokes number with increasing the vortex nder diameter is quite reasonable, as the centrifugal force affecting particles attenuates when the swirl intensity (maximum tangential velocity) decreases (Fig. 5.8). The Stokes number slightly increases as the vortex nder length is increased (cyclone S1 - S4). The insignicant change of Stokes number is quite reasonable because of very limited changes in the ow pattern (cf. Fig. 5.7). The small increase in the Stokes number can be explained with the aid of the contours plots given in Fig. 5.6. The zone of peak axial velocity increases with increasing S . Consequently, the possibility of carrying bigger particles to escape with the upward ow slightly increased. Cyclone S5 depicts a sudden reduction in the Stokes number 104
Stk50 x 10
values in comparison with cyclone S4. This drop is due to separation space increase in cyclone S5, and the possibility of particles to be captured increases. Although, the vortex strength decreased with a small amount due to the reduction in the maximum tangential velocity. Figure 5.9 presents the variation of the cyclone performance parameters with the ratio Dx /S . It is clear that the effect of changing Dx /S depends on the variables (Dx or S ). Consequently, there are two curves per performance parameter (Euler number and Stokes number). The common point in Fig. 5.9 is cyclone D5 (=S1).
12 11 10 9
Euler number (Dx) Stk50 x 10 3 (Dx) Euler number (S) Stk50 x 10 3 (S)
2.6 2.4 2.2 2
Euler number
8 7 6 5 4 3
1.8 1.6 1.4 1.2 1 0.8
0.6
0.8
D x /S
Figure 5.9: The variation of the Euler number and the Stokes number with the x ratio of D (cyclone S5 is excluded). S
The variation of the performance parameters with the ratio of (h S )/D is presented in Fig. 5.10 for cyclones S1 - S5. Increasing (h S )/D decreases both the Euler number and Stokes number. However, the values of the performance parameters depend on the value of the barrel height as is clear from the values for cyclone S5. In order to obtain the Euler number-Stokes number relationship, Fig. 5.11 has been drawn. It indicates a general relationship (trend) between the two dimensionless numbers irrespective to the geometrical parameters values. Two second-order polynomials have been proposed by Elsayed and Lacor [54, 56], Eqs. 5.1 and 5.2. Equation 5.2 presents a good matching for the performance parameters 105
Stk50 x 10
3.8
Euler number (S1-S4) Stk50 x 10 3 (S1-S4) Euler number (S5) 3 Stk50 x 10 (S5)
2.4
2.3
Euler number
2.2
3.4 2.1 3.2 2 3
0.5
0.625
0.75
0.875
(h-S)/ D
Figure 5.10: The variation of the Euler number and the Stokes number with the S . ratio of hD
for cyclones D1 -D5. For smaller values of Euler number, there is underestimation of the Stokes number if the Eq. 5.2 is applied. The difference between the two correlations can be referred to two reasons. Firstly, Eq. 5.1 is based on the CFD simulations data for both Euler number and Stokes number, whereas Elsayed and Lacor correlation [56] (Eq. 5.2) was obtained from experimental data for Euler number values and Iozia and Leith model for the Stokes number values. Secondly, Eq. 5.1 is limited to only four geometrical parameters. The other three factors are xed, h = 1.5, S = 0.5 and Bc = 0.375. Consequently, Eq. 5.1 is not suitable to t the Euler number - Stokes number relationship for cyclones S1 - S5 because the values of S is away from the range of applicability. Stk50 = 100.3533(log10 (Eu)) Stk50 = 100.3016(log10 (Eu))
2
1.1645log10 (Eu)2.3198
Stk50 x 10
3.6
(5.1) (5.2)
0.9479log10 (Eu)2.5154
5.4 Conclusions
Nine cyclones of different vortex nder dimensions (diameter and length) have been simulated using the large eddy simulation (LES) methodology, 106
5.4. Conclusions
3
Simulation Correlation 1 Correlation 2
2.5
Stokes number x 10 3
1.5
Euler number
Figure 5.11: The variation of the Stokes number with the Euler number for cyclones D1D5.
to study the effect of the vortex nder dimensions on the performance and ow pattern. The following conclusions have been obtained. The maximum tangential velocity in the cyclone decreases with increasing the vortex nder diameter. A negligible change is noticed with increasing the vortex nder length. Very limited axial variations in the ow variables are reported with changing the vortex nder dimensions for the same cyclone. Increasing the vortex nder length, makes a small change in both the static pressure, axial and tangential velocity proles. However, decreasing the vortex nder diameter gradually changes the axial velocity prole from the inverted W to the inverted V prole. Decreasing the cyclone vortex nder diameter, increases the maximum tangential velocity. The maximum tangential velocity approaches asymptotically 1.589 times the inlet velocity when decreasing the vortex nder diameter. The Euler number (dimensionless pressure drop) decreases with increasing the vortex nder diameter Dx . Increasing the vortex nder length S slightly increases the Euler number. The Stokes number increases with increasing the vortex nder diameter, because the centrifugal force affecting particles attenuates when the swirl intensity (maximum tangential velocity) decreases. The Stokes number slightly increases as the vortex nder length is 107

increased (cyclone S1 - S4). The insignicant change of Stokes number is quite reasonable because of very limited changes in the ow pattern. The effect of changing Dx /S on the performance parameters depends on the variables (Dx or S ). Increasing (h S )/D decreases both the Euler number and Stokes number. However, the values of the performance parameters depend on the value of the barrel height.
108
Chapter 6 The Inlet Dimensions

6.1 Introduction
The effects of cyclone inlet section dimensions on the cyclone performance (pressure drop and cut-off diameter) have been reported in many articles. Casal and Martinez-Benet [21] proposed the following empirical formula for the dimensionless pressure drop (Euler number),
a b DD Dx D 2
Eu = 11.3
+ 2.33
(6.1)
implying proportionality with the square of the inlet area. Ramachandran et al. [139] on the other hand proposed, Eu = 20 ab 2 Dx
S D H h Bc DD D 1/3
(6.2)
i.e. a linear relation with the inlet area. Iozia and Leith [84, 85] presented a correlation to estimate the cut-off diameter d50 and found proportionality to (a b)0.61 . The importance of inlet dimensions becomes clearer after the study of natural length (or vortex length) by several researchers, e.g., Alexander [1]. The cyclone has two spiral motions, outer and inner. In the reverse ow cyclone, the outer vortex weakens and changes its direction at a certain axial distance Ln from the vortex nder [29]. This distance is usually called the turning length, natural length or vortex length of the cyclone. The inlet area is one of the relevant parameters inuencing the natural length. Alexander [1] found that Ln decreased proportionally to the inlet area (Ai = a b) but the opposite trend has been also reported [29]. 109
Chapter 6. The Inlet Dimensions

Numerous studies have been performed for the effect of geometrical parameters on the ow pattern and performance [e.g., 15, 62, 102, 140, 184] while the effect of cyclone inlet dimensions remained largely unexplored. The articles investigating the effect of cyclone geometry report only briey on the effect of inlet section dimensions on the cyclone performance without sufcient details about their effects on the ow pattern and velocity proles. A new trend is the use of multi-inlet cyclone [e.g., 103, 187, 195].
The effects of cyclone inlet on the ow eld and performance of cyclone separators have been numerically investigated by Zhao et al. [198]. They compared the performance of two types of cyclones with the conventional single inlet and spiral double inlets using the Reynolds stress turbulence model. The results show that the new type cyclone separator with spiral double inlet can improve the symmetry of gas ow pattern and enhance the particle separation efciency. While their nding is for double inlets cyclone, it supports the importance of the effect of the inlet section dimensions on the performance of cyclone separator. The signicant effects of the cyclone inlet dimensions on the cyclone performance have been acknowledged in many articles [e.g., 5]. For two inlets cyclone separators, Zhao [193] reported the possibility of increasing the cyclone efciency without signicantly increasing the pressure drop by improving the inlet geometry of the cyclone. The effect of inlet section angle has been tested by many researchers. Qian and Zhang [137] computationally investigated the effect of the inlet section angle. The pressure drop of the cyclone decreases to a 30% lower value than that for conventional cyclone, if the inlet section angle becomes 45 . However, Qian and Wu [136] reported only 15% reduction in the pressure drop for = 45 .
In summary, all articles mentioned above did not study the effect of the inlet height or width dimensions on the performance and ow pattern but they studied the effect of the inlet congurations (inclined instead of tangential), or the effect of the number of inlets (single or double) or the shape of the inlet section (rectangular duct or nozzle). The present study is intended to computationally investigate the effect of increasing the cyclone inlet width and height on the pressure drop and cut-off diameter and obtaining more details about the ow eld pattern and velocity proles. The study will be done using RANS and the Reynolds stress turbulence model (RSM) will be used as a closure turbulence model. 110

6.2.1 Conguration of the ve cyclones
The numerical simulations were performed on ve cyclones with different inlet dimensions. Figure 8.1 and Table 6.1 give the cyclones dimensions. Table 6.2 gives more details for the used cyclones including the number of cells, geometric swirl number, cyclone volume, ow residence time and the inlet velocity for each cyclone. In swirling ow, the swirl number usually characterizes the degree of swirl. In cyclone separators, the swirling ow is characterized by the geometric swirl number. The geometric swirl number Sg is a measure for the ratio of tangential to axial momentum [75, 154], dened by [74], Sg = Dx D 4 Ain (6.3)
where Dx is the vortex nder diameter, D is the cyclone body diameter, and Ain is the inlet cross-sectional area. For industrial cyclones, the geometrical swirl number usually varies between 1 and 5 [74]. Table 6.2 shows that, the tested cyclones cover this range. Implying that the obtained results can be applied to the industrial cyclones.
6.2.2 Boundary conditions

A velocity inlet boundary condition is used at the cyclone inlet, i.e., a velocity normal to the inlet is specied. An outow boundary condition is used at the outlet. The no-slip boundary condition is used at the other boundaries. The air volume ow rate Qin =50 l/min for all cyclones, air density 1.0 kg/m3 and dynamic viscosity 2.11E-5 Pa s. The turbulent intensity equals 5% and the characteristic length equals 0.07 times the inlet width [74].
6.2.3 Selection of the time step

The average residence time in the cyclone is determined from the cyclone dimensions and gas ow rate [43]. The residence time tres = V /Qin where V is the cyclone volume and Qin is the gas ow rate. This value is used to select the time step. The time step for the unsteady simulation should be a tiny fraction of the average residence time [23]. The tres 0.08s (for all tested cyclones) as shown in Table 6.2. So a time step of 5E-4 is 111

Table 6.1: The geometrical dimensions of the tested cyclones
Dimension Gas outlet diameter, Dx Vortex nder insertion length, S Cone tip-diameter, Bc Cylinder height, h Cyclone height, Ht Inlet height, a Inlet Width, b b/D=0.2625 a/D=0.375 Cyclone dimension/D 0.5 0.5 0.375 1.5 4.0 0.25 0.375 0.50 0.15 0.2625 0.375 b/a
A1 A2 A3 B1 B2 B3
1.05 0.7 0.525 0.4 0.7 1.0
Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = D from the cyclone center.
Cyclones A2 and B2 are identical.
an acceptable value for the current simulation for accurate results and achieve scaled residuals less than 1e-5 for all variables. The selected time step results in an average inlet Courant number of 29.25, 28.88, 21.67, 30.40, 21.45 for cyclones A1, A2, A3, B1 and B3 respectively. However, as the solver is a segregated implicit solver, there is no limitations on the Courant number for stability.
6.2.4
CFD grid
Figure 6.1(b) shows the surface grid of cyclone A2 used in this study. The hexahedral computational grids were generated using GAMBIT grid generator. A grid independency study has been performed for the ve tested cyclones. Three grid levels for each cyclone have been tested, to be sure that the obtained results are grid independent. For example, for cyclone A2 (B2) three levels of meshes with respectively 490164, 714029 and 1174029 cells have been used. The computational results are presented in Table 6.3. As seen, the maximum difference between the results is less than 5%, so the grid with 490164 cells can be considered as adequate [146]. It has been obTable 6.2: The details of the ve tested cyclones Cyclone Number of cells Sg Cyclone volume x105 [m3 ] tres [s] Inlet velocity [m/s]
A1 705088 5.984 6.878 0.0825 13.214
A2=B2 714029 3.989 6.95 0.0834 8.809
A3 820362 2.992 7.017 0.0842 6.607
B1 706370 6.981 6.875 0.0825 15.416
B3 816714 2.793 7.012 0.0841 6.166
The total number of hexahedral cells after the grid independence study
112
6.3. Results
Dx
Le
h Li D b
Ht
Bc
(b) The surface mesh for cyclone A2 (B2)
Figure 6.1: Schematic diagram and surface mesh for the cyclone separator
served that even 490164 grid provides a sufcient grid independency. However, for excluding any uncertainty, computations have been performed using the 714029 cells grid, where the total number of grid points was not that critical with respect to the computation overhead [10].
Table 6.3: The details of the grid independence study for cyclone A2
Number of cells 490164 714029 1174029 % difference
Static pressure drop [N/m2 ] 95.99 100.922 98.68 2.73
Cut-off diameter [m] 0.99 1.0 1.02 2.94
6.3 Results
6.3.1 The axial variation of the ow properties
The tangential velocity is the dominant component of the gas ow in cyclones, which results in the centrifugal force for particle separation [186]. 113
Moreover, the development of axial velocity prole in axial direction will be analyzed for the ve cyclones. Nine sections are used to plot the velocity proles as shown in Table 6.4. Figures 6.2, 6.3 and 6.4 present the radial proles of the time-averaged static pressure, tangential and axial velocity at 9 axial stations. As expected, the tangential velocity proles exhibit the so-called Rankine vortex, which consists of two parts, an outer free vortex and an inner solid rotation in the center (Fig. 6.3). The tangential velocity distribution in the inner region is rather similar at different sections for the same cyclone. In the outer region, due to the sharp drop in velocity magnitude in the near wall region, the distribution is different and the change in the value of maximum tangential velocity is rather limited. Generally, the tangential velocity distribution varies only slightly with axial positions for the same cyclone, which is also reported in other articles [e.g., 66, 127, 163, 186]. This means that, if the tangential velocity increases at one section of the cyclone, it will increase at all other sections. The same conclusion can be drawn from the radial prole of static pressure with higher values of pressure drop expected for cyclone A1 in comparison with cyclones A2 and A3. Cyclone B1 also depicts higher values of pressure in comparison with cyclones B2 and B3 (Fig. 6.2). The axial velocity prole has the shape of an inverted W for all cyclones except B3 with the shape of an inverted V, as a result of the change in the ow eld pattern due to very wide inlet section.
Table 6.4: The position of different plotting sections Section z`/D
S1 2.75
S2 2.5
S3 2.25
S4 2
S5 1.75
S6 1.5
S7 1.25
S8 1.0
S9 0.75
z`measured from the inlet section top
To evaluate the effect of increasing the cyclone inlet width on the tangential and axial velocity proles, the tangential and axial velocity proles at section S9 (close to the inlet section) for the three cyclones (both for A cyclones and B cyclones) are compared in Fig. 6.5. As is clear from Fig. 6.5 the variation of axial velocity close to the wall is limited when changing the inlet width or height. The axial velocity proles for the three cyclones are very similar except at the central region. The most important is the effect of cyclone inlet height or width on the tangential velocity (proportional to the centrifugal force, which is the main force in the separation process). Increasing the cyclone inlet width or height decreases the maximum tangential velocity. Cyclone A1 and B1 have the maximum tangential velocity in comparison with other cyclones. This means that decreasing the cyclone inlet dimension will enhance the collection efciency. 114
6.3. Results
6.3.2 The ow pattern

Figure 6.6 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones A1, A2 and A3. It is observed that, the highest value of the static pressure decreases with increasing the inlet height. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value decreases with increasing the inlet height, so that a better collection efciency can be expected when decreasing the inlet height. The axial velocity patterns for the three cyclones have the shape of an inverted W prole. Figure 6.7 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones B1, B2 and B3. It can be seen that, the non-symmetry of the ow increases with increasing the inlet width. The highest value of the static pressure decreases with increasing the inlet width. The tangential velocity pattern is very similar for all cyclones (Rankine prole). The highest value decreases with increasing the inlet width, so that a better collection efciency can be expected when decreasing the inlet width. The axial velocity patterns for cyclones B1 and B2 have the shape of an inverted W prole while that of cyclone B3 has an inverted V prole. In cyclone B3, the inlet width (b/D=0.375) is wider than the gap between the cyclone barrel and the vortex nder wall. As a result, some part of the incoming ow will impact the vortex nder and hence will not experience any swirling motion around the vortex nder. This results in different axial velocity prole in cyclone B3 in comparison with the other two cyclones. Furthermore, this will cause excessive stresses on the vortex nder, vibrations and noise. From the inspection of Fig. 6.7 (also Fig. 8.2 page 152, Fig. 8.15 page 181 & Fig. 8.24 page 199), two observations can be made: the ow is asymmetric near the cone bottom and there is a large variation in the time-averaged axial velocity in this region. These observations are also found in the literature; for example, Gronald and Derksen [69, Fig. 5], Kaya and Karagoz [92, Fig. 8] and Wang et al. [178, 179, Fig. 6]. However, no explanation for this ow pattern is given. The author thinks this ow pattern can be explained as follows: (1) The cyclone has only one inlet; consequently, the ow should be asymmetric. However, this asymmetry should diminish far away from the inlet section. (2) Another effect comes from the cyclone bottom, which is closed in these simulations (no ow from the cyclone bottom). As a consequence of this boundary condition, the descending ow reverses its direction at the cone bottom resulting in a large asymmetry 115

A1
300 350 300 250
Static pressure [N/m ]
2
B1
250
200
2
200 150 100 50 0 -50 -1

S1 S2 S3 S4 S5 S6 S7 S8 S9
150
100
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -1
-0.5
0
Radial position [-]
0.5
-0.5
0
Radial position [-]
0.5
A2
200 200
B2
150
Static pressure [N/m ] Static pressure [N/m ]
150
100
100
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -1
-0.5
0
Radial position [-]
0.5
-50 -1
-0.5
0
Radial position [-]
0.5
A3
200 125
B3
150
Static pressure [N/m ] Static pressure [N/m ]
100
75
100
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
25
-50 -1
-0.5
0
Radial position [-]
0.5
-25 -1
-0.5
0
Radial position [-]
0.5
Figure 6.2: The radial prole for the time-averaged static pressure at different sections. Note: A2=B2.
116
6.3. Results
A1 B1
A2
B2
A3
B3
Figure 6.3: The radial prole for the time-averaged tangential velocity at different sections. Note: A2=B2.
117

A1 B1
A2
B2
A3
B3
Figure 6.4: The radial prole for the time-averaged axial velocity at different sections. Note: A2=B2.
118
6.3. Results
300
250
a/D=0.25 a/D=0.375 a/D=0.5
350 300 250

2
b/D=0.15 b/D=0.2625 b/D=0.375
200
2
200 150 100 50 0 -50 -1
150
100
50
-50 -1
-0.5
0
Radial position [-]
0.5
-0.5
0
Radial position [-]
0.5
119
at this region. The author believes that the extent of asymmetry is not just a function of the cyclone geometry and boundary condition at the cone bottom, but also the operating condition like the gas ow rate. This asymmetry causes the obtained axial velocity variations (spot-like distribution close to the cone bottom). (3) Another reason can be the boundary condition at the gas outlet (vortex nder exit). A fully developed ow (outow boundary condition) is assumed. Actually, there is no guarantee that the ow is fully developed. May be the ow is still developing throughout a long distance after the vortex nder exit. However, Wang et al. [178] reported that the effect of the gas outlet length on the simulated velocity becomes insignicant after a distance of 1/2 the cyclone diameter. In spite of the above-mentioned interpretations, the following parameters need to be investigated to accurately explain the reason of this ow behavior: (i) The effect of including the dustbin in these simulations. (ii) The effect of the gas outlet tube length. (iii) The effect of the operating condition; e.g., the gas Reynolds number (it includes the effect of inlet velocity, gas density, and viscosity), the operating pressure, and temperature. Moreover, it may be worthwhile to perform these investigations experimentally (using the particle image velocimetry (PIV) technique or any other ow visualizations technique) to check also if these phenomena (observations) come only from the numerical solution (schemes, settings, boundary conditions, etc.).
6.3.3
The cyclone performance
In order to estimate the effect of cyclone inlet dimensions on the performance parameters, the pressure drops have been calculated and presented in Fig. 6.8. A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 micron. 6.3.3.1 The effect of the inlet height Figure 6.8 shows a rapid decrease in the pressure drop when increasing the inlet height for 0.25 a/D 0.4 and a smaller decrease for a/D 0.4. This behavior can be explained as follows. The pressure drop in the cyclone is composed of three main contributions: (1) the pressure drop at the inlet section (decreased by increasing the inlet dimensions). (2) the pressure drop in the cyclone body due to swirling motion and due to wall friction, this contribution decreases with increasing the cyclone inlet height (as the vortex strength will decrease). (3) the main contribution to the cyclone 120
6.3. Results
A1
The static pressure N/m2
A2
A3
Figure 6.6: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom : the static pressure N/m2 , the tangential velocity m/s and the axial velocity m/s. From left to right cyclone A1 through cyclone A3.
The axial velocity m/s
The tangential velocity m/s
121
B1
The static pressure N/m2
B2
B3
Figure 6.7: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom : the static pressure N/m2 , the tangential velocity m/s and the axial velocity m/s. From left to right cyclone B1 through cyclone B3.
The axial velocity m/s
The tangential velocity m/s
122
6.3. Results
pressure drop is the energy loss in the exit tube, which mainly depends on the maximum tangential velocity in the cyclone. As is clear from Fig. 6.5 the maximum tangential velocity decreases with increasing cyclone inlet height. Consequently, in general the total cyclone pressure drop will decrease with increasing the cyclone inlet height. Figure 6.8 also depicts the pressure drop for the three cyclones using four different mathematical models; Muschelknautz method of modeling (MM) [29, 175], Casal and Martinez-Benet [21], Shepherd and Lapple [157], Stairmand [166] indicating the same overall trend obtained using RSM simulations for the pressure drop with better matching between the MM model and the current RSM simulation. The effect of the cyclone inlet height on the cut-off diameter (particle diameter of 50% collection efciency) is shown in Fig. 6.8. The general trend is an increase of the cut-off diameter with increasing cyclone inlet height, due to weakness of the vortex. Figure 6.8 also depicts the cut-off diameter for the three cyclones using two mathematical models; Iozia and Leith [85] and Rietema [142] indicating the same overall trend obtained using RSM simulations for the cut-off diameter with exact matching between the current RSM results and Rietema model.
6.3.3.2 The effect of the inlet width Figure 6.8 shows a rapid decrease in the pressure drop when increasing the inlet width for 0.15 b/D 0.27 and a smaller decrease for b/D 0.27. Generally, both the cyclone pressure drop and the cut-off diameter decrease with increasing the cyclone inlet width. The used mathematical models again indicate the same overall trend obtained using RSM simulations for cut-off diameter and pressure drop. Figure 6.8 indicates the need of applying a multi-objective optimization procedure to get the optimum value for the inlet dimensions. Both the collection efciency (cut-off diameter) and the pressure drop in cyclone separator are important objective functions to be optimized simultaneously [146]. The effects of changing the cyclone inlet dimensions on pressure drop and collection efciency are opposite. Increasing the inlet width will save more driving power but leads to reduced collection efciency (larger cut-off diameter). From the graph, the optimum value will be close to b/D = 0.25 with large reduction in pressure drop (energy losses) and small increase in cut-off diameter. Elsayed and Lacor [50] estimated the optimum value for inlet width (b/D) equals 0.236. 123
6.3.3.3 The inlet height versus the inlet width Figure 6.9 shows that, the impact of changing either the inlet height or width on the pressure drop is almost the same. However, the effect of changing the inlet width on the cut-off diameter is more signicant in comparison with that of the inlet height. When plotting the pressure drop against the ratio of inlet width to inlet height, it becomes clear the optimum range of b/a is from 0.5 until 0.7.
6.4 Conclusions
Five cyclones of different inlet width and height have been simulated, using the Reynolds stress model (RSM), to study the effect of cyclone inlet dimensions on the cyclone separator performance and ow pattern. The following conclusions have been obtained. The maximum tangential velocity in the cyclone decreases with increasing both the cyclone inlet width and height. No acceleration occurs in the cyclone space (the maximum tangential velocity nearly constant throughout the cyclone). The axial variation of both the static pressure and axial velocity is very limited. Increasing the cyclone inlet width or height decreases the pressure drop at the cost of increasing the cut-off diameter. So an optimization procedure is needed to estimate the optimum value of inlet dimensions. Wider inlet cyclones (b/D > gap between the cyclone barrel and the vortex nder are not preferred. The effect of changing the inlet width on the cut-off diameter is more signicant in comparison with that of the inlet height. The optimum ratio of the inlet width to the inlet height b/a is from 0.5 to 0.7.
124
6.4. Conclusions
1000 900
RSM MM Casal Shepherd Stairmand RSM Iozia Rietma
2
Pressure drop Cut-off diameter
1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.55
Pressure drop [N/m 2]
800 700 600 500 400 300 200 100 0 0.2
0.25
0.3
0.35
0.4
0.45
0.5
a/D
1000 900
RSM MM Casal Shepherd Stairmand RSM Iozia Rietma Pressure drop Cut-off diameter
2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.4
800 700 600 500 400 300 200 100 0 0.1
0.15
0.2
0.25
0.3
0.35
b/D
Figure 6.8: The effect of inlet dimensions on the pressure drop and cut-off diameter using CFD simulations and different mathematical models.
125
Inlet width b/D

0.15 1000 900
Effect of inlet height Effect of inlet width Pressure drop Cut-off diameter
0.2
0.25
0.3
0.35
0.4 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.55
800 700 600 500 400 300 200 100 0 0.2
0.25
0.3
0.35
0.4
0.45
0.5
Inlet height a/D
400
Pressure drop [N/m ] Cut-off diameter [micron]
2
2.6 2.4 2.2
300
2 1.8 1.6
200 1.4 1.2 100 1 0.8 0.6 0 0.2 0.4 0.6 0.8 1 0.4 1.2
b/a
Figure 6.9: Comparison between the effect of inlet height and width on the pressure drop and cut-off diameter using CFD simulations.
126
Chapter 7 The Cyclone Height

7.1 Introduction
Numerous studies have been performed for the effect of geometrical parameters on the ow pattern and performance [15, 52, 62, 102, 140, 184] whereas only limited number of studies have been devoted to the effect of the cyclone height. Zhu and Lee [200] have conducted detailed experiments on cyclones of different height and found that the cyclone height can inuence considerably the separation efciency of the cyclones. However, they did not provide any information about the ow pattern nor an explanation for the efciency results. Hoffmann et al. [76] investigated the effect of the cyclone height on the separation efciency and the pressure drop experimentally and theoretically. They found improvement in cyclone performance with increasing the total height Ht up to 5.5 times cyclone diameters beyond this length the separation efciency was dramatically reduced. However, they did not present any contour plot or velocity prole to support the explanation for the effect of the cyclone height on the performance. Recently, Xiang and Lee [186] have repeated the study of Zhu and Lee [200] for the effect of cyclone height via steady three-dimensional simulations using the Reynolds stress turbulence model (RSM). They found that the tangential velocity decreases with increasing the cyclone height, which is responsible for the lower separation efciency observed in long cyclones. The reason for this behavior however, was not analyzed in detail. Moreover, no particle tracking study was presented. The present study is intended to computationally investigate the effect of increasing cyclone (barrel and cone) height on the pressure drop and cutoff diameter and to obtain more details about the ow eld pattern and velocity proles. A RANS approach with the Reynolds stress turbulence 127
Chapter 7. The Cyclone Height
(b) The surface mesh for cyclone C1 (B2)
Figure 7.1: Schematic diagram for the cyclone separator
model (RSM) is used.

7.2.1 Conguration of the tested cyclones
The numerical simulations were performed on four cyclones with different barrel heights (at constant cone height) and three cyclones with different cone heights (at constant barrel height). Figure 7.1 and Table 7.1 give the cyclones dimensions.
7.2.2
Boundary conditions
Table 7.2 shows the details of the boundary conditions. The air volume ow rate Qin =50 L/min for all cyclones, air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 Pa s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [52, 74]. 128

Table 7.1: The geometrical dimensions of the tested cyclonesa
Dimension Inlet height, a Inlet width, b Gas outlet diameter, Dx Vortex nder insertion length, S Cone tip-diameter, Bc Cone height, hc h/D = 1.5 Cyclone b dimension/D 0.375 0.2625 0.5 0.5 0.375 2.5 3.5 4.5 1.0 1.5 2.0 2.5 hc /h Ht c
C1 C2 C3 B1 B2 B3 B4
1.666 2.333 3.0 2.5 1.666 1.25 1.0
4.0 5.0 6.0 3.5 4.0 4.5 5.5
Barrel height, h
hc /D = 2.5
a Body diameter, D = 31 mm. The outlet section is above the cylindrical barrel surface by Le = 0.5D. The inlet section located at a distance Li = D from the cyclone center. b The cone height is hc = Ht h, where Ht is the total cyclone height. Cyclone C1 and B2 are identical. c
Cyclone C2 and cyclone B4 are equal in Ht /D but they are different in both h/D and hc /D.
Grid independency study
A grid independency study has been performed for the tested cyclones. Three different grids have been tested for each cyclone, to be sure that the obtained results are mesh independent. For example, meshes with respectively 490164, 714029 and 1174029 cells have been used for cyclone B2. The computational results on the three grids are presented in Table 7.4. As it can be seen the maximum difference between the results is less than 5%, so the results on the 490164 cells grid can already be considered as sufciently accurate. However, for excluding any uncertainty, computations have been performed using the 714029 cells grid, where the total number of grid points was not that critical with respect to the computation overhead [10, 52]. Figure 7.1(b) shows the surface grid of cyclone C1(B2) used in this study. Table 7.3 gives the total number of cells used for each cyclone after the grid independency study. The hexahedral grids were generated using the GAMBIT grid generator.
Table 7.2: The boundary conditions Boundary Condition Inlet Velocity inlet Outlet Outow Cone tip Wall (no-slip) Other surfaces Wall (no-slip)
129

Table 7.3: The details of the tested cyclonesa Cyclone Number of cells
a
B1 688170
B2 (C1) 714029
B3 712183
B4 786865
C2 770556
C3 820362
The total number of hexahedral cells after the grid independency study
7.3 Results
The ow velocity can be decomposed into three components. The tangential and the axial velocity components are the major velocity components in comparison with the radial velocity component. Xiang and Lee [186] stated that the tangential velocity is the dominant gas velocity in gas cyclones, which results in the centrifugal force for particle separation. The axial component is responsible for the two ow streams (downward and upward).
7.3.1
The axial variation
Figures 7.2 - 7.4 present the radial proles of the time-averaged tangential and axial velocity and static pressure at nine axial stations (cf. Table 7.5). As expected, the tangential velocity proles exhibit the so-called Rankine vortex, which consists of two parts, an outer free vortex and an inner solid rotation in the center (Fig. 7.2). The tangential velocity distribution in the inner region is rather similar at different sections for the same cyclone. In the outer region, due to the sharp drop in velocity magnitude in the near wall region, the distribution is different but the maximum tangential velocity is similar at all sections for the same cyclone. The axial velocity proles at nine different stations are shown in Fig. 7.3. All cyclones show the inverted W prole but the central dip decreases with increasing the
Table 7.4: The details of the grid independence study for cyclone C1 (B2)
Number of cells 490164 714029 1174029 % differencec
a
Eua 3.475 3.654 3.573 2.74
Stk50 x103 b 1.188 1.2 1.224 2.94
2 1 Euler number is the dimensionless pressure drop Eu = P/( 2 Vin ) where P is the static pressure drop, is the gas density, Vin is the gas inlet velocity.
b The Stokes number based on the cut-off diameter; Stk50 = p x2 50 Vin /(18D ) [37]. It is the ratio between the particle relaxation time; p x2 50 /(18) and the gas ow integral time scale; D/Vin where p is the particle density =860 kg/m3 , is the gas viscosity. c
130
7.3. Results
Table 7.5: The position of different plotting sections Section z`/Da
a
S1 2.75
S2 2.5
S3 2.25
S4 2
S5 1.75
S6 1.5
S7 1.25
S8 1.0
S9 0.75
z`measured from the inlet section top (cf. Fig. 7.1(a)).
barrel (or cone) height (Except at the cyclone bottom for cyclones C2 and C3, where the axial velocity shows an inverted V prole). The radial proles of the time averaged static pressure are given in Fig. 7.4. Like for the tangential velocity, the axial variations of the static pressure are very small for the same cyclone. The variations become negligible with increasing barrel (or cone) height. Furthermore, the maximum value of the static pressure decreases when the barrel height is increased.
7.3.2 The radial variation

The tangential and axial velocity proles at section S6 (as a representative for the other sections, because the axial variations in the ow variables are small) for the six cyclones are compared in Fig. 7.5. The variation of axial velocity with changing barrel height is limited close to the wall, especially in the cylindrical part. In the central region, the change in axial velocity prole is more pronounced with the dip in the axial velocity decreasing with increasing the barrel height. This is the result of the atting of the pressure distribution results in a smaller pressure force. This may explain also the change of the axial velocity from cyclones B1 to B4. Increasing the cyclone barrel height decreases the maximum tangential velocity. Cyclone B1 has the maximum tangential velocity in comparison with the other cyclones. The effect of increasing the barrel height on the maximum tangential velocity is limited. The variation of axial velocity in the three C cyclones is limited close to the wall with changing the cone height. The axial velocity proles are very similar except at the central region due to change in the axial velocity prole. Increasing the cyclone cone height decreases the maximum tangential velocity. Cyclone C1 has the maximum tangential velocity in comparison with the two other cyclones (C2 and C3). From the comparison between the radial proles of the four B cyclones, the minimum pressure at the cyclone center is almost the same for all cyclones (B1 - B4). The slope of the static pressure radial prole becomes atter with increasing the barrel height. The minimum pressure at the cyclone center is almost the same for all C cyclones. The static pressure radial proles of cyclones C2 and C3 are very close. Increasing the cyclone 131

C1 (B2)
12
12
B1
10
10
4
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
Radial position (m)
C2
12
12
B3
10
10
4
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
Radial position (m)
C3
12
12
B4
10
10
4
S1 S2 S3 S4 S5 S6 S7 S8 S9
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2 -0.015
-0.01
-0.005
0.005
0.01
0.015
0 -0.015
Radial position (m)
Figure 7.2: The radial prole for the time-averaged tangential velocity at different sections. Note: C1=B2.
132
7.3. Results
C1 (B2)
8
S1 S2 S3 S4 S5 S6 S7 S8 S9
8
B1
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
Radial position (m)
0.01
0.015
C2
8
S1 S2 S3 S4 S5 S6 S7 S8 S9
8
B3
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
Radial position (m)
0.01
0.015
C3
8
S1 S2 S3 S4 S5 S6 S7 S8 S9
8
B4
S1 S2 S3 S4 S5 S6 S7 S8 S9
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
Radial position (m)
0.01
0.015
Figure 7.3: The radial prole for the time-averaged axial velocity at different sections. Note: C1=B2.
133

C1 (B2)
200
300 250
B1
150
200
100
50
150
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
Radial position (m)
C2
200
300 250
B3
150
200
100
50
150
100
S1 S2 S3 S4 S5 S6 S7 S8 S9
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
Radial position (m)
C3
150
300 250
B4
100
2
200
50
S1 S2 S3 S4 S5 S6 S7 S8 S9
150
100
50
-50
S1 S2 S3 S4 S5 S6 S7 S8 S9
-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
Radial position (m)
Figure 7.4: The radial prole for the time-averaged static pressure at different sections. Note: C1=B2.
134
7.3. Results
height (either barrel or cone) decreases the pressure drop, the maximum tangential velocity (vortex strength) and the dip in the axial velocity prole. The effect of increasing the cone height on the axial velocity prole is predominant with respect to the barrel height. The swirling motion of the gas generates a strong radial pressure gradient, the pressure being low in the centre of the vortex and high at the periphery. As the strongly swirling gas enters the connes of the vortex nder on its way out of the cyclone, the swirl is attenuated through friction with the wall. This means that further up the vortex nder the pressure in the centre is higher than at the exit of the separation space: a reverse pressure gradient is present [78] as is clear from Fig.7.5. This drives an axial ow with dip in the centre of the vortex nder (inverted W prole); this core ow prevails throughout the entire separation space of the cyclone in spite of the attenuation of swirl in the conical part of the cyclone. With increasing the cone height the pressure distribution becomes atter consequently the pressure force causes the dip in the axial velocity at the center line becomes less and less. That may explain also the change of the axial velocity from cyclones C1 to C3, Figs. 7.5 and 7.6.
7.3.3 The ow pattern

Figure 7.6 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones C1-C3. The time-averaged static pressure decreases radially from the wall to the center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, whereas the gradient in axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low-pressure zone at the cyclone center (twisted cylinder). Two vortical motions are exist one moving down (outer vortex) and the other moving up (inner vortex). The highest value of the static pressure decreases with increasing the cone height. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value decreases with increasing the cone height but the differences between cyclones C2 and C3 are small, so that a better collection efciency can be expected when decreasing the cone height. Figure 7.7 shows the contour plots of the time-averaged static pressure, tangential and axial velocity for cyclones B1-B4. The highest value of the static pressure decreases with increasing the barrel height. The tangential velocity pattern is very similar in all cyclones (Rankine prole). The highest value decreases with increasing the barrel height, but the differ135

250
300
h c/D=2.5 h c/D=3.5 h c/D=4.5
250
200
200
150
150
100
100
50
50
0
-50
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5

-50 -0.015
-0.01
-0.005
0.005
0.01
0.015
-100 -0.015
Radial position (m)
14 12 10
12
h c/D=2.5 h c/D=3.5 h c/D=4.5
10
8 6 4 2 0 -2 -0.015
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5

-0.01
-0.005
0.005
0.01
0.015
0 -0.015
Radial position (m)
h c/D=2.5 h c/D=3.5 h c/D=4.5
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5
-2
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
-4 -0.015
-0.01
Radial position (m)
0.01
0.015
136
7.3. Results
ences between the four cyclones are small. The axial velocity patterns for the four cyclones have the shape of an inverted W prole.
7.3.4 The performance

In order to estimate the effect of cyclone height on the performance parameters, the Euler number (the dimensionless pressure drop) have been calculated. A discrete phase modeling (DPM) study has been performed by injecting 104 particles from the inlet surface with a particle density of 860 kg/m3 and with a particle size ranging from 0.025 until 5 micron. Figure 7.8 and Table 7.6 show a sharp decrease of the Euler number with increasing the barrel height until h/D > 1.8 (Ht /D > 4.3) and a gradual decrease beyond. This behavior can be explained as follows. The pressure drop in the cyclone is composed of three main contributions [52]: (1) the pressure drop at the inlet section. (2) the pressure drop in the cyclone body due to swirling motion and due to wall friction, this contribution may increase with increasing the cyclone height as the wall friction will increase due to friction with a larger wall surface, or decreases as the vortex strength will decrease because the maximum tangential velocity decreases. (3) the main contribution to the cyclone pressure drop is the energy loss in the exit tube, which mainly depends on the maximum tangential velocity in the cyclone. As is clear from Fig. 7.5 the maximum tangential velocity decreases with increasing cyclone barrel height. As the inlet section is the same in all cyclones, the pressure drop in the inlet section does not vary with increasing barrel height. The sharp decrease of the Euler number between cyclones B1 and B2 is due to the decrease in the maximum tangential velocity. There are two competing contributions: increase of the pressure drop due to friction and decrease of the pressure drop due to the reduction in the vortex strength. At the beginning, the wall friction effect is small in comparison with the effect of vortex strength. For longer cyclones, this effect becomes larger (but still less than that of vortex strength decay). This explains the small variation of the Euler number with the barrel height for h/D > 1.8, which is clear from Fig. 7.5 where the maximum tangential velocity of cyclones B3 and B4 are very close. The behavior of the Stokes number curve as a function of barrel height is quite reasonable with increasing barrel height (separation space), the possibility of particles to be captured increases due to the increased cyclone space. However, the vortex strength decreased with a small amount due to the reduction of the maximum tangential velocity, the main contribution here is the collecting surface. The Stokes number curve becomes 137
Figure 7.6: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From left to right : the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From top to bottom cyclone C1-C3.
138
139
7.3. Results
Figure 7.7: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom: the static pressure (N/m2 ), the tangential velocity (m/s) and the axial velocity (m/s). From left to right cyclone B1 - B4.
nearly at between cyclones B3 and B4 due to the small changes in both the axial and the tangential velocity prole between the two cyclones. Both Ramachandran [139] and Iozia [84] models agree with the CFD results in the trend of decreasing both the Euler number and Stokes number with increasing barrel height, but differ in slope and values.
6 Euler number (CFD) Euler number (Ramachandran model) Stk50 x 10 3 (CFD) 3 Stk50 x 10 (Iozia model) 2.4 2.2 2 5 1.8 1.6 4.5 1.4 1.2 1 3.5 0.8 0.6 3 3.5 4
5.5
Euler number
H t /D
4.5
Figure 7.8: The effect of barrel height on the Euler number and the Stokes number.
Table 7.6: The Euler numbers and Stokes numbers for cyclones B1-B4 Cyclone h/D hc /h Ht /D Eu Eu (Ramachandran model [139]) Stk50 x 103 Stk50 x 103 (Iozia model [84]) B1 1.0 2.5 3.5 4.39 5.71 1.32 1.94 B2 1.5 1.666 4 3.654 4.77 1.2 1.82 B3 2.0 1.25 4.5 3.33 4.17 1.01 1.72 B4 2.5 1.0 5.0 3.09 3.73 0.95 1.64
Figure 7.9 and Table 7.7 show a sharp decrease of both the Euler number and the Stokes number with increasing the cone height until hc /D=3.3 and a gradual decrease when 3.3 > hc /h > 4.0. This behavior can be explained as follows. As the inlet section is the same in all cyclones, the pressure drop in the inlet section may not vary with increasing the cone height. 140
Stk50 x 10 3
7.3. Results
The sharp decrease in the Euler number between cyclones C1 and C2 is due to the huge decrease in the pressure drop in the cyclone body due to the drop in the maximum tangential velocity (the decay of the vortex strength). There are two competing contributions, increase of the pressure drop due to friction and decrease of the pressure drop due to the reduction in the vortex strength. At the beginning, the decay in the pressure drop due to the decay of the vortex strength overrides the effect of increasing the pressure drop due to wall friction for longer cyclones. This explains the small variation of the Euler number with the cone height for hc > 4.0, which is clear from Fig. 7.5 where the maximum tangential velocities of cyclones C2 and C3 are very close. The trend of decreasing Stokes number with increasing cone height is quite reasonable, as more separation space exists, and the possibility of particles to be captured increases. Although, the vortex strength decreased with a small amount - due to the reduction in the maximum tangential velocity- we estimate that the main contribution to the collection efciency comes from the increased collecting surface with increasing the cone height. The reason of nearly constant Stokes number after Ht /D =5.5 is the change in the axial velocity prole. Figure 7.5 shows a higher kinetic energy of the ow at the cyclone bottom for cyclone C3, that will enhance re-entrainment of some of the captured particles to escape with the upward ow. Consequently low collection efciency and higher cut-off diameter (Stokes number). Because, the differences between the axial and tangential velocity prole between cyclones C2 and C3 are limited.
Table 7.7: The Euler numbers and Stokes numbers for cyclones C1-C3 Cyclone hc /h hc /D Ht /D Eu Stk50 x 103 C1 1.666 2.5 4 3.654 1.2 C2 2.333 3.5 5 2.749 0.465 C3 3.0 4.5 6 2.584 0.315
7.3.5 The cone height versus the barrel height

Increasing the cyclone total height (either by increasing the cone or the barrel height) will decrease the maximum tangential velocity. The effect of cone height on the axial velocity prole is much larger than that of the barrel height. Both the Euler and the Stokes numbers decrease with increasing the 141
1.5 3.5 4 4
2 4.5
hc /h H t /D
5
2.5 5.5
3 6
Euler number Stk50 x 10 3
3.5 6.5 4
3.5
3.5 3
Euler number
2.5
2 1.5 1
1.5 0.5 1 0
2.5
3.5
4.5
hc /D
Figure 7.9: The effect of cone height on the pressure drop (Euler number) and cut-off diameter (Stokes number).
h c/D=2.5 h c/D=3.5 h c/D=4.5
-2
-4 -0.015
-0.01
-0.005
0.005
0.01
0.015
Radial position (m)
142
Stk50 x 10
2.5
7.3. Results
total height either by increasing the barrel or cone height. The effect of changing the cone height is more important than that of changing the barrel height (cf. Table 7.6, 7.7 and Fig. 7.11). The effect of increasing the ratio of cone to barrel height hc /h on the cyclone performance depends on the dependent variable. If hc /h increases at constant cone height, the Euler number increases linearly with decreasing the barrel height. If hc /h increases at constant barrel height, the Euler number decreases with increasing the cone height. The effect of changing hc /h at constant cone height has a negligible effect on the Stokes number. Increasing hc /h at constant barrel height decreases the Stokes number. This effect becomes negligible for hc /h > 2.75 (Fig. 7.11). This behavior can be explained by inspecting the variation of the time averaged static pressure, tangential and axial velocity proles with cone and barrel height, Fig. 7.12. As is clear from Fig. 7.12 the differences between the maximum tangential velocity in cyclones C2 and C3 are negligible. Furthermore, the effect of the cone height on the ow eld is more signicant than that of the barrel height.
143

5 4.5 4 3.5 4 Euler number (Barrel height) Euler number (Cone height) Stk50 x 10 3 (Barrel height) 3 Stk50 x 10 (Cone height) 6 5.5 5 4.5
Euler number
3.5 3 2.5 2 1.5
2.5 2 1.5 1
1 0.5 0 3 3.5 4 4.5 5 5.5 6 6.5 0.5
H t /D
6 Euler number (Barrel height) Euler number (Cone height) 3 Stk50 x 10 (Barrel height) 3 Stk50 x 10 (Cone height) 6
Euler number
0 0.75
1.25
1.5
1.75
2.25
2.5
2.75
0 3.25
hc /h
Figure 7.11: The variation of the Euler number and the Stokes number with the barrel and cone height.
144
Stk50 x 10 3
Stk50 x 10 3
7.3. Results
300 300
250
250
200
200
150
150
100
100
50
50
-50
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5 hc/D=3.5 hc/D=4.5

-50
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5 hc/D=3.5 hc/D=4.5

-100 -0.015
-100 -0.015
12
12
10
10
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5 hc/D=3.5 hc/D=4.5

h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5 hc/D=3.5 hc/D=4.5

-2 -0.015
-2 -0.015
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5 hc/D=3.5 hc/D=4.5
h/D=1.0 h/D=1.5 h/D=2.0 h/D=2.5 hc/D=3.5 hc/D=4.5

-2
-2
-4 -0.015
-4 -0.015 -0.01 -0.005 0 0.005 Radial position (m) 0.01 0.015
Figure 7.12: The radial prole for the time averaged static pressure, tangential and axial velocity at three different sections for the six cyclones. From top to bottom : static pressure, tangential and axial velocity. From left to right: S6 - S9. Note: h/D = 1.5 also represents hc /D = 2.5
145
7.4 Conclusions
Six cyclones of different barrel and cone height have been simulated using the Reynolds stress model (RSM), to study the effect of cyclone height on the performance and ow pattern. The following conclusions have been obtained. The maximum tangential velocity in the cyclone decreases with increasing the cyclone (barrel or cone) height. No acceleration occurs in the cyclone space (the maximum tangential velocity nearly constant throughout the same cyclone). Increasing the barrel height, makes a small change in the axial velocity. Increasing the cyclone barrel height decreases the pressure drop and the cut-off diameter. The changes in the performance beyond h/D = 1.8 are small. Increasing the cone height makes a considerable change in the axial velocity. Both the pressure drop and the cut-off diameter decrease with increasing the cyclone cone height. The performance improvement stops after hc /D = 4.0 (Ht /D = 5.5). The effect of changing the barrel height is less signicant on the performance and the ow pattern in comparison with the effect of the cone height.
146
Chapter 8 Optimization
8.1 Introduction
Mathematical optimization refers to the selection of a best element from some set of available alternatives. In the simplest case, this means solving problems in which one seeks to minimize or maximize a real function by systematically choosing the values of real or integer variables from within an allowed set. Generally, the use of the word optimization implies the best result under the circumstances [176]. In 1951, Stairmand [166] presented one of the most popular design guidelines for the high-efciency cyclone separators [147]. Stairmand presented the geometrical ratios for the seven geometrical parameters as: a/D = 0.5, b/D = 0.2, Dx /D = 0.5, Ht /D = 4.0, h/D = 1.5, S/D = 0.5 and Bc /D = 0.375. These values have been obtained based on the Stairmand model [165] which suffers from many shortages [50]. In the Stairmand model [165], the velocity distribution has been obtained from a moment-of-momentum balance, estimating the pressure drop as entrance and exit losses combined with the loss of static pressure in the swirl. The main drawbacks of the Stairmand model are: (1) neglecting the entrance loss by assuming no change of the inlet velocity occurs at the inlet area; (2) assuming a constant friction factor; (3) the effect of the particle mass loading on the pressure drop is not included. All these drawbacks are overcome in the Muschelknautz method of modeling (MM) [77] introduced by Muschelknautz and Trefz [116, 117]. The main benet of MM over other models is its ability to take the following effects into account: a) wall roughness due to both the physical roughness of the materials of construction and to the presence of collected solids. b) the effect of the mass loading and the Reynolds number on cyclone performance. c) the change 147
Chapter 8. Optimization
of ow velocity throughout the cyclone [77]. Consequently, the optimization procedure using a data obtained from MM model via response surface methodology will result in better results than the simple Stairmand model. An alternative approach is using the available experimental data in literature using the articial neural networks (ANNs) approach for the optimization process. The third source of data is the CFD simulations to train the ANNs. A general unconstrained optimization problem can be expressed as [182]: min subject to fi (x), x = [x1 , x2 , . . . , xn ]T , x X where x = [x1 , x2 , . . . , xn ]T is the vector of design variables (cyclone dimensions) with total number of n, X denotes the design space. The fi (x) stands for objective functions with total number of I . I should be larger than or at least equal to 1, which suggests the multi-objective or single objective optimal problem. In Eq. 8.1, f1 (x) is the Euler number and f2 (x) is the cut-off diameter. The design variables are seven geometrical parameters, namely, Dx /D, a/D, b/D, Ht /D, h/D, S/D and Bc /D. The side constraints are: 0.3 Dx /D 0.65, 0.2 a/D 0.65, 0.15 b/D 0.3, 3 Ht /D 6, 0.75 h/D 2, 0.4 S/D 0.75 and 0.1 Bc /D 0.45. There are numerous optimization techniques that can be used for the geometry optimization in the cyclone separator. However, only two techniques have been used in this thesis (Nelder-Mead technique and the genetic algorithms (GA)) for robustness, and availability in the used software. In cases of multi-objective optimization studies, two approaches have been applied, the desirability function and the non-sorted dominated genetic algorithm (NSGA-II) techniques have been selected due to robustness, and availability. More details about the used optimization techniques are given in appendix B. A detailed literature review for the previous optimization studies is given in Sec. 2.5.6, page 21. The present study is an attempt to obtain new optimized cyclone separator designs based on the MM model, experimental data sets and CFD simulations data sets and to investigate the effect of each cyclone geometrical parameter on the cyclone performance using response surface methodology (RSM) and CFD simulations. Table 8.1 summaries the conducted optimization studies. 148 i = 1, 2, . . . , I (8.1)
8.2. Single-objective using MM model

Table 8.1: Summary of optimization studies
No. 1 2 3 4 Design variables 7 7 7 4 Objective functions 1 1 2 2 Source of data MM model Experimental data (Exp.) Exp. and Iozia and Leith model CFD simulations data Meta-model RSM ANN ANN ANN Optimization technique Nelder-Mead Nelder-Mead GA and NSGA-II Desirability function, GA and NSGA-II
8.2 Optimization of the Euler number using MM model and Nelder-Mead technique
The present study aims to obtain a new optimized cyclone separator for minimum pressure drop based on the MM model and to investigate the effect of each cyclone geometrical parameter on the cyclone performance using CFD simulations. This section is an extension to the sensitivity analysis study presented in Sec. 4.1, page 45.
Dx
Le
h Li D b
Ht
Bc
Figure 8.1: Schematic diagram for Stairmand cyclone separator
Table 8.2 gives the optimum values for cyclone geometrical parameters for minimum pressure drop estimated by MM using the downhill simplex optimization technique available in Statgraphics XV software. 149
Table 8.2: The geometrical parameters for minimum pressure drop using MM model
Factor a/D b/D Dx /D Ht /D h S Bc Low 0.5 0.14 0.2 3 1.0 0.4 0.2 High 0.75 0.4 0.75 7 2.0 2 0.4 Optimum 0.618 0.236 0.622 4.236 1.618 0.620 0.382
8.2.1
CFD comparison between the two designs
Numerical settings The air volume ow rate Qin =0.08 m3 /s for the two cyclones (inlet velocity for Stairmand design is 19 m/s and 13.1 m/s for the new design), air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 P a s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. Velocity inlet boundary condition is applied at inlet, outow at the gas outlet and wall boundary condition at all other boundaries. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001 s. The residence time (cyclone volume/gas volume ow rate) of the two cyclones are close ( 0.25 s). The grid renement study shows that a total number of about 134759 hexahedral cells for Stairmand cyclone and 154746 hexahedral cells for the new design are sufcient to obtain a grid-independent solution, and further mesh renement yields only small, insignicant changes in the numerical solution. These simulations were performed on an eight nodes CPU Opteron 64 Linux cluster using Fluent commercial software. The geometrical values are given in Table 8.3 for the two cyclones (cf. Fig. 8.1).
Table 8.3: The values of geometrical parameters for the two designs (D=0.205 m)
Cyclone Stairmand design New design a/D 0.5 0.618 b/D 0.2 0.236 Dx /D 0.5 0.622 Ht /D 4 4.236 h/D 1.5 1.618 S/D 0.5 0.620 Bc /D 0.36 0.382 Li /D 1.0 1.0 Le /D 0.618 1.618
150

Table 8.4: The position of different sectionsa
Section z/Db
a
S1 2.75
S2 2.5
S3 2.25
S4 2.0
S5 1.75
S6 1.5
S7 0.25
Sections S1S5 are located in the conical section, section S6 at the cylindrical part and S7 located through the inlet section. b z measured from the inlet section top
Results and discussion

The pressure eld Figure 8.2 shows the contour plot at Y=0 and at section S7 (at the middle of inlet section, Table 8.4). In the two cyclones, the time-averaged static pressure decreases radially from the wall to center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, while the gradient in the axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low pressure zone at the cyclone center (twisted cylinder). However, the two cyclones have almost the same ow pattern, but the highest pressure of the Stairmand design is nearly twice that of the new design, implying that the new design has a lower pressure drop. The pressure distributions presented in Figs. 8.3 and 8.4 of the two cyclones at sections S1S6 depict the two parts pressure prole (for Rankine vortex). Once again, the highest static pressure for Stairmand design is more than twice that of the new design at all sections while the central value is almost the same for the two cyclones irrespective to the section location. This indicates that, the new design has a lower pressure drop with respect to the Stairmand design. The velocity eld Based on the contour plots of the timeaveraged tangential velocity, Fig. 8.2, and the radial proles at sections S1S6 shown in Figs. 8.3 and 8.4, the following comments can be drawn. The tangential velocity prole at any section is composed of two regions, an inner and an outer one. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer part of the prole (free vortex). This prole is a so-called Rankine type vortex as mentioned 151
Figure 8.2: The contour plots for the time averaged ow variables at sections Y=0 and S7. From top to bottom: static pressure [N/m2 ], tangential velocity [m/s] and axial velocity [m/s]. From left to right Stairmand design and new design respectively.
152

Table 8.5: The performance parameters for the two cyclones
Stairmand design New design Method MM CFD MM CFD Eu [-] 5.79 6.592 5.24 5.672 p [N/m2 ] 1045 1190 450 487 x50 [m] 1.54 1.0 1.77 1.6
before, including a quasi-forced vortex in the central region and a quasifree vortex in the outer region. The maximum tangential velocity may reach twice the average inlet velocity and occurs in the annular cylindrical part. The tangential velocity distributions for the two cyclones are nearly identical in pattern and values (dimensionless), with the highest velocity occurring at 1/4 of the cyclone radius for both cyclones. This implies a nearly equal collection efciency for both cyclones, as the centrifugal force is the main driving force for particle collection in the cyclone separator. The axial velocity proles for the two cyclones are also very close, exhibiting a M letter shape (also known as inverted W axial velocity prole in some other literatures (cf. Horvath et al. [79])). Part of the ow in the central region moves downward in the two cyclones. This phenomena has been shown in the axial velocity pattern in other published articles [e.g., 79, 159].
The DPM results In order to calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a velocity equals the inlet gas velocity. The particle density p is 860 kg/m3 . The grade efciency curves for the two designs are plotted in Fig. 8.5. The DPM analysis results and the pressure drops for the two cyclones are depicted in Table 8.5. An acceptable agreement between the CFD results and the MM mathematical model has been obtained. While the difference between the two cyclone cut-off diameters is small, the saving in pressure drop is considerable (nearly half the value of Stairmand cyclone). Based on the ow pattern analysis and the DPM results. One can conclude that the cyclone collection efciency for the two cyclones should be very close, with the advantage of low pressure drop in the new design. The authors want to emphasis that only small changes in the geometrical dimensions of the two designs led to this improvement in the performance. 153
8.2.2
Conclusions
Both mathematical modeling (the Muschelknautz method of modeling (MM)) and CFD investigation have been used to understand the effect of the cyclone geometrical parameters on the cyclone performance and a new optimized cyclone geometrical ratios based on MM model has been obtained. The new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, but superior in low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly one-half the pressure drop obtained by the old Stairmand design at the same volume ow rate.
154
Figure 8.3: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S1S3 . From top to bottom: section S1S3. From left to right: time-averaged static pressure, tangential velocity and axial velocity respectively.
155
Figure 8.4: The radial prole for the timeaveraged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S4S6 . From top to bottom: section S4S6. From left to right: time-averaged static pressure, tangential velocity and axial velocity respectively.
156
1 Stairmand design New design
0.75
Collection efficiency [-]
0.5
0.25
10
-1
10
10
Figure 8.5: The grade efciency curves for the two designs
157
8.3 Optimization of the Euler number using RBFNN and Nelder-Mead technique
Recently, Articial neural networks (ANNs) have been widely applied in the elds of modeling, prediction, fault detection and process control. In the eld of performance evaluation for cyclone separators, unfortunately, ANNs have not been paid enough attention on their algorithmic advantages [197]. There are few articles about the application of neural network in the eld of cyclone separator. There are six main objectives of this study. (1) Application of the articial neural network to model the pressure drop using experimental dataset. (2) Investigation of the effect of the seven geometrical parameters on the pressure drop based on the trained RBFNN. (3) A detailed comparison between the experimental pressure drop values and the estimated values obtained from different mathematical models. In order to recommend the best mathematical model for future use. (4) Application of the response surface methodology to study the effect of each geometrical parameter on the pressure drop and test the interaction between these parameters using the trained RBFNN. (5) Obtaining the optimum design (geometrical ratios) for minimum pressure drop. (6) CFD study of the new cyclone separator and compare its performance and ow pattern with the Stairmand design.
8.3.1
Radial basis function neural networks (RBFNN)
Why RBFNN Radial basis function neural networks (RBFNNs) are powerful and interesting networks due to their rapid training, generality and simplicity [16]. Girosi and Poggio [65] and Hartman and Keeler [72] proved that RBFNNs are universal approximators and can approximate any continuous function with arbitrary accuracy. Training of these networks is very fast, and they are very good at interpolation [190]. Niros and Tsekouras [120] stated that radial basis function neural networks (RBFNN) have certain advantages over other types of neural networks including better approximation capabilities, simple network structure, and faster learning. Zhao and Su [197] tested three different types of articial neural network to model the pressure drop in cyclone separators, viz. the back propagation neural network (BPNN), the radial basis function neural network (RBFNN) and the generalized regression neural network (GRNN). They stated that compared with the BPNN and GRNN, the RBFNN provides superior prediction performance criteria, better capability of approximation and high robustness. 158
8.3. Single-objective using RBFNN
In this study, the radial basis function neural network (RBFNN) is employed to model the pressure drop in the cyclone separator. The structure of RBFNN The radial basis function neural network (RBFNN) is a kind of 3-layered forward network with multi inputs and multi outputs. The rst layer has (m) inputs, while the second is a hidden layer with (L) units, and the third layer has (n) outputs, Fig. 8.6. The transformation function from the inputs to the hidden units, varied radial basis functions (RBF) on different occasions, is nonlinear, whereas the mapping of the hidden units to the outputs is linear [180]. Theoretically, RBFNN has the ability of approaching nonlinear mapping arbitrarily, fr : Rm Rn , as dened by:
L
y = fr (x) = W0 +
i=1
Wi ( x ci )
(8.2)
where x Rm is the input vector, y Rn is the output vector, ci Rm (i = 1, 2, L) is the center vector, is the Euclidian norm, Wi Rm (i = 1, 2, L) is the weight vector, W0 Rm is the bias vector, and () is the Gaussian function, a non negative and nonlinear function with radial symmetry and attenuation versus center, in the form of: (v ) = exp(v 2 /2 2 ) where is a width constant (spread factor ) [180]. In order to comprehensively compare the model performance, the evaluation parameters, normalized mean squared error E 2 and correlation coefcient R are employed as follows [197]: E2 = 1 n 2 (yN i y N i ) n i=1 (8.4) (8.3)
where n is the number of test cases, y is the actual variable, y is the RBFNN output variable and N i is the neurons number of input layer in the RBFNN. N i y N n i=1 (yN i y N ) y (8.5) R= 2 2 n ( y y n y ) y N i N i N N i=1 i=1 159
Input layer
Hidden layer
Output layer
W0 x1
y1
x2
y2
xm
yn Wi
Figure 8.6: Schematic diagram for the radial basis function neural network
Modeling procedures
The ANN modeling procedure can be divided into the following steps: selecting the variables, dividing the sample, optimizing parameters, training and testing simulation and evaluating performance [197]. In this study, all calculations are carried out on a MacBook pro laptop with the hardware congurations: processor, Intel Core 2 Duo (2.4 GHz); memory, 4.0 GB (DDR3-1067 2G 2); hard drive, 320 GB (7200 rpm); with Mac OS X 10.5 system. The Euler number The pressure drop across a cyclone separator essentially depends on the dimensions and operating conditions. Generally, it is proportional to the average dynamic pressure at the inlet and is often dened as [197] P = Eu
2 1 2 g Vin
(8.6)
where Eu is Euler number (the dimensionless pressure drop also called the pressure drop coefcient [197]). The Euler number is a complex nonlinear function of the cyclone geometrical dimensions and is not affected by operating conditions in the high Reynolds number ( Re > 5E4) [50, 77]. The Euler number will be constant for any cyclone conguration regardless of size as long as the dimension ratios remain the same, although the pressure drop varies with different operating conditions (due to the effect 160
of g and Vin ). Therefore, pressure drop can be established by determining experimentally or theoretically for a particular cyclone design and also be modied by the semi-empirical correlations to take the effect of solid loading [197]. In order to determine the Euler number more accurately, all eight dimensions of the cyclone are selected to establish the ANN models because they have the effect on the Euler number to different extent [139, 197]. Usually, these dimensions can be characterized by the barrel diameter D and expressed as seven dimensionless geometric ratios [197]: Eu = f D x a b S Ht h B c , , , , , , D D D D D D D (8.7)
According to Eq. 8.7, seven independent dimensionless geometrical variables and one dependent variable (the Euler number of the cyclone) are selected as respectively the input and output parameters in the ANN model, as presented in Table 8.6. For simplicity, the division of each factor by the barrel diameter D will be dropped. A dataset of 98 samples obtained from the measurements of pressure drop for different cyclone designs available in the literature [42, 139, 197] is used in the present investigation to evaluate the prediction performance of the ANN models. Table 8.7 presents more details about the used dataset including the minimum, mean, maximum and range of the seven dimensionless geometrical ratios. Due to the large difference in the order of magnitude of the value (cf. Table 8.7), the available dataset is transformed into -1 to 1 interval using the Matlab intrinsic function; mapminmax in order to avoid solution divergence [197]. The ANN calculations have been performed using the neural network toolbox available from Matlab commercial software 2010a. Descriptive statistical parameters of the input dataset Table 8.8 shows Pearson product moment correlations between each pair of variables. These correlation coefcients range between -1 and +1 and measure the strength of the linear relationship between the variables. Moreover, shown in parentheses is the P-value which tests the statistical signifTable 8.6: The input and output variables for ANN model Variables Specication Input parameters X1 X2 X3 X4 Dx a b S X5 Ht X6 h X7 Bc Output diameter y Eu
161
Table 8.7: Descriptive statistical parameters for the training dataset
Variable Minimum Mean Maximum Range Dx 0.25 0.429 0.667 0.417 a 0.113 0.630 1.0 0.887 b 0.067 0.211 0.4 0.333 S 0.39 0.891 3.052 2.662 Ht 1.158 3.283 10.97 9.812 h 0.501 1.189 3.5 2.999 Bc 0.14 0.342 1.0 0.86
icance of the estimated correlations. P-values below 0.05 indicate statistically signicant non-zero correlations at the 95.0% condence level. The following pairs of variables have P-values below 0.05: (1) Dx with a and Bc ; (2) a with b, Ht and Bc ; (3) b with Ht and Bc ;(4) S with Ht , h and Bc ; (5) Ht with h and Bc ; (6) h with Bc . These conclusions can be obtained also from Fig. 8.7. From this analysis, however there are some correlations between the input variables, but this dataset is still reliable. The same dataset has been used successfully by other researchers [e.g., 194, 197]. Furthermore, it is the only available experimental data set in the literature. K-fold cross validation For the calibration of the RBFNN, the spread factor plays an important role in the regression model. To obtain the optimum value for a multistep search technique is used. In the multi-step search technique, the seeking for the optimum value is performed in two steps [197]. The rst step is a coarse search to determine the best range of values. The second search is a ne search in the best range. To avoid overtting, the K-fold cross validation has been employed [8, 80, 197]. The original sample of data is randomly portioned into K subsamples. A single subsample is used for the validation (testing) and the other remaining K-1 subsamples are used for the training. The process of training and testing is then repeated for each of the K possible choices of the subset omitted from the training. The average performance on the K omitted subsets is then our estimate of the generalization performance. This procedure has the advantage that is allows us to use a high proportion of the available data (a fraction 1 1/K) for training, while making use of all the data points in estimating the
Table 8.8: Correlations between each pair of input variables
Dx a Dx -0.377 (0.000) a -0.377 (0.000) b - 0.121 (0.234) 0.442 (0.000) S 0.092 (0.367) 0.148 (0.145) Ht 0.189 (0.061) 0.288 (0.004) h 0.096 (0.346) -0.052 (0.610) Bc 0.199 (0.049) 0.270 (0.007) The bold value shown in parentheses is the P-value. b -0.121 (0.234) 0.442 (0.000) 0.170 (0.093) 0.239 (0.017) 0.194 (0.054) 0.243 (0.015) S 0.092 ( 0.367) 0.148 (0.145) 0.170 (0.093) 0.378 (0.000) 0.685 (0.000) 0.526 (0.000) Ht 0.189 (0.061) 0.288 (0.004) 0.239 (0.017) 0.378 (0.000) 0.393 (0.000) 0.555 (0.000) h 0.096 (0.346) -0.052 (0.610) 0.194 (0.054) 0.685 (0.000) 0.393 (0.000) 0.470 (0.000) Bc 0.199 (0.049) 0.270 (0.007) 0.243 (0.015) 0.526 (0.000) 0.555 (0.000) 0.470 (0.000)
162
Figure 8.7: Qualitative representation of the correlations between each pair of input variables. For uncorrelated pair of variables, the data will be well distributed and no linear correlation can be obtained e.g., Dx and b (rst row, third column). The high correlation between Ht and h is clear (fth row, sixth column), where one can t easily t a straight line.
generalization error. The disadvantage is that we need to train the network K times. Typically K = 10 is considered reasonable and most widely used [18]. For the radial basis function neural networks (newrb in Matlab 2010a), the learning process is a must to obtain the weights (the width of the radial basis function units). In order to obtain the optimum value of the spread factor , the multi-step search technique with 10-fold cross validation from the interval of (0 - 1) with the performance goal of 1E-5, the maximum number of neurons in the hidden layer equals the training sample size of 98 neurons, the number of neurons to add between displays is 10. Based on the mentioned settings, the optimum value of equals 0.191 (E 2 =1.1321E-06), this value is different than that obtained by Zhao and Su [197] ( equals 0.32 (E 2 =5.84E-04)) the reason can be referred to the lower goal used in the current study (the goal used in Zhao and Su [197] was 1E-4, and all other settings are identical). 8.3.1.1 Fitting the ANN Table 8.9 presents more details about the validation of the used RBFNNs. Both the average, minimum, maximum and range of the input (Euler number) and the predicted Euler number are given. It is clear from Table 8.9 that the ANNs preserved the descriptive statistical parameters of the in163
Table 8.9: Validation of the used RBFNN
Experimental x y 23.268 23.268 2.3 1.745 155.3 155.985 153.0 154.24 0.999 1.311E-4 0.017 0.999 MM x 15.150 1.34 138.0 136.66 y 15.150 1.164 137.235 136.071 0.996 1.212E-4 0.012 0.999 Stairmand x y 20.206 20.206 2.88 2.892 132.0 133.521 129.12 130.629 0.999 9.185E-5 0.011 0.999 Ramachandran x y 22.543 22.543 1.85 1.793 153.0 150.854 151.15 149.061 0.999 1.442E-4 0.020 0.999 Shepherd x y 17.774 17.774 0.957 1.197 92.2 92.543 91.243 91.346 0.999 5.411E-5 0.005 1.0
Average Minimum Maximum Range Correlation Coefcient, R Mean squared error, E 2 Intercept Slope
x is the input to the RBFNN and y is the predicted value. Both x and y represent the Euler number. The values of R, E 2 , intercept and slope are that for the testing stage.
put data. The correlation coefcient between the input and the output and the mean squared error are given for each RBFNN. The intercept and the slope of the adjusted line between the input and the predicted value of the ANN are also given. The congured RBFNN predictions versus experimental data and four other models for cyclone Euler number are shown in Fig. 8.8. According to Fig. 8.8, it can be seen that the ANN models are able to attain the high training accuracy. The training mean square errors for the experimental values and the four mathematical models (MM, Stairmand, Ramachandran and Shepherd and Lapple), have the values 1.311E-4, 1.212E4, 9.185E-5, 1.442E-4 and 5.411E-5 respectively (Fig. 8.8). This indicates that, compared with traditional models of curve tting, the models based on articial intelligence algorithm have a superior capability of nonlinear tting. Especially, the RBFNN has its unique and optimal approximation characteristics in learning process [197]. Figure 8.8 illustrates the agreement between the ANN input and output. The obtained relation is a typical linear relation with a coefcient of correlation close to 1 (R > 0.999). The agreement between the input and output of the ANN is also clear from the value of the mean squared error E 2 (< 1.5E 4). That means, the trained neural network predicts very well the Euler number values and can be used in cyclone design and performance estimation. Table 8.9 and Fig. 8.8 present different performance indicators as a validation of the proposed model for experimental values.
8.3.2
Evaluation of different mathematical models
In order to evaluate the performance of the four tested mathematical models in comparison with the experimental values, the percentage residual 164

150 Data point Linear fit
Predicted value
100
50
50
100
150
Input value
(a) Experimental values
150
Data point Linear fit
150
Predicted value
50
Predicted value
50 100 150
100
100
50
50
100
150
Input value
Input value
(b) MM model
(c) Stairmand
100 150 Data point Linear fit 80 Data point Linear fit
Predicted value
Predicted value
50 100 150
100
60
40
50 20
20
40
60
80
100
Input value
Input value
(d) Ramachandran model
(e) Shepherd and Lapple model
Figure 8.8: Linear regression of the RBFNN.
165
400
Residual error %
300 200 100 0 -100 -200 0
MM Stairmand Ramachandran Shepherd
20
40
60
80
100
Run number
Figure 8.9: Percentage residual error for the four tested models based on 98 dataset.
error (Eq. 8.8) for each model has been plotted in Fig. 8.9. % error = Model value - Experimental value 100 Experimental value (8.8)
Figure 8.9 depicts that MM model underestimate the pressure drop by around 50%. Also the percentage errors for other models are between 50%. The peaks in error are almost the same for all models. The residual error of MM model is almost the lowest. The reasons of these peaks may be due to the high values of geometrical swirl number for some of the available dataset (cf. Ramachandran et al. [139]), where Sg = Dx D/(4a b). Sg varies between 1 and 5 for industrial cyclones [74]. The Sg values in the 98 dataset used in this study varies from 2.18 to 92.67. As the swirl number is a measure for the ratio of tangential to axial momentum [74, 75, 154], the high values of Sg may cause violation of the simplied assumptions used in the models. Figure 8.9 seems to indicate that the MM model among the other mathematical models is best suited for estimation of the pressure drop. The effect of geometrical parameters on the Euler number The effects of the geometrical parameters on the Euler number are depicted in Fig. 8.10. To study the effect of each parameter, the tested RBFNN model has been used by varying one parameter at a time from its minimum to maximum values of the available 98 dataset, while the other parameters are kept constant at their mean values (cf. Table 8.7). Figure 8.10(a) indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b and the total height Ht . Less effect is due to the cylinder height h (for h > 2.5) and the inlet height a (for a > 0.55). 166
The effect of the geometrical parameters on the Euler number obtained from the MM model (Fig. 8.10(b)) is very close to that obtained from the analysis of the experimental dataset but with underestimation of the Euler number. This supports the use of the MM model in the Euler number estimation for cyclone separators [29, 50, 77]. The situation for the Stairmand model is not the same as that for the experimental data or MM model. Here, the effect of many geometrical parameters attenuated. The effect of the inlet width becomes insignicant. That is not realistic, as for high values of inlet width, a considerable part of the incoming ow will impact directly the vortex nder which increases the entrance loss and consequently, the total pressure drop in the cyclone separator. The Ramachandran model was initially constructed by curve tting based on the used 98 dataset (cf., Ramachandran et al. [139] for more details). So it depicts a better agreement with the experimental values of the Euler number than that of Stairmand model. The Ramachandran model predicts almost the same Euler number variation with the inlet width. In the Shepherd and Lapple model only the inlet dimensions (aandb) and vortex nder diameter 2 Dx affect the pressure drop (Eu = 16ab/Dx ) as is clear from Fig. 8.10(e). Figure 8.11 compares the effect of each individual geometrical parameter using predictions with the ANN based on respectively the experimental data and the four tested models. The following conclusions can be drawn from the analysis of Fig. 8.11: Dx : All the models (except Shepherd and Lapple) show the same variation of the Euler number with increasing Dx . For Dx > 0.5 there is an exact matching between the results of MM model and the experimental values. a : Both the Ramachandran and the Shepherd and Lapple models predict a (nearly) linear relation between the inlet height and the Euler number. The effect of changing a predicted by the Stairmand model is in accordance with the experimental values for a > 0.8. b : For small values of b (b < 0.15), the MM model results are in good agreement with the experimental values. Also the trends of all models in this range are matching the trend of the experimental values. Beyond this range, Both the Ramachandran and the Stairmand models agree well with the experimental values trend. S : The trend of the MM model results is similar to that of the experimental values, although there is a shift in the values of the Euler number. The Shepherd and Lapple model does not present any effect of changing S on the Euler number (for Shepherd and Lapple model, 2 Eu = 16ab/Dx ), while the MM model always underestimates the ef167
Ht
6
10
h, S
0.5 1
Dx a b S Ht h Bc
1.5
2.5
3.5
40
Euler number
30
20
10
0 0.2 0.4 0.6 0.8 1
D x, a, b, Bc
(a) Experimental values
Ht
6
10
Ht
6
10
h, S
0.5 1
Dx a b S Ht h Bc
h, S
2.5 3 3.5 0.5 1
Dx a b S Ht h Bc
1.5
1.5
2.5
3.5
40
40
Euler number
Euler number
0.4 0.6 0.8 1
30
30
20
20
10
10
0 0.2
0 0.2 0.4 0.6 0.8 1
D x, a, b, Bc
D x, a, b, Bc
(b) MM model
(c) Stairmand model
Figure 8.10: The effect of geometrical parameters on the Euler number using the trained neural networks based on experimental dataset and four different mathematical models.
168
Ht
6
10
Ht
6
10
h, S
0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2
h, S
2.5 3 3.5
40
Euler number
20
Euler number
30
Dx a b S Ht h Bc
40
30
Dx a b S Ht h Bc
20
10
10
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
D x, a, b, Bc
D x, a, b, Bc
(d) Ramachandran model
(e) Shepherd and Lapple model
Figure 8.10: (continued) The effect of geometrical parameters on the Euler number using the trained neural networks based on experimental dataset and four different mathematical models.
fect of S . Both the Stairmand and the Ramachandran models may over/underestimate its effect on Euler number. Ht : Both the Stairmand and the Ramachandran models do not show a signicant effect of Ht on the Euler number. The trends of both MM and Ramachandran model are almost the same as that for the experimental values. h : For higher values of h all models (except the Shepherd and Lapple model) give nearly the same value of Euler number. Bc : Nearly, all the models (except the Shepherd and Lapple model) show the same trend in the changing the Euler number with the cone tip diameter.
8.3.3 Design of experiment (DOE)

Table 8.10 depicts the parameters ranges selected for the seven geometrical parameters. The study was planned using Box-Behnken design, with 64 combinations. A signicant level of P < 0.05 (95% condence) was used in all tests. Analysis of variance (ANOVA) was followed by an F-test of the 169
Table 8.10: The values of the independent variables used in the design of experiment Variables Vortex nder diameter, Dx =X1 Inlet height, a=X2 Inlet width, b=X3 Vortex nder length, S =X4 Total cyclone height, Ht =X5 Cylinder height, h=X6 Cone tip diameter, Bc =X7 minimum 0.2 0.4 0.14 0.4 3.0 1.0 0.2 center 0.475 0.55 0.27 1.2 5.0 1.5 0.3 maximum 0.75 0.7 0.4 2.0 7.0 2.0 0.4
individual factors and interactions. Fitting the model Analysis of variance (ANOVA) showed that the resultant quadratic polynomial models adequately represented the experimental data with the coefcient of multiple determination R2 being 0.965843. This indicates that the quadratic polynomial model obtained was adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (see Table 8.12). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables. Based on the ANOVA results presented in Table 8.12, the variable with the largest effect on the pressure drop (Euler number) was the linear term of vortex nder diameter, followed by the linear term of inlet width and vortex nder length (P < 0.05); the other four linear terms (inlet height, barrel height, cyclone total height and cone tip diameter) did not show a signicant effect (P > 0.05). The quadratic term of vortex nder diameter, vortex nder length and cyclone total height also had a signicant effect (P < 0.05) on the pressure drop; however, the effect of the other four quadratic terms was insignicant (P > 0.05). Furthermore, the interaction between the inlet dimensions and vortex nder diameters (P < 0.05) also had a signicant effect on the pressure drop, while the effect of the remaining terms was insignicant (P > 0.05). Analysis of response surfaces For visualization of the calculated factor, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve 170
Table 8.11: The geometrical dimensions and Euler number for the used cyclones
Exp. No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 New design Stairmand design X1 a/D 0.7 0.55 0.55 0.55 0.7 0.55 0.55 0.55 0.55 0.4 0.55 0.55 0.55 0.7 0.7 0.7 0.4 0.7 0.55 0.55 0.55 0.55 0.55 0.55 0.4 0.4 0.4 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.55 0.4 0.55 0.55 0.7 0.55 0.7 0.55 0.55 0.4 0.4 0.7 0.4 0.4 0.55 0.7 0.4 0.55 0.7 0.55 0.4 0.55 0.55 0.7 0.55 0.55 0.55 X2 b/D 0.27 0.27 0.27 0.27 0.4 0.14 0.4 0.27 0.27 0.4 0.27 0.4 0.14 0.4 0.27 0.27 0.4 0.27 0.4 0.27 0.27 0.14 0.4 0.27 0.14 0.27 0.27 0.27 0.27 0.27 0.4 0.4 0.14 0.27 0.14 0.27 0.14 0.27 0.27 0.27 0.27 0.14 0.14 0.27 0.4 0.27 0.27 0.27 0.27 0.27 0.27 0.27 0.14 0.27 0.14 0.27 0.27 0.14 0.27 0.27 0.27 0.27 0.4 0.27 X3 Dx /D 0.75 0.475 0.475 0.2 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.75 0.2 0.475 0.2 0.475 0.475 0.75 0.475 0.475 0.2 0.75 0.2 0.475 0.475 0.2 0.2 0.475 0.475 0.75 0.475 0.2 0.2 0.75 0.475 0.475 0.475 0.75 0.75 0.475 0.475 0.475 0.75 0.2 0.75 0.2 0.75 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.475 0.2 0.475 0.75 0.475 0.475 0.475 0.475 0.475 X4 Ht /D 5 7 3 7 7 5 5 5 3 7 5 5 5 3 5 5 3 5 5 7 3 5 5 5 7 5 5 5 5 7 5 5 5 3 5 5 5 3 5 5 7 3 5 5 5 7 5 5 5 5 5 3 7 5 5 5 3 3 7 3 5 5 5 7 X5 h/D 1 2 2 1.5 1.5 1 1 1.5 1 1.5 1.5 1.5 1.5 1.5 1 1.5 1.5 2 2 1 1.5 1.5 1.5 1.5 1.5 1 2 1.5 1.5 1.5 1 1.5 1.5 1.5 2 1.5 2 1.5 1 1.5 1 1.5 1.5 2 1.5 1.5 2 1.5 1.5 1.5 1.5 2 1.5 1.5 1 1.5 1.5 1.5 1.5 1 1.5 1.5 2 2 X6 S/D 1.2 2 0.4 1.2 1.2 1.2 1.2 1.2 2 1.2 1.2 0.4 0.4 1.2 1.2 0.4 1.2 1.2 1.2 0.4 1.2 0.4 0.4 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 2 2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 2 1.2 2 1.2 2 1.2 1.2 2 2 0.4 0.4 2 1.2 2 1.2 0.4 1.2 1.2 1.2 0.4 2 1.2 1.2 0.4 X7 Bc /D 0.3 0.3 0.3 0.2 0.3 0.4 0.4 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.4 0.3 0.3 0.2 0.3 0.2 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.2 0.2 0.3 0.3 0.4 0.4 0.3 0.2 0.2 0.3 0.3 0.3 0.3 0.3 0.3 0.3 0.4 0.3 0.4 0.4 0.2 0.4 0.3 0.3 0.2 0.2 0.2 0.4 0.3 0.4 0.3 0.2 0.3 0.4 0.3 Y1 Euler Number 4.16 6.93 9.30 65.70 11.90 4.34 9.19 7.94 8.87 5.88 7.94 3.24 24.50 13.40 105.00 10.30 7.05 4.07 9.19 7.50 77.50 2.42 130.00 7.94 3.17 41.90 40.70 7.94 7.94 3.35 9.45 128.00 23.80 3.82 4.22 7.94 4.33 3.86 2.99 7.94 7.09 6.03 2.23 103.00 3.05 64.10 2.92 5.70 9.64 6.34 6.12 8.51 4.70 5.88 4.51 10.60 76.30 4.06 3.25 9.77 9.91 7.94 9.01 7.30
0.618 0.5
0.236 0.2
0.618 0.5
4.236 4
1.618 1.5
0.618 0.5
0.3819 0.36
5.24 5.79
The new design based on MM model and downhill simplex optimization scheme (Euler number =5.24, cut-off diameter =1.77m ). The standard Stairmand high efciency cyclone design, (Euler number =5.79, cut-off diameter =1.54m based on MM model).
171
Table 8.12: Analysis of variance of the regression coefcients of the tted quadratic equationa
Variable 0 Linear 1 2 3 4 5 6 7 Quadratic 11 22 33 44 55 66 77 Interaction 12 13 14 15 16 17 23 24 25 26 27 34 35 36 37 45 46 47 56 57 67 R2
a
Regression coefcient 138.604 -485.694 20.2354 325.517 -10.7497 -8.7097 -19.1361 -167.261 515.706 113.392 -13.0018 10.9817 1.04634 6.06181 232.346 -158.258 -345.242 -4.21438 8.42261 31.8848 95.7341 -143.344 20.5571 -3.23107 -16.3805 -6.62067 -26.3282 0.0205865 32.7526 -28.156 -0.62389 -2.06662 -18.6467 -2.60425 2.35394 -5.32565 0.965843
F-Ratio
P-Value
96.89 0.06 59.28 6.87 0.06 0.57 0.60 489.19 2.09 0.02 15.89 5.63 0.74 1.74 7.61 27.22 0.15 3.83 3.43 1.24 1.40 1.09 0.17 0.27 0.00 1.34 0.00 0.81 0.02 0.18 0.12 0.40 1.21 0.04 0.01
0.0000 0.8152 0.0000 0.0140 0.8161 0.4575 0.4464 0.0000 0.1590 0.9017 0.0004 0.0247 0.3974 0.1983 0.0101 0.0000 0.6981 0.0602 0.0744 0.2752 0.2473 0.3060 0.6851 0.6076 0.9668 0.2569 0.9982 0.3758 0.8782 0.6763 0.7294 0.5334 0.2803 0.8437 0.9112
172
is proportional to the size of the effect and the direction of the curve species a positive or negative inuence of the effect [50, 61] (Fig.8.12(a)). Based on the main effect plot, the most signicant factors on the Euler number are: (1) the vortex nder diameter Dx , with a secondorder curve with a wide range of inverse relation and a narrow range of direct relation, (2) the inlet width b linearly related to the Euler number, (3) the cyclone total height Ht , (4) the vortex nder length S , whereas the other factors have an insignicant effect. Pareto charts were used to graphically summarize and display the relative importance of each parameter with respect to the Euler number [50]. The Pareto chart shows all the linear and second-order effects of the parameters within the model and estimates the signicance of each with respect to maximizing the Euler number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto charts judges whether each effect is statistically signicant within the generated response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 8.12(b)) and ANOVA table (Table 8.12) there are ve signicant parameters (eight terms in the ANOVA table; Table 8.12) at a 95% condence level: the vortex nder diameter Dx , the inlet width b, the total cyclone height Ht , the vortex nder length S and the inlet height a (due to interaction with Dx ). Therefore, the Pareto chart is a perfect supplement to the main effects plot. To visualize the effect of the independent variables on the dependent ones, surface response of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range while holding the other factors at their central values [189] as shown in Fig. 8.13. Thus, Fig. 8.13(b) was generated by varying the inlet height a and the inlet width b while keeping the other ve factors constant. The trend of the curve is linear, with a more signicant effect for the inlet width b, and a weak interaction between the inlet height a and width b. The response surface plots of Figs. 8.13(a), 8.13(c) and 8.13(d) show that there are strong interactions between the vortex nder diameter Dx and respectively the inlet height a, the cyclone total height Ht and the vortex nder length S . The effect of vortex nder length S is less signicant with respect to the vortex nder diameter Dx , but its effect is higher than that of the barrel height h and the cone tip diameter Bc (cf., Fig. 8.12). 173
8.3.3.1 Optimization (Nelder-Mead method) In this optimization problem, the objective function is the Euler number (f1 (x) in Eq. 8.1). Table 8.13 gives the optimum values for cyclone geometrical parameters for minimum pressure drop. It is clear from Table 8.13 that the new optimized design is very close to the Stairmand design in many geometrical parameters, whereas the new ratios will lead to the minimum pressure drop. To understand the effect of this small change in the geometrical ratios on the ow eld pattern and performance, a CFD study for the two designs is needed.
Table 8.13: The geometrical parameters for minimum pressure drop using RBFNN Factor Dx a b S Ht h Bc Low 0.2 0.5 0.14 0.4 3.0 1.0 0.2 High 0.75 0.75 0.4 2.0 7.0 2.0 0.4 Stairmand design 0.5 0.5 0.2 0.5 4.0 1.5 0.375 Optimum 0.487 0.629 0.203 0.733 4.852 1.633 0.383
174
50
Exp. MM Stairmand Ramchandran Shepherd
25
40
20
Euler number
30
Euler number
0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
15
20
10
10
0 0.2
0.2
0.4
0.6
0.8
Dx
(a) Dx
(b) a
30
65 60 55 50
25
20
45
Euler number
Euler number
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
40 35 30 25 20
15
10
15 10
5 0 0.5 1 1.5 2 2.5 3
(c) b
(d) S
Figure 8.11: Comparison between the effect of each geometrical parameters on the Euler number using different models and experimental values based on RBFNN.
175
20
Exp. MM Stairmand Ramchandran Shepherd Exp. MM Stairmand Ramchandran Shepherd
30
15
Euler number
20
Euler number
2 4 6 8 10 12
10
10 5
0.5
1.5
2.5
3.5
Ht
(e) Ht
(f) h
40
30
Euler number
20
10
0.2
0.4
0.6
0.8
Bc
(g) Bc
Figure 8.11: (continued) Comparison between the effect of each geometrical parameters on the Euler number using different models and experimental values based on RBFNN.
176
(a) Main effects plot
(b) Pareto chart
Figure 8.12: Analysis of design of experiment
177
(a) Dx versus a
(b) a versus b
(c) Dx versus Ht
(d) Dx versus S
Figure 8.13: The response surface plots.
178
8.3.4 CFD Comparison between the two designs

Numerical settings The air volume ow rate Qin =0.08 m3 /s for the two cyclones (inlet velocity for Stairmand design is 19 m/s and 14.9 m/s for the new design), air density 1.0 kg/m3 and dynamic viscosity of 2.11E-5 P a s. the turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. Velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall boundary condition at all other boundaries. The Reynolds stress turbulence model has been used to reveal the turbulence characteristics in the cyclone separators. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001 s. The grid renement study using different levels of grid shows that a total number of 134759 hexahedral cells for the Stairmand cyclone and 381709 hexahedral cells for the new design are sufcient to obtain a grid-independent solution, and further mesh renement yields only small, insignicant changes in the numerical solution. The hexahedral meshes have been obtained using the GAMBIT commercial software. These simulations were performed on an 8 nodes CPU Opteron 64 Linux cluster using Fluent 6.3.26 commercial software. The geometrical values for the two cyclones (cf. Fig. 8.1) are given in Table 8.14. The surface mesh for Stairmand cyclone is given in Fig. 8.14(a).
Table 8.14: The values of geometrical parameters for the two designs (D=0.205 m)
Cyclone Stairmand design New design a/D 0.5 0.628 b/D 0.2 0.203 Dx /D 0.5 0.487 Ht /D 4 4.852 h/D 1.5 1.633 S/D 0.5 0.732 Bc /D 0.375 0.382
The outlet section is above the cyclone surface by Le = 0.618D. The inlet section located at a distance Li = D from the cyclone center.
179
(a) The Stairmand design
(b) The new design
Figure 8.14: The surface meshes for the two designs

The pressure eld Figure 8.15 shows the contour plot at Y=0. In the two cyclones, the timeaveraged static pressure decreases radially from the wall to center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, while the gradient in the axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low pressure zone at the cyclone center (twisted cylinder). However, the two cyclones have almost the same ow pattern, but the highest pressure of the Stairmand design is nearly 1.5 times that of the new design, implying that the new design has a lower pressure drop. The pressure distribution presented in Figs. 8.16 and 8.17 of the two cyclones at sections S1S6 depict the two parts pressure prole (for Rankine vortex). Again, the highest static pressure for Stairmand design is more than 1.5 times that of the new design at all sections while the central value is almost the same for the two cyclones irrespective of the section location. This indicates that, the new design has a lower pressure drop with respect to the Stairmand design. 180
Figure 8.15: The contour plots for the time averaged ow variables at sections Y=0 (cf. Fig.8.1). From top to bottom: Stairmand design and the new design respectively. From left to right: the static pressure (N/m2 ), the tangential and axial velocity (m/s). Note: both cyclones have the same barrel diameter and air volume ow rate.
181
2500
New design Stairmand design
2.5
1
2000
0.8
Axial velocity/ Inlet velocity

-0.05 0 0.05 0.1
0.6
1500
1.5
0.4
1000
0.2
500
0.5
0 -0.1 -0.05 0 0.05 0.1
0 -0.1
-0.2 -0.1 -0.05 0 0.05 0.1
Distance from center (m)
2500
2.5
1
0.8
2000

-0.05 0 0.05 0.1
0.6
1500
1.5
0.4
1000
0.2
500
0.5
-0.2 0 -0.1 -0.05 0 0.05 0.1 0 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
0.8
2000

-0.05 0 0.05 0.1
0.6
1500
1.5
0.4
1000
0.2
500
0.5
-0.2 0 -0.1 -0.05 0 0.05 0.1 0 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
Figure 8.16: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S1S3 . From top to bottom: section S1S3. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
182
2500
2.5
1
0.8
2000

-0.05 0 0.05 0.1
0.6
1500
1.5
0.4
1000
0.2
500
0.5
-0.2 0 -0.1 -0.05 0 0.05 0.1 0 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
0.8
2000

-0.05 0 0.05 0.1
0.6
1500
1.5
0.4
1000
0.2
0.5
500
-0.2 0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.4 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
2000
0.8

-0.05 0 0.05 0.1
0.6
1500
1.5
0.4
1000
0.2
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1
-0.2 -0.1 -0.05 0 0.05 0.1
Figure 8.17: The radial prole for the timeaveraged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S4S6 . From top to bottom: section S4S6. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
183
Section z`/D
a b b
S1 2.75
S2 2.5
S3 2.25
S4 2.0
S5 1.75
S6 1.5
Sections S1S5 are located in the conical section, section S6 at the cylindrical part. z` is measured from the top of the inlet section (cf. Fig. 8.1).
The velocity eld Based on the contour plots of the timeaveraged tangential velocity, Fig. 8.15, and the radial proles at sections S1S6 shown in Figs. 8.16 and 8.17, the following conclusions can be drawn. The tangential velocity prole at any section is composed of two regions, an inner and an outer one. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer part of the prole (free vortex). This prole is a so-called Rankine type vortex as mentioned before, including a quasi-forced vortex in the central region and a quasifree vortex in the outer region. The maximum tangential velocity may reach twice the average inlet velocity and occurs in the annular cylindrical part. The tangential velocity distribution for the two cyclones are approximately nearly identical in pattern and values (dimensionless), with the highest velocity occurring at 1/4 of the cyclone radius for both cyclones. This implies a nearly equal collection efciency for both cyclones, as the centrifugal force is the main driving force for particle collection in the cyclone separator. The axial velocity proles for the two cyclones are also very close, exhibiting the inverted W axial velocity prole [79]. Part of the ow in the central region moves downward in the two cyclones. This phenomena has been shown in the axial velocity pattern in other published articles [e.g., 79, 159]. The DPM results In order to calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a velocity equals the air inlet velocity. The particle density p is 860 kg/m3 and the maximum number of time steps for each injection was 200000 steps. The DPM analysis results and the pressure drops for the two cyclones are depicted in Table 8.16. Although, the difference between the two cyclone cut-off diameters is small, the saving in the pressure drop is considerable (nearly 25% the value of Stairmand cyclone). Based on the ow pattern analysis and the DPM results, one can conclude that the cyclone collection efciency for the two cyclones is very close, with 184

Table 8.16: The pressure drop and the cut-off diameter for the two cyclones
Stairmand design New design p (N/m2 ) 1190 864 x50 (m) 1.0 0.8
the advantage of low pressure drop in the new design. The authors want to emphasis that only small changes in the geometrical dimensions of the two designs lead to this improvement in the performance.
8.3.5 Conclusions
In order to predict accurately the complexly non linear relationships between pressure drop and geometrical dimensions, a radial basis neural network (RBFNN) is developed and employed to model the pressure drop for cyclone separators. The neural network has been trained and tested by the experimental data available in literatures. The following conclusions can be drawn from analysis of the obtained results: The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone pressure drop. Four mathematical models (Muschelknautz method MM, Stairmand, Ramachandran and Shepherd and Lapple) have been tested against the experimental values. The residual error of MM model is the lowest. Also, one can multiply the calculated value using MM by 1.5 to get the experimental value (as a rough approximation) The analysis indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b and the total height Ht . The response surface methodology has been used to t a second order polynomial to the RBFNN. The second-order polynomial has been used to get a new optimized cyclone for minimum pressure drop using Nelder-Mead technique. A comparison between the new design and the standard Stairmand design has been performed using CFD simulation with the Reynolds stress model to get a clear vision of the ow eld pattern in the new design. CFD results show that, the new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, and superior for low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly 75% of the pressure drop obtained by the old Stairmand design at the same volume ow rate. 185
8.4 Single and multi-objective optimization using RBFNN and GA

Both the pressure drop and the cut-off diameter in a cyclone separator can be decreased or increased by varying the cyclone dimensions. For an accurate optimal design of a cyclone, it is quite necessary to use a reliable model for its performance parameters. Optimization of gas cyclone is, indeed, a multi-objective optimization problem rather than a single objective optimization problem that has been considered so far in the literature [53, 146]. Both the pressure drop and the collection efciency in gas cyclones are important objective functions to be optimized simultaneously in such a real-world complex multi-objective optimization problem [54]. These objective functions are either obtained from experiments, empirical models or computed using very timely and high-cost computational uid dynamic (CFD) approaches. Modeling and optimization of the parameters are investigated in the present study, by using radial basis function articial neural networks and multi-objective genetic algorithm optimization technique in order to maximize the collection efciency (minimize the cutoff diameter) and minimize the pressure drop. The optimization studies conducted in the previous sections used the NelderMead technique [119] which suffer from two drawbacks, (1) the nal solution may be affected by the start point, (2) the obtained optimum may be a local minimum. To avoid this situation, the application of evolutionary method like the genetic algorithm must.
Study objectives There are four objectives of this study. (1) Investigation of the effect of the seven geometrical parameters on the cyclone separator performance (the pressure drop and cut-off diameter) based on the experimental data for the pressure drop and the most robust mathematical models for the cut-off diameter. (2) Study the possible interaction between the seven geometrical parameters affecting the cyclone performance using response surface methodology. (3) Multi-objective optimization to obtain new geometrical ratios for optimum performance (minimum pressure drop and minimum cut-off diameter). (4) Obtaining the optimum design (geometrical ratios) of the cyclone separator for minimum pressure drop using the genetic algorithm optimization technique, followed by a comparison of the numerical simulations of the optimal design and the Stairmand design using the Reynolds stress turbulence model. 186
8.4. Multi-objective optimization using GA
8.4.1 Articial neural network (ANN) approach

In this study, the radial basis function neural network (RBFNN) has been used to model the effect of cyclone dimensions on the Euler number (pressure drop). The details of this step have been explained in Sec. 8.3.1, page 160. The cut-off diameter (Stokes number) The source of the training data for the cut-off diameter has been obtained from the application of Iozia and Leith model [85]. This model has been approved as an acceptable approach for calculating the cyclone cut-off diameter [84, 85]. The cut-off diameter x50 is the particle diameter which produces 50% collection efciency. A question may appear here, why the authors employed the cut-off diameter instead of the collection efciency, like in the study of Ravi et al. [141]. Firstly, for low mass loading cyclone separator, the cut-off diameter can replace the collection efciency, since one can t the grade efciency curve using the cut-off diameter via some correlations, cf. Hoffmann and Stein [77, page 91] for more details. Moreover, many models can predict well the cut-off diameter but exhibit different grade efciency curves [77, page 97]. Secondly, the cut-off diameter (instead of the collection efciency) has been used as an objective function in many recent publications [e.g., 50, 148]. Moreover, the selection of the cut-off diameter or the collection efciency for low mass loading cyclones can be considered as a researcher choice. The Iozia and Leith model [84] is similar to the model of Barth [9] both based on the equilibrium-orbit theory (Force balance). The cut-off diameter can be calculated as [84]: x50 = (9 C ) / HCS p V2 max
1/2
(8.9)
where is the gas viscosity, Q is the gas volume ow rate, p is the particle density. HCS is the core height (height of the control surface, Fig. 8.18), Vmax is the maximum tangential velocity, that occur at the edge of the control surface CS . In this model, the value of core diameter dc , and the tangential velocity at the core edge; Vmax are calculated from regression of experimental data using Eq. A.34. Vmax = 6.1Vin a b/D2
0.61
(Dx /D)
0.25
0.74
(Ht /D)
1.53
0.33
(8.10) (8.11)
dc = 0.52D a b/D2 187
(Dx /D)
Figure 8.18: The control surface (core edge) used in the Iozia and Leith model [85].
where Vin is the area-average inlet velocity. The following expression obtained from trigonometry relations and can be used to calculate HCS . (R
Dx 2 )(Ht Bc
2
HCS
= =
R( (Ht S )
h)
+ (h S )
if Bc > Dx if Bc Dx (8.12)
It is clear from this model that the most important geometrical parameters that affect the cyclone cut-off diameter are the vortex nder diameter,the ratio of inlet area to exit area, cyclone height. Based on this model, the cut-off diameter is a function of the inlet gas velocity (i.e., a function of both gas volume ow rate, Barrel diameter, inlet section height and width), gas viscosity and particle density. For this particular study presented in Fig.8.20. The following values have been used: Barrel diameter D =0.1 m, air ow rate =0.8333 l/s, air viscosity 1.0E-5 Pa s and particle density 860 kg/m3 . This means that the obtained results will be valid for this particular case (used for just demonstration). However, the authors believe the variation of the cut-off diameter due to variations of cyclone geometrical dimensions is superior to the effect of these operating parameters, which is quite difcult to cover their range of operating conditions. The cut-off diameter x50 for cyclone separator is always given in units of m. Another way to represent x50 is using a dimensionless number; Stokes number. The Stokes number based on the cut-off diameter; Stk50 = p x2 50 Vin /(18D ) [37]. It is the ratio between the particle relaxation time; p x2 50 /(18) and the gas ow integral time scale; D/Vin . 188

Table 8.17: Validation of the used RBFNN to model the Euler number
x y Average 23.268 23.268 Minimum 2.3 1.745 Maximum 155.3 155.985 Range 153.0 154.24 Correlation Coefcient, R 0.999 Mean squared error, E 2 1.311E-4 Intercept 0.017 Slope 0.999 x is the input to the RBFNN and y is the predicted value. Both x and y represent the Euler number. The values of R, E 2 , intercept and slope are that for the testing stage.
Fitting the ANNs Tables 8.17 and 8.18 present more details about the validation of the used RBFNNs. Both the average, standard deviation, minimum, maximum and range of the input and the predicted values are given. It is clear from Tables 8.17 and 8.18 that the ANNs preserved the descriptive statistical parameters of the input data. The correlation coefcient between the input and the output and the mean squared error are given for each RBFNN. The intercept and the slope of the adjusted line between the input and the predicted value of the ANN are also given. The congured RBFNN predictions versus experimental data for the Euler number and the Iozia and Leith mathematical model for the Euler number and cut-off diameter are shown in Fig. 8.19. According to Fig. 8.19, it can be seen that the ANN models are able to attain the high training accuracy. The training mean square errors for the experimental values and the Iozia and Leith model have the values 1.311E-4 and 3.258E-4 respectively (Fig. 8.19). This indicates that, compared with traditional models of curve tting, the models based on articial intelligence algorithm have a superior capability of nonlinear tting. Especially, the RBFNN has its unique and optimal approximation characteristics in learning process [197]. Figure 8.19 illustrates the agreement between the ANNs input and output. The obtained relation is a typical linear relation with a coefcient of correlation close to 1 (R > 0.999). The agreement between the input and output of the ANN is also clear from the value of the mean squared error E 2 . That means, the trained neural networks predict very well both the Euler number and cut-off diameter values and can be used in cyclone design and performance estimation. Tables 8.17, 8.18 and Fig. 8.19 present different performance indicators as a validation of the proposed model for experimental values. 189
Table 8.18: Validation of the used RBFNN to model the cut-off diameter
Average Minimum Maximum Range Correlation Coefcient, R Mean squared error, E 2 Intercept Slope x 8.219 3.64 15.3 11.66 0.99915 3.258E-4 0.014 0.999 y 8.219 3.716 15.405 11.689
x is the input to the RBFNN and y is the predicted value. Both x and y represent the cut-off diameter. The values of R, E 2 , intercept and slope are that for the testing stage.
14
120
Predicted value (y)
Predicted value (y)
y=0.999 x + 0.0167 R=0.99964 E 2= 1.311E-4 80
12
y=0.998 x + 0.014 R=0.99915 E 2= 3.258E-4
10
40 6
4 40 80 120 4 6 8 10 12 14
Input value (x)
Input value (x)
(a) The Euler number
(b) The cut-off diameter
Figure 8.19: Linear regression of the RBFNNs for the Euler number and the cut-off diameter.
8.4.1.1 The effect of geometrical parameters on the cut-off diameter based on RBFANN The effect of the geometrical parameters on the Euler number based on the trained RBFNN has been presented in Sec. 8.3.2, page 166. The study acknowledge the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b and the total height Ht . Less effect is due to the cylinder height h (for h > 2.5) and the inlet height a (for a > 0.55). The effects of the geometrical parameters on the cut-off diameter are depicted in Fig. 8.20. To study the effect of each parameter, the tested RBFNN model has been used by varying one parameter at a time from its minimum to maximum values of the available 98 dataset, while the other parameters 190

Ht
6
10
h, S
0.5 14 12 10 1 1.5 2 2.5 3 3.5
X 50 [micron]
8 6 4 2 0
Dx a b S Ht h Bc
0.2
0.4
0.6
0.8
D x, a, b, Bc
Figure 8.20: The effect of geometrical parameters on the cut-off diameter based on the Iozia and Leith model [85]. Note: The plotted curves are obtained for a test case with the following settings, Barrel diameter=0.1 m, air ow rate=0.8333l/s, air viscosity=1.0E-5 Pa s, particle density=860 kg/m3 .
are kept constant at their mean values (cf. Table 8.7). Figure 8.20 indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b, the inlet height a and the total height Ht . Less effect is due to the cylinder height h and the cone tip diameter Bc . More analysis is given in Table 8.19.
The signicant geometrical parameters on the cut-off diameter (Stokes number) using the response surface methodology (RSM) Table 8.20 represents the parameters ranges selected for the seven geometrical parameters. The study was planned using BoxBehnken design, with 64 combinations. A signicant level of P < 0.05 (95% condence) was used in all tests. Analysis of variance (ANOVA) was followed by an F-test of the individual factors and interactions [53]. 191
Table 8.19: The variation of the cut-off diameter with cyclone dimensions using the RBFNN model (cf., Fig. 8.20)
Factor
Analysis The vortex nder diameter has the most signicant effect on the cut-off diameter x50 (the highest slope in Fig. 8.20). The slope is very high until Dx = 0.5 and any further increase in Dx produces a small change in x50 . In general, increasing Dx increases x50 (decreasing the collection efciency), this is one of the main reasons of the trade-off between the Euler number and the cut-off diameter objectives. This makes the optimization of cyclone geometry a multi-objective procedure. The variation of x50 with the inlet width is similar in trend and signicance to that for Dx but here the slope changes at b = 0.25. The effect of the vortex nder length and the inlet section height on the cut-off diameter is almost paralleled up to S = 1.5 and a = 0.6 afterwards they lose their signicance and become nearly constant. Increasing the barrel height slightly decreases the cut-off diameter with nearly linear relation. This trend has been reported by other researchers using CFD simulations, e.g., Elsayed and Lacor [51]. The effect of the cyclone total height is basically due to two effects the cone height and barrel height. The curve can be subdivided into four main regions. Sharp decrease in x50 up to Ht = 3.25, no valuable difference between 3.25 and 5.25, sharp increase between 5.25 and 8, and insignicant effect beyond 8. The effect of the cone-tip diameter on the cut-off diameter is quite small. First, increasing the cone-tip diameter slightly decreases the cut-off diameter up to Bc = 0.55 and any further increment increases the cut-off diameter. This trend has been reported by other researchers, e.g., Elsayed and Lacor [46, 55].
Dx
S and a
Ht
Bc
Due to the interaction between the geometrical parameters, especially between Ht with S and Dx (cf., Fig. 8.22 and Table 8.21), the obtained conclusions may not be applicable generally and the application of response surface methodology to analysis the effect of each particular parameter must.
192

Table 8.20: The values of the independent variables used in the design of experiment
Variables Inlet height, a=X1 Inlet width, b=X2 Cone tip diameter, Bc =X3 Vortex nder diameter, Dx =X4 Barrel height, h=X5 Total cyclone height, Ht =X6 Vortex nder length, S =X7 minimum 0.4 0.14 0.2 0.2 1.0 3.0 0.4 center 0.55 0.27 0.3 0.475 1.5 5.0 1.2 maximum 0.7 0.4 0.4 0.75 2.0 7.0 2.0
Analysis of variance (ANOVA) Analysis of variance (ANOVA) showed that the resultant quadratic polynomial models adequately represented the experimental data with the coefcient of multiple determination R2 being 0.984099 (cf., Table 8.21). This indicates that the quadratic polynomial model obtained was adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (see Table 8.21). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables [50, 53]. Based on the ANOVA results presented in Table 8.21, the variable with the largest effect on the Stokes number (cut-off diameter) was the linear term of vortex nder diameter, followed by the linear term of the cyclone total height, the vortex nder length and the inlet width (P < 0.05); the other three linear terms (inlet height, barrel height, and cone tip diameter) did not show a signicant effect (P > 0.05). The quadratic term of cyclone total height, vortex nder diameter and vortex nder length also had a signicant effect on the pressure drop; however, the effect of the other four quadratic terms was insignicant. Furthermore, the interaction between Dx with (Ht , S , b) and between S with (h, Ht ) also had a signicant effect on the Stokes number, while the effect of the remaining terms was insignicant. Analysis of response surfaces For visualization of the calculated factor, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve is proportional to the size of the effect and the direction of the curve species a positive or negative inuence of the effect [50, 61], Fig.8.21(a). Based on the main effect plot, the most signicant factors on the Stokes number are: (1) the vortex nder diameter Dx , with a secondorder curve of direct relation. (2) the cyclone total height Ht inversely related to the Stokes number. (3) the vortex nder length S with direct relationship. (4) the inlet 193
dimensions width b and height a inversely related to the Stokes number. Whereas the other factors have an insignicant effect. The main effect plot supports the analysis given in Table 8.19, except for Ht where the strong interaction between the cyclone total height and the vortex nder length affected the trend given in Fig. 8.20. Pareto charts were used to summarize graphically and display the relative importance of each parameter with respect to the Stokes number [50]. The Pareto chart shows all the linear and second-order effects of the parameters within the model and estimates the signicance of each with respect to maximizing the Stokes number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto charts judges whether each effect is statistically signicant within the generated response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 8.21(b)) and ANOVA table (Table 8.21) there are ve signicant parameters at a 95% condence level: the vortex nder diameter Dx , the total cyclone height Ht , the vortex nder length S and the inlet dimensions a and b. Therefore, the Pareto chart is a perfect supplement to the main effect plot. To visualize the effect of the independent variables on the dependent ones, surface response of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range while holding the other factors at their central values (cf., Table 8.20) [189] as shown in Fig. 8.22. Thus, Fig. 8.22(a) was generated by varying the total height Ht and the vortex nder length S while keeping the other ve factors constant. The response surface plots presented in Fig. 8.22 illustrate the strong interactions between Ht with (S and Dx ) and Dx with (S and b).
8.4.2
Single objective optimization using GA
The genetic algorithm optimization technique has been applied to obtain the geometrical ratios for minimum pressure drop (Euler number). The objective function is the Euler number (using the trained radial basis function neural network presented in Sec. 8.3.1.1, page 163). The design variables are the seven geometrical dimensions of the cyclone separator. 194
(b) Pareto chart. A=a, B=b, C=Bc , D=Dx , E=h, F=Ht , G=S , FG=Ht *S , etc.
Figure 8.21: Analysis of design of experiment for the Stokes number
195
(a) Ht versus S
(b) Dx versus Ht
(c) Dx versus S
(d) b versus Dx
Figure 8.22: The response surface plots for the Stokes number. Note: the stokes number values are multiplied by 1000.
196
GA settings Table 8.22 presents the settings used to obtain the optimum design for minimum pressure using global optimization Matlab toolbox (Matlab 2010a commercial package). The evolution of the cost function for the best individual is given in Fig. 8.23. After 70 iterations (14400 function counts), the Matlab code stops generating a new population since the average change in the tness value becomes less than 1E-6. The total calculation time for this optimization problem was 102 seconds. Table 8.23 gives the optimum values for cyclone geometrical parameters for minimum pressure drop estimated by the articial neural network using the genetic algorithm optimization technique. It is clear from Table 8.23 that the new optimized design is very close to the Stairmand design in many geometrical parameters, whereas the new ratios will result in minimum pressure drop. To understand the effect of this small change in the geometrical ratios on the ow eld pattern and performance, a CFD study for the two designs is needed [53].
8
7.5
Fitness value
6.5
5.5
4.5
10
20
30
40
50
60
70
Generation
Figure 8.23: Evolution of the cost function for the best individual
8.4.2.1 Comparison between the two cyclone designs using CFD Numerical settings The air volume ow rate Qin =0.08 m3 /s for the two cyclones (inlet velocity for Stairmand design is 19 m/s and 16 m/s for the new design), air density 197
1.0 kg/m3 and dynamic viscosity of 2.11E-5 P a s. the turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. Velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall boundary condition at all other boundaries [53]. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001 s. The grid renement study using different levels of grid shows that a total number of 134759 hexahedral cells for the Stairmand cyclone and 378963 hexahedral cells for the new design are sufcient to obtain a grid independent solution, and further mesh renement yields insignicant changes in the numerical solution. The hexahedral meshes have been obtained using the GAMBIT commercial software. These simulations were performed on an 8 nodes CPU Opteron 64 Linux cluster using Fluent 6.3.26 commercial software. The geometrical values for the two cyclones are given in Table 8.24.

The pressure eld Figure 8.24 shows the contour plot at Y=0. In the two cyclones, the timeaveraged static pressure decreases radially from the wall to center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, while the gradient in the axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the lowpressure zone at the cyclone center (twisted cylinder). However, the two cyclones have almost the same ow pattern, but the highest pressure of the Stairmand design is nearly 1.5 times that of the new design, implying that the new design has a lower pressure drop. The pressure distributions presented in Figs. 8.25 and 8.26 of the two cyclones at sections S1S6 depict the two parts pressure prole (for Rankine vortex). Again, the highest static pressure for Stairmand design is more than 1.5 times that of the new design at all sections while the central value is almost the same for the two cyclones irrespective of the section location. 198
Figure 8.24: The contour plots for the time averaged ow variables at sections Y=0. From top to bottom: Stairmand design and the new design respectively. From left to right: the static pressure (N/m2 ), the tangential and axial velocity (m/s). Note: both cyclones have the same barrel diameter and air volume ow rate.
199
2500
2.5
1
2000

-0.05 0 0.05 0.1
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
2000

-0.05 0 0.05 0.1
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
2000

-0.05 0 0.05 0.1
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
Figure 8.25: The radial prole for the time averaged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S1S3 (cf., Table 8.25). From top to bottom: section S1S3. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
This indicates that, the new design has a lower pressure drop with respect to the Stairmand design. The velocity eld Based on the contour plots of the timeaveraged tangential velocity presented in Fig. 8.24, and the radial proles at sections S1S6 shown in Figs. 8.25 and 8.26, the following conclusions can be drawn. The tangential velocity prole at any section is composed of two regions, an inner and an outer one. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer 200

2500
2.5
1
2000

-0.05 0 0.05 0.1
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
2000

-0.05 0 0.05 0.1
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
2500
2.5
1
2000

-0.05 0 0.05 0.1
0.5
1500
1.5
1000
0.5
500
0 0 -0.1 -0.05 0 0.05 0.1 -0.1 -0.5 -0.1 -0.05 0 0.05 0.1
Figure 8.26: The radial prole for the timeaveraged tangential and axial velocity at different sections on the X-Z plane (Y=0) at sections S4S6 (cf., Table 8.25). From top to bottom: section S4S6. From left to right: time-averaged static pressure, tangential and axial velocity respectively.
part of the prole (free vortex). This prole is a so-called Rankine type vortex. The maximum tangential velocity may reach twice the average inlet velocity and occurs in the annular cylindrical part. The tangential velocity distributions for the two cyclones are approximately nearly identical in pattern and values (dimensionless), with the highest velocity occurring at 1/4 of the cyclone radius for both cyclones. This implies a nearly equal collection efciency for both cyclones, as the centrifugal force is the main driving force for particle collection in the cyclone separator. The axial velocity proles for the two cyclones are also very close, exhibiting the inverted W axial velocity prole.
201
1 Stairmand design New design
0.75
Collection efficiency [-]
0.5
0.25
10
-1
10
10
Figure 8.27: The grade efciency curves for the two designs
The DPM results In order to calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a velocity equals the air inlet velocity and particle density p = 860kg/m3. The grade efciency curves for the two designs are plotted in Fig. 8.27. The DPM analysis results and the pressure drops for the two cyclones are depicted in Table 8.26. Although, the difference between the two cyclone cut-off diameters is small, the saving in pressure drop is considerable (nearly 32.5% the value of Stairmand cyclone). Based on the ow pattern analysis and the DPM results, one can conclude that the cyclone collection efciency for the two cyclones is very close, with the advantage of low pressure drop in the new design. The authors want to emphasis that only small changes in the geometrical dimensions of the two designs lead to this improvement in the performance.
8.4.3
Optimal cyclone design for best performance
NSGA-II settings Table 8.27 presents the genetic operators and parameters for multi-objective optimization. The Euler number values have been obtained from the articial neural network trained by experimental values. The Stokes number values are obtained from Iozia and Leith model [85]. In order to investigate the effect of different geometrical and operational parameters on the Pareto front, sixteen test cases with different barrel diameter, gas ow rate and particle density have been tested, cf. Table 8.28. The sixteen test cases covers: 1) Two barrel diameters, 31 mm and 205 mm. 2) Four levels of air ow rates, 50, 60, 70 and 80 l/min. 3) Five values of particle density, 860, 202
1000, 1500, 1750 and 2000 kg/m3 . Pareto front The Pareto front (non dominated points) for the sixteen test cases are presented in Fig. 8.28(a). Figure 8.28(a) clearly demonstrate tradeoffs in objective functions Euler number and Stokes number from which an appropriate design can be compromisingly chosen by the designer [148]. All the optimum design points in the Pareto front are non-dominated and could be chosen by a designer as optimum cyclone separator [148]. The corresponding geometrical ratios of the Pareto front shown in Fig. 8.28(a) are given in Table 8.29 for test case 1 and Table 8.30 for test case 9. Three points A, B and C are indicated in Figs. 8.28(c) and 8.28(c) and Tables 8.29 and 8.30. Point A indicates the point of minimum Euler number and maximum Stokes number. Point B indicates the point of maximum Euler number and minimum Stokes number. Point C indicates an optimal point for the multiobjective optimization problem. In order to obtain the Euler number- Stokes number relationship, Fig. 8.28(b) has been drawn. It indicates a general relationship (trend) between the two dimensionless numbers irrespective to the barrel diameter, gas ow rate, particle density. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number, Eq. 8.13. The obtained correlation can t the data with a coefcient of correlation R2 = 0.98643 as shown in Fig. 8.28(b). Stk50 = 100.3016(log10 (Eu)) Bubble plots for Pareto front For visual inspection of the effect of the seven geometrical parameters on the two conicting performance parameters, the bubble plots on Pareto front points have been drawn for each geometrical parameter. However, only gures for test case 1 (Fig. 8.29) and 9 (Fig. 8.30) are presented, but all other cases depict the same results (trend). Figure 8.29 indicates that: a) Decreasing the vortex nder diameter Dx decreases the Stokes number and increases the Euler number, Fig. 8.29(b). b) Generally speaking, increasing the inlet height a increases the Euler number and decreases the Stokes number. c) A similar trend is exhibited 203
2
0.9479log10 (Eu)2.5154
(8.13)
by the inlet width b but due to interaction with other geometrical and operational variables, one could see a range of bubble sizes in the region of best performance (lower values for both the Euler and Stokes numbers). d) The higher values of total cyclone height Ht will produce less Stokes number, intermediate values could produce less Euler number, smallerintermediate values could produce the optimum performance due to interaction with other variables. e) Short barrels will produce better collection efciency (low Stokes number) and higher Euler numbers. Intermediate values results in low Euler number values. Long barrels can produce the best performance. f) Short vortex nder may produce higher values of Euler numbers or higher values of Stokes number due to strong interaction with other variables. Long vortex nder can produce the optimum performance. g) Generally speaking, the variation of the cone-tip diameter Bc has no effect on the performance parameter. The above comments is restricted to the range of each geometrical variables located on the Pareto front and not for the whole range of values (cf., Fig. 8.29 for the range of each geometrical parameters).
8.4.4
Conclusions
To predict the complex non-linear relationships between the performance parameters and the geometrical dimensions, two radial basis neural networks (RBFNNs) are developed and employed to model the Euler number and Stokes number for cyclone separators. The neural networks have been trained and tested by the experimental data available in literature for Euler number (pressure drop) and Iozia and Leith model [85] for the Stokes number (cut-off diameter). The effects of the seven geometrical parameters on the Stokes number have been investigated using the trained ANN. To declare any interaction between the geometrical parameters affecting the Stokes number, the response surface methodology has been applied. The trained ANN has been used as an objective function to obtain the cyclone geometrical ratios for minimum Euler number using the genetic algorithms optimization technique. A CFD comparison between the new optimal design and the Stairmand design using the Reynolds stress turbulence model has been performed. A multi-objective optimization technique using NSGA-II technique has been applied to determine the Pareto front for the best performance cyclone separator. The following conclusions can be drawn from analysis of the obtained results: 204

The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone performance. The analysis indicates the signicant effect of the vortex nder diameter Dx and the vortex nder length S , the inlet width b, the inlet height a and the total height Ht on the cyclone performance. The response surface methodology has been used to t a second-order polynomial to the RBFNN for the cut-off diameter. The analysis of variance of the cut-off diameter indicates a strong interaction between Dx with (Ht , S , b) and between S with (h, Ht ). The trained RBFNN has been used to get a new optimized cyclone for minimum pressure drop (Euler number) using the genetic algorithm optimization technique. A comparison between the new design and the standard Stairmand design has been performed using CFD simulation with the Reynolds stress turbulence model to get a clear vision of the ow eld pattern and performance in the new design. CFD results shows that, the new cyclone design are very close to the Stairmand high efciency design in the geometrical parameter ratio, and superior in low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly 68% of the pressure drop obtained by the old Stairmand design at the same volume ow rate. This conrms that the obtained design using the genetic algorithm is better than that obtained using Nelder-Mead technique which results in 75% of the Stairmand pressure drop [53]. The two trained RBFNNs have been used in a multi-objective optimization process using NSGA-II technique. Sixteen test cases with different barrel diameter, gas ow rate and particle density have been tested and plotted. The Pareto fronts for the 16 test cases are very close. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number to obtain a general 2 formula, Stk50 = 100.3016(log10 (Eu)) 0.9479 log10 (Eu)2.5154 with a coefcient of correlation R2 = 0.98643. This formula can be used to obtain the Stokes number if the Euler number is known.
205
Table 8.21: Analysis of variance and the regression coefcients of the tted quadratic equation for the Stokes numbera
Source 0 Linear 1 2 3 4 5 6 7 Quadratic 11 22 33 44 55 66 77 Interaction 12 13 14 15 16 17 23 24 25 26 27 34 35 36 37 45 46 47 56 57 67 R2 Regression coefcient -0.0470554 0.942933 -1.3178 2.10188 8.3493 -0.843633 -0.527695 1.46453 -1.17696 1.44904 -3.50314 2.33086 0.281211 0 0.243505 0.726218 0 -2.09219 0 0.121646 -0.187708 0 -4.84457 0 0.262495 -0.466053 0 0 0 0 0 -0.869946 1.44156 0 0 -0.370128 0.984099 F-Ratio P-Value
6.65 24.12 0 1145.05 0 243.98 137.75 0.21 0.18 0.37 9.38 1.49 39.06 7.33 0.03 0 1.25 0 0.22 0.09 0 5.03 0 0.78 0.39 0 0 0 0 0 38.41 16.87 0 39.06 58.83
0.0154 0.0000 1 0.0000 1 0.0000 0.0000 0.6489 0.6737 0.5476 0.0048 0.232 0.0000 0.0114 0.8558 1 0.2732 1 0.6401 0.7726 1 0.0330 1 0.3842 0.5352 1 1 1 1 1 0.0000 0.0003 1 0.0000 0.0000
a Bold numbers indicate signicant factors as identied by the analysis of variance (ANOVA) at the 95% condence level.
Table 8.22: Genetic operators and parameters for single objective optimization
Population type: Initial range: Fitness scaling: Selection operation: Elite count: Crossover fraction: Crossover operation: Mutation operation: Maximum number of generations (iterations): Population size: Double vector [0.2 0.1 0.1 0.3 2.0 0.65 0.05 ; 0.8 0.8 0.6 0.8 8.0 2.5 0.75] Rank Tournament (tournament size equals 4) 2 0.8 Intermediate crossover with the default value of 1.0 The constraint dependent default 1400 200
206
Table 8.23: The optimized cyclone separator design for minimum pressure drop
Factor Dx a b S Ht h Bc Low 0.2 0.5 0.14 0.4 3.0 1.0 0.2 High 0.75 0.75 0.4 2.0 7.0 2.0 0.4 Stairmand design 0.5 0.5 0.2 0.5 4.0 1.5 0.375 Optimum design 0.549 0.595 0.201 0.595 4.549 1.411 0.275
Table 8.24: The geometrical parameters for the two designs (D=0.205 m)
Cyclone Stairmand design New design a/D 0.5 0.595 b/D 0.2 0.201 Dx /D 0.5 0.549 Ht /D 4 4.549 h/D 1.5 1.411 S/D 0.5 0.595 Bc /D 0.375 0.275
The outlet section is above the cyclone surface by Le = 0.618D. The inlet section located at a distance Li = D from the cyclone center.
Section z`/D b
a b
S1 2.75
S2 2.5
S3 2.25
S4 2.0
S5 1.75
S6 1.5
Sections S1S5 are located in the conical section, section S6 at the cylindrical part. z` is measured from the top of the inlet section.
Table 8.26: The performance parameters for the two cyclones

Design New design Method CFD ANN Ramachandran model [139] CFD ANN Ramachandran model [139] p (N/m2 ) 803 584.4 877.98 1190 1015.8 699.66 Euler number 6.338 4.613 5.523 6.567 5.606 4.846 x50 (m) 0.804 2.938 1.0 3.314 Stokes number x 103 0.114 1.815 0.209 1.931
Stairmand design
Table 8.27: Genetic operators and parameters for multi-objective optimization

Population type: Population size: Initial range: Selection operation: Crossover fraction: Crowding distance fraction Crossover operation: Number of generations (iterations): Double vector 105 (i.e., 15* number of variables which is the default in Matlab) [0.2 0.1 0.1 0.3 2.0 0.65 0.05 ; 0.8 0.8 0.6 0.8 8.0 2.5 0.75] tournament (tournament size equals 2) 0.8 0.35 Intermediate crossover with the default value of 1.0 1400 (i.e., 200* number of variables which is the default in Matlab)
207
Table 8.28: The diameters, air ow rates and the particle densities for the sixteen test cases Case 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 D [mm] 205 205 205 205 205 205 205 205 31 31 31 31 31 31 31 31 Q [l/min] 50 60 70 80 50 50 50 50 50 60 70 80 50 50 50 50 p [kg/m3 ] 860 860 860 860 1000 1500 1750 2000 860 860 860 860 1000 1500 1750 2000
208
Table 8.29: The seven geometrical parameters and the obtained Euler number and Stokes number for the nondominated points (Pareto-front) for test case 1 (cf. Table 8.28)
point 1 2B 3 4 5C 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36A 37 Minimum Maximum A Dx 0.306 0.306 0.618 0.326 0.360 0.670 0.585 0.306 0.306 0.305 0.303 0.476 0.312 0.308 0.592 0.320 0.598 0.309 0.430 0.306 0.306 0.326 0.308 0.680 0.307 0.307 0.514 0.310 0.559 0.399 0.516 0.413 0.306 0.307 0.646 0.692 0.558 0.303 0.692 a 0.659 0.688 0.229 0.331 0.295 0.230 0.226 0.685 0.666 0.516 0.286 0.240 0.622 0.686 0.229 0.570 0.229 0.441 0.307 0.667 0.392 0.331 0.686 0.229 0.653 0.368 0.235 0.596 0.246 0.308 0.244 0.256 0.507 0.644 0.229 0.228 0.216 0.216 0.688 b 0.385 0.398 0.213 0.240 0.253 0.213 0.217 0.299 0.361 0.271 0.318 0.217 0.277 0.379 0.213 0.273 0.213 0.266 0.231 0.389 0.277 0.209 0.348 0.213 0.296 0.338 0.217 0.275 0.221 0.229 0.241 0.280 0.352 0.333 0.219 0.213 0.230 0.209 0.398 S 0.410 0.404 0.411 0.451 0.443 0.410 0.419 0.401 0.407 0.439 0.449 0.434 0.419 0.427 0.413 0.423 0.412 0.450 0.438 0.417 0.441 0.451 0.427 0.409 0.421 0.439 0.429 0.424 0.453 0.438 0.422 0.444 0.437 0.423 0.409 0.408 0.439 0.401 0.453 Ht 6.957 6.993 6.774 6.668 6.664 6.777 6.763 6.985 6.965 6.840 6.611 6.743 6.943 6.974 6.831 6.898 6.770 6.769 6.735 6.950 6.708 6.668 6.974 6.787 6.972 6.708 6.709 6.918 6.785 6.707 6.728 6.667 6.811 6.960 6.797 6.819 6.802 6.611 6.993 h 1.779 1.779 1.885 1.885 1.910 1.859 1.901 1.789 1.782 1.868 1.930 1.901 1.909 1.781 1.895 1.909 1.893 1.881 1.898 1.779 1.882 1.900 1.807 1.857 1.816 1.895 1.901 1.899 1.898 1.891 1.904 1.927 1.848 1.807 1.863 1.855 1.922 1.779 1.930 Bc 0.387 0.317 0.495 0.444 0.459 0.496 0.492 0.425 0.398 0.428 0.462 0.471 0.429 0.459 0.494 0.434 0.494 0.435 0.465 0.367 0.443 0.471 0.459 0.495 0.420 0.430 0.479 0.429 0.455 0.460 0.486 0.472 0.460 0.401 0.497 0.498 0.486 0.317 0.498 Euler number 23.843 27.322 1.026 6.485 4.892 0.879 1.159 18.474 22.300 12.977 8.404 1.915 14.816 23.145 1.117 12.767 1.095 10.629 3.211 24.941 9.963 5.515 21.132 0.849 17.606 11.399 1.596 14.387 1.486 3.732 1.797 3.510 16.110 19.929 0.963 0.815 1.311 0.815 27.322 Stokes number x 103 0.552 0.540 2.785 0.798 0.937 3.478 2.089 0.576 0.558 0.633 0.699 1.524 0.619 0.554 2.491 0.658 2.565 0.674 1.229 0.550 0.676 0.823 0.564 3.640 0.588 0.659 1.718 0.622 1.915 1.102 1.674 1.162 0.602 0.575 3.113 3.805 1.945 0.540 3.805
indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.28(c)).
209
Table 8.30: The seven geometrical parameters and the obtained Euler number and Stokes number for the nondominated points (Pareto-front) for test case 9 (cf. Table 8.28)
point 1A 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26B 27 28 29C 30 31 32 33 34 35 36 37 Minimum Maximum A Dx 0.686 0.308 0.308 0.585 0.314 0.309 0.578 0.309 0.310 0.473 0.308 0.432 0.321 0.308 0.308 0.407 0.309 0.452 0.678 0.625 0.309 0.313 0.664 0.309 0.608 0.308 0.309 0.308 0.330 0.658 0.308 0.520 0.631 0.549 0.316 0.686 0.380 0.308 0.686 a 0.236 0.655 0.692 0.235 0.330 0.486 0.237 0.272 0.538 0.238 0.375 0.238 0.327 0.606 0.643 0.255 0.333 0.236 0.236 0.237 0.596 0.325 0.237 0.365 0.235 0.692 0.436 0.427 0.290 0.245 0.362 0.251 0.236 0.242 0.270 0.236 0.239 0.235 0.692 b 0.236 0.390 0.390 0.235 0.236 0.335 0.235 0.347 0.369 0.235 0.382 0.236 0.236 0.383 0.381 0.240 0.353 0.236 0.236 0.236 0.379 0.262 0.236 0.345 0.235 0.390 0.381 0.360 0.252 0.239 0.365 0.244 0.236 0.240 0.341 0.236 0.237 0.235 0.390 S 0.478 0.423 0.423 0.418 0.431 0.426 0.421 0.424 0.429 0.478 0.423 0.445 0.431 0.424 0.424 0.470 0.424 0.448 0.473 0.449 0.425 0.430 0.476 0.425 0.438 0.423 0.424 0.424 0.465 0.477 0.424 0.470 0.449 0.430 0.425 0.478 0.477 0.418 0.478 Ht 6.910 6.902 6.902 6.925 6.837 6.869 6.684 6.897 6.882 6.923 6.901 6.785 6.839 6.896 6.895 6.899 6.882 6.793 6.893 6.800 6.894 6.847 6.911 6.876 6.921 6.902 6.899 6.886 6.890 6.910 6.883 6.906 6.796 6.909 6.891 6.910 6.890 6.684 6.925 h 1.915 1.995 1.995 1.727 1.864 1.931 1.975 1.967 1.984 1.971 1.989 1.947 1.862 1.990 1.986 1.907 1.958 1.945 1.919 1.944 1.988 1.884 1.914 1.946 1.797 1.995 1.989 1.962 1.955 1.924 1.668 1.684 1.945 1.783 1.966 1.915 1.922 1.668 1.995 Bc 0.473 0.471 0.472 0.471 0.470 0.471 0.471 0.472 0.467 0.471 0.471 0.472 0.470 0.471 0.472 0.472 0.471 0.472 0.473 0.472 0.471 0.471 0.473 0.471 0.472 0.472 0.464 0.471 0.471 0.473 0.467 0.472 0.472 0.471 0.472 0.473 0.473 0.464 0.473 Euler number 1.008 21.688 22.898 1.355 6.568 13.947 1.356 8.025 16.799 2.105 12.168 2.494 6.223 19.736 20.775 3.119 10.021 2.255 1.029 1.187 19.039 7.197 1.076 10.740 1.261 22.898 14.120 13.112 5.639 1.146 11.924 1.996 1.164 1.613 7.482 1.008 3.327 1.008 22.898 Stokes number x 103 3.743 0.558 0.551 2.436 0.750 0.621 2.008 0.697 0.597 1.481 0.634 1.300 0.777 0.571 0.565 1.165 0.665 1.394 3.615 2.891 0.578 0.730 3.404 0.656 2.691 0.551 0.616 0.626 0.818 3.290 0.648 1.671 2.964 1.823 0.726 3.743 1.071 0.551 3.743
indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.28(d)).
210
4 3.5
3
Stokes number x 10
3 2.5 2 1.5 1 0.5 0 5 10 15 20 Euler number
Stokes number x 10
Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7 Case 8 Case 9 Case 10 Case 11 Case 12 Case 13 Case 14 Case 15 Case 16
7 6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5

3
Pareto points Stk50=10^{0.3016[LOG10(Eu)]^2-0.9479 LOG10(Eu)-2.5154}
25
0.5
10
15
20 25 30 35
Euler number
(a) Pareto fronts for 16 test cases, linear scale
(b) Pareto fronts for 16 test cases, log scale
4 3.5
3
4 3.5
Stokes number x 10
3 2.5 2 1.5 1 C 0.5 5 10 15 20 Euler number 25 B
Stokes number x 10
3 2.5 2 1.5 1 C 0.5 5 10 15 20 Euler number 25 B
(c) Pareto front for test case 1
(d) Pareto chart for test case 9
Figure 8.28: Pareto fronts for different test cases.
211
(a) Dx , range: 0.303 - 0.692
(b) a, range: 0.216 - 0.688
(c) b, range: 0.209 - 0.398
212
(d) Ht , range: 6.611 - 6.993 (e) h, range: 1.779 - 1.930 (f) S , range: 0.401 - 0.453
Figure 8.29: Bubble plots for different geometrical parameters for test case 1 (cf. Fig. 8.28(c) and Table 8.29).
(a) Dx , range: 0.308 - 0.686
(b) a, range: 0.235 - 0.692
(c) b, range: 0.235 - 0.390
213
(d) Ht , range: 6.684 - 6.925 (e) h, range: 1.668 - 1.995 (f) S , range: 0.418 - 0.478
Figure 8.30: Bubble plots for different geometrical parameters for test case 9 (cf. Fig. 8.28(d) and Table 8.30).
Table 8.31: The seven geometrical parameters and the obtained Euler number and Stokes number for the three points, A, B and C for the 16 test cases (cf. Table 8.28)
Point A1 B1 C1 A2 B2 C2 A3 B3 C3 A4 B2 C4 A5 B5 C5 A6 B6 C6 A7 B7 C7 A8 B8 C8 A9 B9 C9 A10 B10 C10 A11 B11 C11 A12 B12 C12 A13 B13 C13 A14 B14 C14 A15 B15 C15 A16 B16 C16 Dx 0.692 0.306 0.360 0.700 0.301 0.415 0.676 0.304 0.363 0.671 0.303 0.347 0.669 0.301 0.343 0.697 0.301 0.380 0.660 0.301 0.401 0.660 0.301 0.333 0.686 0.308 0.330 0.699 0.302 0.375 0.687 0.300 0.367 0.694 0.301 0.392 0.700 0.300 0.348 0.692 0.305 0.398 0.676 0.301 0.332 0.684 0.314 0.348 a 0.228 0.688 0.295 0.204 0.697 0.414 0.209 0.624 0.278 0.205 0.622 0.237 0.224 0.672 0.228 0.207 0.671 0.241 0.214 0.690 0.350 0.214 0.690 0.278 0.236 0.692 0.290 0.211 0.636 0.298 0.212 0.670 0.259 0.202 0.686 0.368 0.235 0.574 0.308 0.202 0.695 0.389 0.237 0.688 0.224 0.217 0.687 0.292 b 0.213 0.398 0.253 0.227 0.399 0.287 0.204 0.398 0.230 0.212 0.399 0.271 0.204 0.387 0.282 0.207 0.400 0.303 0.205 0.322 0.258 0.205 0.322 0.263 0.236 0.390 0.252 0.205 0.399 0.250 0.232 0.399 0.282 0.202 0.400 0.281 0.201 0.365 0.239 0.203 0.398 0.272 0.205 0.367 0.278 0.202 0.388 0.245 S 0.408 0.404 0.443 0.446 0.561 0.449 0.431 0.581 0.459 0.427 0.467 0.582 0.537 0.416 0.475 0.400 0.500 0.486 0.432 0.459 0.451 0.432 0.459 0.457 0.478 0.423 0.465 0.409 0.424 0.465 0.568 0.481 0.461 0.512 0.431 0.455 0.602 0.408 0.426 0.464 0.434 0.440 0.437 0.562 0.450 0.424 0.400 0.457 Ht 6.819 6.993 6.664 6.422 6.989 6.803 4.378 6.983 6.404 6.524 6.767 6.769 6.946 6.851 6.925 6.734 6.982 6.715 6.962 6.988 6.928 6.962 6.988 6.869 6.910 6.902 6.890 6.499 6.854 6.726 6.702 6.988 6.688 6.999 6.840 6.874 6.752 6.743 6.577 6.963 6.881 6.899 6.779 6.975 6.645 6.725 6.943 6.850 h 1.855 1.779 1.910 1.811 1.840 1.776 1.980 1.116 1.354 1.824 1.975 1.913 1.895 1.905 1.822 1.974 1.228 1.845 1.934 1.918 1.920 1.934 1.918 1.940 1.915 1.995 1.955 1.937 1.614 1.875 1.900 1.693 1.857 1.794 1.191 1.731 1.810 1.907 1.925 1.901 1.785 1.849 1.879 1.704 1.799 1.962 1.792 1.887 Bc 0.498 0.317 0.459 0.470 0.449 0.443 0.496 0.344 0.445 0.467 0.479 0.476 0.479 0.463 0.487 0.465 0.477 0.427 0.495 0.450 0.454 0.495 0.450 0.450 0.473 0.472 0.471 0.444 0.488 0.445 0.465 0.484 0.424 0.499 0.439 0.472 0.481 0.454 0.477 0.463 0.498 0.467 0.469 0.480 0.456 0.380 0.495 0.477 Euler number 0.815 27.322 4.892 0.820 28.222 6.052 0.868 32.296 4.791 0.819 22.621 4.873 0.900 23.579 4.647 0.706 29.367 4.557 0.810 20.739 4.735 0.810 20.739 5.582 1.008 22.898 5.639 0.741 23.886 4.619 0.955 25.971 4.801 0.736 29.948 5.822 0.901 19.174 5.047 0.722 24.604 5.614 0.890 25.765 4.912 0.823 21.743 4.969 Stokes number x 103 3.805 0.540 0.937 4.260 0.538 1.037 4.108 0.561 1.000 3.796 0.556 0.931 3.654 0.538 0.893 4.142 0.538 1.026 3.391 0.557 1.039 3.391 0.557 0.832 3.743 0.551 0.818 4.328 0.543 0.997 4.035 0.534 0.972 4.071 0.533 0.980 4.286 0.565 0.895 4.161 0.543 0.992 3.717 0.549 0.869 4.299 0.565 0.891
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem, cf. Fig. 8.28(c).
214
8.5. Multi-objective optimization using CFD data
8.5 Single and multi-objective optimization using CFD data

8.5.1 Design variables and approaches
The sensitivity analysis presented in Sec. 4.1 reported that the most signicant factors affecting the cyclone performance are, the vortex nder diameter Dx , the inlet width b, the inlet height a and the total cyclone height Ht . The effects of the barrel height h, the vortex nder length S and the cone-tip diameter Bc are insignicant. Therefore, only four signicant geometrical factors have been used in this study. The three other factors are xed based on the Stairmand high efciency design, i.e. Bc /D = 0.375, h/D = 1.5 and S/D = 0.5. The selection of the values for both h and S are also based on the conclusion of Zhu and Lee [200] (when both the pressure drop and the particle collection efciency are considered, a cyclone which has (h-S )/D of 1.0 would be an optimum design). Optimization of a gas cyclone is, indeed, a multi-objective optimization problem rather than a single objective optimization problem that has been considered so far in the literature [146]. Both the pressure drop and the cut-off diameter in gas cyclones are important objective functions to be optimized simultaneously in a multi-objective optimization problem. These objective functions values are either obtained from experiments, empirical models or computed using computational uid dynamic (CFD) approaches. CFD modeling and optimization of the parameters are investigated in the present study and multi-objective Nedler-Mead optimization algorithms are used in order to maximize the collection efciency (minimize the cut-off diameter) and minimize the Euler number (dimensionless pressure drop). The desirability function approach has been used to handle the two objective function. The application of the Nelder-Mead technique suffers from one big disadvantage. The optimal point may be local minimum because the technique depends on the starting point. This was the motivation to use also the genetic algorithm technique. Furthermore, the radial basis function neural networks can attain high accuracy as curve tting approach than the response surface [53]. Consequently, the application of the radial basis function neural networks to model the relationship between the performance parameters and the four geometrical parameters will result in more accurate results. The optimization studies given in the previous sections were applied for the seven geometrical parameters but the current study is focusing only on the most signicant four geometrical parameters. The above-mentioned studies used meta-models using mathematical models and experimental measurements but this study is based only on CFD 215
simulations performed on sampling cyclone with a diameter of 31E-3 m.
Study objectives There are four objectives of this study. (1) Investigation of the effect of the four geometrical parameters on the cyclone separator performance based on CFD simulations results. (2) Study the possible interaction between the four geometrical parameters using the response surface methodology. (3) Application of the multi-objective optimization technique to obtain new geometrical ratios for minimum pressure drop and minimum cut-off diameter, followed by a comparison of the numerical simulations for the new design and the Stairmand design using the Reynolds stress turbulence model. (4) Application of the genetic algorithm technique instead of the desirability function approach using the two trained radial basis function neural networks as tness functions.
Study outline This study is performed in six stages. The rst stage is the application of the response surface methodology (Box-Behnken design) to design an experiment to study the effect of four geometrical parameters (the inlet height and width, the vortex nder diameter, and the cyclone total height) on the cyclone performance using Statgraphics statistical software. Secondly, the obtained 27 test cases (designs) have been computationally simulated using the Reynolds stress turbulence model and discrete phase modeling with the Fluent solver. In the third stage, the Euler numbers and the cut-off diameters obtained are used to t a second order polynomial (response surface) for each response (the Euler number and the cut-off diameter). The obtained polynomials have been used to study the variation of the two responses with the four geometrical parameters. Furthermore, the obtained polynomials have been used to obtain new cyclone geometrical ratios using the multi-objective optimization between the two conicting objectives (the Euler number and the cut-off diameter) using the desirability function approach. The fourth stage is a computational investigation: a numerical comparison between the new design and the Stairmand design has been performed. Fifthly, replace the response surface methodology with the articial neural networks approach and study the effect of each geometrical parameter on the cyclone separator performance. The last stage is an optimization study using the genetic algorithm instead of desirability function. 216
Design of experiment
In order to model a complex multivariate process where the responses are inuenced by several variables, the response surface statistical technique seems the best approach [50]. The steps are as follows: (1) construct the design of experiment by identifying the four tested geometrical parameters (minimum and maximum values) and also to decide upon the dependent variables (the Euler number and the cut-off diameter). Statgraphics commercial statistical software gives 27 runs to be performed (cf. Table 8.32) using the Box-Behnken design of experiment method [118]. (2) Once the runs have been conducted (using CFD simulations) and recorded data inserted in the table; Statgraphics software ts a second order polynomial to this data (one response surface per dependent variable) [30]. The secondorder polynomial (response surface) has the form [50]:
4 4
Yk = 0 +
i=1
i X i +
i=1
ii Xi2 +
i<j
ij Xi Xj
(8.14)
where 0 , i , ii , and ij are the regression coefcients for intercept, linear, quadratic and interaction terms, respectively. Xi and Xj are the independent variables, and Yk is k the response variable (k = 1 for the Euler number and k = 2 for the cut-off diameter). (3) The third step, is the analysis of the response surface plot, main effect plots, Pareto chart and interaction plots. Analysis of response surface plots Analysis of variance (ANOVA) showed that the resultant quadratic polynomial models adequately represent the used data with the coefcient of determination R2 , being 0.94284 and 0.973468 for the Euler number and cut-off diameter respectively. This indicates that the quadratic polynomial models obtained were adequate to describe the inuence of the independent variables studied [189]. Analysis of variance (ANOVA) was used to evaluate the signicance of the coefcients of the quadratic polynomial models (cf. Tables 8.33 and 8.34). For any of the terms in the models, a large F-value (small P-value) would indicate a more signicant effect on the respective response variables. Based on the ANOVA results presented in Table 8.33, the variable with the largest effect on the pressure drop (Euler number) was the linear term of vortex nder diameter (3 ), the linear terms of inlet height and width (P < 0.05) (1 and 2 respectively), whereas the linear term of cyclone total height did not show a signicant effect (P > 0.05). The quadratic 217
Table 8.32: The Box-Behnkens design matrix and the responses of the Euler number and cut-off diametera
Run No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 a X1 a/D 0.250 0.500 0.375 0.375 0.375 0.250 0.375 0.375 0.375 0.500 0.250 0.375 0.375 0.375 0.500 0.375 0.250 0.375 0.500 0.250 0.375 0.375 0.500 0.250 0.500 0.375 0.375 X2 b/D 0.2625 0.2625 0.3750 0.2625 0.2625 0.2625 0.1500 0.1500 0.2625 0.3750 0.2625 0.1500 0.1500 0.3750 0.2625 0.3750 0.2625 0.2625 0.2625 0.3750 0.2625 0.3750 0.1500 0.1500 0.2625 0.2625 0.2625 X3 Dx /D 0.50 0.75 0.25 0.50 0.75 0.50 0.50 0.50 0.25 0.50 0.75 0.50 0.75 0.50 0.50 0.50 0.25 0.25 0.25 0.50 0.50 0.75 0.50 0.50 0.50 0.50 0.75 X4 Ht /D 3 4 4 4 3 5 4 5 5 4 4 3 4 3 5 5 4 3 4 4 4 4 4 4 3 4 5 Euler number (-) 3.500 2.827 27.257 3.475 2.333 2.952 2.726 2.530 17.712 9.086 1.413 3.000 1.211 7.500 5.904 6.326 12.720 21.000 25.440 4.543 4.770 3.029 3.634 1.817 7.000 4.770 1.968 Cut-off diameterb (m) 1.546 2.541 1.158 1.683 2.444 1.364 1.353 1.284 0.956 2.163 1.939 1.455 1.826 2.081 1.787 1.836 0.860 1.084 1.127 1.651 1.683 2.610 1.513 1.155 2.025 1.683 2.156
The values of h/D = 1.5, S/D = 0.5 and Bc /D = 0.375 are identical to that of Stairmand high efciency design, so the variation in the total cyclone height is due to the variations of the cone height.
b The value of cut-off diameter depends on the gas velocity and density, particle density. In this study, D = 31mm, gas volume ow rate Qin = 50 l/min, gas density = 1.0, gas viscosity = 2.11E 5 Pa s and particle density p = 860 kg/m3 .
term of vortex nder diameter also had a signicant effect on the pressure drop; however, the effect of the other three quadratic terms was insignicant. Furthermore, the interaction between the inlet dimensions and vortex nder diameter also had a signicant effect on the pressure drop (13 and 23 ), whereas the effect of the remaining terms was insignicant. Table 8.34 conrms the signicant effect of all linear terms of the inlet dimensions, vortex nder diameter and total cyclone height on the cut-off diameter. Moreover, the interaction between the inlet width and the vortex nder diameter also had a signicant effect on the cut-off diameter, whereas the effect of the remaining terms was insignicant. For the visualization of the results of the analysis, main effects plot, Pareto chart and response surface plots were drawn. The slope of the main effect curve is proportional to the size of the effect, and the direction of the curve species a positive or negative inuence of the effect [50, 61] (Fig.8.31(a)). Based on the main effect plot and the Pareto chart shown in Fig. 8.31, the most signicant factors on the Euler number are: (1) the vortex nder diameter Dx , with a non-linear relation with a wide range where it is inversely proportional to Euler number and a narrow range with a direct 218

Table 8.33: Analysis of variance and regression coefcients for the Euler number
Variable 0 Linear 1 2 3 4 Quadratic 11 22 33 44 Interaction 12 13 14 23 24 34 R2
Regression coefcient 1.88508 41.3522 134.36 -49.1776 -4.14082 16.9991 -44.8511 87.9024 0.368311 48.4622 -90.448 -1.096 -185.69 -1.56444 2.923 0.94284
F-Ratio
P-Value
14.21 33.78 209.03 0.94 0.1 0.5 44.97 0.2 0.41 7.01 0.02 28.5 0.03 0.47
0.0012 0.0000 0.0000 0.3430 0.7509 0.4878 0.0000 0.6602 0.5306 0.0155 0.8992 0.0000 0.8708 0.5017
proportional relation. (2) the inlet width b and the inlet height a almost linearly related to the Euler number, (3) the cyclone total height Ht has an insignicant effect. Pareto charts were used to summarize graphically and display the relative importance of each parameter with respect to the Euler number [50]. The Pareto chart shows all the linear and secondorder effects of the parameters within the model and estimates the significance of each with respect to maximizing the Euler number response. A Pareto chart displays a frequency histogram with the length of each bar proportional to each estimated standardized effect [30]. The vertical line on the Pareto charts judges, whether each effect is statistically signicant within the generated response surface model; bars that extend beyond this line represent effects that are statistically signicant at a 95% condence level. Based on the Pareto chart (Fig. 8.31(b)) there are six signicant parameters at a 95% condence level: the vortex nder diameter Dx ; the 2 inlet width b; the inlet height a and the combinations a Dx , b Dx and Dx . Therefore, the Pareto chart is a perfect supplement to the main effects plot. To visualize the effect of the independent variables on the dependent ones,response surfaces of the quadratic polynomial models were generated by varying two of the independent variables within the experimental range whereas holding the other factors at their central values [189] as shown in Fig. 8.31. Thus, Fig. 8.31(e) was generated by varying the inlet height a and the inlet width b whereas keeping the other factors xed at their central values. The trend of the curve is linear, with a more signicant effect for the inlet width b, and a weak interaction between the inlet height a
219
Table 8.34: diameter
Variable 0 Linear 1 2 3 4 Quadratic 11 22 33 44 Interaction 12 13 14 23 24 34 R2
Analysis of variance and regression coefcients for the cut-off

Regression coefcient 0.365539 1.84206 -1.23737 0.494019 -0.00682157 -2.25164 -0.708116 -0.307782 0.00841817 2.73778 2.68 -0.112 7.07071 -0.164444 -0.16 0.973468 F-Ratio P-Value
67.21 89.76 524.79 15.14 0.82 0.06 0.25 0.05 0.59 2.77 0.08 18.61 0.14 0.63
0.0000 0.0000 0.0000 0.0009 0.3763 0.8152 0.6237 0.8308 0.4531 0.1116 0.7837 0.0003 0.7170 0.4360
and width b. The response surface plots of Figs. 8.31(c), 8.31(d) and 8.31(f) show that there are strong interactions between the vortex nder diameter Dx and the inlet dimensions a and b. From the analysis of the design of experiment for the cut-off diameter, the effect of variation of the vortex nder diameter Dx on the cut-off diameter is opposite to that on the Euler number (cf. Fig. 8.32(a)). The Pareto chart given in Fig. 8.32(b) indicates the signicance effect of the cyclone total height Ht on the cut-off diameter, but its effect is minor in comparison with the three other factors. The signicant interaction exists only between the vortex nder diameter and the inlet dimensions.
8.5.2
Multi-objectives optimization using the desirability function
The desirability function approach From the previous analysis, it is observed that the optimal values for the geometrical parameters that minimize the pressure drop are different from the values that minimize the cut-off diameter (cf. Figs. 8.31 and 8.32). As a result, a multi-objective optimization procedure is needed. The uti220
(b) Pareto chart
Figure 8.31: Analysis of design of experiment for the Euler number
lization of desirability function proposed by Harrington [71] is the most popular and strongly suggested method for multiple response optimization problems [20] to convert the problem into single-objective. The Statgraphics statistical package uses this desirability function approach for optimization of multiple response problems. The desirability function is rst dened for each response. The desirability function d(y ) expresses the desirability of a response value equal to y on a scale of 0 (if the response value is in an unacceptable range) to 1 (for the optimum value), for minimization of response variable. The procedure will then nd the settings of the experimental factors that maximize a combined desirability function, which is a function that expresses the desirability of a solution involving m, where m here equals 2 (one for the Euler number and the other for the
221
(c) Dx versus a
(d) Dx versus b
(e) a versus b
(f) Dx versus Ht
Figure 8.31: (continued) Analysis of design of experiment for the Euler number. Note: In Figs 8.31(c) - 8.31(f), for each plot of two independent variable, all other variables are hold at their central values.
222
(b) Pareto chart
Figure 8.32: Analysis of design of experiment for the cut-off diameter
cut-off diameter), responses through the function of the form [131],
Im 1 I2 D = dI 1 d2 . . . dm
1/(
m j =1
Ij )
(8.15)
where dj is the calculated desirability of the jth response and Ij is an impact coefcient that ranges between 1 and 5 [131]. It represents the importance relative of each response over the other responses [88]. Where the default value is 3. In this study, more importance is given to the Euler number (I1 = 5 and I2 = 3). When a response is to be minimized, the 223
(c) Dx versus a
(d) Dx versus b
(e) a versus b
(f) Dx versus Ht
Figure 8.32: (continued) Analysis of design of experiment for the cut-off diameter. Note: In Figs 8.32(c) - 8.32(f), for each plot of two independent variable, all other variables are hold at their central values.
224

Table 8.35: The optimized cyclone separator design for best performance using the desirability function
Factor a b Dx Ht Minimum 0.25 0.15 0.25 3 Center 0.375 0.2625 0.5 4 Maximum 0.5 0.375 0.75 5 Stairmand design 0.5 0.2 0.5 4.0 Optimum 0.256 0.151 0.415 4.56
desirability of a predicted response equal to y j is dened as, y j < lowj 1 dj =

lowj highj 0 y j highj
lowj y j highj y j > highj
(8.16)
where lowj and highj are the minimum and maximum values of jth response. In this study, the minimum and maximum values obtained in the data sheet have been used (cf. Table 8.32). The desirability plots are given in Fig. 8.33. For more information about statistical model used for multiple response optimization refer to Ref. [131]. Table 8.35 presents the optimum values of geometry parameters that minimize the values of the Euler number and the cut-off diameter, which gives optimum desirability D = 0.83 to minimize the Euler number and cut-off diameter. It is clear from Table 8.35 that the new optimized design differs from the Stairmand design in many geometrical parameters, whereas the new ratios will result in minimum pressure drop and minimum cut-off diameter. The ratio of inlet width to height b/a = 0.589 lays in the optimal cyclone lies in the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. To understand the effect of these changes in the geometrical ratios on the ow eld pattern and performance, a CFD study for the two designs is needed. 8.5.2.1 CFD comparison between the Stairmand and optimal design The Fluent solver has many turbulence models available for simulating turbulent ow. It is generally recognized that only the Reynolds stress model (RSM) and large eddy simulation (LES) can capture the main features of the highly complicated swirling ow in cyclone separators [11, 23, 50, 52, 6264, 74, 91, 159, 186, 198]. The Reynolds stress turbulence model has been used in this study to reveal the turbulent ow in the two cyclone separators. For the detailed governing equation for both the Reynolds averaged Navier-Stokes equation (RANS) and the discrete phase modeling 225
(a) Dx versus a
(b) Dx versus b
(c) a versus b
(d) Dx versus Ht
Figure 8.33: The desirability plots. Note: for each plot of two independent variable, all other variables are hold at their optimal values.
226
(DPM) the reader can refer to Elsayed and Lacor [52]. The geometrical values are given in Table 8.36 for the two cyclones (cf. Fig. 8.34). Numerical settings The air volume ow rate Qin =50 l/min for the two cyclones, air density is 1.0 kg/m3 and dynamic viscosity 2.11E-5 P a s. The turbulent intensity equals 5% and characteristic length equals 0.07 times the inlet width [75]. A velocity inlet boundary condition is applied at inlet, outow at gas outlet and wall boundary conditions at all other boundaries. The nite volume method has been used to discretize the partial differential equations of the model using the SIMPLEC (Semi-Implicit Method for Pressure-Linked Equations-Consistent) method for pressure velocity coupling and QUICK scheme to interpolate the variables on the surface of the control volume. The implicit coupled solution algorithm was selected. The unsteady Reynolds stress turbulence model (RSM) was used in this study with a time step of 0.0001s.
(b) The surface mesh for the new design
Figure 8.34: The cyclone geometry and the surface mesh for the new design.
Table 8.36: The values of geometrical parameters for the two designs (D=31E-3 m)
Cyclone Stairmand design New design a/D 0.5 0.256 b/D 0.2 0.151 Dx /D 0.5 0.415 Ht /D 4 4.56 h/D 1.5 1.5 S/D 0.5 0.5 Bc /D 0.375 0.375 Li /D 1.0 1.0 Le /D 0.5 0.5
227
Grid independency study The grid independence study has been performed for the tested cyclones. Three levels of grid for each cyclone have been tested, to be sure that the obtained results are grid independent. The hexahedral computational grids were generated using Gambit grid generator and the simulations were performed using Fluent 6.3.26 commercial nite volume solver on a 8 nodes CPU Opteron 64 Linux cluster. To evaluate accuratelyto estimate accurately the numerical uncertainties in the computational results, the concept of grid convergence index (GCI) was adopted using three grid levels per cyclone. Grid convergence index (GCI) Table 8.37 presents the grid convergency calculations using GCI method for three grid levels for each cyclone. The following conclusions have been obtained from the GCI analysis [55]: The results are in the asymptotic range, because the obtained values for are close to unity. The ratio R is less than unity this means monotonic convergence [2]. There is a reduction in the GCI value for the successive grid renements ne (GCIne 12 < GCI23 ) for the two variables (Eu and x50 ). This indicates that the dependency of the numerical results on the cell size has been reduced. Moreover, a grid independent solution has been achieved. Further renement of the grid will not give much change in the simulation results. For the two variables (Eu and x50 ), the extrapolated value is only slightly lower than the nest grid solution. Therefore, the solution has converged with the renement from the coarser grid to the ner grid [2]. Figure 8.35 presents a qualitative proof that the obtained results are in the asymptotic range. The value of i,i+1 represent the relative change in each value from coarse to medium and from medium to ne mesh. For example, 1,2 = 0.0256 for the Euler number in the new design means the percentage change in the Euler number when the mesh becomes 986748 cells instead of 717353 cells equals 2.56%. This means no need to use the ne mesh and the usage of the medium mesh of 717353 cells is sufcient. Another example, 1,2 = 0.0052 for the cut-off diameter in the Stairmand design means the percentage change in the cutoff diameter when the mesh becomes 848783 cells instead of 622253 228
cells equals 0.52%. This means no need to use the ne mesh and the usage of the medium mesh of 622253 cells is sufcient. In summary, the grid renement study shows that a total number of about 622253 hexahedral cells for Stairmand cyclone and 717353 hexahedral cells for the new design are sufcient to obtain a grid-independent solution, and further mesh renement yields only small, insignicant changes in the numerical solution.
Table 8.37: Grid convergency calculations using GCI method using three grid levels for each cyclone.
Cut-off diameter Euler number Cut-off diameter Euler number i 0c 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3 Ni 986748 717353 362679 986748 717353 362679 848783 622253 356181 848783 622253 356181 fi 2.5195 2.5380 2.6030 1.2553 4.6170 0.6621 0.6710 1.1121 0.6900 1.2553 0.9820 5.4310 5.4860 1.1090 5.6060 1.2044 6.8330 1.7187 1.7210 1.1090 1.7300 1.2044 1.9150 0.1850 0.1069 0.8173 0.0090 0.0052 0.1678 0.0489 1.0052 1.2270 0.2189 3.9010 0.1200 0.0219 1.2521 0.0999 1.0219 0.2920 0.4232 5.0609 0.0190 0.0283 1.6647 0.0669 1.0283 2.0140 0.7737 4.0099 ri,i+1 ei,i+1 i,i+1 GCIi,i+1 % Ra b
1.1121
0.0650
0.0256
0.9112 0.0331 1.0256
Stairmand design
New design
a R= / . 12 23 b = rp GCI 12 /GCI23 . 12 c The value at zero grid space (h 0). i=1, 2 and 3 denote the calculations at the ne, medium and coarse mesh respectively.

Flow eld pattern Figure 8.36 shows the contour plot at Y=0 and throughout the inlet section. In the two cyclones, the time-averaged dimensionless static pressure (Euler number) decreases radially from the wall to center. A negative pressure zone appears in the forced vortex region (central region) due to high swirling velocity. The pressure gradient is largest along the radial direction, whereas the gradient in the axial direction is very limited. The cyclonic ow is not symmetrical as is clear from the shape of the low-pressure zone at the cyclone center (twisted cylinder). The ow asymmetry is more pronounced in the new cyclone. However, the two cyclones have almost the 229
6 1.5 5
New design (Eu) New design (X50) Stairmand design (Eu) Stairmand design (X50)
1 3
1E-06
2E-06
(h --> 0)
-1
Figure 8.35: Qualitative representation of the grid independency study. The Euler number and the cut-off diameter for the two cyclones using the three grid levels. N 1 is the reciprocal of the number of cells, h 0 means the value at zero grid size (cf. Table 8.37). To obtain a smooth curve; the spline curve tting has been applied in Tecplot post-processing software.
same ow pattern, but the Euler number of the Stairmand design is nearly twice that of the new design. The dimensionless static pressure distribution presented in Fig. 8.37 for the two cyclones indicates that the highest dimensionless static pressure for the Stairmand design is more than twice that of the new design at all sections whereas the central value is almost the same for the two cyclones. This indicates that, the new design has a lower dimensionless pressure drop than the Stairmand design. However, these results are obtained at different inlet velocity for the two cyclones (to have the same air ow rate). The same Euler number values would be obtained if the two cyclones work at the same inlet velocity because the Euler number is not a function of ow velocity if the Reynolds number is higher than 2E4 [50]. The tangential velocity prole is composed of two regions. In the inner region, the ow rotates approximately like a solid body (forced vortex), where the tangential velocity increases with radius. After reaching its peak the velocity decreases with radius in the outer part of the prole (free vortex). The tangential velocity distributions for the two cyclones are nearly identical in pattern (Rankine prole). The inner part of the tangential velocity distribution of the two cyclones is very similar. The outer part for the new design is atter in comparison with the Stairmand cyclone. This implies that there is more space in the optimal cyclone where the particles are subjected to high centrifugal force. Whereas, the maximum dimensionless tangential velocity for Stairmand cyclone is higher than that for the 230
Cut-off diameter
Euler number
new design, the cyclone performance is not only affected by the maximum tangential velocity but also with the separation space (the new design is longer than the Stairmand design) The axial velocity proles for the two cyclones are different in values and shape. Stairmand cyclone exhibit the inverted W axial velocity prole. The new design exhibit the inverted W axial velocity prole away from the inlet section and the inverted V elsewhere. This is due to the very high swirl exist at the inlet section in case of the new design. Performance parameters To calculate the cut-off diameters of the two cyclones, 104 particles were injected from the inlet surface with a particles density p = 860kg/m3 and the maximum number of time steps for each injection was 2E9 steps. The DPM analysis results and the Euler number for the two cyclones are given in Table 8.38. Table 8.38 introduces also a comparison between the CFD results and four different mathematical models viz., the Ramachandran model [139], the Muschelknautz method of modeling (MM) [50, 116, 174], the Iozia model [85], the Ritema model [142](cf. Hoffmann and Stein[77] for more details about these mathematical models). However, no-good matching between the two approaches (CFD and mathematical models) is obtained, they agree in the trend of superior performance of the new design. The Euler number and cut-off diameter for the new design is approximately half that of Stairmand design.
8.5.3 Articial neural network (ANN) approach

Articial neural networks (ANNs) have become an attractive approach for modeling highly complicated and nonlinear system [53, 180, 197]. In this study, the radial basis function neural network (RBFNN) has been used to model the effect of cyclone dimensions on both the pressure drop and the cut-off diameter. For more details about the radial basis function neural networks, the reader can refer Sec. 8.3.1, page 158. Two RBFNNs have been trained using the 27 data set obtained from the CFD simulations for both the Euler number (the dimensionless pressure drop) and the cut-off diameter. In this study, the performance parameters are assumed as functions of only the four geometrical parameters, whereas all other parameters kept constant, Eq. 8.17. 231
Figure 8.36: The contour plots for the time averaged ow variables at sections Y=0 and throughout the inlet section. From top to bottom: the dimensionless static pressure (divided by the dynamic pressure at inlet), the dimensionless tangential velocity, axial velocity. Note: both cyclones have the same air volume ow rate.
232

Table 8.38: The cyclone performance parameters using CFD simulations and four mathematical models
Parameter Euler number Cut-off diameter, m Cyclone Stairmand design New design Stairmand design New design Statgraphics 5.673 2.221 1.706 0.865 CFD 5.606 2.603 1.73 0.69 Ramachandran 4.85 2.71 MM 5.33 2.99 Iozia 1.69 0.95 Ritema 1.55 0.71
4.5
Static pressure / Dynamic pressure at inlet
4 3.5 3 2.5 2 1.5 1 0.5 0 -0.5 -1
1.2
Tangential velocity / Inlet velocity

-0.5 0 0.5 1
0.8
0.6
0.4
0.2
0 -1
-0.5
0.5
Figure 8.37: The radial proles of the time averaged static pressure and tangential velocity at z/D=1.5.
Eu = f1 x50 = f2
a b D x Ht , , , D D D D a b D x Ht , , , D D D D
(8.17)
Due to the large difference in the order of magnitude of the value (cf. Table 8.35), the available dataset is transformed into -1 to 1 interval using the Matlab intrinsic function; mapminmax in order to avoid solution divergence [53, 197]. The RBFNN calculations have been performed using the neural network toolbox available from Matlab commercial software 2010a. The cut-off diameter x50 for cyclone separator is always given in units of m. Another way to represent x50 is using a dimensionless number; Stokes number Stk50 = p x2 50 Vin /(18D ) [37]. It is the ratio between the particle relaxation time; p x2 50 /(18) and the gas ow integral time scale; D/Vin . Fitting the RBFNNs The congured RBFNNs predictions versus the CFD data for the Euler number and cut-off diameter are shown in Fig. 8.38. It can be seen that 233
the RBFNN models are able to attain the high training accuracy. The training mean square errors are zeros (i.e., identical matching between the input and output, the reason behind that may refer to the consistency between the used data in the simulation using the design of experiment, which is not the case for the study of Elsayed and Lacor [53] using experimental data set), Fig. 8.38. This indicates that, in comparison with traditional models of curve tting, the models based on an articial intelligence algorithm have a superior capability of nonlinear tting. Especially, the RBFNN has its unique and optimal approximation characteristics in learning process [53, 197]. Figure 8.38 illustrates the agreement between the RBFNNs input and output. The obtained relation is a typical linear relation with a coefcient of correlation close to 1 (R > 0.999). The agreement between the input and output of the RBFNN is also clear from the value of the mean squared error E 2 . Consequently, the trained neural networks predict very well both the Euler number and cut-off diameter values and can be used in cyclone design and performance estimation. Figure 8.38 present different performance indicators as a validation of the proposed models for the trained data.
26 24 22 20
2.5
18 16 14 12 10 8 6 4 2
Predicted value (y)
Predicted value (y)
y=0.999 x -2.2E-15 R=0.99999 2 E = 0.0
y=0.999 x +4.03E-16 R=0.99999 2 E = 0.0
1.5
10
15
20
25
1.5
2.5
Input value (x)
Input value (x)
(a) The Euler number
(b) The cut-off diameter
Figure 8.38: Linear regression of the RBFNNs for the Euler number and the cut-off diameter.
234

Ht
3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
2.4
Ht
3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
2.2 15
Euler number
10
Dx a b Ht
X 50 [micron]
1.8
1.6
1.4
Dx a b Ht
1.2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.2
0.4
0.6
D x, a, b
D x, a, b
(a) Euler number
(b) Cut-off diameter
Figure 8.39: The effect of geometrical parameters on the cyclone performance.
The effect of the four geometrical parameters on the cyclone performance based on RBFNNs The effects of the geometrical parameters on both the Euler number and the cut-off diameter are depicted in Figs. 8.39(a) and 8.39(b). To study the effect of each parameter, the tested RBFNNs models have been used by varying one parameter at a time from its minimum to maximum values of the available CFD dataset, whereas the other parameters are kept constant at their mean values (cf. Table 8.35). Figures 8.39(a) and 8.39(b) indicate the signicant effect of the vortex nder diameter Dx , the inlet width b, the inlet height a. Less effect is assigned to the total cyclone height Ht . More analysis is given in Tables 8.39 and 8.40.
8.5.4 Optimization Using Genetic Algorithms

Optimal cyclone design for minimum pressure drop The genetic algorithm optimization technique has been applied to obtain the geometrical ratios for minimum pressure drop (Euler number). The objective function is the Euler number (using the trained radial basis function neural network). The design variables are four geometrical dimensions, the inlet height a, the inlet width b, the vortex nder diameter Dx and the total cyclone height Ht . These four variables are the most signicant factors which affect the cyclone performance [50]. 235
Table 8.39: The variation of the Euler number with cyclone dimensions using the RBFNN model (cf., Fig. 8.39(a)) Factor Analysis The most signicant effect is that of the vortex nder diameter Dx with inverse relationship when increasing Dx up to 0.65 after which the relation becomes direct, Fig. 8.39(a). This can be explained as follows: Although the pressure loss in the vortex nder decreases with increasing the vortex nder diameter like the case of viscous ow in a pipe, the Euler number in the cyclone body instead will increase due to the decrease of the ow area just after the ow entrance from the inlet region (the annular space between the barrel and the vortex nder). This analysis indicates the large contribution of the pressure loss in the vortex nder to the total Euler number (the pressure loss at the entrance, the pressure loss in the cyclone body, and the pressure loss in the vortex nder [50]). The inlet height a and width b is almost linearly related to the Euler number. But, why the relation is direct (at the same ow rate increasing the inlet dimensions decreases the inlet velocity. Consequently,the loss in the vortex nder will decrease)? The reason is the increase in the pressure drop at the inlet section, due to deviation of the inlet ow when it mixes with the swirling ow. The effect of changing the total cyclone height Ht on the Euler number is very small.
Dx
a and b
Ht
Table 8.41 presents the settings used to obtain the optimum design for minimum pressure using global optimization Matlab toolbox in Matlab 2010a commercial package. Table 8.42 gives the optimum values for cyclone geometrical parameters for minimum pressure drop estimated by the articial neural network using the genetic algorithm optimization technique. It is clear from Table 8.42 that the new optimized design is very close to the Stairmand design in many geometrical parameters, whereas the new ratios will result in the minimum pressure drop.
Optimal cyclone design for best performance using NSGA-II In case of cyclone separator geometry optimization for minimum Euler number and minimum cut-off diameter, the objectives are conicting with each other. There is no best solution for which all objectives are optimal simultaneously [181]. The increase of one objective will lead to the de236

Table 8.40: The variation of the cut-off diameter with cyclone dimensions using the RBFNN model (cf., Fig. 8.39(b)) Factor Analysis The vortex nder diameter has the most signicant effect on the cut-off diameter x50 (the highest slope in Fig. 8.39(b)). The slope is very high for Dx = 0.4 0.55 and any further increase or decrease in Dx beyond the above range produces a small change in x50 . Increasing the vortex nder diameter decreases the swirling intensity in the cyclone (i.e. Reduction in the centrifugal force). Consequently, low collection efciency (higher x50 ) is obtained. In the meantime, the increase of Dx decreases the pressure drop. This is one of the main reasons of the trade-off between the Euler number and the cut-off diameter objectives. This makes the optimization of cyclone geometry a multi-objective procedure. The variation of x50 with the inlet width and height are similar in trend and signicance to that for Dx but here the slope is high in the range of a = 0.3 0.4 b = 0.2 0.275. The effect of the cyclone total height on the cut-off diameter is due to the effect of the cone height as the barrel height is xed in this study. The slope is very small, with a general trend of inverse relation (increasing the separation space, enhances the collection efciency).
Dx
a and b
Ht
crease of the other objective. Then, there should be a set of solutions, the so-called Pareto optimal set or Pareto front, in which one solution cannot be dominated by any other member of this set [56]. Recently, a number of multi-objective genetic algorithms (MOGAs) based on the Pareto optimal concept have been proposed. The well known nondominated sorting genetic algorithm II (NSGA-II) proposed by Deb et al. [34] is one of the most widely used MOGAs since it provides excellent results as compared with other multi-objective genetic algorithms proposed [25]. Table 8.43 presents the genetic operators and parameters for multi-objective optimization. The Euler number and the Stokes number values have been obtained from the articial neural network trained by the CFD data set. The Pareto front (non dominated points) is presented in Fig. 8.40(a) and Table 8.44. Figure 8.40(a) clearly demonstrate tradeoffs in objective functions (Euler number and Stokes number). All the optimum design points in the Pareto front are non-dominated and could be chosen by a designer as optimum cyclone separator [148]. This set of designs makes the Pareto 237
Table 8.41: Genetic operators and parameters for single objective optimization
Population type Initial range Fitness scaling Selection operation: Elite count Crossover fraction Crossover operation Mutation operation Maximum number of generations: Population size Double vector [0.25 0.15 0.25 3 ; 0.5 0.375 0.75 5 ] for a, b, Dx and Ht respectively Rank Tournament (tournament size equals 4) 2 0.8 Intermediate crossover with the default value of 1.0 The constraint dependent default 800 200
front approach more preferred than the desirability function approach which gives only one design point. Three points A, B and C are indicated in Fig. 8.40(a). Point A indicates the point of minimum Euler number (maximum Stokes number). Point B indicates the point of maximum Euler number (minimum Stokes number). Point C indicates an optimal point for the multi-objective optimization problem. In order to obtain the Euler number-Stokes number relationship, Fig. 8.40(b) has been drawn. It indicates a general relationship (trend) between the two dimensionless numbers irrespective to the four geometrical parameters values. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number, Eq. 8.18. The obtained correlation can t the data with a coefcient of determination R2 = 0.99613 as shown in Fig. 8.40(b). Elsayed and Lacor [56] presented another correlation between the Euler number and the Stokes number, Eq. 8.19. Equation 8.19 presents a good matching for only high values of Euler numbers. For smaller values of Euler number, there is underestimation of the Stokes number. The difference between the two correlations can be referred to two reasons. Firstly, the new correlation is based on the CFD simulations data for both Euler number and Stokes number, whereas Elsayed and Lacor
Table 8.42: The optimized cyclone separator design for minimum pressure drop using genetic algorithm Factor a b Dx Ht Euler number Cut-off diameter Low 0.25 0.15 0.25 3.0 High 0.5 0.375 0.75 5.0 Stairmand design 0.5 0.2 0.5 4.0 5.606 1.706 Optimum design 0.492 0.158 0.617 4.535 2.369 1.704
The values for the Euler number and the cut-off diameter for the Stairmand design have been obtained from CFD simulations, cf. Table 8.38. The values for the Euler number and the cut-off diameter for the new optimal design have been obtained from the trained RBFNN.
238

Table 8.43: Genetic operators and parameters for multi-objective optimization
Population type Population size Initial range Selection operation Crossover fraction Crowding distance fraction Crossover operation Maximum number of generations Double vector 60 (i.e., 15* number of variables which is the default in Matlab) [0.25 0.15 0.25 3 ; 0.5 0.375 0.75 5 ] for a, b, Dx and Ht respectively tournament (tournament size equals 2) 0.8 0.35 Intermediate crossover with the default value of 1.0 800 (i.e., 200* number of variables which is the default in Matlab)
correlation was obtained from experimental data for Euler number values and Iozia and Leith model for the Stokes number values. Secondly, the new correlation is limited to only four geometrical parameters. The other three factors are xed, h = 1.5, S = 0.5 and Bc = 0.375. Consequently, the new correlation (Eq. 8.18) is valid only for these values. Stk50 = 100.3533(log10 (Eu)) Stk50 = 100.3016(log10 (Eu))
2
1.1645log10 (Eu)2.3198
(8.18) (8.19)
0.9479log10 (Eu)2.5154
8.5.5 Conclusions
CFD simulations data have been used to understand the effect of four geometrical parameters on the cyclone performance and to optimize the cyclone geometry. Two meta-models have been used viz., the response surface and the radial basis function neural network approaches. Two optimization techniques have been applied, the desirability function with Nelder-Mead technique and the non-sorted dominated genetic algorithm NSGA-II. The response surface methodology has been used to t two secondorder polynomials to the Euler number and cut-off diameter obtained from CFD simulations. The analysis of variance of the Euler number indicates a strong interaction between Dx with (a, b) and between Dx with b only for the cut-off diameter. The bi-objective functions have been converted to single-objective function using the desirability function approach. A new optimal design has been obtained using the Nelder-Mead technique available in Statgraphics commercial software. The ratio of inlet width to height b/a = 0.589 lays in the optimal cyclone lies in the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. The new design and the Stairmand design have been computationally compared 239
Stokes number x 10
4 3
2 1
C 5 10 15 20 Euler number 25 30
B 35
(a) Pareto front in linear scale
10 8 6
Stk50=10^{0.3533 (LOG10(Eu))^2- 1.1645 (LOG10(Eu) -2.3198} 2 R =0.99613
Stokes number x 10 3
Non-dominated points Elsayed & Lacor correlation Curve-fitting 5 10 15 20 25 30
Euler number
(b) Pareto front in log scale. Note: The correlation (Eq. 8.19) given by Elsayed and Lacor [45] are based on seven geometrical parameters.
Figure 8.40: Pareto front plots obtained from the variation of four geometrical ratios.
240

Table 8.44: The seven geometrical parameters and the obtained Euler number and Stokes number for the nondominated points (Pareto-front).
point 1 2 3 A 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18B 19 20 21C 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 Minimum Maximum a 0.500 0.391 0.374 0.250 0.252 0.250 0.320 0.363 0.259 0.391 0.419 0.251 0.455 0.251 0.298 0.376 0.472 0.500 0.255 0.252 0.286 0.351 0.251 0.265 0.493 0.312 0.352 0.486 0.252 0.252 0.251 0.426 0.261 0.272 0.253 0.251 0.266 0.25 0.50 b 0.375 0.296 0.270 0.150 0.152 0.150 0.214 0.268 0.159 0.284 0.327 0.151 0.365 0.151 0.189 0.270 0.358 0.375 0.155 0.152 0.183 0.233 0.151 0.154 0.321 0.206 0.234 0.365 0.152 0.153 0.151 0.333 0.150 0.161 0.154 0.151 0.284 0.150 0.375 Dx 0.250 0.250 0.252 0.750 0.362 0.732 0.300 0.300 0.472 0.252 0.256 0.711 0.251 0.695 0.441 0.262 0.251 0.250 0.608 0.497 0.346 0.295 0.638 0.460 0.250 0.320 0.295 0.250 0.730 0.676 0.650 0.252 0.747 0.553 0.553 0.711 0.252 0.250 0.750 Ht 4.994 4.979 4.961 4.899 4.893 4.846 4.929 4.945 4.960 4.966 4.983 4.853 4.976 4.858 4.910 4.960 4.986 4.994 4.810 4.879 4.907 4.949 4.897 4.881 4.991 4.869 4.950 4.992 4.840 4.861 4.895 4.977 4.838 4.880 4.880 4.853 4.966 4.810 4.995 h 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 1.5 S 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 Bc 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 0.375 Euler number 33.750 20.806 17.883 0.756 3.326 0.799 8.617 12.160 2.086 19.737 23.514 0.854 29.693 0.895 3.277 16.657 30.324 33.750 1.214 1.759 4.954 10.568 1.050 2.193 28.441 7.136 10.728 31.833 0.817 0.957 1.013 25.126 0.798 1.630 1.451 0.854 13.428 0.7555 33.7497 Stokes number x 103 0.511 0.569 0.595 6.958 1.262 6.446 0.840 0.778 2.080 0.581 0.568 5.903 0.528 5.504 1.554 0.630 0.525 0.511 3.820 2.380 1.103 0.789 4.354 1.780 0.531 0.941 0.784 0.518 6.387 5.071 4.565 0.551 6.822 2.945 3.022 5.903 0.633 0.511 6.958
A indicates the point of minimum Euler number and maximum Stokes number. B indicates the point of maximum Euler number and minimum Stokes number. C indicates an optimal point for the multi-objective optimization problem. (cf. Fig. 8.40(a)).
to get a clear vision for the differences between the ow eld pattern and performance in the two designs. The CFD simulations results and four mathematical models conrmed the better performance of the new design in comparison with the Stairmand design. The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone performance better than the response surface methodology. The used RBFNN presented zero mean squared error and almost unity coefcient of determination. The analysis indicates the signicant effect of the vortex nder diameter Dx and the inlet dimensions a and b on the cyclone performance. Moreover, the range of high inuence is given for each geometrical parameters using the trained RBFNNs. The trained RBFNN for the Euler number has been used to get a new optimized cyclone for minimum pressure drop (Euler number) 241
using the genetic algorithm optimization technique. The new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, and superior in low pressure drop at nearly the same cut-off diameter. But, the optimal design obtained from the desirability function results in a better collection efciency (smaller cut-off diameter) as is clear from Table 8.38 because there the cut-off diameter is included in the desirability function, whereas, the obtained result are for single objective (Euler number). The two trained RBFNNs have been used in a multi-objective optimization process using NSGA-II technique. The Pareto front is presented for the designer with a wide choice for selection. A second-order polynomial has been tted between the logarithms of Euler number and Stokes number to obtain a general formula, 2 Stk50 = 100.3533(log10 (Eu)) 1.1645 log10 (Eu)2.3198 with a coefcient of determination R2 = 0.98643. This formula can be used to obtain the Stokes number if the Euler number is known at h = 1.5, S = 0.5 and Bc = 0.375.
242
Chapter 9 Conclusions and Future Directions

9.1 Conclusions
9.1.1 The most signicant geometrical factors
The geometrical parameters in cyclone separators affect signicantly the ow eld and performance parameters. There are seven geometrical parameters, which can be classied into four classes, namely, the inlet dimensions (height and width), the vortex nder dimensions (diameter and length), the cyclone height (cone and barrel) and the cone-tip diameter. These dimensions do not have the same inuence on the cyclone performance. Consequently, the rst step was to identify the most signicant parameters and the possible interaction between them. The Muschelknautz method of modeling (MM)) has been used to understand the effect of the cyclone geometrical parameters on the cyclone performance. The most signicant geometrical parameters are: 1. The vortex nder diameter 2. The inlet section width 3. The inlet section height 4. The cyclone total height (cone height). The effect of both the barrel height and the vortex nder length on the cyclone separator performance are small in comparison with these most signicant geometrical parameters. There are strong interactions between 243
Chapter 9. Conclusions and Future Directions

the effects of inlet dimensions and the vortex nder diameter on the cyclone performance. Large eddy simulation methodology has been used to study the effect of the cone-tip diameter on the cyclone ow eld and performance. The analysis of results indicates the insignicant effect of the cone-tip diameter on the ow pattern and performance. Four cyclones have been simulated using the Reynolds stress model (RSM), to study the effect of the dust outlet geometry on the cyclone separator performance, ow pattern and velocity proles. The results approved that if the main target of the CFD investigation is the performance parameters, one can safely simulate only the main separation space (cylinder on cone). However, if the aim is to investigate the ow eld pattern, the dust outlet geometry should be included in the simulation domain.
9.1.2
The impact of geometry
The vortex nder dimensions: The maximum tangential velocity in the cyclone decreases with increasing the vortex nder diameter. A negligible change is noticed with increasing the vortex nder length. Increasing the vortex nder length, makes a small change in both the static pressure, axial and tangential velocity proles. However, decreasing the vortex nder diameter gradually changes the axial velocity prole from the inverted W to the inverted V class. Decreasing the cyclone vortex nder diameter, increases the maximum tangential velocity. The Euler number (dimensionless pressure drop) decreases with increasing the vortex nder diameter. Increasing the vortex nder length slightly increases the Euler number. The Stokes number increases with increasing the vortex nder diameter, because the centrifugal force affecting particles attenuates when the swirl intensity (maximum tangential velocity) decreases. The Stokes number slightly increases as the vortex nder length is increased. The inlet dimensions: The maximum tangential velocity in the cyclone decreases with increasing both the cyclone inlet width and height. Increasing the cyclone inlet width or height decreases the pressure drop at the cost of increasing the cut-off diameter. The effect of changing the inlet width on the cut-off diameter is more signicant in comparison with that of the inlet height. The optimum ratio of the inlet width to the inlet height is from 0.5 to 0.7. The cyclone heights: The maximum tangential velocity in the cyclone decreases with increasing the cyclone (barrel or cone) height. Increasing the barrel height, makes a small change in the axial veloc244
9.1. Conclusions
ity. Increasing the cyclone barrel height decreases the pressure drop and the cut-off diameter. The changes in the performance beyond h/D = 1.8 are small. Increasing the cone height makes a considerable change in the axial velocity. Both the pressure drop and the cutoff diameter decrease with increasing the cyclone cone height. The performance improvement stops after hc /D = 4.0 (Ht /D = 5.5). The effect of changing the barrel height is less signicant on the performance and the ow pattern in comparison with the effect of the cone height.
9.1.3 Optimization
Several new optimized cyclone geometrical ratios have been obtained and presented in this thesis. All the new geometrical ratios result in better performance than the Stairmand design. For example, The new cyclone design ratios obtained using MM model and Nelder-Mead technique for minimum pressure drop, are very close to the Stairmand high efciency design in the geometrical parameter ratios, and superior for low pressure drop at nearly the same cut-off diameter. The new cyclone design results in nearly one-half the pressure drop obtained by the Stairmand design at the same volume ow rate. Articial neural networks In order to accurately predict the complexly non linear relationships between pressure drop and geometrical dimensions, a radial basis neural network (RBFNN) is developed and employed to model the pressure drop for cyclone separators. The neural network has been trained and tested by the experimental data available in literatures. The result demonstrates that articial neural networks can offer an alternative and powerful approach to model the cyclone performance. Four mathematical models (Muschelknautz method MM, Stairmand, Ramachandran and Shepherd and Lapple) have been tested against the experimental values. The residual error of MM model is the lowest. Moreover, one can multiply the calculated value using MM by 1.5 to get the experimental value (as a rough approximation). The response surface methodology has been used to t a second-order polynomial to the RBFNN. The second-order polynomial has been used to get a new optimized cyclone for minimum pressure drop using the Nelder-Mead technique. 245

A comparison between the new design and the standard Stairmand design has been performed using CFD simulation to obtain a clear vision of the ow eld pattern in the new design. The new cyclone design results in nearly 75% of the pressure drop obtained by the Stairmand high efciency design at the same volume ow rate.
9.1.4
Multi-objective optimization
Two multi-objective optimization approaches have been applied in this thesis, namely the desirability function and the non-dominated sorted genetic algorithm (NSGA-II). Moreover, two sources of data have been used, namely: (1) The experimental data for the pressure drop (Euler number) and the Iozia and Leith model for the cut-off diameter (Stokes number) for the optimization of the seven geometrical parameters. (2) CFD simulations for the optimization of only four geometrical parameters. Seven geometrical parameters Two radial basis neural networks (RBFNNs) are developed and employed to model the Euler number and the Stokes number for cyclone separators. The neural networks have been trained and tested by the experimental data available in literature for Euler number (pressure drop) and the Iozia and Leith model [85] for the Stokes number (cut-off diameter). The effects of the seven geometrical parameters on the Stokes number have been investigated using the trained ANN. To declare any interaction between the geometrical parameters affecting the Stokes number, the response surface methodology has been applied. The trained ANN has been used as an objective function to obtain the cyclone geometrical ratios for minimum Euler number using the genetic algorithms optimization technique. A CFD comparison between the new optimal design and the Stairmand design has been performed. A multi-objective optimization technique using NSGAII technique has been applied to determine the Pareto front for the best performance cyclone separator. The trained RBFNN has been used to get a new optimized cyclone for minimum pressure drop (Euler number) using the genetic algorithm optimization technique. The new cyclone design results in nearly 68% of the pressure drop obtained by the Stairmand design at the same volume ow rate. This 246
9.1. Conclusions
conrms that the obtained design using the genetic algorithm is better than that obtained using Nelder-Mead technique which results in 75% of the Stairmand pressure drop [53]. The two trained RBFNNs have been used in a multi-objective optimization process using NSGA-II technique. Sixteen test cases with different barrel diameter, gas ow rate and particle density have been tested. The Pareto fronts for these test cases are very close. A second-order polynomial has been tted between the logarithms of the Euler number and the Stokes number to obtain a general for2 mula, Stk50 = 100.3016(log10 (Eu)) 0.9479 log10 (Eu)2.5154 . This formula can be used to obtain the Stokes number if the Euler number is known. Four geometrical parameters CFD simulations data have been used to understand the effect of four geometrical parameters on the cyclone performance and to optimize the cyclone geometry. Two meta-models have been used viz., the response surface and the radial basis function neural network approaches. Two optimization techniques have been applied, the desirability function with Nelder-Mead technique and the non-sorted dominated genetic algorithm NSGA-II. The response surface methodology has been used to t two secondorder polynomials to the Euler number and the cut-off diameter obtained from CFD simulations. The analysis of variance of the Euler number indicates a strong interaction between Dx with (a, b) and between Dx with b only for the cut-off diameter. The bi-objective functions have been converted to single-objective function using the desirability function approach. A new optimal design has been obtained using the Nelder-Mead technique. The ratio of inlet width to height b/a = 0.589 i.e., in the optimal cyclone lies in the recommended range of b/a from 0.5 to 0.7 proposed by Elsayed and Lacor [52]. The new design and the Stairmand design have been computationally compared to get a clear vision for the differences between the ow eld pattern and performance in the two designs. The CFD simulations results conrmed the superior performance of the new design in comparison with the Stairmand design. The result demonstrates once more that articial neural networks can offer an alternative and powerful approach to model the cyclone performance better than the response surface methodology. The used 247

RBFNN presented zero mean squared error and almost unity coefcient of determination. The analysis indicates the signicant effect of the vortex nder diameter Dx and the inlet dimensions a and b on the cyclone performance. Moreover, the range of high inuence is given for each geometrical parameters using the trained RBFNNs. The trained RBFNN for the Euler number has been used to get a new optimized cyclone for minimum pressure drop (Euler number) using the genetic algorithm optimization technique. The new cyclone design is very close to the Stairmand high efciency design in the geometrical parameter ratio, and superior in low pressure drop at nearly the same cut-off diameter. But, the optimal design obtained from the desirability function results in a better collection efciency (smaller cut-off diameter) because there the cut-off diameter is included in the desirability function, whereas, the obtained result are for single objective (Euler number). The two trained RBFNNs have been used in a multi-objective optimization process using NSGA-II technique. The Pareto front is presented for the designer with a wide choice for selection. A second-order polynomial has been tted between the logarithms of the Euler number and the Stokes number to obtain a general for2 mula, Stk50 = 100.3533(log10 (Eu)) 1.1645 log10 (Eu)2.3198 . This formula can be used to obtain the Stokes number if the Euler number is known at h = 1.5, S = 0.5 and Bc = 0.375.
9.2 Future Directions

As a recommendation of future work, the same study is to be performed but at different ow rates and different particle densities. Furthermore, the effect of the cyclone dimensions on natural vortex length and precessing vortex core needs more investigation using large eddy simulation methodology. The present study on the dust outlet geometry is to be enlarged to include the following. (1) The effect of dustbin and dipleg shape, length, diameter, and interior details (like the inner cone at the dustbin inlet) and optimization of these dimensions. (2) The effect of air and dust ow rates, and particle properties. (3) perform the same study but using large eddy simulation (LES) to check the effect of dust outlet geometry on the vortex core precession. 248
9.2. Future Directions

The following issues still need more investigation. (1) test more mathematical models against the experimental value. (2) comparison between the support vector machine approach and neural networks approach for cyclone separator performance estimation. (3) create a neural network model to design the cyclone separator and estimate its performance parameters. (4) generate performance curves for each geometrical and operating parameters that affect the cyclone performance to help the designer in predicting the change of the performance due to change in the cyclone loading and operating conditions. Furthermore, many parameters can be taken into consideration in the optimization process, e.g., erosion rate for harsh environments, surface roughness, interaction between particles using two-phase ow modeling. Another idea is to perform an geometry optimization (aeroacoustic) study to reduce the cyclone-noise especially for domestic applications.
249
250
Appendix A Mathematical models

Since the rst application of aerocyclones in 1886 [3], theories for the estimation of both particle collection efciency and pressure drop of cyclone have been developed by many contributors using different methods with various simplifying assumptions. During the past 50 years, interest in particle collection and pressure theories has steadily increased [196]. The most widely used mathematical models for the cut-off diameter and pressure drop estimation are: Barth model [9] The Muschelknautz method of modeling (MM) [29, 77, 114116, 174, 175] Stairmand model [165] Shepherd and Lapple model [157] Casal and Martinez-Bent model [21] Ramachandran model [139] Iozia and Leith model [84] Rietema model [142]
A.1 General assumptions

Some simplifying assumptions are common to all these models. They can be considered as offering a good compromise between accurate prediction and simplication of the equations [3]. They are: The particles are spherical. The particle motion is not inuenced by the presence of neighboring particles. The radial velocity of the gas equals zero. 251
APPENDIX A. Mathematical models
Figure A.1: The control surface concept in the equilibrium-orbit model.
The radial force on the particle is given by Stokes law.
A.2 Barth model

In 1956, Barth [9] proposed a simple model based on force balance (classied as one of the equilibrium-orbit models [77]). This model enables to obtain the cut-off diameter and the pressure drop values. Barth proposed calculating the wall velocity (the tangential velocity near the wall of the cyclone surface) and the tangential velocity at the control surface CS ( Fig. A.1) through two steps: 1. calculate the wall velocity vw (the velocity outside CS ) from the inlet velocity vin . 2. use its value to calculate the tangential velocity, vCS at CS . In a cyclone with a slot type of rectangular inlet, the inlet jet is compressed against the wall, resulting in a decrease in the area available for the incoming ow and an increase in the velocity. Barth accounted for this by introducing , which is dened as the ratio of the moment-of-momentum of the gas in the inlet and the gas owing along the wall. = vin Rin vw R 252 (A.1)
A.2. Barth model
Figure A.2: Inlet ow pattern for tangential inlet cyclone [77]
where Rin is the radial position of the center of the inlet (Fig. A.2), For a b slot inlet Rin = R 2 , where b is the inlet width and R = D/2 is the cyclone radius. Hoffmann and Stein [77] gave an algebraic relations for , the simplest of which is b R
0.5
= 1 0 .4
(A.2)
Rin So vw can be calculated as vw = vin R , as the inlet velocity is given or Q calculated from vin = a b where Q is the gas volume ow rate, a is the inlet height. To get from vw to the tangential velocity at CS vCS , Barth gave the following relation R vw ( R ) x
vCS =
1+
HCS Rf vw Q
(A.3)
Introducing Eq. (A.1), one obtains vCS = Rin Rx vx a b + HCS f Rin (A.4)
x Where Rx is the vortex nder radius Rx = D 2 , vx is the mean axial velocity Q in the vortex nder vx = R2 , and HCS is the height of the control surface x extending from the bottom of the vortex nder to the cyclone bottom.
253

For the friction factor f , the following relation can be used for hydraulically smooth cyclone surface according to Hoffmann and Stein [77], f = 0.05(1 + 3 co ) (A.5)
Where co is the mass ratio of dust feeding the cyclone to the gas ow rate (dimensionless).
Estimation of the pressure drop

Barth subdivided the pressure drop into three contributions: 1. the inlet losses (Barth assumed that this loss could be effectively avoided by good design). 2. the losses in the cyclone body. 3. the losses in the vortex nder. The pressure drop in the cyclone body can be estimated from, 1 2 v 2 x Dx D 1
vx vCS
Pbody =
H S 0.5Dx f
vCs vx
(A.6)
where f is the friction factor (calculated from Eq. A.5). This model accounts for the effect of solid loading upon pressure loss via the total friction factor f . The pressure drop in the vortex nder can be estimated using a semi- empirical approach as, Px = 1 2 v 2 x vCS vx
2
+K
vCS vx
4 3
(A.7)
where K is the vortex nder entrance factor (K = 3.41 for rounded edge and K = 4.4 for sharp edge) The total pressure drop P = Pbody + Px can be made dimensionless Q using the average inlet velocity vin = ab leading to the so called Euler number Eu Eu = P 2 0.5vin (A.8)
254
A.2. Barth model
Estimation of the cut-off diameter

As mentioned above, the Barths model [9] is based on an equilibrium-orbit model [77]. This model considers the imaginary cylindrical surface CS that is formed by continuing the vortex nder wall to the bottom of the cyclone Fig. A.1. It is based on the force balance of a particle that is rotating in CS x at radius Rx = D 2 . The outwardly directed centrifugal force is balanced against the inward drag caused by the gas owing through surface CS into the inner part of the vortex. Fcentrif ugal = ma = mp
2 vCS v2 = p 6 x3 CS Rx Rx
(A.9)
where the subscript p is for particle properties and x is the particle diame xv ter. Since, for a creep ow, Rep 1 where Rep = gg rp , the drag coefcient 24 . Hence, the Stokesian drag force can be estimated as: for sphere CD = Re p Fdrag = 3g vrCS x (A.10)
where vrCS is the uniform radial gas velocity in the surface of CS given by: vrCS = Q Dx HCS (A.11)
The following expression obtained from trigonometry relations can be used to calculate HCS (R Rx )(H h) + (h S ) c R (B 2 ) (H S )
HCS
= =
if Bc > Dx if Bc Dx (A.12)
Large particles are therefore centrifuged out to the cyclone wall (because the centrifugal force is larger than the drag force. Whereas, small particles are dragged in and escape out through the vortex nder. The particle size for which the two forces balance the particles that orbit in equilibrium in CS is taken as the cyclones cut-off diameter x50 ; it is the particle size that stands a 50 - 50 chance of being captured. This particle size is called of fundamental importance and is a measure of the intrinsic separation capability of the cyclone. Here, all the gas velocity components are assumed constant over CS for the computation of the equilibrium-orbit size. Equating the forces, in Eqs. A.9 and A.10 gives the cut-off diameter x50 as: x50 = 9 vrCS g Dx 2 p vCS (A.13)
255
A.3 The Muschelknautz method of modeling (MM)

Hoffmann and Stein [77] stated that the most practical method for modeling cyclone separators at the present time is the Muschelknautz method (MM) [29, 77, 114116, 174, 175]. The roots of the Muschelknautz method (MM) extend back to an early work performed by Barth [9] as it is based on the equilibrium orbit model [77].
The pressure loss in cyclone

According to the MM model, the pressure loss across a cyclone occurs, primarily, as a result of friction with the walls and irreversible losses within the vortex core, the latter often dominating the overall pressure loss, p = pbody + px . In a dimensionless form, it is dened as the Euler number. Eu = 1
1 2 2 vin
[pbody + px ]
(A.14)
The wall loss, or the loss in the cyclone body is given by, pbody = f AR 1.5 (vw vCS ) (0.9 Q) 2 (A.15)
where is the gas density, Q is the gas volume ow rate, AR is the total inside area of the cyclone contributing to frictional drag. It encompasses the inside area of the roof, the barrel cylinder, the cone, and the external surface of the vortex nder.
AR = Aroof + Abarrel + Acone + Avortex f inder

2 = R2 Rx + 2Rh + (R + Rb )
(Ht h)2 + (R Rb )2
+2Rx S (A.16)
where R = D/2, Rx = Dx /2, Rb = Bc /2. As in the Barths model [9, 77], the tangential gas velocity in the entire space between the wall and the vortex nder can be signicantly higher than the inlet velocity due to constriction of the inlet jet (see Fig. A.2). For tangential inlet, the inlet jet is compressed against the wall, resulting in a decrease in the area available for the incoming ow, and an increase in the velocity. Barth accounts for this by introducing (see Eq. A.1). 256
A.3. The Muschelknautz method of modeling (MM)

Muschelknautz computes the entrance constriction coefcient for a conventional slot-type inlet from the following empirical formula, Eq.A.17 [77]. 1 1 = 2)
1+4
(1
2 )(2 1 + c0
(A.17)
where = b/R, c0 is the ratio of the mass of the incoming solids to the mass of the incoming gas in the stream feeding the cyclone. Knowing , along with vin , Rin and R, one can compute the wall velocity, vw (velocity in the vicinity of the wall), vw = (vin Rin )/(R). Now compute the geometric mean radius, Rm = Rx R which is needed in the computation of a wall axial velocity vzw . vzw = 0 .9 Q 2 ) (R2 Rm (A.18)
Trefz and Muschelknautz [174] found that, approximately 10% of the incoming gas short-circuits the cyclone and ows radially inwards in a spiral like manner along the roof and down the outside of the vortex nder. As a consequence, approximately 90% of the incoming ow Q directly participates in the ow along the walls and in the formulation of the inner vortex [77]. To calculate the friction factor f , Muschelknautz and Trefz [77] dened the cyclone body Reynolds number ReR as: ReR = Rin Rm vzw Ht (A.19)
with and are the gas phase density and absolute viscosity, respectively. Ht is the cyclone total height. The friction factor of the clean gas fair = ks ks , ReR where is the wall relative roughness of the cyclone wall f R R ks (ks = 0.046 [mm] for commercial steel pipe, < 6E 4 for non-negative R logarithm in Eq. A.22). The gas friction factor can be expressed as the sum of two components, one for smooth wall, fsm , plus an added contribution due to wall roughness fr . fair = fsm + fr
0.623 fsm = 0.323Re R
(A.20) (A.21)
257
(A.22) The total frictional drag f within a cyclone consists of two components in the MM, that due to drag on the (pure) gas phase fair and that due to an additional drag imposed by the moving strand of solids, which is present at the walls. The total friction factor f becomes, f = fair + 0.25 R Rx
0.625
fr = log
1 .6 ks 0.0005999 R
2.38 2
1 +
2.25E 5 Re2 R ks 0.000599 R
0.213
c0 Frx str
(A.23)
The second term in this equation is the frictional contribution due to the solids, where is the overall efciency, that is the fraction of incoming solids collected by the cyclone (Hoffmann and Stein [77] suggested to asx ) sume a value of 0.9 to 0.99). Frx is the Froude number (Frx = 2vR x g where vx is the average axial velocity through the vortex nder vx (vx = Q ), str term represents the bulk density of the dust or strand layer at R2 x the walls and can be taken as 0.4bulk where bulk is the bulk density of the solid [77]. The tangential velocity of the gas at the inner core radius Rcs (see Fig. A.3) is given by,
R Rx
vcs = vw 1+
(A.24)
R Rx
f AR 2Q
The second contribution to pressure drop is the loss in the core and in the vortex nder and is given by, px = 2 + vcs vx
2
+3
vcs vx
4/3
1 2 v 2 x
(A.25)
Cut-off diameter
A very fundamental characteristic of any lightly loaded cyclone is its cut-off diameter x50 produced by the spin of the inner vortex. This is the practical diameter that has a 50% probability of capture. The cut-off diameter is 258
A.4. Stairmand model for pressure drop

Q
Vx
0.1Q
Vcs Vz Vw
Rcs
Rc
Figure A.3: Geometric parameters and velocities used in MM model [77]
analogous to the screen openings of an ordinary sieve or screen [77]. In lightly loading cyclones, x50 exercises a controlling inuencing on the cyclones separation performance. It is the parameter that determines the horizontal position of the cyclone grade-efciency curve (fraction collected versus particle size). For low mass loading, the cut-off diameter can be estimated in MM using Eq. A.26 [77].
x50 =
9 (0.9 Q) 2 (H S ) (p )vcs t
(A.26)
A.4 Stairmand model for pressure drop

Stairmand [165] estimated the pressure drop as entrance and exit losses combined with the static pressure loss in the swirl.
2
Eu = 1 + 2q 2
2 (D b) 1 +2 Dx 259
4ab 2 Dx
(A.27)

where q is given by,
Dx 2(Db) 0.5
q=
2AR G ab
Dx 2(Db)
4AR G ab
0.5
(A.28)
where AR is the total wall area of the cyclone body, including the inner walls of the lid, the cylindrical and the conical sections and the outer wall of the vortex nder, given by Eq. A.29. G(= f /2 where f is the friction factor) is a wall friction factor, which Stairmand set equal to 0.005.
0.5
2 D 2 Dx (D + Bc ) AR = + Dh + Dx S + 4 2
(H h) +
D Bc 2
(A.29)
A.5 Purely empirical models for pressure drop

1. Sphered and Lapple model [157] 16 a b 2 Dx
Eu =
(A.30)
2. Casal and Martinez-Bent model [21] ab 2 Dx

2
Eu = 3.33 + 11.3 3. Ramachandran model
(A.31)
The Ramachandran et al. [139] model was developed through a statistical analysis of pressure drop data for ninety-eight cyclone designs. The model is shown to perform better than the pressure drop models of Shepherd and Lapple [157], and Barth [9] in comparison with experimental results.
S D H h Bc D D D 1/3
Eu = 20
ab 2 Dx 260
(A.32)
A.6. Iozia and Leith model for the cut-off diameter
Figure A.4: The control surface used in Iozia and Leith model [85].
A.6 Iozia and Leith model for the cut-off diameter

The Iozia and Leith model [84] is similar to the model of Barth [9] as it is also based on the equilibrium-orbit theory (Force balance). Iozia and Leith [84] gave the following expression for the cut-off diameter, x50 = (9 Q) / HCS p V2 max
1/2
(A.33)
where HCS is the core height (height of the control surface of Barths model) Vmax is the maximum tangential velocity, that occurs at the edge of the control surface CS , Fig. A.4. In this model however the value of the core diameter dc and the tangential velocity at the core edge Vmax are calculated from regression of experimental data using the following equations. Vmax = 6.1vin ab/D2
0.61
(Dx /D)
0.74
(Ht /D)
0.33
(A.34) (A.35)
dc = 0.52D ab/D2
0.25
(Dx /D)1.53
It is clear from this model that the most important geometry parameters that affect the cyclone collection efciency are the vortex nder diameter, the ratio of inlet area to exit area and the cyclone height. 261
A.7 Rietema model for cut-off diameter

The Rietema model [142] estimates the cut-off diameter x50 using Eq. A.36. x50 = Q (p ) Ht P (A.36)
This model relates, the separation cut-off diameter x50 to the pressure drop. Hence, the pressure drop needs to be predicted to use the model. A good pressure drop model for this purpose is that of Shepherd and Lapple. The interested reader can refer to Hoffmann and Stein [77] for more details.
262
Appendix B Optimization Techniques

Two optimization techniques have been used throughout this thesis, namely the Nelder-Mead technique and the genetic algorithms. In the following sections, the details of these techniques will be explained.
B.1 Nelder-Mead
The Nelder-Mead method, also known as downhill simplex method is a commonly used nonlinear optimization technique, The technique was proposed by Nelder and Mead [119] and is a technique for minimizing an objective function in a many-dimensional space. It requires only function evaluations, and no calculation of derivatives [134]. The downhill simplex nonlinear optimization technique has been used by many researchers [e.g., 4, 12]. According to Bernon et al. [12] Powells algorithm and the downhill simplex one are ones of the most used minimization algorithms; the downhill-simplex algorithm became the most performant. Further more, the Statgraphics XV package has been used for design of experiment and optimization, with the only available optimization technique is the NelderMead technique. In this study, the target is to obtain the global optimum values. Consequently, no linear constrains applied. The idea is to employ a moving simplex in the design space to surround the optimal point and then shrink the simplex until its dimensions reach a specied error tolerance [96]. In n-dimensional space, a simplex is a gure of n +1 vertices connected by straight lines and bounded by polygonal faces. If n = 2, a simplex is a triangle; if n = 3, it is a tetrahedron. For two variables, the simplex is a triangle and the method is a pattern search that compares function values at the three vertices of a triangle. The worst vertex where f (x, y ) is largest, is rejected and replaced with a 263
APPENDIX B. Optimization Techniques
new vertex. A new triangle is formed, and the search is continued. The process generates a sequence of triangles (which might have different shapes), for which the function values at the vertices get smaller and smaller. The size of the triangles is reduced, and the coordinates of the minimum point are found [108]. The algorithm is stated using the term simplex (a generalized triangle in N dimensions) and will nd the minimum of a function of N variables. It is effective and computationally compact. In the following paragraphs, a brief explanation is given for this technique. The interested reader can refer to Mathews and Fink [108].
!
R d M d W G Reection B W G Expansion d d M B R E d
W G B R M C1 W G Contraction W G Shrinkage C2 S M Original Simplex B
Figure B.1: Basic operations in the downhill simplex method for two dimensions space [108]
The initial triangle BGW

Let f (x, y ) be the function that is to be minimized. To start, we are given three vertices of a triangle: V k = (xk , yk ), k = 1, 2, 3. The function f (x, y ) is then evaluated at each of the three points zk = f (xk , yk ) for k = 1, 2, 3. The subscripts are reordered so that z1 z2 z3 . We use the notation [108], B = (x1 , y1 ), G = (x2 , y2 ), and W = (x3 , y3 ) (B.1)
to help remember that B is the best vertex, G is good (next to best), and W is the worst vertex. 264
B.1. Nelder-Mead
Midpoint of the good side

The construction process uses the midpoint of the line segment joining B and G. It is found by averaging the coordinates: M= B+G = 2 x1 + x2 y1 + y2 , 2 2 (B.2)
Reection using the point R

The function decreases as we move along the side of the triangle from W to B, and it decreases as we move along the side from W to G. Hence it is feasible that f (x, y ) takes on smaller values at points that lie away from W on the opposite side of the line between B and G. We choose a test point R that is obtained by Reecting the triangle through the side BG. To determine R, we rst nd the midpoint M of the side BG. Then draw the line segment from W to M and call its length d. This last segment is extended a distance d through M to locate the point R (see Fig. B.1). The vector formula for R is [108] R = M + (M W ) = 2M W (B.3)
Expansion using the point E

If the function value at R is smaller than the function value at W, then we have moved in the correct direction toward the minimum. Perhaps the minimum is just a bit further than the point R. So we extend the line segment through M and R to the point E. This forms an expanded triangle BGE. The point E is found by moving an additional distance d along the line joining M and R (see Fig. B.1). If the function value at E is less than the function value at R, then we have found a better vertex than R. The vector formula for E is [108] E = E + (R M ) = 2R M (B.4)
Contraction using the point C

If the function values at R and W are the same, another point must be tested. Perhaps the function is smaller at M, but we cannot replace W with M because we must have a triangle (for 2-D space). Consider the two midpoints C1 and C2 of the line segments W M and M R, respectively (see Fig. B.1). The point with the smaller function value is called C, and the new triangle is BGC. 265
Shrink toward B
If the function value at C is not less than the value at W, the point G and W must be shrunk towards B (see Fig. B.1). The point G is replaced with M, and W is replaced with S, which is the midpoint of the line segment joining B with W [108].
Logical decisions for each step

A computationally efcient algorithm should perform function evaluation only is needed. In each step, a new vertex is found, which replace W. As soon as it is found, further investigation is not needed, and the iteration step is completed. The logical details for two-dimensional cases are explained in Fig. B.2.
B.2
Genetic algorithms (GA)
The genetic algorithm is an optimization technique for solving both constrained and unconstrained optimization problems that is based on natural selection, the process that drives biological evolution [109]. The genetic algorithm repeatedly modies a population of individual solutions. At each step, the genetic algorithm selects individuals at random from the current population to be parents and uses them to produce the children for the next generation. Over successive generations, the population evolves toward an optimal solution. The genetic algorithm can be used to solve a variety of optimization problems that are not well suited for standard optimization algorithms, including problems in which the objective function is discontinuous, no differentiable, stochastic, or highly nonlinear [109]. The genetic algorithm uses three main types of rules at each step to create the next generation from the current population [109]: Selection rules select the individuals, called parents, that contribute to the population at the next generation. Crossover rules combine two parents to form children for the next generation. Mutation rules apply random changes to individual parents to form children. The genetic algorithm differs from a classical, derivative-based, optimization algorithm in two main ways, as summarized in Table B.1. 266
Start a new cycle
es
If f (R) < f (G)
No
Case (i) either reect or extend
Case (ii) either contract or shrink
Yes Yes
If f (B) < f (R)
If f (R) < f (W)
No
No
Replace W with R Compute C-(W+M)/2 or C=(M+R)/2 and f (C)
Compute E and f (E)
Replace W with R
Compute E and f (E)
267
Start next cycle ...
Start next cycle ...
Yes
If f (E) < f (B)
No
B.2. Genetic algorithms (GA)
Yes
If f (C) < f (W)
No
Replace W with E
Replace W with R
Replace W with C
Compute S and f (S) Replace W with S Replace G with M
Start next cycle ... Start next cycle ... Start next cycle ... Start next cycle ...
Figure B.2: The logical decisions for the Nelder-Mead algorithm

Table B.1: Comparison between the classical algorithms and the genetic algorithm Classical Algorithm
It generates a single point at each iteration. The sequence of points approaches an optimal solution. It selects the next point in the sequence by a deterministic computation. In most cases, it nd a relative (local) optimum that is closest to the starting point.
Genetic Algorithm
It generates a population of points at each iteration. The best point in the population approaches an optimal solution. It selects the next population by computation, which uses random number generators. It can nd the global optimum solution with a high probability.
B.2.1 Description of the genetic algorithm process

Genetic algorithms use a population of congurations, called individual, to evolve over a number of generations. Each individual is represented by its genetic material, called a chromosome and every variable in each individual is termed as the gene. For optimization purpose, the chromosome is described by the design variables. The process starts with an initial population of n individuals. The performance of each individual is then evaluated in regard to the tness function and the handling of constraints (if some are considered) [28]. A selection is done in the population to identify valuable parents. Higher is the performance of an individual, higher is its probability to become parent [28]. Two parents are match randomly to exchange their genetic materials to form the offspring for the next generation. This exchange process is called crossover. If this process does not happen, the parents are directly transferred to the next generation meaning the cloning of these individuals. After the crossover operator and before forming the next generation, all the individuals are forced to undergo a mutation process [28]. The evolution procedure is repeated until the population converges to a certain level or simply if the maximum number of generation is reached, Fig. B.3.
B.2.2 Genetic operators Population

Population type species the type of the input to the tness function [109]. There are two options in Matlab, double vector and Bit string (binary chromosome). The double vector option has been used in this study. A double vector chromosome is simply a row vector of double values. These values can be thought of as genes. Thus, 7 genes/chromosome means a double 268
Start
Create initial random population
Evaluate tness for each individual
Store best individuals (chromosomes)
Creating the mating pool (selection of parents)
Create the next generation by applying crossover & mutation (new generation: ospring)
Evaluate tness for each individual
No
Optimal solution found? Stooping criteria met?
Yes Stop
Figure B.3: Flow chart for the genetic algorithms process
vector of 7 elements [27]. The population size n species how many individuals there are in each generation.
Fitness scaling
The scaling function converts raw tness scores returned by the tness function to values in a range that is suitable for the selection function. Scaling function species the function that performs the scaling. One can choose from the following functions [109]: Rank, proportional, Top or Shift linear. The rank option has been used in this study. Rank scales the raw scores based on the rank of each individual, rather than its score. The rank of an individual is its position in the sorted scores. The rank of the ttest individual is 1, the next ttest is 2, and so on. Rank tness scaling removes the effect of the spread of the raw scores. 269
Selection
Different types of selections are implanted in the optimization toolbox in Matlab 2010a [109], but only the tournament selection is used in this study (it is the only available selection type for multi-objective optimization in the Matlab toolbox). The tournament selection randomly identies some competitors from the population to compete against each other. The one with the highest performance win a parent status. The tournament selection permits to control the selective pressure put on the population. The population diversity is adjusted by modication of the competitor number. Greater competitor numbers in the tournament increase the chances to focus the search over the best individuals meaning a greater selective pressure. On the other hand, with only two competitors, the possibility of becoming a parent remains open to a larger band of the population meaning a lesser selective pressure [28]. In other words, the tournament size can take values between 1 and n (population size). Larger values give more chances to the best samples to be selected and to create offsprings. It favors a rapid, although perhaps premature, convergence to a local optimum. Very small values result in a more random selection of parents [173]. The default value of tournament size of four has been used.
Reproduction
Reproduction options determine how the genetic algorithm creates children at each new generation. Elite count species the number of individuals who are guaranteed to survive to the next generation. the Matlab manual [109] suggested to use Elite count to be two (a positive integer less than or equal to population size). Crossover fraction species the fraction of the next generation that crossover produces. Mutation produces the remaining individuals in the next generation. Matlab manual [109] suggested to use a crossover fraction of 0.8 (a fraction between 0 and 1).
Crossover
Crossover combines two individuals, or parents, to form a new individual, or child (offspring), for the next generation [109]. the Matlab optimization toolbox offers six functions: Scattered, single point, two point, intermediate, heuristic and arithmetic. The intermediate function has been used in this study. Intermediate (also, called weighted crossover [28]) creates children by a random weighted average of the parents. Intermediate crossover is controlled by a single parameter ratio: child1 = parent1+ rand*Ratio*(parent2 - parent1). If the ratio is in the range [0,1], the children produced are within the hypercube dened by the parents locations 270

at the opposite vertices. If Ratio is in a larger range, say 1.1, children can be generated outside the hypercube. Ratio can be a scalar or a vector of length equals the number of variables. If the ratio is a scalar, all the children lie on the line between the parents. If the ratio is a vector, children can be any point within the hypercube.
Mutation
Mutation functions make small random changes in the individuals in the population, which provide genetic diversity and enable the genetic algorithm to search for a broader space. To specify the function that performs the mutation in the mutation function eld, one can choose from the following functions: constraint dependent default, gaussian, uniform and adaptive feasible. The constraint dependent default chooses: Gaussian if there are no constraints or adaptive feasible otherwise. 1) Gaussian adds a random number to each vector entry of an individual. This random number is taken from a gaussian distribution centered on zero. The standard deviation of this distribution can be controlled with two parameters. The scale parameter (default value of 1.0) determines the standard deviation at the rst generation. The shrink parameter controls how the standard deviation shrinks as generations go by. If the shrink parameter is 0, the standard deviation is constant. If the Shrink parameter is 1 (the default value), the standard deviation shrinks to 0 linearly as the last generation is reached. 2) Uniform is a two-step process. First, the algorithm selects a fraction of the vector entries of an individual for mutation, where each entry has the same probability as the mutation rate of being mutated. In the second step, the algorithm replaces each selected entry by a random number selected uniformly from the range for that entry. 3) Adaptive feasible randomly generates directions that are adaptive with respect to the last successful or unsuccessful generation. A step length is chosen along each direction so that linear constraints and bounds are satised. The constraint dependent default has been used in this study.
Stopping criteria
The stopping criteria determines what causes the algorithm to terminate. Matlab optimization toolbox has the following stopping criteria, the default values are given in parentheses. Generations specic the maximum number of iterations the genetic algorithm performs (100). Time limit species the maximum time in seconds the genetic algorithm runs before stopping (). 271
Fitness limit - If the best tness value is less than or equal to the value of tness limit, the algorithm stops (). Stall generations - If the weighted average change in the tness function value over stall generations is less than Function tolerance, the algorithm stops (50). Stall time limit - If there is no improvement in the best tness value for an interval of time in seconds specied by Stall time limit, the algorithm stops (). Function tolerance - If the cumulative change in the tness function value over stall generations is less than Function tolerance, the algorithm stops (1E-6).
B.3
Multi-objective optimization
In case of cyclone separator geometry optimization for minimum Euler number and minimum cut-off diameter, the objectives are conicting with each other. There is no best solution for which all objectives are optimal simultaneously [181]. The increase of one objective will lead to the decrease of the other objective. Then, there should be a set of solutions, the so-called Pareto optimal set or Pareto front, in which one solution cannot be dominated by any other member of this set. The denition of domination is as given in Wang et al. [181] For minimal problem, a solution a X dominates a solution b X (a b) if and only if it is superior or equal in all objectives and at least superior in one objective. This can be expressed as follows [181]: a b, if i 1, 2, . . . , m : fi (a) fi (b) j 1, 2, . . . , m : fj (a) < fj (b)
where m is the number of objective functions. Each solution of Pareto optimal set is called a non-inferior solution, which is corresponding to one point on the Pareto front. A general example of the Pareto front with two objectives is illustrated in Fig. B.4. In this example, the Pareto front is composed of six points, which are indifferent to each other (denoted by lled circles in Fig. B.4). While, points with hollow circles are not belonging to the Pareto front, since they are dominated by the Pareto front points. 272
B.3. Multi-objective optimization
Figure B.4: Pareto front for two objective functions.
Nondominated sorting genetic algorithm (NSGA-II)

Recently, a number of multi-objective genetic algorithms based on the Pareto optimal concept has been proposed. The well known nondominated sorting genetic algorithm (NSGA-II) proposed by Deb et al. [34] is one of the most widely used multi-objective genetic algorithm since it provides excellent results as compared with other multi-objective genetic algorithms proposed [25]. A brief description of processes of NSGA-II is presented below, Fig. B.5.
Initially, a parent population of size n is generated randomly. All individuals in this population are sorted into different front levels based on the domination of pair comparison. Each front level is assigned a tness (or a rank) which equals its non-domination level. Level 1 is the top level in which the individual is dominated by none of the other individuals; level 2 is the secondary level in which the individual is dominated by some individuals only in level 1, and so on. In the same front level, the location of the nite number of solutions is expected to be distributed uniformly. In other words, a large diversity of the individuals can prevent the results from sticking into a local optimum. Therefore, another feature, called crowding distance, is adopted to evaluate the local aggregation of individuals. The 273

denition of crowding distance is [34, 181]:
Nobj
Cj =
i=1
(Fij )max (Fij )min
Fij +1 Fij 1
(B.5)
where Cj is the crowding distance of point j on the Pareto front, (Fij ) is the value of the tness function i at point j (i = 1, 2 for two objective functions). For boundary points, the crowding distance is set to the maximum value of the system in order to ensure that these points can survive to the next generation. In the same rank level, individuals who have larger crowding distance also have more opportunities to be selected [182].
Simulated binary crossover

In NSGA-II, a simulated binary crossover operator is used, which simulates the working principle of the single point crossover operator on binary 2,t ,t strings [182]. Let x1 i and xi denote two parent individuals for ith selection in generation t. The procedure of computing the children individuals ,t+1 ,t+1 ,t 2,t x1 and x2 from parent individuals x1 i i i and xi is described below. A spread factor i is dened as the ratio of the absolute difference in children values to that of parent values: i =
,t+1 ,t+1 x1 x2 i i ,t 1,t x2 i xi
(B.6)
Firstly, a random number ui [0, 1] is generated, whereafter, from a specied probability distribution function, the ordinate qi is found so that the area under the probability curve from 0 to qi is equal to the chosen random number ui . The probability distribution used to create a child individual is derived to have a similar search power as that in a single-point crossover in binary-coded GAs and is given as follows [182]: 0.5( + 1) , if i 1 i (B.7) P (i ) = 0.5( + 1)/ +2 , otherwise
i
where, is the distribution index which can be any nonnegative real number. A lager value of gives a higher probability for creating near parent individuals and a small value of allows distant individuals to be selected as children individuals. Using Eq.B.7, the qi can be calculated as follows [182]: 1 (2ui ) +1 , if ui 0.5 1 (B.8) qi = +1 1 , otherwise 2(1ui ) 274
B.3. Multi-objective optimization
Thereafter, the children individuals can be obtained by [182]:

,t+1 x1 i ,t+1 x2 i
= =
,t 2,t 0.5 (1 + qi )x1 i + (1 qi )xi
,t 2,t 0.5 (1 + qi )x1 i + (1 qi )xi
(B.9)
Note that two children individuals are symmetric about the parent individuals. This is deliberately used to avoid any bias towards any particular parent individual in a single crossover operation.
Mutation Operator
Let xk be the component of an individual {xi }, which is going to be mul tated. xu k and xk stand for the maximum and minimum value of this component in all individuals, respectively. The mutated individual yk can be calculated as follows [182]:
l yk = xk + q xu k xk
(B.10)
otherwise (B.11) where m is the mutation index which is set to 20 (1- crossover fraction) in general, and r is a random number. The intermediate variables 1 and 2 stand for: 1 2 = = xk xl k l xu x k k u xk xk l xu k xk (B.12) (B.13)
Here q is a mutation parameter, which stands for: 2r + (1 2r)(1 )m +1 1/(m +1) 1, 1 q = 1 2r(1 r) + 2(r 0.5)(1 2 )m +1 1/(m +1)
if r 0.5
275
Figure B.5: Flow chart for the NSGA-II process
276
List of Publications
Journal Articles
Published
1. K. Elsayed, C. Lacor. Modeling and Pareto Optimization of Gas Cyclone Separator Performance Using RBF Type Articial Neural Networks and Genetic Algorithms. Powder Technology. In Press, Accepted Manuscript, Available online 17 October 2011 doi: http://dx.doi.org/10.1016/j.powtec.2011.10.015 2. K. Elsayed, C. Lacor. Numerical and Empirical Modeling of the Flow Field and Performance in Cyclones of Different Cone-Tip Diameters. Computers & Fluids, Vol. 51, No. 1, pp. 48 - 59, 2011. doi: http://dx.doi.org/10.1016/j.compfluid.2011.07.010 3. K. Elsayed, C. Lacor. Modeling, Analysis and Optimization of Aircyclones Using Articial Neural Network, Response Surface Methodology and CFD Simulation Approaches. Powder Technology, Vol. 212, No. 1, pp. 115 - 133, 2011. doi: http://dx.doi.org/10.1016/j.powtec.2011.05.002 4. K. Elsayed, C. Lacor. The Effect of Cyclone Inlet Dimensions on the Flow Pattern and Performance. Applied Mathematical Modelling, Vol. 35, No. 4, pp. 1952 - 1968, 2011. doi: http://dx.doi.org/10.1016/j.apm.2010.11.007 5. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geometry for Minimum Pressure Drop Using Mathematical Models and CFD Simulations.Chemical Engineering Science, Vol. 65, No. 22, pp. 6048 - 6058, 2010. doi: http://dx.doi.org/10.1016/j.ces.2010.08.042 (listed in the Top 25 Hottest Articles Chemical Engineering > Chemical Engineering Science, October to December 2010).
Accepted with revision

1. K. Elsayed, C. Lacor. The Effect of Cyclone Vortex Finder Dimensions on the Flow Pattern and Performance using LES. Computers & Fluids, 2011. 277
With reviewers
1. K. Elsayed, C. Lacor. CFD Modeling and Multi-Objective Optimization of Cyclone Geometry Using Desirability Function, Articial Neural Networks and Genetic Algorithms. Applied Mathematical Modelling. Submitted, 2011. 2. K. Elsayed, C. Lacor. A CFD Study of the Effect of the Dust Outlet Geometry on the Performance and Hydrodynamics of Gas Cyclones. Computers & Fluids. Submitted, 2011. 3. K. Elsayed, C. Lacor. Numerical Study of the Effect of Cyclone Cone & Barrel Height On The Flow Pattern And Performance. Applied Mathematical Modelling. Submitted, 2011.
International conference proceedings

1. K. Elsayed, C. Lacor. Single and Multi-Objective Optimization of the Cyclone Separator Geometry Using Articial Neural Network and Genetic Algorithm, EUROGEN 2011, Evolutionary and Deterministic Methods for Design, Optimization and Control with Applications to Industrial and Societal Problems, Capua, Italy, 14-16 September 2011. 2. K. Elsayed, C. Lacor. The Effect of Dust Outlet Shape on the Flow Pattern and Performance of Cyclone Separators, 10th International Symposium on Experimental and Computational Aerothermodynamics of Internal Flows (ISAIF10), Brussel, Belgium, 4-7 July 2011. 3. K. Elsayed, C. Lacor. The Effect of the Cyclone Separator Cone Height on the Performance Using Articial Neural Network Model and CFD Simulations, 10th International Symposium on Experimental and Computational Aerothermodynamics of Internal Flows (ISAIF10), Brussel, Belgium, 4-7 July 2011. 4. K. Elsayed, C. Lacor. A CFD Study of the Effect of Cyclone Barrel Height on Its Performance Parameters, 8th International Conference on Computational Fluid Dynamics in the Oil & Gas, Metallurgical and Process Industries, Trondheim, Norway, 21-23 June 2011. 5. K. Elsayed, C. Lacor. Numerical Investigations of the Effect of Different Dust Outlet Designs on the Cyclone Performance and Flow Pattern, 14th International Conference on Aerospace Sciences and Aviation Technology (ASAT-14), Cairo, Egypt, 24-26 May 2011. 6. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geometry for Minimum Pressure Drop Based on Articial Neural Network Model And CFD Simulation, ECCOMAS thematic conference, CFD & Optimization, Antalya, Turkey, 23-25 May 2011. 278
7. K. Elsayed, C. Lacor. Multi-Objective Optimization of Gas Cyclone Based On CFD Simulation, ECCOMAS thematic conference, CFD & Optimization, Antalya, Turkey, 23-25 May 2011. 8. K. Elsayed, C. Lacor. The effect of cyclone height on the ow pattern and performance using LES, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3003, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 9. K. Elsayed, C. Lacor. The effect of cyclone inlet width on the ow pattern and performance, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3085, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 10. K. Elsayed, C. Lacor. Numerical study on the effect of cyclone inlet height on the ow pattern and performance, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG3068, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 11. K. Elsayed, C. Lacor. The effect of cyclone dustbin on the ow pattern and performance, Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3092, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 12. K. Elsayed, C. Lacor. Application of Response Surface Methodology for Modeling and Optimization of the Cyclone Separator for Minimum Pressure Drop, Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June 2010. 13. K. Elsayed, C. Lacor. The Effect of Vortex Finder Diameter on Cyclone Separator Performance and Flow Field, Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June 2010. 14. V. Agnihotri, K. Elsayed, C. Lacor, S. Verbanc. Numerical Study of Particle Deposition in the Human Upper Airways With Emphasis on Hot Spot Formation and Comparison Of LES and RANS Models, Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD10), Lisbon, Portugal, 14-17 June 2010. 15. K. Elsayed, C. Lacor. Modeling of the Gas and Particle Flow in the Cyclone Separator Using LES, RANS and Mathematical Models, 14th International Conference on Applied Mechanics and Mechanical Engineering (AMME-14), Military Technical College, Cairo, Egypt, 2527 May 2010. 16. K. Elsayed, C. Lacor. Optimization of the Cyclone Separator Geometry Based On CFD Simulation, ERCOFTAC day, Vrije Universiteit Brussel, Brussels, Belgium, 3rd December 2009. 279
17. K. Elsayed, C. Lacor. A CFD Study of the Effects of Cone Dimensions on the Flow Field of Cyclone Separators Using LES, 13th International Conference on Aerospace Sciences & Aviation Technology (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009. 18. K. Elsayed, C. Lacor. Investigation of the Geometrical Parameters Effects on the Performance and the Flow-Field of Cyclone Separators Using Mathematical Models and Large Eddy Simulation, 13th International Conference on Aerospace Sciences & Aviation Technology (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009.
National Conference proceedings

1. K. Elsayed, C. Lacor. Study Of the Effects of Geometrical Parameters on the Performance of Cyclone Separators, 8th National Congress on Theoretical and Applied Mechanics (NCTAM2009), Vrije Universiteit Brussel, Brussels, Belgium, 28-29 May 2009. 2. K. Elsayed, C. Lacor. Effects of Geometry Parameters on the Flow Field of Cyclone Separator, Poster day, the GRAduate School in MECHanics (GRASMECH 2008), Royal Military Academy, Brussels, Belgium, 3rd October 2008.
Technical reports
1. K. Elsayed, S. Jayaraju, C. Lacor. Wood Pellet Transport in the Biomass Burner of the Rodenhuize Power Plant, Research report, Laborelec Company, Belgium, April 2009. 2. K. Elsayed, C. Lacor. The State of the Art for Flow in Cyclone Separator, Internal report, Vrije Universiteit Brussel, Brussels, Belgium, April 2008.
280
Bibliography
[1] R. M. Alexander. Fundamentals of cyclone design and operation. In Proceedings of the Australian Institute of Mineral and Metallurgy, number 152, pages 203228, 1949. [2] M. S. M. Ali, C. J. Doolan, and V. Wheatley. Grid convergence study for a two-dimensional simulation of ow around a square cylinder at a low reynolds number. In Seventh International Conference on CFD in the Minerals and Process Industries, CSIRO, Melbourne, Australia, 9-11 December 2009. [3] S. Altmeyer, V. Mathieu, S. Jullemier, P. Contal, N. Midoux, S. Rode, and J. P. Leclerc. Comparison of different models of cyclone prediction performance for various operating conditions using a general software. Chemical Engineering and Processing, 43:511522, 2004. [4] N. Amoura, H. Kebir, S. Rechak, and J.M. Roelandt. Axisymmetric and two-dimensional crack identication using boundary elements and coupled quasi-random downhill simplex algorithms. Engineering Analysis with Boundary Elements, 34(6):611618, 2010. [5] A. Avci and I. Karagoz. Theoretical investigation of pressure losses in cyclone separators. International Communications in Heat and Mass Transfer, 28(1):107117, 1 2001. [6] Mehdi Azadi, Mohsen Azadi, and Ali Mohebbi. A CFD study of the effect of cyclone size on its performance parameters. Journal of Hazardous Materials, 182(1-3):835 841, 2010. [7] M. I. Bakari and F. Hamdullahpur. Optimization of the overall performance of cyclone separators. In 2nd Trabzon International Energy and Environment Symposium, 26-29 July, Turkey, 1998. [8] M. Balcilar, A.S. Dalkilic, and S. Wongwises. Articial neural network techniques for the determination of condensation heat transfer characteristics during downward annular ow of R134a inside a vertical smooth tube. International Communications in Heat and Mass Transfer, 38(1):75 84, 2011. 281
[9] W. Barth. Design and layout of the cyclone separator on the basis of new investigations. Brennstow-Waerme-Kraft (BWK), 8(4):19, 1956. [10] A. C. Benim, K. Ozkan, M. Cagan, and D. Gunes. Computational investigation of turbulent jet impinging onto rotating disk. International Journal of Numerical Methods for Heat & Fluid Flow, 17(3):284 301, 2007. [11] S. Bernardo, M. Mori, A.P. Peres, and R.P. Dionisio. 3-D computational uid dynamics for gas and gas-particle ows in a cyclone with different inlet section angles. Powder Technology, 162(3):190 200, 2006. [12] J. L. Bernon, V. Boudousq, J. F. Rohmer, M. Fourcade, M. Zanca, M. Rossi, and D. Mariano-Goulart. A comparative study of powells and downhill simplex algorithms for a fast multimodal surface matching in brain imaging. Computerized Medical Imaging and Graphics, 25(4):287297, 2001. [13] G. E. P. Box and K. B. Wilson. On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society, 13:1 45, 1951. [14] F. Boysan, W.H. Ayer, and J. A. Swithenbank. Fundamental mathematical-modelling approach to cyclone design. Transaction of Institute Chemical Engineers, 60:222230, 1982. [15] F. Boysan, B. C. R. Ewan, J. Swithenbank, and W. H. Ayers. Experimental and theoretical studies of cyclone separator aerodynamics. IChemE Symp Series, 69:305 320, 1983. [16] D. S. Broomhead and D. Lowe. Multivariable functional interpolation and adaptive networks. Complex Systems, 2:321355, 1988. [17] H. S. Bryant, R.W. Silverman, and F. A. Zenz. How dust in gas affects cyclone pressure drop. Hydrocarbon Processing, 62(4):8790, 1983. [18] John A. Bullinaria. Introduction to neural networks, 2004. http://www.cs.bham.ac.uk/jxb/NN/l10.pdf. [19] H. Buttner. Size separation of particles from aerosol samples using impactors and cyclones. Particle & Particle Systems Characterization, 5(2):8793, 2004. [20] W. M. Carlyle, D. C. Montgomery, and G. C. Runger. Optimization problems and methods in quality control and improvement. Journal of Quality Technology, 32(1):117, 2000. [21] J. Casal and J. M. Martinez-Benet. A better way to calculate cyclone pressure drop. Chemical Engineering, 90(2):99100, 1983. 282
[22] J. Chen and M. Shi. A universal model to calculate cyclone pressure drop. Powder Technology, 171(3):184 191, 2007. [23] T. G. Chuah, J. Gimbun, and T. S.Y. Choong. A CFD study of the effect of cone dimensions on sampling aerocyclones performance and hydrodynamics. Powder Technology, 162:126 132, 2006. [24] T. J. Chung. Computational Fluid Dynamics. Cambridge University Press, 1st edition, 2002. [25] C. A. C. Coello. Twenty years of evolutionary multi-objective optimization: a historical view of the eld. IEEE Computational Intelligence Magazine, 1:2836, 2006. [26] C. A. C. Coello, G. B. Lamont, and D. A. Van Veldhuizen. Evolutionary Algorithms for Solving Multi-Objective Problems. Springer, 2nd edition, 2007. [27] D. Cooley. Using optimtool GA in Matlab, 2011. digital.cs.usu.edu/cooley/CS6665_10_Optimtool_GA.pptx. [28] G. Corriveau, R. Guilbault, and A. Tahan. Genetic algorithms and nite element coupling for mechanical optimization. Advances in Engineering Software, 41(3):422 426, 2010. [29] C. Cort es and A. Gil. Modeling the gas and particle ow inside cyclone separators. Progress in Energy and Combustion Science, 33(5):409 452, 2007. [30] J. S. Cowpe, J. S. Astin, R. D. Pilkington, and A. E. Hill. Application of response surface methodology to laser-induced breakdown spectroscopy: inuences of hardware conguration. Spectrochimica Acta Part B, 62:1335 1342, 2007. [31] J. S. Curtis and B. V. Wachem. Modeling particle-laden ows: A research outlook. Wiley InterScience, 50-11, 2004. [32] L. Davidson. Lecture notes. Dept. of Thermo and Fluid Dynamics, Chalmers University of Technology, G oteborg, Sweden, 2000. [33] J. W. Deardorff. The use of subgrid transport equations in a threedimensional model of atmospheric turbulence. Journal of Fluids Engineering, 95:429438, 1973. [34] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transaction Evolitionery Computation, 6:182197, 2002. [35] W. Deconinck. Design and application of discrete explicit lters for large eddy simulation of compressible turbulent ows. Masters thesis, University of Toronto, 2008. 283
[36] J. J. Derksen. Separation performance predictions of a stairmand high-efciency cyclone. AIChE Journal, 49(6):13591371, 2003. [37] J. J. Derksen, S. Sundaresan, and H. E. A. van den Akker. Simulation of mass-loading effects in gassolid cyclone separators. Powder Technology, 163:5968, 2006. [38] J. J. Derksen and H. E. A. Van den Akker. Simulation of vortex core precession in a reverse-ow cyclone. AIChE Journal, 46(7):1317 1331, 2000. [39] J. J. Derksen, H. E. A. van den Akker, and S. Sundaresan. Twoway coupled large-eddy simulations of the gas-solid ow in cyclone separators. AIChE Journal, 54(4):872885, April 2008. [40] D. B. Dias, M. Mori, and W. P. Martignoni. Boundary condition effects in CFD cyclone simulations. In 8th World Congress of Chemical Engineering (WCCE8), Montreal, August 2009. [41] K. K. J. R. Dinesh and M. P. Kirkpatrick. Study of jet precession, recirculation and vortex breakdown in turbulent swirling jets using LES. Computers & Fluids, 38:12321242, 2009. [42] J. Dirgo. Relationship between cyclone dimensions and performance. PhD thesis, Harvard University, USA, 1988. [43] J. Dirgo and D. Leith. Cyclone collection efciency: Comparison of experimental results with theoretical predictions. Aerosol Science and Technology, 4:401415, 1985. [44] S. E. Elghobashi. On predicting particle-laden turbulent ows. Applied Scientic Research, 52:309329, 1994. [45] K. Elsayed and C. Lacor. Modeling and pareto optimization of gas cyclone separator performance using RBF type articial neural networks and genetic algorithms. Powder Technology, Accepted 9 October, 2011. http://dx.doi.org/10.1016/j.powtec.2011.10.015. [46] K. Elsayed and C. Lacor. Investigation of the geometrical parameters effects on the performance and the ow-eld of cyclone separators using mathematical models and large eddy simulation. In 13th Aerospace Sciences & Aviation Technology (ASAT-13), Military Technical College, Cairo, Egypt, 26-28 May 2009. [47] K. Elsayed and C. Lacor. The effect of cyclone dustbin on the ow pattern and performance. In Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3092, 2010, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. 284
[48] K. Elsayed and C. Lacor. The effect of cyclone height on the ow pattern and performance using LES. In Tenth International Congress of Fluid Dynamics (ICFD10), ASME, Egypt, ICFD10-EG-3003, Ain Soukhna, Red Sea, Egypt, 16-19 December 2010. [49] K. Elsayed and C. Lacor. The effect of vortex nder diameter on cyclone separator performance and ow eld. In Fifth European Conference on Computational Fluid Dynamics (ECCOMAS CFD 2010), Lisbon, Portugal, 14-17 June 2010. [50] K. Elsayed and C. Lacor. Optimization of the cyclone separator geometry for minimum pressure drop using mathematical models and CFD simulations. Chemical Engineering Science, 65(22):60486058, 2010. [51] K. Elsayed and C. Lacor. A CFD study of the effect of cyclone barrel height on its performance parameters. In 8th International Conference on Computational Fluid Dynamics in the Oil & Gas, Metallurgical and Process Industries, Trondheim, Norway, 21-23 June 2011. [52] K. Elsayed and C. Lacor. The effect of cyclone inlet dimensions on the ow pattern and performance. Applied Mathematical Modelling, 35(4):19521968, 2011. [53] K. Elsayed and C. Lacor. Modeling, analysis and optimization of aircyclones using articial neural network, response surface methodology and CFD simulation approaches. Powder Technology, 212(1):115133, 2011. [54] K. Elsayed and C. Lacor. Multi-objective optimization of gas cyclone based on CFD simulation. In ECCOMAS thematic conference, CFD & Optimization, Antalya, Turkey, 23-25 May 2011. [55] K. Elsayed and C. Lacor. Numerical modeling of the ow eld and performance in cyclones of different cone-tip diameters. Computers & Fluids, 51(1):4859, 2011. [56] K. Elsayed and C. Lacor. Single and multi-objective optimization of the cyclone separator geometry using articial neural network and genetic algorithm. In EUROGEN 2011, Evolutionary and Deterministic Methods for Design, Optimization and Control with Applications to Industrial and Societal Problems, Capua, Italy, 14-16 September, 2011. [57] Ferit Ficici, Vedat Ari, and Murat Kapsiz. The effects of vortex nder on the pressure drop in cyclone separators. International Journal of the Physical Sciences, 5(6):804813, 2010. [58] M. W. First. Cyclone dust collector design. ASME Annual General Meeting, Paper No. 49A127 (1949). 285
[59] Fluent. FLUENT 6.3 users guide. Fluent Incorporated, Lebanon, 2006. [60] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamic subgrid-scale eddy viscosity model. Physics of Fluids A, 3(7):1760 1765, 1991. [61] M. Gfrerer and E. Lankmayr. Screening, optimization and validation of microwave-assisted extraction for the determination of persistent organochlorine pesticides. Analytica Chimica Acta, 533:203 211, 2005. [62] J. Gimbun, T. G. Chuah, T. S. Y. Choong, and A. Fakhrul-Razi. A CFD study on the prediction of cyclone collection efciency. International Journal for Computational Methods in Engineering Science and Mechanics, 6(3):161 168, 2005. [63] J. Gimbun, T.G. Chuah, T.S. Choong, and Y.A. Fakhrul-Razi. Prediction of the effects of cone tip diameter on the cyclone performance. Aerosol Science and Technology, 36:10561065, 2005. [64] J. Gimbun, T.G. Chuah, A. Fakhrul-Razi, and T. S. Y. Choong. The inuence of temperature and inlet velocity on cyclone pressure drop: a CFD study. Chemical Engineering & Processing, 44(1):712, 2005. [65] F. Girosi, T. Poggio, and A. Caprile. Extensions of a theory of networks for approximations and learning: outliers and negative examples. In NIPS-3: Proceedings of the 1990 conference on Advances in neural information processing systems 3, pages 750756, Denver, Colorado, United States, 1990. [66] A. Gorton-Hulgerth, J. Woisetschlager, Graham Wigley, and G. Staudinger. Investigation of the ow eld in the upper part of a cyclone with laser and phase doppler anemometry. Particle & Particle Systems Characterization, 17:2127, 2000. [67] A. D. Gosman and E. Ioannides. Aspects of computer simulation of liquid-fuelled combustor. AIAA Journal, 81-0323, 1981. [68] W. D. Grifths and F. Boysan. Computational uid dynamics (CFD) and empirical modelling of the performance of a number of cyclone samplers. Journal of Aerosol Science, 27(2):281304, 1996. [69] G. Gronald and J. J. Derksen. Simulating turbulent swirling ow in a gas cyclone: A comparison of various modeling approaches. Powder Technology, 205(1-3):160 171, 2011. [70] P. O. Haaland. Experimental design in biotechnology. Marcel Dekker, New York, 1989. 286
[71] E. C. Jr. Harrington. The desirability function. Industrial Quality Control, 21:494498, 1965. [72] Eric Hartman, James D. Keeler, and Jacek M. Kowalski. Layered neural networks with gaussian hidden units as universal approximations. Neural Computation, 2(2):210215, 1990. [73] J. O. Hinze. Turbulence. McGraw-Hill, New York, 2nd ed. edition, 1975. [74] A. J. Hoekstra. Gas ow eld and collection efciency of cyclone separators. PhD thesis, Technical University Delft, 2000. [75] A. J. Hoekstra, J. J. Derksen, and H. E. A. Van Den Akker. An experimental and numerical study of turbulent swirling ow in gas cyclones. Chemical Engineering Science, 54:20552065, 1999. [76] A. C. Hoffmann, M. de Groot, W. Peng, H. W. A. Dries, and J. Kater. Advantages and risks in increasing cyclone separator length. AIChE Journal, 47(11):24522460, 2001. [77] A. C. Hoffmann and L. E. Stein. Gas cyclones and swirl tubes: Principle, Design and Operation. Springer, 2nd edition, 2008. [78] A.C. Hoffmann, M. De Groot, and A. Hospers. The effect of the dust collection system on the ow pattern and separation efciency of a gas cyclone. The Canadian Journal of Chemical Engineering, 74:464470, August 1996. [79] A. Horvath, C. Jordan, and M. Harasek. Inuence of vortex-nder diameter on axial gas ow in simple cyclone. Chemical Product and Process Modeling, 3(1), 2008. [80] Y. H. Hu and J. Hwang, editors. Handbook of Neural Network Signal Processing. CRC Press, Washington, D.C., 2002. [81] S. Huang, Q. S. Li, and S. Xu. Numerical evaluation of wind effects on a tall steel building by CFD. Journal of Constructional Steel Research, 63(5):612 627, 2007. [82] S. H. Huang, Q. S. Li, and J. R. Wu. A general inow turbulence generator for large eddy simulation. Journal of Wind Engineering and Industrial Aerodynamics, 98(10-11):600 617, 2010. [83] P. Hutchinson, G. F. Hewitt, and A. E. Dukler. Deposition of liquid or solid dispersions from turbulent gas streams: a stochastic model. Chemical Engineering Science, 26:419439, 1971. [84] D. L. Iozia and D. Leith. Effect of cyclone dimensions on gas ow pattern and collection efciency. Aerosol Science and Technology, 10(3):491500, 1989. 287
[85] D. L. Iozia and D. Leith. The logistic function and cyclone fractional efciency. Aerosol Science and Technology, 12(3):598606, 1990. [86] S. T. Jayaraju. Study of the air ow and aerosol transport in the human upper airway using LES and DES methodology. PhD thesis, Vrije Universiteit Brussel, 2009. [87] Zhang Jian and Jin You-hai. CFD analysis on collection performance of swirl tube with different vortex nder geometries. Journal of System Simulation, 21(4):121417, 2009. [88] Viviana P. Jofre, Mariela V. Assof, Martn L. Fanzone, Hctor C. Goicoechea, Luis D. Martnez, and Mara F. Silva. Optimization of ultrasound assisted-emulsication-dispersive liquid-liquid microextraction by experimental design methodologies for the determination of sulfur compounds in wines by gas chromatography-mass spectrometry. Analytica Chimica Acta, 683(1):126 135, 2010. [89] W. P. Jones and M. Wille. Large-eddy simulation of a plane jet in a cross-ow. International Journal of Heat and Fluid Flow, 17-3:296 306, 1995. [90] I. Karagoz and A. Avci. Modelling of the pressure drop in tangential inlet cyclone separators. Aerosol Science and Technology, 39(9):857 865, 2005. [91] F. Kaya and I. Karagoz. Performance analysis of numerical schemes in highly swirling turbulent ows in cyclones. Current Science, 94(10):1273 1278, 2008. [92] F. Kaya and I. Karagoz. Numerical investigation of performance characteristics of a cyclone prolonged with a dipleg. Chemical Engineering Journal, 151:3945, 2009. [93] L. C. Kenny and R. A. Gussman. A direct approach to the design of cyclones for aerosol-monitoring applications. Journal of Aerosol Science, 31(12):14071420, 2000. [94] Abolfazl Khalkhalia and Hamed Sakhania. Pareto based objective optimization of a cyclone vortex nder using GMDH type neural networks and genetic algorithms. neering Optimization, , Available online: 05 Jul 2011, http://dx.doi.org/10.1080/0305215X.2011.564619. multiCFD, Engi2011.
[95] J. C. Kim and K. W. Lee. Experimental study of particle collection by small cyclones. Aerosol Science and Technology, 12:10031015, 1990. [96] J. Kiusalaas. Numerical Methods in Engineering With MATLAB. Cambridge University Press, Cambridge, UK, 2nd edition, 2010. 288
[97] C. Lacor. Lecture notes. Dept. Fluid Mechanics, Vrije Universiteit Brussel, Brussels, Belgium, 2007. [98] B. E. Launder, G. J. Reece, and W. Rodi. Progress in the development of a Reynolds-stress turbulence closure. Journal of Fluid Mechanics, 68(03):537566, 1975. [99] D. Leith and W. Licht. The collection efciency of cyclone type particle collectors: a new theoretical approach. A.I.Ch.E. Symposium Series, 68:196206, 1972. [100] A. Leonard. Energy cascade in LES of turbulent uid ows. Advances in Geophysics, 18:237248, 1974. [101] D. K. Lilly. A proposed modication of the germano subgrid-scale closure method. Physics of Fluids A, 4:633635, 1992. [102] K. S. Lim, H. S. Kim, and K. W. Lee. Characteristics of the collection efciency for a cyclone with different vortex nder shapes. Journal of Aerosol Science, 35(6):743754, 2004. [103] K. S. Lim, S. B. Kwon, and K. W. Lee. Characteristics of the collection efciency for a double inlet cyclone with clean air. Journal of Aerosol Science, 34(8):10851095, 2003. [104] A. J. Linden. Investigations into cyclone dust collectors. Proceedings of the Institution of Mechanical Engineers Part B Engineering Manufacture (Professional Engineering Publishing), 160:233251, 1949. [105] Z. Liu, Y. Zheng, L. Jia, J. Jiao, and Q. Zhang. Stereoscopic PIV studies on the swirling ow structure in a gas cyclone. Chemical Engineering Science, 61:4252 4261, 2006. [106] C. Liyana-Pathirana and F. Shahidi. Optimization of extraction of phenolic compounds from wheat using response surface methodology. Food Chemistry, 93:4756, 2005. [107] P. Worth Longest and Samir Vinchurkar. Effects of mesh style and grid convergence on particle deposition in bifurcating airway models with comparisons to experimental data. Medical Engineering & Physics, 29(3):350 366, 2007. [108] J. H. Mathews and K. D. Fink. Numerical Methods Using MATLAB. Prentice Hall, 3rd edition, 1999. [109] Matlab. Global Optimization Toolbox Users Guide, Matlab R2010a. The MathWorks, Inc., 2010. [110] P. Moin and J. Kim. Numerical investigation of turbulent channel ow. Journal of Fluid Mechanics, 155:441, 1982. 289
[111] M. E. Moore and A. R. McFarland. Performance modeling of singleinlet aerosol sampling cyclones. Environmental Science and Technology, 27(9):18421848, September 1993. [112] S. A. Morsi and A. J. Alexander. An investigation of particle trajectories in two-phase ow systems. Journal of Fluid Mechanics, 55(02):193208, 1972. [113] S. Murakami. Overview of turbulence models applied in CWE-1997. Journal of Wind Engineering and Industrial Aerodynamics, 74-76:1 24, 1998. [114] E. Muschelknautz. Die berechnung von zyklonabscheidern fur gas. Chemie Ingenieur Technik, 44:6371, 1972. [115] E. Muschelknautz and W. Kambrock. Aerodynamische beiwerte des zyclon abscheiders aufgrund neuer und verbesserter messungen. Chem-Ing-Tech, 42(5):24755, Chemie Ingenieur Technik 1970. [116] E. Muschelknautz and M. Trefz. Design and calculation of higher and design and calculation of higher and highest loaded gas cyclones. in Proceedings of Second World Congress on Particle Technology Kyoto, Japan, pages 5271, October 1990. [117] E. Muschelknautz and M. Trefz. VDI-Warmeatlas. Druckverlust und Abscheidegrad in Zyklonen, 6 Auage:Lj1 Lj9, 1991. [118] Raymond H. Myers, Douglas C. Montgomery, and Christine M. Anderson-Cook. Response Surface Methodology: Process and Product Optimization Using Designed Experiments. John Wiley & Sons, Inc., 3rd edition, 2009. [119] J. A. Nelder and R. Mead. A simplex method for function minimization. The Computer Journal (1965), 7(4). [120] Antonios Niros and George Tsekouras. A fuzzy clustering algorithm to estimate the parameters of radial basis functions neural networks and its application to system modeling. In John Darzentas, George Vouros, Spyros Vosinakis, and Argyris Arnellos, editors, Articial Intelligence: Theories, Models and Applications, volume 5138 of Lecture Notes in Computer Science, pages 194204. Springer Berlin / Heidelberg, 2008. [121] Stefan Obermair, Christoph Gutschi, Jakob Woisetschlager, and Gernot Staudinger. Flow pattern and agglomeration in the dust outlet of a gas cyclone investigated by phase doppler anemometry. Powder Technology, 156:3442, 2005. [122] Stefan Obermair and Gernot Staudinger. The dust outlet of a gas cyclone and its effects on separation efciency. Chemical Engineering Technology, WILEY-VCH Verlag GmbH, 24(12):12591263, 2001. 290
[123] Stefan Obermair, Jakob Woisetschlager, and Gernot Staudinger. Investigation of the ow pattern in different dust outlet geometries of a gas cyclone by laser doppler anemometry. Powder Technology, 138:239251, 2003. [124] T. ODoherty, A. J. Grifths, N. Syred, P. J. Bowen, and W. Fick. Experimental analysis of rotating instabilities in swirling and cyclonic ows. Developments in Chemical Engineering and Mineral Processing, 7:245267, 1999. [125] T. J. Overcamp and S. E. Scarlett. Effect of Reynolds number on the Stokes number of cyclones. Aerosol Science and Technology, 19(3):362370, 1993. [126] S. H. R. Pasandideh and S. T. A. Niaki. Multi-response simulation optimization using genetic algorithm within desirability function framework. Applied Mathematics and Computation, 175(1):366 382, 2006. [127] W. Peng, A. C. Hoffmann, P. J. A. J. Boot, A. Udding, H.W.A. Dries, A. Ekker, and J. Kater. Flow pattern in reverse-ow centrifugal separators. Powder Technology, 127:212 222, 2002. [128] R. H. Perry and D. W. Green, editors. Perrys Chemical Engineers Handbook. McGraw-Hill, 1997. [129] U. Piomelli, Y. Yu, and R. J. Adrain. Sub-grid scale energy transfer and near-wall turbulence structure. Physics of Fluids, 8:215 224, 1996. [130] S. I. Pishbin and M. Moghiman. Optimization of cyclone separators using genetic algorithm. International Review of Chemical Engineering (I.RE.CH.E.), 2(6):683690, 2010. [131] N. W. Polhemus. How to perform an optimization experiment using statgraphics centurion, 2005. http://www.sigmaplus.fr/telecharg/StatG/howto9.pdf. [132] S. B. Pope. Turbulent ows. Cambridge university press, New York, 2000. [133] L. Prandtl. Bericht uber die entstehung der turbulenz. Zeitschrift Angewandte Mathematik und Mechanik, 5:136139, 1925. fur [134] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical Recipes in FORTRAN 77: The Art of Scientic Computing. Cambridge University Press, Cambridge, UK, 1992. [135] William H. Press, Brian P. Flannery, Saul A. Teukolsky, and William T. Vetterling. Numerical Recipes in C The Art of Scientic Computing. Cambridge University Press, Cambridge, UK, 1992. 291
[136] F. Qian and Y. Wu. Effects of the inlet section angle on the separation performance of a cyclone. Chemical Engineering Research and Design, 87(12):15671572, 2009. [137] F. Qian and M. Zhang. Effects of the inlet section angle on the ow eld of a cyclone. Chemical Engineering & Technology, 30(11):1521 4125, 2007. [138] Fuping Qian, Jiguang Zhang, and Mingyao Zhang. Effects of the prolonged vertical tube on the separation performance of a cyclone. Journal of Hazardous Materials, B(136):822829, 2006. [139] G. Ramachandran, D. Leith, J. Dirgo, and H. Feldman. Cyclone optimization based on a new empirical model for pressure drop. Aerosol Science and Technology, 15:135148, 1991. [140] A. Raou, M. Shams, M. Farzaneh, and R. Ebrahimi. Numerical simulation and optimization of uid ow in cyclone vortex nder. Chemical Engineering and Processing, 47:128137, 2008. [141] G. Ravi, Santosh K. Gupta, and M. B. Ray. Multiobjective optimization of cyclone separators using genetic algorithm. Ind. Eng. Chem. Res., 39:42724286, 2000. [142] K. Rietema. Het mechanisme van de afscheiding van jnverdeelde stoffen in cyclonen (in dutch). De Ingenieur, 71(39):ch59ch65, 1959. [143] P. J. Roache. Perspective: A method for uniform reporting of grid renement studies. Journal of Fluids Engineering, 116(3):405413, 1994. [144] Patrick J. Roache. Verication of codes and calculations. AIAA Journal, 36(5):696702, 1998. [145] PJ Roache. Quantication of uncertainty in computational uid dynamics. Annual Review of Fluid Mechanics, 29:123160, 1997. [146] H. Sakhani, M. A. Akhavan-Behabadi, N. Nariman-Zadeh, and M. J. Mahmood Abadi. Modeling and multi-objective optimization of square cyclones using CFD and neural networks. Chemical Engineering Research and Design, 89(3):301 309, 2011. [147] H. Sakhani, M.A. Akhavan-Behabadi, M. Shams, and M. H. Rahimyan. Numerical simulation of ow eld in three types of standard cyclone separators. Advanced Powder Technology, 21(4):435 442, 2010. [148] H. Sakhani, A. Hajiloo, M.A. Ranjbar, and N. Nariman-Zadeh. Modeling and multi-objective optimization of cyclone separators using CFD and genetic algorithms. Computers & Chemical Engineering, 35(6):10641071, 2011. 292
[149] H. Sakhani, S.A. Nourbakhsh, and N. Nariman-zadeh. Modeling and multi-objective optimization of cyclone vortex nder using CFD and neural networks. In 2nd International Conference on Engineering Optimization, September 6-9, 2010, Lisbon, Portugal, pages 19, 2010. [150] P. Sagaut. Large-Eddy Simulation for Incompressible Flows. Scientic Computation, Springer, 2002. [151] B.E. Saltzman and J.M. Hochstrasser. Design and performance of miniature cyclone for respirable aerosol sampling. Environmental Science & Technology, 17(7):418424, July 1983. [152] S. Schmidt, H. M. Blackburn, M. Rudman, and I. Sutalo. Simulation of turbulent ow in a cyclonic separator. In Third international conference on CFD in the minerals and process industries, CSIRO, Melbourne, Australia, 10-12 December 2003. [153] St. Schmidt, H. M. Blackburn, and M. Rudman. Impact of outlet boundary conditions on the ow properties within cyclone. In 15th Australasian Fluid Mechanics Conference, The University of Sydney, Sydney, Australia, 2004. [154] H. Shalaby. On the Potential of Large Eddy Simulation to Simulate Cyclone Separators. PhD thesis, Chemnitz University of Technology, Chemnitz, Germany, 2007. [155] H. Shalaby, K. Pachler, K. Wozniak, and G. Wozniak. Comparative study of the continuous phase ow in a cyclone separator using different turbulence models. International Journal for Numerical Methods in Fluids, 48(11):11751197, 2005. [156] H. Shalaby, K. Wozniak, and G. Wozniak. Particle-laden ow simulation in a cyclone separator. WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim, 6(1):547548, 2006. [157] C. B. Shepherd and C. E. Lapple. Flow pattern and pressure drop in cyclone dust collectors cyclone without intel vane. Industrial & Engineering Chemistry, 32(9):12461248, September 1940. [158] Sujeet Kumar Shukla, Prashant Shukla, and Pradyumna Ghosh. Evaluation of numerical schemes using different simulation methods for the continuous phase modeling of cyclone separators. Advanced Powder Technology, 22(2):209219, 2010. [159] M. D. Slack, R. O. Prasad, A. Bakker, and F. Boysan. Advances in cyclone modeling using unstructured grids. Trans IChemE., 78 Part A, (2000), 2000. [160] John W. Slater. Examining spatial (grid) convergence, 2008. http://www.grc.nasa.gov/WWW/wind/valid/tutorial/spatconv.html . 293
[161] J. Smagorinsky. General circulation experiments with the primitive equations: The basic experiment. Monthly weather review, 91(3):99 164, 1963. [162] A. Smirnov, S. Shi, and I. Celik. Random ow generation technique for large eddy simulations and particle-dynamics modeling. Journal of Fluids Engineering, 123(2):359371, 2001. [163] G. Solero and A. Coghe. Experimental uid dynamic characterization of a cyclone chamber. Experimental Thermal and Fluid Science, 27:8796, 2002. [164] M. Sommerfeld, A. Ando, and D. Wennerberg. Swirling, particleladen ows through a pipe expansion. Journal of Fluids Engineering, 114:648656, 1992. [165] C. J. Stairmand. Pressure drops in cyclone separators. Industrial and Engineering Chemistry, 16(B):409411, 1949. [166] C. J. Stairmand. The design and performance of cyclone separators. Industrial and Engineering Chemistry, 29:356383, 1951. [167] G. Steenackers, R. Versluys, M. Runacres, and P. Guillaume. Reliability-based design optimization of computation-intensive models making use of response surface models. Quality and Reliability Engineering International, 27(4):555568, 2011. [168] Gunther Steenackers. Finite Element Model Updating and Optimization of Mechanical Systems Making Use of Regressive Techniques. PhD thesis, Vrije Universiteit Brussel, 2008. [169] Gunther Steenackers and Patrick Guillaume. Bias-specied robust design optimization: A generalized mean squared error approach. Computers & Industrial Engineering, 54(2):259 268, 2008. [170] Gunther Steenackers, Patrick Guillaume, and Steve Vanlanduit. Robust optimization of an airplane component taking into account the uncertainty of the design parameters. Quality and Reliability Engineering International, 25(3):255282, 2009. [171] A.C. Stern, K.J. Caplan, and P.D. Bush. Cyclone dust collectors. New York: American Petroleum Institute, 1955. [172] P. K. Swamee, N. Aggarwal, and K. Bhobhiya. Optimum design of cyclone separator. American Institute of Chemical Engineers (AIChE ), 55(9):22792283, 2009. [173] Dominique Thevenin and Gabor Janiga. Optimization and computational uid dynamics. Springer Verlag, Berlin, 2008. 294
[174] M. Trefz. Die vershiedenen abscheidevorgange im hoher un hoch beladenen gaszyklon unter besonderer berucksichtigung der sekundarstromung. Forschritt-Berichte VDI; VDI-Verlag GmbH: Dusseldorf, Germany (1992), 295. [175] M. Trefz and E. Muschelknautz. Extended cyclone theory for gas ows with high solids concentrations. Chemie Ingenieur Technik, 16:153160, 1993. [176] P. Venkataraman. Applied optimization with MATLAB programming. John Wiley and Sons, 2nd edition, 2009. [177] H.K. Versteeg and W. Malalasekera. Introduction to Computational Fluid Dynamics, the Finite Volume Method. Longman Scientic & Technical, 1995. [178] B. Wang, D.L. Xu, K.W. Chu, and A.B. Yu. Numerical study of gassolid ow in a cyclone separator. Applied Mathematical Modelling, 30:13261342, 2006. [179] B. Wang, D.L. Xu, G. X. Xiao, K. W. Chu, and A. B. Yu. Numerical study of gas-solid ow in a cyclone separator. In Third International Conference on CFD in the Minerals and Process Industries, pages 371376, CSIRO, Melbourne, Australia, December 2003. [180] Lei Wang, Cheng Shao, Hai Wang, and Hong Wu. Radial basis function neural networks-based modeling of the membrane separation process: Hydrogen recovery from renery gases. Journal of Natural Gas Chemistry, 15:230234, 2006. [181] X.D. Wang, C. Hirsch, Sh. Kang, and C. Lacor. Multi-objective optimization of turbomachinery using improved NSGA-II and approximation model. Computer Methods in Applied Mechanics and Engineering, 200(9-12):883 895, 2011. [182] Xiaodong Wang. CFD Simulation of Complex Flows in Turbomachinery and Robust Optimization of Blade Design. PhD thesis, Vrije Universiteit Brussel, 2010. [183] D. C. Wilcox. Turbulence Modeling for CFD. DCW Industries, Inc, 2nd ed. edition, 1994. [184] R. Xiang, S. H. Park, and K. W. Lee. Effects of cone dimension on cyclone performance. Journal of Aerosol Science, 32(4):549561, 2001. [185] R. B. Xiang and K. W. Lee. Numerical simulation of ow patterns in cyclones of different cone dimensions. Part. Part. Syst. Charact., 22(3):212218, 2005. 295
[186] R. B. Xiang and K. W. Lee. Numerical study of ow eld in cyclones of different height. Chemical Engineering and Processing, 44:877 883, 2005. [187] H. Yoshida, Y. Inada, K. Fukui, and T. Yamamoto. Improvement of gas-cyclone performance by use of local uid ow control method. Powder Technology, 193:614, 2009. [188] Jin You-hai, Ji Guang-qin, Cao Qing-yun, and Wang Jian-jun. Numerical simulation of gas-phase ow eld in vortex nder of cyclone separators. Journal of China University of Petroleum, 32(6):10912, 2008. [189] Y. Yuan, Y. Gao, L. Mao, and J. Zhao. Optimisation of conditions for the preparation of -carotene nanoemulsions using response surface methodology. Food Chemistry, 107:13001306, 2008. [190] Aishe Zhang and Ling Zhang. RBF neural networks for the prediction of building interference effects. Computers & Structures, 82(27):23332339, 2004. [191] B. Zhang and S. Hui. Numerical simulation and PIV study of the turbulent ow in a cyclonic separator. In International Conference on Power Engineering, Hangzhou, China, October 2007. [192] B. Zhao. A theoretical approach to pressure drop across cyclone separators. Chemical Engineering Technology, 27:11051108, 2004. [193] B. Zhao. Experimental investigation of ow patterns in cyclones with conventional and symmetrical inlet geometries. Chemical Engineering & Technology, 28(9):969 972, 2005. [194] B. Zhao. Modeling pressure drop coefcient for cyclone separators a support vector machine approach. Chemical Engineering Science, 64:41314136, 2009. [195] B. Zhao, H. Shen, and Y. Kang. Development of a symmetrical spiral inlet to improve cyclone separator performance. Powder Technology, 145:4750, 2004. [196] B. Zhao and Y. Su. Particle collection theory for cyclone separators: summary and comparison. Particle & Particle Systems Characterization, 23:484488, 2006. [197] B. Zhao and Y. Su. Articial neural network-based modeling of pressure drop coefcient for cyclone separators. chemical engineering research and design, 88:606613, 2010. [198] B. Zhao, Y. Su, and J. Zhang. Simulation of gas ow pattern and separation efciency in cyclone with conventional single and spiral double inlet conguration. Chemical Engineering Research and Design, 84:11581165, 2006. 296
[199] Y. Zhu, M. C. Kim, K. W. Lee, Y. O. Park, and M. R. Kuhlman. Design and performance evaluation of a novel double cyclone. Aerosol Science and Technology, 34(8):373380, 2000. [200] Y. Zhu and K. W. Lee. Experimental study on small cyclones operating at high ow rates. Journal of Aerosol Science, 30:13031315, 1999.
297

PHD Thesis Khairy Elsayed

Diunggah oleh

Informasi Dokumen

Deskripsi Asli:

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

PHD Thesis Khairy Elsayed

Diunggah oleh

Hak Cipta:

Format Tersedia

FACULTY OF ENGINEERING Department of Mechanical Engineering

Vrije Universiteit Brussel October 2011

Advisor: Prof. Dr. Ir. Chris Lacor

Prof. dr. ir. Rik Pintelon Vrije Universiteit Brussel

Prof. dr. ir. Gunther Steenackers Vrije Universiteit Brussel

Prof. dr. ir. Gert Desmet Vrije Universiteit Brussel

Prof. dr. ir. Herman Deconinck Von Karman Institute

Prof. dr. ir. Chris Lacor Vrije Universiteit Brussel

89 . 89 . 92 . 92 . 92 . 93 . 93 . 95 . 95 . 96 . 99 . 102 . 106 . . . . . . 109 109 111 111 111 111 112

Filters Electrostatic precipitators

100 101 Particle size (micron)

1.2 Cyclone separators: types and principals

1.2. Cyclone separators: types and principals

Figure 1.2: Classication of dust collectors

Figure 1.3: Typical cyclone separator

1.2. Cyclone separators: types and principals

1.2.1 Advantages and disadvantages of cyclones

Figure 1.4: Applications of cyclone separator

1.2. Cyclone separators: types and principals

1.2.3 Principals of cyclonic separation

1.2. Cyclone separators: types and principals

Distance from center

Figure 1.6: The tangential velocity distribution in a real vortex [77]

Factors affecting the cyclone performance

2. Particle properties Density Shape Diameter and distribution 10

1.3. Motivation of this work

Mass loading 3. Gas properties Velocity Density Viscosity Temperature Pressure

1.3 Motivation of this work

1.4. Outline of the thesis

1.4 Outline of the thesis

Chapter 2 Literature Review

2.1 Classication of study approaches

Chapter 2. Literature Review

Figure 2.1: The cyclone separator dimensions

works become a tool to study the effect of cyclone geometry on performance.

2.2 Mathematical models

2.3. Experimental methods

2.3 Experimental methods

2.4 Computational uid dynamics (CFD)

Chapter 2. Literature Review

2.5 Discrepancy in the previous studies

The cone tip diameter

2.5. Discrepancy in the previous studies

2.5.2 The dust outlet geometry

2.5.3 The inlet dimensions

Chapter 2. Literature Review

The vortex nder dimensions

2.5. Discrepancy in the previous studies

2.5.5 The cyclone heights

2.5.6 Previous optimization studies

Chapter 2. Literature Review

2.6. Summary and research plan

2.6 Summary and research plan

Chapter 2. Literature Review

Chapter 3 Governing Equations

The energy cascade

Chapter 3. Governing Equations

Figure 3.2: Energy spectrum for a turbulent ow [183]

Chapter 3. Governing Equations

3.2 The governing equations for the gas phase

(ui ) = 0 + t xi p ij (ui uj ) = + (ui ) + t xj xi xj (Hui ) = (E ) + t xi (ji uj qi ) xi

(3.7) (3.8) (3.9)