Study on Slurry Flow Modelling in Pipeline (Eth) [LAHIRI, Sandir Kumar] [Nat. Inst. of Techn. Durgapur; 2009] {329s}

From the SelectedWorks of Dr. Sandip KumarLahiri

March 2010

My PhD thesis : Study on slurry flow modelling inpipeline

ContactAuthor

Start Your OwnSelectedWorks

Notify Meof New Work

Available at: http://works.bepress.com/sandip_lahiri/23

i

SSTTUUDDYYOONNSSLLUURRRRYYFFLLOOWWMMOODDEELLIINNGGIINNPPIIPPEELLIINNEE

SSaannddiipp KKuummaarr LLaahhiirrii

ii

SSTTUUDDYYOONNSSLLUURRRRYYFFLLOOWWMMOODDEELLIINNGGIINNPPIIPPEELLIINNEE

THESIS

Submittedinpartialfulfillmentofthe

Requirementsforthedegreeof

DOCTOR OF PHILOSOPHY

By

SANDIP KUMAR LAHIRI

UndertheSupervisionof

DR. K.C.GHANTA

Professor,DepartmentofChemicalEngineering

NATIONALINSTITUTEOFTECHNOLOGY,DURGAPUR

DEPARTMENTOFCHEMICALENGINEERING

DURGAPUR713209,INDIA

2009

iii

DEDICATED

TO

MY PARENTS

&

MY WIFE

iv

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

v

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

vi

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)

vii

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)

viii

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)

ix

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)

x

List of Contents Contents Page

1. Introduction 1.1 Introduction 2

1.2 Typical flow regime 5

1.3 Critical velocity 5

1.4 Hold up 6

1.5 Pressure drop 6

1.6 Concentration and velocity profile 6

1.7 Motivation of present work 7

1.8 Scope of present work

1.8.1 Advanced numerical modeling 9

1.8.2 Semi empirical modeling 9

1.8.3 Multiphase computational fluid dynamics (CFD) based 9

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

xi

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

xii

3.3.1 Data Collection 91

3.3.2 Regime identification 92

3.4 Scope of present work 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

3.6.1. Input selection and data collection 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


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


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


5.3. Development of the artificial neural network (ANN) and support

vector regression (SVR) based correlation 141

xiii



5.4. Results and discussion 144


5.4.2 Comparison of hybrid ANN-DE model with ANN model 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


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



6.4 Results and discussion 169


6.4.2 Prediction performance of hybrid ANN-GA model 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

xiv

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


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 221

8.3.1 Eulerian Model 222

8.3.1.1 Continuity Equation 222

8.3.1.2 Momentum Equations 223

8.3.1.3 Fluid-solid momentum equations 223

xv

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 231

8.4 Description of CFD simulation 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

xvi

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) 99

Table 3.6: 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 4.3: Typical input and output data for ANN or SVR training 122



correlations(i=7,j=8,k=1) 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 5.2: Typical input and output data for ANN & SVR training 143

Table 5.3: Prediction error by the hybrid ANN-DE based model 144


correlations (i=7, j=8,k=1) 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

xvii

Table 5.8: Comparison of performance of SVR-DE hybrid model Vs SVR

model 148


Table 6.2: Typical input and output data for ANN training 170



correlations (i=7, j=8, k=1) 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

xviii

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 51

Figure 2.10: 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. 84

Figure 3.2: 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 109

Figure 4.3: 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

algorithm 124

Figure 4.6: Variation of critical velocity with density ratio 129

xix

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 155

Fig 6.3: Experimental Vs predicted pressure drop by Marquard Levenburg

algorithm 172

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

based on typical Roco & Shook (1984) data 191

Figure 7.2: 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 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 199



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 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

xx



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 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 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. 239

Figure 8.4: 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. 240

Figure 8.5: 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. 241

Figure 8.6A: 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 245

Figure 8.6B: Comparison of experimental and calculated vertical concentration



xxi













Figure 8.9: Experimental and calculated vertical concentration profile for flow

of 125 micron glass beads in 54.9 mm diameter pipe 251

Figure 8.10A: Experimental and calculated vertical concentration profile for flow


Figure 8.10B: Experimental and calculated vertical concentration profile for flow










Figure 8.13: Concentration profiles in the vertical plane for slurry of 125 micron

particle size 258

Figure 8.14: Concentration profiles in the vertical plane for slurry of 440 micron

particle size 259

Figure 8.15A: 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

xxii

Figure 8.15B: CFD predicted solid phase vertical velocity profile for flow of 125





concentration and flow velocity. 262



concentration and flow velocity. 263







Figure 8.18: CFD predicted solid phase vertical velocity profile for flow of 125

micron glass beads in 54.9 mm diameter pipe 266

Figure 8.19: CFD predicted solid phase vertical velocity profile for flow of 440

micron glass beads in 54.9 mm diameter pipe 267

Figure 8.20: Comparison of vertical velocity profile at a) 1 m/s, b) 3 m/s and

c) 5 m/s for different efflux concentration 268

Figure 8.21: Parity plot of predicted Vs experimental pressure drop for slurry flow

at different overall area-average concentrations and flow velocities 269

Figure 8.22: Pressure drop for slurry of 125 m particle size at different overall

area-average concentrations and flow velocities 270

Figure 8.23: Pressure drop for slurry of 440 m particle size at different overall

area-verage concentrations and flow velocities 270

Figure 8.24: Contours of volume fraction of solid [Cvf- 9.4 %, Vm-1 m/s, d-125

micron] 271

Figure 8.25: Contours of solid velocity magnitude (m/s) [Cvf- 9.4 %, Vm-1 m/s,

d-125 micron] 271

Figure 8.26: Contours of volume fraction of solid [Cvf- 10.41 %, Vm-3 m/s,

d-125 micron] 272

xxiii

Figure 8.27: Contours of solid velocity magnitude (m/s) [Cvf- 10.41 %,

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


d-125 micron] 273




d-125 micron] 274

Figure 8.31: Contours of volume fraction of liquid [Cvf- 20.4 %, Vm-3 m/s,

d-125 micron] 274



Figure 8.33: Contours of volume fraction of solid [Cvf- 20.45 %,

















Vm-2 m/s] , d-125 micron] 279





xxiv




















d-440 micron] 285

Figure 9.1: Contribution of present thesis 294

xxv

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.

xxvi

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.

Chapter1

1

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

2

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

Chapter1

3

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

Introduction

4

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

5

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

6

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,

Chapter1

7

(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.

Introduction

8

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

9

1.8 Scope of present work

The problem considered in this thesis is the formulation of a theory for the flow of dense, non-

colloidal, 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

Introduction

10

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 ANN-

DE 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.

Chapter1

11

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).

Mathematical Tools

12

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).

Chapter2

13

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.

Mathematical Tools

14

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 input-

multiple 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:

Fr

ap

Co

to

Be

pr

2.

A

en

rom literature

pplications to

ompared to em

o noise and inco

ecause of the f

rediction, the a

.2.2 Overview

Applications of

ngineering fall

Classi

Pred

Assoc

Filt

Optim

Concept

Figu

survey it is ev

chemical engi

mpirical and cu

omplete inform

fact that in this

applications in t

w of classific

f neural netwo

l into two m

fication

diction

ciation

ering

mization

tualization

Ch

ure 2.1: Differ

vident that neu

ineering probl

urve-fitting mo

mation and can

s work we are

these two area

cation neural

orks to classifi

ajor areas: (1

useinputvandlabora

useinputvtemperaturate.

learnassocdatathatcrecognizen

smoothan

determinetravellings

Analyzedawithmanyn

hapter2

ent applicatio

ural networks

lems that are

odels, neural n

thus deal with

going to use n

s are explored

l networks

ication problem

) identificatio

valuestopredictatoryresults,deter

valuestopredictare,humidityandw

ciationsoferrorfrontainserror;e.g.noisyinputpattern

inputsignale.g.,s

optimalvaluee.gsalesperson.

taanddetermineattributessothat

on of ANN

hold much pr

complex, non

networks are re

h problems with

neural network

in detail from

ms in bio proc

on of process

categoricaloutpurminethemostlik

noutputvalues;ewindvelocity,pred

reeoridealdata,t.,learnfiveidealpnsasoneoffivep

smoothanoisyele

.,determineminim

conceptualrelatiotgroupingrelation

romise for sig

nlinear and un

elatively less s

h noisy data.

ks in classificat

literatures.

cessing and ch

faults based

ut;e.g.,givensympkelydisease.

e.g.,givendicttheevaporatio

henclassifyassocpatternsandthenatterns.

ectocardiosignal

mumlengthtripo

onshipe.g.,clustenshipscanbeinfer

15

nificant

ncertain.

ensitive

tion and

hemical

on the

ptoms

on

iate

fa

rdatarred.

Mathematical Tools

16

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-bias-

function 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)

Chapter2

17

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

Mathematical Tools

18

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.

2. W

of

.2.5 Limitatio

While neural ne

f neural networ

Long trimpract

network

technol

Large process

such da

but all

small t

essentia

f

p

Figure 2

ons of neural

etworks have m

rks as follows:

training times:

tical. Most sim

k and comple

logy of powerf

amount of tra

s, we may rec

ata. In addition

the training d

training data s

al.

Potentialforonline

use

mimicshumanlearningprocesses

Ch

2.2: Strength a

l networks

many advantage

Training can

mple problem

ex problems c

ful PC, this lim

aining data: I

onsider the us

n, we may also

data are very s

sets. Thus, a b

AN

abilitylearn

processnoisyd

hapter2

and character

es, they are no

n take long en

ms require at l

can require up

mitations getting

f little input-o

se of neural ne

o have a situat

similar, causin

broad based

N

yton

Aa

singdata

ristics of ANN

ot a cure-all. So

nough to make

least 1000 tim

pto 75000. Ho

g wiped out da

output data ex

etworks, since

tion where ther

ng the same pr

data set or ex

extensiveknowledgeindexing

Automatedabstraction

ome of the lim

e the neural n

me steps to tr

owever with th

ay by day.

xist on a prob

they rely hea

re is a large da

roblems as in

xperimental de

19

mitations

network

rain the

he new

blem or

avily on

atabase,

having

esign is

Mathematical Tools

20

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 input-

Chapter2

21

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,

Documents

Study on Slurry Flow Modelling in Pipeline (Eth) [LAHIRI, Sandir Kumar] [Nat. Inst. of Techn. Durgapur; 2009] {329s}