Study On Slurry Flow Modelling in Pipeline (Eth) (LAHIRI, Sandir Kumar) (Nat. Inst. of Techn. Durgapur 2009) (329s)

From the SelectedWorks of Dr.
Sandip Kumar
Lahiri
March 2010
My PhD thesis : Study on slurry flow modelling in

pipeline
Contact
Author
Start Your Own

SelectedWorks
Available at: http://works.bepress.com/sandip_lahiri/23
Notify Me
of New Work
STUDYONSLURRYFLOWMODELINGINPIPELINE
Sandip Kumar Lahiri
STUDYONSLURRYFLOWMODELINGINPIPELINE
THESIS
Submittedinpartialfulfillmentofthe
Requirementsforthedegreeof
D O CTORO F P HILOSOPHY
By
S ANDIP K UMAR L AHIRI
UndertheSupervisionof
D R . K.C.G HANTA
Professor,DepartmentofChemicalEngineering
NATIONALINSTITUTEOFTECHNOLOGY,DURGAPUR
DEPARTMENTOFCHEMICALENGINEERING
DURGAPUR713209,INDIA
2009
ii
DEDICATED
TO
MY PARENTS
&
MY WIFE
iii
NATIONAL INSTITUTE OF TECHNOLOGY

DURGAPUR
DEPARTMENT OF CHEMICAL ENGINEERING
CERTIFICATE
This is to certify that the thesis entitled Study on slurry flow modeling in
pipeline submitted by Sandip Kumar Lahiri in fulfillment of the requirement
of the Degree of Doctor of Philosophy is a record of bonafide research work
carried out by him, in the Department of Chemical Engineering, National
Institute of Technology, Durgapur, under my guidance and supervision. In my
opinion, the thesis has reached the standard fulfilling the requirements of the
Ph. D. degree as prescribed in the regulations of this institute.
(Dr. K.C. Ghanta)

Department of Chemical Engineering
National Institute of Technology
Durgapur, 713209
India
iv
Acknowledgement
Finally, a successful end to the long episode of my PhD study is reached. My journey towards
the PhD degree could not have been better than this. I appreciate the support provided by many
people, which made this journey more exciting and delightful.
I am highly grateful to my supervisor Prof. K.C. Ghanta, who helped me in shaping up the
research work presented in this thesis. Sir, thank you for all the help inside and outside the
campus you have provided to enable me to reach the target. Your simplicity and attitude of
life long learning always kept me motivated and encouraged me to learn the different
aspects of science and technology.
I would like to express my deepest gratitude to the authority of NIT, Durgapur, IPCL,
Nagothane and Sabic , Saudi Arabia to allow me to pursue my dream and helping me to see
this day. I wish to thank to my colleagues Mr Nadeem Khalfe and Mr. Chnimaya Lenka for
their continuous support and encouragement.
I wish to extend the deepest sense of respect to my parents who always inspired and motivated
me to go for the best, in spite of many testing times at their end. Heartiest thanks to my wife ,
whose patience, understanding and constant encouragement enabled me to complete this work.
Sandip Kumar Lahiri
Paper published based on present thesis

Regime identification
1.
Lahiri, Sandip K. and Ghanta, Kartik Chandra ,Development of a Hybrid Artificial

Neural Network and Genetic Algorithm Model for Regime Identification of Slurry
Transport in Pipelines, Chemical Product and Process Modeling: Vol. 4 : Iss. 1,
Article 22. (2009)
2.
Lahiri S.K and Ghanta K.C , Development of a Support Vector Classification Method
For Regime Identification of Slurry Transport in Pipelines , Hydrocarbon
Processing, To be published in Sept issue ,(2009)
Critical velocity
3.
Lahiri S.K and Ghanta K.C , The support vector regression with the parameter
tunning assisted by a differential evolution technique: Study of the critical velocity of
a slurry flow in a pipeline, Chemical Industry & Chemical Engineering Quarterly 14
(3) ,191203 (2008)
4.
Lahiri S.K and Ghanta K.C , Minimize power consumption in slurry transport,
Hydrocarbon Processing,December issue(2008)
5.
Lahiri, Sandip K. and Ghanta, Kartik Chandra ,Hybrid Support Vector Regression
and Genetic Algorithm Technique A Novel Approach in Process Modeling,
Chemical Product and Process Modeling: Vol. 4 : Iss. 1, Article 4. (2009)
Hold up
6.
Lahiri S.K and Ghanta K.C , Development of an artificial neural network correlation
for prediction of hold-up of slurry transport in pipelines, Chemical Engineering
Science 63 ,1497 1509 (2008)
vi
7.
Lahiri S.K and Ghanta K.C ,Artificial neural network model with the parameter
tunning assisted by a differential evolution technique: The study of hold up of the
slurry flow in a pipeline, Chemical Industry & Chemical Engineering Quarterly 15
(2009)
8.
Lahiri S.K and Ghanta K.C , Genetic algorithm tuning improves artificial neural
network models , Hydrocarbon Processing, January issue(2009)
Pressure drop
9.
Lahiri S.K and Ghanta K.C, Development of an artificial neural network correlation
for prediction of pressure drop of slurry transport in pipelines, International J. of
Math. Sci. & Engineering Applications, Vol 2,No 1,1-21(2008)
10. Lahiri S.K and Ghanta K.C , Prediction of Pressure drop of Slurry Flow in Pipeline
by Hybrid Support Vector Regression and Genetic Algorithm Model, Chinese Journal
of Chemical Engineering, 16(6) ,(2008)
11. Lahiri, S.K. and Ghanta, Kartik Chandra, Artificial Neural Network Model with
Parameter Tuning Assisted by Differential Evolution Technique: Study of Pressure
Drop of Slurry Flow in Pipeline, Chemical Product and Process Modeling: Vol.
3 : Iss. 1, Article 54. (2008)
12. Lahiri S.K and Ghanta K.C , Support vector regression with parameter
tuning assisted by differential evolution technique: Study on pressure drop of
slurry flow in pipeline, paper accepted for publication(schedule october09
issue), Korean journal of chemical engineering(2009)
CFD
13. Lahiri S.K and Ghanta K.C , Computational Fluid Dynamics Simulation of the Solid
liquid slurry flow in a pipeline, Hydrocarbon Processing, may issue ,(2009)
vii
14. Lahiri S.K and Ghanta K.C, Computational technique to predict the velocity and
concentration profile for solid-liquid slurry flow in pipelines. Hydro transport 17, The
17th International Conference on the hydraulic transport of solids. The Southern
African Institute of Mining and metallurgy and the BHR Group, 2007, 149-175
(August 2007)
Commercial application
15. Lahiri, S.K. and Khalfe N , Process modeling and optimization of industrial ethylene
oxide reactor by integrating support vector regression and genetic algorithm, The
Canadian Journal of Chemical Engineering, Volume 87, Issue 1, Pages 118-128,
(February 2009)
16. Lahiri S.K , Khalfe N and Garawi M , Process modeling and optimization strategies
integrating neural networks and differential evolution, Hydrocarbon Processing, Oct
issue ,35-50(2008)
17. Lahiri S.K and Khalfe N , Novel approach for process plant monitoring,
Hydrocarbon Processing, march issue ,(2008)
18. Lahiri, Sandip Kumar and Khalfe, Nadeem) ,Process Modeling and Optimization
Strategies Integrating Support Vector Regression and Differential Evolution: A Study
of Industrial Ethylene Oxide Reactor, Chemical Product and Process Modeling:
Vol. 3 : Iss. 1, Article 57. (2008)
Paper under review

19. Lahiri S.K and Ghanta K.C , Critical velocity of slurry flow, under review, Asia
pacific journal of chemical engineering(2009)
viii
20. Lahiri S.K and Ghanta K.C , Prediction of pressure drop and concentration
profile by semi empirical correlations--modification of Wasp model, under
review, Chemical engineering progress(2009)
21. Lahiri S.K and Ghanta K.C , Prediction of hold up of Slurry Flow in Pipeline
by Hybrid Support Vector Regression and differential evolution Model,
under review, Powder technology(2009)
ix
List of Contents
Contents
Page
1. Introduction
1.1 Introduction
1.2 Typical flow regime
1.3 Critical velocity
1.4 Hold up
1.5 Pressure drop
1.6 Concentration and velocity profile
1.7 Motivation of present work
1.8 Scope of present work

1.8.1 Advanced numerical modeling
1.8.2 Semi empirical modeling
1.8.3 Multiphase computational fluid dynamics (CFD) based
modeling
1.9
Thesis purview
10
2. Mathematical tools
2.1 Introduction
13
2.2 Artificial neural network
14
2.2.1 Background works
14
2.2.2 Overview of classification neural networks
15
2.2.3 Overview of prediction neural networks
16
2.2.4 Strengths of ANN
17
2.2.5 Limitations of neural networks
19
2.3 Comparison of neural networks to empirical modeling
20
2.4 Artificial neural network (ANN) based modeling
21
2.4.1 Network Architecture
21
2.4.2 Training
22
2.4.3 Back propagation algorithm (BPA)
22
2.4.3.1 Different back propagation algorithm
25
2.4.4 Performance measures of ANN model
25
2.4.5 Generalizability
27
2.4.6 Step-wise Procedure for Developing an Optimal MLP Model
27
2.5 Tuning parameters of ANN
28
2.6 Optimization of ANN model
30
2.7 Genetic algorithm
32
2.7.1 Literature survey of genetic algorithm
33
2.7.2 Techniques used in GA
33
2.8 GA-based Optimization of ANN Models
36
2.9 Differential evolution
39
2.9.1 Literature survey of Differential evolution
39
2.9.2 Steps performed in DE
40
2.10 DE-based Optimization of ANN Models
41
2.11 Support Vector Machines
44
2.12 Background works
45
2.13 The basic idea behind SVM modeling
46
2.13.1 Support vector machine for classifications
46
2.13.2 Mathematics behind SVM algorithm for classification
48
2.13.3 Training, testing and Generalizability
51
2.13.4 Mathematics behind SVM algorithm for regression
52
2.14 Tuning parameters of SVR
55
2.15 Optimization of SVM model
58
2.16 GA-Based Optimization of SVR Models
60
2.17 DE based optimization of SVR model
61
2.18 Conclusion
64
3. Regime identification of slurry transport in pipelines

3.1 Introduction
82
3.2 Background work
83
3.2.1 Flow regimes in slurry flow
85
3.2.2 Head loss correlations for separate flow regimes
87
3.2.3 Flow regime boundaries (Turian and Yuans approach)
88
3.3 Performance check of Turian Yuans approach
91
xi
3.3.1 Data Collection
91
3.3.2 Regime identification
92
92
3.5. Development of the artificial neural network (ANN) based correlation

3.5.1. Input selection and data collection
96
3.5.2 Prediction performance of hybrid ANN-DE model
96
3.5.3 Prediction performance of hybrid ANN-GA model
99
3.5.4 Comparison of hybrid ANN-DE model with ANN model
100
3.6 Development of the support vector machine (SVM) based correlation

101
3.6.2 Prediction performance of hybrid SVM-DE model
101
3.7. Conclusion
102
4. Critical velocity of slurry flow in pipeline

4.1 Introduction
107
110
4.3 Development of the artificial neural network (ANN) and support

vector regression (SVR) based correlation
120
4.3.1 Collection of data
120
4.3.2 Identification of input parameters
120
4.4 Results and discussion

123
4.4.2 Comparison of hybrid ANN-DE model with ANN model:
125
4.4.3 Prediction performance of hybrid SVR-GA model prediction
126
4.5 Dependence of critical velocity with model input parameters
128
4.6 Comparison with other published correlations
132
4.7. Conclusion
132
5. Hold up in slurry flow in pipeline

5.1 Introduction
138
139
5.3. Development of the artificial neural network (ANN) and support

vector regression (SVR) based correlation
141
xii
141
142
5.4. Results and discussion
144
144
146
5.4.3. Prediction performance of hybrid SVR-DE model
147
5.5 Conclusion
148
6. Pressure drop of slurry flow in pipeline

6.1 Introduction
152
154
6.2.1 Methods Based on the Drag Coefficient of Particles
157
6.2.2 Models Based on Terminal Velocity
159
6.2.3 Friction losses for compound mixture in horizontal

heterogeneous flows
159
6.2.4 Stratified flows
161
6.2.5 Two layer mode
163
6.2.6 Modified Wasp model
165
6.2.7 Summary of literature survey
166
6.3. Development of the artificial neural network (ANN) and

support vector regression (SVR) based correlation
167
168
169
169
169
171
6.4.3 Comparison of hybrid ANN-GA model with ANN model
173
6.4.4 Prediction performance of hybrid SVR-DE model
173
6.4.5 Comparison with other published correlations
175
6.5 Conclusion
176
xiii
7. Semi-empirical method for pressure drop and concentration
profile
7.1 Introduction
183
7.2 Background work
184
7.3 Wasp et al. (1977) model
187
7.4 Comparison of pressure drop prediction by Wasp model with

experimental data
189
7.5 Modified Wasp model

7.5.1 Modifications incorporated in Wasp model
192
7.5.2 Steps to implement modified Wasp model
193
7.6 Results and discussions
196
7.7 Conclusion
205
8. Computational Fluid Dynamics modeling of the Solid liquid
slurry flow in a pipeline

8.1 Introduction
212
214
8.2.1 Multiphase modeling
215
8.2.2 Approaches to Multiphase Modeling
216
8.2.2.1 The Euler-Lagrange Approach
216
8.2.2.2 The Euler-Euler Approach
217
8.2.2.2.1 The VOF Model
217
8.2.2.2.2 The Mixture Model
218
8.2.2.2.3 The Eulerian Model
218
8.2.3 Choosing a Multiphase Model
218
8.2.4 Effect of Particulate Loading
218
8.2.5 Significance of the Stokes Number
220
8.2.6 Guidelines for choosing appropriate model
220
8.3 Formulation of multiphase CFD model

8.3.1 Eulerian Model
221
222
8.3.1.1 Continuity Equation
222
8.3.1.2 Momentum Equations
223
8.3.1.3 Fluid-solid momentum equations
223
xiv
8.3.1.4 Interphase exchange co-efficient
224
8.3.1.5 Fluid-Solid Exchange Coefficient
224
8.3.1.6 Lift Forces
227
8.3.1.7 Solid pressure
228
8.3.1.8 Radial Distribution Function
228
8.3.1.9 Solids Shear Stresses
229
8.3.2 Turbulent model

8.4 Description of CFD simulation
231
232
8.4.1 Two dimensional simulation
232
8.4.2 Validation of CFD simulation
236
8.4.3 Comparison between measured and predicted

concentration profiles based on Syamlal-O'Brien model model,
Wen and Yu model and Gidaspow model
236
8.4.4 Description of modified model
237
8.4.5 Comparison between measured and predicted

concentration profiles based on modified model
238
8.4.6 Three dimensional simulation
239
8.4.7 Results and discussion of 3D simulation
241
8.4.7.1 Concentration profile
241
8.4.7.2 Velocity profile
242
8.4.7.3 Pressure drop
242
8.4.7.4 Contours of solid concentration and velocity
243
8.5 Conclusion
286
9. Contribution of present thesis and future scope

9.1 Contribution of present thesis
293
9.2 Future scope
298
xv
List of tables
Table 2.1:
Different activation function
29
Table 2.2:
Different Kernel type
59
Table 2.3:
Different Loss function
60
Table 3.1:
System and parameter studied
93
Table 3.2:
Some of the input and output data for ANN & SVM training
94
Table 3.3:
Regime identification by Turian and Yuan (1977) approach
95
Table 3.4:
Prediction error by hybrid ANN-DE based model
98
Table 3.5:
Set of equations and fitting parameters for neural network

correlations (i=7, j=7, k=1)
Table 3.6:
99
Comparison of performance of ANN-DE hybrid model Vs ANN

Model
100
Table 3.7:
Prediction error by hybrid SVM-DE based model
102
Table 4.1:
Performance of different correlations to predict critical velocity
119
Table 4.2:
121
Table 4.3:
Typical input and output data for ANN or SVR training
122
Table 4.4:
124
Table 4.5:

correlations(i=7,j=8,k=1)
Table 4.6:
125

Model
126
Table 4.7:
Prediction error by hybrid SVR-GA based model
127
Table 4.8:
Optimum parameters obtained by hybrid SVR- GA algorithm
127
Table 5.1:
142
Table 5.2:
Typical input and output data for ANN & SVR training
143
Table 5.3:
Prediction error by the hybrid ANN-DE based model
144
Table 5.4:

correlations (i=7, j=8,k=1)
Table 5.5:
145

Model
146
Table 5.6:
Prediction error by hybrid SVR-DE based model
147
Table 5.7:
Optimum parameters obtained by hybrid SVR- DE algorithm
148
xvi
Table 5.8:
Comparison of performance of SVR-DE hybrid model Vs SVR

model
148
Table 6.1:
168
Table 6.2:
Typical input and output data for ANN training
170
Table 6.3:
171
Table 6.4:

correlations (i=7, j=8, k=1)
Table 6.5:
172

Model
173
Table 6.6:
Prediction error by SVR based model
174
Table 6.7:
Optimum parameters obtained by hybrid SVR- DE algorithm
175
Table 6.8:
Comparison of performance of SVR-DE hybrid model Vs SVR model 175
Table 6.9:
Performance of different correlations to predict pressure drop
176
Table 7.1:
Drag relationships
188
Table 7.2:
System and parameter studied collected from the literature
190
Table 7.3:
Coal water slurry data collected from Roco & Shook (1984)
190
Table 7.4:
Comparison of pressure drop and concentration profile prediction
197
Table 7.5:
Comparison of correlation co-efficient (R) for concentration profile

prediction by Wasp model and modified Wasp model.
205
Table 8.1:
Experimental data used in the present study
233
Table 8.2:
Different inputs for simulation in FLUENT
234
Table 8.3:
Data used in 3D CFD simulation
241
xvii
List of figures
Figure 2.1:
Different application of ANN
15
Figure 2.2:
Strength and characteristics of ANN
19
Figure 2.3:
Architecture of feed forward network with one hidden layer
23
Figure 2.4:
Different ANN algorithms published in various literatures
26
Figure 2.5:
Structure of different activation function
29
Figure 2.6:
Schematic for hybrid ANN-GA algorithm implementation
38
Figure 2.7:
Schematic for hybrid ANN-DE algorithm implementation
43
Figure 2.8:
Separation of two classes by SVM
47
Figure 2.9:
Non-linear transformation from input to a higher-dimensional feature

space
Figure 2.10:
51
A schematic diagram of support vector regression using -sensitive loss

function
53
Figure 2.11:
Schematic for hybrid SVR-GA algorithm implementation
62
Figure 2.12:
Schematic for hybrid SVR-DE algorithm implementation
63
Figure 2.13:
A simplified three layer feed forward perceptron network
65
Figure 3.1:
Heterogeneous flow regimes in terms of speed versus volumetric

concentration.
Figure 3.2:
84
Flow regimes of heterogeneous flows in terms of particle size versus

mean velocity
84
Figure 3.3:
Four regimes of flow of settling slurries in horizontal pipeline
86
Figure 3.4:
Decision tree for establishing flow regimes
90
Figure 3.5:
Experimental Vs predicted flow regime for BFGS algorithm
97
Figure 4.1:
Plot of transitional mixture velocity with pressure drop
108
Figure 4.2:
Schematic representation of the boundaries between the flow regimes

for settling slurries in horizontal pipelines
Figure 4.3:
109
Simplified concept of particle distribution in a pipe as a function of

volumetric concentration and velocity
110
Figure 4.4:
Major critical velocity correlations available in literatures
113
Figure 4.5:
Experimental Vs predicted critical velocity for Marquard Levenburg
Figure 4.6:
algorithm
124
Variation of critical velocity with density ratio
129
xviii
Figure 4.7:
Variation of critical velocity with pipe diameter
130
Figure 4.8:
Variation of critical velocity with particle diameter
130
Figure 4.9:
Variation of critical velocity with solid concentration
131
Figure 4.10:
Variation of critical velocity with dimensionless group
131
Figure 5.1:
Experimental Vs predicted hold up for FletcherReeves update

algorithm
145
Figure 6.1:
Major correlations for pressure drop published in literature
154
Figure 6.2:
Transfer of momentum between the fluid and the wall during slurry
flows through a pipe
Fig 6.3:
Experimental Vs predicted pressure drop by Marquard Levenburg

algorithm
Figure 7.1:
172
Pressure drop prediction by Wasp model and modified Wasp model

based on typical Roco & Shook (1984) data
Figure 7.2:
155
191
Experimental and calculated concentration profile using modified

Wasp
model for some typical data of Roco and Shook (1983).
a)Run A1:Vm= 1.66 m/s, Cvf=8.37%,D=0.0515 m ,

b)Run A2:Vm= 3.78 m/s, Cvf=9.2%, D=0.0515m , c)Run A3:
Vm= 1.66 m/s, Cvf=18.7%, D=0.0515 m
Figure 7.3:
198

Wasp model for some typical data of Roco and Shook (1983). a)Run
A4:Vm= 4.17 m/s, Cvf=18.9%,D=0.0515 m , b)Run A5:Vm= 1.66 m/s,
Cvf=28%, D=0.0515m , c)Run A6: Vm= 4.33 m/s, Cvf=28.6%,
D=0.0515 m
Figure 7.4:
199

Wasp model for some typical data of Roco and Shook (1983).
a)Run A7:Vm= 2.9 m/s, Cvf=10.3%,D=0.263m b)Run A8:Vm= 3.5 m/s,
Cvf=10%, D=0.263m, c)Run A9: Vm= 2.9 m/s, Cvf=19 %,
D=0.263m
Figure 7.5:
200

a)Run A10:Vm= 3.5 m/s, Cvf=18.4%,D=0.263m b)Run A11:
Vm= 2.9 m/s, Cvf=27%, D=0.263m, c)Run A12: Vm= 3.5 m/s,
Cvf=26.8 %, D=0.263m
201
xix
Figure 7.6:

a)Run A13:Vm= 2.9 m/s, Cvf=34.1%,D=0.263m b)Run A14:
Vm= 3.5 m/s, Cvf=33.8%, D=0.263m, c)Run A15: Vm= 3.16 m/s,
Cvf=10.4 %, D=0.495 m
Figure 7.7:
202

a)Run A16:Vm= 3.76 m/s, Cvf=10.0 %,D=0.495 m b)Run A17:
Vm= 3.07 m/s, Cvf=18.7%, D=0.495 m, c)Run A18: Vm= 3.76 m/s,
Cvf=18.4 %, D=0.495 m
Figure 7.8:
203

a)Run A19:Vm= 3.16 m/s, Cvf=27.3 %,D=0.495 m b)Run A20:
Vm= 3.76 m/s, Cvf=26.9%, D=0.495 m
204
Figure 8.1:
Different approaches to multiphase modeling
217
Figure 8.2:
Guidelines for choosing multiphase model
219
Figure 8.3:
Measured (by Kaushal) and predicted (by present model, Syamlal

model Gidaspow model and Wen and Yu model) concentration
profiles at different efflux concentrations and flow velocity for the
flow of zinc tailing slurry through a 105-mm-diameter pipe.
Figure 8.4:
239
Measured (by Mukhtar) and predicted (by present model)

concentration profiles at different efflux concentrations and flow
velocity for the flow of zinc tailing slurry through a
105-mm-diameter pipe.
Figure 8.5:
240
Measured (by Kaushal) and predicted (by present model) concentration

profiles at different efflux concentrations and flow velocity for the
flow of zinc tailing slurry through a 105-mm-diameter pipe with a
velocity of 2.75 m/s.
Figure 8.6A:
241
Comparison of experimental and calculated vertical concentration

profile for flow of 125 micron glass beads in 54.9 mm diameter pipe
at different efflux concentration and flow velocity
Figure 8.6B:
245

246
xx
Figure 8.7A:

Figure 8.7B:
247

Figure 8.8A:
248

Figure 8.8B:
249

Figure 8.9:
Experimental and calculated vertical concentration profile for flow

of 125 micron glass beads in 54.9 mm diameter pipe
Figure 8.10A:
258
Concentration profiles in the vertical plane for slurry of 440 micron

particle size
Figure 8.15A:
257
Concentration profiles in the vertical plane for slurry of 125 micron

particle size
Figure 8.14:
256

Figure 8.13:
255

Figure 8.12B:
254

Figure 8.12A:
253

Figure 8.11B:
252

Figure 8.11A:
251

Figure 8.10B:
250
259
CFD predicted solid phase vertical velocity profile for flow of 125
micron glass beads in 54.9 mm diameter pipe at different efflux
concentration and flow velocity
260
xxi
Figure 8.15B:
Figure 8.16A:
261
concentration and flow velocity.
Figure 8.16B:
262
concentration and flow velocity.
Figure 8.17A:
263
Figure 8.17B:
264
Figure 8.18:
micron glass beads in 54.9 mm diameter pipe
Figure 8.19:
271
Contours of solid velocity magnitude (m/s) [Cvf- 9.4 %, Vm-1 m/s,

d-125 micron]
Figure 8.26:
270
Contours of volume fraction of solid [Cvf- 9.4 %, Vm-1 m/s, d-125

micron]
Figure 8.25:
270
Pressure drop for slurry of 440 m particle size at different overall

area-verage concentrations and flow velocities
Figure 8.24:
269
Pressure drop for slurry of 125 m particle size at different overall

area-average concentrations and flow velocities
Figure 8.23:
268
Parity plot of predicted Vs experimental pressure drop for slurry flow

at different overall area-average concentrations and flow velocities
Figure 8.22:
267
Comparison of vertical velocity profile at a) 1 m/s, b) 3 m/s and

c) 5 m/s for different efflux concentration
Figure 8.21:
266
micron glass beads in 54.9 mm diameter pipe
Figure 8.20:
265
271
Contours of volume fraction of solid [Cvf- 10.41 %, Vm-3 m/s,

d-125 micron]
272
xxii
Figure 8.27:
Contours of solid velocity magnitude (m/s) [Cvf- 10.41 %,

Vm-3 m/s, d-125 micron]
Figure 8.28:

d-125 micron]
Figure 8.29:
279

Figure 8.43:
279
Contours of volume fraction of solid [Cvf- 41.1 %,

Vm-2 m/s] , d-125 micron]
Figure 8.42:
278

Figure 8.41:
278

Figure 8.40:
277

Figure 8.39:
277

Figure 8.38:
276

Figure 8.37:
276

Figure 8.36:
275

Figure 8.35:
275

Figure 8.34:
274

Figure 8.33:
274
Contours of volume fraction of liquid [Cvf- 20.4 %, Vm-3 m/s,

d-125 micron]
Figure 8.32:
273

d-125 micron]
Figure 8.31:
273

Figure 8.30:
272
280

280
xxiii
Figure 8.44:

Figure 8.45:

Figure 8.46:
Figure 9.1:
284

Figure 8.53:
284

Figure 8.52:
283

Figure 8.51:
283

Figure 8.50:
282

Figure 8.49:
282

Figure 8.48:
281

Figure 8.47:
281
285

d-440 micron]
285
Contribution of present thesis
294
xxiv
Abstract
Many large slurry pipelines were built and operating around the world. Pipeline transport is
considered economical and environment friendly as compared to rail and road transport. To
design the pipelines and its associated facilities (pumps etc) designers need accurate
information regarding pressure drop, hold up, critical velocity, flow regimes etc at the early
design phase. Also the operating engineers need to know accurately the critical velocity so
that he can adjust the slurry flow to have a minimum pressure drop to ensure minimum
operating cost. Such flows are complex and presently very little known about the two-phase
interaction of solid liquid behavior inside pipeline. The correlations presently available in open
literatures for the above mentioned parameters have a prediction error of 25-35%. This much
of error in design and slurry operation has serious cost implication and is totally unacceptable
in present day competitive business scenario. This study was performed in order to develop
model for flow of slurries through pipelines so that the error % can be reduced. This thesis can
be considered as a step forward for better understanding of flow behavior in slurry pipelines.
Attempt has been made in this thesis to utilize the computational capability of two recent
advanced numerical technique namely artificial neural network (ANN) and support vector
regression (SVR) in slurry flow modeling. This thesis has build some simple and superior
correlations of pressure drop, hold up, critical velocity, flow regimes which can be readily
used by design engineers to design slurry pipelines and pumps.
There are some model parameters both in ANN and SVR that are to be tuned by the expert
user during model building time. A new approach was developed in this thesis to tune these
parameters automatically using differential evolution (DE) and genetic algorithm (GA). The
method employs a hybrid approach for minimizing the generalization error. The proposed
hybrid technique relieves the non-expert users to choose the meta parameters of ANN or SVR
algorithm for the used case study and find out the optimum value of these meta parameters on
its own.
In the present study existing Wasp model (1977) for pressure drop has been modified by
alleviating some of the restrictive assumptions used in that model. A new method was also
developed to calculate concentration profile using Wasp model as a starting point. The
concentration profile and pressure drop data predicted by modified model were compared with
the experimental one collected from literature.
xxv
In this study the capability of computational fluid dynamics (CFD) is explored to model
complex solid liquid slurry flow in pipeline. A comprehensive CFD model was developed to
gain deeper insight of the solid liquid slurry flow in pipelines. The theoretical model
developed in this work represents the synthesis of hydrodynamic and interparticle interaction
effects within the framework of equation of conservation of momentum and mass. Two and
three-dimensional model problems are developed using CFD to understand the influence of the
particle drag coefficient on solid concentration profile. It is found that the commercial CFD
software is capable to successfully model the solid liquid interactions in slurry flow and the
predicted concentration profiles show reasonably good agreement with the experimental data.
xxvi
Chapter1
CHAPTER 1
Introduction
Abstract
Slurry transport through pipeline is a widely practiced field in mining and allied industries.
Background of slurry flow modeling, its evolution over the years and limitations of our current
knowledge in this field is presented in this introductory note. Designers need accurate
information regarding pressure drop, hold up, critical velocity, flow regimes etc. at the early
stage of designing the pipelines and its associated facilities (pumps etc). The meaning and
implications of these design parameters in slurry flow modeling, motivation and the scope of
present thesis are described in this chapter with little details. At the end a road map is given
regarding how to read this thesis.
Keywords: Artificial neural network (ANN); Differential evolution (DE); Slurry flow regime,
Slurry critical velocity
Introduction
1.1 Introduction
Pipeline transport has been a progressive technology for conveying a large quantity of bulk
materials. This includes long distance hauling of coal, minerals, ore and solid commodities,
dredging and filling, collection and disposal of solid waste and material processing. It is
possible in the present day technology to incorporate sequences of processing operations into
the overall slurry transport operation, thereby leading to elimination of processing steps and
savings in capital investment; for example, integrating microbial desulfurization into the long
distance coal slurry pipeline transport operation. Compared to a mechanical transport, the use
of a pipeline ensures a dust free environment, demands substantially less space, makes
possible full automation and requires a minimum number of operating staff. On the other hand,
it needs higher operational pressures and demands considerably high quality pumping
equipment and control system.
The behavior of solids and liquids flowing through pipelines has been the subject of
continuing investigation since the turn of the century. In the 1950s significant technical
progress was made in several countries through a strong research effort. In United Kingdom,
experimental work was conducted particularly on the handling of coarse coal slurries. This
work was mainly carried out by worldwide well known British Hydraulic Research
Association (BHRA) in conjunction with the National Coal Board,UK. In France, Durand and
Condolios (1952) carried out a large amount of work on the hydraulic transport of aggregates.
During 1960s several countries became involved in developing hydraulic transport for mining
and a number of coal mine haulage systems were installed. Rigby (1982) took a wide historical
view concerning slurry transport, referring to its early development in the American gold rush
of the mid-nineteenth century. Since then, the Ohio Cadiz coal line, built in 1957 pioneered the
large scale (147 km long254 mm diameter) transport of material at high throughputs (1.5
Mtpa). It was later overtaken by the Black Mesa, 439 km457 mm5 Mtpa, supplying coal to
the Mohave power station in southern Nevada. The first iron ore concentrate line (Savage
River) was built in Tasmania (1967), with conservative slope specifications to cope with the
solids specific gravity. It too has been overtaken in scale by the Brazilian Samarco line,
carrying 7 Mtpa of iron ore concentrate over 400 km. Other materials have included limestone
(UK), gold slime (Australia, South Africa), phosphate (Canada, South Africa), copper
concentrate (Papua New Guinea), copper tailings (Chile), and zinc sulphites (Japan), to name
but a few. Requirements are now placed on public utilities and mining companies to
2
Chapter1
incorporate effective means of waste disposal into their future plans. Slurry systems are used
in flue gas desulphurization and in waste material transport. Hydraulic transport of waste
material, such as bauxite residue, coal mine tailings and sewage, is also a widely accepted and
practiced application.
At the present time there are many organizations throughout the world carrying out research
and development in the field of slurry transport. It is understood that the greatest interest is
been shown in these major lines because of their huge capital investment and substantial
engineering content. It must be noted, however, that there are many small slurry pipelines
being designed and built particularly in the mining, chemical and food processing industries,
for which the details remain unpublished.
A basic understanding of the underlying phenomena is vital to the control of slurry transport
system. Literature survey reveals that studies concerned with solid-liquid mixture flows have
followed one of these three major approaches: 1) the empirical approach 2) the rheological
based continuum approach and 3) the multiphase flow modeling approach. The empirical
approach seems to have received the most attention, perhaps as a concession to the complexity
of slurry flows. Because of its long history a large body of empirical studies dealing with
slurry transport has accumulated the correlations for prediction of pressure drop and for
delineation of flow regimes which constitute two major elements of this body of empirical
work. The rheological approach has emerged in a major way in the mid fifties. It is, however,
strictly applicable to slurries of ultra fine non-colloidal particles, capable of meaningful
rheological characteristics. The multiphase flow modeling approach, which accounts for
liquid, particle and boundary interaction effects, provides the most rational framework for
describing such heterogeneous solid-liquid mixture flows. It requires basic information
regarding the effects of the particle on the structure of the turbulent flow, the particle-particle
and the particle-boundary interactions and other effects, and this approach commonly entails
substantial computational effort.
All of the above methods have their own limitations generating out of inherent complexity and
poor understanding of two-phase flow systems. Despite of extensive research in slurry
technology, our present knowledge of the fundamentals of solid-liquid flow does not satisfy
engineering needs. The need to control processes involving solid-liquid slurry and to design
pipelines, pumps has resulted in the accumulation of a great deal of experimental data. In turn
these data have help developed the body of empirical relations and practical guidelines. But it
is difficult to integrate this body of knowledge into a framework leading to the design
correlation. A predictive model with sound understanding of the fundamentals of particle laden
3
Introduction
turbulent flow, including all significant interactions and the ability to integrate these
quantitatively is not so successful till today as seen from various literatures.
Some of the limitations of this body of published work include:
The slurries investigated and the problems addressed are so specific as to cover a very
limited range of the variables involved.
The particle size distribution (PSD) considered in the experiment is very narrow,
whereas the PSD is very broad in industrial scenario.
Most of the experiments were carried out in very low solid concentration (less than
10%) and correlations developed from such experiments failed to produce reasonable
results at higher solid concentrations (above 25%).
The average error for pressure drop prediction was 35% during Wasp (1977) and now
reduced to 20% with maturity in slurry modeling. This 20% error is not even
acceptable in todays scenario as the slurry transport is very energy extensive
operation and whole economics of the transportation depends on pressure drop.
The slurry systems are not well defined with respect to, say, solids shape and particle
size distributions.
The published works contain incomplete data, which are often impossible to retrieve
and reconstruct from the published versions.
The body of published empirical correlations on slurry pressure drop and critical
velocity is extensive but largely conflicts with each other.
There is limited range of applicability and validity. As for example the correlations
developed for coal-water slurry are found not fit for sand-water slurry.
Because of the resulting uncertainties, extensive pilot pipeline tests are conducted for major
projects around the world. But even then the knowledge of the detailed in situ conditions is
often beyond reach. Because of the complexity of the process, mathematical solution to the
general hydrodynamic problem of slurry flows in pipeline has been a forbidden task for many
decades. Advances in our understanding of turbulent flows and their modeling in recent years
have provided the basic framework for development of mathematical models of slurry flow.
Furthermore, emergence of powerful numerical technique like artificial neural network,
support vector regression, computational fluid dynamics etc. along with the accessibility of
powerful computers has made possible to test such models and to carry out investigations of
the basic phenomena using computer simulation.
Chapter1
1.2 Typical flow regime
In slurry transport different patterns of solids movements are observed, depending upon the
nature of the slurry and the prevailing flow condition. In horizontal pipes these may
conveniently be classified according to the following four regimes:
Homogeneous flow: This regime is also named as symmetric flow characterizing uniform
distribution of solids about the horizontal axis of the pipe, although not necessarily exactly
uniform. In this regime, turbulent and other lifting forces are capable of overcoming the net
body forces as well as the viscous resistance of the particles.
Heterogeneous flow: With decrease in the slurry velocity, intensity of turbulence and lift
forces are decreased. As a result there is distortion of the concentration profile of the particles,
with more of the solids, particularly the larger particles, being contained in the lower part of
the pipe. Thus there is a concentration gradient across the pipe cross section with a larger
concentration of solids at the bottom. This flow is also called asymmetric flow.
Saltation flow: This type of flow takes place at low velocities and is one in which solid
particles tend to accumulate on the bottom of the pipe, first in the form of separated dunes
and then as a continuous moving bed.
Stationary bed flow: As the slurry velocity is further reduced, the lowermost particles of the
bed become nearly stationary, the bed thickens and bed motion is limited to the uppermost
particles tumbling over one another (saltation). Eventually, with continued reduction in the
mixture velocity and build up of the bed, pressure gradient increases very rapidly to maintain
the flow and in the absence of an abnormally high applied pressure, blockage of pipe occurs.
1.3 Critical velocity

The critical velocity is defined as the minimum velocity demarcating flows in which the solids
form a bed at the bottom of the pipe from fully suspended flows. It is the transition velocity
between heterogeneous flow and saltation flow. The critical velocity is one of the important
parameter that must be accurately known for the optimized design of a slurry transportation
pipeline. The significance of this velocity is that it represents the lowest speed at which slurry
pipelines can operate and corresponds to lowest pressure drop in slurry transport.
Introduction
1.4 Hold up
In solid liquid slurry flow in pipelines different layers of solids move with different speeds.
Hold-ups are due to velocity slip of layers of particles of larger sizes, particularly in the
moving bed flow pattern. Due to this slip in velocity, in-situ concentrations are not same as the
concentrations in which the phases are introduced or removed from the system. The variation
of in-situ concentrations from the supply concentrations is referred to as hold up phenomenon.
Very few correlations with limited applicability are available in literatures to predict hold up
ratio in solid liquid slurry flow. Obviously hold up plays an important role in the failure of
many empirical correlations for predicting head loss in flow regimes involving bed formation.
1.5 Pressure drop
The design of a slurry pipeline entails predicting the power requirement per unit mass of solids
delivered over a unit distance. It is vital in this context to be able to relate head gradient to the
independent design parameters. Power consumption and subsequently the whole economics of
the hydro-transport depend on it. There are large number of empirical and semi empirical
correlations available in literatures to predict pressure drop. Most of these equations have been
developed based on limited data comprising of uniform or narrow size-range particles with
very low to moderate concentrations. These correlations are prone to great uncertainty as one
departs from the limited database that supports them. When all the major correlations are
exposed to the large experimental data bank collected from open literature (800 measurements
covering a wide range of pipe dimensions, operating conditions and physical properties), the
average prediction error is found in the range of 25 to 50%. This is definitely not acceptable in
todays scenario.
1.6 Concentration and velocity profile
Wear is a very important consideration in the design and operation of slurry systems, as it
affects both the initial capital costs and the life of components. It may be defined as the
progressive volume loss of material from a surface, due to erosion, abrasion or other causes.
Kawashima et al. (1978) indicated that wear is proportional to volume concentration,
6
Chapter1
(C v)
0.822.0
and velocity, (V)
0.85-4.5
from the review of various laboratory test results. It is
reasonable to assume that wear will depend on the number of particle impacts on the surface,
which in turn depends on the concentration and velocity. Thus to understand the wear
phenomena it is very important to know the detailed in-situ velocity profiles and concentration
profiles of solid, liquid and mixture. Such information is basic to a fundamental understanding
of the mechanisms of dense slurry transport.
1.7 Motivation of present work
Large number of long slurry pipelines was already built around the world and lot more still to
come up. Designers need accurate information regarding pressure drop, hold up, critical
velocity, flow regimes etc. at the early stage to design the pipelines and its associated facilities
(pumps etc). Despite of significant research efforts, prediction of pressure drop, hold up,
critical velocity and other design parameters to ensure optimum pipeline design is still an open
problem for design engineers. The major empirical equations regarding pressure drop, hold up,
critical velocity and flow regime identification when tested with experimental data of different
systems collected from open literatures, it was found that prediction error ranges above 25%
on an average. The capital investments for a slurry pipeline are tremendously high and
naturally 25% error in design or a few per cent error in operating conditions may have critical
cost implications. Even though the error for pressure drop has come down to 20% from 35%
during Wasp(1977), the same is not acceptable in todays business scenario as the slurry
transport is very energy extensive operation and whole economics of the transportation
depends on pressure drop. With this background, the motivation of the present work is to
develop more generalized correlations of pressure drop, hold up, critical velocity having
reasonably low prediction error over a wide range of study. Therefore, this work explores the
possibility of application of two recent advance computational techniques namely artificial
neural network (ANN) methodology and support vector machine (SVM) methodology in
slurry flow modeling. ANN and SVM have emerged as two attractive tools for nonlinear
modeling especially in situations where the development of phenomenological or conventional
regression models becomes impractical or cumbersome. The sole objective is to quickly build
the simple and superior correlations which can be readily used by design engineers to design
slurry pipelines and pumps.
7
Introduction
The second motivation is how the limited range of applicability and validity of existing
correlations can be overcome. Presently the need of the industry is to transport slurry at
maximum concentration as possible (above 30%) to reduce the water consumption and make
the transport economically more viable. Obviously, objective of the present work is to push
this concentration envelop upto 50% and to develop a generalized correlation applicable to
any slurry systems.
The third motivation is to explore the possibility of improving some existing correlations
available in literature. Among various pressure drop models available in literature Wasp model
was found most promising and versatile and can be applied with suitable modification over a
wide range of slurry system. The method developed by Wasp et al. (1977) has been used very
successfully over the last 25 years for Newtonian slurries and large number of long distance
pipelines across the world has been designed using this model. In the present study Wasp
model for pressure drop has been modified by alleviating some of the restrictive assumptions
used in the model.
The fourth motivation is to find the applicability of Computational Fluid Dynamics (CFD) in
slurry flow modeling. The limitations of empirical equations or ANN/ SVM based correlations
are that they do not provide deeper insight of complex phenomena of slurry flow. In recent
years, CFD becomes a powerful tool for predicting fluid flow, heat/mass transfer, chemical
reactions and related phenomena by solving mathematical equations that govern these
processes using a numerical algorithm on a computer. A brief review of recent literature shows
little progress in simulating flow for a slurry pipeline using computational fluid dynamics
(CFD). For solid liquid multiphase flows, the complexity of modeling increases considerably
and this remains as an area for further research and development. Due to the inherent
complexity of multiphase flows, from a physical as well as a numerical point of view,
general applicable codes of computational fluid dynamics (CFD) are non-existent.
Considering the limitations in the published studies, the present work has been concentrated on
a systematic development of a CFD based model to predict the solid concentration profile,
velocity profile and pressure drop in slurry pipeline. The objective is to gain insight of
complex solid-liquid slurry flow.
Chapter1
The problem considered in this thesis is the formulation of a theory for the flow of dense, noncolloidal, settling slurries through horizontal pipelines and the development of appropriate
numerical scheme to carry out computer simulations of these flows based on the theoretical
model.
In the present thesis all the three approaches namely advance numerical modeling (artificial
neural networks (ANNs) and support vector regression (SVR)), semi-empirical modeling and
detailed phenomenological i.e. multiphase CFD modeling have been developed to predict
pressure drop, concentration profile, velocity profile, hold up, critical velocity and flow regime
for slurry flow in pipeline.
1.8.1
Advanced numerical modeling
In the present study, two advanced techniques namely ANN and SVR are applied for slurry
flow modeling. There are some model parameters both in ANN and SVR that are to be tuned
by the expert user during model building time. A new approach was developed in this thesis
to tune these parameters automatically using differential evolution (DE) and genetic algorithm
(GA). The method employs a hybrid approach for minimizing the generalization error. The
proposed hybrid technique also relieves the non-expert users to choose the meta parameters of
ANN or SVR algorithm for the used case study and find out the optimum value of these meta
parameters on its own.
1.8.2
Semi empirical modeling
In the present study Wasp model (1977) for pressure drop has been modified by alleviating
some of the restrictive assumptions used in the model. A new method was also developed to
calculate concentration profile using Wasp model as a starting point. The concentration profile
and pressure drop data predicted by modified model were compared with the experimental one
collected from literature.
1.8.3
Multiphase computational fluid dynamics (CFD) based modeling
A comprehensive computational fluid dynamics (CFD) model was developed in the present
study to gain deeper insight of the solid liquid slurry flow in pipelines. The theoretical model
developed in this work represents the synthesis of hydrodynamic and interparticle interaction
9
Introduction
effects within the framework of equation of conservation of momentum and mass. Two and
three-dimensional model problems were developed using CFD to understand the influence of
the particle drag coefficient on solid concentration profile.
1.9 Thesis purview

Chapter 2 of the thesis contains an overview of different mathematical tools used in this
work. Initially the basic model building steps using ANN and SVR methodology was
discussed. Thereafter automatic tuning of various model parameters of both ANN and SVR by
hybrid techniques was discussed. The thesis has developed four new hybrid modeling
technique namely 1)hybrid SVR-DE technique 2)hybrid SVR-GA technique 3)hybrid ANNDE technique 4)hybrid ANN-GA technique to minimize the generalization error. The detail
steps for hybrid model building and tuning of model parameters have been presented in this
section while the application of models and their performance were discussed in subsequent
chapters.
Chapter 3 describes the application of ANN and SVM model to identify regimes in slurry
flow in pipelines. Initially, different flow regimes in solid-liquid flow are discussed. The
method of Turian and Yuan correlations to classify different regimes were summarized and
performance of this correlation assessed. Later on, the classification capability of ANN and
SVM model were explored to identify different regimes of slurry flow based on experimental
data collected from open literature.
Chapter 4 discussed on different critical velocity correlations available in literatures and
significance of the critical velocity in slurry transport. The prediction performance of ANN
and SVR based correlations for critical velocity was presented and comparison was made with
other major correlations available in literature. The parametric study of critical velocity against
its model input parameters was discussed at the end of the chapter.
Chapter 5 presents emergence of hold up phenomena in slurry flow and capability of ANN
and SVR techniques to predict this phenomenon.
Chapter 6 contains an overview of historical pressure drop correlations in slurry transport.
The application of hybrid ANN and SVR models to develop pressure drop correlations is
discussed in this chapter. The superiority of ANN and SVR model over the major historic
correlations are shown at the end.
10
Chapter1
Chapter 7 describes the Wasp model for pressure drop prediction and its performance on
large database. The chapter presents how Wasp model for pressure drop has been modified by
alleviating some of the restrictive assumptions. A comparative study on concentration profile
and pressure drop predicted by modified model is shown in the chapter.
Chapter 8 explores the applicability of CFD technique to predict the concentration profile,
velocity profile and pressure drop for slurry flow in pipeline. Initially different approaches of
multiphase modeling and detail steps of Eulerian model building are discussed. Later on, how
2D and 3D CFD models perform against experimental data was shown in this chapter.
Chapter 9 summarizes the contribution of the present work. The chapter ends with the future
directions and scopes of the studies in slurry flow modeling.
References
1.
Durand, R. and Condolios, G., The hydraulic transport of coal and solids in pipes,
Colloquium on Hydraulic Transport , National Coal Board, London. (1952)
2.
Kawashima, T. et al. ,Wear of pipes for hydraulic transport of solids.Proc. Hydro

transport 5 Conf. , Paper E3, BHRA, Cranfield. (1978)
3.
Rigby, G.R. Slurry pipelines for the transportation of solids, Mechanical Engineering
Transactions , Paper M1173, pp. 1819, I.E. Aust. (1982)
4.
Wasp, E.J., Kenny, J.P., Gandhi, R.L., Solid Liquid Flow Slurry Pipeline
Transportation, Trans.Tech. Publications, Clausthal, Germany. (1977).
11
Mathematical Tools
Chapter 2
Mathematical tools
Abstracts
This section describes two recent computational techniques namely artificial neural network
(ANN) and support vector regression (SVR) which have emerged as attractive tools for
nonlinear modeling especially in situations where the development of phenomenological or
conventional regression models becomes impractical or cumbersome. In the present study,
these two advanced numerical techniques are applied for slurry flow modeling and detail steps
for model building were discussed. The sole objective is to build up quick, simple and superior
correlations which can be readily used by design engineers to design slurry pipelines and
pumps.
There are some model parameters both in ANN and SVR model that has to be tuned by the
expert user during model building time. The model prediction performance greatly depends
on the optimum tuning of these parameters. Most of the earlier approaches use trial and error
procedures for tuning the ANN or SVR parameters while trying to minimize the training and
test errors. Such an approach apart from consuming enormous time may not really obtain the
best possible performance. A new approach was developed in this thesis to tune these
parameters automatically using differential evolution (DE) and genetic algorithm (GA). This
thesis has developed four new hybrid modeling technique namely 1)hybrid SVR-DE technique
2)hybrid SVR-GA technique 3)hybrid ANN-DE technique 4)hybrid ANN-GA technique and
applied them for minimizing the generalization error. The detail steps for hybrid model
building were presented in this section whereas the application of models and their
performance were discussed in subsequent chapters. The proposed hybrid techniques relieve
the non-expert users to choose the meta parameters of ANN or SVR algorithm for the case
study and find out optimum value of these meta parameters on its own.
Keywords: Artificial neural network (ANN), support vector regression (SVR), support vector
machine (SVM), differential evolution (DE), genetic algorithm (GA).
12
Chapter2
2.1 Introduction
Conventionally, two approaches namely phenomenological (first principles) and empirical, are
employed for slurry flow modeling. In phenomenological modeling, the detailed knowledge of
the solid liquid interaction and associated heat, momentum and mass transport phenomena are
required to represent mass, momentum, and energy balances. The advantages of a
phenomenological model are: (i) since it represents physico-chemical phenomenon underlying
the process explicitly, it provides a valuable insight into the process behavior, and (ii) it
possesses extrapolation ability. Owing to the complex nature of many multiphase slurry
processes, the underlying physico-chemical phenomenon is seldom fully understood. Also,
collection of the requisite phenomenological information is costly, time-consuming and
tedious, and therefore development of phenomenological process models poses considerable
practical difficulties. Moreover, nonlinear behavior being common in multiphase slurry
processes, it leads to complex nonlinear models, which in most cases are not amenable to
analytical solutions; thus, computationally intensive numerical methods must be utilized for
obtaining solutions. Difficulties associated with the construction and solution of
phenomenological models necessitate exploration of alternative modeling formalisms.
Modeling using empirical (regression) methods is one such alternative. In conventional
empirical modeling, appropriate linear or nonlinear models are constructed exclusively from
the process input-output data without invoking the process phenomenology. A fundamental
deficiency of the conventional empirical modeling approach is that the structure (functional
form) of the data-fitting model must be specified a priori. Satisfying this requirement,
especially for nonlinearly behaving processes is a cumbersome task since it involves selecting
heuristically an appropriate nonlinear model structure from numerous alternatives.
In the last decade, artificial neural networks (ANNs) and more recently support vector
regression (SVR) have emerged as two attractive tools for nonlinear modeling especially in
situations where the development of phenomenological or conventional regression models
becomes impractical or cumbersome. In the present study, these two advanced techniques are
applied for slurry flow modeling. The sole objective is to build the quick, simple and superior
correlations which can be readily used by design engineers to design slurry pipelines and
pumps.
13
Mathematical Tools
2.2 Artificial neural network
Over the past decade, neural networks have received a great deal of attention among the
scientists and engineers and they are being touted as one of the greatest computational tools
ever developed. Much of this excitement is due to the apparent ability of neural network to
emulate the brains ability to learn by examples, which in turn enables the networks to make
decisions and draw conclusions when presented with complex, noisy, and/or incomplete
information. ANN is a computer modeling approach that learns from examples through
iterations without requiring a prior knowledge of the relationships of process parameters and,
is consequently, capable of adapting to a changing environment. It is also capable of dealing
with uncertainties, noisy data, and non-linear relationships. ANN modeling have been known
as effortless computation and readily used extensively due to their model-free approximation
capabilities of complex decision-making processes. The advantages of an ANN-based model
are: (i) it can be constructed solely from the historic process input-output data (example set),
(ii) detailed knowledge of the process phenomenology is unnecessary for the model
development, (iii) a properly trained model possesses excellent generalization ability owing to
which it can accurately predict outputs for a new input data set, and (iv)even multiple inputmultiple output (MIMO) nonlinear relationships can be approximated simultaneously and
easily.
The goal of a neural network is to map a set of input patterns onto a corresponding set of
output patterns. The network accomplishes this mapping by first learning from a series of past
examples defining sets of input and output corresponding to the given system. Based on this
learning, the network then applies to a new input pattern to predict the appropriate output.
2.2.1 Background works
Application of neural networks to chemical engineering has increased significantly since 1988.
One of the first application papers was by Hoskins and Himmelblau (1988), who applied a
neural network to the fault diagnosis of chemical reactor system. Since then, the number of
research publications and network applications in bio processing and chemical engineering has
increased significantly. Baugmann and Liu (1995) provide a good overview of potential
application of neural network, as listed below:
14
Ch
hapter2
Classification
useinputvvaluestopredictacategoricaloutpu
ut;e.g.,givensymp
ptoms
andlaboratoryresults,deterrminethemostlikkelydisease.
useinputvvaluestopredictanoutputvalues;ee.g.,given
temperature,humidityandw
windvelocity,pred
dicttheevaporatio
on
rate.
Pred
diction
Assocciation
Filtering
learnassocciationsoferrorfrreeoridealdata,thenclassifyassociate
datathatcontainserror;e.g..,learnfiveidealp
patternsandthen
noisyinputpattern
nsasoneoffivepatterns.
recognizen
smoothaninputsignale.g.,ssmoothanoisyeleectocardiosignal
Optim
mization
Concepttualization
n
determineoptimalvaluee.g.,determineminim
mumlengthtripofa
travellingssalesperson.
Analyzedataanddetermineconceptualrelatio
onshipe.g.,clusterdata
nshipscanbeinferrred.
withmanyattributessothattgroupingrelation
Figu
ure 2.1: Different applicatioon of ANN
Frrom literature survey it is evvident that neuural networks hold much prromise for significant
ap
pplications to chemical engiineering probllems that are complex, nonnlinear and unncertain.
Coompared to em
mpirical and cuurve-fitting moodels, neural networks
n
are reelatively less sensitive
too noise and incoomplete inform
mation and can thus deal withh problems withh noisy data.
Beecause of the fact
f that in thiss work we are going to use neural
n
networkks in classificattion and
prrediction, the applications
a
in these
t
two areas are explored in detail from literatures.
2..2.2 Overview
w of classific
cation neurall networks
A
Applications
off neural netwoorks to classifiication problem
ms in bio proccessing and chhemical
en
ngineering falll into two major areas: (1) identificatioon of process faults based on the
15
Mathematical Tools
operating conditions of a given process and (2) prediction of the most likely categorical group
for a given input pattern, for example, identification of cell growh phase categorizing
(induction phase, growth phase, stationary phase, death phase) fermentation processes.
A partial list of various reported application includes:
Fault diagnosis on a CSTR (Venkatasubramanian, 1990)
Fault diagnosis on a chemical reactor catalytically converting heptanes to toluene

(Hoskins and Himmenblau, 1990)
Classification networks produce Boolean output responses .i.e., zero indicates that the input
patterns are not within the specific class and one indicates that it is. The actual output from the
neural network is a numerical value between 0 and 1, and can represent the probability that
the input pattern corresponds to a specific class. Classification networks used for feature
categorization activate only one output response for any input pattern and select that category
based on which output response has the highest value. In comparison, fault diagnosis networks
allow multiple faults to occur for a given set of operating conditions and can therefore activate
multiple output responses for a given input pattern. From literature survey, the radial-biasfunction network is the most frequently used network architecture for classification problems.
Radial-bias-function networks outperform back propagation networks for most of the case
studies found in literatures.
2.2.3 Overview of prediction neural networks

There is lot of literatures where neural networks are applied to predict the values of process
performance variables from independent operating variables based on laboratory or plant data
in bio processing and chemical engineering. A partial list of various reported application
includes:
Prediction of remote temperature measurements in aluminium manufacturing

(Wizzard & Fehrman,1991)
Estimation of mass transfer co-efficient in electrochemical refining of metals (

Reisner et al.,1993)
Prediction of the silicon contents in the pig iron from blast furnace data (Bulsari &
Saxen,1991)
Prediction of complex kinetics in metallurgical and mineral processing (Reuter et

al.,1993)
16
Chapter2
Prediction of overall gas holdup in bubble column reactors (Ashfaq Shaikh,

Muthanna Al-Dahhan , 2003)
Calculation of the friction factor in pipeline flow of Bingham plastic fluids (Sablania,
S.S. , Walid H.S & Kacimovc, A.,2003)
Prediction performance of a drying system ( Huang & Mujumdar,1993)
From literatures it is evident that neural networks have been very effective in predicting and
optimizing performance data from processes and analytical instrumentation that are complex
and ill defined by first principles.
2.2.4 Strengths of ANN

ANN has a number of properties that give them advantages over other computational
techniques as described below and shown in figure 2.2.
Information is distributed over a field of nodes. This distribution provides greater

flexibility than one finds in symbolic processing, where information is held in one
fixed location.
Neural networks have the ability to learn
Neural networks allow extensive knowledge indexing: Knowledge indexing is the

ability to store a large amount of information and access it easily. ANN can easily
recall, for example, diverse amounts of information associated with a chemical name,
a process, or a set of process conditions. The network stores and retains knowledge in
two forms: a) the connections between nodes and b) the weight factors of these
connections. Because it has so many interconnections, the network can index and
house large amounts of information corresponding to the interrelations between
variables.
ANN is better suited for processing noisy, incomplete or inconsistent data: No single
node within a neural network is directly responsible for associating a certain input
with a certain output. Instead, each node encodes a micro feature of the input-output
pattern. The concept of micro feature implies that each node affects the input-output
pattern only slightly, thus minimizing the effects of noisy or incomplete data in any
given node. Only when we assemble all the nodes together into a single coordinated
network, these micro features map the macroscopic input-output pattern. Other
17
Mathematical Tools
computational techniques do not include this micro feature concept. In empirical
modeling, for instance, each variable used has a significant impact in most models.
Consequently, if the value of one variable is off, the model will most likely yield
inaccurate results. In ANN, however, if the value of one variable is off, the model
will not be affected substantially.
ANN mimics human learning processes: Most human learning and problem solving
occurs by trial and error. For example, if a piece of equipment is not operating
correctly we observe its symptoms and recommend corrective actions. Based on the
results of those actions, we recommend additional corrections. ANN functions in
same fashion. We can train them by iteratively adjusting the strength of the
connections between the nodes. After numerous iterative adjustments, the network
can properly predict the cause and effect relationships.
Automated abstraction: ANN can determine the essentials of input-output

relationships automatically. We do not need a domain expert, that is, an expert in a
particular problem solving domain (e.g. Slurry specialist) to develop the knowledge
base that expert systems require. Through training with direct (and sometimes
imprecise) numerical data, the network can automatically determine cause-and-effect
relations and develop its own knowledge base.
Potential for online use: ANN may take a very long time to train, but once trained,
they can calculate results from a given input very quickly. Since a trained network
may take less than a second to calculate results, it has the potential to be used online
in a control system.
18
Ch
hapter2
abilityyto
learn
n
mimics
human
learning
processes
p
extensive
knowledge
indexing
ANN
Potential
fforonline
use
Automated
A
a
abstraction
processsing
noisyd
data
2 Strength aand characterristics of ANN

Figure 2.2:
2..2.5 Limitatio
ons of neurall networks
W
While
neural neetworks have many
m
advantagees, they are noot a cure-all. Soome of the lim
mitations
off neural networrks as follows:
Long training
tr
times: Training cann take long ennough to makee the neural network
n
impracttical. Most sim
mple problem
ms require at least
l
1000 tim
me steps to trrain the
network
k and compleex problems ccan require up
pto 75000. Hoowever with th
he new
technollogy of powerfful PC, this lim
mitations gettingg wiped out daay by day.
Large amount of traaining data: If little input-ooutput data exxist on a probblem or

processs, we may reconsider the usse of neural neetworks, since they rely heaavily on
such daata. In additionn, we may alsoo have a situattion where therre is a large daatabase,
but all the training data
d
are very ssimilar, causinng the same prroblems as in having
small training
t
data sets.
s
Thus, a bbroad based data set or exxperimental deesign is
essentiaal.
19
Mathematical Tools
No guarantee of optimal results: Most training techniques are capable of tunning
the network, but they do not guarantee that the network will operate properly. The
training may bias the network, making it accurate in some operating regions, but
inaccurate in others. In addition, one may in advertently get trapped in local
minima during training.
No guarantee of 100% reliability: While this applies to all computational

applications, this point is particularly true for neural networks with training data.
Good set of input variables: Selection of input variables that give the proper output
mapping is often difficult. It is not always obvious which input variables of those
variables (e.g. Log, inverse etc) obtain the best results. Some trial and error in
selecting input variables is often required.
2.3 Comparison of neural networks to empirical modeling

Consider the multilayer ANN shown in figure 2.3 where each layer (input, hidden and output)
has three nodes. This network has a total of eighteen connections and eighteen weight factors
to adjust during train the network. An engineer may say, Hold on here! If you give me
eighteen variables, I can curve fit almost anything. This neural network is nothing but
empirical modeling, which has been around for more than fifty years. You are just doing some
fancy curve-fitting. There is truth in that claim (Mah, 1991).
A neural network is an
empirical modeling tool and it does operate by curve fitting. However some notable
differences exist between neural networks and typical empirical models. As a result, ANN
offer distinct advantages in some areas, as explained above, but have limitations in other areas.
First, ANN has a better filtering capacity than empirical models because of the micro feature
concept as discussed earlier. Because each node encodes only a micro feature of the overall
input output pattern, it affects the input-output pattern slightly. Moreover, neural networks
are also massively parallel, so that each node operates independently. We can view each node
as a processor in its own right and these processors all operate in parallel. As a result, the
network does not depend on a single node as heavily as, for instance; an empirical model
depends on an independent variable. Because of this parallelism, ANN has a better filtering
capacity and generally performs better than empirical models with noisy or incomplete data.
Second, neural network are more adaptive than empirical models. ANN has specified training
algorithms, where we adjust weight factors between nodes until we achieve the desired input20
Chapter2
output pattern. If conditions change such that the network performance is inadequate, we can
train the neural network further under these new conditions to correct its performance. In
addition, we can design the network to periodically update its input-output performance,
resulting in a continuous, online, self correcting model. Typical empirical models do not have
this ability.Third, ANN is truly multi input and multi output (MIMO) systems. Most empirical
modeling tools map one, or at most two or three dependent variables. Neural networks can
map many independent variables with many dependent variables as needed.
2.4 Artificial neural network (ANN) based modeling

Neural networks are computer algorithms inspired by the way information is processed in the
nervous system. An ANN is a massively parallel-distributed processor that has a natural
propensity for storing experimental knowledge and making it available. An important
difference between neural networks and standard Information Technology (IT) solutions is
their ability to learn. This learning property has yielded a new generation of algorithms. An
ANN paradigm is composed of a large number of highly interconnected processing elements,
analogous to biological neurons that are tied together with weighted connections that are
analogous to synapses. Learning in biological systems involves adjustments to the synaptic
connections between the neurons. This is true for ANNs as well. Learning typically occurs
through training or exposure to a true set of input and output data where the training algorithm
iteratively adjusts the connection weights. These connection weights represent the knowledge
necessary to solve specific problem.
2.4.1 Network Architecture

The multilayer perceptron (MLP) network used in the model development is depicted in
Figure 2.3. As shown, the network usually consists of three layers of nodes. The layers
described as input, hidden and output layers, comprise N, L and K number of processing nodes
respectively. Each node in the input layer is linked to all the nodes in the hidden layer using
weighted connections. In addition to the N and L number of input and hidden nodes, the MLP
architecture also houses a bias node (with fixed output of +1) in its input and hidden layers; the
bias nodes are also connected to all the nodes in the subsequent layer and they provide
additional adjustable parameters (weights) for the model fitting. The number of nodes (N) in
21
Mathematical Tools
the MLP networks input layer is equal to the number of inputs in the process whereas the
number of output nodes (K) equals the number of process outputs. However, the number of
hidden nodes (L) is an adjustable parameter whose magnitude is determined by issues, such as
the desired approximation and generalization capabilities of the network model.
2.4.2 Training
Training a network consists of an iterative process in which the network is given the desired
inputs along with the correct outputs for those inputs. It then seeks to alter its weights to try
and produce the correct output (within a reasonable error margin). If it succeeds, it has learned
the training set and is ready to perform upon previously unseen data. If it fails to produce the
correct output it re-reads the input and again tries to produce the correct output. The weights
are slightly adjusted during each iteration through the training set (known as a training cycle)
until the appropriate weights have been established. Depending upon the complexity of the
task to be learned, many thousands of training cycles may be needed for the network to
correctly identify the training set. Once the output is correct the weights can be used with the
same network on unseen data to examine how well it performs.
2.4.3 Back propagation algorithm (BPA)

The back propagation algorithm modifies network weights to minimize the mean squared error
between the desired and the actual outputs of the network. Back propagation uses supervised
learning in which the network is trained using data for which input as well as desired outputs
are known. Once trained, the network weights are frozen and can be used to compute output
values for new input samples.
22
Ch
hapter2
Figuree 2.3: Architeccture of feed forward

f
netwoork with one hidden
h
layer
olves presentingg an input dataa to input layerr neurons that pass
p the
Thhe feed forwarrd process invo
innput values ontto the first hidden layer. Eacch of the hiddeen layer nodes computes a weighted
w
suum of its inputt and passes thhe sum throughh its activationn function and presents the result
r
to
thhe output layerr. The goal is to find a set of weights thaat minimize m
mean squared error.
e
A
tyypical back propagation algorrithm can be given as followss:
Thhe MLP netwoork is a nonlin
near function-m
mapping devicee that determinnes the K dimeensional
noonlinear function vector f whhere f: X - Y.
Y Here, X is a set of N-dimeensional input vectors
(X
X = {xp}; p = 1, 2, , P and x = [x1, x2, , xn, , xN]T ), and Y is the seet of corresponnding Kdiimensional outtput vectors (Y
Y = {yp}; p = 1,
1 2, , P wheere y = [y1, y2, , yk, , yK]T ). The
prrecise form of f is determinedd by : (i) netwoork topology, (ii)
( choice of tthe activation function
f
ussed for computting outputs off the hidden annd output nodees, and (iii) nettwork weight matrices
m
WH and WO (theey refer to the weights betweeen input and hidden
h
nodes, aand hidden andd output
noodes, respectively). Thus, thee nonlinear mappping can be exxpressed as:
23
Mathematical Tools
y = y (x; W)
[2.1]
H
Where, W = {W , W }. This equation suggests that y is a function of x, which is parametrized

by W It is now possible to write the closed-form expression of the input-output relationship
approximated by the three-layered MLP as:
~L
~ N H
y k = f 2 wlk o f1 wnl
xn ; k = 1,2,........K
n = 0
l = 0
Where
yk
functions;
[2.2]
refers to the kth network output; f 2 , f 1 denote the nonlinear activation
wlk o
th
refers to the weight between l hidden node and k
th
output node; wnl
is
the weight between nth input and lth hidden node, and xn represents the nth network input.
Note that in Eq. 2.2, the bias node is indexed as the zeroth node in the respective layer. In
order that an MLP network approximates the nonlinear relationship existing between the
process inputs and the outputs, it needs to be trained in a manner such that a prespecified error
function is minimized. In essence, the MLP-training procedure aims at obtaining an optimal
set (W) of the network weight matrices WH and WO, which minimize an error function. The
commonly employed error function is the root-mean-squared-error (RMSE) defined as:
[2.3]
Where, NP refers to the number of patterns used in the training; i denotes the index of the input
pattern (vector) and Ei represents the sum-squared-error (SSE) pertaining to the ith training
pattern.
The SSE is expressed as:
Where,
and
[2.4]
are the desired (target) and predicted values of the kth output node,
respectively.
The most widely used formalism for the error minimization is the error-back-propagation
(EBP) algorithm utilizing a gradient-descent technique known as the generalized delta rule
(GDR). In the EBP methodology, the weight matrix set, W, is initially randomized. Thereafter,
an input vector from the training set is applied to the networks input nodes and the outputs of
the hidden and output nodes are computed. The outputs are computed as follows. First the
24
Chapter2
weighted-sum of all the node-specific inputs is evaluated, which is then transformed using a
nonlinear activation function, such as the logistic sigmoid. The outputs from the output nodes
are then compared with their target values and the difference is used to compute the sum
squared error defined in Eq. 2.4. Upon SSE-computation, the weight matrices WH and WO are
updated using the GDR framework ( refer Appendix 1). This procedure when repeated with
the remaining input patterns in the training set, completes one network training iteration. For
the RMSE minimization, several training iterations are usually necessary.
2.4.3.1 Different back propagation algorithm
There are several different back propagation training algorithms published in literatures.
Figure 2.4 shows some of those algorithms used in the present study. They have a variety of
different computation and storage requirements, and no one algorithm is best suited to all
locations. The basic differences between these algorithms are how they handle the weight upgradation to reduce error and how they modify learning rate () to reduce convergence time.
The special features and comparative advantages of various algorithms as reviewed from
literature (Baugmann & Liu ,1995; matlab online documentation,2009) are summarized in
appendix 2.
2.4.4 Performance measures of ANN model
There are different measures by which ANN performance is assessed; validation and leaveone-out error estimates being the most commonly used one. Here the total available data was
divided as training data (75% of data) and test data (25 % data) chosen randomly. While ANN
algorithm was trained on training data but the ANN performance is estimated on test data.
The statistical analysis of ANN prediction is based on the following performance criteria:
1.
The average absolute relative error (AARE) on test data should be minimum
2.
[2.5]
The standard deviation of error() on test data should be minimum

25
Mathem
matical Toolss
[2.6]
3.
nd output should be around unity.

u
The cross-correlation co-eefficient (R) beetween input an
[2.7]
AN learning is
ANN
i considered successful
s
onlyy if the system
m can perform
m well on test data on
w
which
the system
m has not beenn trained.
Batchgrradient
dece
ent
Resiliient
backprop
pagation
FletcherReeves
updaate
Conjuggate
gradient
PolkaR
Ribiere
updaate
ANNalgo
orithm
PowellBeale
Restaarts
BFGSalggorithm
QuasiNeewton
algorithm
Levenb
burg
Marqu
uardt
Onestep
psecant
algorithm
nt ANN algorrithms publish
hed in various literatures
Figurre 2.4: Differen
26
Chapter2
2.4.5 Generalizability
The AARE minimization procedure by itself does not ensure that the trained network would
possess the much desired generalization ability owing to which the network also predicts
accurately the outputs corresponding to a new set of input data. To ensure that the MLP model
possesses satisfactory generalization capability, a care must be exercised to avoid what is
known as network overfitting. It occurs when: (i) the network is trained excessively (overtraining), and/or (ii) the network architecture houses more hidden nodes than necessary (overparameterization). Neural learning is considered successful only if the system can perform
well on test data on which the system has not been trained. This capability of a network is
called generalizability. Given a large network, it is possible that repeated training iterations
successively improve performance of the network on training data e.g. by memorizing
training samples, but the resulting network may perform poorly on test data (unseen data).
This phenomenon is called overtraining. The proposed solution is to constantly monitor the
performance of the network on the test data. Hecht- Nielsen (1990) proposes that the weight
should be adjusted only on the basis of the training set, but the error should be monitored on
the test set. Here we apply the same strategy: training continues as long as the error on the test
set continues to decrease and is terminated if the error on the test set increases. Training may
thus be halted even if the network performance on the training set continues to improve.
The procedure for obtaining an optimal network architecture and the corresponding optimal
weight matrix set W is described in next section (for more details refer Baugmann & Liu
,1995, Tambe et al.,1996, and Jack & Nandi,2002)
2.4.6 Step-wise Procedure for Developing an Optimal MLP Model
The stepwise procedure for obtaining an optimal network architecture and the weight matrix
set W, is presented below. The procedure ensures that the network is not over fitted.
1. Partitioning the available example input-output data into two sets, namely, training and test
sets
2. Fixing the number of nodes in MLPs hidden layer to a small number (1 or 2).
3. Initializing the network weight set, W, randomly.
4. Training the network over several iterations using input-output vectors in the training set
such that the RMSE with respect to the test set (defined as Etst, and computed at the end of
27
Mathematical Tools
each iteration) is minimized. The weight matrix that has resulted in the lowest Etst magnitude
are then Stored.
5. Repetition of steps 3 and 4 many times (say 10-15) by changing the seed value of the
random number generator. This ensures that the weights are initialized differently in each
repeated training run, which helps in exploring the error-surface rigorously and thereby
locating the deepest local or the global minimum on the error surface.
6. Going back to step 2 and repeating the steps 3 to 5 by systematic incrementing the number
of hidden nodes (L); the network architecture and the weight set, W, that result in the overall
least Etst magnitude are taken as optimal.
2.5 Tuning parameters of ANN

It is well known that ANN generalization performance (estimation accuracy) depends on a
good setting of meta-parameters listed below:
No of nodes in hidden layer: The number of nodes in hidden layer has a profound effect on
ANN performance. Too few nodes could not learn the relationship in data properly and too
large number of nodes increases the network complexity and execution time. From literature it
is found that the optimal number of nodes in hidden layer normally calculated by trial and
error method. Such an approach apart from consuming enormous time may not really obtain
the best possible performance.
The activation functions in input and output layer: Each hidden node and output node
applies the activation function to its net input. Five types of activation function reported in
literature and used in this work are shown in table 2.1 and figure 2.5
28
Chapter2
Table 2.1: Different activation function
Figure 2.5: Structure of different activation function
29
Mathematical Tools
There is no consensus in literature which type of activation function is to be used. It depends
on type of input training data and the case under investigation. For a new user it is difficult to
choose the activation function for his data as he has no guidelines to choose. Multilayer
networks typically use sigmoid transfer functions in the hidden layers. Sigmoid functions are
characterized by the fact that their slope must approach zero, as the input gets large. This
causes a problem when using steepest descent to train a multilayer network with sigmoid
functions, since the gradient can have a very small magnitude; and therefore, cause small
changes in the weights and biases, even though the weights and biases are far from their
optimal values.
The learning rate: The performance of the back propagation algorithm can be improved if we
estimate the optimal learning rate. For a new user choosing the optimal learning rate is very
difficult. The learning rate is multiplied with the negative of the gradient to determine the
changes to the weights and biases. The larger the learning rate, the bigger the step. If the
learning rate is made too large, the algorithm becomes unstable. If the learning rate is set too
small, the algorithm takes a long time to converge.
Training algorithm: Apart from above parameters the ANN performance also depends upon
the training algorithm used for back propagation. Over the years different researchers have
developed many ANN training algorithms to reduce execution time and computer storage
requirements. There are several different back propagation training algorithms published in
literatures. Figure 2.4 shows some of those algorithms used in the present study.
Most of the available ANN software require the above four parameters as user inputs or these
are calculated on trial and error basis. This needs long execution time to explore all the
combinations of above parameters to really find the best possible solutions. In the present
study genetic algorithm and differential evolution technique is used to find out the optimum
values of these parameters.
2.6 Optimization of ANN model

This study is motivated by a growing popularity of ANN for process modeling and regression
problems. However, many ANN regression application studies are performed by expert users
having good understanding of ANN methodology. Since the quality of ANN models depends
on a proper setting of ANN architecture and ANN meta-parameters, the main issue for
30
Chapter2
practitioners trying to apply ANN regression is how to set these parameter values (to ensure
good generalization performance) for a given data set. Non expert users face difficulty to find
a optimum ANN architecture and confused about how to choose the ANN meta parameters.
While existing sources on ANN regression give some recommendations on appropriate setting
of ANN parameters, there is clearly no consensus and plenty of contradictory opinions. Also
there is lot of ANN algorithms available in literatures with their respective advantages. Some
algorithms requires less computation time and computer storage requirement, some others
have more accurate prediction capability. It is difficult for a non expert users to choose the best
(robust, accurate and fast) algorithm for a case study. Present thesis addresses this issue and
develop a new hybrid procedure to find the optimum ANN architecture and tune the ANN
parameters and thus relieve the non expert users.
Basically the setting of optimum ANN architecture and tuning of ANN meta parameters can
be viewed mathematically as a optimization problem where test set errors (generalization
error) has to be minimized. In commonly used deterministic optimization techniques, solution
to an optimization problem is represented in the form of a vector consisting of values of
decision variables at which gradient of the objective function with respect to the decision
variables becomes zero. Thus, gradient computation is an integral feature of such optimization
paradigms. Additionally, most gradient-based techniques require the objective function to be
smooth, continuous and differentiable. In the case of an ANN, it is possible to express the
nonlinear mapping that it executes, in terms of a generic closed-form function. It may be noted
that the nonlinear mapping ability of ANNs is due to the nonlinear activation function used for
computing the node-specific outputs. For computing an output, the nonlinear activation
function makes use of the arguments comprising a number of network parameters (weights)
and node-specific inputs. Consequently, the mapping executed by an ANN attains complex
nonlinear character that cannot be guaranteed to simultaneously fulfill the smoothness,
continuity and differentiability criteria for the objective function. This feature of ANN models
poses difficulties in using the conventional deterministic techniques for optimizing their input
space. Thus, the gradient-based methods cannot be efficiently used for optimizing the input
space of an ANN model and, therefore, it becomes necessary to explore alternative
optimization formalisms, which are lenient towards the form of the objective function.
In the recent years, Genetic Algorithms (GAs) and Differential Evolution (DE) that are
members of the stochastic optimization formalisms have been used with a great success in
solving problems involving very large search spaces. The GAs and DE were originally
developed as the genetic engineering models mimicking population evolution in natural
31
Mathematical Tools
systems. Specifically, both GAs and DE enforce the survival-of the-fittest and genetic
propagation of characteristics principles of biological evolution for searching the solution
space of an optimization problem. The principal features possessed by GAs and DE are: (i)
they require only scalar values and not the second and/or first-order derivatives of the
objective function, (ii) capability to handle nonlinear and noisy objective functions, (iii) they
perform global search and thus are more likely to arrive at or near the global optimum, and (iv)
GAs and DE do not impose pre-conditions, such as smoothness, differentiability and
continuity, on the form of the objective function. Owing to these attractive features, both GAs
and DE are being increasingly used for solving diverse optimization problems in chemical
engineering.
In the present thesis, ANN and GA or ANN and DE formalisms are integrated to arrive at a
hybrid process modeling and optimization strategy. The strategy (henceforth referred to as
ANN-GA or ANN-DE) uses an ANN as the process modeling paradigm, and the GA or
DE for optimizing the input space of the ANN model such that an improved prediction is
realized.
2.7 Genetic algorithm

Most of the traditional optimization techniques based on gradient methods have the possibility
of getting trapped at local optimum depending upon the degree of non-linearity and initial
guess. Hence, these traditional optimization techniques do not ensure global optimum and also
have limited applications. In the recent past, non-traditional search and optimization methods
based on natural phenomena (evolutionary computation) such as simulated annealing
(SA)(Kirkpatrick et al.,1983), Genetic Algorithms (Goldberg, 1989) and differential
evolution(DE)(Price & Storn,1997) have been developed to overcome this problem. These
algorithms are stochastic in nature with probabilistic transition rules. These are found to have a
better global perspective than the traditional methods (Deb 2000). Genetic Algorithms (GA)
are computerized search and optimization algorithms based on the mechanics of natural
genetics and natural selection. They mimic the survival of the fittest principle of nature to
make a search process.
2.7.1 Literature survey of genetic algorithm
32
Chapter2
Professor John Holland of the University of Michigan, Ann Arbor envisaged the concept of
genetic algorithm way back in 1960s (Holland, 1975). In late 1980s further developments
were made and the methodology was conceptualized (Goldberg, 1989). Genetic algorithms are
unique by themselves and stand out from the traditional methods of optimization. These are
closely related to the natural selection and the survival of the fittest theories of Charles
Darwin. There are certain striking features of GA that establishes its credibility. The most
essential feature is that GAs work with a population of points instead of a single point. This
enables the search to come up with a region of global solutions. Secondly, the initial
population is chosen randomly. Therefore the search can proceed in any direction. GAs use
probabilistic transition rules rather than deterministic procedures (Goldberg, 1989). Hence the
search will be multidirectional.
Since their inception, GAs have evolved like the species they try to mimic and have been
applied successfully in many diverse fields. The various applications of GAs are: process
design and optimization (Androulakis & Venkatasubramanian, 1991), computer-aided
molecular design (Venkatasubramanian et al., 1994), optimal design of ammonia synthesis
reactor (Upreti & Deb, 1996), estimation of heat transfer parameters in trickle bed reactors
(Babu & Vivek, 1999) etc. Apart from these, there are many applications in different
disciplines ( Goldberg 1989; Deb2000).
2.7.2 Techniques used in GA
For the implementation of GAs, there are certain well defined steps.
Coding
Coding is the method by which the variables xi are coded into string structures. A linear
mapping rule is used for the purpose of coding:
(decoded value)
[2.8]
The decoded value (si, the binary digit in the coding) is given by the rule,
where
0,1
[2.9]
For example, for a binary code (0111), the decoded value will be
(0111)=(1)20+(1)21+(1)22+(0)23=7
[2.10]
l
It is also evident that for a code of length l, there are 2 combinations or codes possible. The
length of the code depends upon the required accuracy for that variable.
33
Mathematical Tools
[2.11]
So, adhering to the above mentioned rules, it is possible to generate a number of guesses, or, in
other words, an initial population of coded points that lie in the given range of the function.
Next step is calculation of fitness.
Fitness
As already been mentioned, GAs work on the principles of survival of the fittest. This in
effect means that the good points or the points that yield maximum values for the function are
allowed to continue in the next generation, while the less profitable points are discarded from
calculations. Depending upon whether the initial objective function needs to be maximized or
minimized, the fitness function is hence defined in the following ways:
F(x)=f(x) for a maximization problem
[2.12]
F(x)=1/(1+f(x)) for a minimization problem
[2.13]
The fitness function value for a particular coded string is known as the strings fitness. This
fitness value is used to decide whether a particular string carries to the next generation or not.
GA operation begins with a population of random strings. These strings are selected from a
given range of the function and represent the design or decision variables. To implement our
optimization routine, three operations are carried out;
1.
Reproduction
2.
Crossover
3.
Mutation
Operators in GA
Reproduction
The reproduction operator is also called the selection operator. This is because it is this
operator that decides the strings to be selected for the next generation. The end result of this
operation is the formation of a mating pool, where the above average strings are copied in a
probabilistic manner. The rule can be represented as,
The probability of selection of the ith string into the mating pool is given by
[2.14]
34
Chapter2
Where, Fi is the fitness of the ith string. Fj is the fitness of the jth string and n is the population
size.
The average fitness of all the strings is calculated by summing the fitness of individual strings
and dividing by the population size, and is represented by the symbol
[2.15]
It is obvious that the string with the maximum fitness will have the most number of copies in
the mating pool. This is implemented using roulette wheel selection. The algorithm of this
procedure is as follows:
Step 1: using Fi, calculate pi.
Step 2; calculate the cumulative probability Pi.
Step 3: Generate n random numbers (between 0 and 1)
Step 4: copy the string that represents the chosen random number in the cumulative probability
range into the mating pool. A string with higher fitness will have a larger range in cumulative
probability and so has more probability of getting copied into the mating pool.
At the end of this implementation, all the strings that are fit enough would have been copied
into the mating pool and this marks the end of the reproduction operation.
Crossover
After the selection operator has been implemented, there is a need to introduce some amount
of randomness into the population in order to avoid getting trapped in local searches. To
achieve this, cross over operation was performed. In the crossover operation, new strings are
formed by exchange of information among strings of the mating pool. For example,
00|000
00|111
11|111
11|000
Parents
Children
Crossoverpoint
Strings are chosen at random and a random crossover point is decided. Crossover is performed
in the method shown above. It is evident that, using this method, better strings or worse strings
may be formed. If worse strings are formed, then they will not survive for long, since
reproduction will eliminate them. But what if majority of the new strings formed are worse?
This undermines the purpose of reproduction. To avoid this situation, we do not select all the
strings in population for crossover. We introduce a crossover probability (pc). Therefore,
(100pc) % of the strings are used in crossover. (1-pc)% of the strings is not used in crossover.
35
Mathematical Tools
Through this we have ensured that some of the good strings from the mating pool remain
unchanged. The procedure can be summarized as follows;
Step 1: Select (100pc) % of the strings out of the mating pool at random.
Step 2: Select pairs of strings at random
Step 3: Decide a crossover point in each pair of strings ( again this done by a random number
generation over the length of the string and the appropriate position is decided according to the
value of the random number)
Step 4: Perform the crossover on the pairs of the strings by exchanging the appropriate bits.
Mutation
Mutation involves making changes in the population members directly, that is, by flipping
randomly selected bits in certain strings. The aim of mutation is to change the population
members by a small amount to promote local searches when the optimum is nearby. Mutation
is performed by deciding a mutation probability pm and selecting strings, on which mutation is
performed, at random. The procedure can be summarized as follows:
Step1: calculate the approximate number of mutations to be performed by
Approximate number of mutations=n/pm
[2.16]
Step 2: generate random numbers to decide whether mutation is to be performed on a

particular population member or not. This is decided by a coin toss. That is, select a random
number range to represent true and one to represent false. If the outcome is true, perform
mutation, if false, do not.
Step 3: if the outcome found in step 2 is true for a particular population member, then generate
another random number to decide the mutation point over the length of the string. Once
decide, flip the bit corresponding to the mutation point.
With the end of mutation, the strings obtained represent the next generation. The same
operations are carried out on this generation until the optimum value is encountered.
2.8 GA-based Optimization of ANN Models

The objective function and the optimal problem of ANN model of the present study are
represented as:
Minimize AARE(X) on test set

36
Chapter2
X {x1, x2, x3, x4, x5}
[2.17]
Where,
x1 = Number of nodes in hidden layer {1,2 ., 100}
x2 = Input layer activation function {1,2,3,4 and 5 corresponds to five activation function in
table 2.1}
x3 = Output layer activation function {1,2,3,4 and 5 corresponds to five activation function in
table 2.1}
x4 = Learning rate{0 to 5}
x5 = Training algorithm {1,2,..,8 corresponds to eight training algorithm as per figure 2.4}
The objective function is minimization of average absolute relative error (AARE) on test set
and X is a solution string representing a design configuration of ANN architecture. The design
variable x1 takes any integer values for number of nodes in the range of 1 to 100, x2 represents
the Input layer activation function taking any values in the range of 1 to 5 corresponds to five
activation function in table 2.1. X3 represents the output layer activation function taking any
values in the range of 1 to 5 corresponds to five activation function in table 2.1. X4 represents
learning rates and can take any value between 0 to 5. The variable x5 takes eight values of the
training algorithm which corresponds to eight ANN training algorithm in figure 2.4.
The total number of design combinations with these variables is 100 x 5 x 5 x 5 x 8 = 100,000.
This means that if an exhaustive search is to be performed it will take at the maximum 105
function evaluations before arriving at the global minimum AARE for the test set ( assuming 5
trials to arrive at optimum learning rate). So, the strategy which takes few function evaluations
is the best one. Considering minimization of AARE as the objective function, genetic
algorithm technique is applied to find the optimum design configuration of ANN model. The
methodology adopted is shown in figure 2.6. In the GA procedure, the search for an optimal
solution (decision) vector, x*, begins from a randomly initialized population of probable
(candidate) solutions. The solutions, usually coded in the form of binary strings
(chromosomes), are then tested to measure their fitness in fulfilling the optimization objective.
Subsequently, a main loop comprising following operations is performed: (i) selection of
better (fitter) parent chromosomes, (ii) production of an offspring solution population by
crossing over the genetic material between pairs of the fitter parent chromosomes, and (iii)
mutation (bit-flipping) of the offspring strings. Implementation of this loop generates a new
population of candidate solutions, which as compared to the previous population, usually fares
37
Mathem
matical Toolss
Figurre 2.6: Schemaatic for hybrid

d ANN-GA alggorithm impleementation
38
Chapter2
better at fulfilling the optimization objective. The best string that evolves after repeating the
above described loop till convergence forms the solution to the optimization problem. The
stepwise procedure for the GA-based optimization of a ANN model is provided in Appendix 3
(also see the flowchart in Figure 2.6).
2.9 Differential evolution

DE is a generic name for a group of algorithms that are based on the principles of GA but have
some inherent advantages over GA. Differential Evolution (DE), an improved version of GA,
is an exceptionally simple evolution strategy that is significantly faster and robust at numerical
optimization and is more likely to find a functions true global optimum. Unlike simple GA
that uses a binary coding for representing problem parameters, DE uses real coding of floating
point numbers. The mutation operator here is addition instead of bit-wise flipping used in GA.
DE uses non-uniform crossover and tournament selection operators to create new solution
strings. Among the DEs advantages are its simple structure, ease of use, speed and robustness.
It can be used for optimizing functions with real variables and many local optima.
2.9.1 Literature survey of Differential evolution
In the recent years, Differential Evolution (DEs) that are members of the stochastic
optimization formalisms have been used with a great success in solving problems involving
very large search spaces. The DEs were originally developed as the genetic engineering
models mimicking population evolution in natural systems. Specifically, DE like genetic
algorithm (GA) enforce the survival-of the- fittest and genetic propagation of
characteristics i.e. principles of biological evolution for searching the solution space of an
optimization problem. DE has been used to design several complex digital filters (Price and
Storn, 1997) and to design fuzzy logic controllers (Sastry et al., 1998). DE can also be used for
parameter estimations e.g. Babu and Sastry (1999) used DE for the estimation of effective heat
transfer parameters in trickle-bed reactors using radial temperature profile measurements.
They concluded that DE takes less computational time to converge compared to the existing
techniques without compromising with the accuracy of the parameter estimates.
39
Mathematical Tools
2.9.2 Steps performed in DE
This paper demonstrates the successful application of Differential Evolution to the practical
optimization problem. As on principal DE is similar, we use a population of points in the
search for the optimum like GA. The population size is denoted by NP. The dimension of each
vector is denoted by D. The main operation is the NP number of competitions that are to be
carried out to decide the next generation.
To start with, we have a population of NP vectors within the range of the objective function.
We select one of these NP vectors as our target vector. We then randomly select two vectors
from the population and find the difference between them (vector subtraction). This difference
is multiplied by a factor F (specified at the start) and added to third randomly selected vector.
The result is called the noisy random vector. Subsequently, crossover is performed between
the target vector and noisy random vector to produce the trial vector. Then, a competition
between the trial vector and target vector is performed and the winner is replaced into the
population. The same procedure is carried out NP times to decide the next generation of
vectors. This sequence is continued till some convergence criterion is met. This summarizes
the basic procedure carried out in differential evolution. The details of this procedure are
described below.
Assume that the objective function is of D dimensions and that has to be optimized. The
weighting constants F and the crossover constant CR are specified.
Step 1 Generate NP random vectors as the initial population: Generate (NP * D) random
numbers and linearize the range between 0 and 1 to cover the entire range of the function.
From these (NP * D) numbers, generate NP random vectors, each of dimension D, by mapping
the random numbers over the range of the function.
Step 2 Choose a target vector from the population of size NP: First generate a random number
between 0 and 1. From the value of the random number decide which population member is to
be selected as the target vector (Xi) (a linear mapping rule can be used).
Step3 Choose two vectors at random from the population and find the weighted difference:
Generate two random numbers. Decide which two population members are to be selected (Xa,
Xb). Find the vector difference between the two vectors (Xa - Xb). Multiply this difference by
F to obtain the weighted difference.
Weighted difference = F (Xa - Xb)
[2.18]
40
Chapter2
Step 4 Find the noisy random vector: Generate a random number. Choose a third random
vector from the population (Xc). Add this vector to the weighted difference to obtain the noisy
random vector (Xc).
Step 5 Perform crossover between Xi and Xc to find Xt , the trial vector: Generate D random
numbers. For each of the D dimensions, if the random number is greater than CR, copy the
value from Xi into the trial vector; if the random number is less than CR, copy the value from
Xc into the trial vector.
Step 6 Calculate the cost of the trial vector and the target vector: For a minimization problem,
calculate the function value directly and this is the cost. For a maximization problem,
transform the objective function f(x) using the rule F(x) = 1/ [1 + f(x)] and calculate the value
of the cost. Alternatively, directly calculate the value of f(x) and this yields the profit. In case
cost is calculated, the vector that yields the lesser cost replaces the population member in the
initial population. In case profit is calculated, the vector with the greater profit replaces the
population member in the initial population.
Step 1 6 are continued till some stopping criterion is met. This may be of two kinds; one may
be some convergence criterion that states that the error in the minimum or maximum between
two previous generations should be less than some specified value. The other may be an upper
bound on the number of generations. The stopping criterion may be a combination of the two.
Either way, once the stopping criterion is met, the computations are terminated.
Choosing DE key parameters NP, F, and CR is seldom difficult and some general guidelines
are available. Normally, NP ought to be about 5 to 10 times the number of parameters in a
vector. As for F, it lies in the range 0.4 to 1.0. Initially F = 0.5 can be tried then F and/or NP is
increased if the population converges prematurely. A good first choice for CR is 0.1, but in
general CR should be as large as possible (Price and Storn, 1997). Already, DE has been
successfully applied for solving several complex problems and is now being identified as a
potential source for accurate and faster optimization.
2.10 DE-based Optimization of ANN Models

The DE based optimization of ANN models is almost same as used in GA. The objective
function and the optimal problem of ANN model of the present study are represented as:
41
Mathematical Tools
Minimize AARE(X) on test set
X {x1, x2, x3, x4, x5}
[2.19]
Where,
x1 = Number of nodes in hidden layer {1,2 ., 100}
x2 = Input layer activation function {1,2,3,4 and 5 corresponds to five activation function in
table 2.1}
x3 = Output layer activation function {1,2,3,4 and 5 corresponds to five activation function in
table 2.1}
x4 = Learning rate{0 to 5}
x5 = Training algorithm {1,2,..,8 corresponds to eight training algorithm as per figure 2.4}
The objective function is minimization of average absolute relative error (AARE) on test set
and X is a solution string representing a design configuration of ANN architecture. The design
variable x1 takes any integer values for number of nodes in the range of 1 to 100, x2 represents
the Input layer activation function taking any values in the range of 1 to 5 corresponds to five
activation function in table 2.1. x3 represents the output layer activation function taking any
values in the range of 1 to 5 corresponds to five activation function in table 2.1. x4 represents
learning rates and can take any value between 0 to 5. The variable x5 takes eight values of the
training algorithm which corresponds to eight ANN training algorithm in figure 2.4.
The total number of design combinations with these variables is 100 x 5 x 5 x 5 x 8 = 100,000.
This means that if an exhaustive search is to be performed it will take at the maximum 105
function evaluations before arriving at the global minimum AARE for the test set (assuming 5
trials to arrive at optimum learning rate). So the strategy which takes few function evaluations
is the best one. Considering minimization of AARE as the objective function, differential
evolution technique is applied to find the optimum design configuration of ANN model. The
methodology adopted is shown in figure 2.7.
42
Chapter2
Figure 2.7: Schematic for hybrid ANN-DE algorithm implementation
43
Mathematical Tools
2.11 Support Vector Machines

The problem of empirical data modeling is germane to many engineering applications. In
empirical data modeling a process of induction is used to build up a model of the system, from
which it is hoped to deduce responses of the system that have yet to be observed. Ultimately
the quantity and quality of the observations govern the performance of this empirical model.
By its observational nature data obtained is finite and sampled; typically this sampling is nonuniform and due to the high dimensional nature of the problem the data will form only a sparse
distribution in the input space. Traditional neural network approaches have suffered
difficulties with generalization, producing models that can over fit the data. This is a
consequence of the optimization algorithms used for parameter selection and the statistical
measures used to select the best model. The foundations of Support Vector Machines (SVM)
have been developed by Vapnik (1995) and are gaining popularity due to many attractive
features, and promising empirical performance. The formulation embodies the Structural Risk
Minimization (SRM) principle, which has been shown to be superior, (Gunn et al., 1997), to
traditional Empirical Risk Minimization (ERM) principle, employed by conventional neural
networks. SRM minimizes an upper bound on the expected risk, as opposed to ERM that
minimizes the error on the training data. It is this difference which equips SVM with a greater
ability to generalize, which is the goal in statistical learning. SVMs were developed to solve
the classification problem, but recently they have been extended to the domain of regression
problems (Vapnik et al., 1996). The salient features of SVR are: (i) like ANNs, SVR is an
exclusively data based nonlinear modeling paradigm, (ii) SVR based models are based on the
principle of structural risk minimization, which equips them with greater potential to
generalize, (iii) parameters of an SVR model are obtained by solving a quadratic optimization
problem, (iv) the objective function in SVR being of quadratic form, it possesses a single
minimum thus avoiding the heuristic procedure involved in locating the global or the deepest
local minimum on the error surface, and (v)in SVR, the inputs are first nonlinearly mapped
into a high dimensional feature space wherein they are correlated linearly with the output.
Although the foundation of the SVR paradigm was laid down in mid 1990s, its chemical
engineering applications such as fault detection (Jack & Nandi,2002) have emerged only
recently.
44
Chapter2

Support vector approach originally proposed by Glucksman (1966) and Vapnik (1995) and
later developed and popularized by number of authors is also known universally as feed
forward networks (Schlkopf, Burges & Vapnik, 1995; Cortes & Vapnik, 1995; Burges, 1998;
Schlkopf et al., 1999; Cristianini & Shawe-Taylor, 2000). It was stated earlier that SVMs
employ structural risk minimization (SRM) principle. This principle is based on the fact that
the error rate of a learning machine on test data (i.e., the generalization error rate) is bounded
by the sum of the training error rate. A term that depends on Vapnik-Chervonenkis (VC)
dimension. VC dimension is a measure of complexity of the decision space. It facilitates
quantitative means of discriminating between the capacities of different classifiers. For nonlinear and non-separable problems, support vector methodology provides a decision space with
a minimal VC dimension and training error so that the classifier has a low probability of
generalization errors. The SVM classification methodology and theory is well documented and
details can be found in number of research papers and technical reports (Gunn et al., 1997;
Burges, 1998; Cristianini and Shawe-Taylor, 2000).
Support vector machines (SVMs) have been successfully applied to a number of applications
such as handwriting recognition, particle identification (e.g., muons), digital images
identification (e.g., face identification), text categorization, bioinformatics (e.g., gene
expression), function approximation and regression, and database marketing, and so on.
Although SVMs have become more widely employed to forecast time-series data (Tay &
Cao,2001; Cao, 2003; Kim, 2003) and to reconstruct dynamically chaotic systems (Mller et
al., 1997; Mukherjee, Osuna, & Girosi,1997; Mattera & Haykin, 1999; Kulkarni, Jayaraman,
& Kulkarni,2003), a highly effective model can only be built after the parameters of SVMs
carefully determined (Duan, Keerthi, & Poo, 2003). Min and Lee (2005) stated that the
optimal parameter search on SVM plays a crucial role in building a prediction model with high
prediction accuracy and stability. The kernel-parameters are the few tunable parameters in
SVMs controlling the complexity of the resulting hypothesis (Cristianini, Campell, & Taylor,
1999). They pointed out that selecting the optimal degree of a polynomial kernel is critical to
ensure good generalization of the resulting support vector machine model. They proposed an
automatic selection for determining the optimal degree of polynomial kernel in SVM by
45
Mathematical Tools
Bayesian and Laplace approximation method and a rule based meta-learning approach. In
addition, to construct an efficient SVM model with RBF kernel, two extra parameters: (a)
sigma squared and (b) gamma, have to be carefully predetermined. However, few studies have
been devoted to optimizing the parameter values of SVMs. Evolutionary algorithms often have
to solve optimization problems in the presence of a wide range of problems (Dastidar,
Chakrabarti, & Ray, 2005; Shin, Lee, Kim, & Zhang, 2005; Yaochu & Branke, 2005; Zhang,
Sun, & Tsang, 2005). In these algorithms, genetic algorithms (GAs) have been widely and
successfully applied to various types of optimization problems in recent years (Goldberg,
1989; Fogel, 1994; Cao, 2003; Alba & Dorronsoro, 2005; Aurnhammer & Tonnies, 2005;
Venkatraman & Yen, 2005).
2.13 The basic idea behind SVM modeling

This technique can be subdivided into two distinct parts:
Learning: consists of training SVM with examples at its disposition
Prediction: where new samples are inserted in which the result is not known.
Each example can be written as a pair (input, output) where input is the data set and output
makes up how they should be catalogued. Mathematically, examples can be regarded as a pair
(x, y) where x is a vector of real numbers and output can be a Boolean value (like yes/no, 1/-1,
true/false), or a real number. In the first case, we speak of a problem of classification and latter
a problem of regression.
For conventional purposes, the group of examples that make up the SVM is called Training
Set, whereas the group which contains the examples for prediction is called Test Set. This
name derives from the fact that generally, error is calculated based on the total results in the
test and this in turn measures the quality of the system.
2.13.1 Support vector machine for classifications
In simple terms, the SVM for classification can be thought of as creating a line, or hyper plane
between two sets of data. If we imagine a two-dimensional case, the action of the SVM can be
shown easily. In Figure 2.8, a series of data points for two different classes of data are shown
as green circles (class A) and brown squares (class B). The SVM attempts to place a linear
boundary between the two different classes, and orient it in such a way that the margin
(represented by the dotted lines) is maximized. In other words, the SVM tries to orient the
46
Ch
hapter2
booundary in succh a way as to

o ensure that thhe distance bettween the bounndary and the nearest
daata point in eacch class is max
ximal. The bouundary is then placed in the middle
m
of this margin
beetween the two
o points. The nearest
n
data pooints are used to
t define the m
margin, and aree known
ass support vectors (representeed by the greey circles and square). Oncee the support vectors
seelected, the resst of the featurre set is not reqquired becausee the support vectors
v
containn all the
innformation need
ded to define thhe classifier.
Figuree 2.8: Separatiion of two classses by SVM
47
Mathematical Tools
2.13.2 Mathematics behind SVM algorithm for classification
Mathematically, SVM can be defined comparatively easily. The explanation that follows is an
overview of the functioning of an SVM (Vapnik, 1998; Vapnik, 1995; Scholkopf, 1999;
Cherkassy, 1999; Agarwal, 2003). For any point that lies on our boundary line, we can write
( w.x ) + b = 1
[2.20]
Where, w is a vector that defines the boundary, x is an input, or data vector, and b is a scalar
threshold value. At the margins, H1 and H2, where the support vectors are situated, the
equations for class A and B, respectively, are
( w.x ) + b = 1 and ( w.x) + b = 1
[2.21]
w and b may have to be scaled for this to be the case. Therefore, as the support vectors
correspond to the extremities of the data for a given class, anything that belongs to class A will
conform to the equation,
( w.x) + b 1 and ( w.x) + b 1
[2.22]
respectively, for class B. Combining these two functions, a decision function can be created to
determine whether a given data point belongs to class A or B. This is defined as,
f ( x) = sign(( w.x) + b)
[2.23]
To find the best boundary between our data sets, we tried to find a solution of w which allows
this. It can be shown that the solution is of the general form,
n
w=
v x
i i
[2.24]
i =1
where xi are the support vectors that have been kept from training.
48
Chapter2
Substituting this into equation (2.23), we get

f ( x) = sign (
v ( x.x ) + b)
i
[2.25]
A case has been constructed for a linear boundary in two dimensions. However, there will be
cases where the linear boundary in input space will be unable to separate two classes properly.
At first this seems a large problem. However, by transforming the data into a higherdimensional space, it is possible to create a hyper plane that allows linear separation in the
higher dimension (which corresponds to a curved surface in the lower dimensional input
space, refer figure 2.9). In SVMs, this is achieved through the use of a transformation, called
Kernel, which transforms the data from an N-dimensional input space into Q-dimensional
feature space:
s = (x)
[2.26]
Figure 2.9 shows the possible effect on some fictitious feature set, and how the separability of
the data changes after the transformation. Substitution of the transformation into equation
(2.25) gives,
f ( x) = sign(
v ( ( x) ( x )) + b)
i
[2.27]
Making transformations into higher-dimensional space in order to perform a dot product on the
results is relatively intensive computationally. A kernel can be used to perform this
transformation and the dot product in one step if the transformations that allowed are restricted
to those which can be replaced by an equivalent kernel function. In this way, it is possible to
reduce the computational load, but retain the effect of the higher-dimensional transformation.
The kernel function is defined as,
K ( x, y ) = ( x) ( y )
[2.28]
Substituting into equation (2.27), we get the final basic form of the SVM:
49
f ( x ) = sign (
Mathematical Tools
viK ( x, xi ) + b)
[2.29]
vi is used as a weighting factor to determine which of the vectors in the input are actually
support vectors i.e. 0 < vi < .Input vectors with a corresponding vi equal to 0 or 1 are not
support vectors, and may be for all intents and purposes, discarded. As an illustration, one of
the Kernel used for the experiments in this thesis is the RBF Kernel, defined by the equation,
K ( x, y ) = e
x y
2 2
[2.30]
The width of the RBF kernel parameter, , can be determined by an iterative process that
selects the optimum width on the basis of the full feature sets. The other kernel types and their
equation are given in Table 2.2. The kernel type was chosen after a number of empirical
experiments, and finally that Kernel was chosen which offer the best performance on data
under study.
The introduction above covers the case of separable data. For non-separable data the case is
slightly different. The values of vi, the support vectors, can be thought of as indicating the
degree of importance that each support vector has to the overall boundary definition. By
limiting the values that vi can take, this therefore limits the effect that any one support vector
has on the overall solution, and can be used to control the generalization properties of the
SVM, by reducing the effect that outliers have on the boundary defined by the support vectors.
For the non-separable case therefore, the constraint is 0 < vi < C .
For a separable case, C can be thought of as equal to infinity, while for a non-separable case,
the value of C may be varied, depending upon the number of allowable errors in the trained
solution. Few errors are permissible for high C values, while low C values will allow a higher
proportion of errors in the solution, although with a comparatively smooth boundary between
the classes.
Thus, to control the generalization capability of the SVM there are a few free parameters; the
limiting term C (also known as the misclassification penalization term), and the kernel
parameters which can be controlled by the user. The number of different kernel parameters is
dependent on the type of kernel in use. The free parameters are set at the beginning of the
training run, and can be used to alter the performance of the SVM as desired.
50
Ch
hapter2
n-linear transfformation from

m input to a higher-dimens
h
sional feature space
Fiigure 2.9: Non
2..13.3 Training, testing an

nd Generaliza
ability
Trraining a suppoort vector macchine consists oof an iterative process in whhich the SVM is
i given
thhe desired inpu
uts along with the correct ouutputs for thosse inputs. It tthen seeks to alter
a
its
m
margin
(w) andd bias (b) to try
t and produce the correctt output (withiin a reasonablle error
m
margin).
If it su
ucceeds, it has learned the traaining set and is ready to perrform upon preeviously
unnseen data. Iff it fails to pro
oduce the correect output, it re-reads
r
the inpput and again tries to
prroduce the corrrect output. Thhe margins andd bias are slighhtly adjusted thhrough the trainning set
fo
or each iteration (known as a training cyclee) until the apppropriate marggins and bias has
h been
esstablished. Deepending uponn the complexiity of the taskk to be learnedd, many thousands of
traaining cycles may
m be needeed for the SVM
M to correctly identify the ttraining set. Once
O
the
ouutput is correctt, the margins (w)
( and bias (bb) can be used with the samee SVM on unseeen data
too examine how
w well it perforrms. SVM leaarning is consid
dered successfful only if the system
51
Mathematical Tools
can perform well on test data on which the system has not been trained. This capability of a
SVM is called generalizability.
2.13.4 Mathematics behind SVM algorithm for regression
Consider a training data set g = {(x1, y1), (x2, y2), (xP, yP )}, such that xi N is a vector of
input variables and yi is the corresponding scalar output (target) value. Here, the
modeling objective is to find a regression function, y = f (x), such that it accurately predicts the
outputs {y} corresponding to a new set of input-output examples, {(x, y)}, which are drawn
from the same underlying joint probability distribution as the training set. To fulfill the stated
goal, support vector regression (SVR) considers the following linear estimation function.
,
[2.31]
Where, w denotes the weight vector; b refers to a constant known as bias; f(x) denotes a
,
function termed feature, and
represents the dot product in the feature space,
l , such
that : x l, w l . The basic concept of support vector regression is to map nonlinearly the
original data x into a higher dimensional feature space and solve a linear regression problem in
this feature space.
The regression problem is equivalent to minimize the following regularized risk function:
[2.32]
Where
,
0,
[2.33]
Equation 2.33 is also called -insensitive loss function. This function defines a -tube. If the
predicted value is within the -tube, the loss is zero. If the predicted value is outside the tube,
the loss is equal to the magnitude of the difference between the predicted value and the radius
of the tube. is a precision parameter representing the radius of the tube located around the
regression function (Fig. 2.10) and the region enclosed by the tube is known as -intensive
zone.
52
Ch
hapter2
d
of su
upport vectorr regression u
using -sensitiive loss
Fiigure 2.10: A schematic diagram
fu
unction
2
By
Thhe SVR algoriithm attempts to position thee tube around the data as shhown in Fig. 2.10.
suubstituting the e-insensitive
e
looss function into Eq. (2.32), the
t optimizatioon object becom
mes:
M
Minimize
[2.34]
W the constraaints,
With
Su
ubject to
[2.35]
W
Where,
the consstant C > 0 stan
nds for the pennalty degree off the sample with
w error exceeeding .
Tw
wo positive slack
s
variablees
reprresent the disstance from aactual values to the

53
Mathematical Tools
corresponding boundary values of -tube. The SVR fits f(x) to the data in a manner such that:
(i) the training error is minimized by minimizing
and, (ii) w 2 is minimized to increase
the flatness of f(x) or to penalize over complexity of the fitting function. A dual problem can
then be derived by using the optimization method to maximize the function,
Maximize
,
[2.36]
Subject to
0 and 0
,
Where,
[2.37]
are Lagrange multipliers. Owing to the specific character of the above-described
quadratic programming problem, only some of the coefficients, (i*- i) are non-zero and the
corresponding input vectors, xi, are called support vectors (SVs). The SVs can be thought of as
the most informative data points that compress the information content of the training set. The
coefficients and * have an intuitive interpretation as forces pushing and pulling the
regression estimate f(xi) towards the measurements, yi.
The SVM for function fitting obtained by using the above-mentioned maximization function is
then given by,
[2.38]
As for the nonlinear cases, the solution can be found by mapping the original problems to the
linear ones in a characteristic space of high dimension, in which dot product manipulation can
be substituted by a kernel function, i.e.
function is used in the SVR. Substituting
. In this work, different kernel

,
in Equation 2.36 allows us
to reformulate the SVM algorithm in a nonlinear paradigm. Finally, we have,
[2.39]
54
Chapter2
2.14 Tuning parameters of SVR

There are different measures by which SVM performance is assessed, but validation and
leave-one-out error estimates are the most commonly used one. Here we divide the total
available data as training data (75% of data) and test data (25 % data) chosen randomly on
which the SVR performance is assessed.
The statistical analysis of SVR prediction is based on the following performance criteria:
1.The average absolute relative error (AARE) on test data should be minimum
[2.40]
2.The standard deviation of error() on test data should be minimum
[2.41]
3.The cross-correlation co-efficient (R) between input and output should be around unity.
[2.42]
SVR learning is considered successful only if the system can perform well on test data on
which the system is not trained. It is well known that SVM generalization performance
(estimation accuracy) depends on a good setting of meta-parameters C,
and kernel
parameters such as kernel type and loss function type. The problem of optimal parameter
selection is further complicated by the fact that SVM model complexity (and hence its
generalization performance) depends on all five parameters..
Selecting a particular kernel type and kernel function parameters is usually based on
application-domain knowledge and also should reflect distribution of input (x) values of the
55
Mathematical Tools
training data. Parameter C determines the trade off between the model complexity (flatness)
and the degree to which deviations larger than are tolerated in optimization formulation. For
example, if C is too large (infinity), then the objective is to minimize the empirical risk only,
without regard to model complexity part in the optimization formulation.
Parameter
controls
the width of the
-insensitive
zone, used to fit the training data
(Cherkassky and Mulier, 1998; Vapnik, 1995; Vapnik, 1998)
can affect the number of support vectors used to construct the regression
function. The bigger and the fewer support vectors are selected. On the other hand, bigger
-values result in more flat estimates. Hence, both C and -values affect model complexity
The value of
(but in a different way).

To minimize the generalization error, these parameters should be properly optimized.
Existing practical approaches to the choice of C and
can
be summarized as follows
(Cherkassy and Ma,2002):

z
Parameters C and
are
selected by users based on a priori knowledge and/or user
expertise (Cherkassky and Mulier, 1998; Vapnik, 1998; Scholkopf,
Burges,
Smola,1999). Obviously, this approach is not appropriate for non-expert users. Based on
observation that support vectors lie outside the
-tube and the SVM model complexity
strongly depends on the number of support vectors. Scholkopf, Burges, Smola,1999

suggest to control another parameter (i.e., the fraction of points outside the
instead of
Under this approach, parameter
has
-tube)
to be user-defined. Similarly,
Mattera and Haykin, 1999 propose to choose - value so that the percentage of support
vectors in the SVM regression model is around 50% of the number of samples. However,
one can easily show examples when optimal generalization performance is achieved with
the number of support vectors larger or smaller than 50%.
z
Smola, Murata, Schlkopf and K. Muller,1998 proposed asymptotically optimal
values proportional to noise variance, in agreement with general sources on SVM

(Cherkassky and Mulier,1998; Vapnik,1998;). The main practical drawback of such
proposals is that they do not reflect sample size. Intuitively, the value of
should be
smaller for larger sample size than for a small sample size (with the same level of noise).
z
Selection of . It is well-known that the value of

noise level, that is
should be proportional to the input
(Cherkassky and Mulier,1998; Vapnik,1998;
Smola,
Murata, Schlkopf and Muller,1998;).

Cherkassky and Ma (2002) propose the following (empirical) dependency:
56
Chapter2
ln n
n
[2.43]
Based on empirical tuning, they proposes constant value
=3
which gives good
performance for various data set sizes, noise levels and target functions for SVM
regression. Here n is the number of training data sample. They assume that the standard
deviation of noise
is known or can be estimated from data which is again a difficult
task for non expert user.

z
Selecting parameter C equal to the range of output values (D. Mattera and S. Haykin,
1999). This is a reasonable proposal, but it does not take into account possible effect of
outliers in the training data.
Cherkassky and Ma (2002) propose following prescription for regularization parameter:
C = max(| y + 3 y |, | y 3 y |)
[2.44]
Where, y is the mean of the training responses (outputs), and y is the standard
deviation of the training response values. They claim this prescription can effectively
handle outliers in the training data. However the proposed value of C-parameter is
derived and applicable for RBF kernels only.
z
Using cross-validation for parameter choice as suggested by Cherkassky and

Mulier,1998; Scholkopf, Burges, Smola,1999. This is very computational and dataintensive.
Several recent references present statistical account of SVM regression ( Smola

and Schlkopf, 1998; Hastie, Tibshirani and Friedman, 2001) where the
- parameter is
associated with the choice of the loss function (and hence could be optimally tuned to
particular noise density) whereas the C parameter is interpreted as a traditional
regularization parameter in formulation that can be estimated for example by crossvalidation (Hastie, Tibshirani and Friedman, 2001).
z
Selecting a particular kernel type and kernel function parameters is usually based on
application-domain knowledge and also should reflect distribution of input (x) values of
the training data. Very little literature is available to throw light on this.
As evident from the above discussion, there is no shortage of (conflicting) opinions on optimal
setting of SVM regression parameters. Existing software implementations of SVM regression
usually treat SVM meta-parameters as user-defined inputs. For a non expert user it is very
difficult task to choose these parameters as he has no prior knowledge for these parameters for
57
Mathematical Tools
his data. In such a situation, user normally relies on trial and error method. Such an approach
apart from consuming enormous time may not really obtain the best possible performance.
2.15 Optimization of SVM model

This study is motivated by a growing popularity of support vector machines (SVM) for
regression problems. However, many SVM regression application studies are performed by
expert users having good understanding of SVM methodology. Since the quality of SVM
models depends on a proper setting of SVM meta-parameters, the main issue for practitioners
trying to apply SVM regression is how to set these parameter values (to ensure good
generalization performance) for a given data set. Though existing sources on SVM regression
give some recommendations on appropriate setting of SVM parameters, there is clearly no
consensus and (plenty of) contradictory opinions.
Conventionally, various deterministic gradient based methods are used for optimizing a
process model. Most of these methods however require that the objective function should be
smooth, continuous, and differentiable. The SVR models can not be guaranteed to be smooth
especially in regions wherein the input-output data (training set) used in model building is
located sparsely. In such situations, an efficient optimization formalism known as genetic
algorithm, which is lenient towards the form of the objective function, can be used.
The above five parameters of SVR are optimized by GA algorithm stated below:
The objective function and the optimization problem of SVR model of the present study are
represented as,
Minimize (AARE(X)) on test set, X {x1, x2, x3, x4, x5})
[2.45]
Subject to, x1 {0, 10+5}, x2 = {0 to 1}, x3 = {1, 2}, x4 = {1, 2,.., 6}, x5 = {1,2,..,6}
The objective is the minimization of average absolute relative error (AARE) on test set and X is
a solution string representing a design configuration. The design variable x1 takes any values
for C in the range of 0.0 to 10,000. x2 represents the (epsilon) taking any values in the range
of 0.0 to 1. x3 represents the loss function types: e insensitive loss function and Huber loss
function represented by the numbers 1 and 2 respectively (Table 2.3). x4 represents the kernel
types: all 6 kernels given in Table 2.2 represented by numbers 1 to 6 respectively. The variable
x5 takes six values of the kernel parameters (represents degree of polynomials etc) in the range
58
Chapter2
1. to 6. The total number of design combinations with these variables is 100 100 2 6 6
= 720,000. This means that if an exhaustive search is to be performed, it will take at the
maximum 720,000 function evaluations before arriving at the global minimum A for the test
set (assuming 100 trials for each to arrive at optimum C and
). So, the strategy which takes
few function evaluations is the best one. For minimization of A as the objective function,
genetic algorithm technique is applied to find the optimum design configuration of SVR
model.
Table 2.2: Different Kernel type
59
Mathematical Tools
Table 2.3: Different Loss function
2.16 GA-Based Optimization of SVR Models

The strategy for GA based optimization of SVR model is principally same as the methodology
adopted in GA based optimization of ANN models. The stepwise procedure for the GA-based
optimization of a SVR model is provided as the flowchart in Figure 2.11. The detail
calculation steps are summarized in appendix 4. In simple term the hybrid methodology is
described as below: Initially GA will select random values of five parameters namely C,, loss
function type, kernel type and kernel parameters within user defined limits for Npop number of
candidates. These values will be then passed to the SVR model and the output from SVR
model will be calculated. This output will be used to calculate the average prediction error
(AARE) on test set which will be considered as fitness of the candidate (lowest AARE will be
the fittest candidate). In the second iterations, another set of population will be selected and
subsequently their fitness will be calculated. The candidates from these two iterations will be
compared and only fittest candidate will survive. All other candidate will be discarded. In
subsequent iterations, the new candidate (GA will utilize the crossover and mutation to
generate these new candidates) will be compared with the fittest candidate from earlier
iterations
and the candidate fitness will be refined based on survival of the fittest
methodology until maximum allowable iterations are reached. At the end, this algorithm will
generate a optimal set of five parameters namely C, , loss function type, kernel type and
kernel parameters within user defined limits which will produce lowest AARE. These five
optimal parameters will be then used in SVR model to predict the output. There are certain
striking features of GA that establishes its credibility. The most essential feature is that GAs
work with a population of points instead of a single point. This enables the search to come up
with a region of global solutions. Secondly, the initial population is chosen randomly.
Therefore the search can proceed in any direction.
60
Chapter2
2.17 DE based optimization of SVR model

The principals for DE based optimization of SVR model is same as the methodology adopted
in DE based optimization of ANN models. The purpose of the hybrid DE based SVR
algorithm is to find out the optimal value of five SVR meta parameters namely C,, loss
function type, kernel type and kernel parameters so that SVR model will generate lowest error
(AARE) on test set. The optimization power of differential evolution was used to arrive at this
optimal solution. The stepwise procedure for the DE-based optimization of a SVR model is
provided as the flowchart in Figure 2.12.
61
Mathematical Tools
Initialize generation index, Ngen= 0
Generate binary coded initial population of Npop chromosomes
Decode jth chromosome (j = 1, 2, , Npop) to obtain the

corresponding solution vector, x j . This xj consist of five
parameters of SVR namely C,, loss function type, kernel type and
kernel parameters.
Apply x j to the SVR model to compute the output, yk.
Evaluate fitness, Aoftestsetof the jth chromosome using {yk}
Rank the chromosomes in the descending order of their fitness
Select (Npop/2) number of parent pairs by
utilizing
Generate offspring population by performing

crossover on parent pairs
Perform mutation on the offspring population
Update generation index: Ngen = Ngen + 1

NO
N gen
?
Yes
Decode the top ranked chromosome to

obtain the optimal solution vector, x*
Figure 2.11: Schematic for hybrid SVR-GA algorithm implementation
62
Chapter2
Figure 2.12: Schematic for hybrid SVR-DE algorithm implementation
63
Mathematical Tools
Matlab implementation
ANN and GA toolbox are already available in matlab. SVM, SVR and DE matlab code was
developed as per the above algorithm. Then four new code was developed by combining
existing matlab toolbox and own code to implement four hybrid methodology described above.
2.18 Conclusion
This section explains ANN and SVR methodologies for a process modeling. The strengths and
weakness of ANN and SVR methodologies were explored. This chapter also described in
detail the hybrid methodology adopted for tuning the meta parameters of ANN and SVR
models. This thesis has developed four new hybrid modeling technique namely 1)hybrid SVRDE technique 2)hybrid SVR-GA technique 3)hybrid ANN-DE technique 4)hybrid ANN-GA
technique and applied them successfully in slurry flow modeling. The proposed four hybrid
techniques relieve the non-expert users to choose the meta parameters of SVR and ANN
algorithm for any case study and find out optimum value of these meta parameters on its own.
64
Chapter2
Appendix-1
Genaralized delta rule (GDR) algorithm
This appendix describes the procedure adopted to update weights in ANN algorithm after each
iteration.
Normalizedinput
Actualoutput
T31=1
T21=1
I1
a1
v11
b1
v21 v31
I2
a2
v22
a3
I3
v33
v12
w11
c1
Desiredoutput
d1
w12
T22=1
b2
w31
T32=1
c2
d2
v32
w32
v13
v23
w13
c3
b3
T23=1
w33
d3
T33=1
Figure 2.13: A simplified three layer feed forward perceptron network
Step1: randomly specify numerical values for all weight factors (vij s and wjk s) within the
interval [-1,+1]. For the delta rule, the internal threshold values must be assigned as follows:
all input-layer thresholds must equal zero i.e T1i=0; all hidden and output layer thresholds
must equal one i.e. T2j=T3k=1 Here i=1,2,..L, where L is the number of nodes in layer 1,
j=1,2,..m, where m is the number of nodes in layer 2, k=1,2,..n, where n is the number of
nodes in layer 3
Step 2: introduce the input Ii into the neural network and calculate the output from the first
layer according to the equations:
Xi=Ii+T1i=Ii+0=Ii
[2.46]
65
Mathematical Tools
F() here is the sigmoid function.
Step 3: knowing the output from the first layer, calculate outputs from the second layer, using
the equation:
[2.47]
Where f() is the sigmoid function. Note that T2i=1 in this algorithm, and we are again adding
it to the weighted inputs. When used in this mode, T2i acts as a bias function rather than an
internal threshold.
Step4: knowing the output from the second layer, calculate the result from the output layer,
according to the equation :
[2.48]
Where f() is the sigmoid function. Note again T3k=1 in this algorithm, and we are again
adding it to the weighted inputs.
Step 5: continue steps 1-4 for P number of training patterns presented to the input layer.
Calculate the mean square error, E, according to the following equations:
[2.49]
Where P is the number of training patterns presented to the input layer, n is the number of
nodes on the output layer , dkp is the desired output value from the kth node in the pth training
pattern and ckp is the actual output value from the kth node in the pth training pattern.
Step 6: knowing the pth pattern, calculate 3kp, the gradient descent term for the kth node in
the output layer (layer 3) for training pattern p. the following equation is used:
[2.50]
Where f() is the sigmoid function:
[2.51]
And its partial derivative is:

66
Chapter2
[2.52]
Step 7: again knowing the pth pattern, calculate 2jp, the gradient descent term for the jth node
on the hidden layer (layer 2). Following equation is used:
[2.53]
Where, the subscript k denotes a node in the output layer. Recall that xj is defined by:
[2.54]
And the partial derivative of the sigmoid function again is:
[2.55]
Step 8: knowing 2ip for the hidden layer and 3kp for the output layer, calculate the weight
changes using the equations:
Where,
[2.56]
[2.57]
is the learning rate and is the momentum co-efficient. Momentum is simple an
added weight used to speed up the training rate. The momentum coefficient, , is usually
restricted such that 0<<1. Thus the momentum terms,
and
, are fractional
values of the weight changes from the previous training iteration.

Step 9: knowing the weight changes, update the weights according to the equations:
[2.58]
[2.59]
67
Mathematical Tools
Repeat steps 2-9 for all training patterns until the squared error is sufficiently low. We have
now completed one iteration. We repeat step2-9, until the mean squared error calculated from
equation 3 below an acceptable limit.
Appendix 2
Different back propagation algorithm
This appendix describes the sailent features of different backpropagation algorithm available
in ANN toolbox in matlab.
Batch Gradient Descent with variable learning rate (traingdx)
The performance of the back propagation algorithm can be improved if we allow the learning
rate to change during the training process. The learning rate is made responsive to the
complexity of the local error surface. The learning rate is multiplied with the negative of the
gradient to determine the changes to the weights and biases. The larger the learning rate, the
bigger the step. If the learning rate is made too large, the algorithm becomes unstable. If the
learning rate is set too small, the algorithm takes a long time to converge.
An adaptive learning rate requires some changes in the training procedure used by simple
gradient decent algorithm. First, the initial network output and error are calculated. At each
epoch new weights and biases are calculated using the current learning rate (typical starting
value is 0.05). New outputs and errors are then calculated. If the new error exceeds the old
error by more than a predefined ratio (typically 1.04), the new weights and biases are
discarded. In addition, the learning rate is decreased (typically by multiplying by 0.7).
Otherwise, the new weights, etc., are kept. If the new error is less than the old error, the
learning rate is increased (typically by multiplying by 1.05).
This procedure increases the learning rate, but only to the extent that the network can learn
without increasing large error. Thus, a near-optimal learning rate is obtained for the local
terrain. When a larger learning rate could result in stable learning, the learning rate is
increased. When the learning rate is too high to guarantee a decrease in error, it gets decreased
until stable learning resumes.
68
Chapter2
Resilient Back propagation (trainrp)
Multilayer networks typically use sigmoid transfer functions in the hidden layers. Sigmoid
functions are characterized by the fact that their slope must approach zero, as the input gets
large. This causes a problem when using steepest descent to train a multilayer network with
sigmoid functions, since the gradient can have a very small magnitude; and therefore, cause
small changes in the weights and biases, even though the weights and biases are far from their
optimal values.
The purpose of the resilient back propagation training algorithm is to eliminate these harmful
effects of the magnitudes of the partial derivatives. Only the sign of the derivative is used to
determine the direction of the weight update; the magnitude of the derivative has no effect on
the weight update. The size of the weight change is determined by a separate update value.
The update value for each weight and bias is increased by a factor whenever the derivative of
the performance function with respect to that weight has the same sign for two successive
iterations. The update value is decreased by another factor whenever the derivative with
respect to weight changes sign from the previous iteration. If the derivative is zero, then the
update value remains the same. Whenever the weights oscillate, the weight change will be
reduced. If the weight continues to change in the same direction for several iterations, then the
magnitude of the weight change will be increased.
Conjugate Gradient Algorithms
The basic back propagation algorithm adjusts the weights in the steepest descent direction
(negative of the gradient). This is the direction in which the performance function is
decreasing most rapidly. It turns out that, although the function decreases most rapidly along
the negative of the gradient, this does not necessarily produce the fastest convergence. In the
conjugate gradient algorithms a search is performed along conjugate directions, which
produces generally faster convergence than steepest descent directions.
Fletcher-Reeves Update (traincgf)
All of the conjugate gradient algorithms start out by searching in the steepest descent direction
(negative of the gradient) on the first iteration.
p0 = g0
[2.60]
The general procedure for determining the new search direction is to combine the new steepest
descent direction with the previous search direction:
69
Mathematical Tools
pk = gk + kpk 1
[2.61]
The various versions of conjugate gradient are distinguished by the manner in which the
constant is computed. For the Fletcher-Reeves update the procedure is,
[2.62]
This is the ratio of the norm squared of the current gradient to the norm squared of the
previous gradient.
Polak-Ribire Update (traincgp)
Polak and Ribire proposed another version of the conjugate gradient algorithm. As with the
Fletcher-Reeves algorithm, the search direction in every iteration is determined by,
pk = gk + kpk 1
[2.63]
For the Polak-Ribire update, the constant is computed by
[2.64]
This is the inner product of the previous change in the gradient with the current gradient
divided by the norm squared of the previous gradient.
Powell-Beale Restarts (traincgb)
For all conjugate gradient algorithms, the search direction will be periodically reset to the
negative of the gradient. The standard reset point occurs when the number of iterations is equal
to the number of network parameters (weights and biases), but there are other reset methods
that can improve the efficiency of training. One such reset method was proposed by Powell
based on an earlier version proposed by Beale. In this technique, the process will restart if
there is very little orthogonality left between the current gradient and the previous gradient.
This is tested with the following inequality,
[2.65]
If this condition is satisfied, the search direction is reset to the negative of the gradient.
70
Chapter2
Quasi-Newton Algorithms
BFGS Algorithm (trainbfg)
Newtons method is an alternative to the conjugate gradient methods for fast optimization. The
basic step of Newtons method is
[2.66]
Where, Ak is the Hessian matrix (second derivatives) of the performance index at the current
values of the weights and biases. Newtons method often converges faster than conjugate
gradient methods. Unfortunately, it is complex and expensive to compute the Hessian matrix
for feed forward neural networks.
There is a class of algorithms that is based on Newtons method, which doesnt require
calculation of second derivatives. These are called quasi-Newton (or secant) methods. They
update an approximate Hessian matrix in each iteration of the algorithm. The update is
computed as a function of the gradient.
Levenberg-Marquardt (trainlm)
Like the quasi-Newton methods, the Levenberg-Marquardt algorithm was designed to

approach second-order training speed without computing the Hessian matrix. When the
performance function has the form of a sum of squares (as typical in training feed forward
networks), then the Hessian matrix can be approximated as,
[2.67]
and the gradient can be computed as,
[2.68]
Where, J is the Jacobian matrix that contains first derivatives of the network errors with
respect to the weights and biases, and e is a vector of network errors.
The Levenberg-Marquardt algorithm uses this approximation to the Hessian matrix in the
following Newton-like update:
71
Mathematical Tools
[2.69]
When the scalar is zero, this is just Newtons method, using the approximate Hessian matrix.
When is large, this becomes gradient descent with a small step size. Newtons method is
faster and more accurate near an error minimum. So, the aim is to shift towards Newtons
method as quickly as possible. Thus, is decreased after each successful step (reduction in
performance function) and is increased only when a tentative step would increase the
performance function. In this way, the performance function will always be reduced in each
iteration of the algorithm.
One Step Secant Algorithm (trainoss)
Since the BFGS algorithm requires more storage and computation in every iteration than the
conjugate gradient algorithms, there is need for a secant approximation with smaller storage
and computation requirements. The one step secant (OSS) method is an attempt to bridge the
gap between the conjugate gradient algorithms and the quasi-Newton (secant) algorithms. This
algorithm does not store the complete Hessian matrix and it assumes that at each iteration, the
previous Hessian was the identity matrix. This has the additional advantage that the new
search direction can be calculated without computing a matrix inverse.
Appendix 3
Stepwise procedure for the GA-based optimization of an ANN model
Step 1 (Initialization): Set generation index (Ngen) to zero and generate a population of Npop
binary strings (chromosomes) randomly; each string consisting of lchr bits is divided into N
segments equal to the number of decision (input) variables to be optimized.
Step 2 (Fitness computation): Decode jth binary-coded chromosome (j = 1, 2, , Npop) to
obtain its equivalent decimal-valued solution vector (xj) using:
;
Where,
and
[2.70]
respectively refer to the lower and upper bounds on xn; ln is the length of
nth binary segment and Sn denotes the decimal equivalent of the nth binary segment. The
72
Chapter2
vector xj actually represents five parameters of ANN model namely number of nodes in
hidden layer, activation function in input and output layer, learning rate and ANN training
algorithm. Next, depending upon the model to be optimized, vector xj is used to compute the
output of an ANN model; this output is subsequently used to calculate the fitness value
(AARE of test set) of the jth candidate solution. Upon computing the fitness scores of Npop
candidate solutions in the current population, the solutions are ranked in the decreasing order
of their fitness scores.
Step 3 (Parent selection): From the current population, choose Npop number of parent
chromosomes to form the mating pool. The members of this pool, which are used to produce
offspring population, possess relatively high fitness scores. The commonly used parent
selection techniques are the Roulette-Wheel (RW) method.
Step 4 (Crossover): Randomly select Npop/2 number of parent pairs from the mating pool and
perform crossover operation on each pair with probability equal to pcross (0 <pcross <1). In
crossover, parent strings are cut at the same randomly chosen crossover point to obtain two
substrings per parent. The substrings are then mutually exchanged between the parents and
combined to form two offspring chromosomes. This crossover operation is performed on all
the parent pairs in the mating pool to obtain an offspring population of the size of the mating
pool.
Step 5 (Mutation): Flip (mutate) the bits of the offspring strings where the probability of a bit
getting flipped (zero to one or vice versa) is equal to pmut. The recommended range of pmut
is [0.01-0.05].
Step 6: Increment the generation index: Ngen = Ngen + 1.
Step 7: Repeat steps 2-6 on the newly generated offspring strings until convergence is
achieved. The criteria for the convergence are: Ngen exceeds its maximum limit ( maxgen N ),
or the fitness score of the best (fittest) string in the offspring population undergoes a very
small or no change over successive generations. After convergence, the string possessing
highest fitness value is decoded to obtain the optimized decision variable vector, x*.
73
Mathematical Tools
Appendix 4
Stepwise procedure for the GA-based optimization of a SVR model
Step 1 (Initialization): Set generation index (Ngen) to zero and generate a population of Npop
binary strings (chromosomes) randomly; each string consisting of lchr bits is divided into N
segments equal to the number of decision (input) variables to be optimized.
Step 2 (Fitness computation): Decode jth binary-coded chromosome (j = 1, 2, , Npop) to
obtain its equivalent decimal-valued solution vector (xj) using:
where,
and
[2.71]
respectively refer to the lower and upper bounds on xn; ln is the length of
nth binary segment and Sn denotes the decimal equivalent of the nth binary segment. This xj
consist of five parameters of SVR namely C, , loss function type, kernel type and kernel
parameters. Next, depending upon the model to be optimized, vector xj is used to compute the
output of an SVR model; this output is subsequently used to calculate the fitness value (AARE
of test set) of the jth candidate solution. Upon computing the fitness scores of Npop candidate
solutions in the current population, the solutions are ranked in the decreasing order of their
fitness scores.
Step 3 (Parent selection): From the current population, choose Npop number of parent
chromosomes to form the mating pool. The members of this pool, which are used to produce
offspring population, possess relatively high fitness scores. The commonly used parent
selection techniques are the Roulette-Wheel (RW) method.
Step 4 (Crossover): Randomly select Npop/2 number of parent pairs from the mating pool and
perform crossover operation on each pair with probability equal to pcross (0 <pcross <1). In
crossover, parent strings are cut at the same randomly chosen crossover point to obtain two
substrings per parent. The substrings are then mutually exchanged between the parents and
combined to form two offspring chromosomes. This crossover operation is performed on all
the parent pairs in the mating pool to obtain an offspring population of the size of the mating
pool.
74
Chapter2
Step 5 (Mutation): Flip (mutate) the bits of the offspring strings where the probability of a bit
getting flipped (zero to one or vice versa) is equal to pmut. The recommended range of pmut
is [0.01-0.05].
Step 6: Increment the generation index: Ngen = Ngen + 1.
Step 7: Repeat steps 2-6 on the newly generated offspring strings until convergence is
achieved. The criteria for the convergence are: Ngen exceeds its maximum limit (maxgen N),
or the fitness score of the best (fittest) string in the offspring population undergoes a very
small or no change over successive generations. After convergence, the string possessing
highest fitness value is decoded to obtain the optimized decision variable vector, x*.
Nomenclature
AARE:
average absolute relative error
Ei:
sum-squared-error corresponding to the ith training pattern
Etrn :
RMSE corresponding to the training set
Etst:
RMSE corresponding to the test set
f:
~
f1 :
vector of nonlinear functions approximated by the ANN

nonlinear activation functions in input layer
~
f2
nonlinear activation functions in output layer
K:
number of nodes in the networks output layer
L:
number of nodes in networks hidden layer
lchr :
length of chromosome
N:
number of nodes in networks input layer
N genmax:
maximum number of generations over which GA simulations are conducted
Ng:
generation index
NP :
number of input-output patterns used for network training
Npop :
size of the candidate solution (chromosome) population
pcr :
crossover probability
pmut :
mutation probability
R:
cross-correlation co-efficient
W:
weight matrix set {WH, WO}

75
Mathematical Tools
WH :
weight matrix between input and hidden nodes
wk :
kth weighting coefficient
W :
H
nl
weight matrix between hidden and output nodes

:
weight between nth input and lth hidden node
wlk o :
weight between lth hidden node and kth output node
x:
N-dimensional vector of process input (decision) variables.
x* :
optimal decision variable vector
xj :
jth decimal valued solution vector
Xn :
nth network input.
y:
K-dimensional vector of process output variables
yj :
ANN model outputs corresponding to xj
yk :
kth network output
Greek symbols
standard deviation of error
Abbreviations
ANN:
Artificial neural network
DE:
Differential evolution
GA:
Genetic algorithm
SVM:
Support vector machine
SVR:
References
1.
Agarwal M.,Jade A. M.,Jayaraman V. K.,Kulkarni B. D., Supprot vector machines: A

useful tool for process engineering applications, Chem. Engg. Prog., 57-62 (2003)
2.
Alba, E., & Dorronsoro, B. ,The exploration/exploitation tradeoff in dynamic cellular

genetic algorithms. IEEE Transactions on Evolutionary Computation, 9(2),126142.
(2005).
3.
Anroulakis, I.P. and Venkatasubramanian, V., A genetic algorithm framework for process
design and optimization, Computers and chemical engineering,15(4),217-228(1991)
76
Chapter2
4.
Aurnhammer, M., & Tonnies, K. D. ,A genetic algorithm for automated

correlation across faults in seismic images. IEEE Transactions on
horizon
Evolutionary
Computation, 9(2), 201210(2005).

5.
Babu B. V.,Sastry K. K. N., Estimation of heat transfer parameters in a trickle-bed reactor

using differential evolution and orthogonal collocation, Comp. Chem. Engng.,, 23, 327 339 (1999)
6.
Babu, B.V. and Vivek, N., Genetic algorithms for estimating heat transfer parameters in
trickle bed reactors, proceedings of the international symposium and 52nd annual session
of
IIChE
(CHEMCON-99),Panjab
University,
Chandigarh,
dec
20-23,
(http://bvbabu.50megs.com /custom.html/#26) (1999)

7.
Baughman D.R.,Liu Y.A, Neural Networks in Bioprocessing and Chemical Engineering,

Academic Press., (1995)
8.
Bulsari, A.B., Medeveder, A., and Saxen. A., Sensor falt detection using state vector
estimator and neural networks applied to a biochemical process, Acta Polytechnic
Scandinavica, Chemical technology and metallurgy series,199,1-20,(1991)
9.
Burges C., A tutorial on support vector machines for pattern recognition, Data Mining and
Knowledge Discovery,, 2(2), 1-47 (1998)
10. Burges, C.J.C. ,A tutorial on support vector machines for pattern recognition.,Data Mining
and Knowledge Discovery, Vol. 2, number 2. (1998).
11. Cao, L. ,Support vector machines experts for time series forecasting. Neurocomputing,
51(1-4), 321339. (2003).
12. Cherkassky V. ,Ma Y., Practical selection of SVM parameters and noise estimation for
SVM regression, Neurocomputing (special issue on SVM), (2002)
13. Cherkassky V.,and Mulier F., Learning from Data: Concepts Theory and Methods, John
Wiley & Sons,, (1998)
14. Christiani, N., & Shawe-Taylor, J. ,An introduction to support vector machines.
Cambridge, England: Cambridge University Press. (2000).
15. Cortes, C., & Vapnik, V. ,Support vector networks. Machine Learning, 20:273-297.
(1995).
16. Cristianini, N., Campell, C., & Taylor, J. S., Dynamically adapting kernels in support
vector machines. Advances in Neural Information Processing Systems,11(2), 204210.
(1999).
77
Mathematical Tools
17. Dastidar, T. R., Chakrabarti, P. P., & Ray, P. ,A synthesis system for analog circuits
based on evolutionary search and topological reuse. IEEE Transactions on Evolutionary
Computation, 9(2), 211224(2005).
18. Deb, K., Optimization for engineering design: Algorithm and examples, 4th edn, Prentice
Hall of India, New Delhi,(2000)
19. Duan, K., Keerthi, S. S., & Poo, A. N. ,Evaluation of simple performance measures for
tuning SVM hyperparameters. Neurocomputing, 51(1-4), 4159. (2003).
20. Fogel, D. B. ,An introduction to simulated evolutionary optimization. IEEE Transactions
on Neural Networks, 5(1), 314. (1994).
21. Glucksman, H. ,On improvement of a linear separation by extending the adaptive process
with a stricter criterion. IEEE Transactions on Electronic Computers, EC-15 (6), 941-944.
(1966).
22. Goldberg, D. E. ,Genetic algorithms in search, optimization and machine learning.
Reading, MA: Addision-Wesley. (1989).
23. Gunn, S. ,Support Vector Machines for Classification and Regression. ISIS Technical
Report. (1997).
24. Hastie T.,Tibshirani R.,and Friedman J., The Elements of Statistical Learning Data
Mining Inference and Prediction, (Springer,, (2001)
25. Hecht-Nielsen R., Theory of the back propagation neural network, Proceedings of the
international joint conference on neural networks,, 1, 593-611 (1989)
26. Holland, J.H., Adaptation in natural and artificial systems,University of Michigan press,
ANN Arbor, Michigan (1975)
27. Hoskins, J.C. and Himmenblau , D.M., Artificial neural network models of knowledge
representation in chemical engineering, computer and chemical engineering,12,881,(1988)
28. Hoskins, J.C. and Himmenblau , D.M.,, Fault detection and diagnosis using artificial
neural networks, Artificial intelligence in process engineering, academic press, San Diego,
CA,123-160 (1990)
29. Huang, B. and
Mujumder A.S., use of neural networks to predict industrial dryer
performance, Drying technology,11,525(1993)

30. Jack L. B., Nandi A. K.,
Fault detection using support vector machines and artificial
neural networks augmented by genetic algorithms, Mech. Sys. Sig. Proc.,, 16, 372-390
(2002)
78
Chapter2
31. Jack, L.B., & Nandi, A.K.. Fault detection using support vector machines and artificial
neural networks, augmented by genetic algorithms. Mechanical Systems and Signal
Processing, 16 (2-3), 373-390. (2002)
32. Kim,
K.
,Financial
time series
forecasting
using
support
vector
machines.
Neurocomputing, 55(1), 307319. (2003).

33. Kirkpatrick, S., Gelatt, C.D., and Vechhi, M.P., Optimization by simulated annealing,
science,220(4568),671-680 (1983)
34. Kulkarni, A., Jayaraman, V. K., & Kulkarni, B. D., Control of chaotic dynamical systems
using support vector machines. Physics Letters A, 317(5), 429435. (2003)
35. Mah, R.S.H, Neural nets, Letter to the editor, Chemical engineering progress,87,6,
January (1991)
36. Mattera D.,and Haykin S., Support Vector Machines for Dynamic Reconstruction of a
Chaotic System , Advances in Kernel Methods: Support Vector Machine,, MIT Press,
(1999)
37. Matlab online documentation, Documentation for MathWorks Products, R2009a,
(http://www.mathworks.com/access/helpdesk/help/helpdesk.html)(2009)
38. Min, J. H., & Lee, Y. C. ,Bankruptcy prediction using support vector machine with
optimal choice of kernel function parameters. Expert Systems with Applications, 28(4),
603614. (2005).
39. Mukherjee, S., Osuna, E., & Girosi, F. ,Nonlinear prediction of chaotic time series using a
support vector machine. In Proceedings of the NNSP97. (1997)
40. Mller, K.-R., Smola, A., Rtsch, G., Schlkopf, B., Kohlmorgen, J., & Vapnik, V.
,Predicting time series with support vector machines. In Proceedings of the ICANN97,
999 (1997).
41. Price K., Storn R., Differential evolution- A simple evolution strategy for fast
optimization,, Dr. Dobb.s Journal,, 2294)18-24.& 78, (1997)
42. Reisner, J., Reuter, M.A..,krugger, J. , modeling of the mass transfer in gas sparged
electrolyses with neural nets, Chemical engineering science,48,1089(1993)
43. Reuter, M.A.., Deventer J.S.J.V, Vanderwalt, T.J., A generalized neural net kinetic rate
equation. Computer and chemical engineering,48,1281 (1993)
44. Sablania, S.S. , Walid H.S & Kacimovc, A.,Explicit calculation of the friction factor in
pipeline flow of Bingham plastic fluids: a neural network approach,Chemical Engineering
Science 58 , 99 106(2003)
79
Mathematical Tools
45. Sastry K. K. N.,Behra L., Nagrath I. J., ,
Differential evolution based fuzzy logic
controller for nonlinear process control, Fundamenta Informaticae:, Special Issue on Soft
Comput., (1998)
46. Scholkopf B,Burges J..,Smola A., ,
Advances in Kernel Methods: Support Vector
Machine, (MIT Press,, (1999)

47. Schlkopf, B., Burges, C., & Vapnik, V. ,Extracting support data for a given task. In U.
M. Fayyad and R. Uthurusamy, editors, Proceedings, First International Conference on
Knowledge Discovery and Data Mining, Menlo Park. AAAI Press. (1995).
48. Schlkopf, B., Platt, J.C., Shawe-Taylor, J., & Smola, A.J. Estimating the support of a
high dimensional distribution. Technical Report MSR-TR-99-87, Microsoft Research,
Redmond, WA, USA. Online version: http://www.kernelmachines.org/papers/oneclasstr.ps.gz(1999).
49. Shaikh,A. & Al-Dahhan,M.,Development of an artificial neural network correlation for
prediction of overall gas holdup in bubble column reactors,Chemical Engineering and
Processing 42 , 599-610,(2003)
50. Shin, S. Y., Lee, I. H., Kim, D., & Zhang, B. T. ,Multiobjective evolutionary optimization
of DNA sequences for reliable DNA computing. IEEE Transactions on Evolutionary
Computation, 9(2), 143158. (2005).
51. Smola A.,Murata N.,Schlkopf B., Muller K., Asymptotically optimal choice of epsilon
loss for support vector machines, Proc. ICANN,, (1998)
52. Smola A.J.,Schlkopf B.,Mller K. R., The connection between regularization operators
and support vector kernels, Neural Networks,, 11, 637-649 (1998)
53. Tambe S. S.,Kulkarni B. D., Deshpande P. B., Elements of Artificial Neural Networks
with selected applications in Chemical Engineering and Chemical & Biological Sciences,
Simulations & Advanced Controls, Louisville, KY (1996)
54. Tay, Francis E. H., & Cao, L. ,Application of support vector machines in financial time
series forecasting. OMEGA The International Journal of Management Science, 29(4),
309317. (2001).
55. Upreiti, S.R. and Deb, K., Optimal design of an ammonia sysnthesis reactor using genetic
algorithms, Computers and Chemical engineering,2191),87-92(1997)
56. Vapnik V., Statistical Learning Theory, John Wiley,, New York., (1998)
57. Vapnik V., The Nature of Statistical Learning Theory, Springer Verlag,, New York.,
(1995)
80
Chapter2
58. Vapnik V.,Golowich S.,Smola A., Support vector method for function approximation
regression estimation and signal processing, Adv. in Neural Inform. Proces. Syst.,, 9, 281287 (1996)
59. Vapnik, V. ,The Nature Of Statistical Learning Theory. Springer. (1995).
60. Venkatasubramanian, V., Chan, K., and Caruthers, J.M., computer-aided molecular design
using genetic algorithms, Computers and Chemical engineering,18(9),833-844(1994)
61. Venkatasubramanian, V., Vaidyanathan, R., and Yamamoto, Y., process fault detection
and diagnosis using neural networks. 1. Steady-state processes, Computers and Chemical
Engineering,14(7),699-712 (1990)
62. Venkatraman, S., & Yen, G. G. ,A generic framework for constrained optimization using
genetic algorithms. IEEE Transactions on Evolutionary Computation, 9(4), 424435.
(2005).
63. Wizzard, j.t. and Fuhrman, M.G., Neural network application to remote temperature
measurements, AIChE national meeting, Houston, TX, april(1991)
64. Yaochu, J., & Branke, J., Evolutionary optimization in uncertain environments-a survey.
IEEE Transactions on Evolutionary Computation, 9(3), 303317. (2005)
65. Zhang, Q., Sun, J., & Tsang, E. ,An evolutionary algorithm with guided mutation for the
maximum clique problem. IEEE Transactions on Evolutionary Computation, 9(2), 192
200. (2005).
81
Chapter3
CHAPTER 3
Regime Identification of Slurry Transport in

Pipelines
Abstract
Four distinct regimes were found existent (namely flow with stationary bed, flow with moving
bed, heterogeneous suspension and homogeneous suspension) in slurry flow in pipeline
depending upon the flow velocity. Literature survey shows that there are very few number of
correlations proposed for identification of these regimes in slurry pipelines and these
available correlations have certain limitation to predict the regime over a wide range of
conditions. As regime identification is an important criterion for slurry pipeline design to
select suitable correlation of pressure drop for that regime, a method has been proposed to
identify the regime using artificial neural network (ANN) & support vector machine (SVM)
modeling. ANN & SVM were applied on a databank of around 800 measurements collected
from the open literature. The method incorporates three hybrid techniques namely 1)artificial
neural network-differential evolution (ANN-DE), 2) artificial neural network-genetic
algorithm (ANN-GA) and 3)support vector machine- differential evolution (SVM-DE) for
efficient tuning of ANN & SVM meta parameters respectively. Statistical analysis showed that
the proposed method has an average misclassification error (AME) of 0.03%. A comparison
with selected correlations in the literature showed that the developed hybrid method
noticeably improve the prediction of regime over a wide range of operating conditions,
physical properties, and pipe diameters.
Keywords: Artificial neural network, Differential evolution, Genetic algorithm, Support

vector machine, Slurry flow regime, Slurry flow
81
3.1 Introduction
Four distinct regimes were found existent (namely homogeneous suspension, heterogeneous
suspension, moving bed, sliding bed) in slurry flow in pipeline depending upon flow velocity.
Great attention has been paid for reduction of hydraulic losses in slurry transport as power
consumption constitutes a substantial portion of operational costs for the overall pipeline
transport. The pressure drop correlations available in literature are applicable to a particular
regime for which they were developed. The correlations show ill prediction of pressure drop
when they apply for other regimes. Thus regime identification becomes important for slurry
pipeline design as they are the prerequisite to select suitable pressure drop correlation in that
regime. The understanding of different regime formation and accurate prediction of pressure
drop of slurry flow in design phase makes possible to optimize pipe diameter and energy
requirements.
Experimental observations have shown that different correlations should be used in each of the
identifiable flow regimes. This approach is logical but not straightforward to apply. The main
problem arises because of difficulty to define the boundaries between the flow regimes. These
boundaries are poorly defined because these are based on visual observations of particle
motions in small laboratory pipelines. Many researchers have attempted to establish
correlations among the relevant experimental variables that can be used to define the
boundaries of the flow regimes. These attempts have met with only limited success and an
approach developed by Turian and Yuan (1977) is found most popular and promising.
The approach of Turian and Yuan (1977) claims to provide a completely self consistent
definition of the flow regime boundaries that results directly from the head loss correlations
and no additional correlations are required to define the boundaries. The method claims that it
is based on a large database of reliable experimental data and consequently the method can be
used with confidence for practical engineering problems. A databank of around 800
measurements collected from the open literature was exposed to Turian and Yuans
calculations and the result of prediction of flow regimes is found very poor. The poor
prediction motivated this work and an attempt has been made to develop a suitable method
which can well identify the flow regime over a wide range of operating conditions.
82
Chapter3
3.2 Background work

Two schools are credited for laying the foundations of modern hydro-transport engineering;
SOGREAH in France, and the British Hydro-mechanic Research Association of the United
Kingdom. Starting in 1952, Durand and Condolios (1952) of SOGREAH published number of
studies on the flow of sand and gravel in pipes up to 900 mm (35.5 in) in diameter.
Based on the specific gravity of particles with a magnitude of 2.65, they proposed to divide the
flows of non-settling slurries in horizontal pipes into four categories based on average particle
size as follows:
Homogeneous suspensions for particles smaller than 40 m (mesh 325)
Suspensions maintained by turbulence for particle sizes from 40 m (mesh 325) to

0.15 mm (mesh 100)
Suspension with saltation for particle sizes between 0.15 mm (mesh 100) and 1.5 mm
(mesh 11)
Saltation for particles greater than 1.5 mm (mesh 11)
This initial classification was refined over the next 18 years by Newitt et al. (1955), Ellis and
Round (1963), Thomas (1964), Shen (1970), and Wicks (1971). Due to the interrelation
between particle sizes and terminal and deposition velocities, the original classification
proposed by Durand has been modified to four flow regimes based on the actual flow of
particles and their size. Referring to Figures 3.1 and 3.2, there are four main regimes of flow in
a horizontal pipe,
1. Flow with a stationary bed. (sometimes called sliding bed regime)
2. Flow with a moving bed and saltation (with or without suspension) ( denoted as moving bed
regime and /or saltation regime)
3. Heterogeneous mixture with all solids in suspension. This regime sometimes called
asymmetric flow regime in some references [ King (2002)].
4. Pseudo homogeneous or homogeneous mixtures with all solids in suspension. This regime
sometimes called symmetric flow regime in some references [ King (2002)].
83
84
Ch
hapter3
Fiigure 3.2 Flow

w regimes of heterogeneou
us flows in teerms of particcle size versuss mean
veelocity
3..2.1 Flow re
egimes in slu
urry flow
Too understand the
t slurry behaavior it is veryy important to know the diffe
ferent flow regimes of
sluurry flow. Thee tendency thatt the solid partticles have to settle
s
under the influence off gravity
haas a significan
nt effect on thee behavior of sslurry that is transported
t
in a horizontal pipeline.
p
Thhe settling tenndency leads to
t a significannt gradation in
n the concentraation of solidss in the
sluurry. The conccentration of solids is higherr in the lower sections of thee horizontal pippe. The
ex
xtent of the accumulation off solids in the llower section depends stronggly on the veloocity of
thhe slurry in thee pipeline. Thee higher the veelocity, the hig
gher is the turb
rbulence level and the
grreater is the abbility of the caarrier fluid to kkeep the particcles in suspenssion. It is the upward
m
motion
of eddyy currents trannsverse to thee main directtion of flow of the slurry that is
reesponsible for maintaining thhe particles inn suspension. At very high turbulence levvel, the
suuspension is almost
a
homogeeneous with vvery good disppersion of thee solids while at low
tuurbulence level, the particlees settle towarrds the floor of the channel. Naturally, at low
tuurbulence, partiicles have tenddency to remainn in contact with the flow annd are transporrted as a
sliiding bed und
der the influennce of the preessure gradiennt in the fluidd. Between theese two
85
extremes of behavior, two other more or less clearly defined flow regimes can be identified.
When the turbulence level is not high to prevent any deposition of particles on the floor of the
channel, the flow regime is described as heterogeneous suspension flow. As the velocity of the
slurry is reduced further a distinct mode of transport known as saltation develops. In the
moving bed regimes, there is a visible layer of particles on the floor of the channel and these
are being continually picked up by turbulent eddies and dropped to the floor again further
down the pipeline. The solids therefore spend some of their time on the floor and the rest in
suspension of the flowing fluid. Under moving bed conditions the concentration of solids is
strongly non uniform. The four regimes of flow are illustrated in figure 3.3.
Flow with a Stationary Bed

When the slurry flow speed is low, the bed thickens. As the fluid above the bed tries to move
the solids by entrainment, they tend to roll and tumble. The particles with the lowest settling
speed move as an asymmetric suspension, whereas the coarser particles build up the bed. As
the speed drops even further, the pressure to maintain the flow becomes quite high and
eventually the pipe blocks up. Flow with saltation and asymmetric suspension occurs above
the speed of blockage. This means that the coarser particles sand up, whereas the finer
particles continue to move.
Figure 3.3 Four regimes of flow of settling slurries in horizontal pipeline

Saltation can eventually lead to blockage of a pipe. It may result in a number of problems,
such as water hammer, wear, and freezing in cold climates. Most engineering specifications
86
Chapter3
require that the pipeline be designed to operate at speeds higher than those associated with
saltation.
Flow with a Moving Bed

When the speed of the flow is low and there are a large number of coarse particles, the bed
moves like desert sand dunes. The top particles are entrained in the moving fluid above the
bed. Consequently, the upper layers of the bed move faster than the lower layers in a
horizontal pipe. If the mixture were composed of a wide range of particles with different sizes
and settling velocities, the bed would be composed of the particles with the highest settling
speed. Particles with a moderate settling speed are maintained in an asymmetric suspension,
with most particles concentrated in the lower half of the pipe, whereas the particles with the
lowest settling speed move as a symmetric suspension.
Heterogeneous suspension or Asymmetric Flow

With increasing flow speed, turbulence is sufficient to lift more solids. All particles move in an
asymmetric pattern with the coarsest at the bottom of a horizontal pipe covered with
superimposed layers of medium- and fine-sized particles. Many particles may strike the
bottom of the pipe and rebound. Wear on the bottom of the pipe must be taken into account in
maintenance schedules and the pipes must be rotated at intervals suggested by the slurry
engineer in order to maintain an even wear pattern of the internal wall of the pipe. Although
the flow is not symmetric, from the point of view of power consumption, this regime may be
the most economical for transporting a certain mass of solids.
Homogeneous mixtures or Symmetric Flow at High Speed

At speeds in excess of 3.3 m/s (10 ft/s), all solids may move in a symmetric pattern (but not
necessarily uniformly). Sometimes this flow is called pseudo-homogeneous because of its
symmetry around the pipe axis. Power consumption is a linear relationship of the static head
multiplied by the velocity, but is proportional to the cube of velocity needed to overcome
friction losses. Power consumption for flow of pseudo-homogeneous mixtures of coarse and
fine particles may be excessive for long pipelines.
3.2.2 Head loss correlations for separate flow regimes
87
The moving bed and heterogeneous suspension regimes have been studied most widely. The
best known correlation for the excess pressure gradient due to the presence of solid particles is
proposed by Durand, Condolios (1952) and is given as,
Pf, sl P
fw
1.5
= C
( C*
D Fr )
Pfw
[3.1]
Where is a constant, C is the volume fraction of solids in the suspension, CD is the drag coefficient at terminal settling velocity and Fr is the Froude number. The value to be used for the
constant is uncertain and values between 65 and 150 are reported in the literature. Because
this correlation does not apply to all regimes of flow and the experimental data cannot be used
to fix the value more precisely. Errors of 100 percent and more in the calculated value of can
result.
While the Durand-Condolios (1952) correlation is useful in the heterogeneous suspension flow
regime, it deviates more and more from actual conditions in the other flow regimes.
identifiable flow regimes. Using the experimental data, Turian and Yuan (1977) established
that the excess pressure gradient in each flow regime can be correlated using an equation of
the form:
f sl f w = KC fw C
*
D
Fr
[3.2]
Where, the coefficients K,,, and have values that are specific to each flow regime and
C*D is the drag co-efficient at the terminal settling velocity of slurry. Using experimental data
gathered from experiments in each flow regime, the best available values of these parameters
in each flow regime are given by,
Sliding bed (regime 0)
fsl fw = 12 . 13 C 0 . 7389
fw 0 . 7717
* 0 . 4054
D
Fr
1 . 096
[3.3]
Moving bed (regime 1)
fsl fw = 107 . 1 C 1 . 018 fw 1 . 046 C
* 0 . 4213
D
Fr
1 .. 354
[3.4]
88
Chapter3
Heterogeneous suspension (regime 2)
fsl fw = 30 . 11 C 0 . 868 fw
1 . 200
* 0 . 1677
D
Fr
0 . 6938
[3.5]
Homogeneous suspension (regime 3)
fsl fw = 8 . 538 C 0 . 5024
fw 1 . 428 C
*
D
0 . 1516
Fr
0 . 3531
[3.6]
A fairly consistent trend in the variation of the correlating parameters is seen in the four
correlations.
3.2.3 Flow regime boundaries (Turian and Yuans approach)
The boundaries of the flow regimes are defined in a self-consistent manner by noting that any
two regimes are contiguous at their common boundary and therefore each of the two
correlations must be satisfied simultaneously. For example, the boundary between the sliding
bed regime (Regime 0) and the moving bed regime (regime 1) must lie along the solution
locus of the equation,
12.13C 0.7389 fw0.7717C D*
0.4054
0.4213
Fr 1.096 = 107.1C 1.018 fw1.046C D*
Fr 1..354
[3.7]
This is simplified to
2
Fr =
V
= 4679 C
Dg [ s 1 ]
1 . 083
fw 1 . 064 C
* 0 . 0616
D
[3.8]
The regime transition number for transitions between regime 0 and regime 1 is defined by,
R 01 =
4679 C
1 . 083
Fr
fw 1 . 064 C
* 0 . 0616
D
[3.9]
and this number must be unity on the boundary between these two regimes.
89
The transition numbers for the other possible transitions are found in the same way and are
given by,
R 02 =
R 12 =
Fr
0 . 1044
6 . 8359
R 13 =
12 . 522 C
R 23 =
0 . 3225
fw
40 . 38 C 1 . 075
* 0 . 5906
D
[3.10]
Fr
fw 0 . 2334
* 0 . 3840
D
Fr
fw 0 . 3820
*
D
0 . 2263
0 . 5153
1 . 065
0 . 5724
Fr
0 . 9375
fw 0 . 6700 C D*
[3.11]
[3.12]
[3.13]
Fr
R 03 =
1 . 6038
C 0 . 3183
fw 0 . 8837
*
D
0 . 7496
[3.14]
90
Chapter3
Figure 3.4 Decision tree for establishing flow regimes

These numbers define the boundaries between any two flow regimes a and b by the
condition Rab=1. It is possible to identify the regime that applies to a particular set of physical
conditions quite simply from knowledge of the transition numbers Rab. If a<b the value of Rab
increase monotonically as the velocity increases. At low velocities, Rab<1 and with increasing
velocity, the value of Rab will eventually pass through the value 1.0. This must signal a
transition out of regime a. The following simple rules will fix the flow regime:
If Rab<1 the regime is not b.
If Rab>1 the regime is not a.
This inequalities must be tested for the combinations of a and b. No more than three of the
transition numbers need to be calculated to fix the flow regime uniquely. Notice that these
rules test the flow regimes negatively and a single test will never suffice to define the flow
regime. It is always necessary to test at least three different combinations of a and b to get a
definitive identification of the flow regime (figure 3.4). The applicable flow regime can be
identified quickly and easily using the figure 3.4 and the appropriate equation can be selected
from equations above to calculate the slurry frictional head loss.
Computer program (C program) of the above method was developed and the results are shown
in table 3.3. Following calculation is used to calculate CD* (King, 2002):
1. Drag co-efficient at terminal settling velocity (CD*) is evaluated by balancing drag and
buoyancy force as per King (2002)
C D*
dp 3 ( s f ) g
f V T2
dp 2
=
6
2
4
C D* = [
4 ( s f ) gdp
]
3 fV T 2
[3.15]
[3.16]
Particle Reynolds number at terminal settling velocity is given by,
Re
*
p
dpV
T f
[3.17]
It is not possible to solve CD* from equation [3.16] directly because it is a function of terminal
settling velocity VT and particle diameter dp through the relationship of Turton-Levenspiel
(1986). This problem can be solved by introducing a dimensionless number 1* as follows:
91
1* = C D* Re
*
p
4 ( s f ) fg
] dp 3
=[
3f 2
[3.18]
As per Karamanev (1996) CD* can be calculated as follows:
C D* =
432
1*
(1 + 0 . 0470 1*
0 . 66
)+
0 . 517
1 + 154
[3.19]
0 . 33
1*
3.3 Performance check of Turian Yuans approach
3.3.1 Data Collection

As mentioned earlier, over the years researchers have amply quantified the flow regime of
slurry flow in pipeline. In this work, about 800 experimental runs have been collected from 20
sources of open literature spanning the years 1950 -2002 (most of them compiled in PhD thesis
report of Hsu (1987)). The data were screened for incompleteness, redundancies and evident
inaccuracies. This wide range of database includes experimental information from mainly two
regimes namely moving bed regime (regime 1) and heterogeneous flow regime (regime 2).
These two regimes were selected as they have great practical significance. Usually, all
practical hydro-transport are carried out in heterogeneous suspension regime because of lowest
pressure drop and subsequent less power requirement. Table 3.1 indicates the range of
collected databank for regime identification and Table 3.2 shows some of the sample data.
3.3.2 Regime identification
Now the data from Table 3.2 was exposed to Turian and Yuans calculation and regimes were
identified as shown in Table 3.3. It is evident from the last two columns of the Table 3.3 that
Turian Yuan approach is not producing promising results and failed to correctly identify the
regimes. The poor performance of Turian and Yuans approach motivated this work and an
attempt has been made to explore the new ANN and SVM classification methodology to
identify the flow regime correctly.
92
Chapter3

To facilitate the design and scale up of pipelines and slurry pumps, there is a need to have
suitable correlation that can predict flow regime over a wide range of operating conditions,
physical properties and particle size distributions. As stated earlier, the pressure drop
correlations available in literature are applicable to a particular regime for which they were
developed. The correlations show poor prediction of pressure drop when they apply for other
regimes. Thus regime identification becomes important for slurry pipeline design. The
understanding of different regime formation and accurate prediction of pressure drop of
slurries in design phase makes possible to optimize energy and water requirements.
Table 3.1: System and parameter studied
[Wilson (1942), Durand and Condolios (1952), Newitt et. Al. (1955), Thomas
(1962), Zandi and Govatos (1967), Shook (1969), Babcock (1970), Turian and
Yuan (1977), Wasp et al (1977), Oroskar and Turian (1980), Gillies et al (1999),
Kaushal (2002)]
Slurry system: Coal-water, Coal-brine, Ash-water, Copper ore-water, Sand-water,
Gypsum-water, Glass-water, Gravel-water, iron-water, iron-kerosene, high density
material-water, iron tailings-water, limestone-water, limonite-water, plastic-water,
potash-brine, sand-ethylene glycol, nickel shot-water, iron powder-water, ore-water
Pipe diameter (m)
0.0127 0.80
Particle diameter (m)
(0.0017 0.868) x 10-2
Liquid density (kg/m3)
770 1350
Solids density (kg/m3)
1150 8900
Liquid viscosity (kg/m-sec)
0.008 1.9 x 10-1
Solids concentration (volume fraction)
0.005 0.561
Velocity (m/s)
0.18 - 4.56
93
Table 3.2: Some of the input and output data for ANN & SVM training
SL. No.
Particle
dia,,
cm
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
0.0130
0.0100
0.2200
0.2200
0.8680
0.1830
0.0300
0.0140
0.0090
0.0060
0.1113
0.3346
0.0390
0.0410
0.0410
0.0540
0.0540
0.0500
0.0500
0.0500
0.0500
0.0200
0.0200
0.0490
0.0490
0.0490
0.0250
0.0250
0.2040
0.0850
0.0850
0.0242
0.0242
0.0242
0.0250
0.0250
0.0400
0.0400
0.0400
0.0242
0.0242
Solid
conc. ,
vol.
fraction
0.0280
0.3000
0.3040
0.3040
0.1650
0.0560
0.1700
0.0220
0.1250
0.0140
0.1920
0.1000
0.4750
0.3220
0.3830
0.3600
0.4200
0.0500
0.1200
0.1800
0.4200
0.1800
0.2400
0.1200
0.1800
0.2400
0.0050
0.0050
0.0250
0.2150
0.2150
0.0080
0.0300
0.0700
0.3000
0.3180
0.0300
0.0620
0.1370
0.0300
0.0700
Solid
density
gm/cc
1.8340
2.8200
1.6700
1.6700
1.5300
1.5300
3.3600
2.6900
3.0000
3.1000
4.5500
4.5500
1.9840
1.9840
1.9840
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6000
2.6000
2.6000
2.6500
2.6500
2.6500
2.6500
2.6500
2.6000
2.6000
Fluid
density,
gm/cc
Fluid
viscosity,
gm/cm-sec
Pipe dia,
cm
Fluid
velocity,
cm/sec
Regime
0.9980
0.9820
1.3250
1.3250
0.9980
0.9980
0.9980
0.9980
0.9980
0.9980
0.9980
0.9980
1.1460
1.1420
1.1420
1.3500
1.3500
1.3500
1.3500
1.3500
1.3500
1.1320
1.1320
1.0950
1.0950
1.0950
0.9980
0.9980
0.9980
0.9980
0.9980
0.9990
1.0000
1.0000
0.9980
0.9980
0.9980
0.9980
0.9980
1.0000
1.0000
0.0098
0.0130
1.5200
1.5200
0.0098
0.0098
0.0098
0.0098
0.0098
0.0098
0.0098
0.0098
0.0114
0.0120
0.0120
0.0560
0.0560
0.0560
0.0560
0.0560
0.0560
0.3820
0.3820
0.0575
0.0575
0.0575
0.0098
0.0098
0.0098
0.0098
0.0098
0.0100
0.0100
0.0101
0.0098
0.0098
0.0098
0.0098
0.0098
0.0100
0.0101
7.6200
5.0000
7.6200
10.1600
20.8000
4.0000
10.3000
20.7000
20.7000
14.9000
8.2500
15.0000
5.2200
10.7600
10.7600
5.2450
5.2450
5.2450
5.2450
5.2450
5.2450
5.2450
5.2450
5.2450
5.2450
5.2450
5.1500
5.1500
15.0000
5.0800
5.0800
7.6000
7.6000
7.6000
10.3000
10.3000
10.3000
10.3000
10.3000
7.6000
7.6000
147.22
332.50
85.98
204.14
254.85
52.55
126.73
210.62
337.16
120.59
260.32
315.00
138.99
117.23
117.23
76.73
78.41
57.60
65.60
65.15
72.39
75.29
66.60
121.61
124.81
125.45
68.80
83.54
209.68
155.10
122.73
86.52
115.90
130.15
102.16
108.70
364.88
207.63
218.93
115.90
130.15
1
-1
-1
-1
1
-1
-1
-1
1
1
1
1
-1
-1
-1
-1
1
1
1
-1
-1
-1
-1
1
1
1
-1
-1
-1
1
1
1
-1
-1
-1
1
1
-1
-1
-1
-1
94
Chapter3
Table
3.3
Regime
identification
by
Turian
Sl.
No.
R01
R02
R03
R12
R13
R23
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
11.89
0.66
0.13
0.75
4.30
1.30
0.32
8.05
3.10
4.07
1.13
2.27
0.51
0.31
0.25
0.08
0.07
0.35
0.18
0.11
0.06
0.14
0.08
0.69
0.47
0.35
8.74
13.57
8.72
0.94
0.55
6.83
3.17
1.64
0.15
0.17
32.74
4.16
1.99
3.17
0.03
1.75
3.16
9.82
0.04
0.01
0.01
0.01
0.03
0.01
0.02
0.01
0.06
0.02
0.02
0.03
0.03
0.01
0.02
0.02
0.03
2.89
2.56
0.06
0.07
0.08
0.00
0.00
0.01
0.03
0.02
0.00
0.01
0.01
0.01
0.01
0.03
0.01
0.02
0.01
0.06
0.55
2.25
7.38
0.02
0.01
0.00
0.01
0.02
0.02
0.01
0.01
0.01
0.01
0.00
0.01
0.01
0.01
0.01
0.01
0.01
3.96
2.90
0.04
0.04
0.03
0.01
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.00
0.00
0.05
0.01
0.01
0.01
0.34
1.19
0.91
3.60
0.25
0.09
0.03
0.11
0.18
0.07
0.10
0.09
0.14
0.05
0.05
0.04
0.04
0.04
0.04
0.04
0.04
0.89
0.66
0.16
0.15
0.14
0.06
0.09
0.12
0.13
0.08
0.06
0.08
0.08
0.03
0.03
0.47
0.13
0.12
0.08
0.24
0.58
1.08
4.09
0.08
0.04
0.01
0.07
0.08
0.07
0.03
0.04
0.04
0.01
0.01
0.02
0.02
0.03
0.02
0.02
0.01
1.67
1.14
0.08
0.07
0.06
0.06
0.09
0.07
0.04
0.03
0.05
0.05
0.04
0.01
0.01
0.25
0.06
0.04
0.05
0.12
0.14
1.51
5.25
0.01
0.01
0.00
0.03
0.02
0.06
0.00
0.01
0.00
0.00
0.00
0.00
0.00
0.01
0.01
0.00
0.00
5.74
3.37
0.02
0.02
0.01
0.06
0.09
0.03
0.00
0.00
0.04
0.02
0.01
0.00
0.00
0.08
0.01
0.01
0.02
and
Yuan
Regime by Turian
approach
1
2
3
3
1
1
0
1
1
1
1
1
0
0
0
0
0
0
0
0
0
3
3
0
0
0
1
1
1
0
0
1
1
1
0
0
1
1
1
1
(1977)
approach
Regime by
experiments
2
2
2
2
1
1
1
1
2
2
2
2
2
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
1
1
1
1
1
1
2
2
2
2
2
Industry needs quick and easily implementable solutions at design phase to accurately
calculate the pressure drop and optimum pipe diameter. The model derived from the first
principle is no doubt the best solution. But in the scenario where the basic principles of regime
95
identification model accounting all the interactions for slurry flow is absent, the numerical
model may be promising to give some quick and easy solutions for prediction of slurry flow
regime.
This work considers a systematic approach using robust hybrid ANN-DE, ANN-GA and SVRDE techniques to build a correlation for regime identification from available experimental
data. A data bank has been made by collecting experimental data from the available literature
(800 measurements covering a wide range of pipe dimensions, operating conditions and
physical properties).
3.5. Development of the artificial neural network (ANN) based

correlation
The development of the ANN-based correlation was started with the collection of a large
databank. The next step is to perform an artificial neural network classification, and to validate
it statistically.
After extensive literature survey all physical parameters that influence identification of
regimes are put in a so-called wish-list. Thus, the input variables such as pipe diameter,
particle diameter, solids concentration, solid and liquid density and viscosity and velocity of
flowing medium have been chosen to predict different regimes in slurry pipeline. Some
portion of the input and output data used for ANN classification is shown in Table 3.2.

Present study focuses on classification and identification of two major regimes namely
heterogeneous suspension and moving bed regime. All commercial slurry transport is in
heterogeneous region as it contributes lowest power cost per ton of slurry transported. The
method developed here can be extended for classification of other regimes also. Initially all the
data related with these two regimes were collected from open literature. Seven parameters
were identified as input (Table 3.2) to ANN and the +1 or -1 is put as target. The output was
designated as +1 for heterogeneous suspension flow regimes data and -1 for moving bed
96
Ch
hapter3
reegimes data. As the magnitudde of inputs grreatly differs frrom each otherr, they are norm
malized
inn -1 to +1 scalle. 75% of totaal dataset was chosen randoomly for traininng and rest 255% was
seelected for validdation and testting.
Thhese data thenn exposed to hybrid
h
ANN-D
DE model described in chappter 2. The AN
NN-DE
m
model
optimizess the five param
meters of ANN
N architecture ( namely numbber of nodes inn hidden
laayer, the activaation functionss in input and output layer, the
t learning raate and ANN training
t
algorithm) and selects
s
on its own
o
the best aarchitecture annd training algoorithm for the present
hown in figure 3.5 and predicction error wass summarized in
i Table
daata . The model output was sh
3.4. Out of all thhe possibilitiess, BFGS algorrithm with seveen number of nnodes in hiddeen layer
nd log sigmoidd and linear fuunction in inpuut and outputt layer has em
merged out as the
t best
an
soolution (with lo
owest AARE) for the presennt case. The othher performancce parameters namely
ex
xecution time and
a storage reqquirement is neeglected, as theey are not impportant for this type of
stuudy. The low
w prediction error (AARE
E-0.03%) mayy be considerred as an exxcellent
peerformance con
nsidering the poor
p
understandding of slurry flow
f
phenomenna and large databank
fo
or training com
mprising of widee varieties of systems.
s
1.5
1
0.5
0.5
1
7
13
19
25
31
37
43
49
55
61
67
73
79
85
91
97
103
109
115
121
127
133
1
1.5
Reecordnumber
Expeerimentalregim
me
Caalculatedregim
me
Figuree 3.5: Experim

mental Vs pred
dicted flow reggime for BFGS algorithm
The output waas designated as +1 for heeterogeneous suspension flow
w regime andd -1 for
(T
saaltation regime)
97
Table 3.4: Prediction error by hybrid ANN-DE based model
Prediction performance by ANN-DE model

training
testing
AARE
0.000315
0.000320
Sigma
0.000833
0.000853
0.9999
0.9999
Optimum number of nodes
Input activation function
Log sigmoid function
Output activation function
Linear
Optimum learning rate
0.042
Best training algorithm
BFGS
Recall that, all the 800 experimental data collected from open literature was exposed earlier to
Turian Yuan formulas for regime identification and AARE calculated was 25%. The present
work has reduced the misclassification error from 25% to 0.03%. The advantage of the present
method is that user doesnt have to calculate the CD, settling velocity, Froude number, R01,
R02, R03, R12, R13, and R23 etc as in the case of Turian Yuans approach to evaluate the
regime. In the present approach, regime will be evaluated more accurately from the basic
slurry flow data (pipe diameter, solids concentration, fluid velocity etc.) effortlessly and user
will be relieved to calculate others parameters stated above. Once regime has been evaluated
correctly, appropriate pressure drop correlations (equation 3.3-3.6) can be selected for the each
regime.
The weight, bias and final equations to determine the regime for any solid-liquid slurry flow in
pipelines are summarized in Table3.5. With the help of the equation at the bottom of Table
3.5, any user can easily identify whether the slurry flow regime falls in heterogeneous regime
98
Ch
hapter3
orr it falls in mov

ving bed regim
me. If it is founnd that the flow
w is in movingg bed regime, velocity
v
off slurry flow can
c be increaseed to make it in
i heterogeneo
ous regime. Thhis will ensuree lowest
prressure drop annd power consuumption.
Table 3.5: Set of
o equations an
nd fitting paraameters for neeural network
k correlations
(i=7,, j=7, k=1)
on performan
nce of hybrid
d ANN-GA model
m
3..5.3 Predictio
In
n a separate stuudy, same inpuut-output data was exposed to ANN-GA m
model as descrribed in
ch
hapter 2. All the
t calculation
ns are same except GA which has been used in place of DE
algorithm to opttimize the AN
NN meta param
meters. The AN
NN-GA algoritthm produces exactly
i Table 3.4. This
T proves thaat both ANN-D
DE and ANN-G
GA algorithm are
a most
saame results as in
likkely to find thee global optimaa as compared to ANN algoriithm.
99

In a separate study, the same data set was exposed to ANN algorithm only (without the DE or
GA algorithm) and the different parameters based on exhaustive search was tried to be
optimized. But, it was not possible to reach the best solutions starting from arbitrary initial
conditions. Specially, the optimum choice of learning rate is very difficult to arrive at after
starting with some discrete value. Many times the solutions got stuck up in sub optimal local
minima. These experiments justified the use of a hybrid technique for ANN parameter tuning.
The best prediction after exhaustive search along with ANN parameters was summarized in
Table 3.6. From this table it is clear that even after 100,000 runs, the ANN algorithm is unable
to locate the global minima and the time of execution is 4.3 hrs in Pentium 4 processor. On the
other hand, the hybrid ANN-DE technique is able to locate the global minima with 2000 runs
within 0.5 hr. The prediction accuracy is also much better. Moreover it relieves the non expert
users to choose the different parameters and find an optimum ANN meta parameters with a
good accuracy.
Table 3.6: Comparison of performance of ANN-DE hybrid model Vs ANN model
Performance criteria
Prediction performance
by Hybrid ANN-DE
by ANN model only
model
AARE
0.000320
0.00451
Sigma
0.000853
0.0022
0.9999
0.98
Execution time (hr)
0.5
4.3
100
Chapter3
3.6 Development of the support vector machine (SVM) based

correlation

The development of the SVM-based correlation was started with the collection of a large
databank. The next step was to perform an SVM classification, and to validate it statistically.
Same set of input-output data (like ANN, refer section 3.5.1) was used to train the SVM
algorithm.
3.6.2 Prediction performance of hybrid SVM-DE model
As the magnitude of inputs greatly differs from each other, they are normalized in -1 to +1
scale. 75% of total dataset was chosen randomly for training and rest 25% was selected for
validation and testing. These data then exposed to hybrid SVM-DE model described in chapter
2. The SVM-DE model optimize the four parameters of SVM architecture (namely C, type of
loss function, kernel type and kernel parameter) and select on its own the best architecture for
the present data. The model output and prediction error was summarized in Table 3.7. Out of
all the possibilities, exponential radial basis function (erbf) with kernel width 1 has emerged
out as the best solution (with lowest AARE) for the present case. The other performance
parameters namely execution time and storage requirement is neglected, as they are not
important for this type of study. The low prediction error (AARE-0.03%) may be considered
as an excellent performance considering the poor understanding of slurry flow phenomena and
large databank for training comprising various systems.
101
Table 3.7: Prediction error by hybrid SVM-DE based model

Prediction performance by SVM-DE model
training
testing
AARE
0.000375
0.000380
Sigma
0.000837
0.000858
0.9999
0.9999
10,000
Kernel type
Exponential radial
Type of loss function
Huber loss function
Optimum kernel parameter
basis function
3.7. Conclusion
ANN regression methodology with a robust parameter tuning procedure has been described in
this work which can be used effortlessly where phenomenological model is difficult to
develop. The method employs three hybrid approaches for minimizing the generalization error.
Superior prediction performances were obtained for the case study of regime identification and
a comparison with selected correlations in the literature showed that the developed ANN or
SVM correlation noticeably improved prediction of slurry flow regime over a wide range of
operating conditions, physical properties, and pipe diameters. The proposed hybrid technique
also relieve the non-expert users to choose the meta parameters of ANN or SVM algorithm for
a case study and find out optimum value of these meta parameters on its own . The results
indicate that ANN or SVM based technique with the DE or GA based parameters tuning
approach can yield excellent generalization and can be advantageously employed for a large
class of regression problems encountered in process engineering.
102
Chapter3
Nomenclature
C:
Volume fraction of solids in the suspension (-)

*
CD :
Drag co-efficient at terminal settling velocity (-)
D:
Pipe diameter (m)
dp:
Particle diameter ( m)
Fr:
Froude number(-)
fsl:
Darch Weisbach friction factor for slurry(-)
fw:
Darch Weisbach friction factor for carrier fluid(-)
K :
Constant used in equation 3.2(-)
Rep :
Particle Reynold number, (-)
VT :
Terminal settling velocity ( m/s)
Greek symbols
:
Constant used in Durand Condolios equation(-)
s :
Viscosity of slurry( Pa. s)
f :
Density of fluid( kg/m3)
s :
Density of solid(kg/m3)
Abbreviations
ANN:
DE:
GA:
Genetic algorithm
SVM:
SVR:
103
References
1.
Babcock H.A., Heterogeneous flow of heterogeneous solids Advances in solid liquid flow
in pipes and its application, Pergamon Press, N.Y,125-148 (1970)
2.
Durand R., Condolios E, Compterendu des Deuxiemes journees de L Hydraulique, Paris,

societe Hydrotechnique de France,29-55 (1952)
3.
Ellis, H. S., and G. F. Round. ,Laboratory studies on the flow of nickelwater

suspensions. Canadian Mining and Metallurgical Bulletin, 56, 773781. (1963).
4.
Gillies R.G.,Hill K.B.,Mckibben M.J.,Shook C.A., Solids transport by laminar Newtonian

flows, Powder Technology, 104, 269277 (1999)
5.
Hsu F.L., Flow of non-colloidal slurries in pipelines: Experiments and numerical

simulations PhD Thesis, University of Illinois, Chicago , (1987)
6.
Karamanev . D.G., Equations for the calculations of terminal velocity and drag coefficients
of
solid
spheres
and
gas
bubbles,
Chemical
engineering
communications,147,75-84.(1996)
7.
Kaushal D.R.,Tomita Yuji, Solids concentration profiles and critical velocity in pipeline
flow of multi sized particulate slurries, International Journal of Multiphase Flow, 28,
1697-1717 (2002)
8.
King . R.P., Introduction to practical fluid flow, Butterworth- Heinemann, Burlingtom,

MA., 59-67.(2002)
9.
Newitt D.M.,Richardson J.F.,Abbott M ,Turtle R.B., Hydraulic conveying of solids in

horizontal pipes, Trans. Instn Chem. Engrs, 33, 93-113 (1955)
10. Oroskar A.R.,Turian R.M., the hold up in pipeline flow of slurries, AIChE J., 26, 550-558
(1980)
11. Shen, H. W. ,Sediment transportation mechanismTransportation of sediments in pipes.
Journal Hydraulics Division Am. Soc. Civ. Eng., 96, 15031538.( 1970.)
12. Shook C. A., Pipelining solids: The design of short distance pipelines Paper presented at
the Symposium on pipeline transport of solids, The Canadian society for chemical
engineering, Toronto., (1969)
13. Thomas D. G., Transport characteristics of suspensions: Part VI Minimum transport
velocity for large particle size suspensions in round horizontal pipes, AIChE J, 8, 373-378
(1962)
104
Chapter3
14. Thomas, D. G. ,Transport characteristics of suspensions, Part IX. Am Inst. Chem. Eng.
Journal, 10, 303308. (1964).
15. Turian R.M.,& Yuan T.F, flow of slurries in pipelines, AIChE J., 23, 232-243 (1977)
16. Turton. R. and Levenspiel. O., A short note on the drag correlations for spheres, Powder
technology,47,83-86, (1986)
17. Wasp E.J,Kenny J.P.,Gandhi R.L., Solid liquid flow slurry pipeline transportation, Trans
tech Publ., Rockport, mass., (1977)
18. Wicks, M. ,Transportation of solids of low concentrations in horizontal pipes. In Advances
in SolidLiquid Flow in Pipes and Its application, edited by I. Zandi. New York:
Pergamon Press. (1971).
19. Wilson W.E., Mechanics of flow with non colloidal inert solids, Trans ASCE, 107, 15761594 (1942)
20. Zandi I.,Govatos G., Heterogeneous flow of solids in pipelines, Proc. ACSE, J. Hydraul.
Div., 93(HY3), 145-159 (1967)
105
Critical velocity
CHAPTER 4
Critical Velocity of Slurry Flow in Pipeline
Abstract
The critical velocity is one of the important parameter that must be accurately known for the
optimum design of a slurry transportation pipeline. The significance of this velocity is that it
represents the lowest speed at which slurry pipelines can operate and corresponds to lowest
pressure drop for transporting slurry. The magnitude of this velocity is strongly dependent on
particle size, particle density, solids concentration and pipe diameter. Many attempts have
been made to propose a generalized correlation for the prediction of critical velocity.
However, available correlations fail to predict the critical velocity over a wide range of
conditions. Based on a databank of around 800 measurements collected from the open
literature, a method has been proposed to predict the critical velocity using artificial neural
network (ANN) and support vector regression (SVR) modeling. The method incorporates
differential evolution and genetic algorithm technique for efficient tuning of ANN & SVR meta
parameters. A comparison with selected correlations in the literature showed that the
developed ANN & SVR method noticeably improved the prediction of critical velocity (i.e.
reduces the prediction error from average 25% to 6.5%) over a wide range of operating
conditions, physical properties, and pipe diameters.
Keywords: ANN; SVR; Critical velocity; Differential evolution; Slurry
106
Chapter4
4.1 Introduction
The critical velocity (Vc) is defined as the minimum velocity at which the solids form a bed at
the bottom of the pipe demarcating flow from fully suspended flows. It is also referred to as
the minimum carrying or the limiting deposition velocity or velocity corresponds to lowest
pressure drop and perhaps the most important transition velocity in slurry transport. Operation
velocities below critical velocity are uneconomical as pressure drop rises considerably. Apart
from its obvious danger of plugging the pipeline, it also produces excessive erosion in the
lower part of the pipe. A system containing particles of wide granulometric range results in
suspension of smaller particles keeping the larger particles deposited in the pipeline. Thus the
long term stability of a heterogeneous flow depends on the reliable prediction of the critical
velocity.
Four distinct regimes were found existent [namely flow with stationary bed, flow with a
moving bed, heterogeneous suspension (sometimes called asymmetric flow) and homogeneous
suspension (sometimes called symmetric flow)] in slurry flow in pipeline depending upon the
average velocity of flow. The four regimes of flow described in chapter 3 can be represented
by a plot of the pressure gradient versus the average velocity of the mixture (Figure 4.1).
The transitional velocities are defined as:
V1: velocity at or above which the bed in the lower half of the pipe is stationary. In
the upper half of the pipe, some solids may move by saltation or suspension.
V2: velocity at or above which coarser particles forming a moving bed.
V3 or Vc: velocity at or above which all particles move as an asymmetric suspension

and below which the solids start to settle and form a moving bed.
V4: velocity at or above which all solids move as a symmetric suspension.
107
Criticcal velocity
ure 4.1: Plot of

o transitional mixture veloccity with presssure drop
Figu
Frrom figure 4.11 it is evident that pressure drop curve paasses through m
minima (pointt 3) and
opperation at thiss velocity will ensure lowest power consum
mption. For alll practical purpposes of
thhe slurry transpport, the mixtuure velocity is kept
k
above thee V3 and thus iit is very impoortant to
knnow this velociity. From poweer consumptionn point of view
w, this is the veelocity where pressure
p
drrop is minimum
m i.e. corresponnds to lowest ppower consump
ption in slurry transport.
Despite of signnificant researrch efforts, pprediction of critical
c
velociity and other design
e
an adeequate solid ssuspension is still an open problem for design
paarameters to ensure
en
ngineers. Desiign of slurry pipelines reliees on empiriccal correlationns obtained frrom the
ex
xperimental daata. These correelations are proone to great unncertainty oncee one departs from
fr
the
lim
mited databasee that supportss them. All off the above em
mpirical equatioons when testeed with
crritical velocity data for differeent systems coollected from op
pen literature, it was found th
hat they
prredicted the crritical velocityy above 25% error on an average.
a
With this backgrouund, the
m
motivation
of thhe present worrk is to developp a critical vellocity correlatiion having reasonably
loow prediction error
e
for a widee range of studyy.
108
Ch
hapter4
Fiigure 4.2: Sch

hematic repressentation of th
he boundaries between the flow
f
regimes for
f
setttling slurries in horizontal pipelines
Too facilitate thee design and scale
s
up of piipelines and sllurry pumps, tthere is a neeed for a
co
orrelation that can predict sluurry critical veelocity over a wide
w
range off operating connditions,
phhysical propertties and particlle size distributtions. Predictioon of critical velocity
v
is very
y crucial
ass operating the pipeline near to critical veloocity ensures lowest power cconsumption. Industry
I
neeeds quick and
d easily implem
mentable solutiions. The modeel derived from
m the first prin
nciple is
noo doubt the beest solution. Bu
ut in the scenaario where the basic principlles of critical velocity
v
m
model
accountin
ng all the interractions for sluurry flow is abbsent, the num
merical model may be
prromising one to
o give quick annd suitable soluutions for the prediction
p
of sllurry critical veelocity.
109
Criticcal velocity
Fiigure 4.3 Simp

plified concept of particle d
distribution in a pipe as a fu
unction of volu
umetric
con
ncentration an
nd velocity
4.2 Backgro
ound works
Thhis section revviews the evolu
ution of criticaal velocity mod
dels of Newtonnian slurry flow
ws from
w graded andd uniform partiicle sizes to coomplex mixturres of coarse aand fine particlles. For
well
thhis purpose, som
me equations are
a listed in a historical conttext, to demonnstrate their evolution.
Thhe sheer number of equationss demonstratess how complicaated heterogeneeous flows are.
In
n the past, engiineers have trieed to simplify the complexityy of slurry flow
ws by defining
g certain
traansition veloccities. Attemptt was made to develop siimplified relattions for trannsitional
veelocities with particle size and
a solid conccentration as shown
s
in figurre 4.2 and figgure 4.3
reespectively.
W
With
the use of modern research tools, there is an em
merging approoach of
rej
ejecting the con
ncept of an abrrupt change froom one state of
o flow to anotther and a tenddency to
co
onsider such a change over a band of the sppeed. Differentt approaches haave been develloped to
ex
xamine the miixture of coarrse and fine pparticles from superimposedd layers to tw
wo-layer
m
models.
Figure 4.4
4 shows som
me major correllations of critical velocity off slurry flow avvailable
inn literature.
110
Chapter4
The transitional velocity V3 is extremely important because it is the speed at which the
pressure gradient is at a minimum. Although there is evidence that solids start to settle at
slower speeds in complex mixtures, operators and engineers often referred to transitional
velocity as the speed of deposition.
Wilson approach (1942): The approach to obtain the magnitude of V3 is basically to conduct
a test and measure pressure drop per unit length of pipe. V3 is considered to occur at the
minima, or the point of minimum pressure drop. W. E. Wilson (1942) expressed the pressure
gradient of non-colloidal solids by referring to clean water and by proposing a correction to the
DarcyWeisbach equation. He expressed the consumed power due to friction by the following
equation:
[4.1]
Where,
= head loss due to friction (in units of length)
= DarcyWeisbach friction factor

C1 = constant
Equation 4.1 may also be expressed as,
[4.2]
By differentiating this equation with respect to V, we obtain for the minimal value
[4.3]
[4.4]
Wilson (1942) defined a factor C3 to determine whether the particles will settle to form a bed:
111
Critical velocity
[4.5]
If C3 > 1 most particles with a terminal velocity Vt will stay in suspension. If C3 < 1 most
particles with a terminal velocity Vt will settle out.
Durand and Condolios correlations (1952): Durand and Condolios, 1952 derived the
following equation for V3 for uniformly sized sand and gravel:
2 gDi ( s l )
Vc = V 3 = FL
0.5
[4.6]
Where,
FL
= is the Durand factor based on grain size and volume concentration
V 3 = the critical transition velocity between flow with a moving bed and a heterogeneous flow
Di = pipe inner diameter (m)
= acceleration due to gravity (9.81 m/s)
s= density of solids in a mixture (kg/m3)

l = density of liquid carrier (kg/m3)
Newitt, Richardson, Abbott & Turtle correlations (1955): While the work of Durand and
Condolios (1952) focused on the drag coefficient for sand and gravel, the equations of Newitt
et al. (1955) and Wilson (1942) focused on the terminal velocity. Newitt et al., 1955 preferred
to express the velocity of transition between moving bed and heterogeneous flow in terms of
the terminal velocity of particles,
V 3 = 17Vt
[4.7]
Zandi and Govatos correlations (1967): Zandi and Govatos, 1967 and Zandi, 1971 extended
the work of Durand to other solids and different mixtures. They defined an index number as,
Ne =
V 2CD 0.5
s
CvDig ( 1)
l
[4.8]
112
Ch
hapter4
Wilson,
11942
Durandand
Condolios,
1952
Kaushall,
2002
N
Newittet
al.,
1955
Gilliese
etal.,
1999
9
Crittical
velo
ocity
Wilson,
1992
Zandiand
Govatos,
1967
Durand,,Con
dolios,W
Wors
ter,
19700
SSchillerand
Herbich,
1991
Oroskarand
Turian,
1980
Waspeta
al.
,1970
Figurre 4.4: Major critical velocitty correlation

ns available in literatures
A the critical value
At
v
of Ne = 40, the flow
w transition beetween saltatioon and heteroggeneous
reegime occurs. This
T
statementt infers that w
when Ne < 40 saltation
s
occurrs, and when Ne
N > 40
heeterogeneous flow
fl develops.
Su
ubstituting the value of Ne = 40, V3 may be expressed as,
s
40 C vD ig ( l 1)
V3=
1/ 4
CD
0 .5
[4.9]
113
Critical velocity
This equation is therefore a modified version of equation given by Durand and Condolios. Due
to its simplicity, this method continues to be used by dredging engineers who usually deal with
sand and gravel mixtures of less than 20% concentration by volume.
Durand, Condolios and Worster approach (1970):The saltation and heterogeneous
suspension regimes have been studied most widely and the best known correlation for the
excess pressure gradient due to the presence of solid particles in the slurry is due to Durand,
Condolios and Worster (1970). The Durand-Condolios-Worster correlation for the excess
pressure gradient is,
,
Where,
[4.10]
is a constant and C is the volumetric fraction of solids in the suspension and
the drag coefficient at terminal settling velocity. The value to be used for the constant
is
is
uncertain and values between 65 and 150 are reported in the literature. As this correlation does
not apply to all regimes of flow, the experimental data cannot be used to fix the value more
precisely. Errors of 100 per cent and more in the calculated value of
can result. This is not
as serious as it might appear at first sight since in many cases the excess pressure drop is only
a small fraction of the total pressure drop along the pipe and errors in the value of
are
correspondingly less important. The pipe Froude number is a dimensionless group that
indicates the relative strengths of the suspension and settling tendencies of the particles in the
slurry.
1
Where V slurry velocity, s is specific gravity of solid and D is pipe diameter.
The Froude number is a useful index for the importance of the momentum transfer process
from particles to the pipe wall relative to the direct transfer of momentum from the fluid to the
pipe wall.
The velocity at minimum pressure drop is easy to calculate by finding the point on the curve
that has zero slope and,
114
Chapter4
[4.11]
.
[4.12]
The variation of fw with flow rate has been neglected in the differentiation. The critical
velocity or velocity at minimum pressure drop is given by the solution to the equation and,
[4.13]
Wasp ,Aude ,Kenny,Seiter & Jacques approach (1970): In an effort to represent for more
diluted concentrations, Wasp et al., 1970 proposed a ratio between the particle diameter and
the pipe diameter. Wasp also proposed the use of a modified factor FL' so that,
0. 5
2 gDi ( s l ) dp
Vc = V 3 = FL '
Di
l
(1 / 6 )
[4.14]
Oroskar and Turian approach (1980): Using an analysis based on balancing the energy
required to suspend the particles with that derived from dissipation of an appropriate fraction
of the turbulent eddies, Oroskar and Turian(1980) derived the following critical velocity
correlation :
[4.15]
Where the exponent n is the hindered settling exponent in multi-particle sedimentation as

given by Richardson and Zaki(1954). The values of n fall approximately in the range 2<n<5.
115
Critical velocity
Schiller and Herbich approach (1991): Schiller and Herbich, 1991 proposed the following
equation for the Durand velocity factor based on the
d 50 of the particles:
(6.9d 50)
FL = (1.3Cv0.125 )(1 exp
)
[4.16]
Where d 50 is expressed in mm.

Wilson approach (1992): Another school of researchers particularly led by Shook, Wilson, and
Gilles in Canada, focused on refining the original models of Newitt, which were based on the
terminal velocity. These authors proposed that the compound mixture of coarse and fine particles
can be simplified to what they called stratified flows, in which the fines slide over a moving
bed of coarse solids. Newitt, as expressed by Equation 4.7, had proposed that the deposition
velocity was a mere factor of 17 times the terminal velocity. Wilson (1991) indicated that this
approach was not confirmed by tests for large pipes. The concept he proposed was that the
flow was going through a gradual transition from a fully stratified to a fully suspended flow,
with a gradual change in the pressure gradient. In this partially stratified model, it is argued that
the lifting of particles by turbulence is strongly influenced by their diameter. Thus, the fines tend
to be transported better by the fluid than the coarse particles at the bottom of the pipe. These
particles are transported by eddies, and the largest eddy would be equal to the pipe diameter. The
pipe diameter was therefore proposed as a reference. The resistance to motion of the solid
particles is proposed to be the result of a contact load associated with mechanical sliding friction
in the coarser bed and fluid friction associated with the friction velocity of the carrier fluid. In
this complex environment, Wilson et al. (1992) defined a new settling velocity as,
0.9
2.7
[4.17]
T is terminal settling velocity (m/s) and is carrier luid viscosity

There is no question that the approach of Wilson et al. (1992) is extremely interesting, but it is
more complex than the methods proposed by earlier researchers. It requires the support of
testing, a database, computer software, and a personal computer.
116
Chapter4
Gillies, Hill, Mckibben & Shook approach (1999): With a more complex approach
considering the actual viscosity of the slurry and the density of the particles, Gillies et al., 1999
developed an equation for the Froude number (Fr) in terms of Archimedean number as,
Fr = aArb
Where,
Ar =
[4.18]
4
3l2
Fr = FL
and
dp3 l( s l) g
For Ar > 540, a = 1.78, b = 0.019

For 160 < Ar < 540, a = 1.19, b = 0.045
For 80 < Ar < 60, a = 0.197, b = 0.4
For Ar < 80, the Wilson and Judge, 1976 equation can be used, which expressed the Froude
number as,
dp
Fr = (20.5 ) 2.0 + 0.30 log 10
DiCD
This correlation is useful in the range of 10
[4.19]
dp
3
<
< 10
D
i
C
D
Kaushal & Tomita approach (2002): In their experiments Kaushal et al. (2002) considered
multisized particulate iron ore slimes, copper tailings and zinc tailings slurries flowing through
105 and 55 mm diameter pipelines. The ratio of volumetric concentration at the bottom of pipe
and volumetric static settled concentration of the solid particles
is calculated and
are plotted against efflux concentration, Cvf .It is observed that Cv(0.1)/Cvss varies linearly
with Cvf at deposition velocity. Hence, a linear equation on the basis of least square curve
fitting method was determined for this variation as follows,
.
[4.20]
117
Critical velocity
or at deposition velocity,
.
[4.21]
Hence, they concluded that the deposition velocity for transportation of commercial slurries
through pipeline is the flow velocity corresponding to the bottom concentration equal to three
times the product of efflux concentration and static settled concentration by volume.
Critical velocity- comparison of published correlations

Different researchers have used different variables and variable groupings in their correlations.
In order to compare and to evaluate these correlations , first they were compiled into a
standard form. From literature it is concluded that critical velocity dependence can be
expressed by,
, , ,
, ,
[4.22]
implies that the effect of gravity to be only
Inclusion of the variable grouping
manifested through net settling forces on the particles. In non-dimensional variables equation
4.22 can be written in the following form
Use of reference velocity
[4.23]
1 in equation 4.23 is motivated by the fact that the
critical velocity dependence on pipe diameter is approximately according to D 0.5.

Published critical velocity relationships and equations suggest that most available correlations
can be expressed under the form:
[4.24]
118
Chapter4
In which f(C, s) is mainly a function of particle concentration and k, l and m take different
numerical values for different correlations. Table 4.1 lists published critical velocity
correlations which have been recast in the format of equation 4.24 . Most of the critical
velocity correlations listed in Table 4.1 is empirical except some like Oroskar and Turian
(1980) and Kaushal et al. (2002) which are based on an assumed model for the particle
suspension process.
Table 4.1: Performance of different correlations to predict critical velocity

Sl
no
1
2
3
4
5
6
7
8
9
10
11
12
Author
f(C,s)
%AARE
Wilson (1942)
1.13[Cs/(Cs+1-C)]0.36
(s-1)-0.36
-0.181
0.098
0.081
49.51
Durand and
Condolios (1952)
Newitt et. al.
(1955)
Thomas (1962)
Zandi and Govatos
(1967)
Shook (1969)
Babcock (1970)
Turian and Yuan
(1977)
Wasp et al (1977)
Oroskar and Turian
(1980)
Kaushal et
al.(2002)
Present work
0.7-1.3
36.53
13.8804
-0.5
0.5
93.43
1.3762(s-1)0.1878
-0.408
-0.183
-0.408
29.95
(20C)0.5
-0.25
50.02
2.43 C 0.33
-0.25
34.50
2.2361 C 0.5
-0.25
46.68
8.948 C 0.4779
-0.027
-0.117
39.97
3.3999 C 0.2156
0.167
26.68
2-0.5 [5C(1-C)2n-1]8/15
0.067
25.94
Cannot recast in
present form
22.01
06.45
All the 800 experimental data collected from open literature was also exposed to different
correlations for critical velocity available in open literature and average error (AARE) were
calculated for each of them (Table 4.1). It is obvious upon comparing the various expressions
in Table 4.1 that the qualitative disparities among the various correlations which have been
proposed are quite broad. Most correlations and the experimental data on which they are
based, indicate that the critical velocity has approximately a square root dependence on pipe
diameter (~D 0.5) and a rather weak dependence on particle diameter (~d0), especially for larger
119
Critical velocity
non-colloidal particles. Accordingly, the correlations of Wilson (1942), Newitt et al. (1955),
Thomas (1962) and Turian and Yuan (1977) are in conflict to varying degrees with
experimental evidence. This disparity and large prediction error (above 22%) for critical
velocity motivates this work. The main objective is to reduce this prediction error by applying
the ANN and SVR technique.
4.3. Development of the hybrid ANN-DE and SVR-GA based correlation

The development of the ANN and SVR-based correlation was started with the collection of a
large databank. The next step was to expose this databank to ANN or SVR, and to validate the
prediction performance statistically.
As mentioned earlier, over the years researchers have amply quantified the critical velocity of
slurry flow in pipeline. In this work, about 800 experimental points have been collected from
20 sources of open literature spanning the years 1942 -2002. The data were screened for
incompleteness, redundancies and evident inaccuracies. This wide range of database includes
experimental information from different physical systems to develop a unified correlation for
critical velocity. Table 4.2 indicates the wide range of the collected databank for critical
velocity.
After extensive literature survey all physical parameters that influence critical velocity are put
in a so-called wish-list.
Out of the number of inputs in wish list, we used artificial neural network to establish the
best set of chosen inputs, which describes critical velocity. The following criteria guide the
choice of the set of inputs:
The inputs should be as few as possible.
Each input should be highly cross-correlated to the output parameter.
120
Chapter4
The selected input set should give the best output prediction, which is checked by
using statistical analysis [e.g. average absolute relative error (AARE), standard
deviation].
[Wilson (1942), Durand and Condolios (1952), Newitt et. Al. (1955), Thomas (1962),
Zandi and Govatos (1967), Shook (1969), Babcock (1970), Govier and
Aziz(1972),Turian and Yuan (1977), Wasp et al (1977), Oroskar and Turian
(1980),Hsu(1987), Gillies et al (1999), Kaushal (2002)]
Gypsum-water, Glass-water, Gravel-water, iron-water, iron-kerosene, high density
material-water, iron tailings-water, limestone-water, limonite-water, plastic-water,
potash-brine, sand-ethylene glycol, nickel shot-water, iron powder-water, ore-water
Pipe diameter (m)
0.0127 0.80
Particle diameter (m)
(0.0017 0.868) x 10-2
Liquid density (kg/m3)
770 1350
Solids density (kg/m3)
1150 8900
Liquid viscosity (kg/m-sec)
0.008 1.9 x 10-1
0.005 0.561
Critical Velocity (m/s)
0.18 - 4.56
While choosing the most expressive inputs, there is a compromise between the number of
inputs and prediction. Based on different combinations of inputs, trial and error method was
used to finalize the input set which gives reasonable low prediction error (AARE) when
exposed to artificial neural network.
Based on the above analysis, the input variables such as pipe diameter, particle diameter,
solids concentration, solid and liquid density and viscosity of flowing medium have been
finalized to predict critical velocity in slurry pipeline. Table 4.3 shows some typical data used
for artificial neural network and support vector regression model.
121
Critical velocity
Table-4.3: Typical input and output data for ANN or SVR training
Particle
dia,cm
0.0130
0.0100
0.2200
0.2200
0.8680
0.1830
0.0300
0.0140
0.0090
0.0060
0.1113
0.3346
0.0390
0.0410
0.0410
0.0540
0.0540
0.0500
0.0500
0.0500
0.0500
0.0200
0.0200
0.0490
0.0490
0.0490
0.0250
0.0250
0.2040
0.0850
0.0850
0.0242
0.0242
0.0242
0.0250
0.0250
0.0400
0.0400
0.0400
0.0242
0.0242
Solid
conc. ,
vol.
fraction
0.0280
0.3000
0.3040
0.3040
0.1650
0.0560
0.1700
0.0220
0.1250
0.0140
0.1920
0.1000
0.4750
0.3220
0.3830
0.3600
0.4200
0.0500
0.1200
0.1800
0.4200
0.1800
0.2400
0.1200
0.1800
0.2400
0.0050
0.0050
0.0250
0.2150
0.2150
0.0080
0.0300
0.0700
0.3000
0.3180
0.0300
0.0620
0.1370
0.0300
0.0700
Solid
density ,
gm/cc
1.8340
2.8200
1.6700
1.6700
1.5300
1.5300
3.3600
2.6900
3.0000
3.1000
4.5500
4.5500
1.9840
1.9840
1.9840
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6500
2.6000
2.6000
2.6000
2.6500
2.6500
2.6500
2.6500
2.6500
2.6000
2.6000
Fluid
viscosity,
Fluid
density, gm/cm- Pipe dia,
sec
cm
gm/cc
0.9980
0.0098
7.6200
0.9820
0.0130
5.0000
1.3250
1.5200
7.6200
1.3250
1.5200
10.1600
0.9980
0.0098
20.8000
0.9980
0.0098
4.0000
0.9980
0.0098
10.3000
0.9980
0.0098
20.7000
0.9980
0.0098
20.7000
0.9980
0.0098
14.9000
0.9980
0.0098
8.2500
0.9980
0.0098
15.0000
1.1460
0.0114
5.2200
1.1420
0.0120
10.7600
1.1420
0.0120
10.7600
1.3500
0.0560
5.2450
1.3500
0.0560
5.2450
1.3500
0.0560
5.2450
1.3500
0.0560
5.2450
1.3500
0.0560
5.2450
1.3500
0.0560
5.2450
1.1320
0.3820
5.2450
1.1320
0.3820
5.2450
1.0950
0.0575
5.2450
1.0950
0.0575
5.2450
1.0950
0.0575
5.2450
0.9980
0.0098
5.1500
0.9980
0.0098
5.1500
0.9980
0.0098
15.0000
0.9980
0.0098
5.0800
0.9980
0.0098
5.0800
0.9990
0.0100
7.6000
1.0000
0.0100
7.6000
1.0000
0.0101
7.6000
0.9980
0.0098
10.3000
0.9980
0.0098
10.3000
0.9980
0.0098
10.3000
0.9980
0.0098
10.3000
0.9980
0.0098
10.3000
1.0000
0.0100
7.6000
1.0000
0.0101
7.6000
Critical
velocity,
cm/sec
140.2100
350.0000
90.5000
214.8800
242.7100
55.3200
133.4000
221.7000
321.1000
114.8500
247.9200
300.0000
146.3000
123.4000
123.4000
80.7700
74.6800
54.8600
62.4800
68.5800
76.2000
79.2500
70.1000
115.8200
118.8700
119.4800
72.4200
87.9400
220.7200
147.7100
116.8900
82.4000
122.0000
137.0000
107.5400
103.5200
347.5000
218.5600
230.4500
122.0000
137.0000
122
Chapter4

As the magnitude of inputs and outputs greatly differ from each other, they are normalized in 1 to +1 scale. 75% of total dataset was chosen randomly for training and rest 25% was selected
for validation and testing.
Six parameters were identified as input (Table 4.3) for ANN and the critical velocity is put as
target. These data then exposed to hybrid ANN-DE model described in chapter 2. After
optimization of five ANN parameters described above, the model output was summarized in
Table 4.4. Out of all the possibilities, Marquard Levenburg algorithm with eight number of
nodes in hidden layer and sigmoidal and linear function in input and output layer has emerged
out as the best solution (with lowest AARE) for the present case. The low AARE (6.5%) may
be considered as an excellent prediction performance considering the poor understanding of
slurry flow phenomena and large databank for training comprising various systems. The parity
plot of experimental and predicted critical velocity is shown in figure 4.5 as a goodness of
hybrid ANN-DE performance.
The weight, bias and final equations to calculate the critical velocity for any solid liquid slurry
flow in pipelines are summarized in table 4.5. After predicting the critical velocity accurately,
the slurry velocity can be maintained at very near (20% more as a rule of thumb, King (2002))
to this critical velocity to minimize the pressure drop and power consumption.
123
Criticcal velocity
T
Table
4.4: Preediction error by hybrid AN
NN-DE based model
m
P
Prediction perrformance by ANN-DE mod
del
training
ttesting
AARE
0.0642
00.0645
Sigma
0.061
00.0625
0.940
00.931
Optimum
m number of nodes
Input activation
a
funcction
siggmoidal functioon
Output activation fun

nction
Linear
Optim
mum learning rate
r
0.028
Best trraining algoritthm
Marrquard Levenbuurg
Fiigure 4.5: Exxperimental Vs predicted

d critical vellocity for Maarquard Levvenburg
allgorithm by ANN-DE
A
based
d model
124
Chapter4
Table 4.5: Set of equations and fitting parameters for neural network correlations
(i=7,j=8,k=1)
W i,j
0.9227
6.5474
4.4409
0.8863
0.1354
-0.0249
-1.8078
41.588
6.9049
7.7732
0.4609
-0.6561
0.3059
1.0487
-5.3788
0.2704
0.2029
-1.6524
10.304
3.5264
-0.5292
-0.9048
-1.4978
-2.8234
-0.4559
-4.0595
2.2811
-1.2910
-0.1691
-1.3236
-2.5551
3.6650
3.1036
0.1726
2.6369
-0.6306
1.2714
4.8205
-5.9361
22.544
-0.4560
-0.6831
2.3384
-0.6111
-1.0833
11.862
-1.6822
2.8742
12.058
-3.7727
-1.5977
5.2927
0.8488
15.600
3.3558
14.882
4.6141
0.3134
-0.2594
-3.6637
-0.8366
0.2123
0.7791
-0.1320
-1.7843
W j,k
Where Ui=input variable, Hj=hiddenlayer output and U7=1, H9=1 (bias)

In a separate study, we exposed the same dataset to ANN algorithm only (without the DE
algorithm) and tried to optimize the different parameters based on exhaustive search. We
found that it was not possible to reach the best solutions starting from arbitrary initial
conditions. Especially, the optimum choice of learning rate is very difficult to arrive at after
starting with some discrete value. Many times the solutions got stuck up in sub optimal local
minima. These experiments justified the use of a hybrid technique for ANN parameter tuning.
The best prediction after exhaustive search along with ANN parameters was summarized in
table 4.6. From the table it is clear that even after 100,000 runs, the ANN algorithm is unable
to locate the global minima and the time of execution is 3.5 hrs in Pentium 4 processor. On the
other hand, the hybrid ANN-DE technique is able to locate the global minima with 800 runs
within 0.5 hr. The prediction accuracy is also much better. Moreover it relieves the non expert
users to choose the different parameters and find an optimum ANN meta parameters with a
good accuracy.
125
Critical velocity
by Hybrid ANN-DE
model
Prediction performance by
ANN model only
testing
testing
AARE
0.0645
0.0751
Sigma
0.0625
0.0655
0.931
0.901
Execution time (hr)
0.5
3.5
4.4.3 Prediction performance of hybrid SVR-GA model prediction

As the magnitude of inputs and outputs greatly differ from each other, they are normalized in 1 to +1 scale by the formula
75% of total dataset was chosen randomly for training and rest 25% was selected for validation
and testing.
Same six parameters were identified as input (Table 4.2) for SVR and the critical velocity is
put as target. These data then exposed to hybrid SVR-GA model described in chapter 2. After
optimization of five SVR parameters described above, the model output was summarized in
Table 4.7. The SVR-GA algorithm also converged to same average error AARE (6.5%) like
ANN. The optimum value of SVR meta parameters were summarized in table 4.8. From table
4.8 it is clear that there exist almost 4 different feasible solutions which lead to same
prediction error.
126
Chapter4
Table 4.7: Prediction error by hybrid SVR-GA based model

Prediction performance by SVR
training
testing
AARE
Sigma
R
0.0642
0.061
0.940
0.0645
0.0625
0.931
Table 4.8: Optimum parameters obtained by hybrid SVR- GA algorithm

Sr No
1
2
3
4
943.82
2298.16
3954.35
6180.98
0.50
0.36
0.48
0.54
Kernel
type
erbf
erbf
erbf
erbf
Type of loss
function
e- insensitive
e- insensitive
e- insensitive
e- insensitive
Kernel
parameter
2
3
4
5
AARE
0.0645
0.0645
0.0645
0.0645
The final equation is as follows: critical velocity=K (i, j)*Beta+Bias

Where K (i, j) = kernel function=
,
for i=1 to n and j=1 to m

Where, n= number of test record and m=number of training record (600 in this case)
tstx(i,6) is a (n X 6) test matrix for testing input( which user wants to test for calculating
critical velocity) and should be arranged in a sequence similar to table 4.3
trnx (j,6) is a (600 X 6) training matrix for training input data which is arranged in a sequence
similar to table 4.3
beta (600,1) is a (600 X 1) matrix calculated during training and Bias is zero for this case
p1 is width of rbfs (which can be found as kernel parameter in table 4.8).
127
Critical velocity
The trnx () and beta () matrix are not reproduced here for sake of brevity and can be supplied
upon request.
4.5 Dependence of critical velocity with model input parameters

To evaluate the dependence of critical velocity with model input parameters, a test matrix (30
X 6) was made where only one input parameter ( say particle diameter) is varied from its
minimum to maximum value and all the remaining five parameters kept at their average
value. This matrix is then exposed to the already developed hybrid SVR-GA model and critical
velocity was calculated for each of the 30 records. Same procedure repeated for all other five
input parameters and variations of critical velocity with each of the input parameters were
calculated. While plotting, critical velocity is plotted slightly differently as shown in figure 4.6
to 4.10. This is to evaluate and compare the SVR-GA model prediction with published
correlations.
Critical velocity vs ratio of density of solid to liquid: Figure 4.6 represents the change of
critical velocity with the ratio of density of solid to liquid. From Figure 4.6 it is clear that more
velocity is needed to keep the particle in suspension when the solid is heavier and this makes
sense from physics.
Critical velocity vs pipe diameter :Most correlations and the experimental data indicate that
the critical velocity has approximately a square root dependence on pipe diameter (~D
0.5
).
Figure 4.7 confirms the findings of the experiments qualitatively with slight change of
dependence on pipe diameter (~D 0.348).
Critical velocity vs particle diameter: The results of Figure 4.8 suggest that the dependence of
the critical velocity on particle diameter is quite weak, which is approximately like d0.042. This
observations are supported by the findings of Oroskar and Turians (1980) correlation where
the critical velocity dependence on particle diameter is given by d 0 and on pipe diameter
D
is
0.6
Critical velocity vs solid concentration: A substantial body of experimental data suggests that
the critical velocity-concentration relation possesses a maximum. SVR-GA model also predict
a maxima as depicted in figure 4.9 and it is consistent with the experimental evidence. The
maximum in the critical velocity-concentration relationship occurs because of increased
particle interaction with concentration. The critical velocity initially increases with
concentration, but as more particles are added hindered settling effects become increasingly
128
Ch
hapter4
innfluential and counteract paarticle settlingg, which explaains the reverrsal in the Vc vs. C
reelationship.
Cr
Critical
velocityy vs dimension
nless parameteer (
): To evaluaate the dependdence of
otther parameters and for comp

parison sake, critical
c
velocity
y is plotted aggainst a dimenssionless
paarameter (
) in figgure 4.10. Thee figure suggessts very weak ddependence off critical
veelocity with diimensionless group

g
(
which suppoorts the experrimental
finndings of Orosskar and Turiann and Wilson where

w
exponennts are 0.067 annd 0.098 respecctively.
Fig 4.6: Variation of critical velocity with

w density raatio
129
Criticcal velocity
Fig 4.7: Varriation of critical velocity with pipe diameeter
Fig 4.8: Variaation of criticaal velocity with

h particle diam
meter
130
Ch
hapter4
Fig
F 4.9: Variattion of criticall velocity with
h solid concenttration
Fig 4.10: Variattion of criticall velocity with dimensionlesss group

131
Critical velocity
4.6 Comparison with other published correlations

All the 800 experimental dataset collected from open literature was also exposed to different
formulas and correlations for critical velocity available in open literature and AARE were
calculated for each of them (Table 4.1). From Table 4.1, it is evident that prediction error of
critical velocity has reduced considerably in the present work.
4.7. Conclusion:
ANN and SVR regression methodology with a robust parameter tuning procedure have been
described in this work. The method employs a hybrid ANN-DE and SVR-GA approach for
minimizing the generalization error. Superior prediction performances were obtained for the
case study of critical velocity. The performance of the present work in comparison with
selected correlations in the literature showed that the developed ANN and SVR correlation
noticeably improved prediction of critical velocity over a wide range of operating conditions,
physical properties, and pipe diameters. The proposed hybrid technique also relieve the nonexpert users to choose the meta parameters of ANN or SVR algorithm for the case study and
find out optimum value of these meta parameters on its own . The results indicate that ANN
and SVR based technique with the DE or GA based parameters tuning approach described in
this work can yield excellent generalization and can be advantageously employed for a large
class of regression problems encountered in process engineering.
132
Chapter4
Nomenclature
:
Drag coefficient at terminal settling velocity.
DarcyWeisbach friction factor (-)
Head loss due to friction (m)
Liquid viscosity (Kg/ms)
C:
The volumetric fraction of solids in the suspension (-)
C1 :
Constant used in equation 4.1 (-)
CD :
Drag co-efficient ( used in equation 4.19 & 4.24)
CV :
Volumetric concentration of solid (-)
Cv(0.1) :
Volumetric concentration of solid at the 0.1D distance from the bottom of

pipe
Cvf :
Efflux concentration
Cvss :
Volumetric static settled concentration of the solid particles (-)
Cw :
Concentration by weight (%)
d :
particle diameter (m)
d50 :
Sieve passage diameter for 50% of the particles (m)
Di :
Inner diameter of pipe (m)
FL :
DurandCondolios coefficient to determine the deposition velocity(-)
Fr:
fw :
Froude number(-)
Fanning friction factor of carrier liquid(-)
133
Critical velocity
g:
Acceleration due to gravity (m/sec2)
L:
length of conduit (m)
Ne :
Index number used in equation 4.8
s:
Specific gravity of solid (-)
V:
Velocity of mixture (m/s)
V1 :
Velocity for which flow with a stationary bed is started (m/s)
V2 :
Velocity for which flow with a moving bed is started(m/s)
V3 :
Deposition velocity or velocity above which solids start to move(m/s)
V4 :
Velocity above which all solids move as a pseudohomogeneous

mixture(m/s)
Vc :
Critical velocity (m/s)
Vmin :
Velocity at minimum pressure drop (m/s)
Vt :
Terminal velocity of solid particle (m/s)
Greek Symbols
l:
Density of liquid carrier (kg/m3)
s:
Density of solids in a mixture (kg/m3)
Constant used in equation 4.10
Pressure loss factor (-)
134
Chapter4
References
1.
Babcock H.A., Heterogeneous flow of heterogeneous solids Advances in solid liquid flow
2.
Durand R., Condolios E, compterendu des Deuxiemes journees de L Hydraulique, Paris,

3.
Gillies R.G.,Hill K.B.,Mckibben M.J.,Shook C.A., Solids transport by laminar Newtonian

4.
Govier,
G. W., Aziz. K. , The Flow of Complex Mixtures in Pipes, New York: Van
Nostrand Reinhold., (1972)

5.
Hsu F.L., Flow of non-colloidal slurries in pipelines: Experiments and numerical

simulations , PhD Thesis, University of Illinois, Chicago , (1987)
6.
Kaushal D.R.,Tomita Yuji, Solids concentration profiles and critical velocity in pipeline
flow of multisized particulate slurries, International Journal of Multiphase Flow, 28, 16971717 (2002)
7.
King R.P.,Introduction to practical fluid flow,Butterworth Heinemann, Burlington,

M.A.(2002)
8.

9.
Oroskar A.R.,Turian R.M., the hold up in pipeline flow of slurries, AIChE J., 26, 550-558
(1980)
10. Richardson, J.F., Zaki, W.M., Sedimentation and fluidization: part-1. Transactions of the
Institution of Chemical Engineers 32,35 53. (1954).
11. Schiller R.E.,Herbich P.E.,
Sediment transport in pipes., Handbook of Dredging,New
York: McGraw-Hill., (1991)

(1962)
14. Turian R.M.,& Yuan T.F, Flow of slurries in pipelines, AIChE J., 23, 232-243 (1977)
135
Critical velocity
16. Wasp E.J.,Aude T.C.,Kenny J.P.,Seiter R.H.,Jacques, R.B.,
Deposition velocities
transition velocities and spatial distribution of solids in slurry pipelines, Proc. Hydro
transport 1, BHRA Fluid Engineering, Coventry, UK, paper H42 53-76 (1970)
17. Wilson K.C., Addie G.R, Clift R, Slurry transportation using centrifugal pumps, New
York: Elsevier Applied Sciences., (1992)
18. Wilson W.E., Mechanics of flow with noncolloidal inert solids, Trans ASCE, 107, 15761594 (1942)
19. Worster. B.,Bain A.G.,Bonnington S.T., The hydraulic transport of solids by pipeline,
Pergamon press, (1970)
20. Zandi I., Advances in solid liquid flow in pipes and its application, Pergamon Press, N.Y.,
(1971)
Div., 93(HY3), 145-159 (1967)
136
Chapter5
Chapter 5
Hold up in slurry flow in pipeline
Abstract
In solid liquid slurry flow in pipelines different layers of solids move with different speed.
Coarse particles move at the bottom of the pipe while finer particles move at the top section of
the pipe. Hold-ups are due to velocity slip of layers of particles of larger sizes, particularly in
the moving bed flow pattern. Due to this slip in velocity, in-situ concentrations are not same as
the concentrations in which the phases are introduced or removed from the system. The
variation of in-situ concentrations from the supply concentrations is referred to holdup
phenomenon. The holdup effect is measured by the holdup ratio, given by ratio of the average
in-situ concentration and mean discharge concentration. Very few correlations with limited
applicability are available in literatures to predict hold up ratio in solid-liquid slurry flow. An
attempt has been made in the present study to develop generalized correlations for hold up
from available experimental data by using advance computational techniques namely artificial
neural network (ANN) and support vector regression (SVR). Differential evolution (DE) and
genetic algorithm (GA) are used to tune the parameters of ANN and SVR. The developed ANN
and SVR correlations well predicted (2.5% error) the hold up over a wide range of operating
Keywords: Hold up, Artificial neural network, Differential evolution, Genetic algorithm,
Support vector machine, Slurry flow regime, Slurry flow
137
Hold up in slurry flow in pipelines
5.1 Introduction
The tendency that the solid particles have to settle under the influence of gravity has a
significant effect on the behavior of slurry that is transported in a horizontal pipeline. The
settling tendency leads to a significant gradation in the concentration of solids in the slurry.
The concentration of solids is higher in the lower sections of the horizontal pipe. The extent of
the accumulation of solids in the lower section depends strongly on the flow velocity of the
slurry in the pipeline. With higher velocity, higher is the turbulence level and greater is the
ability of the carrier fluid to keep the particles in suspension. It is the upward motion of eddy
currents transverse to the main direction of flow of the slurry that is responsible for
maintaining the particles in suspension. At very high turbulence levels the suspension is
almost homogeneous with very good dispersion of the solids while at low turbulence levels the
particles settle towards the floor of the channel and can in fact remain in contact with the flow
and are transported as a sliding bed under the influence of the pressure gradient in the fluid.
Thus the various flow patterns described in earlier chapter are observed in slurry flow in
pipeline. Hold-ups are basically due to velocity slip of layers of particles of larger sizes,
particularly in the moving bed flow pattern.
In solid-liquid multiphase flow, the separate phases move at different average velocities and
the in-situ concentrations are not same as the concentrations in which the phases are
introduced or removed from the system. The variation of in-situ concentrations from the
supply concentrations is referred to hold up phenomenon. The hold up effect is measured by
the hold up ratio, given by the ratio of the average in-situ concentration and mean discharge
concentration as,
.
.
[5.1]
Though there are large numbers of literatures available regarding slurry flow, but most of them
deal with pressure drop and critical velocity. Detailed measurement of the hold up in slurry
systems are scarce and has only been addressed in a handful of studies. There has been no
generally applicable method for the prediction of hold up for all flow patterns of solid liquid
slurry flow. In the homogeneous suspension flow regime, it is common to assume that hold up
is negligible and therefore in-situ and the input concentration are considered identical. This is
not practically correct but is a suitable approximation in many cases. Hold up has been clearly
observed, especially in the flow regimes involving bed formation. For coarse particles, the
138
Chapter5
difference in fluid and particle mean velocities are quite appreciable and this difference is
referred as the slip velocity. It is probable that a large contribution to the apparent slip velocity
is due to fluid flowing at high velocity through the portion of low solids concentration in the
pipe. Obviously hold up plays an important role in the failure of many empirical correlations
in predicting head loss in flow regimes involving bed formation.
Despite of research efforts, prediction of hold up and other design parameters to ensure an
adequate solid suspension is still an open problem for design engineers. Design of slurry
pipeline relies on empirical correlations obtained from the experimental data. These
correlations are prone to great uncertainty once one departs from the limited database that
supports them. All of the available empirical equations when tested with hold up data for
different systems collected from open literature, it was found that they predicted the hold up
with more than 25% error on an average. This may be due to the fact that empirical equations
derived for coal-water slurry flow may not be fitted well for sand-water slurry flow and vice
versa. With this background, the motivation of the present work is to develop a hold up
correlation having reasonably low prediction error for a wide range of study.

As might be expected from the great variety and complexity of the flow behavior of liquidsolid and gas-solid systems, a host of empirical correlations relating to the flow patterns and
the pressure gradient have been developed. Also some correlations of restricted applicability
related to the holdup phenomenon have been proposed. None of these correlations is generally
applicable to all flow patterns and to both gas and liquid carriers.
Newitt et al. (1955) conducted speed measurements of a slurry mixture in a horizontal pipe. In
the case of light Plexiglas pipe, zircon or fine sand did not result in local slip; particles and
water moved at the same speed. However, for coarse sand and gravel, they observed
asymmetric suspension and a sliding bed. They also observed that in the upper layers of the
horizontal pipe, the concentrations of larger particles were the same as for finer solids, but
were marked by differences in the magnitude of the discharge rate of the lower layers. This
means that a local slip velocity exists and increases from zero in the upper 1/3 or so of the pipe
to a maximum value at the bottom of the pipe.
139

The holdup phenomenon has been clearly observed, especially in the moving bed flow pattern,
for the flow of both liquid-solid and gassolid mixtures, but relatively few reliable quantitative
measurements of it have been made. Bonnington (1959) studied the sand-water slurry flow
having a mean size of 0.048 inch in a 1.5 inch pipe. The marked increase in hold up, as the
liquid velocity is decreased and as the moving bed flow pattern is reached, is evident from
their data.
Soo and Trezek (1966) reported hold up data for the dilute phase transport of 30 micron
magnesia particles in air at loadings below a mass ratio of three. The holdup ratio is less than
unity because the dispersed solid particles tend to be more concentrated near the center of the
pipe and their average velocity exceeds that of the carrier gas.
Hsu (1987) has a detailed study of hold up in his PhD dissertation. The findings of Hsu are
listed below: He has studied the velocity and concentration profile of 190 micron coal slurry
(concentration 32.6 volume %) in 20.85 cm pipeline at different velocity range (1.37 to 2.59
m/sec). He concluded that with the decrease of slurry velocity, hold up is marginally increased.
In his another study (Hsu,1989), he considered fine gypsum particles ( 38.2 micron) in 5.263
cm pipeline with different velocity range (1.11 to 4.70 cm) and solid concentration (10.7 to
30.6 volume%). He concluded that despite large concentration change, hold up ratio is near to
unity and practically no slip velocity exists due to homogeneous suspension by very small
particles. Hsu also studied the slurry of sand particles (165 micron) in different velocity (1.663.76 m/s) and concentration range (8.4 27.2 volume %) with various pipe diameters (5.15 cm
to 49.50cm). He concluded that hold up ratio decreases with increasing average concentration.
This surprising finding results from the fact that at a given mean mixture velocity the
concentration gradient is less steep for the higher average concentration and consequently the
mixture velocity distribution is less asymmetric across the vertical plane. The contribution to
the integrand
the pipe cross section, where
in equation 5.1 from the middle and upper regions of
,
,
is larger, is relatively smaller for the case
of a lower average concentration because the local value of C is low. Consequently the holdup
is higher. Moreover, the local concentrations in the lower part of the pipe cross section are
about the same for both low and high average concentrations and therefore the contribution to
the integrand from this region of the pipe cross section further trends to make the value of the
integral smaller in the case of the lower average concentration. He also noticed that the
average solid concentration increases to very high values and the resulting severe damping of
the turbulence may lead to reduction in particle transport and therefore a reversal in the trend.
140
Chapter5

As seen from the literatures, the correlations to calculate the hold up ratio in solid liquid slurry
are rare and very few experimental studies have been made to develop hold up data. Prediction
of hold up in slurry flow remains as unexplored area of research. In the present work we
collected all the available hold up data from open literature and tried to make ANN and SVR
based correlations which will help the design engineer to predict hold up at early design phase.
DE and GA were used to tune the various meta parameters of ANN and SVR models so that
optimum architecture or models can be evolved.
The development of the ANN or SVR based correlation was started with the collection of a
large databank.

To test the suitability of different correlations we have collected 200 experimental hold up data
from open literature spanning over the years 1959 to 1987. The majority of data are from
Schrick (1973); Scarlet and Grimley (1974); Roco and Shook (1982); Gillies (1983) etc. listed
and summarized in PhD thesis of Hsu (1987).The data were screened for incompleteness,
redundancies and evident inaccuracies. This wide range of database includes experimental
information from different physical systems to develop a unified correlation for the hold up.
Table 5.1 indicates the wide range of the collected databank for the hold up.
141

Table 5.1: System and parameter studied*
Slurry system: Coal water, Copper ore water, Sand water, Gypsum water, Glass
propanol and Gravel water
0.04 0.495
Pipe diameter (m)

-6
Particle diameter (m)*10

3
Liquid density (kg/m )
38.3 13000
1000 1250
Solids density (kg/m )
1370 2844
-3
Liquid viscosity (Pa.s)*10
0.12 4
Velocity (m/s)
1.05 4.81
0.0372 0.333
Max packing conc. (volume fraction)
0.58-0.8
Pressure drop (Pa/m)
99.9 4727.7
* References : Bonnington 1959; Soo and Trezek 1966 ; Schrick et al.,1973; Scarlet &
Grimley,1974; Roco & Shook, 1982; Govier & Aziz, , 1982; Gillies et al,1983; Hsu,1987

After the extensive literature survey, all physical parameters that influence the holdup are put
in a so-called wish-list. Based on different combinations of inputs, a trial and error method
was used to finalize the input set which gives a reasonably low prediction error (AARE) when
solids concentration, solid and liquid density and viscosity of flowing medium, flow velocity,
solid concentration, maximum packing concentration and pressure drop have been finalized to
predict the hold up in a slurry pipeline. Table 5.2 shows some typical data used for ANN &
SVR training.
142
Chapter5
Table-5.2: Typical input and output data for ANN & SVR training
Input parameters
Output
Pipe
Particle
Liquid
Solids
Liquid
Max
Velocity
Solids
Pressur
dia
dia
density
density
viscosi
packing
(m/s)
concentra
e drop
Hold
up
(cm)
(micron)
(gm/cc)
(gm/cc)
ty (cp)
concentra
tion
(Pa/m)
ratio
tion
(volume
(volume
fraction)
fraction)
5.26
38.3
2.33
1.00
0.69
1.11
0.1070
294.1
1.001
5.26
38.3
2.33
1.00
0.69
1.33
0.3060
542.9
1.000
20.85
190.0
1.37
1.14
0.78
2.59
0.3260
266.5
1.002
20.85
190.0
1.37
1.14
0.78
2.34
0.3270
226.3
1.003
20.85
190.0
1.37
1.14
0.78
2.01
0.3330
177.3
1.007
20.85
190.0
1.37
1.14
0.78
1.37
0.3270
99.9
1.064
5.15
165.0
2.65
1.00
0.58
1.66
0.0741
666.2
1.133
5.15
165.0
2.65
1.00
0.58
1.66
0.1694
901.3
1.104
26.30
165.0
2.65
1.00
0.58
2.90
0.0932
261.6
1.105
26.30
165.0
2.65
1.00
0.58
3.50
0.0921
334.1
1.086
26.30
165.0
2.65
1.00
0.58
2.90
0.1759
305.7
1.080
26.30
165.0
2.65
1.00
0.58
3.50
0.1726
382.1
1.066
26.30
165.0
2.65
1.00
0.58
3.50
0.3241
526.1
1.043
49.50
165.0
2.65
1.00
0.58
3.16
0.0943
143.0
1.103
49.50
165.0
2.65
1.00
0.58
3.76
0.0923
186.1
1.083
49.50
165.0
2.65
1.00
0.58
3.76
0.2602
254.7
1.034
15.85
190.0
2.65
1.30
0.58
2.50
0.1365
475.2
1.099
15.85
190.0
2.65
1.30
0.58
2.50
0.2899
630.9
1.035
15.85
190.0
2.65
0.12
0.58
3.00
0.1267
648.9
1.184
15.85
190.0
2.65
0.12
0.58
2.90
0.2775
866.7
1.081
4.00
580.0
1.25
2.27
4.00
0.65
1.05
0.0612
845.0
1.307
26.30
13000
2.65
1.00
0.8
3.20
0.0372
842.5
2.420
26.30
13000
2.65
1.00
0.8
4.00
0.0440
989.5
2.045

5.4.1 Prediction performance of hybrid ANN-DE model:
As the magnitude of inputs and outputs greatly differ from each other, they are normalized in 1 to +1 scale. 75% of total dataset was chosen randomly for training and the rest 25% was
selected for validation and testing.
143
Nine parameters were identified as input (Table 5.2) for ANN and the hold up is put as a
target. These data were then exposed to a hybrid ANN-DE model described in chapter 2. After
optimization of five ANN parameters described earlier, the model output was summarized in
Table 5.3. Out of all the possibilities, Fletcher Reeves update algorithm with 6 (six) nodes in a
hidden layer and a log sigmoid and linear function in the input and output layer has emerged as
the best solution (with the lowest AARE and along with the highest correlation coefficient,
R) for the present case. The low AARE (2.5%) may be considered as an excellent prediction
performance considering the poor understanding of the slurry flow phenomena and a large
databank comprising various systems. The parity plot of the experimental and predicted hold
up is shown in Figure 5.1 as a goodness of the hybrid ANN-DE performance.
Table 5.3: Prediction error by the hybrid ANN-DE based model
Prediction performance by ANN
training
testing
AARE
0.024
0.025
Sigma
0.028
0.030
0.985
0.980
Log sigmoid function
Linear
0.045
FletcherReeves update
144
Ch
hapter5
Fiigure 5.1: Exp

perimental Vs predicted holld up for FletccherReeves u
update algorith
hm
Table 5.4: Set
S of equation
ns and fitting parameters fo
or neural netw
work correlatiions
145

The weight, bias and final equations to calculate the hold up for any solid liquid slurry flow in
pipelines are summarized in Table 5.4. Same data was applied to ANN-GA algorithm also and
the solution is found converge to almost same results as ANN-DE solution. After predicting
the holdup accurately, the slurry velocity can be maintained to reduce this hold up to minimize
the pressure drop and power consumption.

In a separate study, we exposed the same dataset to ANN algorithm only (without the DE or
GA algorithm) and tried to optimize different parameters based on the exhaustive search. We
found out that it was not possible to reach the best solutions starting from arbitrary initial
conditions. The optimum choice of the learning rate is especially very difficult to arrive at
after starting with some discrete value. Many times the solutions got stuck up in sub optimal
local minima. These experiments justified the use of a hybrid technique for ANN parameter
tuning. The best prediction after the exhaustive search along with ANN parameters was
summarized in Table 5.5. From this table it is clear that even after 100,000 runs, the ANN
algorithm is unable to locate the global minima and the time of execution is 4 hrs in Pentium 4
processor. On the other hand, the hybrid ANN-DE technique is able to locate the global
minima with 2000 runs within 1 hr. The prediction accuracy is also much better. Moreover, it
relieves the non-expert users to choose the appropriate parameters and find optimum ANN
meta parameters with a good accuracy.
Hybrid ANN-DE model
Prediction
performance by ANN
model only
testing
testing
AARE
0.025
0.030
Sigma
0.030
0.035
0.980
0.979
Execution time (hr)
146
Chapter5
5.4.3. Prediction performance of hybrid SVR-DE model

Similarly, as in ANN inputs and outputs are normalized in -1 to +1 scale. 75% of total dataset
was chosen randomly for training and rest 25% was selected for validation and testing. Same
nine parameters were identified as input (Table 5.2) for SVR and the hold up is put as target.
These data then exposed to hybrid SVR-DE model described in chapter 2. After optimization
of five SVR parameters described earlier, the model output was summarized in Table 5.6. As
expected, SVR also converged to global minima and AARE was equal to 2.5%, similarly as
observed from ANN results. The optimum value of SVR meta parameters were summarized in
Table 5.7. From table 5.7 it is clear that there exist almost 4 different feasible solutions which
lead to same prediction error.
Table 5.6: Prediction error by hybrid SVR-DE based model
Training
Testing
AARE
0.024
0.025
0.028
0.030
0.985
0.980
In a separate study, we exposed the same dataset to SVR algorithm only (without the DE
algorithm) and tried to optimize the different parameters based on exhaustive search. We
found that it was not possible to reach the best solutions starting from arbitrary initial
conditions. Specially the optimum choice of C and is very difficult to arrive at after starting
with some discrete value. The best prediction after exhaustive search along with SVR
parameters was summarized in Table 5.8. From the table it is clear that even after 720,000
runs, the SVR algorithm is unable to locate the global minima and the time of execution is 4
hrs in Pentium 4 processor. On the other hand, the hybrid SVR-DE technique is able to locate
the global minima with 2000 runs within 1 hr. The prediction accuracy is also much better.
Moreover it relieves the non expert users to choose the different parameters and find optimum
SVR meta parameters with a good accuracy.
147
H
Hold
up in slurrry flow in pip
pelines
Table 5.7:
5 Optimum
m parameters oobtained by hyybrid SVR- DE
D algorithm
Sr No
Kernel
Type of loss
l
Keernel
type
function
paraameter
AA
ARE
3.82
3
0.0051
erbf
e- insensittive
0.0
025
5.16
5
0.0037
erbf
e- insensittive
0.0
025
7.35
7
0.0045
erbf
e- insensittive
0.0
025
8.98
8
0.0054
erbf
e- insensittive
0.0
025
Table 5.8: Comparison of

o performancce of SVR-DE
E hybrid modeel Vs SVR mod
del
Prediction
nperformanceby
hybridSVRDEmode
Performanccecriteria
AARE
Sigma
R
Executiiontime(hr)
0.0225
0.0330
0.9880
1
Predictio
onperformancebySVR
modelw
withoutDEtuningafter
exxhaustivesearch
0.027
0.035
0.979
2.5
sion
5.5 Conclus
Thhere is dearth of
o correlations with wide appplicability to prredict hold up ratio in slurry flow in
piipeline. Hold up
u ratio is requuired at early phase of designn of pipelines. In
I the present study it
w shown thatt ANN and SVR can be a superior alterrnative of emppirical correlattions to
was
prredict hold up ratio in slurry flow. The devveloped ANN and
a SVR correelations well prredicted
thhe hold up (2.5
5 % error) overr a wide rangee of operating conditions, phhysical propertiies, and
piipe diameters.
148
Chapter5
Nomenclature
AARE:
Average absolute relative error
C:
Capacity control factor in SVR algorithm
Cdischarge :
Discharge solid concentration
Cin-situ :
In-situ solid concentration
Erbf:
Exponential radial basis function
R:
Cross-correlation co-efficient
Ui :
ith input of ANN model
Vm,average :
Average mixture velocity
Vm,in-situ
In situ mixture velocity
Wi,j :
Weightage of input layer of ANN model
Wj,k :
Weightage of output layer of ANN model
Greek symbols
:
Standard deviation of error
SVR parameter
Abbreviations
ANN:
DE:
GA:
Genetic algorithm
SVM:
SVR:
149
References
1.
Bonnington, S.T., Experiments on the hydraulic transport of mixed-size solids, British

Hydromechanics research Association,637, Cranefield,Bedford,(1959)
2.
Gillis, R., Husband, W.H.W., Small, M., and Shook, C.A., Concentration distribution
effects in a fine particle slurry, proc 8th technical conference, slurry transport
association,131-142,(1983)
3.
Govier, G.W., Aziz, K., The Flow of Complex Mixtures in Pipes. Krieger, Malabar, FL.(
1982)
4.
Hsu FL, Turian RM, Ma TW, Flow of non colloidal slurries in pipelines. AIChE journal :
35: 429-442, (1989)
5.
Hsu, F.L, Flow of non-colloidal slurries in pipelines: Experiments and numerical

simulations, PhD Thesis, University of Illinois, Chicago (1987).
6.

7.
Roco, M.C. and Shook, C.A., New approach to predict concentration distribution of fine
particle slurry flows, PCH Physicochem hydrodynamics journal,2,special issue,(1982)
8.
Scarlett,B. and Grimley, A., Particle velocity and concentration profiles during hydraulic
transport in a circular pipe, Proceedings of the 3rd International Conference on Hydraulic
Transport of Solids, BHRA Fluid Engineering,, Cranfield UK Paper 3, pp 23 (1974)
9.
Schriek, W,Smith, L.G., Hass D.B. and husband W.H.W, Experimental studies on the
hydraulic transport of limestone, report E73-10, saskatchewan research council,
Saskatoon, Canada, august (1973)
10. Soo, S.L., and Trezek, G.J. , Ind. Eng. Chem. Fundam.,5,388(1966)
150
Chapter6
CHAPTER 6
Pressure drop of slurry flow in pipeline

Abstract
Pressure drop is the most significant parameter for solid liquid transport in mineral and other
solids handling industry. Power consumption and subsequently the whole economics of the
hydro-transport depend on it. For this reason, pipeline design is mainly based on optimization
in pressure drop and initial investment. There are large number of empirical and semi
empirical correlations available in literatures to predict pressure drop. Most of these
equations have been developed based on limited data comprising of uniform or narrow sizerange particles with very low to moderate concentrations. As the correlations are empirical in
nature, their applicability is limited e.g. the correlations developed for sand-water slurry flow
do not produce promising results when they were applied on coal-water slurry flows. These
correlations are prone to great uncertainty as one departs from the limited database that
supports them. When all the major correlations are exposed to the large experimental data
collected from open literature (800 measurements covering a wide range of pipe dimensions,
operating conditions and physical properties), the average prediction error found is 25-50%.
This poor prediction basically motivated this work. An attempt has been made in the present
study to develop generalized correlations for pressure drop by using advance computational
technique namely artificial neural network (ANN) and support vector regression (SVR).
Differential evolution and genetic algorithm is used to tune the parameters of ANN and SVR. A
comparison with selected correlations in the literature showed that the developed ANN and
SVR correlations noticeably improved the prediction of pressure drop (reduce the error to
12.7%) over a wide range of operating conditions, physical properties, and pipe diameters.
Keywords: Artificial neural networks, Genetic algorithm, Support vector regression,

Differential evolution; Slurry pressure drop
151
6.1 Introduction
An engineer exposed with the literature on slurry flows would be astonished by the number of
different equations related to pressure drop. Since the work of the French scientists Durand
and Condolios in 1952, and British scientists Newitt et al. in 1955, engineers and scientists
have tried continuously to develop appropriate equations for pressure drop and friction losses.
A number of factors interact in a horizontal pipe. The flow takes the form of different regimes,
and includes everything from a simple stationary bed at low speed to a pseudo homogeneous
flow at high speeds. For each regime, equations have been developed over the years to account
for the mean particle diameter, the diameter of the conduit, the density of the particles, their
drag coefficient, the speed of flow, etc. As a result, there are many angles from which a
heterogeneous flow of settling solids can be examined.
Two principal schools of heterogeneous slurry flows have been developed over the last 50
yearsone around the DurandCondolios approach (1952) and the other around the Newitt
approach (1955). The former evolved gradually until Wasp (1977) and modified for multilayer
compound systems. The latter gradually evolved to yield the two-layer model. There was no
consensus as to which model to use. Some have argued that the Wasp method was more
suitable for coal, whereas the two-layer model was better for sand. This is based on the
number of papers published on two-layer models for sand slurry mixtures, emanating from the
great interest in Canadian oil sands.
Despite the large area of application, the available models describing the suspension
mechanism do not completely satisfy engineering needs. The behavior of solids in liquid
flowing through pipelines has been the subject of continuing investigation since the turn of the
19th century. From the literature (Kaushal et al., 2002; Gillies et al., 1991, 1999, 2000; Ghanta
et al., 1996; Doron et al., 1987; Govier et al., 1982), it was found that attempts to solve slurry
flow problems may be divided into two main categories. In the first approach, one begins from
the experimental facts and generalizes known correlations for some parameters by dimensional
analysis, without providing an insight into the flow mechanism (e.g. Newitt et al.,1955; Zandi
& Govatos, 1967 ; Turian & Yuan, 1977; and many others). In the second approach, one starts
from the basic equations of motion and numerically solves these for some situations with
physical or mathematical assumptions for different terms. Numbers of such phenomenological
modeling have been proposed by Wilson (1976, 1988), Wilson & Pugh (1988), Roco & Shook
(1984, 1985), Doron et al. (1987) and many others.
152
Chapter6
Both of the above methods have their own limitations generating out of inherent complexity
and poor understanding of two-phase flow systems. A successful predictive model with sound
understanding of the fundamentals of particle laden turbulent flow, including all significant
interactions and the ability to integrate these quantitatively is not developed till today
satisfactorily as seen from various literatures.
Pressure drop is the most significant parameter for solid liquid transport in mineral and other
solids handling industry. Power consumption and subsequently the whole economics of the
hydro-transport depend on it. Therefore, pipeline design significantly stresses on optimization
in pressure drop and initial investment. In spite of the lack of detailed fundamental knowledge
required for the formulation and modeling of multiphase turbulent flows, the need to predict
slurry behavior handled in various industries has motivated the work, aimed at obtaining quite
reliable solutions. Various investigations (Wasp et al., 1970; Doron et al., 1987; Wilson et al.,
1988; Gillies et al., 1991; Ghanta, 1996 etc.) have tried to show and propose correlation
relating pressure drop with other most probable parameters of solid-liquid flow namely solids
density, liquid density, particle size, solids concentration, pipe diameter, viscosity of flowing
media, velocity of suspension etc.
To facilitate the design and scale up of pipelines and slurry pumps, there is a need for a
correlation that can predict slurry pressure drop over a wide range of operating conditions,
physical properties and particle size distributions. When all the major correlations are exposed
to the large experimental data collected from open literature (800 measurements covering a
wide range of pipe dimensions, operating conditions and physical properties), the average
prediction error is found to be 25-50%. This large prediction error has motivated the work .An
attempt has been made in the present study to develop a generalized correlation for pressure
drop by using advance computational technique namely artificial neural network and support
vector regression. Industry needs quick and easily implementing solutions. The model derived
from the first principle is no doubt the best solution. But in the scenario, where the basic
principles for pressure drop modeling accounting all the interactions for slurry flow is absent,
the numerical model may be promising to give some quick and easy solutions for prediction of
slurry pressure drop.
This paper presents a systematic approach using robust hybrid ANN-DE and SVR-DE
techniques to build a pressure drop correlation from available experimental data. This
correlation has been derived from a broad experimental data bank collected from the open
literature (800 measurements covering a wide range of pipe dimensions, operating conditions
and physical properties).
153
Preessure drop of slurry flow in pipeline
6.2 Backgro
ound works
Thhis section reviews the evoluution of modelss for Newtoniaan slurry flows from well gradded and
unniform particlee sizes to compplex mixtures of
o coarse and fine
fi particles. F
For this purpose, some
eq
quations are listed in a histoorical context, to demonstraate their evoluttion (figure 6.1). The
shheer number off equations dem
monstrate how complicated heterogeneous fflows are.
1952
Duranda
D
ndCondolios
1955
Newittet
N
al.
1967
Zandiand
Z
dGovatos
1977
Turianan
T
dYuan
1977
Waspeta
W
al.
1991
Shookan
S
dRoco
1997
Wilson
W
2002
Kaushal&
K
&Tomita
c
foor pressure drrop published in literature
Figure 6.1 Major correlations
W
When
settling slurry
s
is transpported, significcant gradients in the solids cconcentration develop
d
unnder the influeence of gravitty. The solid particles that are present inn the slurry generate
g
ad
dditional momeentum transferr processes andd it must be coonsidered to deevelop models for the
traansfer of mom
mentum from thhe slurry to the pipe wall. Thee presence of solid
s
particles increase
i
thhe rate of mom
mentum transferr between the fluid and the containing
c
wallls of the conduuit. The
traansported partiicles frequentlyy strike the waalls. Doing so, the particles trransfer the mom
mentum
too the wall and dissipate
d
somee of their kinetiic energy. The particles also transfer some of their
m
momentum
to thhe fluid if theyy move faster than the fluid in their neighbbourhood and receive
m
momentum
from
m the fluid whhen move slow
wer than the fluid
f
in their nneighbourhoodd. These
154
Ch
hapter6
prrocesses ensurre a continuouus exchange oof momentum between the fluid and thee walls,
beetween the fluiid and the partiicles and betweeen the particlees and the walll. This is illusttrated in
Fiigure 6.2. The net result of this
t
model is tthe existence of
o an additionaal path throughh which
m
momentum
can be transferred from the fluidd to the solid wall
w and that is the indirect paath from
fluuid to particlees and from particles
p
to thee wall. This path acts in paarallel with thee direct
traansfer path from the fluid to the walls. This additional transfer meechanism leads to an
inncrease in the pressure
p
drop.
mentum betweeen the fluid and the wall during slurryy flows
Fiigure 6.2 Traansfer of mom
t
through
a pipe
Siince the first slurry
s
pipeliness were built, eengineers and scientists
s
havee tried to correelate the
loosses with slurrry to those of an equivalent volume of waater. The hydraaulic friction gradient
g
fo
or the slurry (iim) is higher thhan the hydrauulic friction grradient for an equivalent vollume of
w
water.
It was innitially assumeed that the friiction losses would
w
increasee in proportionn to the
voolumetric conccentration of solids. A term im was then defined
d
as the frictional head
d of the
m
mixture
in equiv
valent meters (or
( feet) of thee carrier fluid (e.g.,
(
water) peer unit of pipe length.
Thhere are numbber of models to predict fricction losses an
nd they are esssentially basedd on the
innteraction forcees between soliids and liquid ccarrier. Some of
o them use thee drag coefficieent
155

while others use the terminal velocity of the solids, and some consider the solids to be moving
as a bed with a layer of liquid and suspended fines above it.
To reflect the increase in friction head due to presence of solids, Durand and Condolios
(1952) proposed a non dimensional ratio,
[6.1]
The frictional gradient,
is defined as
[6.2]
Using the approach developed by Durand and Condolios, the Darcy-Weisbach friction factor
for the slurry is correlated with the friction factor for an equivalent volume of water by the
following equation:
[6.3]
Wasp et al. (1977) deducted that the coefficient Kf is between 80 and 150, depending on the
slurry. The most common value is actually 81 for most sands according to Govier and Aziz
(1972).
Babcock (1968) was very critical over all equations of pressure gradients based on
the work of Durand and Condolios or their followers. Geller and Gray (1986) did not
agree with Babcocks criticisms and spelled out some of the misgivings. Govier and Aziz
(1972) did confirm that errors of the order of 40% have occurred in predicted values of Z,
but for all the purposes, these equations were the best available till the early
1970s. Herbich (1991) agreed with the value of 81 for most dredged sands and gravel.
Sand and gravel are typically dredged and then pumped at a volumetric concentration smaller
than 20%.
156
Chapter6
6.2.1 Methods Based on the Drag Coefficient of Particles

Based on the analysis of test data from 11 references of sand for particle sizes ranging up to 1
inch (25.4 mm), in pipes with a diameter range from 38.1 mm to 558.8 mm, and in volumetric
concentration up to 22%, Zandi and Govatos (1967) derived an equation for the index number
Ne in terms of the volumetric concentration, and some empirical parameters:
[6.4]
where
They plotted this function against a parameter to express head loss as,
[6.5]
where K and m are constants. On a logarithmic scale they found,

For
>10, K=6.3 and m=-0.354
For
<10, K=280 and m=-1.93
The dramatic change in values of K and m at
= 10 has encouraged researchers to develop
more sophisticated models.

Despite its simplicity, this method continues to be used by dredging engineers who
usually deal with sand and gravel mixtures of less than 20% concentration by volume.
Shook et al. (1991) modified Zandis equation by proposing in-situ concentration of particles
Ct rather than volumetric concentration:
[6.6]
[6.7]
157

They measured a magnitude of m = -1 for one single type of coal in different pipe sizes. They
measured different values of K for different coals. The in-situ concentration Ct remained
constant with speed, but the volumetric concentration of solids Cv that could be increased with
V.
identifiable flow regimes. Using the experimental data, Turian and Yuan (1977) established
that the excess pressure gradient in each flow regime can be correlated using an equation of
the form,
f sl f w = KC fw C
*
D
Fr
[6.8]
the coefficients K,,, and have values that are specific to each flow regime. Using
experimental data gathered from experiments in each flow regime, the best available values of
these parameters in each flow regime are given by
Sliding bed (regime 0)
fsl fw = 12 . 13 C 0 . 7389
fw 0 . 7717
* 0 . 4054
D
Fr
1 . 096
[6.9]
Saltation (regime 1)
fsl fw = 107 . 1 C 1 . 018 fw 1 . 046 C
* 0 . 4213
D
Fr
* 0 . 1677
D
Fr
1 .. 354
[6.10]
Heterogeneous suspension (regime 2)
fsl fw = 30 . 11 C 0 . 868 fw
1 . 200
0 . 6938
[6.11]
Homogeneous suspension (regime 3)
fsl fw = 8 . 538 C 0 . 5024
fw 1 . 428 C
*
D
0 . 1516
Fr
0 . 3531
[6.12]
Fairly consistent trends in the variation of the correlating parameters can be seen in the four
correlations. Here C*D is drag co-efficient at the terminal settling velocity of slurry.
158
Chapter6
6.2.2 Models Based on Terminal Velocity

Newitt et al. (1955) conducted tests in pipes smaller than 150 mm (6 in) and proposed to
express Z in terms of the terminal velocity (instead of the drag coefficient).
[6.13]
For small pipes, K2 = 1100.

For solids of different sizes, Newitt suggested a weighted mean diameter as
[6.14]
Hayden and Stelson (1968) proposed a modification of the Durand-Condolios equation using the
terminal velocity instead of the drag coefficient:
100
[6.15]
Geller and Gray (1986) pointed out that the equations of Durand, Newitt, and Babcock
converged when m = -1.
Newitt et al. (1955) minimized the importance of lift forces when a bed cannot form because
of lift forces on particles. However, the work of Bagnold (1954, 1955, and 1957) indicated that
the submerged weight of particles separated from the bed was transmitted to the bed or the pipe
wall under the same conditions. Thus, mechanical friction can contribute to head loss.
6.2.3 Friction losses for compound mixture in horizontal heterogeneous flows
Many slurries resulting from dredging, cyclone underflow, and tailings disposal are not
pumped with single-sized particles. Some authors such as Newitt et al. (1955) proposed the
use of a weighted average particle diameter but Hill et al. (1986) proposed that the particles
159

should be divided. The finer particles would move as a heterogeneous flow, while the
coarser particles would move as a bed by saltation. Hill et al. (1986), Wasp et al. (1977), and
Gaesler (1967) demonstrated that this approach worked well when applied to pumping watercoal mixtures.
The compound or heterogeneous-homogeneous system is the most important and
most common in slurry transportation. It involves coarse and fine particles both. The fines
move as a homogeneous mixture while the remainder moves as a heterogeneous mixture.
To conduct this analysis, the rheological and physical properties of the solids must be
known.
This method was pioneered by Wasp et al. (1977) and further developed in some respects by the
stratification model described later on. Wasp et al. (1977) extended the Durand approach to a
concept of multiple and superimposed layers of particles of different sizes and volumetric
concentration, with a logarithmic concentration distribution. This method uses the drag
coefficient of the particles as an important parameter. The heterogeneous mixture or bed motion
is based on the method of concentration in relation to a reference layer.
The method proposed by Wasp et al. (1977) can be summarized as follows:
1. The total size fraction is divided into a homogeneous fraction using Durands equation.
2. The friction losses are calculated for the homogeneous fraction based on the rheology of the
slurry, assuming Newtonian flow.
3. The friction losses are then calculated for the heterogeneous fraction using Durands equation.
4. A ratio C/CA for the size fraction of solids based on friction losses estimated in steps 2 and 3 is
determined.
5. Based on the value of C/CA, the fraction size of solids in homogeneous and heterogeneous
flows are determined.
Steps 2 to 5 are re-iterated till convergence of the friction loss.
This method developed by Wasp et al. (1977) has been used very successfully over the last 25
years for Newtonian slurries.
When it was developed in the early 1970s, the Wasp method was considered as a state of the art.
It does, however, ignore or minimize one important parameter, namely the shear stress between
the different superimposed layers. This is a topic that the two-layer method attempts to tackle. It
does, therefore, tend to predict pressure losses higher than those from stratified flows in certain
160
Chapter6
circumstances of bimodal (fine and coarse) distribution. Nevertheless, the Wasp method remains
a very useful method in this day for the design of pipelines, particularly when it is supported by
laboratory tests.
6.2.4 Stratified flows
Another school of researchers, particularly led by Shook, Wilson, and Gilles in Canada, focused
on refining the original models of Newitt et al (1955). They proposed that the compound mixture
of coarse and fine particles can be simplified to what they called stratified flows, in which the
fines slide over a moving bed of coarse solids.
Newitt had proposed that the deposition velocity was a mere factor of 17 times the terminal
velocity. Wilson (1992) indicated that this approach was not confirmed by tests for large pipes.
The concept he proposed was that the flow goes through a gradual transition from a fully
stratified to a fully suspended flow, with a gradual change in the pressure gradient.
The method is based on the fractional increase in pressure gradient due to the presence of the
solid
[6.16]
decreases with increasing average velocity of the slurry in the pipe. Suspended solids are
transported more efficiently than contact load and this causes the decrease in excess frictional
energy dissipation as velocity increases. An empirical correlation proposed by Wilson et al.
(1997) is,
C
[6.17]
is the velocity at the midpoint of the transition from fully stratified to homogenous flow.
Like the Durand-Condolios-Worster correlation this method postulates that
is directly
proportional to the volume fraction of solid in the slurry. The parameter M accounts for
distribution of particle sizes in the slurry to some extent and is calculated from the 50% passing
size d50 and the 85 % passing size d85 of the particle population. This relationship is quite
complex and is given by,
161
0.25
13
, 1.7
[6.18]
[6.19]
and
are related to the terminal settling velocities of the particles of size
and
respectively.
0.9
2.7
[6.20]
is calculated from equation 6.20 by substituting

by substituting
for particles of size d50 and
for particles of size d85.
is a function of the terminal settling velocity of the d50 particle and the ratio d50/D.
Determining
cosh 60
[6.21]
is no easy task. In this partially stratified model, it is argued that the lifting of
particles by turbulence is strongly influenced by their diameter. Thus, the fines

tend to be transported better by the fluid compared to the coarse solids at the bottom of the pipe.
These particles are transported by eddies, and the largest eddy would be equal to the pipe
diameter. The pipe diameter was therefore proposed as a reference. The resistance to motion of
the solid particles is proposed to be the result of a contact load associated with mechanical
sliding friction in the coarser bed and fluid friction associated with the friction
velocity of the carrier fluid.
There is no question that the approach of Wilson et al. (1992) is extremely interesting, but it is
more complex than the methods proposed by Equations 6.5 or 6.8 It requires the support of
testing, a database, a software, and a personal computer.
162
Chapter6
6.2.5 Two layer model

Khan and Richardson (1996) explained that Shook and Roco (1991) developed a two-layer
model for stratified slurry flows, which may be summarized as follows:
The slurry flow of heterogeneous mixtures is considered to consist of two layers, each
with its own velocity of motion and volumetric concentration; but it is assumed that
there is no slip between liquid and solid phases.
The solids in the upper layer are fully suspended at a volumetric concentration C1, and
move at a velocityV1.
The coarser solids in the lower layer are considered to be packed. However, because of
their irregular shape, there is a certain void fraction between the particles. For sand, the
lower layer is considered to be at a maximum volumetric concentration of 60% and
move at a velocityV2.
The model consists essentially of mass balances and force balances for the two superimposed
slurry layers. The lower layer consists of contact load contributing to sliding friction at the pipe
wall and upper layer comprises of suspended load for which the immersed weight is transferred
to the carrier fluid... The mass balances relate the mean velocity V and delivered solids volume
fraction Cv to the mean velocities V1 and V2 and concentrations C1 and Clim in upper layer
and lower layer, respectively
[6.22]
[6.23]
The concentration C1 is considered to represent particles which are suspended by fluid lift
forces. Concentration Clim is comprised of particles suspended by these lift forces and particles
which produce Columbic friction. The force balances relate the axial pressure gradient to the
stresses at the perimeters of the layers. The upper layer is considered to contribute only kinetic
friction, so that for horizontal flow:
[6.24]
163
For the lower layer:
[6.25]
Assuming the volume fraction of suspended particles is the same in both phases, and that the
suspended particles contribute a buoyant effect, the force F2z is given by Wilson (1976) as:
0.5
/ 1
[6.26]
The boundary stresses for upper and lower layers were calculated from density of liquid
and suspended particles in each layer using a Fanning friction factor f1 :
0.5
[6.27]
0.5
[6.28]
The interfacial stress 12 is calculated from friction factor f12 and difference in velocity between
the layers:
0.5
[6.29]
The empirical features of the model are the friction factor f12, the friction coefficient
, ,the
concentration limit Clim and the fraction of the in-situ concentration Cr which contributes
Columbic friction. f12 is calculated using following equation suggested by Wilson (1988):
3.36
[6.30]
Where,
5
For
1.86
[6.31]
is greater than 0.002 and Y = 0 otherwise and d12 is the particle diameter at
the hypothetical interface.
164
Chapter6
On the basis of experimental measurements, Gillies et al. (1991) determined the value of
as
0.5 for the sand and gravel slurries. Using Wilson symbol Cc (contact load) for particles, which
are not suspended by fluid lift forces, the quantity of interest is,
[6.32]
On the basis of experimental data base, Cc =Cr was found to be a function of the ratio of the
slurry bulk velocity to particle terminal falling velocity at infinite dilution:
Where
exp
[6.33]
is the unhindered settling velocity of particle of diameter d50 and d50 is the mass
median particle diameter of (+74 m) solids.

For the calculation of Clim, following correlation is used:
0.074
[6.34]
On the basis of experiments, Gillies et al. (1991) suggested the value of Cmax for narrow and
broad distributions as 0.6 and 0.75, respectively.The two-layer models have gained wide
acceptance in the oilsand industry.
6.2.6 Modified Wasp model

Kaushal and Tomita (2002) have studied the pressure drop and concentration profile for multisized zinc tailings slurry flowing through 105 mm diameter pipe at five efflux concentrations
ranging 4% to 26% using eight flow velocities in the range 1.2 to 4.0 m/s for each efflux
concentration. Wasp et al. (1977) model for pressure drop prediction has been modified by them
by alleviating some of the restrictive assumptions. Comparison of modified model with
experimental data shows good agreement. Kaushal et al. also compare their pressure drop data
with the predictions by two layer model of Gillies et al. (1991) and found a satisfactory
agreement.
165
6.2.7 Summary of literature survey

It was shown that two principal schools of heterogeneous slurry flows have developed over the
last 50 yearsone around the DurandCondolios approach and the other around the Newitt
approach. The former evolved gradually until Wasp modified it for multilayer compound
systems. The latter gradually evolved to yield the two-layer model. The science of slurry flows
is continuously evolving and needs to be understood in the context of its evolution. It would be
a mistake to favor one school of thought over another. Many successful pipelines have been
designed using either model. The slurry engineer should appreciate that these models are
effective tools to be used with correct empirical coefficients obtained by experimental testing of
samples in pumping loops.
The mixture of coarse and fine particles can lead to a concentration gradient of solids in
horizontal pipe. The coarse particles tend to flow in the lower layers, whereas the fines flow in
the upper layers. Some of the authors recommend determining the pressure losses of fine solids
separately from coarse solids at the corresponding volumetric concentration and particle
diameter of each size. Methods based on concentration ratio for each layer have been
developed. A process of iteration is needed to achieve a final estimation of the pressure loss due
to the bed.
The practicing engineer needs to understand the limitations of each method and those models
have often been developed for a certain range of particle sizes based on experimental data. It is
wise to check the original data. Because the new methods of stratified flows or two-layer
models use the actual hydraulic diameter of the bed, whereas the Wasp and Durand methods use
the actual pipe diameter and therefore it is easy to get confused. In fact, some of the proponents
of the two-layer models leave the impression that Wasp and Durand used the wrong pipe
diameter. This is not the correct way of approach. It is wiser to recognize that the Wasp and
Durand methods are useful tools for the range of slurries for which they were developed. This
includes concentrations of coarse particles up to 20% (volume fraction). This covers, in fact,
most dredged gravels and sands, coal in a certain range of sizes, as well as crushed rocks. It is
also important to appreciate that the work of Zandi and Govatos (1967) was based on sands up
to 22% (volume concentration), and it would be erroneous to push the envelope of application
of their equations beyond such a range.
In the last thirty years, considerable progress has been made with stratified flows. The new twolayer models push the envelope of understanding beyond the limits of the models of Durand,
166
Chapter6
Zandi, and Wasp into the range of volumetric concentrations of 30%. But these models have
their limitations too. Certainly it would be unwise to use the two-layer model when the d50 is
smaller than 74 micrometers. There is still a considerable amount of experimental data
associated with these stratified models needed to obtain the Columbic friction factor and to
determine the difference in velocity between the upper and lower layers. When the particles are
not too coarse, such a difference is not too large, and approximations such as those proposed by
Richardson and Khan are justified, but when the d50 is around 400 micrometers or higher, the
differences in velocity and shear between the upper and lower layers become important.

Despite significant research efforts, prediction of pressure drop in slurry pipeline is still an open
problem for design engineers. Owing to the complex nature of many multiphase slurry
processes, the underlying physico-chemical phenomenon is seldom fully understood. Also,
collection of the requisite phenomenological information is costly, time-consuming and tedious,
and therefore development of phenomenological models poses considerable practical
difficulties. Moreover, nonlinear behavior being common in multiphase slurry processes, it
leads to complex nonlinear models, which in most cases are not amenable to analytical
solutions; thus, computationally intensive numerical methods must be utilized for obtaining
solutions. Difficulties associated with the construction and solution of phenomenological
models necessitates the exploration of alternative modeling formalisms.
Artificial neural networks (ANNs) and more recently support vector regression (SVR) have
emerged as two attractive tools for nonlinear modeling especially in situations where the
development of phenomenological or conventional regression models becomes impractical or
cumbersome. In the present work, we propose four hybrid approach (namely Artificial neural
networks differential evolution (ANN-DE), support vector regression- genetic algorithm
(SVR-GA), Artificial neural networks-genetic algorithm (ANN-GA) and support vector
regression-differential evolution (SVR-DE)) for tuning the meta parameters of ANN and SVR
models and illustrate its application for predicting the pressure drop of solid liquid flow.
The development of ANN-based or SVR-based correlation was started with the collection of a
large databank. The next step was to perform a rigorous regression using ANN or SVR, and to
validate it statistically.
167

As mentioned earlier, over the years researchers have amply quantified the pressure drop of
slurry flow in pipeline. In this work, about 800 experimental points have been collected from
20 sources of open literature spanning the years 1942 -2002. The data were screened for
incompleteness, redundancies and evident inaccuracies. This wide range of database includes
experimental information from different physical systems to develop a unified correlation for
pressure drop. Table 6.1 indicates the range of the collected databank for pressure drop.
[Wilson (1942), Durand and Condolios (1952), Newitt et al. (1955), Thomas
(1962), Zandi and Govatos (1967), Shook (1969), Babcock (1970), Turian and
Yuan (1977), Wasp et al. (1977), Oroskar and Turian (1980), Gillies et al.
(1999), Kaushal (2002)]
Gypsum-water, Glass-water, Gravel-water, iron-water, iron-kerosene, high
density material-water, iron tailings-water, limestone-water, limonite-water,
plastic-water, potash-brine, sand-ethylene glycol, nickel shot-water, iron powderwater, ore-water
Pipe diameter (cm)
1.9 49.5
Particle diameter(micron)
38.3 13000
Liquid density(gm/cc)
1 1.25
Solid density(gm/cc)
1.37 2.844
Liquid viscosity(cp)
0.12 4 cp
Velocity(m/s)
0.86 4.81
Solid volume fraction
0.014 0.333
168
Chapter6

After extensive literature survey all physical parameters that influence pressure drop are put
together in a so-called wish-list.
Out of the number of inputs in wish list, we used artificial neural network to establish the
best set of chosen inputs, which describes pressure drop. The following criteria guide the
choice of the set of inputs:
The inputs should be as few as possible.
Each input should be highly cross-correlated to the output parameter.
These inputs should be weakly cross-correlated to each other.
The selected input set should give the best output prediction, which is checked by
using statistical analysis [e.g. average absolute relative error (AARE), standard
deviation].
While choosing the most expressive inputs, there is a compromise between the number of
inputs and prediction. Based on different combinations of inputs, trial and error method was
used to finalize the input set which gives reasonably low prediction error (AARE) when
solids concentration, solid and liquid density , viscosity of flowing medium and mixture
velocity have been selected to predict pressure drop in slurry pipeline. Table 6.2 shows some
typical data used for ANN & SVR correlations.

As the magnitude of inputs and outputs greatly differ from each other, they are normalized in 1 to +1 scale. 75% of total dataset was chosen randomly for training and rest 25% was selected
for validation and testing.
Seven parameters were identified as input (Table 6.2) for ANN and the pressure drop is put
as target. These data then exposed to hybrid ANN-DE model described in chapter 2. After
optimisation of five ANN parameters described earlier, the model output was summarised in
Table 6.3. Out of all the possibilities, Marquard Levenburg algorithm with ten numbers of
169

Table 6.2: Typical input and output data for ANN training
Input
Output
pipe dia
(cm)
particle dia
(micron)
liquid
density
(g/cc)
solid
density
(g/cc)
liquid
viscosity
(cp)
slurry
velocity
(m/s)
solid conc.
(vol. frac.)
dpdz
(Pa/m)
2.54
2.54
2.54
1.90
1.90
1.90
5.26
5.26
5.26
5.26
5.26
5.26
20.85
20.85
20.85
20.85
20.85
20.85
5.15
5.15
5.15
5.15
5.15
5.15
26.30
26.30
26.30
26.30
26.30
26.30
26.30
26.30
49.50
49.50
49.50
49.50
49.50
49.50
15.85
15.85
15.85
15.85
5.07
5.07
4.00
4.00
4.00
4.00
26.30
26.30
40.1
40.1
40.1
40.1
40.1
40.1
38.3
38.3
38.3
38.3
38.3
38.3
190.0
190.0
190.0
190.0
190.0
190.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
165.0
190.0
190.0
190.0
190.0
520.0
520.0
580.0
580.0
580.0
580.0
13000.0
13000.0
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.25
1.25
1.25
1.25
1.00
1.00
2.84
2.84
2.84
2.84
2.84
2.84
2.33
2.33
2.33
2.33
2.33
2.33
1.37
1.37
1.37
1.37
1.37
1.37
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.65
2.27
2.27
2.27
2.27
2.65
2.65
0.85
0.85
0.85
0.85
0.85
0.85
1.00
1.00
1.00
1.00
1.00
1.00
1.14
1.14
1.14
1.14
1.14
1.14
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.30
1.30
0.12
0.12
1.00
1.00
4.00
4.00
4.00
4.00
1.00
1.00
1.06
1.13
1.15
1.10
1.50
1.70
1.11
3.01
4.81
1.33
3.12
4.70
2.59
2.34
2.01
1.78
1.59
1.37
1.66
3.78
1.66
4.17
1.66
4.33
2.90
3.50
2.90
3.50
2.90
3.50
2.90
3.50
3.16
3.76
3.07
3.76
3.16
3.76
2.50
2.50
3.00
2.90
1.90
2.00
2.88
2.70
2.01
1.05
3.20
4.00
0.07
0.07
0.07
0.07
0.07
0.07
0.11
0.11
0.11
0.31
0.31
0.31
0.33
0.33
0.33
0.33
0.32
0.33
0.07
0.09
0.17
0.19
0.27
0.29
0.09
0.09
0.18
0.17
0.26
0.26
0.33
0.32
0.09
0.09
0.17
0.17
0.26
0.26
0.14
0.29
0.13
0.28
0.09
0.21
0.16
0.14
0.10
0.06
0.04
0.04
560.00
600.00
600.00
1250.00
1500.00
1700.00
294.10
1651.30
3822.90
542.90
2352.60
4727.70
266.50
226.30
177.30
147.00
123.40
99.90
666.20
2449.20
901.30
3428.90
1136.40
4408.10
261.60
334.10
305.70
382.10
355.60
453.60
414.40
526.10
143.00
186.10
157.70
210.60
193.00
254.70
475.20
630.90
648.90
866.70
1175.60
1763.40
3926.00
3580.00
2217.00
845.00
842.50
989.50
170
Chapter6
nodes in hidden layer and log sigmoidal and linear function in input and output layer
respectively has emerged out as the best solution (lowest AARE) for the present case. The low
AARE (12.7%) may be considered as an excellent prediction performance considering the
poor understanding of slurry flow phenomena and large databank comprising of wide varieties
of systems. The parity plot of experimental and predicted pressure drop is shown in figure 6.3
as a goodness of hybrid ANN-DE performance. The weight, bias and final equations to
calculate the pressure drop for any solid liquid slurry flow in pipelines are summarized in
Table 6.4.

Same dataset was exposed to hybrid ANN-GA model. It is found that ANN-GA solution is
also converged almost same solution similar to table 6.3. It only take some more time to
converge as compared to ANN-DE algorithm.
Table 6.3: Prediction error by hybrid ANN-DE based model
Prediction performance by ANN-DE model

training
testing
AARE
0.127
0.129
Sigma
0.169
0.171
0.977
0.975
10
Log sigmoidal
Linear
0.049
Marquard
Levenburg
171
Preessure drop of slurry flow in pipeline
Fiig 6.3: Experim

mental Vs preedicted pressu
ure drop by Marquard
M
Leveenburg algoriithm by
A
ANN-DE
modeel
Table 6.4: Set

S of equation
ns and fitting parameters fo
or neural netw
work correlatiions
(i=7, j=10, k=1)
W i,j
10
-0.9399
0.806
-0.9229
-1.332
0.1665
0.295
0.0049
-2.026
-8.0025
1.174
-0.1766
0.231
-1.1001
0.363
-3.2882
1.169
-0.5517
-1.226
-0.2273
-0.948
0.5833
1.884
0.9990
-0.272
-1.1449
-0.907
-0.3371
0.165
-2.9975
1.341
1.0900
-2.825
0.6550
-2.857
-1.1554
-0.312
-1.5511
3.629
-0.6619
-2.468
-1.2177
-0.820
-0.8448
1.028
-0.9994
-0.696
-0.3325
1.460
1.112
-0.236
-1.0000
-0.186
0.4228
2.628
-1.7447
-1.387
1.6686
1.024
1.4454
0.692
-0.6944
1.221
-0.4661
0.309
-0.9885
-0.536
0.0036
0.493
0.173
0.458
2.6388
3.482
2.1332
-0.957
-1.3777
-1.135
-1.0056
1.307
-4.5573
3.396
10
11
-1.0811
1.495
1.1663
-0.839
-0.0005
-0.882
-1.149
-0.889
4.101
-1.052
0.642
W j,k
Wherre Ui=input varriable, Hj=hiddden layer outpuut and U8=1, H11=1 (bias)
172
Chapter6
6.4.3 Comparison of hybrid ANN-GA model with ANN model

As ANN-GA algorithm takes slightly longer time than ANN-DE algorithm, ANN-GA
algorithm is tested against ANN algorithm only. Same dataset is exposed to ANN algorithm
only (without the GA algorithm) and tried to optimize the different parameters based on
exhaustive search. We found that it was not possible to reach the best solutions starting from
arbitrary initial conditions. Many times the solutions got stuck up in sub optimal local minima.
These experiments justified the use of a hybrid technique for ANN parameter tuning. The best
prediction after exhaustive search along with ANN parameters was summarized in Table 6.5.
On the other hand, the hybrid ANN-GA technique is able to locate the global minima with
2000 runs within 1.2 hr. The prediction accuracy is also much better. Moreover it relieves the
non expert users to choose the different parameters and find an optimum ANN meta
parameters with a good accuracy.
Table 6.5: Comparison of performance of ANN-GA hybrid model Vs ANN model
Prediction by Hybrid
ANN-GA model
Prediction by ANN
model only
testing
testing
AARE
0.127
0.132
Sigma
0.169
0.170
0.977
0.975
Execution time (hr)
1.2
4.9
6.4.4 Prediction performance of hybrid SVR-DE model

After building successful ANN model, same data was tried to build hybrid SVR-DE model.
Same seven parameters were identified as input (Table 6.1) for SVR and the pressure drop is
put as target. These data then exposed to hybrid SVR-DE model described in chapter 2. After
optimization of five SVR parameters described earlier, the model output was summarized in
Table 6.6. The low AARE (12.7%) may be considered as an excellent prediction performance
173

considering the poor understanding of slurry flow phenomena and large databank comprising
of various systems. The optimum value of SVR meta parameters were summarized in Table
6.7. From Table 6.7 it is clear that there exist almost 4 different feasible solutions which lead
to same prediction error.
Table 6.6: Prediction error by SVR based model

training
testing
AARE
0.123
0.129
Sigma
0.163
0.169
0.978
0.977
Prediction performance was also checked for SVR-GA model in a separate attempt and it
converges similar to value given in table 6.6. In a separate study, we exposed the same dataset
to SVR algorithm only (without the DE algorithm) and tried to optimize the different
parameters based on exhaustive search. We found that it was not possible to reach the best
solutions starting from arbitrary initial conditions. Specially, the optimum choice of C and is
very difficult to arrive at after starting with some discrete value. Many times the solutions got
stuck up in sub optimal local minima. These experiments justified the use of a hybrid
technique for SVR parameter tuning. The best prediction after exhaustive search along with
SVR parameters was summarized in Table 6.8. From the table it is clear that even after
720,000 runs, the SVR algorithm is unable to locate the global minima and the time of
execution is 4 hrs in Pentium 4 processor. On the other hand, the hybrid SVR-DE technique is
able to locate the global minima with 2000 runs within 1 hr. The prediction accuracy is also
174
Chapter6
much better. Moreover it relieves the non expert users to choose the different parameters and
find a optimum SVR meta parameters with a good accuracy.
Table 6.7: Optimum parameters obtained by hybrid SVR- DE algorithm
Sr
No
1
2
3
4
5043.82
6698.16
7954.35
8380.98
0.50
0.36
0.48
0.54
Kernel
type
erbf
erbf
erbf
erbf
Type of loss
function
e- insensitive
e- insensitive
e- insensitive
e- insensitive
Kernel
parameter
2
3
4
5
AARE
0.129
0.129
0.129
0.129
Table 6.8: Comparison of performance of SVR-DE hybrid model Vs SVR model

Prediction
performance by
SVR model without DE
hybrid SVR-DE model
tuning after exhaustive search
testing
testing
AARE
0.129
0.132
Sigma
0.169
0.170
0.977
0.975
Execution time (hr)
Performance criteria
6.4.5 Comparison with other published correlations

All the 800 experimental data set collected from open literature was also exposed to different
correlations for pressure drop available in open literature and AARE were calculated for each
of them (Table 6.9). From table 6.9, it is evident that prediction error of pressure drop has
reduced considerably in the present work.
175

Table 6.9: Performance of different correlations to predict pressure drop
Sl no
1
2
3
4
5
6
7
8
9
10
Author
Wilson (1942)
Durand and Condolios (1952)
Newitt et al. (1955)
Zandi and Govatos (1967)
Shook (1969)
Turian and Yuan (1977)
Wasp et al. (1977)
Gillies et al. (1999)
Kaushal (2002)
Present work
% AARE
49.51
36.53
93.43
50.02
34.50
39.97
26.68
22.31
22.01
12.70
6.5 Conclusion
In this study it is demonstrated that ANN and SVR based correlations can be an easy
alternative of empirical equations for pressure drop prediction in slurry flow in pipeline.
Superior prediction performances were obtained for the case study of pressure drop and a
comparison with selected correlations in the literature showed that the developed ANN and
SVR correlation noticeably improved prediction of pressure drop over a wide range of
operating conditions, physical properties, and pipe diameters. The method incorporates hybrid
differential evolution and genetic algorithm technique for efficient tuning of ANN and SVR
meta parameters. The proposed hybrid techniques also relieve the non-expert users to choose
the meta parameters of ANN and SVR algorithm for the case study and find out the optimum
value of these meta parameters on its own. The results indicate that ANN or SVR based
technique with the GA or DE based parameters tuning approach described in this work can
yield excellent generalization and can be advantageously employed for a large class of
regression problems encountered in process engineering.
176
Chapter6
Nomenclature
A:
Cross-sectional area of pipe(m2)
A1:
Area of upper layer in two layer model(m2)
A2:
Area of lower layer in two layer model (m2).
C:
Volume fraction of solids in the suspension(-)
C1:
Solid volume fraction in upper layer of two layer model(-)
Cc :
Contact load for particles, which are not suspended by fluid lift forces(-)
CD :
Drag co-efficient
CD :
Drag co-efficient at terminal settling velocity
Clim:
Solid volume fraction in lower layer of two layer model(-)
Cr :
Fraction of the in situ concentration which contributes Coulombic
Ct :
In-situ concentration of particles (-)
Cv :
Volumetric concentration of solids (-)
D:
Pipe diameter(m)
d50:
Sieve passage diameter for 50% of the particles(m)
d85:
Sieve passage diameter for 85% of the particles(m)
Di :
Pipe inner diameter (m)
dp :
Particle diameter, m
dpi :
Particle diameter of ith fraction used in equation 6.14 (m)
dpm :
Weighted mean particle diameter used in equation 6.14 (m)
fD:
DarcyWeisbach friction factor (-)
F2z :
Coulombic resisting force exerted by the wall on those particles which do not
contribute to the kinetic friction in two layer model( Newton).
fDL :
Darcy friction factor for equivalent volume of water(-)
fDm :
Darcy friction factor for mixture(-)
Fr:
Froude number(-)
fsl :
DarcyWeisbach Friction factor for slurry(-)
fw :
DarcyWeisbach Friction factor for carrier fluid(-)
g:
Acceleration due to gravity (m/s2)
im :
pressure gradient for the slurry mixture in meters of water (m)
K :
K2 :
Experimentally determined constant used in equation 6.13(-).

177

Kf :
Co-efficient in the Durand equation for pressure drop ( equation 6.3)(-)
m:
Power coefficient in Zandis models(-)
mi :
The mass fraction of solids with particle diameter of dp(-)
mt:
Total mass of particles(Kg)
Ne :
Index number in equation 6.4 (-)
S1 :
Partial perimeter of upper layer in two layer model (m)
S2 :
Partial perimeter of lower layer in two layer model (m)
S12 :
Width of interface in two layer model (m)
V:
Velocity of mixture (m/s)
Vm :
Mean velocity of mixture
VT :
Terminal settling velocity , m/s
:
Z:
Average velocity of slurry in pipe (m/s)

Nondimensional parameter to express difference between friction losses due
to the slurry and an equivalent volume of water.(-)
Greek symbols
:
Pressure gradient for an equivalent volume of water or carrier fluid in meters

of water (m)
:
unhindered settling velocity of particle of diameter d50 (m/s)
Terminal settling velocities of the particles of size
(m/s)
Terminal settling velocities of the particles of size
. (m/s)
s :
Viscosity of slurry(Pa. s)
0 :
Angle defining lower layer (radians),
Pfw :
Pressure gradient of equivalent water (Pa/m)
Psl :
Pressure gradient of slurry (Pa/m)
s :
Coefficient of friction between particle and pipe(-)
1:
Density of liquid in two layer model (kg/m3)

178
Chapter6
2f :
Density of (fluid +suspended solids) in lower layer(kg/m3).
f :
Density of liquid combined with (-74 m) fines in two layer model(kg/m3)
s :
Density of solids(kg/m3)
1 :
Wall stress of upper layer(Pa)
12 :
Stress at hypothetical interface (Pa)
2k :
Kinetic stress at the partial perimeter of lower layer(Pa)
product of index number and volumetric concentration (refer equation 6.4)(-)
Pressure loss factor
t :
Pressure loss factor based on in-situ concentration of particles (-)
References
1. Babcock H.A., Heterogeneous flow of heterogeneous solids Advances in solid liquid flow
2. Babcock H.A., Heterogeneous flow of heterogeneous solids, Paper presented at
International Paper presented at International Symposium on Solid Liquid Flow in Pipes,
March 37, University of Pennsylvania, Philadelphia. , (1968)
3. Bagnold R.A, Gravity-free dispersion of large spheres in a Newtonian fluid under shear,
Proc. Royal Soc. A, 225, 4963 (1954)
4. Bagnold R.A, Some flume experiments on large grains but little denser than the
transporting fluid, and their implications., Proc. Inst. Civ. Eng., 4, 3, 174205 (1955)
5. Bagnold R.A, The flow of cohesion less grains in fluids., Phil. Trans. Roy. Soc., 249, 235
297. (1957)
6. Doron P.,Granica D.,Barnea D., Slurry flow in horizontal pipesexperimental and
modeling, Int. J. of Multiphase Flow, 13, 535547 (1987)
7. Durand R., Condolios E, compterendu des Deuxiemes journees de L Hydraulique, Paris,
8. Gaessler
H,
Experimentelle
und
Theoretische
Untersuchungen
uber
die
Stromungsvorgange Beim Transport von Festoffen in Flassigkeiten durch Horizontale

Rohrleitungen. Doctoral thesis. , Karlsruhe, Germany, (1967)
179

9. Geller, L. B.,. Gray. W.M., Selected Theoretical Studies Made in Conjunction with the
Joint Canada/FRG Research Project on Coarse Slurry, Short Distance Pipeline. CANMET
SP 86-16 EGovernment of Canada Publications, (1986)
10. Ghanta K.C, Studies on rhelogical and transport characteristic of solid liquid suspension in
pipeline, PhD Thesis, IIT Kharagpur., (1996)
11. Gillies R.G.,Hill K.B.,Mckibben M.J.,Shook C.A., Solids transport by laminar Newtonian
12. Gillies R.G.,Shook C.A., Modeling high concentration settling slurry flows, The Canadian
Journal of Chemical Engineering, 78, 709716 (2000)
13. Gillies R.G.,Shook C.A.,Wilson K.C., An improved two layer model for horizontal slurry
pipeline flow, The Canadian Journal of Chemical Engineering, 69, 173178 (1991)
14. Govier,
G. W., Aziz. K. , The Flow of Complex Mixtures in Pipes, New York: Van
Nostrand Reinhold., (1972)

15. Govier G.W.,Aziz K., The Flow of Complex Mixtures in Pipes, Krieger Publication,
Malabar, FL (1982)
16. Hayden J.W, Stelson. T.E, Hydraulic conveyance of solids in pipes, Paper presented at the
International Symposium on SolidLiquid Flow in Pipes, March 37, University of
Pennsylvania,Philadelphia., (1968)
17. Herbich H, Handbook of Dredging Engineering. New York: McGraw-Hill., (1991)
18. Hill R.A, Snock P.E, Gandhi. R.L., Hydraulic transport of solids., The Pump Handbook,
2nd ed. Edited by I. J. Karassik et al. New York: McGraw-Hill., (1986)
19. Kaushal D.R.,Tomita Yuji, Solids concentration profiles and critical velocity in pipeline
flow of multisized particulate slurries, International Journal of Multiphase Flow, 28, 16971717 (2002)
20. Khan, A.R., Richardson. J.F., Comparison of coarse slurry pipeline models., Proceedings
of hydrotransport,Cranfield, UK: BHR Group., 13, 259281 (1996)
21. Newitt D.M.,Richardson J.F.,Abbott M ,Turtle R.B., Hydraulic conveying of solids in
22. Oroskar A.R.,Turian R.M., The critical velocity in pipeline flow of slurries, AIChE J., 26,
550-558 (1980)
23. Roco M.C.,Shook C.A.,
Computational methods for coal slurry pipeline with
heterogeneous size distribution, Powder Technology, 39, 159176 (1984)

180
Chapter6
24. Roco M.C.,Shook C.A., Turbulent flow of incompressible mixtures, J. Fluids Engg., 107,
224-231 (1985)
26. Shook C.A, Roco. M.C., The two layer model., Slurry Flow: Principles and Practice.,
Newton, MA: Butterworth-Heinemann., (1991)
(1962)
28. Turian R.M.,& Yuan T.F, flow of slurries in pipelines, AIChE J., 23, 232-243 (1977)
30. Wasp E.J.,Aude T.C.,Kenny J.P.,Seiter R.H.,Jacques, R.B.,
Deposition velocities
transition velocities and spatial distribution of solids in slurry pipelines, Proc. Hydro
transport 1, BHRA Fluid Engineering, Coventry, UK, paper H42 53-76 (1970)
31. Wilson
K.C., Addie G.R, Clift R, Slurry transportation using centrifugal pumps, New
York: Elsevier Applied Sciences., (1992)

32. Wilson
K.C., Addie G.R,Sellgren A., Clift R, Slurry transportation using centrifugal
pumps, Blackie Academic and Professional, 2nd edition, (1997)

33. Wilson K.C,& Pugh F.J.,
Dispersive-force modeling of turbulent suspension in
heterogeneous slurry flow, Can. J. Chem. Engg, 66, 721-727 (1988)

34. Wilson K.C., A unified physically-based analysis of solidliquid pipeline flow,
Proceedings of the 4th International Conference on Hydraulic Transport of Solids, BHRA
Fluid Engineering, Cranfield UK Paper A2, 116 (1976)
35. Wilson K.C., Evaluation of interfacial friction for pipeline transport models, In Proc. 11th
Int. Conf. on the hydraulic transport of solids in pipes, Stratford-upon-Avon U.K., paper
B4PP107-116 (1988)
36. Wilson W.E., Mechanics of flow with noncolloidal inert solids, Trans ASCE, 107, 15761594 (1942)
Div., 93(HY3), 145-159 (1967)
181
Semi-empirical method for pressure drop and concentration profile
CHAPTER 7
Semi-empirical method for pressure drop and

concentration profile
Abstract
In the literature, several correlations have been proposed for pressure drop prediction in
slurry pipelines. However, these correlations fail to predict pressure drop over a wide range
of conditions. Among various models available in literature Wasp model was found most
promising, versatile and can be applied for slurry system having wide range of system
parameters with suitable modification. In the present study Wasp model for pressure drop has
been modified by alleviating some of the restrictive assumptions of the model.
A new method was also developed to find out concentration profile using Wasp model as a
starting point. The concentration profile and pressure drop data predicted by modified model
were compared with the experimental one collected from literature. The model showed a quite
well agreement and the average prediction error for pressure drop was found less than that by
original Wasp model.
Keywords: Pressure drop; Concentration profile; Particle diffusivity; Slurry flow.
182
Chapter7
7.1 Introduction
Transportation of slurries through pipeline is common in solid handling, mineral and
petrochemical industries and its huge power consumption are drawing attention in recent
years. Need and benefit of accurately predicting the vertical solid concentration profiles and
pressure drop of slurry pipelines during design phase is enormous as it gives better selection of
slurry pumps, optimization of power consumption and helps to maximize economic benefit.
Despite the large area of application, the available published models describing the slurry
pressure drop and concentration profile do not completely satisfy engineering needs. Efforts
are still on to develop more reasonable model for the prediction of pressure drop and
concentration profile in pipes and in this direction, the work of Wasp (1977), Roco and Shook
(1983, 1984), Gillies et al. (1999, 2000) and Kaushal et al. (2002) is worth mentioning.
Despite the significant research efforts, prediction of pressure drop and solid concentration
profile in slurry pipeline is still an open problem for design engineers. Design of slurry
pipelines relies on empirical correlations obtained from the experimental data. These
correlations are prone to great uncertainty if one departs from the limited database that
supports them. Most of the equations available in literature for predicting pressure drop and
vertical solids concentration profiles in slurry pipeline are empirical in nature and have been
developed based on limited data on materials having equisized or narrow size-range particles.
Moreover, at higher solids concentration, very few experimental data on local solids
concentration is available because of the difficulties in the measurement techniques.
Considering this, it would be most useful to develop computational models, which will allow
a priori estimation of pressure drop and the solid concentration profile over the pipe cross
section.
In heterogeneous flow regime, pressure drop is heavily dependent on how solids are
distributed in vertical plane. Therefore, pressure drop correlations which have not taken into
account the solid concentration profile are not expected to perform well. Wasp model (1977) is
first one which developed pressure drop model by taking into account the concentration
profile. This semi-empirical model is distinct from other pressure drop correlations in a sense
that it tries to incorporate the underlying physics (phenomenology) in its model development.
Many big slurry pipelines around the world were designed with this model for last 30 years.
However, when this model was exposed to a large databank of pressure drop of various
183

systems, it predicts pressure drop above 25% error. This 25% error in design has serious cost
implications in todays business and can make it noncompetitive as most of the slurry
pipelines are designed by optimizing the initial investment and working capital.
In the light of the above shortcomings, an attempt has been made in the present study to
modify Wasp model for pressure drop by alleviating some of the restrictive assumptions used
in the model so that prediction accuracy can be improved. Moreover, this work explores the
possibility to utilize this modified Wasp model to predict concentration profile in slurry
pipelines.
7.2 Background work

In reality, slurries often consist of a wide distribution of particles. The coarser ones tend to
move towards bottom layer in the horizontal pipe, and the finer ones move above these bottom
layers. Understanding of the distribution of these particles in layers after layers is essential for
a correct estimation of the friction losses.
OBrien (1933) and Rouse (1937) were the first ones to propose the simple diffusion
type model for the prediction of concentration profile of solids in turbulent streams. Ismail
(1952) modified the simple diffusion model by correlating the mass transfer coefficient to
shear stress/velocity gradient which was derived from von Karman universal velocity profile
and then compared with his own data generated in a rectangular channel. He found that the
von Karman constant was a function of solids concentration and decreased with increasing
solids concentration.
Ismail(1952) focused initially on rectangular closed conduits. His work demonstrated
that the concentration can be expressed in terms of exponential function as,
[7.1]
Where,
Es = the mass transfer coefficient
a = height of layer A above bottom of the conduit
y = distance from the pipe bottom
C = volumetric concentration of the particle size under consideration
CA = volumetric concentration of solid at pipe axis
Vt=settling velocity
184
Chapter7
Hsu (1986) reexamined the work of Ismail (1952) by proposing a polar coordinate system (r,
) for the analysis of concentration distribution in a pipe:
,
[7.2]
Where,
= the angle from the horizontal
= angle from the vertical starting at the lowest quadrant point
RI = inner diameter of pipe
r = local radius for a point in the flow
For many pipes, C/ CA is considered by Wasp et al. (1977) to be 0.08 D, from the bottom of
the pipe. Wasp et al. (1977) examined the concentration distribution in the coal pipeline of
Consolidation Coal Company, Ohio at a height of 8% from the bottom of the conduit and at
8% from the top of the conduit. They reinterpreted the work of Ismail (1952) and devised the
following equation:
.
[7.3]
Where,
Uf is the friction velocity
k is the von Karman constant
= constant of proportionality
Wasp (1963) assuming a value of k to be 0.35 used in Ismails equation found that the
prediction of concentration profile agreed reasonably well with his experimental results.
Subsequently, Wasp et al. (1970) analyzed their own data and Ismails (1952) data at y =0.50
D and 0.08D, where y is the distance from the pipe bottom and D is the pipe diameter, to come
up with a modified equation for the prediction of concentration profile.
If is assumed to be unity and there is no suppression for the von Karman constant, i.e., K =
0.4, then,
4.5
[7.4]
Thomas (1962) commented that the DurandCondolios approach was limited to sand and
similar solids and proposed a more general criterion for evaluating flow of slurries in terms of
the ratio Vt/Uf or ratio of free-fall velocity to friction velocity. He indicated that when,
185

0.2,
[7.5]
then the solids would be transported as heterogeneous slurry. Parallel to Wasps work (1963),
Shook and Daniel (1965) and Shook et al. (1968) were involved in extensive research in the
area of slurry flow in pipes. They found that the equation proposed by OBrien (1933) and
Rouse (1937) does not account for normal dispersive forces. Karabelas (1977) in the late
seventies developed an empirical model to predict the vertical composite concentration and
particle distributions based on Hunts (1954) formulation for solid-liquid flows through pipe.
Karabelas (1977) has also shown the applicability of his closed form expression by comparing
the predictions with experimental results. Seshadri et al. (1982) used the equation developed
by Wasp et al. (1970) to predict the overall concentration profile and individual size
distribution. They found a quite well agreement of concentration profile and size distribution
at the top of the pipe whereas large deviations were observed at the bottom of the pipe. Efforts
are still on to develop more reasonable model for the prediction of concentration profile in
pipes and in this direction, the work of Roco and Shook (1983, 1984), Gillies et al. (1999,
2000), Mukhtar (1991) and Kaushal et al. (2002) is worth mentioning.
From literatures it is quite clear that the concentration profiles have profound effect
on pressure drop of heterogeneous slurry flow. The pressure drop correlations without any
account of concentration profile are expected to fail when they are subjected to predict
pressure drop out of their limited database on which they were made. As discussed and
reviewed in chapter 6.2, there are lot of correlations available for pressure drop in literatures
since 1942 to today. These large number of pressure drop correlations can be distinctively
divided into two major classes such as empirical correlations and semi-empirical correlations.
In the empirical correlations approach, one begins from the experimental facts and generalizes
known correlations for some parameters by dimensional analysis, without providing an insight
into the flow mechanism (e.g. Newitt et al.1955; Zandi & Govatos, 1967; Turian & Yuan,
1977; Ghanta, 1999 and many others). In the semi-empirical correlations approach, one starts
from the basic equations of motion and numerically solves these for some situations with
physical or mathematical assumptions for different terms. These semi-empirical correlations
try to understand and include the physics (phenomenology) of slurry behavior within them.
Numbers of such phenomenological modeling have been proposed by Wasp (1977), Wilson
(1976, 1988), Wilson & Pugh (1988), Roco & Shook (1984, 1985), Doron et al. (1987) and
many others. Among various models available in literature Wasp model was found most
promising, versatile and can be applied with suitable modification over a wide
186
Chapter7
range of slurry system e.g. slurries resulting from dredging, cyclone underflow, and tailings
disposal. It involves coarse and fine particles. The fines move as a homogeneous mixture
while the remainder moves as a heterogeneous mixture. Based on this concept, the pressure
drop calculation method developed by Wasp et al. (1977) has been used very successfully over
the last 30 years for Newtonian slurries.
7.3 Wasp et al. (1977) model

The method is based on the assumption that the total pressure drop in two phase flow can split
into two parts; pressure drop due to vehicle (homogeneously distributed particles) and excess
pressure drop due to bed formation (heterogeneously distributed particles). The method
suggested is an iterative one. In the first iteration, the particular suspension is assumed to be
completely homogeneous. Based on this assumption the pressure gradient is computed using
DarcyWeisbach formula which for slurry flow given by,
[7.6]
where ivehicle is the pressure gradient due to homogeneously distributed particles (vehicle) and
fm is the Fanning friction factor which is evaluated by Woods equation:
[7.7]
Where,
0.026
0.133
22
1.62
[7.8]
/D is the pipe roughness which is determined from the velocity against pressure drop data for
water. Rem is the Reynolds number for the slurry, which is calculated using Thomas (1965)
correlation for slurry viscosity as given below:
[7.9]
187
2.5

.
10.05
0.00273
[7.10]
Where m is the mass density of slurry, Vm is the flow velocity of the slurry, g is the
acceleration due to gravity, l is the viscosity of carrier liquid and Cvf is the slurry efflux
concentration by volume.
In the second iteration particles in slurry are divided into 4 to 6 sizes. For each size fraction the
percentage of solids in vehicle is calculated from,
1.8
[7.11]
where C/CA is the ratio of volumetric concentration of solids at 0.08D from top to that at pipe
axis, is the dimensionless particle diffusivity and is taken as 1.0, k is defined as von Karman
constant, u* is the friction velocity calculated from the pressure drop in first iteration and wj is
the settling velocity of jth size particle calculated using standard drag relationships given in
Table 7.1.
Table 7.1: Drag relationships
Range of particle Reynolds number
Relation for drag co-efficient (CD)
(Rep)
Stokes law ,
Intermediate,
Newtons law,
Rep 1.0
1<Rep1000
1000<Rep2*105
CD=24 Rep-1
CD=24rep-1(1+0.15Rep0.687)
CD=0.44
After computing the vehicle portion for each particle size, the total percentage of solids in
vehicle is calculated. The vehicle pressure drop is calculated thereafter for slurry of this
concentration using DarcyWeisbach equation. The pressure drop due to remaining solids for
each particle size which are in bed is calculated using Durand type of relationship given as:
188
Chapter7
82
[7.12]
Where, iw is the pressure gradient due to flow of carrier liquid, Cvjbed is the volumetric
concentration of jth size particles in the bed portion and S is the specific gravity of solids.
The total pressure drop due to bed is the sum of pressure drop due to each particle size in the
bed:
[7.13]
The total pressure drop due to slurry is the sum of pressure drop due to vehicle and pressure
drop due to bed. If the difference in pressure drops in two successive iterations is greater than
5%, the iterative scheme is continued. In the subsequent iteration, steps of the last iteration are
repeated with modified frictional velocity u* based on pressure drop in that iteration. Pressure
drop obtained from the latest iteration is again compared with the pressure drops from the
previous iteration and if the difference is still not within acceptable limit (5%), scheme is
repeated.
Wasp method was considered as a state of the art when it was developed in the early
1970s. It does, however, ignore or minimize one important parameter, namely the shear stress
between the different superimposed layers. This is the area that the two-layer method (Gillies
et al., 1991) attempted to tackle. It does, therefore, tend to predict pressure losses higher than
those from stratified flows in certain circumstances of bimodal (fine and coarse) distribution.
Nevertheless, the Wasp method remains a very useful method to this day for the design of
pipelines, particularly when it is supported by laboratory tests.
7.4 Comparison of pressure drop prediction by Wasp model with

experimental data
Computer program for calculation of pressure drop using Wasp model has been developed.
Wasp model was tested with 500 experimental slurry flow data of different slurry system
collected from Hsu (1986) and Roco & Shook (1984). The broad range of the experimental
data set used in this work is given in table 7.2.
189

Table 7.2 System and parameter studied collected from the literature
Slurry system: coal-water, copper ore-water, sand-water, gypsum-water, glass-water, gravelwater
Pipe diameter D, m
0.019 0.495
Particle diameter dp, m
38.3 13000
-3
Liquid density L, kgm
1000 1250
Solids density S, kgm-3
1370 2844
Liquid viscosity , mPas
0.12 4
-1
Velocity U, ms
0.86 4.81
Solids concentration (volume fraction) f
0.014 0.333
A typical 20 data of coal-water slurry collected from Roco & Shook (1984) (Table 7.3) was
selected to test the performance of Wasp model. These data was chosen purposefully as both
experimental pressure drop and concentration profile are available for these set of data. The
prediction error by Wasp model on pressure drop for the typical 20 data of Roco &
Shook(1984) is shown in figure 7.1.
Table 7.3: Coal water slurry data collected from Roco & Shook (1984)
Run
number
Pipe dia, m
A1
A2
A3
A4
A5
A6
A7
A8
A9
A10
A11
A12
A13
A14
A15
A16
A17
A18
A19
A20
0.0515
0.0515
0.0515
0.0515
0.0515
0.0515
0.2630
0.2630
0.2630
0.2630
0.2630
0.2630
0.2630
0.2630
0.4950
0.4950
0.4950
0.4950
0.4950
0.4950
Mixture
velocity,
m/s
1.66
3.78
1.66
4.17
1.66
4.33
2.90
3.50
2.90
3.50
2.90
3.50
2.90
3.50
3.16
3.76
3.07
3.76
3.16
3.76
Efflux
concentration, vol
fraction
0.0837
0.0920
0.1870
0.1890
0.2800
0.2860
0.1030
0.1000
0.1900
0.1840
0.2700
0.2680
0.3410
0.3380
0.1040
0.1000
0.1870
0.1840
0.2730
0.2690
Pressure
drop, Pa/m
666.20
2449.20
901.30
3428.90
1136.40
4408.10
261.60
334.10
305.70
382.10
355.60
453.60
414.40
526.10
143.00
186.10
157.00
210.60
193.00
254.70
190
Ch
hapter7
It is found that the Wasp moddel predicts preessure gradientt with reasonaable accuracy at
a lower
effflux concentraations for all thhe flow velocitiies. However, Wasp
W
model ovverestimates thhe same
fo
or higher efflux
x concentrations at larger floow velocities. From the com
mparison, it is cleared
thhat Wasp model is not takin
ng into considderation the chhanges in fluidd and flow prooperties
w
which
occur witth increase in efflux
e
concentraation and flow
w velocity.
Wasp model an
nd modified W
Wasp model baased on
Fiigure 7.1: Pressure drop prrediction by W
typiical Roco & Shook (1984) data
d
Thhe causes forr non-suitabilitty of Wasp m
model at high
her efflux conncentrations annd flow
veelocities were investigated
i
annd may be due to the followinng:
The dim
mensionless paarticle diffusivvity (), which is not equal to
t unity as connsidered
by Wassp.
The vaalue of von Kaarman constantt, k was assum

med constant (=
=0.4) by Wasp
p which
may noot be true at th
he higher flow
w velocities, wh
hen the vehicle pressure dro
op has a
major share
s
in total prressure drop.
191

The free settling velocity was used by Wasp. However, at higher concentration
settling velocity is greatly affected by neighboring particles and hindered settling
velocity should be used.
The almost same conclusion was drawn by Kaushal & Tomita (2002) while applying Wasp
model on their experimental data.
7.5 Modified Wasp model

7.5.1
Modifications incorporated in Wasp model
In the light of shortcomings of Wasp model following modifications were made in the existing
model.
a) Kaushal & Tomita (2002) has already expressed the correlation of dimensionless particle
diffusivity () for their experimental data by taking into consideration the effect of solids
concentration. Also, Ismail (1952) found that the von Karman constant is a function of solids
concentration and decreased with increasing solids concentration.
Based on extensive analysis of solids concentration profile and pressure drop data of large
databank consisting of coal-water, copper ore-water, sand-water, gypsum-water, glass-water,
gravel-water system published in literature (Hsu, 1986) a new constant
modified
(=*k) of
Wasp model was developed by taking into consideration the effect of solids concentration on
and k and expressed as follows,
modified =430.78Cvf3 - 110.98Cvf2 + 9.5439Cvf - 0.1728
[7.14]
b) At higher flow velocities, the vehicle pressure drop has a major share in total pressure drop.
Mukhtar (1991) has proposed a modified Woods equation for vehicle friction factor by taking
into consideration the effect of concentration for multisized particulate in slurry transportation
through pipe. The SwainJain equation was used instead of Woods equation in the range of
5000 < Re < 108 to determine the friction coefficient of the homogeneous portion of the
mixture and then Mukhtar (1991) modification was incorporated.
c) Hindered settling velocity was used as per Richardson and Zakis (1954) correlation instead
of free settling velocity used by Wasp.
192
Chapter7
7.5.2 Steps to implement modified Wasp model

Computer program was developed based on following steps to implement the modified Wasp
model:
1.
Total size fraction is divided into number (5) of homogeneous fraction.
2.
Friction loss of the homogeneous fractions was calculated based on the rheology of the
slurry, assuming Newtonian flow and using Darcy- Weisbach formula.

As discussed earlier, SwainJain equation was used to determine the friction coefficient of the
homogeneous portion of the mixture. The equation is as given:
.
/
[7.15]
Where fD =Darcy friction factor

Mukhtar (1991) modification was incorporated as follows:
1
0.33
[7.16]
Where, Cwf is the efflux concentration by weight.
3.
[7.17]
Particle Reynolds number for each size fraction was calculated as,
[7.18]
4.
Drag co-efficient was calculated using Turton(1986) equation,
5.
0.173
.
.
[7.19]
Terminal (Vto) velocity was calculated using following equation,
[7.20]
193

6.
Richardson and Zakis co-efficient (Z) was calculated as,
4.65
0.002
1.95
4.35
17.5
4.45
18.0
0.2
0.2
1.0
1.0
[7.21]
Hindered settling velocity was determined by,

V
7.
Durands equation was used to calculate the friction losses of the heterogeneous fraction:
8.
[7.22]
82
[7.23]
The total pressure drop due to bed is the sum of pressure drop due to each particle size in
the bed:
[7.24]
The total pressure drop due to slurry is the sum of pressure drop due to vehicle and pressure
drop due to bed.
[7.25]
If the difference in pressure drop in the two successive iterations is greater than 5%, the
iterative scheme is continued further.
9.
In the subsequent iteration, steps of the previous iteration are repeated with modified
frictional velocity u* based on pressure drop in previous iteration.
[7.26]
194
Chapter7
[7.27]
10. A ratio C/CA was defined for the size fraction of solids based on friction losses estimated
in steps 2.
10
[7.28]
Where modified was calculated by the equation as discussed earlier.
[7.29]
Pressure drop obtained from the last iteration is again compared with the pressure drop from
the previous iteration and if the difference is still not within acceptable limit (5%), scheme is
repeated.
11. Once the difference in pressure drop in two successive iterations is less than 5 %, final
pressure drop and C/CA was freezed. Now, this C/CA is used further to verify the predicted
concentration profile with the experimental one.
Based on large experimental data on concentration profile collected from open literature, it is
found that concentration profile can be approximated by
[7.30]
=Y/D, vertical distance from the bottom of the pipe, a and b are empirical
Where
constants.
Recalling that C/CA is the ratio of volumetric concentration of solids at 0.08D from top of the
pipe to that at pipe axis, it can be written as,
.
.
[7.31]
Once C/CA is known from step 10 constant b can be solved from the above equation for each
size fraction of solids.
195

12. Another constant a in equation 7.30 can be solved from the following boundary
condition:
.
[7.32]
Where, Ci is the individual solids volume fraction.

13. a and b are thus calculated for each size fraction and then concentrations are obtained
from step 11. Finally overall concentration profiles were obtained by adding the individual
concentrations.
7.6 Results and discussions

The main aim of this section is to expose the above calculation scheme to the experimental
data collected from open literature and estimate the prediction error for pressure drop and
concentration profile. Pressure drop for all the 20 run of Roco & Shook data was calculated
using Wasp model and present model (modified Wasp model).
The statistical analysis of prediction is performed on the following performance criteria:
1.
The average absolute relative error (AARE) on test data should be minimum
2.
[7.33]
The standard deviation of error() on test data should be minimum
3.
[7.34]
The cross-correlation co-efficient (R) between input and output should be around unity.
[7.35]
196
Chapter7
Based on the particle size distribution of original sand in Roco & Shook (1984) experiment,
the total size fraction was divided into 5 fractions as 350, 250, 182.5, 132.5 and 75 micron.
These five fractions were then exposed to calculations stated above. Pressure drop is
calculated for each of 20 run separately by Wasp model and modified Wasp model. The
pressure drop prediction error for each run is shown in figure 7.1. From this figure it is evident
that pressure drop prediction accuracy is increased considerably after modification was applied
on Wasp model.
Table 7.4 summarizes the AARE and for pressure drop prediction. This table shows that the
modified Wasp model able to bring down prediction error from 37.5% to 5.8%. Also there is a
significant increase in correlation co-efficient (R) for concentration profile prediction.
Concentration profiles were calculated for all 20 runs by taking C/CA from Wasp model and
from modified Wasp model and their agreement with experimental concentration profile was
shown in table 7.4 and 7.5. The marked increase in R-value in modified Wasp model justifies
its goodness over Wasp model.
Typical plot for experimental and calculated concentration profile by modified Wasp model
were shown in figure 7.2 to figure 7.8. From these plots, it is observed that calculated
concentrated profiles closely approximate the experimental concentration profiles. It can also
be concluded that Modified Wasp model is able to predict vertical concentration profile apart
from its main use of pressure drop prediction.
Table 7.4: Comparison of pressure drop and concentration profile prediction

Wasp
Performance parameter
model
Modified
Wasp
model
0.375
0.058
Standard deviation on pressure drop prediction
0.430
0.056
Average
0.154
0.902
Average absolute relative error (AARE) on

pressure drop prediction
Correlation
co-efficient
for
concentration profile
197
Seemi-empirical method
m
for preessure drop and
d concentrationn profile
a
calculated
d concentratiion profile ussing modified
d Wasp
Fiigure 7.2: Exxperimental and
m
model
for som
me typical data of Roco aand Shook (11983). a)Run
n A1:Vm= 1.666 m/s,
Cvf=8.37%,D=00.0515 m , b)R
Run A2:Vm= 3.78 m/s, Cvf=99.2%, D=0.05115m , c)Run A3:
A Vm=
1..66 m/s, Cvf=188.7%, D=0.0515 m
198
Ch
hapter7
a
calculated
d Wasp
m
model
for som
n A4:Vm= 4.117 m/s,
Cvf=18.9%,D=00.0515 m , b)R
Run A5:Vm= 11.66 m/s, Cvf=2
28%, D=0.05115m , c)Run A6:
A Vm=
4..33 m/s, Cvf=288.6%, D=0.0515 m
199
m
a
calculated
d Wasp
m
model
for som
me typical da
ata of Roco and Shook (1983).
(
a)Ru
un A7:Vm= 2.9 m/s,
Cvf=10.3%,D=00.263m b)Run
n A8:Vm= 3.55 m/s, Cvf=10%
%, D=0.263m,, c)Run A9: Vm= 2.9
m Cvf=19 %, D=0.263m
m/s,
200
Ch
hapter7
a
calculated
d Wasp
m
model
for som
n A10:Vm= 3.5 m/s,
Cvf=18.4%,D=00.263m b)Run
n A11:Vm= 2.99 m/s, Cvf=27%
%, D=0.263m, c)Run A12: Vm= 3.5
m Cvf=26.8 %,
m/s,
% D=0.263m
201
m

a
calculated
d Wasp
m
model
for som
n A13:Vm= 2.9 m/s,
Cvf=34.1%,D=00.263m b)Run
n A14:Vm= 3.5 m/s, Cvf=33.8%, D=0.2633m, c)Run A15: Vm=
3..16 m/s, Cvf=100.4 %, D=0.49
95 m
202
Ch
hapter7
a
calculated
d Wasp
m
model
for som
me typical datta of Roco and Shook (19
983). a)Run A16:Vm= 3.776 m/s,
Cvf=10.0 %,D=
=0.495 m b)R
Run A17:Vm= 3.07 m/s, Cvff=18.7%, D=00.495 m, c)Ru
un A18:
Vm= 3.76 m/s, Cvf=18.4 %, D=0.495 m
203
m
a
calculated
d Wasp
m
model
for som
me typical datta of Roco and Shook (19
983). a)Run A19:Vm= 3.116 m/s,
Cvf=27.3 %,D=0
0.495 m b)Ru
un A20:Vm= 3.76 m/s, Cvf=2
26.9%, D=0.4995 m
204
Chapter7
Table 7.5: Comparison of correlation co-efficient (R) for concentration profile prediction
by Wasp model and modified Wasp model.
Run
number
R by
Wasp
model
A1
A2
A3
A4
A5
A6
A7
A8
A9
A10
A11
A12
A13
A14
A15
A16
A17
A18
A19
A20
0.16
0.21
0.19
0.24
0.19
0.18
0.12
0.14
0.13
0.13
0.09
0.11
0.17
0.14
0.13
0.16
0.16
0.17
0.11
0.16
R by
modified
Wasp
model
0.96
0.99
0.95
0.97
0.91
0.85
0.96
0.94
0.87
0.91
0.91
0.84
0.83
0.83
0.93
0.89
0.91
0.95
0.79
0.85
7.7 Conclusion
Following conclusion can be drawn on the basis of the present study:
1.
Wasp model for pressure drop is modified by alleviating some of the restrictive
assumptions used in that model. The predicted pressure drop by modified Wasp
model has been compared with the experimental pressure drop data collected from
literature. A quite well agreement was observed and average prediction error was far
less than that by Wasp model.
2.
A new method was developed to calculate concentration profile using Wasp model as
a starting point. The predicted concentration profile by modified model was
compared with the experimental one collected from literature and found satisfactory.
205
Nomenclature
a:
height of layer A above bottom of the conduit
b:
Dimensionless particle diffusivity
C:
volumetric concentration of the particle size under consideration
C/CA :
Ratio of volumetric concentration of solids
CA :
volumetric concentration at pipe axis
Ci :
Individual solids volume fraction
Cvf :
Slurry efflux concentration by volume
Cvjbed :
Volumetric concentration of jth size particles in the bed portion
Cwf :
Efflux concentration by weight.
D:
Pipe diameter (m)
Es:
the mass transfer coefficient
fD :
Darcy friction factor
fm
Fanning friction factor
g:
Acceleration due to gravity
ivehicle
pressure gradient due to homogeneously distributed particles (vehicle)
iw :
Pressure gradient due to flow of carrier liquid
k:
von Karman coefficient
R :
Correlation co-efficient
r:
local radius for a point in the flow
Rem :
Reynolds number
RI :
inner diameter of pipe

206
Chapter7
S :
Specific gravity of solids.
u* :
Friction velocity
Uf:
Friction velocity (m/s)
Vm :
Flow velocity
Vt :
settling velocity
Vto :
Terminal velocity
wj :
Settling velocity of jth size particle
y:
distance from the pipe bottom
y:
Y/D, vertical distance from bottom of pipe
Z :
Richardson and Zakis co-efficient
the angle from the horizontal
The dimensionless particle diffusivity
Greek symbols
/D :
Pipe roughness
angle from the vertical starting at the lowest quadrant point
m :
The mass density of slurry
:
l :
Total mass of particles (Kg)

Viscosity of carrier liquid
207
References
1.
Brien M.P., Review of the theory of turbulent flow and its relations to sediment
transport. Transaction of the American Geophysical Union 14,487 491. (1933).
2.
Doron, P., Granica, D., Barnea, D., Slurry flow in horizontal pipes experimental
and modeling.Int.J.of Multiphase Flow 13,535 547. (1987).
3.
Ghanta, K.C., Purohit, N.K., Pressure drop prediction in hydraulic transport of bidispersed particles of coal and copper ore in pipeline. The Canadian Journal of
Chemical Engineering 77,127 131. (1999).
4.
Gillies, R.G., Hill, K.B., Mckibben, M.J., Shook, C.A., Solids transport by laminar
Newtonian flows. Powder Technology 104,269 277. (1999).
5.
Gillies, R.G., Shook, C.A., Modelling high concentration settling slurry flows. The
Canadian Journal of Chemical Engineering 78,709 716. (2000).
6.
Gillies, R.G., Shook, C.A., Wilson, K.C.,An improved two layer model for horizontal
slurry pipeline flow. The Canadian Journal of Chemical Engineering 69,173 178.
(1991)
7.
Hsu, Flow of non-colloidal slurries in pipeline.Ph.D. Thesis, University of Illionos.

(1986).
8.
Hunt, J.N., The turbulent transport of suspended sediment in open channels. Royal
Society of London, Proc., Series A 224 (1158), 322 335. (1954).
9.
Ismail, H.M., Turbulent transfer mechanism and suspended sediment in closed

channels. Transactions of ASCE 117,409 446. (1952).
10. Karabelas, A.J., Vertical distribution of dilute suspensions in turbulent pipe flow.
AIChE Journal 23,426 434. (1977).
11. Kaushal, D.R, Tomita, Y., Dighade, R.R., Concentration at the pipe bottom at
deposition velocity for transportation of commercial slurries through pipeline.
Powder Technology 125,89 101. (2002).
12. Kaushal, D.R, Tomita, Y., Solid concentration profiles and pressure drop in pipeline
flow of multisized particulate slurries. International journal of multiphase flow
28,1697-1717(2002).
13. Mukhtar, A., Investigations of the flow of multisized heterogeneous slurries in
straight pipe and pipe bends.Ph.D. Thesis, IIT, Delhi. (1991).
208
Chapter7
14. Newitt D.M.,Richardson J.F.,Abbott M ,Turtle R.B., , Hydraulic conveying of solids

in horizontal pipes, Trans. Instn Chem. Engrs, 33, 93-113 (1955)
15. Richardson, J.F., Zaki, W.M., Sedimentation and fluidization: part-1. Transactions of
the Institution of Chemical Engineers 32,35 53. (1954).
16. Roco M.C.,Shook C.A., Turbulent flow of incompressible mixtures, J. Fluids Engg.,
107, 224-231 (1985)
17. Roco, M.C., Shook, C.A., Computational methods for coal slurry pipeline with
heterogeneous size distribution. Powder Technology 39,159 176. (1984).
18. Roco, M.C., Shook, C.A., Modelling of slurry flow, the effect of particle size. The
Canadian Journal of Chemical Engineering 61,494 503. (1983).
19. Rouse, H., Modern conceptions of the mechanics of fluid turbulence. Transactions of
ASCE 102,463 505. (1937).
20. Seshadri, V., Malhotra, R.C., Sundar, K.S., Concentration and size distribution of
solids in slurry pipeline. Proc., 11th National Conference on Fluid Mechanics and
Fluid Power, BHEL, Hyderabad, India, pp.110 123. (1982).
21. Shook, C.A., Daniel, S.M., Flow of suspensions of solids in pipeline:I. Flow with a
stable stationary deposit. The Canadian Journal of Chemical Engineering 43,56 72.
(1965).
22. Shook, C.A., Daniel, S.M., Scott, J.A., Holgate, J.P., Flow of suspensions in
pipelines. The Canadian Journal of Chemical Engineering 46,238 244. (1968).
velocity for large particle size suspensions in round horizontal pipes, AIChE J,, 8,
373-378 (1962)
24. Thomas, D.G., Transport characteristics of suspensions: VIII.A note on the viscosity
of Newtonian suspensions of uniform spherical particles. Journal of Colloidal Science
20,267 277. (1965).
25. Turian R.M.,& Yuan T.F, Flow of slurries in pipelines, AIChE J., 23, 232-243 (1977)
26. Turton. R. and Levenspiel. O., A short note on the drag correlations for spheres,
Powder technology,47,83-86, (1986)
27. Wasp, E.J., Aude, T.C., Kenny, J.P., Seiter, R.H., Jacques, R.B., Deposition
velocities, transition velocities and spatial distribution of solids in slurry
pipelines.Proc.Hydrotransport 1,BHRA Fluid Engineering, Coventry, UK, paper
H4.2, pp.53 76. (1970).
209

28. Wasp, E.J., Cross country coal pipeline hydraulics. Pipeline News 35,20 25. (1963).
29. Wasp, E.J., Kenny, J.P., Gandhi, R.L., Solid Liquid Flow Slurry Pipeline
Transportation, rst ed.Trans.Tech. Publications, Clausthal, Germany. (1977).
30. Wilson K.C,& Pugh F.J., Dispersive-force modeling of turbulent suspension in
heterogeneous slurry flow, Can. J. Chem. Engg, 66, 721-727 (1988)
31. Wilson K.C., A unified physically-based analysis of solidliquid pipeline flow,
Proceedings of the 4th International Conference on Hydraulic Transport of Solids,
BHRA Fluid Engineering,, Cranfield UK, Paper A2 pp 116 (1976)
32. Wilson, K.C., Evaluation of interfacial friction for pipeline transport models.
Proceedings of the 11th International Conference on Hydraulic Transport of Solids,
BHRA Fluid Engineering, Cran field, UK, pp.107 116. (1988).
33. Zandi I.,Govatos G., Heterogeneous flow of solids in pipelines, Proc. ACSE, J.
Hydraul. Div.,, 93(HY3), 145-159 (1967)
210
Chapter8
CHAPTER 8
Computational Fluid Dynamics modeling of the Solid
liquid slurry flow in a pipeline
Abstract
An attempt has been made in the present study to develop generalized slurry flow model using
CFD and utilize the model to predict concentration profile. The purpose of CFD model is to
gain better insight into the solid liquid slurry flow in pipelines. Initially a two-dimensional
model problem was developed to understand the influence of the particle drag coefficient on
solid concentration profile. The preliminary simulations highlighted the need for the correct
modeling of the inter phase drag force. The various drag correlations available in literature
(with a new proposed correlation) were incorporated in a two-fluid model (Euler-Euler) along
with the standard k- turbulence model with mixture properties to simulate the turbulent solidliquid flow in a pipeline. The computational model was mapped on to a commercial CFD
solver FLUENT6.2 (of Fluent Inc., USA). The model predictions were compared with the
published experimental data of Kaushal et al. (2002) and Mukhtar (1991). The predicted
results show reasonably good agreement with the experimental data. After encouraging results
of two dimensions, simulations were set up in three dimensions because of the axial asymmetry
of the two-phase flow. To push the envelope of applicability of simulation, the recent data of
Kaushal (2005) (with solid concentration up to 50%) was selected to validate the three
dimensional simulation. The experimental data consists of water- glass bead slurry at two
different particle sizes (125 and 440 m) with different flow velocity (from 1 to 5 m/s) and
overall concentration up to 10 to 50% by volume. The predicted pressure drop and
concentration profile were validated by experimental data and showed excellent agreement.
Interesting findings were come out from the parametric study of velocity and concentration
profiles. The computational model and results discussed in this work would be useful for
extending the applications of CFD models for simulating large slurry pipelines.
Keywords: CFD, slurry flow, drag co-efficient, concentration profile, velocity profile
211
CFD Modeling
8.1 Introduction
Particle transport through pipes is an important operation in many industries including foods,
pharmaceuticals, chemicals, oil, mining, construction and power generation industries. In
many of these applications the carrier fluid may be highly viscous and may have a Newtonian
or non-Newtonian rheology and flow is usually turbulent. It has been the serious concern of
researchers around the world to develop accurate models for pressure drop and concentration
distribution in slurry pipeline over the years.
Need and benefit of accurately predicting the velocity profiles, concentration profiles and
pressure drop of slurry pipelines during design phase is enormous as it gives better selection of
slurry pumps, optimization of power consumption and thereby helps to maximize economic
benefit. Concentration distribution may be used to determine the parameters of direct
importance (mixture and solid flow rates) and the secondary effects such as wall abrasion and
particle degradation. The recent works of Kaushal and Tomita (2002) and Kumar et al. (2003)
are worth mentioning in the field of concentration distribution in slurry pipeline.
In spite of the inadequate fundamental knowledge required for the formulation and modeling
of multiphase turbulent flows, the need to predict slurry behavior handled in various industries
has motivated work, aiming at obtaining approximate solutions. Efforts are still on to develop
more reasonable correlation based model for the prediction of concentration profile in pipes
and in this direction, the work of Roco and Shook (1983, 1984), Gillies et al. (1999, 2000),
Mukhtar (1991) and Kaushal et al. (2002) is worth mentioning. Most of the equations available
in literature for predicting vertical solids concentration profiles in slurry pipeline are empirical
in nature and have been developed based on limited data for materials having equi-sized or
narrow size-range particles with very low concentrations. Most of the earlier studies on slurry
pipeline systems are based on moderate volumetric concentrations of solids (say up to 20%).
Much larger concentrations now coming into common use show more complicated behaviour.
Also in any practical situation, the solids are coarser in size with broad particle grading being
transported at large flow velocities.
Despite significant research efforts, prediction of solid concentration profile in slurry pipeline
is still an open problem for design engineers. Design of slurry pipelines relies on empirical
correlations obtained from the experimental data. These correlations are prone to great
uncertainty as one departs from the limited database that supports them. Moreover, for higher
values of solid concentration, very few experimental data on local solid concentration are
212
Chapter8
available because of the difficulties in the measurement techniques. Considering this, it would
be most useful to develop computational models, which will allow a priori estimation of the
solid concentration profile over the pipe cross section.
An attempt has been made in the present study to develop generalized slurry flow model using
CFD and utilize the model to predict concentration profile. A comprehensive computational
fluid dynamics (CFD) model was developed in the present study to gain insight into the solid
liquid slurry flow in pipelines.
In recent years, CFD becomes a powerful tool being used in the area like fluid flow, heat/mass
transfer, chemical reactions and related phenomena by solving mathematical equations that
govern these processes using a numerical algorithm on a computer.
A brief review of recent literature shows little progress in simulating flow in a slurry pipelines
using CFD.
For solid liquid multiphase flows, the complexity of modeling increases
considerably and this remains an area for further research and development. Due to the
inherent complexity of multiphase flows, from a physical as well a numerical point of view,
general applicable CFD codes are non-existent. The reasons for the lack of fundamental
knowledge on multiphase flows are three-fold:
(1) Multiphase flow is a very complex physical phenomenon where many flow types can occur
(solid liquid, gassolid, gasliquid, liquidliquid, etc.) and within each flow type several
possible flow regimes can exist (e.g. in slurry flow four regimes exist namely homogeneous
flow, heterogeneous flow, flow with moving bed, stationary bed).
(2) The complex physical laws and mathematical treatment of phenomena occurring in the
presence of the two phases (interface dynamics, coalescence, break-up, drag, solid liquid
interaction, . . .) are still largely under developed. For example, to date there is still no
agreement on the governing equations. In addition, proposed constitutive models are empirical
but often lack experimental validation for the conditions they are applied under.
(3) The numeric for solving the governing equations and closure laws of multiphase flows are
extremely complex. Very often multiphase flows show inherent oscillatory behavior, requiring
costly transient solution algorithms. Almost all CFD codes apply extensions of single-phase
solving procedures, leading to diffusive or unstable solutions, and require very short timesteps.
213
CFD Modeling
In spite of the major difficulties mentioned above, attempts have been made to simulate the
solid-liquid flow in pipelines. Few number of studies are focused on predicting the solid
concentration distribution in the experimental slurry pipelines. Although some degree of
success is reported, a number of limitations are apparent. Considering the limitations in the
published studies, the present work has been undertaken to systematically develop a CFD
based model to predict the solid concentration profile in slurry pipeline. The aim is to explore
the capability of CFD to model such complex flow.
In the present work, the solid suspension in a fully developed pipe flow was simulated. The
two-fluid model based on the Eulerian-Eulerian approach along with a standard k- turbulence
model with mixture properties was used. The computational model developed in this work was
used to simulate solid-liquid flow in the experimental setup used by Kaushal et al (2002) . The
model predictions were evaluated by comparing predictions with the experimental data.

CFD studies on solid liquid slurry flow in pipelines have not been widely studied as observed
from the literature and majority of the documented data focuses on empirical correlations of
concentration profile of water-based slurries of fine particles. There is, therefore, a clear need
for experimental data and CFD models to describe the flow of large particles in Newtonian
fluids as they are relevant to a number of industrial applications such as the conveying of
particulate food mixtures, gravel, and coal lumps.
The use of CFD, however, has been hampered by lack of understanding of the complex solid
liquid flows and that result has only been addressed in a handful of studies (Charles and
Charles (1971); Duckworth et al. (1983, 1986); Chhabra and Richardson (1985); Ghosh and
Shook,(1990); Fairhurst et al.(2001); Barigou et al.(2003); Gradeck et al.(2005); Legrand et
al. (2007); Eesa and Barigou( 2008)).
Detailed measurements of the flow field and pressure drop in these systems are scarce. Some
limited studies experimented with magnetic resonance imaging (McCarthy et al. 1996) and
ultrasound Doppler velocimetry Guer et al. (2003). Barigou et al. (2003) used positron
emission particle tracking (PEPT) to study the flow of coarse (d = 5 or 10mm) particle nearly
neutrally-buoyant alginate particles in shear-thinning carboxymethyl cellulose (CMC) fluids
and reported information on flow regimes, solid phase velocity profiles, and particle passage
time distributions. Computational modeling in this specific area has been even more limited. In
214
Chapter8
a rare attempt, Krampa-Morlu et al. (2004) used CFD to study the flow features of coarse
aqueous solidliquid slurries in turbulent upward flow including velocity profiles. The model,
implemented using the commercial CFD software CFX4.4 (ANSYS Inc.), was tested using
experimental data from Sumner et al. (1990). The particles had a density of 2650 kg/m3 and a
diameter of 0.47 or 1.7 mm and were simulated at concentrations upto 30% by volume. The
authors concluded that, using the default settings, the code failed to accurately predict
important features of the flow. Recently, Eesa and Barigou, 2008 investigated the capabilities
of CFD to model the flow of coarse nearly-neutral buoyant particles in shear-thinning CMC
fluids in a horizontal pipe for a limited number of flow cases (Eesa and Barigou, 2008). CFD
results of particle velocity profiles were validated with experimental data obtained by PEPT,
while pressure drop predictions were compared with a number of selected correlations from
the literature. In a recent paper (Eesa and Barigou, 2009), a CFD model based on the
commercial code ANSYS CFX 10.0, is used to conduct a detailed parametric study of the
transport of nearly neutral-buoyant coarse particles in laminar
non Newtonian both
horizontal and vertical flow.
8.2.1 Multiphase modeling

Literatures were explored to track the development of multiphase modeling using CFD. In
spite of the major difficulties mentioned above, significant progress has been made in various
areas of multiphase flow CFD. In their famous article from T. B. Anderson and Jackson
(1967) derive the continuum equations of motion for gasparticle flow. Later, many authors
improve the constitutive models and/or perform simulations for various gassolid flows. Ishii
(1975) derives governing equations for multiphase fluidfluid system from first principles,
going through the details of the derivation. The inherent assumptions of this derivation are
slightly different from the assumptions applied by Anderson and Jackson (1967), which
constrain the types of multiphase flow where they can be applied. Ishii (1975) also gives
appropriate closure models for various gasliquid flow conditions.
Up till the early 1980s, most computational models for multiphase flow on an application level
describe both the continuous phase and the dispersed phase by a so-called Eulerian model. In
these models the dispersed phase, like the continuous phase, is described as a continuous fluid
with appropriate closures. Hence, one calculates only the average local volume fraction,
velocity, etc. and not the properties of each individual dispersed particle or droplet. With the
increase of computational power, however, also the Lagrangian formulation of the dispersed
215
CFD Modeling
phase becomes feasible, but unfortunately the amount of dispersed particles or droplets that
can be tracked is, even today and in the near future, very limited. A separate class of methods
to track interfaces in fluidfluid flow are volume-of-fluid (VOF), level-set or front tracking
methods. Although the idea behind these methods is similar, their numerical implementation,
and thus behavior, may differ greatly. These methods describe both fluids with one set of
equations and solve another equation for the evolution of the interfaces between the two fluids.
The strength and weakness of various methods are discussed below as found from literatures.
8.2.2 Approaches to Multiphase Modeling

Advances in computational fluid mechanics have provided the basis for further insight into the
dynamics of multiphase flows.
Currently there are two approaches for the numerical
calculation of multiphase flows: the Euler-Lagrange approach and the Euler-Euler approach
(Figure 8.1).
8.2.2.1 The Euler-Lagrange Approach
The fluid phase is treated as a continuum by solving the time-averaged Navier-Stokes

equations, while the dispersed phase is solved by tracking a large number of particles, bubbles,
or droplets through the calculated flow field. The dispersed phase can exchange momentum,
mass, and energy with the fluid phase. A fundamental assumption made in this model is that
the dispersed second phase occupies a low volume fraction, even though high mass loading is
acceptable. The particle or droplet trajectories are computed individually at specified intervals
during the fluid phase calculation. This makes the model appropriate for the modeling of
spray dryers, coal and liquid fuel combustion, and some particle-laden flows, but inappropriate
for the modeling of liquid-liquid mixtures, fluidized beds, or any application where the volume
fraction of the second phase is not negligible.
216
Ch
hapter8
Appro
oachestoMu
ultiphasseMod
deling
Eu
ulerLagrange
Approacch
Eu
ulerEuler
A
Approach
VOF
Model
Mixture
Model
Eulerian
n
Model
Figure 8.1: Different

D
apprroaches to multiphase modeeling
8.2.2.2 The Eule
er-Euler Appro
oach
In
n the Euler-Euller approach, thhe different phases are treatedd mathematicaally as interpennetrating
co
ontinua. Sincee the volume off a phase cannot be occupied
d by the other phases,
p
the conncept of
phhasic volume fraction
f
is introoduced. Thesse volume fracctions are assum
med to be conntinuous
fu
unctions of spaace and time and
a their sum iis equal to onee. Conservatioon equations for
f each
phhase are deriveed to obtain a set of equatiions, which haave similar struucture for all phases.
Thhese equations are closed by providing relations thatt are obtainedd from empiriical info
ormation or, in the case of graanular flows, by application of
o kinetic theoory.
In
n literatures, th
hree different Euler-Euler
E
muultiphase models are availablee: the volume of fluid
(V
VOF) model, thhe mixture moddel, and the Euulerian model.
8.2.2.2.1 The VO
OF Model
Thhe VOF modeel is a surfacee-tracking techhnique appliedd to a fixed E

Eulerian meshh. It is
deesigned for tw
wo or more imm
miscible fluidss where the poosition of the interface betw
ween the
fluuids is of interrest. In the VO
OF model, a single set of moomentum equaations is shared
d by the
fluuids, and the volume fractioon of each off the fluids in each computaational cell is tracked
thhroughout the domain.
d
Applications of thee VOF model include
i
stratifieed flows, free--surface
floows, filling, slooshing, the mootion of large bubbles
b
in a liquuid, the motionn of liquid afteer a dam
brreak, the prediiction of jet brreakup (surfacee tension), andd the steady orr transient traccking of
an
ny liquid-gas innterface.
217
CFD Modeling
8.2.2.2.2 The Mixture Model
The mixture model is designed for two or more phases (fluid or particulate). Likewise in the
Eulerian model, this model treats the phases as interpenetrating continua. The mixture model
solves the mixture momentum equation and prescribes relative velocities to describe the
dispersed phases. Applications of the mixture model include particle-laden flows with low
loading, bubbly flows, sedimentation, and cyclone separators. The mixture model can also be
used without relative velocities for the dispersed phases to model homogeneous multiphase
flow.
8.2.2.2.3 The Eulerian Model
The Eulerian model is the most complex of the multiphase models. It solves a set of n
momentum and continuity equations for each phase.
Coupling is achieved through the
pressure and interphase exchange coefficients. The manner in which this coupling is handled
depends upon the type of phases involved. Granular (fluid-solid) flows are handled differently
than non-granular (fluid-fluid) flows. For granular flows, the properties are obtained from
application of kinetic theory. Momentum exchange between the phases is also dependent
upon the type of mixture being modeled. Applications of the Eulerian multiphase model
include bubble columns, risers, particle suspension, and fluidized beds.
8.2.3 Choosing a Multiphase Model

The first step in solving any multiphase problem is to determine which of the regimes best
represents our flow. Literatures were cited to search some broad guidelines for determining
appropriate models for each regime and summarized in figure 8.2. Fluent documents (2001)
provide details about how to determine the degree of interphase coupling for flows involving
bubbles, droplets, or particles, and the appropriate model for different amounts of coupling.
8.2.4 Effect of Particulate Loading

Particulate loading has a major impact on phase interactions. The particulate loading is defined
as the mass density ratio of the dispersed phase (d) to that of the carrier phase (c):
[8.1]
218
Ch
hapter8
Thhe material deensity ratio
is greater tthan 1000 for gas-solid flow

ws, about 1 forr liquid-
soolid flows, and less than 0.001 for gas-liquidd flows.

U
Using
these parameters it is possible
p
to estimate the averaage distance between the inddividual
paarticles of the particulate
p
phaase. An estimaate of this distaance has been given by Crow
we et al.
(1998)
1
:
[8.2]
W
Where,
dispersedphase
model
volumefractions Discreetephasem
<1
10%
dispersedphase
Mixturemodel
volumefractions
Eulerianmodel
>1
10%
Pneu
umatic
tran
nsport
Slurryyflowor
hydrotransport
Mixturemodelfo
orhomogen
neousflow
w
Eulerianmodelfo
orgranularflow
Mixturemodel
Eulerianmodel
m
mod
del
Figure 8.2:: Guidelines foor choosing multiphase
nformation aboout these paraameters is impportant for dettermining how
w the dispersedd phase
In
shhould be treateed. For examp
ple, for a gas--particle flow with a particuulate loading of
o 1,the
innterparticle spaace is about 8 and
a the particlee can therefore be treated as iisolated (i.e., very
v
low
paarticulate loadiing). Dependinng on the partiiculate loading
g, the degree of
o interaction between
b
thhe phases can be
b divided into three categoriees:
219
CFD Modeling
For very low loading, the coupling between the phases is one-way; i.e., the fluid
carrier influences the particles via drag and turbulence, but the particles have no
influence on the fluid carrier. The discrete phase, mixture, and Eulerian models can
all handle this type of problem correctly. Since the Eulerian model is the most
expensive, the discrete phase or mixture model is recommended
For intermediate loading, the coupling is two-way; i.e., the fluid carrier influences the
particulate phase via drag and turbulence, but the particles in turn influence the carrier
fluid via reduction in mean momentum and turbulence. The discrete phase, mixture,
and Eulerian models are all applicable in this case, but one need to take into account
other factor (Stokes number) in order to decide which model is more appropriate.
For high loading, there is two-way coupling plus particle pressure and viscous
stresses due to particles (four-way coupling). Only the Eulerian model will handle
this type of problem correctly.
8.2.5 Significance of the Stokes Number

For systems with intermediate particulate loading, estimating the value of the Stokes number
can help to select the most appropriate model. The Stokes number can be defined as the
relation between the particle response time and the system response time:
[8.3]
Where,
ts is based on the characteristic length (Ls) and the characteristic velocity (Vs) of the system
under investigation.
8.2.6 Guidelines for choosing appropriate model
For St
1.0, the particle will follow the flow closely and any of the three models
(discrete phase, mixture, or Eulerian) is applicable. One can therefore choose the
least expensive (the mixture model, in most cases), or the most appropriate
considering other factors.
220
Chapter8
For St > 1.0, the particles will move independently of the flow and either the discrete
phase model or the Eulerian model is applicable.
For St 1.0, again any of the three models is applicable. One can choose the least
expensive or the most appropriate considering other factors.
If there is a wide distribution of the dispersed phases, the mixture model may be
preferable. If the dispersed phases are concentrated just in portions of the domain, one
should use the Eulerian model instead.
If interphase drag laws that are applicable to system available (either within FLUENT
or through a user-defined function), the Eulerian model can usually provide more
accurate results than the mixture model.
If the interphase drag laws are unknown or their applicability to system is

questionable, the mixture model may be a better choice.
If one wants to solve a simpler problem, which requires less computational effort, the
mixture model may be a better option, since it solves a smaller number of equations
than the Eulerian model.
If accuracy is more important than computational effort, the Eulerian model is a

better choice. Keep in mind, however, that the complexity of the Eulerian model can
make it less computationally stable than the mixture model.
8.3 Formulation of multiphase CFD model

The EulerianEulerian two-fluid model was adopted here, whereby both the liquid and solid
phases are considered as continua. Treating the solid phase as an Eulerian phase is possible
provided that the inter-phase interactions are adequately modeled. In fact, the Eulerian
approach has been reported to be efficient for simulating multiphase flows once the interaction
terms are included (Hu et al., 2001). The EulerianLagrangian model, however, which
simulates the solid phase as a discrete phase and so allows particle tracking, is in principle
more realistic. However, after evaluating the relevant literature as well as conducting a number
of simulation trials, it was concluded that the number of dispersed particles that can be tracked
within all the different commercial CFD software available is currently very limited, thus
restricting the applicability of the EulerianLagrangian model to only dilute mixtures well
below 5% by volume (van Wachem and Almstedt, 2003).
221
CFD Modeling
The turbulent flow of coarse particles in a Newtonian fluid is assumed to be governed by the
following equations which form the basis of the EulerianEulerian CFD model used.
8.3.1 Eulerian Model

In the Eulerian-Eulerian approach, two phases are considered to be interpenetrating continua.
For the present CFD simulations, the Eulerian-Eulerian multiphase model implemented in the
commercial code Fluent 6.2 was used. With this approach, the continuity and the momentum
equations are solved for each phase and therefore, the determination of separate flow field
solutions is allowed. The Eulerian model is the most complex and computationally intensive
among the multiphase models. It solves a set of n momentum and continuity equations for
each phase. Coupling is achieved through the pressure and interphase exchange coefficients.
For granular flows, the properties are obtained from application of kinetic theory.
8.3.1.1 Continuity Equation

The continuity equations for a generic phase q is given by
(qq ) + (q qq ) = 0
t
[8.4]
The solution of this equation for each secondary phase, along with the condition that the
volume fractions sum to one, allows for the calculation of the primary-phase volume fraction.
222
Ch
hapter8
8..3.1.2 Momen
ntum Equatio
ons
Flluid-fluid mom
mentum equatiions
Thhe conservationn of momentum
m for a fluid phhase q is,
[8.5]
is an exteernal body forcce,
is a lift forcee,
is a virtual mass force,
is an intteraction force between phasees, and p is thee pressure shareed by all phasees.
8..3.1.3 Fluid-s
solid momen
ntum equations
FL
LUENT uses a multi-fluid granular modeel to describe the flow behaavior of a fluuid-solid
m
mixture.
The soolid-phase streesses are derivved by making
g an analogy between the random
paarticle motion arising
a
from paarticle-particlee collisions andd the thermal m
motion of molecules in
a gas, taking innto account thee inelasticity oof the granularr phase. As thhe case for a gas,
g the
inntensity of the particle
p
velocitty fluctuations determines the stresses, visccosity, and presssure of
thhe solid phasee. The kineticc energy assocciated with thhe particle velocity fluctuattions is
reepresented by a pseudo therm
mal" or granular temperaturee which is propportional to thhe mean
sqquare of the ranndom motion of
o particles.
Thhe conservationn of momentum
m for the
solid phase is,,

=
[8.6]
223
CFD
D Modeling
W
Where,
ps is the
solids prressure, Kls = Ksl is the momeentum exchangge coefficient between

b
fluuid phase l annd solid phase s, N is the tottal number of phases. The lift
l force Flift,s and the
viirtual mass forrce Fvm,s have been neglecteed in the calcu
ulations, becauuse they give a minor
co
ontribution to the
t solution witth respect to thhe other terms.
hase exchange co-efficie

ent
8..3.1.4 Interph
It can be seen inn Equations 8.5
5 & 8.6 that moomentum exchaange between tthe phases is based
b
on
thhe value of the fluid-fluid excchange coefficiient Kpq and, fo
or granular flow
ws, the fluid-soolid and
soolid-solid exchaange coefficien
nts Kls.
8..3.1.5 Fluid-S
Solid Exchan
nge Coefficie
ent
Thhe fluid-solid exchange
e
coeffficient Ksl is in the following general form:
[8.7]
W
Where
f is defiined differentlly for the diffferent exchang
ge-coefficient models
m
(as deescribed
beelow), and
, the ``particullate relaxation time'', is defineed as,

[8.8]
W
Where,
d s is thee diameter of particles of phasse s.
A definitions of
All
o f include a drag
d
function ((CD) that is bassed on the relattive Reynolds number
(R
Res). It is the drag
d
function that differs am
mong the exchhange-coefficient models. Following
thhree models foound in literatuure which is ppromising andd widely used for calculatinng solid
liqquid interaction
n in slurry flow
w.
Syamla
al-O'Brien moodel (1989)
Fo
or this model f is defined as,
[8.9
9]
224
Ch
hapter8
W
Where
the drag function has a form derived by
b Dalla Valle (1948)
[8.10]
Thhis model is based

b
on meassurements of thhe terminal veelocities of parrticles in fluid
dized or
seettling beds, with
w
correlatio
ons which aree function of the volume fraction and relative
Reeynolds numbeer:
8.11]
[8
W
Where
the subsscript l is for the
t
diiameter of the
fluid phase,
p
s is for the
solid phase, and d
is the
solid phasse particles.

e
coeffficient has the form
[8.12]
W
Where
v r, s is thee terminal velo
ocity correlatioon for the solid phase:
[8.13]
w
with
[8.14]
nd
an
225
CFD
D Modeling
[8.15]
[
for
nd
an
for
[8.16]
[
Thhis model is apppropriate wheen the solids shear

s
stresses are
a defined acccording to Syaamlal et
al. (1993)
nd Yu (1966) model
Wen an
or the model of
o Wen and Yu
u (1966), the flluid-solid exchhange coefficieent is of the following
Fo
fo
orm:
[8.17]
W
Where,
[8.18]
Thhis model is apppropriate for dilute

d
systems.
pow model (19992)

Gidasp
Thhe Gidaspow model

m
(1992) is a combinatioon of the Wen and
a Yu model (1966) and thee Ergun
eq
quation (1952).
e
coeffficient K sl is off the following
g form:
[8.19]
226
Ch
hapter8
W
Where
[8.20]
W
When
,
[8.21]
Thhis model is reecommended foor dense fluidizzed beds.

So
olid-Solid Exchange Coefficcient
Thhe solid-solid exchange
e
coeffficient Kls has tthe following form
f
(Syamlal,, 1987)
[8.22]
W
Where,
elss -co-efficient of restitution
Cfr,ls
- coefficiennt of friction beetween the lth aand sth solid-p
phase particles (Cfr,ls = 0)
f
dl- the diameter of the particless of solid l
g 0,ls- the radial distribution
d
coeefficient
8..3.1.6 Lift Forces
Fo
or multiphase flows, FLUEN
NT can include the effect off lift forces onn the secondaryy phase
paarticles (or dro
oplets or bubbbles). These lifft forces act on
n a particle m
mainly due to velocity
v
grradients in thee primary-phasse flow field. T
The lift force will be more significant for larger
paarticles, but thee FLUENT moodel assumes tthat the particlee diameter is much
m
smaller than
t
the
innter-particle spacing. Thus, th
he inclusion of
o lift forces iss not appropriaate for closely packed
paarticles or for very
v
small particles.
Thhe lift force actting on a seconndary phase, p and a primary phase, q is com
mputed from
[8.23]
227
CFD Modeling
The lift force Flift will be added to the right-hand side of the momentum equation for both
phases. In most cases, the lift force is insignificant compared to the drag force, so there is no
reason to include this extra term. If the lift force is significant (e.g., if the phases separate
quickly), it may be appropriate to include this term. In this simulation, the effect of lift force
has been checked.
8.3.1.7 Solid pressure
For granular flows in the compressible regime (i.e., where the solids volume fraction is less
than its maximum allowed value), a solids pressure is calculated independently and used for
the pressure gradient term, in the granular-phase momentum equation. Because a Maxwellian
velocity distribution is used for the particles, a granular temperature is introduced into the
model, and appears in the expression for the solids pressure and viscosities. The solids
pressure is composed of a kinetic term and a second term due to particle collisions:
[8.24]
Where, ess is the coefficient of restitution for particle collisions, g0,ss is the radial distribution
function, and is the granular temperature. FLUENT uses a default value of 0.9 for ess, but
the value can be adjusted to suit the particle type.
The function g0,ss is a distribution function that governs the transition from the compressible"
condition with s <
s,max,
where the spacing between the solid particles can continue to
decrease, to the incompressible" condition with

s = s,max, where no further decrease in the spacing can occur. A value of 0.63 is the default
for s,max, but it can be modified during the problem setup.
8.3.1.8 Radial Distribution Function
The radial distribution function, g0, is a correction factor that modifies the probability of
collisions between grains when the solid granular phase becomes dense. This function may
also be interpreted as the non-dimensional distance between spheres:
228
Chapter8
[8.25]
Where, s is the distance between grains. From Equation it can be observed that for a dilute
solid phase s , and therefore g0 1. When the solid phase compacts, s 0 and g0. The
radial distribution function is closely connected to the factor of Chapman and Cowling's
(1990) theory of non-uniform gases. is equal to one for a rare gas, and increases and tends to
infinity when the molecules are so close together that motion is not possible.
In the literature there is no unique formulation for the radial distribution function. FLUENT
employs as follows:
[8.26]
8.3.1.9 Solids Shear Stresses
The solids stress tensor contains shear and bulk viscosities arising from particle momentum
exchange due to translation and collision. A frictional component of viscosity can also be
included to account for the viscous-plastic transition that occurs when particles of a solid
phase reach the maximum solid volume fraction.
The collision and kinetic parts, and the optional frictional part, are added to give the solids
shear viscosity:
[8.27]
Collisional Viscosity
The collisional part of the shear viscosity is modeled as Gidaspow et al.,1992 & Syamlal et
al,1993
229
CFD Modeling
[8.28]
Kinetic Viscosity
FLUENT provides two expressions for the kinetic part. The default expression is from
Syamlal et al. (1993)
[8.29]
The following optional expression from Gidaspow et al. (1992) is also available:
[8.30]
Bulk Viscosity
The solids bulk viscosity accounts for the resistance of the granular particles to compression
and expansion. It has the following form (Lun et al. (1984)):
[8.31]
Frictional Viscosity
In dense flow at low shear, where the secondary volume fraction for a solid phase nears the
packing limit, the generation of stress is mainly due to friction between particles. If the
frictional viscosity is included in the calculation, FLUENT uses Schaeffer's (1987)
expression:
[8.32]
230
Ch
hapter8
W
Where,
ps is thhe solids pressure, is the angle of interrnal friction, aand I2D is the second
innvariant of the stress
s
tensor.
nt model
8..3.2 Turbulen
Fo
or the turbulen
nt multiphase fllow, a standardd k- turbulencce model was uused. In this stu
udy, the
sim
mplest k- tuurbulence mod
del was assum
med, referred too the Mixture Model where only a
co
ouple of k and equations arre solved and the
t physical prroperties of thee mixture are adopted
a
Thhe two phasess are assumed to share the same
s
k and values and theerefore the intterphase
tuurbulence transfer is not consiidered.
Thhe k and
equuations describbing this modell are as followss:
[8.33]
nd
an
[8.34]
W
Where,
the mixtture density and
d velocity,
and
, are computed froom

[8.35]
an
nd
[8.36]
Thhe turbulent viscosity,
, is computed from,
231
CFD
D Modeling
[8.37]
an
nd the productiion of turbulennce kinetic enerrgy, G k, m, is coomputed from,
[8.38]
Thhe model consttants
and
have the folllowing values :
8.4 Descripttion of CFD

D simulatio
on
mensional sim
mulation
8..4.1 Two dim
In
nitially simulation was set upp in two dimennsions for com
mmercial softwaare package FL
LUENT
(F
Fluent Inc., USA). To validatee the CFD moddel 15 sets of Kaushal
K
.(20022) data and 15 sets
off Mukhtar (1991). data for multisized
m
partticulate zinc taailings slurriess flowing throuugh 105
m diameter pip
mm
pelines have beeen consideredd. The particle size range usedd in all these data
d was
w
wide
enough to cover the ran
nge expected inn commercial slurries. The rrange of experrimental
daata taken is summarized
s
inn Table 8.1 . In Kaushal (2002) and M
Mukhtar (1991
1) data,
co
oncentration prrofiles have beeen reported foor different effflux concentrattions at differeent flow
veelocities. Thesse data are used
u
for propposing the CFD
C
model foor the prediction of
co
oncentration profile.
p
The detail
d
of expeerimental set up and methhod of collecction of
ex
xperimental datta can be foundd in reference K
Kaushal (20022).
232
Chapter8
Table 8.1 Experimental data used in the present study

Author
material
specific
Gravity
pipe
dia,
mean particle
dia of efflux
(mm) sample (m)
efflux
Flow
Pressure
conc.
Velocity
drop
range(%)
range(m/s)
range
(m/mwc)
Kaushal
(2002)
Zinc
2.820
105
34.95
3.8-26
2.597
105
69.24
4.09-25.8
2-3.5
0.062-0.214
tailings
Mukhtar
Zinc
(1991)
tailings
1.48-3.08
0.036-0.147
The geometry used consisted of a pipe of diameter D = 105 mm. The pipe length, L, was much
greater than the maximum entrance length, Le, required for fully developed flow. In singlephase Newtonian laminar flow, Le can be estimated from (Shook and Roco, 1991)
Where,
is the tube Reynolds number and um is the mean flow velocity. There
is no correlation available for estimating Le in two-phase solidliquid flow. However, for

almost all the particles used here, above equation should give a reasonable estimate of Le.
Whilst such estimates were used as a guide, a series of numerical trials were conducted using
different pipe lengths. For all of the cases simulated here, a pipe length of 3 m was sufficient
to give a fully developed solidliquid flow through most of the pipe length whilst keeping
computational cost low. Using a longer pipe did not affect the results. The geometry was
meshed in to approximately 1.8105 tetrahedral cells.
233
CFD Modeling
Table 8.2: Different inputs for simulation in FLUENT
Category
Description
Input
Model
Multiphase
Multiphase model
Eulerian
Number of phases
Inter phase mass transfer
Off
Solver
Segregated
Space
2D
Formulation
Implicit
Velocity formulation
Absolute
Time
Steady
Model
k-epsilon ( 2 eqn)
Viscous
k-epsilon model
Standard
model
near wall treatment
standard wall function
k-epsilon multiphase model
Mixture
Drag co-efficient
User defined
Lift co-efficient
None
Restitution co-efficient
0.9
Phase 1
Water
Phase 2
Zinc
phase 2
Granular
particle diameter
Constant
Granular viscosity
Syamlal-obrien
Granular bulk viscosity
Lun-et-al
Frictional viscosity
None
Packing limit
0.63
Operating
Operating pressure
101325 Pa
condition
Gravitational acceleration
9.81 m/s2 at -Y direction
Velocity specification method at velocity
Magnitude normal to boundary
Model
Solver
Phase
interaction
Phase details
inlet
Primary
Boundary
condition
and
secondary
phase
velocity
Constant
Volume fraction of secondary phase at
Constant
magnitude
velocity inlet
Back flow volume fraction of secondary
Constant
phase at pressure outlet
Solution
control
Gauge pressure at pressure outlet
Pressure velocity coupling
Phased coupled SIMPLE
Momentum
First order upwind
Volume fraction
First order upwind
Turbulent kinetic energy
First order upwind
Turbulent dissipation rate
First order upwind
234
Chapter8
For Eulerian multiphase calculations, we use the Phase Coupled SIMPLE (PC-SIMPLE)
algorithm (Vasquezand Ivanov ,2000) for the pressure-velocity coupling. PC-SIMPLE is an
extension of the SIMPLE algorithm to multiphase flows. The velocities are solved coupled by
phases, but in a segregated fashion. The geometry was meshed into approximately 50000
quadrilateral cells in GAMBIT 2.2 pre-processor. Dense computational grid was used because
of the pilot-scale pipe dimensions. The initial conditions were: a) uniform fully developed
velocity profile at pipe inlet and b) the solid particles uniformly distributed at pipe inlet. The
first-order upwind discretization scheme was used for the volume fraction, momentum
equations, turbulence kinetic energy (k), and turbulence dissipation rate (). The various inputs
to build the model in FLUENT solver were summarized in Table 8.2. All the simulations were
performed in double precision.
Simulations of the carrier fluid flowing alone were performed first to serve both as an initial
validation of the code and the numerical grid, and to reveal the effects of solid particles on the
liquid velocity (by deselecting the volume fraction equations). Once the initial solution for the
primary phase was obtained, the volume fraction equations were turned back on and the
calculation continued with all phases. An inlet flow rate boundary condition was used at the
pipe inlet, while static pressure was specified at the outlet. The homogeneous volumetric
fraction of each phase was specified at the inlet. Using flow rate as a boundary condition is the
common way of formulating pipe flow problems, i.e. one designs a system to deliver a given
flow rate. It is noted however that using a pressure-specified inlet boundary condition is a
stricter way of testing the CFD code as a flow rate boundary condition may be perceived as a
way of helping to steer the simulation towards the right solution. This pressure option was
tested but it did not affect the results of the CFD computations .The usual no-slip boundary
condition was adopted at the pipe wall. The solution was assumed to have converged when the
mass and momentum residuals reached 104 for all of the equations solved. This typically
required 150 iterations.
Due to the complexity of the solidliquid flows considered here, simulations initially required
a great deal of experimentation and optimization. Of primary importance was the appropriate
modeling of forces and interactions between the two phases. The drag force was modeled
using the Syamlal-O'Brien model (1993) , Wen and Yu (1966) model and Gidaspow
model (1992). The lift force was modeled using Equation 8.23 with the default co-efficient =
0.5. A range of co-efficient values were tested but they did not affect the results.
235
CFD Modeling
8.4.2 Validation of CFD simulation
As pointed out above, detailed measurements of the flow field in solidliquid suspensions are
very scarce due to the lack of suitable measurement techniques. As pointed out earlier, 15 set
of Kaushal (2002) and 15 set of Mukhtar (1991) vertical concentration profile data have been
considered to validate the CFD model. The validation of velocity profile and pressure drop in
single-phase fluid flow simulations, however, was based on exact analytical equations
available in the literature.
8.4.3 Comparison between measured and predicted concentration profiles

based on Syamlal-O'Brien model (1993) model, Wen and Yu model (1966) and
Gidaspow model (1992)
CFD simulations were done using all the 30 cases of Kaushal (2002) and Mukhtar (1991) data
for all the above three model and some typical concentration profile shown in Figure 8.3-8.5.
These figures present the comparison between measured and predicted concentration profiles
for Kaushal (2002) and Mukhtar (1991) data, where y/D is the reduced vertical coordinate
from pipe bottom, Cv( y/D) the concentration by volume at y /D and Vm the flow velocity of
slurry. It is seen that for almost all the data, above three models fail to predict the
concentration profile accurately. All the three models predicted the concentration profile as flat
and look like homogeneous profile. However the experimental concentration profile is not flat
and more solid concentrate at the bottom. From the experimental solid concentration profile as
seen in Figure 8.3-8.5, it is clear that the settling tendency leads to a significant gradation in
the concentration of solids in the slurry. The concentration of solids is higher in the lower
sections of the horizontal pipe. The extent of the accumulation of solids in the lower section
depends strongly on the velocity of the slurry in the pipeline. With higher velocity, higher is
the turbulence level and greater is the ability of the carrier fluid to keep the particles in
suspension. It is the upward motion of eddy currents transverse to the main direction of flow of
the slurry that is responsible for maintaining the particles in suspension. At very high
turbulence levels the suspension is almost homogeneous with very good dispersion of the
solids while at low turbulence levels the particles settle towards the floor of the channel and
can in fact remain in contact with the flow. At this stage, solids are transported as a sliding bed
under the influence of the pressure gradient in the fluid. When the turbulence level is not high
236
Chapter8
to prevent any deposition of particles on the floor of the channel, the flow regime is described
as being heterogeneous suspension.
From the quantitative comparison, it is clear that these three models are not taking into
consideration the changes in fluid and flow properties that occur with increase in solid
concentration at the pipe bottom. All the above three models could not capture the variations
of drag co-efficient with the increase in solid concentration at the pipe bottom. In light of the
above shortcomings, an effort was made in the present study to modify the existing models to
incorporate the effect of solid concentration on drag co-efficient.
8.4.4 Description of modified model
As discussed earlier, the estimation of drag is critical for accurate prediction of solid
concentration distribution. In a solid-liquid pipeline, the interphase drag coefficient, CD, is a
complex function of a drag coefficient in a stagnant liquid, CD0, and the prevailing turbulence
level. In the present work, we have critically examined the available information to select the
appropriate interphase drag formulation.
Computer program (user defined function) was developed based on following steps to
implement the modified model:
1.
Particle Reynolds number for size fraction was calculated as,
[8.39]
where
2.
is the relative velocity of particles with liquid.
Drag co-efficient was calculated using Turton-Levenspiel ( 1986 ) equation,

1
0.173
.
.
[8.40]
This equation is chosen purposefully from literature and it is found that this equation predicts
accurately the experimental value of drag co-efficient in the range of Rep < 2*105
3.
Terminal (Vto) velocity was calculated using following equation,
[8.41]
4.
If the surrounding liquid is turbulent, as in the case of a slurry pipeline, the prevailing
turbulence and solid particles is expected to influence the effective settling velocity of the
237
CFD Modeling
particles. However unhindered settling velocity in the calculations is not accounting for
the effect of concentration, particle size distribution and pipe walls. Richardson and Zaki
(1954) have already given a correlation for hindered settling velocity by taking into
consideration the above factors.
Richardson and Zakis co-efficient (Z) was calculated as,
4.65
5.
V
6.
Ksl
1.95
4.35
17.5
4.45
18.0
0.002
.
.
0.2
1.0
0.2
1.0
[8.42]
Hindered settling velocity was determined by,

1
[8.43]
Fluid-solid exchange coefficient is calculated from

CD
[8.44]
The above user defined function was coupled in Fluent solver.
8.4.5 Comparison between measured and predicted concentration profiles

based on modified model
The overall concentration profiles are predicted by modified model for 30 sets of experimental
data reported by Kaushal (2002) and Mukhtar (1991) and some of them are shown in Figures
8.38.5. It is observed that for almost all the data, the modified model gives an almost exact fit
between the measured and predicted overall concentration profiles. It may be recalled that in
the earlier predictions by the Syamlal-O'Brien model , Wen and Yu model and Gidaspow
model (Figs. 8.1), the predicted concentration profiles were flat and large deviations observed
from experimental profile. It shows that the predictions by modified model for overall
concentration profile are better than the predictions by above three model. The modified model
238
Ch
hapter8
is actually takeen care of chhange of settling velocity and

a
drag coefffient with thhe solid
co
oncentration att the bottom off the pipe. Acttually when wee plot the conttours of concen
ntration
prrofile (figure 8.5) in CFD wee found that sollid concentratiion is very low
w at the top of the
t pipe
an
nd settled at th
he bottom of thhe pipe as expected. Due to this differencee the drag co-eefficient
an
nd settling velo
ocity is not connstant throughoout the pipe cro
oss section andd they vary along with
thhe concentratioon. This non uniform drag co-efficient and
a
settling velocity
v
gives rise to
diifferent solid liquid exchangee co-efficient accross pipe crosss section as peer equation 8.44
8..4.6 Three dimensional simulation

A
After
encouragiing results of two dimensions, simulationns were set upp in three dim
mensions
beecause of the axial
a
asymmettry of the two--phase flow. Thhe purpose is twofold: to geet better
innsight of the innherent physicss of solid liquid interaction and
a to verify hhow the availab
ble drag
co
o-efficient moddels perform in
i three dimennsions. The thrree dimensionn simulations helps
h
to
viisualize better about
a
the distriibution of solidd and liquid at pipe outlet.
239
CFD
D Modeling
Too push the env
velope of appllicability of sim
mulation, the recent
r
data of Kaushal(2005
5) (refer
taable 8.3) was selected
s
to vallidate the three dimensional simulation. T
The experimen
nts were
co
onducted in 544.9 mm diametter horizontal pipe for two sizes
s
of glass beads
b
of whicch mean
diiameter and geometric
g
stand
dard deviationn are 440 mm
m and 1.2 andd 125 mm an
nd 1.15,
reespectively. Floow velocity waas up to 1-5 m
m/s and overall concentration was up to 10 to 50%
byy volume for each
e
velocity. Initially,
I
the drrag force was modeled usingg the Syamlal-O
O'Brien
m
model
(1989) , Wen
W and Yu (1966) model aand Gidaspow model (1992).. All the modeels have
peerformed succeessfully and litttle difference w
was found at thhe final results..
240
Ch
hapter8
Tablee 8.3: Data useed in 3D CFD simulation

Au
uthor
material
specific
G
Gravity
Kaaushal et al
Glass
G
beads
2.47
2
pipe
mean particle
efflux
F
Flow
dia,
dia of efflux
conc.
V
Velocity
(mm)
sample (m)
range(%)
rrange(m/s)
54.9
125 & 440
100-50
1-5
(20005)
mulation
8..4.7 Results and discussion of 3D sim
8..4.7.1 Concentration proffile

Fiigure 8.6 to figure
f
8.9 show
w the experim
mental and CF
FD predicted vertical
v
concen
ntration
prrofile of slurry
y of 125 micrron glass beadds in 54.9 mm
m diameter pippe at differentt efflux
co
oncentration an
nd flow velocitty. Figure 8.100 to 8.12 show same for slurrry of 440 microon glass
beeads at differen
nt efflux concentration and flow velocity. The agreemennt between callculated
an
nd experimentaal concentrationn profiles is quuite well as evident from thesse curves.
Fiigure 8.13 andd 8.14 show co
oncentration prrofiles in the vertical
v
plane ffor slurry of 125 m,
4440 m respectiively, by Cv/Cvvf vs. y/D, whhere Cv is the volumetric
v
conncentration at y = y/D,
y being distancee from the pipe bottom and D the pipe diameter.
buted in the vertical plane with
w the
It is observed thhat the particlees are asymmeetrically distrib
mmetry increasiing with increease in particle size becausee of the graviitational
deegree of asym
efffect. It is also observed thatt the degree off asymmetry foor the same ovverall concentraation of
sluurry increases with decreasinng flow velocitty. This is expected because with decrease in flow
veelocity there will
w be a decreasse in turbulent energy, whichh is responsiblee for keeping thhe
241
CFD Modeling
solids in suspension. From these figures, it is also observed that for a given velocity,
increasing concentration reduces the asymmetry because of enhanced interference effect
between solid particles. The effect of this interference is so strong that the asymmetry even at
lower velocities is very much reduced at higher concentrations. Therefore it can be concluded
that degree of asymmetry in the concentration profiles in the vertical plane depends upon the
particle size, velocity of flow and overall concentration of slurry.
Measured concentration profiles show a distinct change in the shape for slurries of coarser
particle size (i.e., 440 m) with higher concentrations at lower velocities (figure 8.11 and
figure 8.12). It is observed that the maximum concentration at bottom does not change and
extends up to centre of the pipeline, thus making a sudden drop in the concentration in the
upper half of the pipeline. The reason for such a distinct change in shape of concentration
profiles may be attributed to the sliding bed regime for coarser particles at lower velocities and
higher concentrations.
8.4.7.2 Velocity profile
Figure 8.15 to figure 8.19 show the corresponding vertical velocity profile across the pipe
cross section at pipe outlet. Due to unavailability of experimental data, the agreement between
experimental and predicted velocity profile could not be judged. However, velocity profile
pattern in those figures quite match the theoretical understanding. Therefore it may be
concluded indirectly that CFD model is capable of validating the velocity profile for slurry
flow. The solid phase velocity profile is generally asymmetrical about the central axis at low
velocity (say 1 m/s). The asymmetry in the solid phase velocity profile is a result of particle
settling due to the density difference between the two phases. The asymmetrical nature of
velocity profile is reduced at higher velocity range (say 3-5 m/s) and velocity profile becomes
symmetrical.
Figure 8.20 shows the comparison of velocity profile at different efflux concentration for
different flow velocity for 125 micron particles. From this figure, it can be concluded that
velocity profile does not change much due to increase in concentration from 10% to 50%.
8.4.7.3 Pressure drop
Parity plot of predicted and experimental pressure drop is shown in figure 8.21. From this
figure it is evident that agreement between calculated and experimental pressure drop are quite
242
Chapter8
well. Calculated pressure drop for slurry of 125 m particles are presented in figure 8.22 at
overall area-average concentrations around 10%, 20%, 30%, 40% and 50%. It is observed that
the pressure drop at any given flow velocity increases with increase in concentration. This
trend is seen for all concentrations at all velocities. The rate of increase in pressure with
concentration is small at low velocities but it increases rapidly at higher velocities. The
pressure drop for slurry of 440 m particles are presented in figure 8.23 at overall area-average
concentrations around 10%, 20%, 30% and 40%. From this figure, it is observed that the
pressure drop at any given flow velocity increases with increase in concentration, but the rate
of increase is comparatively smaller at higher flow velocities. Furthermore, at lower velocities,
pressure drop remains constant at lower concentrations and decreases with flow velocity at
higher concentrations. From figure. 8.22 and 8.23, it is observed that finer particle size has less
pressure drop at lower flow velocities and has more pressure drop at higher flow velocities
than coarser particles. Such an increase in pressure drop for coarser particle size at lower
velocity is due to the increased amount of particles moving in the bed due to gravitational
effect, while, in case of finer particle size at higher velocities, pressure drop is more due to
greater surface area causing more frictional losses in suspension.
8.4.7.4 Contours of solid concentration and velocity
Figure 8.24 to 8.55 show contours of volume fraction of solid and contours of solid velocity at
the pipe outlet at different flow velocities and efflux concentration. These pictures help
visualize to understand the solids distribution across pipe cross section. This is one of the
biggest advantages of CFD which helps to generate such type of concentration and velocity
contour. Figure 8.24 shows how solid was settled at the bottom of the pipe (indicated by red
colour at the bottom). Solid concentration at the top of the pipe is very low and at the top most
points is virtually absent (indicated by the blue colour at the top of the pipe). Large area at the
centre portion of the pipe, solid concentration remains uniform (indicated by green colour).
Figure 8.25 shows the velocity contours across pipe cross section at the outlet of the pipe.
From this picture it is clear that maximum velocity is not at the centre of the pipe but slightly
lower than the pipe axis. This is due to solids tending to settle at the bottom of the pipe due to
low velocity (1 m/s) of slurry and thus makes the velocity profile asymmetrical. A comparison
of velocity profile of figure 8.25 and 8.29 reveals the fact that at higher velocity (5 m/s), the
velocity profile becomes symmetrical and maximum velocity occurs at the pipe axis. This may
be due to flat concentration profile for higher turbulence at higher velocity. Again marked
243
CFD Modeling
difference of velocity profile of figure 8.25 and 8.36, shows that with the increase of solid
concentration (from 9.4% to 30.3%) at same slurry velocity (1 m/s), the asymmetrical nature
of velocity profile increases and maximum velocity location moves more towards the bottom
of the pipe.
Comparison of concentration profile contour of figure 8.28, figure 8.33, figure 8.39, figure
8.43 and figure 8.49 reveals that at a fixed flow velocity ( say 5 m/s), when solid concentration
increases from 10% to 50% , the asymmetry nature of concentration profile reduces because
of enhanced interference effect between solid particles.
Measured concentration profiles show a distinct change in the shape for slurries of coarser
particle size (i.e., 440 m) at lower velocities (Fig. 8.24 and Fig. 8.47). It is observed that most
of the solid settled down at the bottom of the pipe for coarser particle resulting in steep
concentration gradient.
244
Ch
hapter8
Fiigure 8.6A: Comparison

C
off experimentaal and calculaated vertical cconcentration profile
fo
or flow of 1225 micron gllass beads in
n 54.9 mm diameter
d
pipee at differentt efflux
co
oncentration and
a flow veloccity
245
CFD
D Modeling
Fiigure 8.6B: Comparison

C
fo
n 54.9 mm diameter
d
co
oncentration and
a flow veloccity
246
Ch
hapter8

C
fo
n 54.9 mm diameter
d
co
oncentration and
a flow veloccity
247
CFD
D Modeling

C
fo
n 54.9 mm diameter
d
co
oncentration and
a flow veloccity
248
Ch
hapter8
C
fo
n 54.9 mm diameter
d
co
oncentration and
a flow veloccity
249
CFD
D Modeling
C
fo
n 54.9 mm diameter
d
co
oncentration and
a flow veloccity
250
Ch
hapter8
Fiigure 8.9: Exp

perimental an
nd calculated vertical conccentration proofile for flow of 125
m
micron
glass beeads in 54.9 mm
m diameter p
pipe
251
CFD
D Modeling
E
and calculated vertical con
ncentration profile for flow of
o 440
Fiigure 8.10A: Experimental
m
micron
m diameter p
pipe
252
Ch
hapter8
Fiigure 8.10B: Experimental

E
and calculated
d vertical concentration proofile for flow of
o 440
m
micron
m diameter p
pipe
253
CFD
D Modeling

E
o 440
m
micron
m diameter p
pipe
254
Ch
hapter8

E
and calculated
o 440
m
micron
m diameter p
pipe
255
CFD
D Modeling

E
o 440
m
micron
m diameter p
pipe
256
Ch
hapter8
E
and calculated
o 440
m
micron
m diameter p
pipe
257
CFD
D Modeling
a)Case C
Cvf=10%
b)
b)Casee Cvf=30%
c) Casse Cvf=50%
Fiigure 8.13- Cooncentration profiles
p
in the vertical planee for slurry off 125 micron particle
p
sizze
258
Ch
hapter8
a)Casee Cvf=10%
b)Casee Cvf=30%
c) Casse Cvf=40%
Fiigure 8.14- Cooncentration profiles
p
in the vertical planee for slurry off 440 micron particle
p
sizze
259
CFD
D Modeling
C
predicted
d solid phase vvertical velociity profile for flow of 125 micron
m
Fiigure 8.15A: CFD
gllass beads in 54.9
5
mm diam
meter pipe at diifferent effluxx concentration
n and flow vellocity
260
Ch
hapter8
Fiigure 8.15B: CFD

C
predicted
d solid phase vvertical velocitty profile for flow
f
of 125 micron
5
mm diam
261
CFD
D Modeling
C
predicted
d solid phase vvertical velociity profile for flow of 125 micron
m
Fiigure 8.16A: CFD
5
mm diam
n and flow vellocity.
262
Ch
hapter8
Fiigure 8.16B: CFD

C
predicted
d solid phase vvertical velocitty profile for flow
f
of 125 micron
5
mm diam
n and flow vellocity.
263
CFD
D Modeling
C
predicted solid phase vertical veloccity profile forr flow of 125 micron
Fiigure 8.17A: CFD
5
mm diam
264
Ch
hapter8
Fiigure 8.17B: CFD

C
predicted
d solid phase vertical veloccity profile forr flow of 125 micron
5
mm diam
265
CFD
D Modeling
FD predicted solid
s
phase veertical velocity
y profile for floow of 125 miccron
Fiigure 8.18: CF
5
mm diam
meter pipe
266
Ch
hapter8
Fiigure 8.19: CF
FD predicted solid
s
phase veertical velocity
y profile for floow of 440 miccron
5
mm diam
meter pipe
267
CFD
D Modeling
mparison of vertical
v
velocitty profile at a)) 1 m/s, b) 3 m/s and c) 5 m//s for
Fiigure 8.20 Com
diifferent efflux concentration
n
268
Ch
hapter8
Fiigure 8.21: Paarity plot of prredicted Vs exxperimental prressure drop ffor slurry flow
w at
diifferent overalll area-averag
ge concentratioons and flow velocities
v
269
CFD
D Modeling
Fiigure 8.22: Prressure drop foor slurry of 1225 m particlee size at differrent overall arreaavverage concen
ntrations and flow
f
velocitiees
f slurry of 440
4 m particlle size at differrent overall arreaFiigure 8.23: Prressure drop for
avverage concen
ntrations and flow
f
velocitiess
270
Ch
hapter8
Figure 8.24: Contours

C
of vo
olume fraction
n of solid [Cvf- 9.4 %, Vm-1 m/s, d-125 miicron]
Fiigure 8.25: Coontours of soliid velocity magnitude (m/s) [Cvf- 9.4 %, Vm-1 m/s, d-1225
m
micron]
271
CFD
D Modeling
ume fraction oof solid [Cvf- 10.41

1
%, Vm-3 m/s, d-125 micron]
m
Fiigure 8.26: Coontours of volu
Fiigure 8.27: Coontours of soliid velocity magnitude (m/s) [Cvf- 10.41 %

%, Vm-3 m/s, d--125
m
micron]
272
Ch
hapter8

1
m

%, Vm-5 m/s, d--125
m
micron]
273
CFD
D Modeling

2 %, Vm-3 m
m/s, d-125 miccron]

ume fraction oof liquid [Cvf- 20.4 %, Vm-3 m/s, d-125 micron]
m
274
Ch
hapter8
m
micron]

2
m
275
CFD
D Modeling
%, Vm-5 m/s, d--125

m
micron]

3 %, Vm-1 m
m/s, d-125 miccron]
276
Ch
hapter8
m
micron]

3
m
277
CFD
D Modeling
%, Vm-3 m/s, d--125

m
micron]

3
m
278
Ch
hapter8

%, Vm-5 m/s, d--125
m
micron]

4 %, Vm-2 m
m/s], d-125 miicron]
279
CFD
D Modeling
m
micron]

3
m
280
Ch
hapter8

%, Vm-5 m/s, d--125
m
micron]

4
m
281
CFD
D Modeling
%, Vm-3 m/s, d--125

m
micron]

4
m
282
Ch
hapter8

%, Vm-4 m/s, d--125
m
micron]

4
m
283
CFD
D Modeling
%, Vm-5 m/s, d--125

m
micron]

9 %, Vm-1 m
m/s, d-440 miccron]
284
Ch
hapter8
m
micron]
F
Figure
8.53: Contours
C
of voolume fraction
n of solid [Cvf- 8.62 %, Vm-5 m/s, d-440 micron]
m
285
CFD Modeling
8.5 Conclusion
In this study the capability of CFD is explored to model complex solid liquid slurry flow in
pipeline. It is found that the commercial CFD software (FLUENT) is capable to successfully
model the solid liquid interactions in slurry flow and the predicted concentration profiles
results show reasonably good agreement with the experimental data. Following conclusions
have been drawn on the basis of present study:
1.
The particle concentration profile is modeled for high concentration slurry transport
where the maximum overall area-average concentration is 50% by volume employing
coarse particles and higher flow velocities up to 5 m/s.
2.
It is observed that the particles are asymmetrically distributed in the vertical plane
with the degree of asymmetry increasing with increase in particle size because of the
gravitational effect. It is also observed that the degree of asymmetry for the same
overall concentration of slurry increases with decreasing flow velocity.
3.
For a given velocity, increasing concentration reduces the asymmetry because of

enhanced interference effect between the solid particles. The effect of this
interference is so strong that the asymmetry even at lower velocities is very much
reduced at higher concentrations.
4.
A distinct change in the shape of concentration profiles was observed indicating the
sliding bed regime for coarser particles at lower flow velocities.
5.
The solid phase velocity profile is generally asymmetrical about the central axis at
low velocity (say 1 m/s). The asymmetry in the solid phase velocity profile is a result
of particle settling due to the density difference between the two phases. The
asymmetrical nature of velocity profile is reduced at higher velocity range (say 3-5
m/s) and velocity profile becomes symmetrical.
6.
Pressure drop at any given flow velocity increases with increase in concentration.
This trend is seen for all concentrations at all velocities. The rate of increase in
pressure with concentration is small at low velocities but it increases rapidly at higher
velocities.
The computational model and results discussed in this work would be useful for extending the
applications of CFD models for simulating large slurry pipelines.
286
Chapter8
Nomenclature
CD :
Drag function
Cfr,ls :
Coefficient of friction between the lth and sth solid-phase particles
d :
Particle dia
dl :
Diameter of the particles of solid l
els :
Co-efficient of restitution
ess :
Coefficient of restitution for particle collisions
g 0,ls:
Radial distribution coefficient
g0,ss :
Radial distribution function
Ksl :
The fluid-solid exchange coefficient
Res:
Relative Reynolds number
St :
Stokes number
um :
Mean flow velocity
Vto :
Terminal velocity
Greek symbols
Particulate loading
Material density ratio

Velocity of phase q
0;
Interphase velocity (=
Velocity of phase s
0;
Interphase velocity (=
Mass transfer from the l
th
th
to s phase.
Mass transfer from thesth to qth phase.
287
CFD Modeling
References
1.
Anderson ,T. B. and Jackson , R., A Fluid Mechanical Description of fluidized Beds.
I & EC Fundam., 6:527-534, (1967).
2.
Barigou, M., Fairhurst, P.G., Fryer, P.J., Pain, J.-P.,. Concentric flow regime of solid
liquid food suspensions: theory and experiment. Chemical Engineering Science 58,
16711686. (2003)
3.
Chapman, S. and Cowling, T. G., The Mathematical Theory of Non-Uniform Gases.

Cambridge University Press, Cambridge, England,3rd edition, (1990).
4.
Charles, M.E., Charles, R.A., Advances in SolidLiquid Flow and its Applications,
Pergamon(1971)
5.
Chhabra, R.P., Richardson, J.F., Hydraulic transport of coarse particles in viscous

Newtonian and non-Newtonian media in a horizontal pipe. Chemical Engineering
Research & Design 63, 390397. (1985)
6.
Crowe, C., Crowe, M. ,Sommer ,E., and Yutaka Tsuji. Multiphase Flows with
Droplets and Particles. CRC Press, (1998).
7.
Dalla Valle, J. M., Micromeritics, Pitman, London, (1948).
8.
Duckworth, R.A., Pullum, L., Addie, G.R., Lockyear, C.F.,. Pipeline transport of
coarse materials in a non-Newtonian carrier fluid. Hydrotransport 10, Paper C2, pp.
6988. (1986)
9.
Duckworth, R.A., Pullum, L., Lockyear, C.F., The hydraulic transport of coarse coal
at high concentration. Journal of Pipelines 3, 251265. (1983)
10. Eesa, M., Barigou, M.,. CFD investigation of the pipe transport of coarse solids in
laminar power law fluids., Chemical Engineering Science 64 , 322 333,(2009).
11. Eesa, M., Barigou, M.,. Horizontal laminar flow of coarse nearly-neutrally buoyant
particles in non-Newtonian conveying fluids: CFD and PEPT experiments compared.
International Journal of Multiphase Flow 34 (11), 9971007. (2008)
12. Ergun, Fluid Flow through Packed Columns, Chem. Eng. Prog. vol. 48, no. 2, , pp.
8994. (1952)
13. Fairhurst, P.G., Barigou, M., Fryer, P.J., Pain, J.P., Parker, D.J., Using positron
emission particle tracking (PEPT) to study nearly neutrally buoyant particles in high
solid fraction pipe flow. International Journal of Multiphase Flow 27, 18811901.
(2001)
288
Chapter8
14. Fluent Incorporated: Fluent 6.0 Documentation. Lebanon, USA, (2001)

15. Ghosh, T., Shook, C.A.,Freight Pipelines, Hemisphere. (1990)
16. Gidaspow, D., Bezburuah, R. and Ding, J., Hydrodynamics of Circulating Fluidized
Beds, Kinetic Theory Approach, Fluidization VII, Proceedings of the 7th Engineering
Foundation Conference on Fluidization, , pp. 7582. (1992)
17. Gidaspow, D., Bezburuah, R., and Ding, J., Hydrodynamics of Circulating Fluidized
Beds, Kinetic Theory Approach. In Fluidization VII, Proceedings of the 7th
Engineering Foundation Conference on Fluidization, pages 75-82, (1992).
18. Gillies , R.G., Shook , C.A. and Willson , K.C. ,An improved two layer model for
horizontal slurry pipeline flow. The Canadian Journal of Chemical Engineering, vol.
69,pp. 173178. (1991)
19. Gillies, R.G. and Shook, C.A., Modelling high concentration settling slurry flows.
The Canadian Journal of Chemical Engineering, vol. 78, pp. 709716. (2000)
20. Gillies, R.G., Hill, K.B., Mckibben, M.J. and Shook, C.A. Solids transport by laminar
Newtonian flows. Powder Technology vol. 104, ,pp. 269277. (1999)
Gradeck, M., Fagla, B.F.Z., Baravian, C., Lebouch, M.,. Technical note:
experimental thermomechanic study of Newtonian and non-Newtonian suspension
flows. International Journal of Heat and Mass Transfer 48, 34773769. (2005)
21. Guer, Y.L.E., Reghem, P., Petit, I., Stutz, B., Experimental study of a buoyant
particle dispersion in pipe flow. Transactions IChemEPart A: Chemical Engineering
Research & Design 81, 11361143. (2003.)
22. Hu, H.H., Patankar, N.A., Zhu, M.Y., Direct numerical simulations of fluidsolid
systems using the arbitrary LagrangianEulerian technique. Journal of Computational
Physics 169, 427462. (2001).
23. Ishii, M., Thermo-Fluid Dynamics Theory of Two-Phase Flow. Eyrolles, Paris.
(1975).
24. Kaushal , D.R, Sato , K. ,Toyota. T.,Funatsu. K., and Tomita. Y. Effect of particle
size distribution on pressure drop and concentration profile in pipeline flow of highly
concentrated slurry. International Journal of Multiphase Flow 31 , 809823.,(2005)
25. Kaushal , D.R, Tomita , Y. and Dighade , R.R. Concentration at the pipe bottom at
deposition velocity for transportation of commercial slurries through pipeline.
Powder Technology, vol. 125, pp.89101.(2002)
289
CFD Modeling
26. Kaushal, D.R, and Tomita , Y. Solid concentration profiles and pressure drop in
pipeline flow of multisized particulate slurries. International journal of multiphase
flow, vol. 28, , pp. 1697-1717. (2002)
27. Krampa-Morlu, F.N., Bergstrom, D.J., Bugg, J.D., Sanders, R.S., Schaan, J.,.
Numerical simulation of dense coarse particle slurry flows in a vertical pipe. In:
Proceedings of the Fifth International Conference on Multiphase Flow, ICMF'04,
Paper No. 460. (2004).
28. Kumar, U., Mishra, R., Singh, S.N., Seshadri, V.,. Effect of particle gradation on flow
characteristics of ash disposal pipelines. Powder Technol. 132, 3951. (2003)
29. Launder, B.E. and Spalding, D.B. Lectures in Mathematical Models of Turbulence,
Academic Press, London, England, (1972).
30. Legrand, A., Berthou, M., Fillaudeau, L., Characterization of solidliquid
suspensions (real, large non-spherical particles in non-Newtonian carrier fluid)
flowing in horizontal and vertical pipes. Journal of Food Engineering 78, 345355.
(2007)
31. Lun ,C. K. K., Savage, S. B., Jeffrey, D. J., and Chepurniy, N., Kinetic Theories for
Granular Flow: Inelastic Particles in Couette Flow and Slightly Inelastic Particles in a
General Flow Field. J. Fluid Mech., 140:223-256, (1984).
32. McCarthy, K.L., Kauten, R.J., Walton, J.H., Dynamics of fluid/particulate mixtures in
tube flow. Short communication. Magnetic Resonance Imaging 14, 995997. (1996).
33. Mukhtar, A. ,Investigations of the flow of multisized heterogeneous slurries in
straight pipe and pipe bends., Ph.D. Thesis, IIT, Delhi. (1991).
34. Richardson, J.R. and Zaki, W.N., Sedimentation and Fluidization: Part I., Trans. Inst.
Chem. Eng. vol. 32, , pp. 3553. (1954)
35. Roco, M.C. and Shook, C.A. Computational methods for coal slurry pipeline with
heterogeneous size distribution. Powder Technology, vol. 39, pp. 159176. (1984)
36. Roco, M.C. and Shook, C.A., Modelling of slurry flow, the effect of particle size. The
Canadian Journal of Chemical Engineering, vol. 61, , pp. 494 503. (1983)
37. Schaeffer, D. G., Instability in the Evolution Equations Describing Incompressible
Granular Flow. J. Diff. Eq., 66:19-50, (1987).
38. Seshadri, V., Malhotra, R.C. , Sundar , K.S. ,Concentration and size distribution of
solids in slurry pipeline. Proc. 11th National Conference on Fluid Mechanics and
Fluid Power, BHEL, Hyderabad, India,pp.110123. (1982)
290
Chapter8
39. Shook, C.A., Roco, M.C.,. Slurry Flow: Principles and Practice. ButterworthHeimemann, (1991).
40. Sumner, R.J., McKibben, M.J., Shook, C.A., Concentration and velocity distributions
in turbulent vertical slurry flows. Ecoulements SolideLiquide 2 (2), 3342. (1990)
41. Syamlal, M. and O ' Brient. J.,Computer Simulation of Bubbles in a Fluidized Bed,
AIChE Symp. Series, vol. 85, pp. 2231. (1989)
42. Syamlal, M., Rogers, W. and O ' Brient. J., MFIX Documentation, Volume 1, Theory
Guide, National Technical Information Service, Springfield, VA, (1993).
43. Syamlal, M., Rogers, W., and O'Brien T. J. ,MFIX Documentation: Volume 1,
Theory
Guide.
National
Technical
Information
Service,Spring_eld,
VA,
DOE/METC-9411004, NTIS/DE9400087. (1993).

44. Syamlal, M., The Particle-Particle Drag Term in a Multiparticle Model of
Fluidization.
National
Technical
Information
Service,Spring_eld,
VA,.
DOE/MC/21353-2373, NTIS/DE87006500. (1987)

45. Turton. R. and Levenspiel. O., A short note on the drag correlations for spheres,
Powder technology,47,83-86, (1986)
46. VanWachem, B.G.M., Almstedt, A.E., Methods for multiphase computational fluid
dynamics. Chemical Engineering Journal 96, 8198. (2003).
47. Vasquez, S.A. and Ivanov, V.A., A Phase Coupled Method for Solving Multiphase
Problems on Unstructured Meshes, Proceedings of ASME FEDSM'00: ASME 2000
Fluids Engineering Division Summer Meeting, Boston, (2000).
48. Wen , C.Y. and Yu , Y.H., Mechanics of Fluidization.Chem. Eng. Prog. Symp. Series,
62, pp. 100111. (1966)
291
Contribution of present thesis and future scope
CHAPTER 9
Abstract
Contribution of present thesis includes development of more accurate ANN & SVR based
models to predict pressure drop, critical velocity, hold up and to identify different flow
regimes. This thesis also modifies existing Wasp model for pressure drop and develops a CFD
based model to predict concentration and velocity profile. Future scope includes the
exploration of CFD capability in different areas of slurry flow in pipeline.
Keywords: Pressure drop, Concentration profile, Particle diffusivity, Slurry flow.
292
Chapter9
9.1 Contribution of present thesis

Many large slurry pipelines were built and operating around the world. Pipeline transport is
considered economical and environment friendly as compared to rail and road transport. To
design the pipelines and its associated facilities (pumps etc) designers need accurate
information regarding pressure drop, hold up, critical velocity, flow regimes etc at the early
design phase. Also the operating engineers need to know accurately the critical velocity so
that they can adjust the slurry flow to has a minimum pressure drop to ensure minimum
operating cost. Such flows is complex and very little known about the two-phase interaction
of solid liquid behavior inside pipeline. The correlations presently available in open literatures
for the above mentioned parameters have a prediction error of 25-35%. This much of error in
design and slurry operation have serious cost implication and are totally unacceptable in
present day competitive business scenario. This study was performed in order to develop a
model for flow of slurries through pipelines so that the error % can be reduced. This thesis can
be considered as a step forward for better understanding of flow behavior in slurry pipelines.
Experimental measurement of the local velocity and concentration distributions in slurry flow
is very difficult. Nonintrusive probes are limited, while internal probes interfere with the flow
and can not withstand the severe erosion resulting from the presence of solids. The modeling
approach offers a meaningful, relatively inexpensive, complementary alternative for
examining all useful details relating to slurry transport. The model developed in this work was
found to give results which agreed very well with available pressure drop and concentration
distribution data. The main contribution of this project is summarized as below:
1.
Development of new modeling techniques :
This thesis has developed four new hybrid modeling technique namely 1)hybrid SVR-DE
technique 2)hybrid SVR-GA technique 3)hybrid ANN-DE technique 4)hybrid ANN-GA
technique and applies them successfully in slurry flow modeling. It seems that the first three
techniques (SVR-DE, SVR-GA and ANN-DE) are completely novel in slurry flow modeling.
The project has built some quick, simple superior correlations for slurry flow which can be
readily used by design engineers to design slurry pipelines and pumps.
293
Contriibution of preseent thesis and future

f
scope
Fournewhybrid
AN
NN&SVR
basedmodelling
teechnique
Developmentof
or
CFDmodelfo
concentration
n
profile
Improve
ementof
Waspm
modelfor
pressurredrop
Autotunninggof
ANN&SV
VR
modelmetta
parameterrs
Regime
identification
errorreduce
from25%to
1.5%
Con
ntribution
Critiicalvelo
ocity
predictionerror
reducefrom25%
to6.5%
Newcorrelatio
on
forpressure
drop
New
wcorrelation
forh
holdupwith
2..5%error
Figure 9.1: Contriibution of pressent thesis

2.
Automatic tuning of mod

del parameterrs:
M
Many
ANN and
d SVR applicattions/ softwaree available in literatures/markkets are perforrmed by
eexpert users having
h
good unnderstanding oof ANN or SV
VR methodology. Non experrt users
faace difficulty to
t find an opttimum ANN oor SVR archittecture and coonfused about how to
ch
hoose the metaa parameters. The
T proposed four
f
hybrid techhniques relievee the non-expeert users
too choose the meta
m
parameterrs of SVR andd ANN algorithhm for his casse study and find
f
out
opptimum value of
o these meta parameters
p
on iits own. The beenefit is obviouus: user does not
n have
too bother about setting the arcchitecture or fixxing the meta parameters of ANN or SVR
R model.
Thhe proposed hybrid
h
techniq
que doesnt reequire user ex
xpertise and m
makes the mo
odel as
eeffortless compputation. The results indicatte that ANN orr SVR based teechnique with the GA
orr DE
294
Chapter9
based parameters tuning approach described in this work can yield excellent generalization
and reduce the processing time for finding the global minima. They can be advantageously
employed for a large class of regression problems encountered in process engineering.
3.
More accurate regime identification:
Four distinct regimes were found existent (namely sliding bed, saltation, heterogeneous
suspension and homogeneous suspension) in slurry flow in pipeline depending upon the
average velocity of flow. In the literature, few numbers of correlations has been proposed for
identification of these regimes in slurry pipelines. Regime identification is important for slurry
pipeline design as they are the prerequisite to apply different pressure drop correlation in
different regime. Designer of slurry pipelines needs to know this information at the early stage
of his design so that he can adjust velocity or pipe diameter to keep the flow regime in
heterogeneous zone. This will ensure the low pressure drop in pipeline and subsequently low
operating cost. This project has successfully developed a new methodology by applying the
support vector classification and artificial neural network model to accurately classify the
different regime in slurry transport. The project manages to reduce the prediction accuracy to
1.5 % as compared to 25 % error of Turian & Yuan correlation. With the proposed ANN and
SVM based correlation, user can easily identify the flow regime accurately (within +/- 1.5%
error range).
4.
Accurate correlations for critical velocity:
The significance of this velocity is that it represents the lowest speed at which slurry pipelines
can operate and corresponds to lowest pressure drop in slurry transport. From operating
engineers point of view, accurate knowledge of this velocity gives him a chance to optimize
the slurry flow to reduce the pressure drop in pipeline & power requirement in pump. This
project developed ANN and SVR based generalized correlations to predict critical velocity. A
comparison with selected correlations in the literature showed that the developed ANN & SVR
method noticeably improved the prediction of critical velocity (i.e. reduces the prediction error
from average 25% to 6.5%) over a wide range of operating conditions, physical properties, and
pipe diameters.
5.
Development of correlations for hold up:
The variation of in-situ concentrations from the supply concentrations is referred to as holdup
phenomenon. Though there are large number of literatures are available regarding slurry flow,
295

but most of them deal with pressure drop and critical velocity. Detailed measurements of the
hold up in slurry systems are scarce and have only been addressed in a handful of studies.
There has been no generally applicable method for the prediction of hold up of solid liquid
slurry flow. In the present study it was shown that ANN & SVR can be a superior alternative
of empirical correlations to predict hold up ratio in slurry flow. The developed ANN &SVR
correlation predicted accurately the hold up (error 2.5%) over a wide range of operating
6.
More accurate pressure drop prediction:
Accurate pressure drop prediction during design phase is key to success in slurry pipeline and
associated facilities design. Hence lot of past works were devoted to predict the pressure drop
accurately and large number of empirical and semi empirical correlations available in
literatures. Most of these equations have been developed based on limited data on materials
having equalized or narrow size-range particles and with very low concentrations. In this
work, a generalized pressure drop correlation is developed which is applicable to any slurry
system. The developed ANN or SVR based correlations reduce the prediction error % from
average 22% to 12.7% for pressure drop. Moreover the applicability range of the proposed
correlations is very large and thus makes it more versatile. The test of the model with the data
on gypsum ( median particle size 38.3 micron) demonstrates that it applies to suspensions
containing particles stretching from the boundary of colloidal behavior at the small particle
size end. At the other extreme, predictions of the model with data on slurries composed of
pebbles having a particle size of 13000 micron were also found to be excellent. Aside from
this enormous particle size range, the model was found to do an excellent job of predicting
data spanning a range of solid concentration from 1.4 % to 50% and pipe diameters from 4 cm
to 49.5 cm. The results indicate that ANN or SVR based technique with the GA or DE based
parameters tuning approach described in this work can be considered as a superior alternative
than empirical based correlations.
7.
Improvement of existing pressure drop model:
After successfully ANN & SVR based correlations were built, an attempt is made in this thesis
to develop some semi-empirical correlations for pressure drop. In this thesis, Wasp model for
pressure drop is modified by alleviating some of the restrictive assumptions used in that
model. The following three major modifications are incorporated. 1) A new variable ( modified)
296
Chapter9
is introduced in place of particle diffusivity and von Karman constant in original Wasp model.
This new modified particle diffusivity is calculated from existing concentration as oppose to
assumed constant in Wasp model. 2) Friction co-efficient of homogeneous part is modified. 3)
Hindered settling velocity is considered as oppose to free settling velocity considered in
original Wasp model.
The predicted pressure drop by modified Wasp model has been compared with the
experimental pressure drop data collected from literature. A quite well agreement was
observed and average prediction error was found far less than that by original Wasp model
(5.8% against 37.5% error by Wasp model).
Moreover, this work explores the possibility to utilize this modified Wasp model to predict
concentration profile in slurry pipelines. A new method was developed to calculate
concentration profile using Wasp model as a starting point. The predicted concentration profile
by modified model was compared with the experimental one collected from literature and
found satisfactory.
8.
Concentration and velocity profile by CFD:
In this study the capability of CFD is explored to model complex solid liquid slurry flow in
pipeline. It is found that the commercial CFD software is capable to successfully model the
solid liquid interactions in slurry flow and the predicted concentration profiles results show
reasonably good agreement with the experimental data.
Initially a two dimensional model is developed with the existing drag models available in
literature. The preliminary simulations highlighted the need for the correct modeling of the
interphase drag force. An effort was made in the present study to modify the existing models
to incorporate the effect of solid concentration at drag co-efficient. The model predictions
were compared with the published experimental data from two different sources. The predicted
results show reasonably good agreement with the experimental data. Later on, the particle
concentration profile and velocity profile are modeled in three dimensional space for high
concentration slurry transport where the maximum overall area-average concentration is 50%
by volume employing coarse particles and higher flow velocities up to 5 m/s. The pressure
drop predictions of this model are also compared with the experimental data and found in good
agreement. Several parametric studies were made to understand the effect of particle size, solid
concentration and velocity on slurry flow. The computational model and results discussed in
this work would be useful for extending the applications of CFD models for simulating large
slurry pipelines.
297

The present slurry flow model and the associated numerical code provide the pressure gradient
as well as the detailed in situ solid, liquid and mixture velocity distributions and concentration
profiles. The agreement between model prediction and experimental measurement is excellent.
The detailed information on velocity and concentration distributions is needed in assessing
pipeline erosion and wear and in extension of the analysis to include continuous heterogeneous
reaction and/or other simultaneous solids processing operations. The further extension of these
detailed calculations to particle size mixtures, which is within the capabilities of the model,
will provide the basis for optimization of the slurry transport.
9.
User friendly numerical code:
The present CFD and ANN or SVR based models including the associated boundary
conditions is complete in the sense that no experimentally determined slurry pipeline flow data
are required as input data. Input data required in the numerical code include: the solid and
liquid properties, i.e., densities, liquid viscosity, particle size and maximum packing
concentration; the system dimensions-the pipe diameter for circular pipes; and the operating
conditions i.e., the mixture discharge velocity or throughput and the delivered or discharge
concentration of solids. All remaining data, including additional model co-efficient are
incorporated within the numerical code.
9.2 Future scope
Validation with complete data set:

During model building time, the necessity for conducting own experiment and developing own
data was felt many times. Moreover, the experimental data collected from literature is not
complete in a sense that complete set of data for pressure drop, velocity profile, concentration
profile, hold up, critical velocity etc was not found for a particular system and particular
experiment. Most of the published data focused on some particular subject like
Kaushals(2002) experimental data focused on pressure drop and concentration profile only
while neglecting the velocity profile, critical velocity and hold up related issues. So it is highly
recommended to build all the required data with own experimental set up and then validate all
of these models developed in this thesis with the complete set of data.
298
Chapter9
Identifications of other flow regimes:

This thesis successfully identifies the slurry flow regimes with the classification capability of
ANN and SVM methodology. However regime identification was done for only heterogeneous
regime and moving bed regime as most of the published data was found for these two regimes
only.
Same
ANN
&
SVM
model
may
be
trained
for
classification
of
homogeneous/heterogeneous regime and moving bed/saltation regime and may be extended in

future to develop generalized correlations for identifications of above two regimes.
Correlations for other transitional velocities:

It is shown in this thesis that how ANN and SVR regression capability can be exploited to
generate correlations for critical velocity. As explained in chapter 4, there are four transitional
velocities for slurry flow in pipeline (refer figure 4.1).
In this thesis, only critical velocity (V3) is calculated as this is the most important transitional
velocity for slurry transport and most of the published data are available for this critical
velocity. Same modeling methodology may be extended in future to develop generalized
correlations for other three transitional velocities.
Exploration of CFD capabilities:

This thesis has developed a preliminary CFD model for concentration profile and pressure
drop in slurry flow in pipeline. The idea was to explore the capability of commercial CFD
software to solve such complex solid liquid slurry flow. This is just a beginning and lot of
work can be done by CFD in future to gain better insight in slurry behavior. The following
points give some of the future scope of CFD modeling in slurry flow in pipeline:
a)
CFD simulations can be extended to slurry containing mixtures of different particle sizes
and different materials.
b) CFD simulations may be extended for prediction of different regimes, hold up and critical
velocity.
c)
Parametric study of hold up, critical velocity can be done to get better understanding.
d) In future, CFD simulation should be validated with complete set of data e.g. concentration
profile, velocity profile, pressure drop, hold up, critical velocity etc.
e)
The CFD modeling capability to be extended to model slurry flow in vertical pipeline,
pipe bend and through pipe fittings and pressure drop for each case can be calculated.
299

f)
The effect of lift force, particle collision, solid liquid interaction, solid solid interaction
and drag force on concentration profile may be studied in detail in future. Currently the
knowledge in this area is very limited.
g) Effect of particle shape, particle surface roughness on pressure drop and concentration
profile may be explored in future to gain better insight. Presently, the publications in this
area are very scarce.
Dependence of particle diffusion co-efficient:

As discussed earlier this thesis has modified the Wasp model to make it more accurate. One
term for particle diffusion coefficient,
modified
was introduced in this modification which is a
function of concentration only (refer equation 7.14). It is recommended for further study to
find out the dependence of particle diffusion coefficient (
modified)
on particle size, particle
shape and other fluidparticle relationships besides efflux concentration of slurry.
Explorations of different strategies:

Several strategies have been proposed in literature for DE implementation. Babu (1980)
proposed almost 10 different strategies and summarized their comparative advantage. In the
present thesis only one strategy for DE was used. All other strategies may be explored to
reduce the computational time or to increase accuracy.
References
1.
Babu B.V. and Munawar. S.A., Optimal design of shell-and-tube heat exchangers using
different strategies of differential evolution, preJournal.com-the faculty lounge, article
number 003873, website: http://www.prejournal.com, march, (2001)
2.
Kaushal, D.R, Tomita, Y,.Solid concentration profiles and pressure drop in pipeline flow
of multisized particulate slurries. International journal of multiphase flow 28,16971717(2002)
300

Study On Slurry Flow Modelling in Pipeline (Eth) (LAHIRI, Sandir Kumar) (Nat. Inst. of Techn. Durgapur 2009) (329s)

Cargado por

Información del documento

Título original

Derechos de autor

Formatos disponibles

Compartir este documento

Compartir o incrustar documentos

Opciones para compartir

¿Le pareció útil este documento?

¿Este contenido es inapropiado?

Copyright:

Formatos disponibles

Study On Slurry Flow Modelling in Pipeline (Eth) (LAHIRI, Sandir Kumar) (Nat. Inst. of Techn. Durgapur 2009) (329s)

Cargado por

Copyright:

Formatos disponibles

From the SelectedWorks of Dr.

My PhD thesis : Study on slurry flow modelling in

Start Your Own

Sandip Kumar Lahiri

S ANDIP K UMAR L AHIRI

NATIONAL INSTITUTE OF TECHNOLOGY

(Dr. K.C. Ghanta)

Sandip Kumar Lahiri

Paper published based on present thesis

Lahiri, Sandip K. and Ghanta, Kartik Chandra ,Development of a Hybrid Artificial

Paper under review

1.2 Typical flow regime

1.3 Critical velocity

1.5 Pressure drop

1.6 Concentration and velocity profile

1.7 Motivation of present work

1.8 Scope of present work

1.8.2 Semi empirical modeling

1.8.3 Multiphase computational fluid dynamics (CFD) based

2.2 Artificial neural network

2.2.1 Background works

2.2.2 Overview of classification neural networks

2.2.3 Overview of prediction neural networks

2.2.4 Strengths of ANN

2.2.5 Limitations of neural networks

2.3 Comparison of neural networks to empirical modeling

2.4 Artificial neural network (ANN) based modeling

2.4.1 Network Architecture

2.4.3 Back propagation algorithm (BPA)

2.4.3.1 Different back propagation algorithm

2.4.4 Performance measures of ANN model

2.4.6 Step-wise Procedure for Developing an Optimal MLP Model

2.5 Tuning parameters of ANN

2.6 Optimization of ANN model

2.7 Genetic algorithm

2.7.1 Literature survey of genetic algorithm

2.7.2 Techniques used in GA

2.8 GA-based Optimization of ANN Models

2.9 Differential evolution

2.9.1 Literature survey of Differential evolution

2.9.2 Steps performed in DE

2.10 DE-based Optimization of ANN Models

2.11 Support Vector Machines

2.12 Background works

2.13 The basic idea behind SVM modeling

2.13.1 Support vector machine for classifications

2.13.2 Mathematics behind SVM algorithm for classification

2.13.3 Training, testing and Generalizability

2.13.4 Mathematics behind SVM algorithm for regression

2.14 Tuning parameters of SVR

2.15 Optimization of SVM model

2.16 GA-Based Optimization of SVR Models

2.17 DE based optimization of SVR model

3. Regime identification of slurry transport in pipelines

3.2 Background work

3.2.1 Flow regimes in slurry flow

3.2.2 Head loss correlations for separate flow regimes

3.2.3 Flow regime boundaries (Turian and Yuans approach)

3.3 Performance check of Turian Yuans approach

3.3.1 Data Collection

3.3.2 Regime identification

3.4 Scope of present work

3.5. Development of the artificial neural network (ANN) based correlation