Final PHD Thesis of Nirjhar Bar 2016

APPLICATIONS OF NEURAL
NETWORKS FOR THE PREDICTION

OF HYDRODYNAMIC PARAMETERS IN
LIQUID AND GAS-LIQUID FLOW
THROUGH PIPELINES
Thesis submitted to
MAULANA ABUL KALAM AZAD UNIVERSITY OF
TECHNOLOGY
(Formerly known as West Bengal University of Technology)
In partial fulfillment of the Degree of
DOCTOR OF PHILOSOPHY
2016
Nirjhar Bar
DEPARTMENT OF PHYSICS
MAULANA ABUL KALAM AZAD UNIVERSITY OF TECHNOLOGY
BF 142, SECTOR 1, SALT LAKE CITY
KOLKATA - 700 064
WEST BENGAL
INDIA
Thesis Certificate
This is to certify that the thesis entitled “Applications of Neural
Networks for the Prediction of Hydrodynamic Parameters in Liquid and
Gas-liquid Flow through Pipelines”, submitted by Nirjhar Bar to the
Maulana Abul Kalam Azad University of Technology, West Bengal
(Formerly known as West Bengal University of Technology) for the
award of the degree of Doctor of Philosophy in Science (Physics), is a
bona fide record of the research work done by him under our supervision
in the duration of five years from July 2008 to July 2013. He has
completed the work truthfully and successfully to the best of our
knowledge. The contents of this thesis, in full or in part, have not been
submitted to any other Institute or University for the award of any degree
or diploma.
Prof. Manindra Nath Biswas Prof. Sudip Kumar Das

(Supervisor) (Co-Supervisor)
Former Professor of Chemical Engineering Professor
IIT Kharagpur and Professor (Acting) Department of Chemical Engineering
Government College of Engineering and Leather Technology University of Calcutta
Kolkata
ACKNOWLEDGEMENTS
It is with real pleasure that, I record my indebtedness to my Academic
Supervisors, Prof. Manindra Nath Biswas and Co-Supervisor Prof. Sudip Kumar Das
for their counsel and guidance during the preparation of this thesis.
I am grateful to Dr. Tarun Kanti Bandyopadhyay, Dr. Asit Baran Biswas, Mr.
Kaushik Ganguly, Dr. Bimal Das, Dr. Gargi Bhattacharya, Mr. Biswajit Singha for
their continuous encouragements and support for the preparation of this thesis, sharing
their knowledge and valuable experience with me irrespective of their busy schedule.
I wish to record my sincere thanks to God Almighty, Mr. Asish Kumar Mitra,
Mrs. Sumita Mitra, Mr. Kuntal Khanra, my parents, wife, sister, other relatives and
those whose name I have forgotten to mention here.
I also wish to thank Prof. A C Gomes, Rev. Dr. Bala Sundar Singh, Dr. Geetha
Singh, and my friends who had given me a lot of encouragements over the years.
My thanks are due to their good wishes and blessings, without which it would
have been a very difficult task for me to go ahead. I pray to God for His blessings and
mercies.
Place: Kolkata
Date: 10th November, 2016 Nirjhar Bar
LIST OF PUBLICATIONS
Paper published/communicated in International Journals and Conference

Proceedings
1) Bar, N., Bandyopadhyay, T.K., Das, S.K. and Biswas, M.N. 2010. Prediction
of pressure drop using artificial neural network for non-Newtonian liquid flow
through piping components, Journal of Petroleum Science and Engineering.
71 187 – 194.( ISSN: 0920 – 4105)
2) Bar, N., Das, S.K. and Biswas, M.N. 2010. Prediction of Pressure Drop Using
Artificial neural Network for Gas non-Newtonian Liquid Flow through Piping
Components, Industrial and Engineering Chemistry Research. 49 9423 –
9429.(ISSN: 0888 – 5885)
3) Bar, N., Biswas, A.B., Das, S.K. and Biswas, M.N. 2011. Gas-non-Newtonian
Liquid Flow through Horizontally Oriented Helical Coils – Prediction of
Frictional Pressure Drop Using ANN, Artificial Intelligent Systems and
Machine Learning 3(7) 412 – 418. (ISSN: 0974 – 9543)
4) Bar, N., Ghosh, T.K., Das, S.K. and Biswas, M.N. 2011. Air-Water Flow
through 3 mm and 4 mm Tubes – Experiment and ANN Prediction, Artificial
Intelligent Systems and Machine Learning 3(8) 531 – 537. (ISSN: 0974 –
9543)
5) Bar, N., Biswas, A.B., Das, S.K. and Biswas, M.N. Holdup Analysis for Gas-
non-Newtonian Liquid Flow through Horizontal Helical Coils – Empirical
Correlation versus ANN Prediction. AIP Conference Proceedings # 1298;
International Conference on Modeling, Optimization and Computing,
(ICMOC 2010), Eds. American Institute of Physics, New York, pp 104 – 109,
2010. (ISSN: 978 – 0 – 7354 – 0854 – 8)
6) Bar, N., Ghosh, T.K., Das, S.K. and Biswas, M.N. Prediction of Flow Pattern
in Small Diameter Tubes Using Artificial Neural Network. Ciit Conference
Proc. Catalog Number: 11 – 5 – 0006. International Conference on Advances
in Engineering and Technology, (ICAET 2011), May 27 – 28, 2011. (ISBN:
978 – 4507 – 6433 – 9) – AWARDED BEST PAPER
7) Bar, N., Das, S.K. and Biswas, M.N. Frictional Pressure Drop Using ANN for
Gas-Non-Newtonian Liquid Flow through 45° Bend. Ciit Conference Proc.
Catalog Number: 11 – 5 – 0006. International Conference on Advances in
Engineering and Technology, (ICAET 2011), May 27 – 28, 2011. (ISBN: 978
– 4507 – 6433 – 9) – AWARDED BEST PAPER
8) Bar, N., Das, S.K. and Biswas, M.N. 2013. Prediction of Frictional Pressure
Drop using Artificial Neural Network for Air-water Flow through U-Bend.
Procedia Technology 10 813 – 821. (ISSN: 2212 – 0173)
9) Bar, N., Biswas, M.N. and Das, S.K. Flow Regime Prediction using Artificial
Neural Networks for Air-Water Flow through 1–5 mm Tubes in Horizontal
Plane. Second International Conference on INformation systems Design and
Intelligent Applications – 2015, (INDIA 2015), Advances in Intelligent and
Soft Computing, Vol 339, Mandal et al(Eds): Kalyani, India, January 8 – 9,
2015. (ISBN: 978 – 81 – 322 – 2250 – 7)
Mr. Nirjhar Bar is the member of IEEE and INNS.
i
PREFACE
The thesis entitled Applications of neural networks for the prediction of
hydrodynamic parameters in liquid and gas-liquid flow through pipelines is
submitted herewith. The present studies deal with the modeling of hydrodynamics of
liquid and gas-liquid flow through pipes and piping components, bend, helical coils
and flow regime prediction using artificial neural network (ANN). The data for this
analysis is acquired from literature and thesis from earlier research works carried out
in our laboratory. The thesis clearly demonstrates the predictive ability of ANN in
hydrodynamics of single and multi-phase flow system.

ii
CONTENTS
Page no.
Acknowledgement
List of publication
Preface i
Contents ii
List of Figures v
List of Tables ix
Summary xii
Chapter - 1 Introduction
1 Introduction 1
Chapter - 2 Artificial neural network methodology
2.1 Introduction 14
2.2 Artificial Neural Network Architecture 15
2.2.1 Input layer 16
2.2.2 Hidden layer 16
2.2.3 Output Layer 17
2.3 Different algorithms used for training 17
2.3.1 Multilayer Perceptron 17
2.3.1.1 Backpropagation algorithm 20
2.3.1.2 Levenberg-Marquardt algorithm 22
2.3.2 Radial Basis Function (RBF) 23
2.3.3 Principal Component Analysis (PCA) 25
2.3.4 Support Vector Machine (SVM) 25
2.4 Optimization of the ANN 26
2.5 Performance of the ANN 28
Chapter - 3 Non-Newtonian liquid flow through piping components in
horizontal plane
3.1 Introduction 33
3.2 Artificial neural network architecture 34
3.4 Results and discussions 35
3.5 Conclusions 36
iii
Chapter - 4 Gas-non-Newtonian liquid flow through pipes and piping

components in horizontal plane
4.1 Introduction 51
4.5 Conclusion 55
Chapter - 5 Gas-non-Newtonian liquid flow through 45° bend
in horizontal plane
5.1 Introduction 81
5.5 Conclusion 84
Chapter - 6 Gas-non-Newtonian liquid flow through helical coils
in horizontal plane
6.1 Introduction 91
6.5 Conclusion 95
Chapter - 7 Prediction of flow regime for air-water flow through
small diameter tube
7.1 Introduction 110

7.2 Data Collection 114
7.3 Results and performance 117
7.3.1 Backpropagation algorithm 118
7.3.2 Levenberg-Marquardt algorithm 119
7.3.3 Radial Basis Function (RBF) 120
7.3.4 Principal Component Analysis (PCA) 120
7.3.5 Support Vector Machine (SVM) 121
7.4 Comparison of ANN Prediction with experimental data 122
7.5 Conclusion 122
iv
Chapter - 8 Conclusions 147

Nomenclature 149
References 151
Appendix 165
List of figures
LIST OF FIGURES
Figure No. Titles
2.1 Schematic diagram of artificial neural network
2.2 Flowchart representing training of the single hidden layer

backpropagation network
3.1–3.4 Variation of the minimum value of MSE with the number of nodes in
the hidden layer
3.5–3.8 Variation of the average MSE for training over 5 different runs vs the
number of epochs
3.9–3.12 Variation of the average MSE for cross-validation over 5 different runs
vs the number of epochs
3.13 Comparison of pressure drop across the elbow for testing
3.14 Comparison of pressure drop across the orifice for testing
3.15 Comparison of pressure drop across the gate valve for testing
3.16 Comparison of pressure drop across the globe valve for testing
3.17 Comparison of pressure drop across the elbow for prediction
3.18 Comparison of pressure drop across the orifice for prediction
3.19 Comparison of pressure drop across the gate valve for prediction
3.20 Comparison of pressure drop across the globe valve for prediction
4.1–4.5 Variation of the minimum value of MSE with the number of

processing elements in the hidden layer for four different transfer
functions
4.6–4.10 Variation of the average MSE for training over 5 different runs vs the
number of epochs in case for four different transfer functions in the
hidden layer
4.11–4.15 Variation of the average MSE for cross-validation over 5 different runs
vs the number of epochs for four different transfer functions in the
hidden layer
4.16 Comparison of two-phase pressure drop across the horizontal pipe for
testing for four different transfer functions in the hidden layer
vi
List of figures
4.17 Comparison of two-phase pressure drop across the elbow for testing
using four different transfer functions in the hidden layer
4.18 Comparison of two-phase pressure drop across the orifice for testing
4.19 Comparison of two-phase pressure drop across the gate valve for
testing using four different transfer functions in the hidden layer
4.20 Comparison of two-phase pressure drop across the globe valve for
testing using four different transfer functions in the hidden layer
4.21 Comparison of two-phase pressure drop across the horizontal pipe for
prediction using four different transfer functions in the hidden layer
4.22 Comparison of two-phase pressure drop across the elbow for

4.23 Comparison of two-phase pressure drop across the orifice for

4.24 Comparison of two-phase pressure drop across the gate valve for
4.25 Comparison of two-phase pressure drop across the globe valve for
5.1 Variation of the minimum value of MSE with the number of

processing elements in the hidden layer for four different transfer
functions
5.2 Variation in the average MSE for training over 5 different runs vs the
number of epochs
5.3 Variation in the average MSE for cross-validation over 5 different runs
vs the number of epochs
5.4 Comparison of the two-phase frictional pressure drop
5.5 Comparison of the two-phase frictional pressure drop for final

prediction

processing elements in the hidden layer for two-phase frictional
pressure drop across the coil

processing elements in the hidden layer for gas holdup in the coil
vii
List of figures
6.3 Variation of the average MSE for training over 5 different runs vs the
number of epochs in case of two-phase frictional pressure drop across
the coil
6.4 Variation of the average MSE for training over 5 different runs vs the
number of epochs in case of gas holdup in the coil
6.5 Variation of the average MSE for cross-validation over 5 different runs
vs the number of epochs in case of two-phase frictional pressure drop
across the coil
6.6 Variation of the average MSE for cross-validation over 5 different runs
vs the number of epochs in case of gas holdup in the coil
6.7 Comparison of two-phase frictional pressure drop across the horizontal

coil for testing with four different transfer functions
6.8 Comparison of the gas holdup in the horizontal coil for testing with
four different transfer functions
6.9 Comparison of two-phase frictional pressure drop across the horizontal

coil for final prediction
6.1 Comparison of gas holdup in the horizontal coil for final prediction
7.1 Graphical representation of data for sample 1
7.6 Variation of minimum value of cross-validation with the number of

processing elements in the hidden layer for MLP network trained with
backpropagation algorithm for all 5 samples
7.7 Cross-validation curve for MLP network trained with BP algorithm for
all 5 samples

Levenberg-Marquardt algorithm for all 5 samples
7.9 Cross-validation curve for MLP network trained with LM algorithm

for all 5 samples
viii
List of figures

processing elements in the hidden layer for RBF network trained with
Levenberg-Marquardt algorithm for all 5 samples
7.11 Cross-validation curve for RBF network trained with LM algorithm for
all 5 samples

processing elements in the hidden layer for PCA network trained with
Levenberg-Marquardt algorithm for all 5 Samples
7.13 Cross-validation curve for PCA network trained with LM algorithm for
all 5 samples
7.14 Cross-validation curve for SVM network for all 5 samples

List of Tables
LIST OF TABLES
Table No. Title
3.1 Description of the data used for neural network analysis
3.2 Range of different data sets for all four different systems
3.3 Optimum number of nodes for different systems
3.4 Performance of neural network for training for 5 runs
3.5 Performance of neural network for cross-validation for 5 runs
3.6 Performance of neural network for testing
3.7 Performance of the neural network for prediction
3.8 Comparison between experimental and ANN prediction for elbow
3.9 Comparison between experimental and ANN prediction for orifice
3.10 Comparison between experimental and ANN prediction for gate valve
3.11 Comparison between experimental and ANN prediction for globe

valve
4.1 Description of the data used for neural network analysis
4.2 Range of different data sets for all five different systems
4.3 Performance of best neural network for training over 5 runs
4.4 Performance of the neural network for cross-validation over 5 runs
4.5 Performance of the best neural network for testing
4.6 Performance of the best neural network for prediction
4.7 Comparison between the experimental and ANN prediction in case of

horizontal pipe for transfer function 3 with 25 processing elements in
the hidden layer

elbow for transfer function 1 with 14 processing elements in the hidden
layer

orifice for transfer function 1 with 15 processing elements in the
hidden layer
x
List of tables

gate valve for transfer function 4 with 13 processing elements in the
hidden layer
4.11 Comparison between the experimental and ANN prediction for globe
valve for transfer function 2 with 25 processing elements in the hidden
layer
5.1 Description of data used for neural network analysis
5.2 Optimum number of processing elements in hidden layer for four

different transfer functions
5.3 Performance of neural network with different transfer functions for

training and cross–validation over 5 runs
5.4 Performance of neural network with four different transfer functions

for testing using optimum number of processing elements

for prediction using optimum number of processing elements
5.6 Comparison of the ANN prediction using the best network and the
experimental data
6.1 Range of different variables investigated for horizontal helical coils
6.2 Optimum number of processing elements used in hidden layer for four

for training and cross-validation over 5 runs for two-phase frictional
pressure drop

for training and cross-validation over 5 runs for gas holdup

two-phase frictional pressure drop during testing

two-phase frictional pressure drop during testing
6.7 Performance of neural network with different transfer functions for the
prediction of two-phase frictional pressure drop across the helical coil
6.8 Performance of neural network with different transfer functions for the
prediction of gas holdup in the helical coil
xi
List of tables
6.9 Comparison of the prediction using the best network and the
experimental data for pressure drop per unit length
6.10 Comparison of the prediction using the best network and the
experimental data for gas holdup
6.11 Comparison of results for the prediction of gas holdup using best
neural network and output calculated using Eq. (6.1)
7.1 Description of data collected from literature for Air-Water flow
7.2 Minimum value of MSE for cross-validation reached during training

for all 5 different ANNs
7.3 Optimum number of processing elements in the hidden layer during

training for 4 different ANNs
7.4 Comparison of ANN predictions with experimental flow regime data

with 5 different ANN used
7.5 Comparison of ANN predictions with experimental flow regime data

for sample 3 using PCA network
Summary
SUMMARY
The single and two-phase flow through piping components, like elbows,
orifice, gate and globe valves, bend, helical coils is more complex than the straight
pipes. The problems to determine the pressure drop is important in designing and
analyzing the machines used for hydrodynamics. The friction between the fluid and
the wall causes this pressure drop. The problem of predicting pressure drop in these
systems is much more uncertain than for the straight pipe because the mechanism of
flow is not clearly defined. At least three types of losses are possible, i.e., skin
friction, loss due to change in flow direction and the constriction of the flow path.
Artificial Neural Networks (ANNs) are computational modeling tools received
extensive acceptance in many disciplines for modeling complex and real-world
problems. It defines as structures comprised of densely interconnected adaptive
simple processing elements, called nodes or neurons that are capable of performing
massive parallel computations for data processing and knowledge representation. The
attractiveness of ANNs comes from the remarkable information processing
characteristics; the system may be nonlinear, highly parallelism, robust, faulty and
failure tolerant, learning, ability to handle large information and their capability to
generalize. The objective of ANN based computing is to develop mathematical
algorithms that will enable ANNs to learn from past information and predict present
information. The ANN models are empirical in nature but it can provide practically
accurate solutions for phenomena that are only understood through experimental data
or field observation.
The design and operation of two-phase flow system one needs to know the
hydrodynamic parameters, i.e., pressure drop, holdup and flow regime. The
xiii
Summary
knowledge of flow pattern is essential for the prediction of pressure drop, holdup and
thermodynamic quantities like heat and mass transfer coefficients (Bell, 1988 and
Spedding and Spence, 1988). The complexity of the two-phase flow system and also
the lack of basic understanding of the underling physics of the problem, the majority
of the two-phase pressure drop and holdup analysis are suggested by using either
empirical or semi-empirical correlation.
With the developments in polymer processing, mineral recovery, food
processing, biomedical engineering, biochemical engineering and waste water
treatment, situation increasingly arises where the liquid characteristics are non-
Newtonian in nature. Hence, there is a need to have a detailed study of hydrodynamic
parameters (pressure drop, holdup, flow regimes etc.) for non-Newtonian liquid and
gas-non-Newtonian liquid flow through various pipes and piping components.
The present study aims to develop a method of predicting the hydrodynamic
parameters using artificial neural networks.
The thesis is divided into eight chapters as,
Chapter - 1: This chapter deals with the importance of liquid and gas-liquid flow
through various pipes and piping components and its usefulness in the various
industries. The research work performed by various researchers is discussed
chronologically. Details literature review is mentioned mainly focusing on the
hydrodynamics of single and two-phase flow. This chapter also gives a brief historical
review of the studies related to artificial neural networks and its development over the
years. The objective of the present study is also mentioned here.
Chapter - 2: It deals with the details description and the various techniques related
to the ANNs used for this study. Two types of Multilayer Perceptrons (MLP), i.e.,
Backpropagation (BP) algorithm and Levenberg Marquardt (LM) algorithm

xiv
Summary
respectively, Radial Basis Function (RBF), Principal Component Analysis (PCA) and
Support Vector Machine (SVM) and their optimization techniques are discussed. The
statistical parameters, e.g. Mean Squared Error (MSE), Average Absolute Relative
Error (AARE), Standard Deviation (σ), Cross-correlation coefficient (R), Chi-square
test (χ²) are used for the evaluation of the performance of the ANN analysis. The
methodology of the above mentioned ANNs and the training, cross-validation, testing
and the final prediction related to these ANNs are described along with the stopping
criterions during training, the values of the various constant parameters ( α , µ , λ etc.)
are also mentioned.
Chapter - 3: This chapter deals with the prediction of pressure drop for non-
Newtonian liquid flow through piping components, namely elbow, orifice, gate valve
and globe valve. The experimental data used for the prediction is collected from the
earlier work carried out in our laboratory (Bandyopadhyay and Das, 2007). Multilayer
Perceptron (MLP) consisting of a single hidden layer is used for the prediction of the
pressure drop across the piping components and is trained with Backpropagation
(gradient-descent) algorithm. The hyperbolic tangent transfer function in the hidden
layer and a linear transfer function in the output layer have been used for updating the
connection weights. This analysis reveals that during the training phase the mean
squared error (MSE) for training and cross-validation kept on decreasing over the
32000 epochs for most of the cases. This is a indicator that suggests that training was
good. To optimize the network before testing and final prediction the number of nodes
or processing elements are varied from 1 to 25. The number of nodes for which the
value of MSE reaches a minimum is considered to be the optimum number. These
optimum numbers of nodes are used for further subsequent analysis. The optimum
number of nodes in the hidden layer is given by 8, 14, 9 and 20 for elbow, orifice,
xv
Summary
gate valve and globe valve respectively. The cross-correlation coefficient value is
nearly 0.99 for each of the four cases for testing and final prediction. The low value of
Average Absolute Relative Error and Standard Deviation also prove the effectiveness
of the ANN analysis.
Chapter - 4: This chapter deals with the prediction of pressure drop for gas-non-
Newtonian liquid flow through horizontal pipe and piping components, namely elbow,
orifice, gate valve and globe valve. The experimental data used for the prediction is
collected from the earlier work carried out in our laboratory and the subsequent
publications (Banerjee and Das, 1998; Samanta et al., 1999; Bandyopadhyay et al.,
2000; Bandyopadhyay, 2002). The proposed prediction is done with a Multilayer
Perceptron (MLP) consisting of a single hidden layer, which is trained with
Backpropagation (gradient-descent) algorithm. Four different transfer functions in the
hidden layer and a linear transfer function in the output layer have been used for
updating the connection weights. This analysis reveals that during the training phase
the mean squared error (MSE) for training and cross-validation kept on decreasing
over the 32000 epochs for most of the cases. This is a indicator that suggests that
training was good. To optimize the network before testing and final prediction the
number of nodes or processing elements are varied from 1 to 25. The number of nodes
for which the value of MSE reaches a minimum is considered to be the optimum
number. These optimum numbers of nodes are used for further subsequent analysis.
The cross-correlation coefficient value is nearly 0.98 for each of the five cases for
testing and final prediction. Because of the closeness of the result the Chi-square (χ²)
test was performed to find the best result.
Chapter - 5: This chapter deals with the prediction of pressure drop for gas-non-
Newtonian liquid flow through 45° bend. The experimental data used for the
xvi
Summary
prediction is collected from the earlier work by Das (1988). The proposed prediction
is done with a Multilayer Perceptron (MLP) consisting of a single hidden layer, which
is trained with Backpropagation (gradient-descent) algorithm. Four different transfer
functions in the hidden layer and a linear transfer function in the output layer have
been used for updating the connection weights. This analysis reveals that during the
training phase the mean squared error (MSE) for training and cross-validation kept on
decreasing over the 32000 epochs for most of the cases. The best predictive network
consists of the hyperbolic tangent transfer function with 20 processing elements in the
hidden layer.
Chapter - 6: In this chapter ANN is used to predict the two-phase frictional pressure
drop and gas holdup for gas-non-Newtonian liquid flow through different helical coils
in horizontal orientation. The experimental data used for the prediction is collected
from the earlier work carried out in our laboratory and the subsequent publications
(Biswas and Das, 2007; Biswas, 2007). The proposed prediction is done with a
Multilayer Perceptron (MLP) consisting of a single hidden layer, which is trained
with Backpropagation (gradient-descent) algorithm. Four different transfer functions
in the hidden layer and a linear transfer function in the output layer have been used for
updating the connection weights. The similar procedure as mentioned in the earlier
chapters is also followed here. The best predictive network consists of the hyperbolic
tangent transfer function with 10 and 15 processing elements in the hidden layer for
two-phase frictional pressure drop and gas holdup respectively.
Chapter - 7: This chapter deals with the prediction of the flow regime in small
diameter (1 – 5 mm) tubes for air-water flow in horizontal tubes. A data bank of 2114
data collected from different literature (Barnea et al., 1983; Triplett et al., 1999;
Coleman and Garimella, 1999; Yang and Shieh, 2001; Chen et al., 2002; Ide et al.,
xvii
Summary
2007; Venkatesan et al., 2010) and 218 data points from experiments performed in
our own laboratory. The study is conducted using 5 different ANN structures, i.e., two
types of Multilayer Perceptrons (MLP) trained with Backpropagation (BP) algorithm
and Levenberg Marquardt (LM) algorithm respectively, Radial Basis Function (RBF),
Support Vector Machine (SVM) and Principal Component Analysis (PCA). The total
data is randomized for five different times to create five samples to eliminate
sampling error. Then each of these five samples are analyzed (training, cross-
validation and final prediction) for all five different neural networks as mentioned
earlier. The output is represented by 7 columns corresponding to 7 different flow
patterns. Then the output is translated from symbolic data to numeric data, i.e., into 7
columns consisting of 0 and 1, where each input row corresponding to a particular air
velocity, water velocity and tube diameter is represented in the output with 6 values of
number 0 and one value of number 1. The row having number 1 corresponds to that
flow pattern. Then each sample is analyzed with the above mentioned ANNs
separately. Finally the mean of all five samples are calculated and it has been
observed that Radial Basis Function (RBF) gives slightly better predictability than the
other ANNs.
Chapter - 8: This chapter deals with the general conclusions of the entire thesis.
CHAPTER 1
Introduction
Introduction
1 Introduction
The Artificial Neural Network or ANN models which emulate human brain
actions in processing information are being extensively applied in recent years in
many disciplines like Physics, engineering, forecasting, financial market prediction
and other real-world problems (Haykin, 1994; Marini, 2009; Feng and Lu, 2010; Oh
and Lee, 2010). In scientific and engineering applications the ANN methodology
yield accurate, robust, fault and failure tolerant results for both linear and nonlinear
applications and has proved capable of handling complex problems involving
modeling, classification, function approximation, pattern recognition, multivariable
data analysis, fault detection, prediction of polymer quality, process control etc. in a
number of practical applications (Hoskins and Himmelblau, 1988; Himmelblau, 2000;
Basheer and Hajmeer, 2000; Cong et al., 2013).
The attractiveness of ANNs come from their remarkable information
processing characteristics; the system may be nonlinear, highly parallel, robust, fault
and failure tolerant, ability to handle large scale data and their capability to generalize
them through experimental data or field observation. The objective of ANN based
computing is to develop mathematical algorithms that will enable ANNs to learn from
past information and predict present information as accurately as possible. The basic
predicative methodology of ANNs comprises of subdividing the control element into
numerous interconnected but simple processing elements, called nodes or neurons that
are capable of performing massive parallel computations for data (Hecht-Nielsen,
1990; Schalkoff, 1997; Basheer and Hajmeer, 2000).
Survey of literature reveals that artificial neural networks (ANN) are now
extensively used in different branches of science and technology. Rumelhart et al.
(1986) proposed the "multilayer perceptron" (MLP) and showed that it was an
2
Introduction
application for parallel distributed processing. From the end of the 80's there has been
explosive growth in applying neural networks to various problems in science and
engineering, like process dynamics, modeling of processes, biotechnology and
bioengineering, data analysis and transformation etc. (Molga, 2003, Kumar et al.,
2010). Literature review shows the applicability of ANN to different industrial
processes, e. g., to control the product, online estimation of product quality, drying
process of food industry etc. (Montagu and Morris, 1994; Ramesh et al., 1996; Liu et
al., 2000).
Single phase and multiphase flow of fluid through piping components is a very
important topic for science and engineering which is amenable to analysis by
theoretical methods only under simple situations. For handling complex situations
extensive data generation is necessary for generating empirical correlations. The ANN
methodology may provide an viable alternative for the prediction of parameters
related to hydrodynamics of single and multiphase flow through pipes and piping
components, e.g., pressure drop, gas or liquid holdup, flow regimes. The model based
on ANN exhibit the input-output mapping without following any specific rule, which
decides its widespread application. ANN models can learn from examples,
incorporate large number of variables, and provide an adequate and quick response to
the new information. However, Individual relations between the input variables and
the output variables are not developed by scientific/engineering judgment but it is
similar to a black box. Some of the advantages of neural networks are as follows:
1) The ability to represent both linear and non-linear relationships.
2) The ability to learn these relationships directly from the data used is an
advantage of ANN.
3
Introduction
3) The Advantages of MLP is that this type of network can be used to create a
model that correctly maps the input to the output using historical data so that
the model can then be used to produce the output when the output is unknown.
4) At least in some cases if not always, i.e., for prediction by the trained network
is alternative to experimentation and save a lot of time which may have been
consumed.
However, ANN technique also has some disadvantages (Sando et al., 2005) like,
1) Minimizing over-fitting requires a great deal of computational efforts.
2) The individual relations between the input variables and the out variables are
not developed by engineering judgment so that the model tends to be a black
box without analytical basis.
Adhikari and Jindal (2000) used ANN approach to predict pressure drop of
non-Newtonian fluid through tubes of 7.51mm to 16.34 mm stainless steel tube.
Larachi et al. (2001) combined the ANN and dimensional analysis to derive
correlations for bed porosity and liquid and gas holdup in three phase fluidization.
Blanco et al. (2001) reported the applicability of ANN modeling to determine several
linear and non-linear physical properties of bitumens. Shaikh and Al-Dahhan (2003)
reported the prediction of the overall gas holdup in bubble column reactors using
artificial neural network using a databank of around 3500 experimental data collected
from the open literature. Xie et al. (2003) used ANN for the prediction of flow regime
in gas-liquid-pulp fiber slurry flow. Sablani et al. (2003) reported ANN approach to
predict the friction factor for flow of bingham plastic. Sablani and Shayya (2003)
developed ANN based non-iterative calculation of the friction factor for power law
fluids. Molga (2003) reported the two types of neural models namely a global neural
model and a hybrid neural model to a conventional CSTR reactor. Zhao et al. (2005)
4
Introduction
reported the use of ANN to calculate the thixotropic stress of waxy crude oil. Torkar
et al. (2008) used ANN to predict the apparent viscosity of alumina-paraffin
suspension. Lahiri and Ghanta (2008) reported the use of ANN for the prediction of
holdup in slurry transportation system. Manickraj and Balasubramanian (2008) used
ANN for the estimation of the heat transfer coefficient in liquid-solid fluidized bed.
Marini (2009) reviewed the applicability of the ANN prediction in foodstuff analysis.
Recently, Singha et al. (2014) successfully used ANN to predict percentage removal
of Cr (VI) from aqueous solution used in adsorption process.
The first article on the application of neural networks in physics was reported
by Gorman and Sejnowski (1988), where they used ANN to classify sonar returns
from two undersea targets, a metal cylinder and a similarly shaped rock. Cutts et al.
(1989) demonstrated that ANN is an important tools for pattern recognition in high-
energy Physics. Lynch et al. (2001) reviewed the existing works on neural networks
in Physics and also suggested some possible applications in Physics. They observed
that data acquisition, device control and robotics are the prime area for the use of
neural networks. In quantum mechanics first Darsey et al. (1991) and then Androsiuk
et al. (1994) used neural network to solve the Schrödinger equation to find
eigenenergies of a two-dimensional harmonic oscillator. Later Sugawara (2001)
developed a new methodology for solving the Schrödinger equation based on the
genetic algorithm and the neural network. Müller (2003) discussed the applications of
artificial intelligence in high energy and nuclear physics. Caetano et al. (2011)
proposed a new approach to solve nonlinear differential equations related to atomic
and molecular physics using ANN derived from density function theory (DFT). In the
field of medical physics, successful classification of ECG signals was achieved by
using a two-stage feed forward neural network (Hosseini et al., 2006). Meneses et al.
5
Introduction
(2011) proposed a method using Energy Minimization via Graph Cuts (EMvGC)
algorithm with artificial neural networks (EMANNs) for correcting misclassifications
due to intensity variation of phase contrast. They concluded that EMANN performed
better than conventional techniques for bone and insect imaging.
Artificial neural networks (ANN) have been used as analytical tools in some
cases related to fluid dynamics. Cai et al. (1994) used self Kohonen self-organizing
feature map model to classify the flow regimes of air-water two-phase horizontal flow.
Tsoukalas et al. (1997) used a neurofuzzy system to classify patterns during air-water
vertical flow. Mi et al. (1998, 2001) investigated two-phase gas-liquid flow regime in
vertical flow using neural network systems. They used signals from electrical
capacitance probes as input to the neural network for the training process of the
supervised neural network system. Dibike and Solomatine (2001) successfully
performed their study on the applicability of ANNs for downstream flow forecasting
in the Apure river basin (Venezuela). The study of Milano and Koumoutsakos (2002)
showed that ANN can be used to get good reconstruction and prediction capabilities
for the near wall velocity fields. Sunde et al. (2005) successfully used a MLP trained
with backpropagation algorithm for the classification of two-phase air-water vertically
upward flow. Trafalis et al. (2005) used MSVM models to predict the flow regimes in
vertical and horizontal two-phase flow in pipes. Their study had shown that MSVN
network predicts more accurately than theoretical correlations. Hernández et al.
(2006) reported the flow regime classification using Probabilistic neural network
(PNN) and self-organized neural network (SONN) for air-water upward two-phase
flow from the signals measured from a conductivity probe during experiment. They
demonstrated successfully the suitability of the ANN-based method developed for
flow regime identification. Sharma et al. (2006) used probabilistic neural network
6
Introduction
based Bayes-Parzen classification theory to predict flow pattern in the air-water flow
through vertical pipelines. Yunlong et al. (2008) used a method based on image multi-
feature fusion and support vector machine (SVM) to identify gas-liquid two-phase
flow regime. They observed 100% accuracy for the overall identification of the flow
regime. Bai et al. (2008) presented the input signals to a linear vector quantization
(LVQ) ANN and they were able to predict flow regime very accurately. Kuwahara et
al. (2008) successfully used radial basis function (RBF) to classify flow regimes for
air-magnetic fluid vertically upward two-phase flow. Alizadehdakhel et al. (2009)
conducted their experiment on a tube of diameter 2 cm and successfully predicted the
pressure drops using CFD and ANN. Tambouratzis and Pázsit (2009) presented a
non-invasive methodology self-organizing map (SOM) to identify Boiling Water
Reactor (BWR) two-phase flow regimes from neutron radiography images (frames) of
coolant flow recordings. Tambouratzis and Pàzsit (2010) presented a general
regression ANN for the identification of the two-phase flow that occurs in the coolant
channels of boiling water reactors and they were able to predict flow regime very
accurately. Recently Yari et al. (2014) used an artificial neural network (ANN) model
to calculate drag force on an axisymmetric underwater vehicle by obtaining dataset
from a computational fluid dynamic analysis.
Bar and Das (2011, 2013a) successfully predicted the two-phase friction
factor for gas-non-Newtonian liquid flow through 90o, 135o and 180o bend using
Backpropagation (BP), Delta-Bar-Delta (DBD), Quick-Prop (QP), Levenberg-
Marquardt (LM), Scaled Conjugate gradient (SCG) training algorithms in a single
hidden and output layer; they used four different transfer functions in their modeling.
Recently Bar and Das (2013b) successfully predicted flow regimes related to air-
water flow in micro channels (0.05 – 0.6 mm) using 3 different ANN structures i.e.,
7
Introduction
two types of Multilayer Perceptrons (MLP) trained with Backpropagation (BP)
algorithm and Levenberg Marquardt (LM) algorithm respectively and Support Vector
Machine (SVM) of diameters in horizontal plane. These studies prove that when we
consider fluid dynamics as a whole, the prediction of hydrodynamic parameters using
ANN is in its infancy. So there it is necessary to find the usefulness of prediction of
hydrodynamic parameters using ANN. From the above study one can draw the
conclusion that artificial neural network has now become a integral part of physics.
The single and two-phase flow through piping components, like elbows,
orifice, gate and globe valves, bend, helical coils is more complex than the straight
pipes. The problems to determine the pressure drop is important in designing and
analyzing the machines used for hydrodynamics. The friction between the fluid and
the wall causes this pressure drop. The problem of predicting pressure drop in these
systems is much more uncertain than for the straight pipe because the mechanism of
flow is not clearly defined. At least three types of losses are possible, i.e., skin
friction, loss due to change in flow direction and the constriction of the flow path.
The simultaneous gas–liquid two-phase flows are featured in numerous
industrial applications such as in areas of chemical, civil, nuclear, mineral, energy,
food, pharmaceutical, metallurgy and physics. For the optimization of design and
operation of two-phase flow system one needs to know the hydrodynamic parameters,
i.e., pressure drop, holdup and flow regime. Hagen (1839) was the first to determined
the pressure drop for capillary flow as sum of viscosity and kinetic energy term. The
classical equation known as Hagen-Poiseuille equation was reported by Poiseuille
(1841). Later it was established that the viscosity is an intrinsic property of the
material (Navier, 1823; Stokes, 1845). These theories had a huge impact in
hydrodynamics and ever since it became a integral part of physics.

8
Introduction
The knowledge of flow pattern is essential for the prediction of pressure drop,
holdup and thermodynamic quantities like heat and mass transfer coefficients (Bell,
1988; Spedding and Spence, 1988). The complexity of the two-phase flow system and
also the lack of basic understanding of the underling physics of the problem, the
majority of the two-phase pressure drop and holdup analysis are suggested by using
either empirical or semi-empirical correlation. Literature reviews suggest that
enormous amount of literature are available in the gas-Newtonian liquid flow through
tubes and are summarized in several books (Wallis, 1969; Govier and Aziz, 1972;
Hestroni, 1982; Crowe, 2006). However, Taitel and Dukler (1976) developed
mechanistic model to predict the flow regime transition and later on other researchers
(Xiao et al., 1990; Ansari et al., 1994; Kaya et al., 1999; Zhang et al., 2003a, b)
developed comprehensive mechanistic models to determine the flow pattern and also
to compute the pressure drop and holdup for gas- liquid flow. Very little work has
been reported with two-phase gas-non-Newtonian liquid flow in comparison with the
gas-Newtonian liquid system. The literature review of Das (1988) indicated that the
gas-non-Newtonian liquid flow behave differently from the hydrodynamics of gas-
Newtonian liquid flow and also it has great practical importance particularly in
transport of non-Newtonian liquid and other industrial applications. Das et al. (1989,
1992) developed empirical correlation for the prediction of frictional pressure drop
and holdup for gas-non-Newtonian liquid flow through horizontal tube. Dziubinski
(1995) developed an expression of drag ratio for two-phase pressure drop for
intermittent flow of gas-non-Newtonian liquid.
The flow through curved geometry is more complex than that of straight pipe,
when flow enters the curved pipe the presence of curvature generates the centrifugal
force that acts right angle to main flow and results in secondary flow. The strength of
9
Introduction
the secondary flow depends on the curvature of the surface. The secondary flow is
superimposed on the main flow and the point of maximum velocity is moved towards
the outer wall. A literature survey indicated that numerous publications could be
found dealing with flow phenomenon and the pressure drop in single-phase flow
through piping components and helical coils (Berger and Talbot, 1983; Das, 1996
a,b). Two-phase gas-liquid flow through curved pipes is more complex in nature.
When the flow enters the curved portion, the heavier density phase is subjected to a
larger centrifugal force and causes the liquid to move away from the center of the
curvature. For helical coils this process is a continuous function of coil geometry.
However, two-phase flow in helical coils has rarely investigated. Mandal and Das
(2002, 2003) reported the extensive literature survey for gas-Newtonian liquid flow
through helical coils. The area of gas-non-Newtonian liquid flow has received
increasing attention in last few decades. The primary reason is the many industrial
applications in which the systems exhibit complex rheological behaviors. But the
literature review (Biswas, 2007) suggested that gas-non-Newtonian liquid flow
through coils have limited study. Mathematical models derived from the physical
description and understanding of the gas-liquid flow through coils is extremely
difficult as the phenomena of momentum transfer between the phases, the wall
friction, the shear at the phase interface, change of flow direction, phase separation
cannot be specified quantitatively. Hence researchers correlate their experimental data
using empirical or semi empirical correlation. In many industrial systems like power
generation, nuclear reactors, petroleum and biochemical processing systems, it is
necessary to monitor the flow regimes during normal and also transient operation
conditions for the safety and overall performance. So the accurate prediction of the
flow regime is an important task. Techniques developed for flow regime detection are
10
Introduction
mostly based on experimental data (Ghanbarzadeh et al., 2012). Traditional methods
of flow pattern identification are generally performed by visual observation for
transparent pipes (Hanafizadeh et al., 2011), while some recent methods are X-ray
and γ-ray tomography, magnetic resonance imaging, electrical probe, electric
capacitance tomography, flow induced vibration and ultrasound techniques (Xu and
Xu, 1998; Jana et al., 2006; Hua et al., 2010; Hernandez et al., 2011; Kreitzer et al.,
2012, 2013). However, the main two methods used by the researchers to identify two-
phase flow regimes are direct observation such as high speed photography and
analysis of two-phase signals like void fraction fluctuation or pressure fluctuation
signals (Jones Jr. and Zuber, 1975; Matsui, 1986; Lowe and Rezkallah, 1999; Sun et
al., 2006; Paranjape et al., 2011;). Empirical flow regime map was first developed by
Baker (1954) and later Taitel and Dukler (1976) proposed flow regime map using
superficial gas velocity and superficial liquid velocity as coordinate system. These
studies prove that when we consider fluid dynamics as a whole, the accurate
prediction of hydrodynamic parameters using empirical correlation or any other
scientific method are extremely difficult for the various complexities of the type of
fluid systems and the geometry of the components. It is also clear that the knowledge
of hydrodynamic parameters are necessary for the manufacturing of various
components used for the flow of fluid. So it is necessary to study the usefulness of
prediction of hydrodynamic parameters using ANN as a method alternative to
empirical correlation.
A critical examination of literature available reveal that artificial neural
network is a very important analytical tool for large scale data processing and it can
be used to predict hydrodynamic parameters. The present study aims at the
development of a useful method of predicting the hydrodynamic parameters, in simple

11
Introduction
to complex flow conditions, using artificial neural networks. The present thesis
reports studies on the following systems,
1) Prediction of pressure drop for non-Newtonian liquid flow through piping
components, namely elbow, orifice, gate valve and globe valve.
2) Prediction of pressure drop for gas-non-Newtonian liquid flow through pipe
and piping components, namely horizontal tube, elbow, orifice, gate valve and
globe valve.
3) Prediction of pressure drop for gas-non-Newtonian liquid flow through 45°
bends.
4) Prediction of frictional pressure drop and gas holdup for gas-non-Newtonian
liquid flow through helical coils in horizontal orientation.
5) Prediction of flow regimes for air-water flow through 1 - 5 mm diameter
horizontal tubes.
The thesis has been presented in 8 Chapters.
Chapter 1 presents the importance of liquid and gas-liquid flow through various pipes
and piping components, its usefulness along with the details literature review
focusing on the hydrodynamics of single and two-phase flow and the research
work done in this field by various researchers in the past. This chapter also
gives a brief historical review of the studies related to ANN, its application in
fluid dynamics and its development over the years.
Chapter 2 presents the details description of two types of Multilayer Perceptrons
(MLP), i.e., Backpropagation (BP) algorithm and Levenberg Marquardt (LM)
algorithm respectively, Radial Basis Function (RBF), Principal Component
Analysis (PCA) and Support Vector Machine (SVM) and their optimization
techniques i.e., their overall methodology. It also gives the detail of the
12
Introduction
statistical parameters used for the measurement of accuracy of the ANN
analysis.
Chapter 3 presents the prediction of pressure drop for non-Newtonian liquid flow
through piping components, namely elbow, orifice, gate valve and globe valve.
The low value of Average Absolute Relative Error and Standard Deviation
and with the cross-correlation coefficient value of nearly 0.99 for each of the
four cases for testing and final prediction proves the effectiveness of the
analysis.
Chapter 4 presents the prediction of pressure drop for gas-non-Newtonian liquid flow
through horizontal pipe and piping components, namely elbow, orifice, gate
valve and globe valve. The low value of Average Absolute Relative Error and
Standard Deviation and with the cross-correlation coefficient value of nearly
0.98 for each of the five cases for testing and final prediction proves the
effectiveness of the analysis. Finally the Chi-square (χ²) test was performed to
find the best result.
Chapter 5 presents the prediction of pressure drop for gas-non-Newtonian liquid flow
through 45° bend. The low value of Average Absolute Relative Error and
Standard Deviation and with the cross-correlation coefficient value of nearly
0.98 for each of the five cases for testing and final prediction proves the
effectiveness of the analysis. Finally the Chi-square (χ²) test was performed to
find the best result.
Chapter 6 presents the prediction of two-phase frictional pressure drop and gas holdup
for gas-non-Newtonian liquid flow through different helical coils in horizontal
orientation. The low value of Average Absolute Relative Error and Standard
Deviation and with the cross-correlation coefficient value of nearly 0.95 for
13
Introduction
each of the five cases for testing and final prediction proves the effectiveness
of the analysis for the prediction of two-phase frictional pressure drop. Finally
the Chi-square (χ²) test was performed to find the best result.
However, for the prediction of gas holdup the cross-correlation
coefficient value was found to be nearly 0.90. To verify the accuracy of the
result for the prediction of gas holdup is copared with the empirical correlation
established by Biswas (2007). Finally it is found that the ANN prediction is
slightly better representation of the experimental data than the empirical
correlation established by Biswas (2007).
Chapter 7 presents the prediction of the flow regime in small diameter (1 – 5 mm)
tubes for air-water flow in horizontal tubes using 5 different ANN structures.
First the total of 2332 data are divided into five samples to eliminate sampling
error. Then each of these five samples are analyzed (training, cross-validation
and final prediction) for all five different neural networks. Finally the mean of
all five samples are calculated and it has been observed that Radial Basis
Function (RBF) gives slightly better predictability than the other networks.
Chapter 8 presents the general conclusions of the entire thesis.

CHAPTER 2
Artificial Neural Network Methodology

2.1 Introduction
This chapter deals with the ANN methodology used for the present study. It
gives brief description of the training algorithms used and their relevant
characteristics.
At present there are few software packages, e.g., Matlab, Neurusolution,
INForm, CAD/Chem, Statistica, EasyNN, WEKA etc. with various versions are
available in the market. Some of them were tried for this study also. Chen et al. (2002)
reported their test performed on four available neural network packages namely,
Neural Shell2, BrainMaker, CAD/Chem and NeuralWorks Professional II/Plus and
concluded that all the software packages have the capacity to generate reasonably
predictive models. They also reported that the optimized neural network architecture
may vary from program to program and claimed that NeuralShell2 had given the best
predictability but the methodology for choice of training algorithm was not properly
mentioned. Hence the comparison was not proper. Plumb et al. (2005) reported their
test performed on three available neural network packages (Matlab, INForm and
CAD/Chem) to compare their predictive ability using four different backpropagation
algorithms. They concluded that the optimized neural network architecture may vary
from software to software but there are no significant differences in the predictive
ability of the three software packages. Neurosolution 5 was used for the analysis of all
the problems reported in this thesis.
Sola and Sevilla (1997) reported the effects of data normalization on the ANN
process and concluded normalization of data yields better prediction. However,
recently it was reported that without normalization of data also yields better results
(Wu and Lo, 2010; Bar and Das, 2011, 2013a,b).

15
The present thesis is the successful demonstration of the predictive property of
different type of artificial neural networks when it is used for function approximation
and classification problems in the field of fluid dynamics. The function approximation
is manifested by the prediction of few hydrodynamic parameters like single-phase
pressure drop, two-phase frictional pressure drop, gas holdup etc. for liquid and gas-
non-Newtonian liquid flow through horizontal pipes and piping components,
horizontally oriented helical coils and bends using backpropagation (BP) algorithm on
a three layered Multilayer Perceptron. The classification property of artificial neural
network is demonstrated by the prediction of flow regime in small diameter (1–5 mm)
for air-water flow using Backpropagation (BP), Levenberg-Marquardt (LM), Radial
Basis Function (RBF), Principal Component Analysis (PCA) and Support Vector
Machine (SVM).
2.2 Artificial Neural Network Architecture
Fig. 2.1 shows the schematic diagram of an artificial neural network (ANN).
Multilayer Perceptron technique was proposed and successfully applied in the parallel
distributed processing system (Rumelhart et al., 1986). Mathematically a multilayer
perceptron spans a family of functions from a finite dimensional input to a finite
dimensional output (Lopez and Oňate, 2006). These functions are parameterized by
all the biases and synaptic weights in the neural network (Šima and Orponen, 2003). It
was observed that a multilayer perceptron with one hidden layer of sigmoid neurons
and an output layer of linear neuron provides a general framework for approximating
any function up to any desired degree of accuracy, provided sufficiently many hidden
neuron are available (Hornik et al., 1989).

16
So the basic structure of ANN (Multilayer Perceptrons) used for function
approximation in this thesis consists of three layers. The MLP used for classification
also consists of three layers.
2.2.1 Input layer
The input layer simply accepts the data from the specific file and after
adjusting the weights they are sent to the hidden layer.
2.2.2 Hidden layer
The number of hidden layers and the number of processing elements or nodes
in them vary, according to the optimization capability using any particular training
algorithm and network architecture. The synapse that connects the hidden layer to the
input layer adjusts the weights and the learning rates. It is always desired that the
number of processing elements in the hidden layer must be kept at a minimum to
reduce the complexity of the network.
The processing elements use a transfer function or activation function. Four
different transfer functions were used for our network. These are represented below as
eβ x − e− β x
Transfer function 1: =y f= βx
( x) tanh= (2.1)
eβ x + e− β x
= β x 1 for β x > 1
Transfer function 2: = ( x) β x where
y f= (2.2)
β x = −1 for β x < −1
= β x 0 for β x < 0
Transfer function 3: = ( x) β x where
y f= (2.3)
= β x 1 for β x > 1
1
Transfer function 4: =y f=
( x) (2.4)
1 + e− β x
where, β is the gain and it is used to control the steepness of the transfer function. The
value of β is mostly unity.

17
2.2.3 Output layer
Since Generalized delta-rule algorithm is used for the backpropagation for this
network so a bias function is used. The number of processing elements or nodes in the
output layer may vary also according to the necessity. For all problems of function
approximation a linear function is used as the output transfer function. It is
mathematically represented as
y= f ( x)= x + b (2.5)
where, b is the bias term. It is the processing element of the output layer each of
which sums the weighted connections from the hidden layer. Thus the bias function
simply adds a fixed number to the summation of the processing elements and the
output is generated. The linear processing element is constructed with a synapse and a
bias axon. The Synapse implements a sum of products and the bias axon adds the bias.
2.3 Different algorithms used for training
The training process is actually a process of learning for the network. There
are three types of learning, i.e., supervised, unsupervised and reinforced learning.
During the training process the all the weight values are adjusted so that we can
obtain the best nonlinear relationship between the input and output parameters. The
success of the ANN analysis primarily depends on the type of network chosen and
then the algorithm used to update the weight values. The most popular known type of
ANN structures are chosen for the studies presented here. A brief description of all
these ANNs and their optimization process are presented here.
2.3.1 Multilayer Perceptron (MLP)
The MLP was first proposed by Werbos (1974) and then implemented
successfully by Rumelhart et al. (1986). To choose an appropriate network
architecture one must keep in mind the size of the network, the complexity, the
18
training time and most importantly the capacity of generalization, i.e., ability to give
satisfactory results for data different from that used in training. To choose the proper
network architecture for MLP one need to consider the choice of the number of
hidden layers, the transfer functions used in those hidden layer and the algorithms
used for training.
Presently, there are many training algorithms reported in the literature, e.g.,
Backpropagation (BP) (Rumelhart et al., 1986), Delta-Bar-Delta (DBD) (Jacobs,
1988), Quick-Prop (QP) (Fahlman, 1988), Extended-Delta-Bar-Delta (EDBD) (Minai
and Williams, 1990), Conjugate gradient (CG) (Charalambous, 1992), Scaled
Conjugate gradient (SCG) (Mӧller, 1993), Levenberg-Marquardt (LM) (Hagan and
Menhaj, 1994), Bees Algorithm (BA) (Pham et al., 2006) etc. As a result of the
various studies done (Chan, 1990; Karras and Perantonis, 1993; Sidani and Sidani,
1994; Alpsan et al., 1995; Hanan and Bishop, 1997) it can be concluded that the
performance of the algorithm depends on the specific problem under consideration,
i.e., performance of the network is not system independent. For this study we have
used Backpropagation (BP) algorithm for function approximation problems.
During the last three decades a network with single hidden layer using some
popular transfer functions like sigmoid, hyperbolic tangent etc. were used to perform
function approximation successfully (Hecht-Nielsen, 1987; Carol and Dickinson,
1989; Hornik et al., 1989; Barron, 1993; Bar and Das, 2011; Bar and Das, 2013a).
MLPs are structured with the above mentioned three layers: an input layer,
hidden layer(s) and an output layer. Processing elements or nodes in the input layer
only transfers the input signals xi =

( i 1, 2, − − −, N ) related to a desired output x to each
of the j (where=
j 1, 2, − − −, N ) processing elements of the hidden layer. The synapse
19
that connects the hidden layer to the input layer adjusts the weights and the learning
rates. Each input signal is then being multiplied by their respective connection
weights wij . Then every processing element in the hidden layer sums its input signals.
If the output of the jth processing element in the hidden layer is h j then
 Nj 
h j = f nh  ∑ wij xi  (2.6)
 j =1 
where, f nh denotes the nth hidden layer transfer function. For this analysis the four
( n = 4 ) different transfer functions in the hidden layer have been used. The transfer
functions used in the form of equation are presented in Eqs. (2.1–2.4). If y is the
network output (a single output, i.e., single and two-phase pressure drop, frictional
pressure drop per unit length) of the processing element of the output layer and w j is
the connection weight corresponding to the jth hidden layer output, i.e., h j , and the
output layer. Then
 Nj 
y = f out  ∑ w j h j  (2.7)
 j =1 
where, f out is the output layer transfer function. It is the linear transfer function and is
presented at Eq. (2.5).
During the training the learning process of the network continues by
modifying the connection weights wij and w j that have been randomly chosen at the
beginning of the training to minimize the sum of the squared difference between the
desired and the network output E with each epoch. Where E is represented as:
=E
2
(
1 2
xin − yin2 ) (2.8)
20
where, i corresponds to the ith input, n corresponds to the nth epoch during training.
The training process continues, i.e., the number of epochs are chosen such that the
network reaches a point where the modifications of the connection weights do not
significantly minimize E or a threshold value of E, which is assigned in the beginning
of the training is reached during training. The modification process of the connection
weights is done using training algorithm. The important aspects of training algorithm
are the type of learning of the network and the speed of training process.
2.3.1.1 Backpropagation (BP)
Fig. 2.2 represents the flowchart of typical Backpropagation process used for
the problems of function approximation. Backpropagation is discussed in detail in our
earlier papers (Bar and Das, 2011, 2012a, 2013a,b). After the output is generated the
criterion function accepts the output(s) of the network and compares them with the
desired output(s). It calculates the error and passes this error to the backpropagation
components, which propagates the errors backward through the network and allows
adaptation of the weights in the hidden layer processing elements and a closed-loop
control system is thus established. The weights are adjusted using a gradient-descent-
based algorithm. Widrow et al. (1988) proposed the Generalized delta-rule algorithm
which is used for the Backpropagation. If wij (t ) is the value of connection weight in
the hidden layer then the weights are updated using the following equation during the
epoch number (t + 1) ,
∂E
−α
∆wij (t ) = (t ) (2.9)
∂wij
The above equation was modified by using a momentum term and is represented by
∂E
∆wij (t ) =−α (t ) + µ∆wij (t − 1) (2.10)
∂wij
21
where, ∆wij (t ) represents the change of connection weights for the jth processing
element in the hidden layer during epoch number t with that of ith input xi . α is
learning rate and µ is the momentum coefficient.
If we take a high value of α then it will cause the training to be accelerated
and it will yield a higher change in the weight value also. This will cause the search to
oscillate on the error surface a bit more than usual, which in turn will make
convergence difficult. In contrast the smaller value of α allows the search to be slow
but steady and directed towards a global minima.
The value of µ helps updating the weight values in such a way that the search
escape local minima and reduce the possibility of instability (Haykin, 1994). A greater
value of µ will make sure that the search is not stuck in a local minima. On the other
hand it can also create the possibility where there is the risk of overshooting the
solution exactly similar to the case of high value of α . According to Henseler (1995)
when µ >1.0 will cause instability. Also a smaller value of µ will slow down the
training.
The initial values of α and µ for BP algorithm are not mentioned in some of
the papers where these algorithms are successfully used (Ubeyli and Guler, 2004;
Ozkan et al., 2010; Bonakdari et al., 2011). Although in some papers the values of
above mentioned parameters are given (Karnin, 1990; Maier and Dundy, 1998;
Udelhoven and Schutt, 2000; Yildiz et al., 2002; Desai et al., 2005; Ho, 2009). From
the above literature survey it is clear that the values are randomly chosen. So it is
clear that these values are dependent on the problem it deals with. The initial values of
these constant parameters mentioned for training of BP are kept constant throughout
this analysis. A few of these values of α and µ for BP algorithm are kept common
22
for this study and the studies by Karnin (1990), Maier and Dundy (1998), Udelhoven
and Schutt (2000), Yildiz et al. (2002), Desai et al. (2005) and Ho (2009). For this
analysis gives more emphasis towards the variation of number of processing elements
in the hidden layer and transfer functions used for the individual algorithms. In the
present study Backpropagation is used for the problems related to function
approximation as well as Classification.
2.3.1.2 Levenberg-Marquardt (LM)
This is a second order learning algorithm that uses the method of optimization.
If N is the number of epochs and n is the number of weights then the weight update
equation is given as:
∆=
w (J T
J + λI ) J Te (2.11)
where, J is ( N ×1) × n Jacobian matrix and e is ( N ×1) ×1 error matrix, I is the identity
matrix and λ is the combination coefficient. Here e is represented as:
 e1 
e 
 2
−
e=  (2.12)
−
−
 
eN 
where, the tth element of error matrix e, i.e., et is represented as:
e=
t xt − yt (2.13)
where, xt and yt are the desired output and actual output during epoch number t
respectively.
The modified Jacobian matrix J (since there is only 1 output) is represented as:
23
 ∂e1 ∂e1 ∂e1 

 ∂w − − −
∂w2 ∂wn 
 1 
 ∂e2 ∂e2 ∂e2 
 − − −
J =  ∂w1 ∂w2 ∂wn  (2.14)

− − − − − − − − − − 
 
 ∂eN ∂eN
− − −
∂eN 
 ∂w1 ∂w2 ∂wn 
In this algorithm the only user dependent parameter is λ , which is set only in the
beginning. It is required for the user to modify the value of λ during the training any
more. When λ is large the algorithm behaves like steepest descent and when it is
small the algorithm behaves like Gauss-Newton. In this way the Levenberg-
Marquardt algorithm combines the best features of these two algorithms but avoids
most of their limitations. In the present study this is used for the problems related to
function approximation as well as Classification.
2.3.2 Radial Basis Function (RBF)
Broomhead and Lowe (1988) were the first to use RBF network. Moody and
Darken (1989), Poggio and Girosi (1990) also used Gaussian as there transfer
function in the network used by them. RBF consists of three layers, namely an input
layer, a hidden and output layer having linear transfer function. The training process
of RBF consists of two stages, the first of which is unsupervised and the next stage is
supervised.
The unsupervised learning stage determines the parameters for the basis
function. It only involves the input data. In this study k-means clustering algorithm is
used. It is used to divide the total input data into k number of subsets in such a way
that the distance between each of the cluster center and the points in there are
minimized. The basis function for the network used in this study is a Gaussian
function represented here as,

24
 n 2
 ∑ ( xi − X ) 
G ( xi ) = exp  − i =1  (2.15)
 2σ 2 
 
where, n is the number of inputs, X is the center of each Gaussian and σ is the RBF
width parameter. This width parameter, σ value varies from one cluster to another.
The value of ( xi − X ) is measured using Euclidean distance.
Each output of the hidden unit is represented by,
∑ ∑ w G (x )
N
h( xi ,σ j ) = i j i (2.16)
i =1 j =1
where, N is the number of basis functions. The optimization of the width σj of the
RBF kernel parameter is done using an iterative process. The first aim of the training
of a RBF network is to determine an optimum central point of each cluster using a
basis function (Gaussian). It should be noted that central point of each cluster may not
be coinciding with any data point. Any supervised learning algorithm can be used for
the optimization of width parameter, σ j . In this analysis Levenberg-Marquardt
algorithm is used. A hyperbolic tangent transfer function presented in Eq. (2.1) is used
as a transfer function during the supervised learning stage in the various processing
elements.
The transformation from input to hidden layer is highly nonlinear and the
hidden layer to output layer is fully linear. This linear output is given by,
F ( xi ) = w0 + ∑ wi {h( xi ,σ j )} (2.17)
i =1
The overall performance of RBF network is dependent on the quality of clustering
and the choice of the value of k apart from the optimization of the value of the width
25
parameter. For the present study RBF analysis is used for problems of Classification,
i.e., flow regime prediction.
2.3.3 Principal Component Analysis (PCA)
The Principal Component Analysis (PCA) was first used by Oja (1982, 1992)
and later on developed also by Sanger (1989). This analysis also involves a
unsupervised stage first and a supervised stage after that.
The Generalized Hebbian Algorithm (GHA) used and popularized by Sanger
(1989) is used for this analysis in the unsupervised stage. It is primarily a process of
feature extraction in the first stage then the classification of this features in the second
stage using Levenberg-Marquardt algorithm as mentioned in Section 2.2.1.2. The
number of principal components can be chosen by the user according to their choice.
For the present study analysis using PCA network is used for the problem of
Classification, i.e., flow regime prediction. The basics of the method and some
modification are also discussed by other researchers (Karmer, 1991; Damiantaras and
Kung, 1996; Chatterjee et al., 2000).
2.3.4 Support Vector Machine (SVM)
The Support Vector Machine (SVM) was first proposed by Glucksman (1966)
and later on it was popularized later by others (Boser et al., 1992; Cortes and Vapnik,
1995). This technique can be understood by considering two hyperplanes in between
two sets of data for classification in such a way that the difference between the two
sets are maximum, i.e., SVM orients a boundary in such a way that the distance
between the boundary and the nearest data in each set is maximum (optimal
separating surface which is equidistant from both the sets). Let us consider two
datasets as A and B. If we consider two hyperplanes describing the boundary of each

26
data set then the boundary on set A is considered to be passing through some points of
the set A and the similar is true for the boundary of set B. The points falling on
(defining) the boundary of A and B are known as support vectors. Once the SVs are
selected the rest of the data can be rejected and by doing so the number of training
data is reduced. For the present study analysis using SVM network is used for the
problem of Classification, i.e., flow regime prediction. The basics of the method are
also discussed by other researchers (Vapnik, 1998; Burges, 1998; Amari and Wu,
1999; Scholkopf et al., 1999).
2.4 Optimization of the ANN
The optimizations of different networks are achieved by trial and error
method. These processes of optimization of the networks depends on the following,
1) the description of data used;
2) the number of processing elements in the hidden layer;
3) the number of epochs for training of the respective networks;
4) stopping criterion of training;
5) the values of the constant parameters used (only for BP method);
Initially the total data is randomized to prevent sampling error. Then 60% data
points are used for training, 20% for cross-validation, 10% for testing and the rest 10%
used for prediction for the problems related to function approximation. It is always
desired that the number of processing elements in the hidden layer must be kept
minimum to reduce the complexity of the network. In the hidden layer the numbers of
processing elements are optimized by varying the number 1 to 25. Similar procedure
is followed in all of our earlier paper (Bar and Das, 2011, 2012a, 2013a). When a
neural network starts to train the connection weights start to modify so that the error is
reduced. However, the continuation of reduction of error in training does not

27
guarantee good results as there is a possibility of over-fitting or over-training (Reed,
1993). To check the performance a good strategy is cross-validation during training.
The cross-validation is done simultaneously with training as it was done earlier by us
(Bar and Das, 2011, 2012a,b, 2013a,b). The stopping criterion of the training is
dependent on the value of the training and cross-validation MSE. For the problems
where at the time of training, both the training and cross-validation error kept on
decreasing all the times the minimum value of MSE for training is considered to be
optimum to chose the number of processing elements or nodes in the hidden layer. If
there is some increase in the value of cross-validation MSE then the point where the
cross-validation MSE achieves its minimum value is considered to be optimum. The
threshold value of the MSE for cross-validation in all cases has been kept as 0.001 for
all the problems related to function approximation (pressure drop and holdup
prediction) and is 0.01 for the problems related to classification (flow regime
prediction). There are also some other stopping criterions dependent on the algorithm
used.
For the hidden layer of Backpropagation (BP) network the value of learning
rate α is 1 and that of momentum coefficient µ is 0.7. For the output layer of BP
network the value of learning rate α is 0.1 and that of momentum coefficient µ is 0.7.
The maximum numbers of epochs used are 32000. Another stopping criterion has
been added, i.e., if there is no improvement of the value of cross-validation MSE for
20000 epochs, then the training is stopped.
For Levenberg-Marquardt (LM) algorithm the initial value of λ is 0.01. The
numbers of epochs are 500. If for 200 epochs there is no improvement in the value of
cross-validation MSE, then training is stopped.

28
For the training of Rradial Basis Function (RBF) network 1000 epochs for
unsupervised learning and maximum of 1000 epochs for supervised learning. The
numbers of clusters (K value) are kept 15. However training is stopped when no
improvement in cross-validation MSE for 200 consecutive epochs during supervised
learning phase have been observed. For the supervised learning Levenberg-Marquardt
algorithm has been used for updating the weights under supervised learning condition.
There are 3 principal components kept for this analysis. For training of
Principal Component Analysis (PCA) network with three principal components, 1000
epochs have been used for unsupervised learning and also 1000 epochs for supervised
learning. However, training is stopped when there is no improvement in cross-
validation MSE for 200 consecutive epochs during supervised learning phase. For the
supervised learning LM algorithm is used for updating the weights.
For the training with Support Vector Machine (SVM) network 1000 epochs
are used. However training is stopped when no improvement in cross-validation MSE
for 300 consecutive epochs is observed.
2.5 Performance of the ANN
The performance of the network is checked using the following parameters:
Mean Square Error (MSE),
N
1
=
MSE
N
∑ (x − y )
i =1
i i
2
(2.18)
Average Absolute Relative Error (AARE),
1 N
( yi − xi )
AARE =
N
∑ i =1 xi
(2.19)
Standard Deviation (σ),

2
N
1  ( yi − xi ) 
=σ ∑ 
i =1 N − 1  xi
− AARE  (2.20)

29
The AARE and Standard Deviation are kept as small as possible for the better
performance of the Neural Network.
Cross-correlation coefficient (R),
∑ ( x − x)( y − y)
i i
R= i =1
(2.21)
N N
i
=i 1 =i 1
∑ ( x − x) ∑ ( y − y )
2
i
2
It has also been verified that the Cross-correlation coefficient between input and
output is as close to unity as possible.
Chi-square test (χ²),
N
( xi − yi ) 2
χ2 = ∑ (2.22)
i =1 yi
The Chi-square test has also been performed to find the best-fit network model. The
lowest value indicates the best model.
The thesis consists of the studies on the applicability of ANN modeling as,
i. Prediction of pressure drop across the piping components for non-
Newtonian liquid flow using Backprogagation training algorithm in
hidden with the Eq. (2.1) and output layer with the Eq. (2.5) as transfer
function.
ii. Prediction of pressure drop across the pipes and piping components for
gas-non-Newtonian liquid flow using Backprogagation training algorithm
in hidden layer with the Eqs. (2.1 – 2.4) and output layer with the Eq.
(2.5) as transfer functions.
iii. Prediction of pressure drop across the 45o bend for gas-non-Newtonian
liquid flow using Backprogagation training algorithm in hidden with the
Eqs. (2.1 – 2.4) and output layer with the Eq. (2.5) as transfer functions.
30
iv. Prediction of pressure drop and holdup across the helical coil in horizontal
orientation for gas-non-Newtonian liquid flow using Backprogagation
training algorithm in hidden with the Eqs. (2.1 – 2.4) and output layer
with the Eq. (2.5) as transfer functions.
v. Prediction of flow regime for air-water flow through small diameter (1 – 5
mm) pipeline using Backprogagation (BP), Levenberg-Marquardt (LM),
Radial Basis Function (RBF), Principal Component Analysis (PCA) and
Support Vector Machine (SVM) training algorithm in hidden and output
layer with the Eq. (2.1) as transfer function.

31
Fig. 2.1 Schematic diagram of artificial neural network

32
START
INPUT LAYER
In i where i = 1, 2, ---- n
Inputs and the corresponding outputs are
presented to the hidden later
Weight update equation is

w ih=
(New) w ih (Old) + ∆w ih
Bias update equation is
θ ih (New)
= θ ih (Old) + ∆θ ih
HIDDEN LAYER
Processing Elements vary from 1 to 25 for each Transfer Function
7
Inputs presented to the hidden using equation I h = ∑ w hi In i +θ h

i=1
Output of hidden layer is calculated using the transfer functions of Eqs

(2.1-2.4) i.e., O nh = f nh (I h ) where I h = x
Weight and bias are updated and stored automatically
OUTPUT LAYER
25
Inputs presented to the output using equation Yo = ∑ w O oi nh

+θ o i vary from 1 to 25
i=1
Output of the output layer is calculated using the following transfer functions i.e.,
Y = f ( Yo ) . If Y = y , then Y= f ( y )= y + b
o o
Weight and bias are updated and stored using the following equation
Weight update equation is
w io=
(New) w io (Old) + ∆w io
Bias update equation is
θ io (New)
= θ io (Old) + ∆θ io
N
1
=
Estimation of error Ei
2
∑ (x i
− yi )
2
i =1
No Ei ≤ 1 × 10-3
Yes
End Training
Fig. 2.2 Flowchart representing training of the single hidden layer backpropagation
network
CHAPTER 3
Non-Newtonian liquid flow through piping components in

horizontal plane
Non-Newtonian liquid flow through piping components in horizontal plane
Numerous investigations have shown that ANN can be used successfully for
correlating experimental data sets for macroscopic single and multiphase flow
characteristics. The approach proved its worth when rigorous fluid mechanics
treatment based on the solution of first principle equations is not tractable. Evaluation
and prediction of the frictional pressure drop across different piping components such
as orifices, gate and globe valves and elbows in 0.0127 m piping components for non-
Newtonian liquid flow is manifested in this chapter. The experimental data used for
the prediction is taken from the earlier work carried out in our laboratory,
Bandyopadhyay and Das (2007). The proposed approach towards the prediction is
done using a multilayer perceptron (MLP), which is trained with backpropagation
algorithm because the function approximation is achieved with very good accuracy
using MLPs.
3.1 Introduction
The pipe fittings for valves, elbows, orifice, etc. are the essential part of any
piping system. Hydrodynamics of any piping components is more complex in nature
than the straight pipes. The problems to determine the pressure drops in pipe fittings
are important in designing and analyzing the machines used for hydrodynamics. The
friction between the fluid and the fitting wall causes this pressure drop. Dissipation of
energy takes place when the flow of fluid takes place through pipes, which can be
detected by the pressure drop across the various cross-sections within the pipes. The
problem of predicting pressure drops in pipes is much more uncertain than for the
straight pipe because the mechanism of flow is not clearly defined. At least two types
of losses are possible i.e., skin friction and loss due to change in flow direction. In the
present paper application of ANN is used to predict the pressure drop in different
34
piping components for the flow of non-Newtonian liquid based on the earlier
experimental work carried out in our laboratory by Bandyopadhyay and Das (2007).
3.2 Artificial neural network architecture
Experimental data has been collected from the study conducted in our
laboratory earlier (Bandyopadhyay and Das, 2007). Table 3.1 represents the
description of the number data used for neural network analysis, and Table 3.2
represents the range of data used.
Input parameters are the physical and operating variables of the system. The
physical variables include the diameter of the tube, Dt, radius of curvature of the
elbow, Reb, angle of the elbow, αeb, ratio of the valve opening to the full opening of
valve (α) in case of globe and gate valve, properties of the non-Newtonian liquid, i.e.,
flow behavior index (nʹ), consistency index (Kʹ), density (ρl), while the operating
variable is liquid flow rate (Ql).
The hyperbolic tangent function presented in Eq. (2.1) is preferred for this
problem because
1) Hyperbolic tangent function has grater slope than that of the logistic function.
2) Hyperbolic tangent function has a negative response for negative input and
positive response for a positive input on the other hand the logistic function
has always a positive response.
Initially the total data is randomized. The first 60% of data points are used for
training, the next 20% for cross-validation, the next 10% for testing and the rest used
for prediction.
In the hidden layer the numbers of nodes are optimized varying the number 1
to 25, and each case the MSE is calculated. Figs. 3.1 – 3.4 show the variation of MSE
35
with the number of nodes for different systems. It is clear from the graph that at a later
stage there is an upward trend of MSE is observed. Similar trends are also observed
by Yetilmezsoy and Demirel (2008). The optimum number of node is that node where
the MSE is minimum. Table 3.3 shows the optimum number of nodes in the training
section of the ANN for different system. These optimum numbers of nodes are used
for further subsequent analysis.
3.4 Result and discussion
For every separate system the optimum result has been achieved using 32000
epochs for training. The cross validation is also done simultaneously. The total data
set was randomized before the network is used.
Figs. 3.5 – 3.8 represent the training curve for the four different systems
namely elbow, orifice, gate valve and globe valve respectively. Figs. 3.9 – 3.12
represent the cross-validation curve for the above mentioned four different systems.
Initially the MSE for each epoch for training and cross-validation in all four cases are
recorded for 5 different runs separately. Each run consists of 32000 epochs. Then the
average of the MSE of all the five different runs is calculated for each of the 32000
epochs for each of the four systems. The gradual decrease of the values of average
MSE in all four cases shows that the training and cross-validation in all four cases are
accurate.
Figs. 3.13 – 3.16 show the comparison between the experimental to the
predicted output for testing for the above mentioned four different systems.
Figs. 3.17 – 3.20 show the comparison between the experimental to the
predicted output. This comparison proves the effectiveness of the Neural Network
analysis.
36
Tables 3.6 and 3.7 represent the performance of neural network for testing and
prediction. It is clear from these tables that the cross-correlation coefficient value is
more than 0.99 for each of the four cases. The low value of the average absolute
relative error (AARE) also shows the accuracy of the result in the different systems.
This result indicates that the performance of the network output is excellent.
3.5 Conclusions
A neural network based model has been developed for the prediction of
pressure drop for non-Newtonian liquid flow through different piping components,
i.e., elbow, orifice gate valve and globe valve. A multilayer perceptron with
backpropagation algorithm has been used for this analysis. The ANN model
accurately predicts the pressure drop across the different piping components as
evident from the cross-correlation coefficient, which is more than 0.99.

37
5.0x10-3
Piping Component: Elbow
4.5x10-3
4.0x10-3
Minimum value of MSE for training
3.5x10-3
3.0x10-3
2.5x10-3
2.0x10-3
1.5x10-3
1.0x10-3
5.0x10-4
5 10 15 20 25
Number of nodes in hidden layer
Fig. 3.1 Variation of the minimum value of MSE with the number of nodes in the
hidden layer
4.0x10-4
Piping component: Orifice
3.0x10-4
Minimum value of MSE for training
2.0x10-4
1.0x10-4
0.0
5 10 15 20 25
Number of nodes in the hidden layers
hidden layer
38
3.0x10-3
Piping component: Gate valve
2.5x10-3
Minimum value of MSE for traning
2.0x10-3
1.5x10-3
1.0x10-3
5.0x10-4
0.0
5 10 15 20 25
Number of nodes in the hidden layers
hidden layer
5.00x10-3
Piping component: Globe valve
2.50x10-3
0.00
5 10 15 20 25
Number of nodes in the hidden layer
hidden layer
39
101
Piping component: Elbow
100
Average MSE for training
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
Fig. 3.5 Variation of the average MSE for training over 5 different runs vs the number
of epochs
102
101
100
Average MSE for traning
10-1
10-2
10-3
10-4
10-5
1 10 100 1000 10000
Number of epochs
of epochs
40
101
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
of epochs
102
101
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
of epochs
41
101
100
Average MSE for cross-validation
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
Fig. 3.9 Variation of the average MSE for cross-validation over 5 different runs vs the
number of epochs
102
101
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
Fig. 3.10 Variation of the average MSE for cross-validation over 5 different runs vs
the number of epochs
42
101
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
102
101
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
43
10
Pressure drop accross the elbow, predicted
1
∆Pelbow, kPa
0.1
0.01
0.01 0.1 1 10
Pressure drop accross the elbow, experimental
∆Pelbow, kPa
Fig. 3.13 Comparison of pressure drop across the elbow for testing
100
Pressure drop accross the orifice, predicted
10
∆Porifice, kPa
0.1
0.1 1 10 100
Pressure drop accross the orifice, experimental
∆Porifice, kPa
Fig. 3.14 Comparison of pressure drop across the orifice for testing
44
100
Pressure drop accross the gate valve, predicted
10
∆Pgate valve, kPa
0.1
0.1 1 10 100
Pressure drop accross the gate valve, experimental
∆Pgate valve, kPa
Fig. 3.15 Comparison of pressure drop across the gate valve for testing
100
Pressure drop accross the globe valve, predicted
10
∆Pglobe valve, kPa
0.1
0.1 1 10 100
Pressure drop accross the globe valve, Experimental
Fig. 3.16 Comparison of pressure drop across the globe valve for testing
45
10
Pressure drop accross the elbow, predicted
1
∆Pelbow, kPa
0.1
0.01
0.01 0.1 1 10
Pressure drop accross the elbow, experimental
∆Pelbow, kPa
Fig. 3.17 Comparison of pressure drop across the elbow for prediction
100
Pressure drop accross the orifice, predicted
10
∆Porifice, kPa
0.1
0.1 1 10 100
Pressure drop accross the orifice, experimental
∆Porifice, kPa
Fig. 3.18 Comparison of pressure drop across the orifice for prediction
46
100
10
∆Pgate valve, kPa
0.1
0.1 1 10 100
∆Pgate valve, kPa
Fig. 3.19 Comparison of pressure drop across the gate valve for prediction
100
Pressure drop accross the globe valve, predicted
10
0.1
0.1 1 10 100
Pressure drop accross the globe valve, experimental
Fig. 3.20 Comparison of pressure drop across the globe valve for prediction
47
Table 3.1 Description of the data used for neural network analysis
Data
Data Data Data
considered
System type considered considered considered
for cross-
for training for testing for prediction
validation
Elbow 67 22 11 11
Orifice 37 13 6 6
Gate valve 184 62 31 30
Globe valve 198 66 33 33
Table 3.2 Range of different data sets for all four different systems
Measurement type Range

Angle of elbow 45° to 135°
Radius of curvature of elbow (m) 0.011 ≤ Rc ≤ 0.022
Linear Length of elbow (m) 0.011 ≤ Leb ≤ 0.016
Orifice diameter (m) 0.0059 ≤ Do ≤ 0.0090
Ratio of the valve opening to the full
0.25≤ α ≤ 1.00
opening of valve
Flow behavior index 0.6015 ≤ n/ ≤ 0.9013
Consistency index (Nsn'/m2) 0.0142 ≤ K/ ≤ 0.7112
Density (kg/m3) 1001.69 ≤ ρl ≤ 1003.83
Concentration of SCMC solution (kg/m3) 0.2 to 0.8
5 3
Liquid flow rate Ql×10 (m /s) 3.75 to 29.83
Pressure drop (Experimental) (kPa)
Elbow 0.0333 ≤ ΔPElbow ≤ 2.8
Orifice 0.2666 ≤ ΔPOrifice ≤ 23.733
Gate valve 0.133 ≤ ΔPGate valve ≤ 31.467
Globe valve 0.666 ≤ Δ PGlobe valve ≤ 36.933
Table 3.3 Optimum number of nodes for different systems
System Number of nodes

Elbow 8
Orifice 14
Gate valve 9
Globe valve 20
Table 3.4 Performance of neural network for training for 5 runs
Measurement type Elbow Orifice Gate valve Globe valve

SD (σ) 0.000048 0.000016 0.000115 0.000085
MSE 0.000769 0.000004 0.000136 0.000431
48
Table 3.5 Performance of neural network for cross-validation for 5 runs

SD (σ) 0.004500 0.000804 0.000811 0.000123
MSE 0.004770 0.000416 0.000535 0.000299
Table 3.6 Performance of neural network for testing
Measurement type Elbow Orifice Gate value Globe valve

AARE 0.115858 0.397462 0.206211 0.057142
SD (σ) 0.120079 0.428693 0.461611 0.097175
MSE 0.009806 0.267522 0.458182 0.250700
CCC (R) 0.985481 0.998365 0.994948 0.998415
Table 3.7 Performance of the neural network for prediction

AARE 0.065341 0.066737 0.117726 0.061164
SD (σ) 0.063186 0.064147 0.129887 0.060828
MSE 0.002403 0.229735 0.141582 0.309966
CCC (R) 0.993125 0.995197 0.998141 0.997769
Table 3.8 Comparison between experimental and ANN prediction for elbow
Concentration Liquid Pressure drop

Elbow of SCMC flow rate ΔPElbow
type solution Ql×105 (kPa)
(kg/m3) (m3/s) Experimental ANN prediction
0.6 14.62 1.2667 1.249759
45° 13.13 0.9000 0.911881
0.8
14.62 1.3333 1.295753
16.81 1.1000 1.022941
0.4
21.93 1.6667 1.596243
90°
7.37 0.3667 0.29727
0.8
14.62 1.0667 1.049036
0.2 11.00 0.2000 0.224098
0.4 14.62 0.5333 0.447197
135°
0.6 3.75 0.1333 0.130684
0.8 9.50 0.5333 0.556886
49
Table 3.9 Comparison between experimental and ANN prediction for orifice
Pressure drop
Ratio of Concentration
Liquid flow ΔPOrifice
orifice to pipe of SCMC
rate Ql×105 (kPa)
diameter solution
(m3/s) ANN
(D0/Dt) (kg/m3) Experimental
prediction
7.10 3.3333 3.387938
0.2
0.4646 13.17 13.3333 12.97115
0.6 4.05 1.3333 1.443126
10.17 2.4000 2.393595
0.2
0.5984 13.17 5.3333 4.417875
0.8 13.17 6.2666 5.638773
Table 3.10 Comparison between experimental and ANN prediction for gate valve
Ratio of the Pressure drop

Liquid flow
valve opening to Concentration of ΔPGate valve
rate
the full opening SCMC solution (kPa)
Ql×105
of valve (kg/m3) ANN
(m3/s) Experimental
(α) prediction
1 2 3 4 5
0.4 17.00 29.600 28.80939
0.25 5.58 3.067 3.692321
0.6
11.83 15.467 15.52798
0.4 26.58 16.933 18.20118
0.375
0.6 22.83 13.067 13.22906
0.4 10.58 0.933 1.340598
0.6 11.83 1.133 1.340662
0.5 14.67 2.267 2.265269
0.8 17.67 3.333 3.214601
19.17 4.000 3.779162
13.83 1.200 1.18915
0.4
26.58 3.600 4.222866
0.625
0.6 21.33 2.733 2.565035
0.8 23.75 3.600 3.688514
10.58 0.800 0.794543
0.4
18.58 1.867 1.656816
19.67 2.000 1.748432
0.75
0.6 22.83 2.667 2.513837
26.00 3.600 3.455987
0.8 17.67 1.600 1.707707
0.4 5.58 0.267 0.145876
16.58 1.200 1.296625
0.6
0.875 22.83 2.400 2.345548
7.00 0.400 0.357105
0.8
23.75 2.267 2.947786
50
1 2 3 4 5
0.4 26.58 3.133 2.749782
13.50 0.800 0.949554
0.6
1.00 19.67 1.733 1.619196
16.67 1.600 1.517764
0.8
10.17 0.533 0.761945
Table 3.11 Comparison between experimental and ANN prediction for globe valve
Ratio of the Pressure drop

Concentration Liquid
valve opening to ΔPGlobe valve
of SCMC flow rate
the full opening (kPa)
solution Ql×105
of valve ANN
(kg/m3) (m3/s) Experimental
(α) prediction
0.4 7.17 6.400 5.85189
0.25
0.8 7.00 6.133 6.52783
0.375 0.6 8.83 5.600 6.513195
15.00 13.200 14.36812
0.5 0.6
21.33 28.267 28.14108
9.08 3.467 3.934271
0.2
10.83 4.933 5.492578
15.42 10.000 11.3533
0.4
18.58 15.867 16.3066
0.625 16.58 14.267 14.40507
0.6
18.17 16.267 17.13177
4.05 1.067 0.854864
0.8 14.67 10.933 11.26347
22.25 23.867 24.90891
0.2 7.25 2.000 2.149102
0.4 17.00 11.867 11.34345
10.33 5.200 5.085809
0.6 15.00 10.667 10.38398
0.75
24.42 26.267 25.85415
8.58 3.733 3.859251
0.8 10.17 5.600 5.236604
17.67 14.400 14.61506
13.83 6.400 6.45859
0.4
18.58 12.533 12.06304
21.33 18.400 18.55921
0.875 0.6 22.83 20.667 21.07895
27.5 30.400 29.21663
5.58 1.467 1.720833
0.8
20.75 18.8 18.67782
7.17 1.867 1.864492
0.4
9.00 2.267 2.780653
1.00
0.6 15.00 8.933 8.654032
0.8 5.58 1.467 1.490792
CHAPTER 4
Gas-non-Newtonian liquid flow through pipes and piping

components in horizontal plane
Gas-non-Newtonian liquid flow through pipes and piping components in horizontal plane
The ANN approach proved its worth when rigorous fluid mechanics treatment
based on the solution of first principle equations is not tractable. Evaluation and
prediction of the frictional pressure drop across different piping components such as
orifices, gate and globe valves, elbows, and horizontal pipe in 0.0127 m diameter for
gas non-Newtonian liquid flow is manifested in this chapter. The experimental data
used for the prediction is taken from the earlier work done in our laboratory, i.e.,
Bandyopadhyay (2002) and the subsequent publications, i.e., Banerjee and Das (1998),
Samanta et al., (1999) and Bandyopadhyay et al., (2000). The proposed approach
toward the prediction is done using a multilayer perceptron (MLP) with one hidden
layer and four different transfer functions, which is trained with backpropagation
algorithm.
4.1 Introduction
The problem of predicting pressure losses in piping components is much more
uncertain than that for straight pipes because the mechanism of the flow is not clearly
defined. At least three types of losses are superposed – skin friction, loss due to
change of flow direction and the constriction of the flow path. There are only few
experimental data available in the literature as mentioned by Das (1996b). Two-phase
gas-liquid flow occurs in many engineering applications, such as in equipment related
to the oil, chemical, process and power generation industries. The hydrodynamics of
co-current gas-liquid flows have received extensive treatment during last few decades.
There are, however, areas which have received little attention; one of these areas is
gas-liquid flow through piping components. From the above studies it is clear that
gas-liquid flow through piping components is even more complex in nature. The gas-
non-Newtonian two-phase flow is gaining important, as it is clear from the review of
Das et al. (1989). Mathematical models derived from the physical description and
52
understanding of the gas-liquid flow through piping components is extremely difficult
as the phenomena of momentum transfer between the phases, the wall friction, the
shear at the phase interface, change of flow direction, phase separation and
constriction cannot be specified quantitatively. Hence, the empirical equations for
pressure losses across the different piping components for gas-non-Newtonian liquid
flow have been derived by Das and his co-workers (Banerjee and Das, 1998; Samanta
et al., 1999; Bandyopadhyay et al., 2000; Bandyopadhyay, 2002) from their
experimental data. In recent years, the concept of artificial neural network (ANN) has
gained widespread application in many engineering problems Himmelblau (2000).
ANN models can learn from examples, incorporate large number of variables, and
provide an adequate and quick response to the new information.
In the present chapter application of ANN is used to predict the two-phase
pressure drop for 0.0127 m horizontal pipe and different piping components based on
the earlier experimental data obtained from (Bandyopadhyay, 2002, Banerjee and Das,
1998, Samanta et al., 1999 and Bandyopadhyay et al., 2000).
4.2 Artificial Neural Network Architechture
Experimental data has been collected from the earlier work carried out in our
laboratory (Bandyopadhyay, 2002, Banerjee and Das, 1998, Samanta et al., 1999 and
Bandyopadhyay et al., 2000). Table 4.1 represents the description of the number of
data used for Neural Network analysis, and Table 4.2 represents the range of data
used.
physical variables include the diameter of the tube (Dt), radius of curvature of the
elbow (Reb), angle of the elbow (αeb), ratio of the valve opening to the full opening of
valve (α) in case of globe and gate valve, properties of the non-Newtonian liquid, i.e.,
53
flow behavior index (n'), consistency index (K'), surface tension of the liquid (σl),
density of liquid (ρl), density of gas (ρg), viscosity of gas (µg) , acceleration due to
gravity (g), while the operating variable is liquid flow rate (Ql) and gas flow rate (Qg).
The values of the physical properties of gas, i.e. density and viscosity, the tube
diameter and acceleration due to gravity are constant so it is ineffective for ANN
programming.
Four different transfer functions as given in Eqs. (2.1–2.4) are used in the
hidden layer for this analysis. These are represented below as
Initially the total data was randomized. The first 60% of data points are used
for training, the next 20% for cross-validation, the next 10% for testing and the rest
used for prediction.
set has been randomized before the network is used. In the hidden layer the numbers
of processing elements are optimized varying the number 1 to 25, and each case the
MSE was calculated. Figs. 4.1 to 4.5 show the variation of MSE with the number of
processing elements for horizontal pipe, elbow, orifice, gate valve and globe valve
respectively. Similar procedure was followed by Yetilmezsoy and Demirel (2008).
The optimum number of processing elements is that where the MSE is minimum.
These optimum numbers of processing elements are used for further subsequent
analysis.
4.4 Results and discussion
Figs 4.6 to 4.10 represent the training curve for the above mentioned five
different piping systems for four different transfer functions used in the hidden layer.
54
Figs. 4.11 to 4.15 represent the cross-validation curve for the above mentioned
five different piping systems for four different transfer functions used in the hidden
layer. Initially the MSE for each epoch for training and cross-validation in all five
cases are recorded for 5 different runs separately. Each run consists of 32000 epochs.
Then the average of the MSE of all the five different runs is calculated for each of the
32000 epochs for each of the five systems. The gradual decrease of the values of
average MSE in all five cases shows that the training was accurate.
Figs. 4.16 to 4.20 show the comparison between the experimental to the
predicted output for testing for the above mentioned five different systems
respectively for all the four different transfer functions used in the hidden layer.
Figs. 4.21 to 4.25 show the comparison between the experimental to the
predicted output for the above mentioned five different systems respectively for four
different transfer functions used in the hidden layer after optimization. This
comparison proves the effectiveness of the neural network analysis.
Tables 4.5 and 4.6 represent performance of neural network for Testing and
Prediction. It is clear from these tables that the Cross Correlation Co-efficient value is
more than 0.97 for each of the five cases for four different transfer functions used in
the hidden layer. The low value of the Average Absolute Relative Error (AARE) also
shows the accuracy of the result in the different systems. This result indicates that the
performance of the network output is excellent.
Since the Cross Correlation Co-efficient value is more than 0.97 for all the
best networks, so the Chi-square (χ²) test was performed to find the best result. Table
4.6 contains the result for the Chi-square test. The Chi-square test confirms that the
best network for prediction of pressure drop for horizontal pipe is the one which has
the transfer function 3 with 25 processing elements in the hidden layer. The Chi-
55
square test confirms that the best network for prediction of pressure drop for elbow is
the one which has the transfer function 1 with 14 processing elements in the hidden
layer. The Chi-square test confirms that the best network for prediction of pressure
drop for orifice is the one which has the transfer function 1 with 15 processing
elements in the hidden layer. The Chi-square test confirms that the best network for
prediction of pressure drop for gate valve is the one which has the transfer function 4
with 13 processing elements in the hidden layer. The Chi-square test confirms that the
best network for prediction of pressure drop for globe valve is the one which has the
transfer function 2 with 25 processing elements in the hidden layer. Tables 4.7 to 4.11
show the comparison between the experimental to the ANN predicted output for the
above mentioned five different piping systems respectively for the optimum number
of transfer functions used in the hidden layer.
4.5 Conclusions
A neural network based model has been developed for the prediction of
pressure drop for gas non-Newtonian liquid flow through different piping components,
i. e., horizontal pipe, elbow, orifice gate valve and globe valve. A multilayer
perceptron with backpropagation algorithm has been used for this analysis with four
different transfer functions used in the hidden layer and a linear function in the output
layer. The ANN model accurately predicts the pressure drop across the horizontal
pipe and different piping components as evident from the Cross Correlation Co-
efficient, which is more than 0.97.
The Chi-square test confirms that the best network for prediction of pressure
drop for horizontal pipe is the one which has the transfer function 3 with 25
processing elements in the hidden layer. The Chi-square test confirms that the best
network for prediction of pressure drop for elbow is the one which has the transfer
56
function 1 with 14 processing elements in the hidden layer. The Chi-square test
confirms that the best network for prediction of pressure drop for orifice is the one
which has the transfer function 1 with 15 processing elements in the hidden layer. The
Chi-square test confirms that the best network for prediction of pressure drop for gate
valve is the one which has the transfer function 4 with 13 processing elements in the
hidden layer. The Chi-square test confirms that the best network for prediction of
pressure drop for globe valve is the one which has the transfer function 2 with 25
processing elements in the hidden layer.

57
1.2x10-2
Piping component: Horizontal pipe
Transfer Function 1
Transfer Function 2
Transfer Function 3
Minimum value of MSE for training Transfer Function 4
6.0x10-3
10 20
Number of processing element in the hidden layer
Fig. 4.1 Variation of the minimum value of MSE with the number of processing elements in
the hidden layer for four different transfer functions
9.0x10-3
Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
6.0x10-3
3.0x10-3
0.0
10 20
Number of processing elements in the hidden layer
58
8.0x10-3
Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
6.0x10-3
Minimum Value of MSE for training
4.0x10-3
2.0x10-3
0.0
10 20
6.0x10-3
Transfer function 1
Transfer function 2
Transfer function 3
Transfer function 4
Minimum Value of MSE for traning
4.0x10-3
2.0x10-3
0.0
5 10 15 20 25
59

Transfer Function 1
3.0x10-3 Transfer Function 2
Transfer Function 3
Minimum Value of MSE for traning Transfer Function 4
2.0x10-3
1.0x10-3
0.0
10 20
101
Transfer Function 1
Transfer Function 2
100 Transfer Function 3
Transfer Function 4
10-1
10-2
10-3
10-4
10-5
1 10 100 1000 10000
Number of epochs
Fig. 4.6 Variation of the average MSE for training over 5 different runs vs the number of
epochs in case for four different transfer functions in the hidden layer
60
102
Transfer Function 1
Transfer Function 2
Transfer Function 3
100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
102
Transfer Function 1
Transfer Function 2
Transfer Function 4
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
61
102
Transfer Function 1
Transfer Function 2
Transfer Function 4
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
102
Transfer Function 1
Transfer Function 2
Transfer Function 4
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
62
101
Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
Average MSE for cross-validation 100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
number of epochs for four different transfer functions in hidden layer
102
Transfer Function 1
Transfer Function 2
Transfer Function 3
100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
number of epochs for four different transfer functions in the hidden layer
63
102
Transfer Function 1
Transfer Function 2
Transfer Function 3
100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
102
Transfer Function 1
Transfer Function 2
Transfer Function 3
100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
64
102
Transfer Function 1
Transfer Function 2
Transfer Function 4
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
100
Transfer Function 1
Two-phase pressure drop accross the horizontal pipe, predicted
Transfer Function 2
Transfer Function 3
Transfer Function 4
10
∆Ptp, kPa
0.1
0.1 1 10 100
Two-phase pressure drop accross the horizontal pipe, Experimental
∆Ptp, kPa
Fig. 4.16 Comparison of two-phase pressure drop across the horizontal pipe for
testing for four different transfer functions in the hidden layer
65
100
Transfer Function 1
Transfer Function 2
Two-phase pressure drop accross the elbow, predicted

Transfer Function 3
Transfer Function 4
10
∆Ptp, kPa
0.1
0.01
0.01 0.1 1 10 100
Two-phase pressure drop accross the elbow, Experimental
∆Ptp, kPa
Fig. 4.17 Comparison of two-phase pressure drop across the elbow for testing using
four different transfer functions in the hidden layer
100
Transfer Function 1
Transfer Function 2
Transfer Function 3
Two-phase pressure drop accross the orifice, predicted
Transfer Function 4
10
∆Ptp, kPa
0.1
0.1 1 10 100
Two-phase pressure drop accross the orifice, Experimental
∆Ptp, kPa
Fig. 4.18 Comparison of two-phase pressure drop across the orifice for testing using
four different transfer functions in the hidden layer
66
100
Transfer Function 1
Two-phase pressure drop accross the gate valve, Predicted

Transfer Function 2
Transfer Function 3
Transfer Function 4
10
∆Ptp, kPa
0.1
0.01
0.01 0.1 1 10 100
Two-phase pressure drop accross the gate valve, Experimental
∆Ptp, kPa
Fig. 4.19 Comparison of two-phase pressure drop across the gate valve for testing
100
Transfer Function 1
Two-phase pressure drop accross the globe valve, Predicted
Transfer Function 2
Transfer Function 3
Transfer Function 4
∆Ptp, kPa
10
1
1 10 100
Two-phase pressure drop accross the globe valve, Experimental
∆Ptp, kPa
Fig. 4.20 Comparison of two-phase pressure drop across the globe valve for testing
67
100
Transfer Function 1
Two-phase pressure drop accross the horizontal pipe, Predicted

Transfer Function 2
Transfer Function 3
Transfer Function 4
∆Ptp, kPa
10
1
1 10 100
Two-phase pressure drop accross the horizontal pipe, Experimental
∆Ptp, kPa
Fig. 4.21 Comparison of two-phase pressure drop across the horizontal pipe for
100
Transfer Function 1
Transfer Function 2
Transfer Function 3
Two-phase pressure drop accross the elbow, Predicted
Transfer Function 4
10
∆Ptp, kPa
0.1
0.1 1 10 100
Two-phase pressure drop accross the elbow, Experimental
∆Ptp, kpa
Fig. 4.22 Comparison of two-phase pressure drop across the elbow for prediction
68
100
Transfer Function 1
Transfer Function 2
Transfer Function 3
Two-phase pressure drop accross the orifice, Predicted

Transfer Function 4
10
∆Ptp, kPa
0.1
0.1 1 10 100
Two-phase pressure drop accross the orifice, Experimental
∆Ptp, kPa
Fig. 4.23 Comparison of two-phase pressure drop across the orifice for prediction
100
Transfer Function 1
Transfer Function 2
Transfer Function 3
Two-phase pressure drop accross the gate valve, Predicted
Transfer Function 4
10
∆Ptp, kPa
0.1
0.1 1 10 100
Two-phase pressure drop accross the gate valve, Experimental
∆Ptp, kPa
Fig. 4.24 Comparison of two-phase pressure drop across the gate valve for prediction
69
100
Transfer Function 1
Transfer Function 2
Two-phase pressure drop accross the globe valve, Predicted
Transfer Function 3
Transfer Function 4
10
∆Ptp, kPa
0.1
0.1 1 10 100
Two-phase pressure drop accross the globe valve, Experimental
∆Ptp, kPa
Fig. 4.25 Comparison of two-phase pressure drop across the globe valve for
70
Table 4.1 Description of the data used for neural network analysis
Data Data considered Data Data

System type considered for for cross- considered for considered for
training validation testing prediction
Horizontal pipe 76 26 12 12
Elbow 216 72 35 35
Orifice 216 72 36 36
Gate valve 432 144 72 72
Globe valve 357 119 59 59
Table 4.2 Range of different data sets for all five different systems

Piping components
Pipe diameter (m) 0.0127
Angle of elbow 45° to 135°
Radius of curvature of elbow (m) 0.011 ≤ Rc ≤ 0.017
Ratio of the valve opening to the full
0.25≤ α ≤ 1.00
opening of valve (gate and globe)
Orifice diameter (m) 0.0059 ≤ Do ≤ 0.0090
Ratio of orifice to pipe diameter 0.4646 ≤ DO/Dt ≤ 0.7087
Physical properties of liquid
n’
Consistency index (Ns /m ) 2
0.0142 ≤ K/ ≤ 0.7112
Density (kg/m3) 1001.69 ≤ ρl ≤ 1003.83
Surface tension (N/m) 0.07834 ≤ σl ≤ 0.0832
Physical properties of air
Density (kg/m3) 1.1614
Viscosity of air (Ns/m2) 1.846 × 105
Flow Rate
5 3
Gas flow rate Qg×105(m3/s) 1.7860 to 44.7530
Pressure drop (Experimental) (kPa)
Horizontal pipe 0.5111 ≤ ΔPtp ≤ 24.3777
Elbow 0.1333 ≤ ΔPtp ≤ 7.6000
Orifice 0.4000 ≤ ΔPtp ≤ 46.4000
Gate valve 0.2667 ≤ ΔPtp ≤ 16.2667
Globe valve 1.7333 ≤ ΔPtp ≤ 45.4667
71
Table 4.3 Performance of best neural network for training over 5 runs
Measurement Transfer Transfer Transfer Transfer

System type
Type function 1 function 2 function 3 function 4
Horizontal SD (σ) 0.000050 0.000026 0.000041 0.000211
pipe MSE 0.000241 0.000028 0.000127 0.000437
SD (σ) 0.000602 0.000385 0.000088 0.000169
Elbow
MSE 0.002200 0.000615 0.001437 0.002682
SD (σ) 0.000345 0.000147 0.000141 0.000076
Orifice
MSE 0.001351 0.000525 0.000839 0.001884
SD (σ) 0.000021 0.000061 0.000086 0.000027
Gate valve
MSE 0.001004 0.000716 0.001026 0.001283
SD (σ) 0.000142 0.000063 0.000078 0.000075
Globe valve
MSE 0.000704 0.000443 0.000750 0.000979
Table 4.4 Performance of the neural network for cross-validation over 5 runs

System type
type function 1 function 2 function 3 function 4
Horizontal SD (σ) 0.001301 0.001558 0.001500 0.001038
pipe MSE 0.003521 0.004575 0.005771 0.008821
SD (σ) 0.000369 0.011327 0.000443 0.000165
Elbow
MSE 0.006549 0.006491 0.006172 0.007163
SD (σ) 0.000188 0.000088 0.000186 0.000161
Orifice
MSE 0.001314 0.001637 0.001148 0.001222
SD (σ) 0.000052 0.000071 0.000032 0.000047
Gate valve
MSE 0.001371 0.001180 0.001482 0.001651
SD (σ) 0.000114 0.000087 0.000058 0.000073
Globe valve
MSE 0.000952 0.000916 0.001011 0.001036
72
Table 4.5 Performance of the best network for testing

System type
AARE 0.139431 0.091344 0.141090 0.102353
Horizontal SD (σ) 0.159955 0.113177 0.221266 0.146129
pipe MSE 1.155644 0.786913 1.941306 1.080488
CCC (R) 0.973274 0.982891 0.954967 0.973249
AARE 0.198541 0.390736 0.164034 0.217509
SD (σ) 0.298257 0.888029 0.145425 0.189990
Elbow
MSE 0.097880 0.184884 0.118365 0.143730
CCC (R) 0.986369 0.977530 0.984129 0.980557
AARE 0.074940 0.146182 0.116760 0.091214
SD (σ) 0.062268 0.202334 0.129884 0.087222
Orifice
MSE 2.473256 6.773724 5.042553 2.915442
CCC (R) 0.994096 0.978081 0.984655 0.993024
AARE 0.123889 0.114436 0.104425 0.105825
SD (σ) 0.241808 0.136457 0.118239 0.142197
Gate valve
MSE 0.139428 0.168411 0.102588 0.116460
CCC (R) 0.988455 0.986104 0.991832 0.990975
AARE 0.091750 0.081437 0.095528 0.090104
SD (σ) 0.106653 0.087487 0.102939 0.099928
Globe valve
MSE 1.519431 1.488616 1.756384 1.576777
CCC (R) 0.993495 0.993687 0.992771 0.993492
73
Table 4.6 Performance of the best neural network for prediction
Transfer Transfer Transfer Transfer

Measurement type
function 1 function 2 function 3 function 4
Horizontal Pipe
AARE 0.049228 0.063092 0.031665 0.036619
SD (σ) 0.040359 0.060217 0.025103 0.040873
MSE 0.503364. 0.417634 0.355275 0.613286
CCC (R) 0.993446 0.994157 0.995086 0.991108
χ² 0.506992 0.598442 0.251003 0.468126
Optimum number of nodes
23 15 25 16
in hidden layer
Elbow
AARE 0.109949 0.179042 0.148015 0.147962
SD (σ) 0.095818 0.176449 0.111625 0.115751
MSE 0.071158 0.075493 0.069816 0.066139
CCC (R) 0.970161 0.969416 0.970622 0.969900
χ² 1.142137 2.648163 1.553100 1.535549
14 21 25 25
in hidden layer
Orifice
AARE 0.102161 0.225771 0.252263 0.108418
SD (σ) 0.135718 0.443963 0.494438 0.160257
MSE 1.056509 1.301362 1.306411 1.289722
CCC (R) 0.996085 0.995262 0.995488 0.995386
χ² 0.097863 –27.40547 –12.71403 5.448912
15 17 24 13
in hidden layer
Gate valve
AARE 0.103767 0.106917 0.102240 0.106946
SD (σ) 0.110178 0.103566 0.097182 0.092749
MSE 0.281146 0.324183 0.291895 0.240387
CCC (R) 0.977378 0.973481 0.976331 0.980929
χ² 4.784179 5.967776 5.079541 3.935746
18 22 21 13
in hidden layer
Globe valve
AARE 0.097179 0.086814 0.097012 0.093106
SD (σ) 0.177516 0.159675 0.209676 0.209544
MSE 3.082688 2.906793 4.158719 4.003748
CCC (R) 0.980708 0.981658 0.973646 0.974839
χ² 11.387228 10.051990 13.35638 12.25122
19 25 24 7
in hidden layer
74
Table 4.7 Comparison between the Experimental and ANN prediction in case of
Horizontal pipe for transfer function 3 with 25 processing elements in the hidden
layer
Concentration Pressure drop ΔPtp (kPa)

Liquid flow rate Gas flow rate
of SCMC
Ql × 105 Qg × 105
Solution Experimental ANN prediction
(m3/s) (m3/s)
(kg/m3)
16.17 20.55 4.933321 4.990828
0.2
16.17 23.58 5.777763 5.274757
0.4 13.17 13.60 10.66664 10.54437
4.05 23.08 9.244421 9.565604
4.05 25.53 9.622198 9.571865
13.17 11.32 13.37774 13.67911
0.6
13.17 13.29 13.11108 13.31841
13.17 25.37 10.44442 10.77923
16.17 25.00 10.26664 10.13056
4.05 16.93 20.84439 22.18126
0.8 4.05 19.81 22.42217 23.00794
7.00 17.56 21.84439 20.62309
75
Table 4.8 Comparison between the experimental and ANN prediction in case of
elbow for transfer function 1 with 14 processing elements in the hidden layer
Concentration Liquid Gas flow Pressure drop

Elbow of SCMC flow rate rate ΔPtp
type solution Ql×105 Qg×105 (kPa)
(kg/m3) (m3/s) (m3/s) Experimental ANN prediction
0.2 14.62 15.70 2.2667 2.065968
3.75 34.50 0.9333 0.973776
0.4 11.00 17.70 1.0667 1.247328
14.62 32.40 3.3333 3.15054
45° 11.00 15.60 1.0667 1.063151
0.6
18.31 22.00 3.2000 3.216127
11.00 16.70 1.3333 1.02367
0.8 11.00 37.20 2.0000 2.133195
18.31 35.00 4.2667 4.964812
3.75 31.30 0.7333 0.894471
0.2
11.00 33.00 2.6667 2.630053
0.4 14.62 35.20 2.6667 3.261253
7.37 31.50 1.2000 1.058144
14.62 27.00 1.8000 1.949367
90° 0.6 18.31 8.10 1.8000 2.015942
18.31 24.70 2.8000 2.815253
21.94 25.70 4.2667 3.620619
3.75 13.80 0.5333 0.529525
0.8 3.75 35.80 0.8000 0.843085
7.37 13.20 0.6000 0.645817
11.00 19.70 0.9333 1.027393
90° 0.8
21.94 6.50 2.4667 2.691336
7.37 27.10 0.6667 0.63931
0.2 14.62 4.10 0.6000 0.731141
14.62 36.20 2.3333 2.289849
3.75 22.20 0.8000 0.498794
7.37 21.90 0.6000 0.587868
0.4 11.00 33.00 1.0000 0.997696
135° 14.62 25.20 1.3333 1.236241
18.31 38.00 2.2667 2.978264
3.75 36.20 0.9333 0.92699
11.00 30.50 1.6667 1.323917
0.6
18.31 9.80 0.9333 1.155331
21.94 4.40 1.3333 1.520084
0.8 7.37 4.00 0.4667 0.514844
76
Table 4.9 Comparison between the experimental and ANN prediction in case of
orifice for transfer function 1 with 15 processing elements in the hidden layer
Ratio of Concentration Liquid flow Gas flow Pressure drop

orifice to pipe of SCMC rate rate ΔPtp
diameter solution Ql×105 Qg×105 (kPa)
(D0/Dt) (kg/m3) (m3/s) (m3/s) Experimental ANN prediction
7.10 17.064 8.8000 8.444915
7.10 11.509 7.8667 8.814428
0.2
10.17 28.735 17.7333 19.41237
16.30 20.387 39.3332 37.41957
9.80 9.627 13.3333 12.94572
0.4 9.80 16.176 15.4667 15.3013
16.17 10.159 33.7332 33.81536
4.05 17.436 4.2667 3.65506
0.4646
7.00 14.320 8.4000 8.415892
13.17 9.428 23.3333 22.85905
0.6
13.17 20.026 30.2667 29.48502
13.17 24.667 30.8000 32.78415
10.17 29.600 22.0000 22.76875
16.17 7.030 32.0000 34.60525
0.8 13.17 12.153 28.8000 26.3082
13.17 26.258 34.1333 35.77834
10.17 24.057 5.8667 6.440958
0.2
13.17 17.559 8.6667 8.105627
9.80 15.944 4.4000 5.188583
0.4 16.17 12.417 12.2667 11.93446
0.5984 16.17 20.060 14.1333 14.48361
7.00 13.770 2.8000 2.996167
0.6 10.17 28.915 6.9333 7.079917
16.17 12.445 13.3333 12.52229
0.8 10.17 32.949 8.2667 7.644342
0.2 4.05 16.680 0.6667 0.792466
4.05 11.148 0.5333 0.360538
4.05 18.181 0.6667 0.999422
0.4 4.05 19.274 0.6667 1.080605
9.80 23.762 3.2000 3.445014
0.7087 16.17 31.479 7.6000 9.007728
4.05 18.061 0.9333 1.104902
0.6 10.17 29.498 5.2000 4.089867
10.17 29.821 5.7333 4.114838
4.05 6.447 0.4000 0.33552
0.8
4.05 22.321 1.2000 0.989421
77
Table 4.10 Comparison between the experimental and ANN prediction in case of gate
valve for transfer function 4 with 13 processing elements in the hidden layer
Ratio of Pressure drop

the valve Liquid Gas ΔPtp
Concentration
opening to flow flow (kPa)
of SCMC
the full rate rate
Solution
opening of Ql×105 Qg×105 ANN
(kg/m3) Experimental
valve (m3/s) (m3/s) prediction
α
1 2 3 4 5 6
9.08 11.983 2.0000 2.074451
12.58 8.214 2.6667 2.876519
0.2 14.33 10.620 3.2000 3.544451
14.33 13.378 2.9333 3.723773
15.83 14.122 4.5333 4.27538
10.58 21.428 2.2667 3.168706
13.83 9.566 3.7333 3.62766
17.00 20.98 5.8667 5.952589
0.4 20.17 13.123 7.2000 6.65964
0.5 20.17 28.319 9.2000 8.454929
23.42 30.040 10.8000 10.60632
26.58 21.408 10.4000 11.25639
7.08 31.612 1.7333 2.481648
10.33 13.834 2.9333 2.847169
0.6 10.33 23.904 3.2000 3.467014
13.53 13.160 4.5333 4.155891
22.83 11.700 7.8666 8.644518
13.17 20.000 4.5333 4.501968
0.8
13.17 23.529 4.6667 4.815414
9.08 34.090 2.6667 2.430084
0.625 0.2
10.83 44.395 2.6667 3.670301
14.33 8.224 2.8000 2.427138
0.2
15.83 32.206 5.0667 4.404266
13.5 9.557 3.0667 2.780066
0.6 19.67 15.143 5.6000 5.319428
0.625
22.83 12.642 5.8667 6.338725
13.17 44.378 7.2000 5.137622
0.8 22.25 28.436 8.0000 7.672372
22.25 33.015 8.6667 8.27856
7.25 7.137 0.6667 0.717877
7.25 20.515 1.0667 1.17547
0.2
12.58 22.296 2.1333 2.489534
15.83 12.607 2.9333 2.679852
0.75
10.58 19.578 2.1333 1.971707
10.58 31.746 3.2000 2.678435
0.4
17.00 35.785 5.2000 5.225328
20.17 18.068 4.9333 4.712396
78
1 2 3 4 5 6
23.42 7.643 4.1333 4.830121
23.42 26.393 6.2667 6.596863
0.4 26.58 15.887 6.2667 6.608151
26.58 22.000 6.9333 7.254222
29.83 17.418 7.2000 7.914456
0.75
16.58 7.435 2.9333 3.37987
0.6
22.83 11.327 5.6000 5.607109
13.17 13.953 2.1333 2.395177
0.8 13.17 16.463 2.2667 2.563076
16.17 32.774 5.6000 4.943768
7.25 1.8428 0.5333 0.50834
9.08 10.534 1.0667 1.135182
0.2
9.08 13.837 1.3333 1.264789
14.33 14.314 2.6667 2.347889
10.58 6.698 1.0667 1.43029
0.4 13.83 20.380 3.2000 2.975331
0.875 17.00 10.420 2.8000 3.197081
7.08 17.482 1.0667 1.294154
0.6 13.5 31.239 4.4000 4.069982
29.08 23.609 8.8000 8.977579
10.17 24.443 2.6667 2.301555
0.8 16.17 30.970 4.6667 4.850897
22.25 25.458 6.0000 6.4031
0.2 14.33 9.1047 1.6000 2.029655
20.17 11.327 4.1333 4.180723
26.58 10.875 5.4667 5.984643
0.4
26.58 14.002 5.8667 6.284438
26.58 20.138 7.2000 6.89688
0.6 4.05 26.117 0.8000 0.965314
1.00
10.17 32.715 2.5333 3.147842
19.17 32.679 5.8667 6.327696
22.25 7.261 4.1333 4.394637
0.8
22.25 19.292 5.3333 5.779977
22.25 42.420 7.8667 8.639272
25.25 36.253 9.4667 9.007789
79
Table 4.11 Comparison between the experimental and ANN prediction for globe
valve for transfer function 2 with 25 processing elements in the hidden layer
Ratio of Pressure drop

the valve Liquid Gas ΔPtp
Concentration
opening flow flow (kPa)
of SCMC
to the full rate rate
solution
opening Ql×105 Qg×105 ANN
of valve (m3/s) (m3/s) prediction
α
1 2 3 4 5 6
0.2 10.83 28.116 12.4000 14.8392
13.83 3.3999 14.5333 15.19884
13.83 14.520 17.8667 18.26845
0.4
17.00 8.197 24.6667 27.15332
20.17 21.089 40.0000 41.41
0.5 7.08 5.382 4.9333 5.080498
10.33 22.813 13.6000 13.12204
0.6
10.33 31.168 15.4667 15.16992
13.50 25.696 20.5333 21.0226
13.17 7.467 16.9333 16.90107
0.8
16.17 26.373 32.1333 29.81873
4.05 12.168 2.0000 2.415581
0.2
17.58 8.326 8.3260 18.01292
7.17 11.560 4.6667 4.537121
13.83 6.220 14.9333 14.43964
0.625 0.4
13.83 25.413 18.2667 19.40204
17.00 2.200 19.7333 18.79036
10.33 5.857 8.9333 7.857937
0.6
10.33 10.517 9.0667 8.953107
4.05 18.732 4.0000 2.698132
10.83 18.695 11.3333 10.82385
0.2
14.33 18.664 14.8000 16.34471
0.75
17.58 14.58 24.5333 20.45872
10.58 24.671 10.6667 11.49279
0.4
17.00 17.416 23.2000 23.68644
20.17 18.039 34.6667 33.05773
0.4
20.17 29.383 38.4000 37.05104
10.33 35.232 13.8667 15.61982
0.6
0.75 19.67 5.519 28.0000 26.71179
10.17 17.241 10.6667 10.24763
0.8 13.17 24.193 16.8000 17.39857
16.17 10.563 20.9333 20.10418
4.05 14.061 2.1333 2.458932
0.2
10.83 12.318 9.2000 8.997853
0.875 10.58 28.476 12.0000 11.86905
0.4 13.83 13.126 16.1333 15.0297
13.83 21.606 17.3333 16.97362
80
1 2 3 4 5 6
7.08 22.099 5.3333 5.90696
0.6 16.58 10.925 20.4000 19.94119
16.58 19.556 22.1333 22.34946
0.875
16.17 10.676 20.6667 19.03051
0.8 16.17 21.352 23.2000 23.32883
19.17 9.909 27.8667 27.61462
10.83 10.274 8.5333 7.869045
14.33 20.270 16.1333 14.79409
0.2
17.58 11.516 20.6667 20.29838
21.00 13.483 27.6000 27.19692
13.83 6.485 12.5333 12.29174
13.83 31.830 16.4000 17.58474
0.4 17.00 12.652 19.2000 19.23248
17.00 19.750 21.7333 21.41444
1.00
17.00 31.813 22.5333 25.67654
0.6 13.50 25.906 16.4400 17.35056
4.05 10.948 2.0000 2.192688
4.05 19.925 2.1333 2.752232
4.05 24.291 3.2000 3.44149
0.8
7.08 2.809 3.2000 4.040541
10.17 14.527 8.6667 8.594569
16.17 13.333 20.0000 19.61771
CHAPTER 5
Gas-non-Newtonian liquid flow through 45° bend in

horizontal plane
Gas-non-Newtonian liquid flow through 45° bend in horizontal plane
Applicability of Artificial Neural Networks (ANN) methodology have been
investigated using experimental data obtained from our earlier experimental studies
on the frictional pressure drop for gas-non-Newtonian liquid flow through 45° bend.
This approach has proved its worth when rigorous fluid mechanics treatment based on
the solution of first principle equations is not tractable. The proposed approach
towards the prediction is done using a Multilayer Perceptron (MLP), which is trained
with backpropagation algorithm with the help of four different transfer functions in a
hidden layer.
5.1 Introduction
The theoretical analysis of the two-phase pressure drop across the bend is
extremely difficult because the phenomena of momentum transfer between the phases,
the wall friction, shear at the phase interface, curvature arises the centrifugal forces
etc. cannot be specified quantitatively. Therefore, the prediction of the frictional
pressure drop by theoretical analysis alone is not easy so the alternating correlating
method is the regression analysis. The present section deals with the applicability of
the ANN prediction for gas-non-Newtonian liquid flow through 45o bend in the
horizontal plane.
Experimental data has been collected from the earlier study by Das (1988).
Table 5.1 represents the range of data used for neural network analysis.
physical variables include the diameter of the tube (Dt), radius of curvature of the
bend (Rb), angle of the bend (θ), linear length of the bend (Lb), properties of the fluids,
i.e., flow behavior index (n'), consistency index (K'), surface tension of the liquid (σl),
density of liquid (ρl), density of gas (ρg), viscosity of gas (µg) and acceleration due to
82
gravity (g), while the operating variables are liquid flow rate (Ql) and gas flow rate
(Qg). The values of the physical properties of gas, i.e. density and viscosity, the tube
diameter, radius of curvature of the bend, angle of the bend, linear length of the bend
and acceleration due to gravity are constant so it is ineffective for ANN programming.
Four different transfer functions given in Eqs. (2.1–2.4) are used in the hidden
layer for this analysis.
5. 3 Optimization of the ANN
Initially the total data was randomized. The first 60% of data points are used
for training, the next 20% for cross-validation, the next 10% for testing and the rest
used for prediction.
For every separate system the optimum result was achieved using 32000
set was randomized before the network is used. In the hidden layer the numbers of
processing elements are optimized varying the number 1 to 25, and each case the
MSE was calculated. Fig. 5.1 shows the variation of MSE with the number of
processing elements. The optimum number of processing elements is that where the
MSE is minimum. These optimum numbers of processing elements are used for
further subsequent analysis.
Table 5.2 shows the optimum number of processing elements in the training
section of the ANN with different transfer functions. These optimum numbers of
processing elements are used for further subsequent analysis.
Fig. 5.2 represents the training curve for four different transfer functions used
in the hidden layer. Fig 5.3 represents the cross-validation curve for four different
transfer functions used in the hidden layer. The optimum number of processing
83
element is estimated on the basis of minimum value of MSE. Initially the MSE for
each epoch for training and cross-validation in all five cases are recorded for 5
different runs separately. Each run consists of 32000 epochs. Then the average of the
MSE of all the five different runs is calculated for each of the 32000 epochs for each
of the five systems. Table 5.3 represents the performance of the network for different
networks.
Testing was done just before the final prediction to check the effectiveness of
training using the above mentioned four networks. Table 5.4 represents the
performance of the testing. The low values of AARE and cross correlation co-
efficient (R) which is greater than 0.97 in each case indicates that the training was
good and the network can be used for final prediction. Fig 5.4 show the comparison
between the experimental to the predicted two-phase frictional pressure drop per unit
length across the 45o bend with optimum number of processing elements in the hidden
layer for testing.
Table 5.5 represents the performance of the neural networks for final
prediction. It is clear from the table that the cross correlation co-efficient (R) value is
more than 0.98 for each of the four cases. The low value of the average absolute
relative error (AARE) also shows the accuracy of the result in the different systems.
The high value of cross correlation co-efficient (R) and the low value of chi-
square (χ²) for the transfer function 1 with 20 optimum number of processing
elements gives the most accurate prediction of the two-phase frictional pressure drop
per unit length across the bend. Fig 5.5 shows the comparison between the
experimental to the predicted two-phase frictional pressure drop per unit length across
the 45° bend with 20 processing elements in the hidden layer and transfer function
84
number 1 in a hidden layer. This comparison is also represented in the Table 5.6. This
result indicates that the performance of the network output is excellent.
5.5 Conclusions
A neural network based model has been developed for the prediction of two-
phase frictional pressure drop per unit length for gas non-Newtonian liquid flow
through 45o bend in the horizontal plane. A multilayer perceptron with
backpropagation algorithm has been used for the analysis. Four different slandered
transfer functions in a hidden layer and a linear output function has been used.
Optimization for each transfer function is carried out in all cases. The ANN model,
using a hidden layer with transfer function 1 and 20 processing elements gives better
predictability of the two-phase frictional pressure drop per unit length across the 45o
bend for gas-non-Newtonian liquid flow.

85
Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
4.0x10-3
Minimum value of MSE during training
2.0x10-3
10 20
Fig. 5.1 Variation of the minimum value of MSE with the number of processing
elements in the hidden layer for four different transfer functions
101
Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
100
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
Fig. 5.2 Variation in the average MSE for training over 5 different runs vs the number
of epochs
86
101
Transfer Function 1
Transfer Function 2
Transfer Function 3
Average MSE for Cross-validation
10-1
10-2
10-3
10-4
1 10 100 1000 10000
Number of epochs
Fig. 5.3 Variation in the average MSE for cross-validation over 5 different runs vs the
number of epochs
100
Two-phase frictional pressure drop per unit length accross the 45o bend,
Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
predicted, ∆Ptp/Lb,(kpa/m)
10
1
1 10 100
o
Two-phase frictional pressure drop per unit length accross the 45 bend,
experimental, ∆Ptp/Lb,(kpa/m)
Fig. 5.4 Comparison of the two-phase frictional pressure drop

87
100
Two-phase frictional pressure drop per unit length accross the 45o bend,
Number of hidden layer: 1

Transfer Function: 1
Number of Processing Elements: 20
predicted, ∆Ptp/Lb,(kPa/m)
10
1
1 10 100
o
Two-phase frictional pressure drop per unit length accross the 45 bend,
experimental, ∆Ptp/Lb,(kPa/m)
Fig. 5.5 Comparison of the two-phase frictional pressure drop for final prediction
88
Table 5.1 Description of data used for neural network analysis

Pipe diameter Dt (m) 0.01905
Bend properties
Bend angle θ 45o
Radius of curvature Rc (m) 0.1195
Linear length Lb (m) 0.13
Consistency index (Nsn'/m2) 0.0295 ≤ K/ ≤ 0.6150
3
Density (kg/m ) 1002.1 ≤ ρl ≤ 1004.5
Surface tension (kN/m) 78 ≤ σl ≤ 82.97
2
Viscosity of air (Ns/m ) 0.00001846
Flow Rate
Liquid flow rate Ql (m3/s) 0.00009 to 0.000431
3
Gas flow rate Qg (m /s) 0.00003646 to 0.00046396
Measuring parameter
Two-phase frictional pressure drop per
3.076 ≤ ΔPtp/Lb ≤ 54.55
unit length (kPa/m)
Total number of data points used
210
for ANN application
Table 5.2 Optimum number of processing elements in hidden layer for four different
transfer functions
Transfer Function in Hidden Layer Optimum Number of Processing Elements

Transfer function 1 20
Table 5.3 Performance of neural network with different transfer functions for training
and cross–validation over 5 runs
Training
Measurement
Transfer Transfer Transfer Transfer
type
function 1 function 2 function 3 function 4
SD (σ) 0.000031 0.000033 0.000072 0.000011
MSE 0.000492 0.000579 0.000745 0.000970
Cross-validation
SD (σ) 0.000104 0.000154 0.000076 0.000032
MSE 0.000866 0.001125 0.001136 0.001376
89
Table 5.4 Performance of neural network with four different transfer functions for
testing using optimum number of processing elements

AARE 0.079934 0.100127 0.101506 0.085616
SD (σ) 0.163785 0.184267 0.148327 0.150166
MSE 2.654034 3.931020 3.126818 2.616015
CCC (R) 0.982176 0.973379 0.979070 0.982349
prediction using optimum number of processing elements

AARE 0.045191 0.071402 0.068001 0.051669
SD (σ) 0.052621 0.106760 0.066799 0.054290
MSE 1.024628 1.257030 1.526097 1.205827
CCC (R) 0.995802 0.995301 0.993728 0.995278
χ² 1.468583 2.168435 2.286890 1.651124
Table 5.6 Comparison of the ANN prediction using the best network and the
experimental data
Conc. of Liquid Gas flow ΔPtp/Lb

SCMC flow rate rate kPa/m
solution Ql Qg ANN
Experimental
kg/m3 m3/s m3/s prediction
1 2 3 4 5
0.000233 8.53E-05 9.844 9.59508
0.000399 0.000312 35.888 35.82978
0.50 0.000431 0.000163 33.838 34.15758
0.000431 0.000228 36.914 37.1266
0.000465 0.000209 41.836 41.49013
0.000266 0.00013 11.997 9.350024
0.000299 6.78E-05 10.561 10.1515
0.67 0.000332 0.000152 13.638 14.32898
0.000332 0.000356 21.02 20.17117
0.000399 0.000207 22.353 22.93654
0.00009 0.000257 3.076 3.27794
0.000115 0.000334 4.819 4.814786
0.83
0.000233 0.000257 11.792 11.16915
0.000299 6.45E-05 11.484 12.56602
90
1 2 3 4 5
0.0001 0.000153 6.152 6.078122
0.000162 0.000147 9.741 10.38693
0.000178 0.000213 12.817 12.66347
1.00 0.000233 0.000146 15.996 16.53258
0.000233 0.000429 22.558 22.16054
0.000299 0.000186 25.635 24.93481
0.000266 0.000363 22.046 25.12663
CHAPTER 6
Gas-non-Newtonian liquid flow through helical coils in

horizontal plane
Gas-non-Newtonian liquid flow through helical coils in horizontal plane
Helical coils are extensively used in different process industries. The two-
phase flow through the coils is more complex than that of straight pipe due to
presence of centrifugal forces; the flow through coils is always developing in nature.
The knowledge of frictional pressure drop and gas holdup is an important
hydrodynamic parameter used for the designing the coil. In the present chapter the
applicability of the Artificial Neural Networks (ANN) methodology has been
investigated using experimental data obtained from our earlier studies on the frictional
pressure drop for gas-non-Newtonian liquid flow through helical coils in horizontal
orientation. Multilayer Perceptron (MLP) trained with backpropagation algorithm
using four different transfer functions in a hidden layer is used in ANN.
6.1 Introduction
The ANN’s capability to predict the two-phase frictional pressure drop and
gas holdup across the coils is one of the best estimation methods with high accuracy
as evident from the previous three chapters. ANN with backpropagation algorithm has
been used to predict the gas holdup. In the present chapter application of ANN is used
to predict the two-phase frictional pressure drop and gas holdup for gas-non-
Newtonian liquid (SCMC) flow through different helical coils in horizontal
orientation based on the earlier experimental data (Biswas and Das, 2006; Biswas,
2007).
Experimental data has been collected from the earlier work carried out in our
laboratory (Biswas and Das, 2007; Biswas, 2007). Table 6.1 represents the range of
variables investigated for the prediction of both two-phase frictional pressure drop
and gas holdup.

92
physical variables include the diameter of the tube (Dt), diameter of the coil (Dc),
properties of the non-Newtonian liquid, i.e., flow behavior index (n'), consistency
index (K'), surface tension of the liquid (σl), density of liquid (ρl), density of gas (ρg),
viscosity of gas (µg) , acceleration due to gravity (g), while the operating variable is
liquid flow rate (Ql) and gas flow rate (Qg). The values of the physical properties of
gas, i.e. density and viscosity and acceleration due to gravity are constant so it is
ineffective for ANN programming.
Four different transfer functions given in Eqs. (2.1–2.4) are used in the hidden
and output layer for this analysis.
6. 3 Optimization of the ANN
Initially the total data is randomized for both two-phase frictional pressure
drop and gas holdup separately. The first 60% of data points are used for training, the
next 20% for cross-validation, the next 10% for testing and the rest used for
prediction.
set has been randomized before the network is used. In the hidden layer the numbers
of processing elements are optimized varying the number 1 to 25, and each case the
MSE is calculated. Figs 6.1 and 6.2 show the variation of MSE for cross-validation
with the number of processing elements in the hidden layer for both two-phase
frictional pressure drop and gas holdup respectively. The optimum number of
processing elements is that where the MSE is minimum. These optimum numbers of

93
Table 6.2 shows the optimum number of processing elements in the training
section of the ANN with different transfer functions. These optimum numbers of
Figs 6.3 and 6.4 represent the training curve for the above mentioned five
different piping systems for four different transfer functions used in the hidden layer.
Figs 6.5 and 6.6 represent the cross-validation curve for four different transfer
functions used in the hidden layer. The optimum number of processing element is
estimated on the basis of minimum value of MSE. Initially the MSE for each epoch
for training and cross-validation in all five cases are recorded for 5 different runs
separately. Each run consists of 32000 epochs. Then the average of the MSE of all the
five different runs is calculated for each of the 32000 epochs for each of the five
systems. Tables 6.3 and 6.4 represent the data for the minimum value of MSE for
training and cross-validation for both two-phase frictional pressure drop and gas
holdup respectively.
Testing was done just before the final prediction to check the effectiveness of
training using the above mentioned four networks. Tables 6.5 and 6.6 represent the
performance during testing for both two-phase frictional pressure drop and gas holdup
respectively. The low values of AARE and cross correlation co-efficient (R) which is
greater than 0.95, indicates that the training was good and the network can be used for
final prediction of frictional pressure drop. In case of gas holdup the low values has
been achieved for AARE and cross correlation co-efficient (R) which is nearly 0.90,
indicates that the training was acceptable but not as good as the prediction of the two-
phase frictional pressure drop and the network has been used for final prediction of
gas holdup. Figs 6.7 and 6.8 show the comparisons between the experimental to the
94
predicted two-phase frictional pressure drop and gas holdup with optimum number of
processing elements in the hidden layer for testing.
prediction of two-phase frictional pressure drop. It is clear from the table that the
cross correlation co-efficient (R) value is nearly 0.97 for each of the four cases for the
prediction of the two-phase frictional pressure drop. The low value of the average
absolute relative error (AARE) also shows the accuracy of the result in the different
systems.
prediction of gas holdup. The low value of the average absolute relative error (AARE)
also shows the accuracy of the result in the different systems. It is clear from the table
that the cross correlation co-efficient (R) value is nearly 0.90 for each of the four
cases for the prediction of the two-phase frictional pressure drop which is statistically
acceptable.
Since the cross correlation co-efficient value is nearly same for all the analysis
so the Chi-square (χ²) test is performed to find the best result. The lowest value of the
Chi-square (χ²) confirms that the best network for prediction of two-phase frictional
pressure drop is the network with transfer function 1 with 10 processing elements in
the hidden layer.
The Chi- square (χ2) value also confirms that the best network for prediction
of gas holdup is the transfer function 1 with 15 processing elements in the hidden
layer. This result indicates that the performance of the network output is excellent
Table 6.9 represents the performance of the network and its comparison to the
experimental data for the prediction of the pressure drop per unit length across the
horizontal coil with transfer function 1 and 10 processing elements in the hidden layer.
95
Table 6.10 represents the performance of the network and its comparison to the
experimental data for the prediction of the gas holdup in the horizontal coil with
transfer function 1 and 15 processing elements in the hidden layer.
To verify the accuracy of the result for the prediction of gas holdup a
comparison was necessary with the empirical correlation established by Biswas
(2007). The empirical correlation is represented below,
 0.9364

−0.2155  Dt 
αg = −0.8326
1 − exp −6.3029 Rel 0.4741
Re g N pl    (6.1)
  Dc  
The variance of estimate and correlation coefficient of the above equation are
0.024635 and 0.894101 respectively.
Table 6.11 shows the comparison between the experimental data with the
empirical correlation, i.e., Eq. (6.1) and the ANN predicted values. It is clear from the
table that ANN prediction is slightly better representation of the experimental data
than the empirical correlation.
6.5 Conclusions
Experimental data on the two-phase frictional pressure drop and gas holdup
for gas-non-Newtonian liquid flow through helical coils are collected from our earlier
studies to examine the applicability of ANN. A multilayer perceptron with
backpropagation algorithm consisting of four different transfer functions in a single
hidden layer and a linear transfer function in the output layer has been used for ANN
analysis. Initially the training is carried out to optimize the number of processing
elements in the hidden layer in all four cases by estimating the minimum value of
MSE. This optimum number of processing elements in the hidden layer is used for
subsequent analysis, i.e., for testing and prediction. In all cases the statistical
96
parameters like AARE, SD, MSE, R and χ² are used to examine the performance of
the network. The best ANN for the prediction of two-phase frictional pressure drop
for gas-non-Newtonian liquid flow through helical coils in horizontal orientation is
the transfer function 1 with 10 processing elements in the hidden layer. The Chi-
square (χ2) value also confirms that the best network for prediction of gas holdup is
the transfer function 1 with 15 processing elements in the hidden layer. The ANN
prediction of gas holdup gives slightly better predictability than the empirical
correlation.
97
0.020
Minimum value of MSE for cross-validation in case of two-phase

Transfer Function 1
Transfer Function 2
Transfer Function 3
Transfer Function 4
0.015
frictional pressure drop
0.010
0.005
10 20
elements in the hidden layer for two-phase frictional pressure drop across the coil
0.025
Transfer Function 1
Transfer Function 2
Transfer Function 3
Minimum value of MSE for cross-validation in case of
Transfer Function 4
0.020
gas holdup
0.015
0.010
0.005
10 20
elements in the hidden layer for gas holdup in the coil
98
101
Transfer Function 1
Transfer Function 2
Transfer Function 3
Average MSE for traning in case of two-phase frictional

Transfer Function 4
pressure drop in horizontal coil 100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
of epochs in case of two-phase frictional pressure drop across the coil
102
Transfer Function 1
Average MSE for traning in case of gas holdup in horizontal coil
Transfer Function 2
Transfer Function 3
Transfer Function 4
101
100
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
of epochs in case of gas holdup in the coil
99
101
Transfer Function 1
Transfer Function 2
Transfer Function 3
Average MSE for cross-validation in case of two-phase

Transfer Function 4
100
frictional pressure drop in horizontal coil
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
number of epochs in case of two-phase frictional pressure drop across the coil
102
Transfer Function 1
Transfer Function 2
Transfer Function 3
Average MSE for cross-validation in case of gas hold up in
Transfer Function 4
101
100
horizontal coil
10-1
10-2
10-3
1 10 100 1000 10000
Number of epochs
number of epochs in case of gas holdup in the coil
100
100
Transfer Function 1
Transfer Function 2
Two-phase frictional pressure drop accross the horizontal coil

Transfer Function 3
Transfer Function 4
predicted, ∆Ptp, kPa/m
10
1
1 10 100
experimental, ∆PHorizontal coil, kPa
Fig. 6.7 Comparison of two-phase frictional pressure drop across the horizontal coil
for testing with four different transfer functions
10
Transfer Function 1
Transfer Function 2
Transfer Function 3
Gas hold up in the horizontal coil, predicted
Transfer Function 4
1
αg
0.1
0.01
0.01 0.1 1 10
Gas holdup in the horizontal coil, experimental
αg
Fig. 6.8 Comparison of the gas holdup in the horizontal coil for testing with four
101
100
Transfer Function 1

Number of hidden layer processing elements: 10
predicted, ∆PHorizontal coil, kPa
10
1
1 10 100
experimental, ∆PHorizontal coil, kPa
Fig. 6.9 Comparison of two-phase frictional pressure drop across the horizontal coil
for final prediction
10
Transfer Function: 1
Numbar of hidden layer processing elements: 15
Gas holdup in the horizontal coil, predicted
1
αg
0.1
0.01
0.01 0.1 1 10
Gas holdup in the horizontal coil, experimental
αg
Fig. 6.10 Comparison of gas holdup in the horizontal coil for final prediction
102
Table 6.1 Range of different variables investigated for horizontal helical coils

Coil Properties
Tube diameter (m) 0.00904 ≤ Dt ≤ 0.0118
Coil diameter (m) 0.1732 ≤ Dc ≤ 0.2680
Consistency index (Nsn’/m2) 0.0142 ≤ K/ ≤ 0.7112
Density (kg/m3) 1001.69 ≤ ρ ≤ 1003.83
Liquid property 0.69685 × 10-7 ≤ Npl ≤ 0.11146 × 10-3
Raynolds number 50 ≤ Rel ≤ 2100
Surface tension (N/m) 0.07834 ≤ σl ≤ 0.0832
Raynolds Number 100 ≤ Reg ≤ 3500
Viscosity of air (Ns/m2) 0.00001846
Flow rate
5 3
Gas flow rate Qg×105(m3/s) 0.70 to 56.13
Measuring parameter
Two-phase frictional pressure drop (kPa/m) 2.8775 ≤ ΔPtp/L ≤ 33.0103
Gas holdup 0.071 ≤ αg ≤ 0.880
Total number of data points 766
Table 6.2 Optimum number of processing elements used in hidden layer for four
Optimum number of
Transfer function in
processing elements
hidden layer
ΔPftc/L αg
Transfer function 1 15 10
103
training and cross-validation over 5 runs for two-phase frictional pressure drop

Training
SD (σ) 0.000494 0.000146 0.000207 0.000302
MSE 0.007560 0.006567 0.007021 0.008393
Cross-validation
SD (σ) 0.001918 0.000335 0.000493 0.000305
MSE 0.009002 0.008427 0.007746 0.007944
training and cross-validation over 5 runs for gas holdup

Training
SD (σ) 0.000178 0.000073 0.000075 0.000145
MSE 0.008316 0.007682 0.007876 0.008454
Cross-validation
SD (σ) 0.000269 0.000277 0.000068 0.000089
MSE 0.008679 0.008457 0.008642 0.008990
Table 6.5 Performance of neural network with different transfer functions for two-
phase frictional pressure drop during testing

AARE 0.135057 0.127849 0.146226 0.144132
SD (σ) 0.125134 0.124990 0.122872 0.128555
MSE 7.565240 7.665425 8.589985 8.199548
CCC (R) 0.956905 0.956572 0.951520 0.953505
Table 6.6 Performance of neural network with different transfer functions for two-
phase gas holdup during testing

AARE 0.103307 0.104937 0.096366 0.090723
SD (σ) 0.091525 0.102116 0.099003 0.080880
MSE 0.004515 0.004898 0.004083 0.003922
CCC (R) 0.909165 0.899457 0.917325 0.920342
104
Table 6.7 Performance of neural network with different transfer functions for the
prediction of two-phase frictional pressure drop across the helical coil

AARE 0.137721 0.136747 0.144295 0.153650
SD (σ) 0.113812 0.116137 0.116709 0.130063
MSE 3.603808 3.548291 3.554401 3.957043
CCC (R) 0.971994 0.972468 0.971562 0.968119
χ² 19.98049 20.06086 20.83103 22.85247
Table 6.8 Performance of neural network with different transfer functions for the
prediction of gas holdup in the helical coil

AARE 0.121936 0.117483 0.125479 0.127843
SD (σ) 0.141958 0.131088 0.146448 0.142306
MSE 0.004746 0.005220 0.006145 0.005636
CCC (R) 0.921583 0.911284 0.894034 0.903627
χ² 0.890912 0.915204 1.016782 1.002043
105
Table 6.9 Comparison of the prediction using the best network and the experimental
data for pressure drop per unit length
Two-phase frictional
Conc.
Tube Coil Liquid Gas flow pressure drop per unit
of
Diameter Diameter flow rate rate length of the coil, ΔPtp/L
SCMC
Dt Dc Ql Qg kPa/m
solution 3
m m m /s m3/s ANN
prediction
1 2 3 4 5 6 7
6.75E-05 4.55E-05 20.0003 17.11244
0.2134
6.75E-05 0.00002 17.4953 17.22868
4.75E-05 0.000138 19.1219 18.92184
0.00904 6.75E-05 0.000075 21.8309 19.6315
0.2633 0.000055 0.000131 21.7778 19.00992
0.00006 4.32E-05 19.5469 19.88434
0.2 0.000075 0.000043 23.4244 19.88572
0.0001 8.63E-05 7.8527 7.570498
0.2158
0.0001 0.000139 8.6984 8.036368
0.0096
0.0001 6.42E-05 8.0225 8.485858
0.2657
0.0001 0.00012 9.4601 8.507354
0.000133 0.000171 4.8798 5.03982
0.0118 0.2177
0.000133 0.000267 5.6205 5.856673
4.75E-05 5.21E-05 12.3263 16.93229
0.2134
0.00904 4.75E-05 0.000174 17.4953 16.52792
0.2633 4.25E-05 0.000221 22.7392 18.40749
0.000117 0.000101 9.7374 7.666817
0.0096 0.2158
0.000117 0.000139 10.8247 8.32775
6.67E-05 0.000214 8.1616 8.739294
0.0096 0.2657 0.000117 9.15E-05 10.6658 8.501113
0.0001 0.000167 10.573 8.792647
0.4
0.0001 0.000156 4.4252 5.663768
0.1776 0.000167 6.01E-05 4.3598 4.580421
0.0002 0.000233 7.5206 6.352849
0.0001 3.89E-05 3.0063 4.510643
0.0118
0.2177 0.000167 0.000242 7.5157 6.257399
0.000167 0.000252 7.6246 6.295989
0.000102 0.000273 8.2921 7.591236
0.268
0.000121 1.72E-05 6.2934 5.050896
4.75E-05 0.000154 30.1774 26.47533
0.6 0.00904 0.1732 6.75E-05 0.000045 32.5177 28.41164
0.000075 0.000012 30.7932 28.89526
106
1 2 3 4 5 6 7
6.75E-05 3.75E-05 20.3582 16.96682
0.2134
0.000055 6.42E-05 16.2229 17.16951
4.25E-05 0.000255 19.7062 18.62569
0.00904 6.75E-05 6.85E-05 22.8401 19.64527
0.2633 6.75E-05 7.19E-05 22.9463 19.64611
0.000055 0.000101 19.0157 19.62539
0.00006 0.000113 21.0341 19.59966
0.0001 2.48E-05 5.2097 5.7776
0.1757
0.000117 6.49E-05 8.4976 6.790514
0.0096
0.000117 0.0001 9.7132 8.221755
0.2158
0.0001 8.44E-05 7.4178 7.806595
0.6
0.0001 0.000518 5.2971 6.004758
0.000133 0.000101 3.7058 5.16314
0.000133 0.000237 4.8394 6.15517
0.1776
0.000133 0.000326 5.6895 6.42475
0.000167 7.83E-05 4.665 4.931507
0.0118 0.0002 0.000226 7.477 6.099578
6.67E-05 0.000204 5.664 6.323518
0.0001 3.15E-05 2.9627 4.584911
0.2177
0.000133 0.000155 5.3155 5.978301
0.0002 0.0001 7.0582 5.448566
0.268 8.52E-05 1.89E-05 3.6734 4.931798
4.25E-05 0.000148 27.7139 27.18926
0.1732
0.000075 3.24E-05 30.1774 28.21439
0.00904 4.75E-05 4.04E-05 14.2348 16.93185
0.2134
4.25E-05 0.000139 16.8989 18.02505
0.2633 4.25E-05 3.54E-05 12.6417 19.14337
4.75E-05 7.42E-05 16.413 19.34961
0.000055 6.37E-05 17.2629 19.30143
0.000055 8.65E-05 19.1751 19.39887
0.00904 0.2633
0.000055 0.000105 19.6 19.45696
0.00006 9.89E-05 21.0341 19.44067
0.000075 1.61E-05 18.1658 19.0119
0.8 0.0001 0.000152 8.5208 10.32356
0.1757 0.000117 3.12E-05 7.0157 6.433031
0.000117 0.00013 10.1416 9.616772
0.0096
0.0001 0.00013 8.6984 9.885412
0.2158
0.000117 0.000141 10.873 10.20156
0.2657 8.34E-05 0.000065 6.9791 8.521074
0.1776 0.0001 0.000512 5.4061 5.720164
0.00005 0.000124 5.359 5.455639
0.2177
0.000167 0.000168 6.3176 5.946173
0.0118
8.52E-05 6.97E-05 4.6728 5.408167
0.268 8.52E-05 0.000036 4.1056 4.720887
0.000121 0.000137 8.2111 7.726494
107
Table 6.10 Comparison of the prediction using the best network and the experimental
data for gas holdup
Conc. of Tube Coil Liquid Gas flow Gas holdup

SCMC Diameter Diameter flow rate rate αg
solution Dt Dc Ql Qg ANN
3 Experimental
(kg/m ) m m m3/s 3
m /s prediction
1 2 3 4 5 6 7
4.75E-05 5.75E-05 0.635 0.563199
0.1732 0.000075 1.75E-05 0.422 0.423063
0.000075 3.14E-05 0.445 0.47665
4.75E-05 6.12E-05 0.619 0.574274
0.2134
0.00904 4.75E-05 0.000045 0.51 0.526086
6.75E-05 0.000075 0.36 0.440606
0.2 0.00006 0.00013 0.52 0.557586
0.2633
0.000075 2.68E-05 0.22 0.279637
0.000075 0.000043 0.307 0.340315
0.0096 0.1757 0.0001 2.81E-05 0.211 0.291583
0.1776 0.000133 3.83E-05 0.271 0.370139
0.0118 0.2177 0.000133 0.000039 0.28 0.493908
0.268 0.000102 0.000116 0.321 0.37952
0.000055 0.00011 0.639 0.640127
0.1732
0.00006 9.09E-05 0.591 0.603486
4.25E-05 6.66E-05 0.62 0.546009
0.2134 4.25E-05 0.000102 0.678 0.628165
0.00904
0.00006 0.00003 0.408 0.423604
6.75E-05 8.06E-05 0.445 0.426565
0.2633 0.000055 0.000105 0.5 0.489786
0.4 0.000075 3.35E-05 0.249 0.260928
0.1757 8.34E-05 1.86E-05 0.21 0.193298
0.0096
0.2657 0.000117 5.71E-05 0.259 0.310654
0.1776 0.0001 0.000367 0.817 0.705253
6.67E-05 0.000106 0.848 0.529612
0.2177
0.0118 0.000167 0.000252 0.679 0.705417
7.28E-05 0.000473 0.797 0.730394
0.268
8.52E-05 1.39E-05 0.18 0.125571
4.25E-05 7.48E-05 0.682 0.634179
4.25E-05 8.13E-05 0.66 0.64959
0.6 0.00904 0.1732
6.75E-05 0.000037 0.566 0.522697
0.000075 2.48E-05 0.458 0.478159
108
1 2 3 4 5 6 7
4.25E-05 4.32E-05 0.569 0.524635
4.25E-05 9.01E-05 0.679 0.661019
0.2134 0.000055 3.81E-05 0.471 0.505612
0.000055 0.000105 0.704 0.691303
0.00006 0.000039 0.571 0.509028
0.00904
4.75E-05 0.000088 0.48 0.439095
4.25E-05 0.000156 0.592 0.553482
0.2633 4.25E-05 0.000255 0.654 0.604615
0.000055 0.00017 0.562 0.566292
0.6
0.000075 6.09E-05 0.347 0.361368
0.0001 0.000115 0.59 0.499135
0.2158
0.0096 0.000117 0.000448 0.46 0.457069
0.2657 0.000117 6.07E-05 0.264 0.334598
0.0001 0.000273 0.704 0.611553
0.1776 0.0002 0.000132 0.407 0.46088
0.0118 0.0002 0.000226 0.485 0.571823
8.52E-05 0.000378 0.711 0.641024
0.268
0.000102 0.000046 0.3 0.259296
4.75E-05 6.56E-05 0.639 0.607206
4.25E-05 9.43E-05 0.663 0.676619
0.1732 6.75E-05 4.92E-05 0.53 0.557105
6.75E-05 4.94E-05 0.542 0.557764
0.8 0.00904 0.00006 0.000049 0.583 0.556445
4.75E-05 0.000125 0.719 0.741818
0.2134 0.000055 4.16E-05 0.548 0.525475
0.00006 4.22E-05 0.571 0.527817
0.2633 4.75E-05 0.000063 0.451 0.406064
6.75E-05 5.76E-05 0.347 0.389003
0.00904 0.2633
0.00006 9.89E-05 0.482 0.497864
0.000117 0.000117 0.434 0.470408
0.1757
0.0096 0.0001 6.06E-05 0.344 0.335503
0.2158 0.0001 1.83E-05 0.1 0.17013
0.0001 0.000317 0.722 0.600452
0.000133 0.00004 0.266 0.138706
0.1776
0.000133 0.000374 0.675 0.654255
0.000167 0.000193 0.475 0.470498
0.8
0.0001 0.000235 0.665 0.712248
0.000133 3.02E-05 0.29 0.250217
0.0118 0.2177 0.000133 0.000161 0.553 0.626388
0.000167 0.000225 0.658 0.703022
0.0002 7.12E-05 0.39 0.411924
0.268 8.52E-05 6.97E-05 0.319 0.295814
8.55E-05 0.000525 0.754 0.624617
0.268 0.000121 0.000032 0.18 0.19242
0.000121 0.000089 0.28 0.172305
109
Table 6.11 Comparison of results for the prediction of gas holdup using best neural
network and output calculated using Eq. (6.1)
Calculated
Measurement Best neural
from Eq.
Type network
(6.1)
AARE 0.121936 0.140263
SD (σ) 0.141958 0.156954
MSE 0.004746 4.185085
CCC (R) 0.921583 0.894101
CHAPTER 7
Prediction of flow regimes for air-water flow through

small diameter tube
Prediction of flow regimes for air-water flow through small diameter tube
Artificial Neural Network (ANN) based techniques for the classifications of
flow regimes in air-water flow through 1 mm to 5 mm tubes are presented. 218 data
points are based on the experimental investigation in 3 mm and 4 mm tubes and 2114
data points from various experimental results from the published literature for air-
water two-phase flow in small diameter tubes have been used. Five different well
known artificial neural network models have been used to predict the flow regime.
The ANN model based Radial Basis Function gives the slightly better predictability
over the other networks used.
7.1 Introduction
Gas-liquid two-phase flow widely exists in industrial processes, such as,
power generation, thermal engineering, chemical processes, petroleum refineries,
petrochemical industries, nuclear industries and many other heat transfer equipment
with phase change. In the two-phase gas-liquid flow the hydrodynamics, heat transfer
and mass transfer characteristics are largely influenced by the flow regimes. In many
industrial systems like power generation, nuclear reactors, petroleum and biochemical
processing systems, it is necessary to monitor the flow regimes during normal and
also transient operation conditions for the safety and overall performance. So the
accurate prediction of the flow regime is an important task.
Gas–liquid two-phase flow patterns in small channels have been investigated
extensively in the past (Suo and Griffith, 1964; Barnea et al., 1983; Barajas and
Panton, 1993; Fukano and Kariyasaki, 1993; Mishima et al. 1995; Mishima and
Hibiki, 1996; Triplett et al., 1999; Coleman and Garimella, 1999; Yang and Shieh,
2001; Zhao and Bi, 2001; Chen et al., 2002; Ide et al., 2007; Venkatesan et al., 2010).
Definitions of flow regimes are generally based on the description and graphical
illustrations. The visual flow pattern classification is based on visual observation,

111
which probably varies for different viewer particularly in the transition from one flow
regime to another. Many literatures are available on the development of different
techniques, such as conductivity probe, radiation attenuation, hot wire anemometer
etc. for flow regime identification. Jones and Zuber (1975) measured void fraction
fluctuations by a linearized X-ray void measurement system and the data used to
calculate the probability density function (PDF) of fluctuations in void fraction that
had helped to identify the main three flow regimes, i.e., bubbly, slug and annular
flow. Some researchers also conducted differential pressure measurement to identify
the flow regimes (Yang and Shieh, 2001; Chen et al., 2002; Ide et al., 2007;
Venkatesan et al., 2010; Akbar et al., 2003; Damianides and Westwater, 1988;
Ghiaasiaan and Abdel-Khalik, 2001; Lowe and Rezkallah, 1999; Rezkallah, 1996;
Serizawa and Feng, 2001; Zhao and Rezkallah, 1993). However, these two
techniques, i.e., the volumetric void fraction and pressure fluctuations were dependent
on each other and/or vary with the test location. Some researchers also proposed
mechanistic models for each transition separately in the flow regimes as these
transitions were associated with different mechanism. But unfortunately different
researchers proposed different equations to identify the same transition and most of
these are associated with the air-water flow through pipelines of different diameters
and inclinations (vertical, horizontal and inclined). Early work on the empirical flow
regime map developed by Baker (1954) and other researchers (Mandhane et al., 1974;
Taitel and Dukler, 1976; Weisman et al., 1979) had attempted to develop a
generalized flow regime map. Taitel and Dukler (1976) proposed a flow regime map
containing predominant four flow regimes namely, stratified, intermittent, bubble and
annular with superficial gas velocity and superficial liquid velocity as co-ordinates
and also proposed physical based criteria for the transition from one regime to the
112
next. Taitel (1990) proposed a unified model for predicting the flow regime
transitions in channels of any orientation based on simple physical criteria using well
known two-phase dimensionless group, Froude number (Fr), the parameter T which
related the liquid pressure drop to buoyancy, and Lockhart-Martinelli parameter, X.
They also developed the regime transition criteria which had reflected the distinction
between surface tension driven flow regimes, such as bubble and intermittent flow
and the shear force driven flow regimes, such as stratified and annular flow. The
models contained very little empiricism so it could be applied to most of fluids and be
extrapolated to conditions other than air-water and steam-water flow and almost all
diameter pipelines. Fukano et al. (1989) studied the flow patterns in capillary tubes
for air-water flow with diameters ranging from 1.0 to 4.9 mm and compared the maps
with the Mandhane et al. (1974) flow regime map. They identified the flow pattern
but did not provide any flow pattern transition lines. Coleman and Garimella (1999)
pointed out after analyzing the experimental data from literature for the flow pattern
that Mandhane et al. (1974) could not predict the flow regime transitions for small
diameter tubes, also the theoretical prediction of Taitel and Dukler (1976) and the
correlations presented by Weisman et al. (1979) were not applicable for small
diameter tubes.
Suo and Griffith (1964) showed the elongated bubble flow pattern in capillary
tubes with diameters ranging from 1.0 to 1.6 mm. They identified and attempted to
correlate the transition lines from elongated bubble to annular and bubbly flow by
using the average volumetric flow of the liquid and gas phases and the velocity of the
bubbles. Barnea et al. (1983) visually observed the two-phase flow patterns in small
horizontal tubes with diameters ranging from 4.0 to 12 mm and classified these flow
patterns according to the four major regimes, i.e., dispersed, annular, intermittent and
113
stratified. They also observed that all transitions except the stratified to non-stratified
transition were satisfactorily described by the Taitel and Dukler (1976) model.
Damianides and Westwater (1988) investigated air-water flow pattern for horizontal
glass tubes of inner diamater in the range 1.0 to 5.0 mm. They observed that there
were pronounced effect on the flow regime transition with the diameter. Taitel and
Dukler (1976) model predicted for the dispersed bubble-intermittent boundary and
annular-intermittent boundary but not for stratified flow. They also observed that the
surface tension was a very important variable for small diameter tubes but less for
large tubes. Fukano and Kariyasaki (1993) investigated air-water flow through
horizontal and vertical tubes with inner diameter vary from 1.00 to 9.00 mm. The
transition boundaries of their results agreed well with the experimental data obtained
by various researchers (Barajas and Panton, 1993; Triplett et al., 1999; Mishima et al.,
1993). The stratified flow pattern was not observed by the investigators. Triplett et al.
(1999) studied air-water flow through 1.1 and 1.45 mm diameter tubes. Their
observation was consistent with the results of Fukano and Kariyasaki (1993) and
Daminanides and Westwater (1988). They concluded that the surface tension is
predominant in capillaries. Coleman and Garimella (1999) investigated air-water flow
patterns in tube with diameter 1.3 to 5.5 mm. Yang and Shieh (2001) studied the flow
pattern of air-water and R-134a refrigerant in tubes with inner diameter from 1.0 to
3.0 mm. They observed six flow regimes, bubble, plug, slug, wavy stratified,
dispersed and annular flow in their studies. However, the transitions for the refrigerant
R-134a were very sharp and clear than the air-water flow. They also observed that
their results agreed well with those by Damianides and Westwater (1988) and
concluded that the surface tension force be an important parameter for flow pattern
determination. Hassan et al. (2005) investigated air-water flow through 800 µm, 1.0
114
mm and 3 mm diameter glass tubes and observed the surface tension dominated flow.
Ide et al. (2007) investigated air-water flow through capillary tubes with diameter 1.0
to 4.9 mm. They observed bubble, intermittent and annular flow regime in the
capillaries. Lee and Lee (2008) studied air-water and air-methanol flow through glass
tubes of 1.46 mm and 1.8 mm diameter, teflon tube of 1.59 mm diameter and
polyurethane tube of 2 mm diameter and observed plug, slug, annular and stratified
flow regime. Venkatesan et al. (2010) investigated air-water flow through 0.6 to 3.4
mm diameter glass tubes. They visualized the flow pattern by using high speed
camera and presented flow pattern maps for each tubes. They observed unique flow
patterns for different tube diameter and confirmed the diameter effect on the flow
pattern.
The present study deals with the creation of data bank from the data collected
from the flow regime studies conducted by various researchers including our own
related to air-water flow and the prediction of the flow patterns using 5 different ANN
structures, i.e., two types of Multilayer Perceptrons (MLP) trained with
Backpropagation (BP) algorithm and Levenberg Marquardt (LM) algorithm
respectively, Radial Basis Function (RBF), Support Vector Machine(SVM) and
Principal Component Analysis (PCA) in circular tubes of diameters ranging from 1–5
mm in horizontal plane.
7.2 Data Collection
The data are collected from the experimental investigation and also from the
literature (Barnea et al., 1983; Triplett et al., 1999; Coleman and Garimella, 1999;
Yang and Shieh, 2001; Chen et al., 2002; Ide et al., 2007; Venkatesan et al., 2010) to
predict the different flow regimes for air-water gas-liquid flow through small diameter
tubes. The other experimental data are collected from literatures and total 2332 data
115
points are used for ANN prediction. The input parameters and their ranges are given
below,
1) Gas flow rate, Ug varies from a minimum of 0.016 m/s to a maximum
of 98.3 m/s;
2) Liquid flow rate Ul varies from a minimum of 0.00155 m/s to a
maximum of 9.64 m/s;
3) Diameter of the tube, D varies from a minimum of 0.001 m to a
maximum of 0.005 m.
Table 7.1 summarizes the data used for ANN studies. Table 7.1 consists of the 15
different types of flow regimes presented by the various authors. The ANN analysis
with 15 outputs can increase the complexity i.e., very long duration of training time is
required.
Annular flow regime (A), Slug-Annular flow regime (SA) and Wavy Annular
flow regime (WA) as described by various authors in their individual papers are taken
as Annular flow regime (A) for this study.
The Bubble or Bubbly flow regime (B), Dispersed Bubbly flow regime (DB or
DiB) and Bubble-train Slug flow regime (BTS) as described by various authors in
their individual papers are taken as Bubbly flow regime (B) for this study.
Intermittent flow regime (I), Slug flow regime (Sl), Plug flow regime (P) and
Elongated Bubble flow regime (EB) as described by various authors in their
individual papers are taken as Intermittent flow regime (I) for this study.
Stratified Wavy flow regime (SW) and Stratified Smooth flow regime (SS) as
described by various authors in their individual papers are taken as Stratified flow
regime (S) for this study.

116
The Churn flow regime (C), Dispersed flow regime (Di) and Wavy flow
regime (W) are left as these are without any change for this study.
Therefore, for this study the 15 types of flow regimes are modified to 7 to
reduce the complexity, i.e., duration of training for the ANN analysis.
In general, researchers used either normalized data or raw data for ANN
analysis. From our previous experience (Bar and Das 2011, 2012a,b, 2013a,b) here
only the raw data is used for ANN analysis. The total data is first randomized for five
different samples to eliminate sampling error. Then each of these five samples is
analyzed (training, cross-validation and final prediction) for all five different neural
networks as discussed earlier. The output is represented by 7 columns corresponding
to 7 different flow patterns. Then the output is translated from symbolic data to
numeric data, i.e., into 7 columns consisting of 0 and 1, where each input row
corresponding to a particular air velocity, water velocity and tube diameter is
represented in the output with 6 values of number 0 and one value of number 1. The
row having number 1 corresponds to that flow pattern.
For the comparison of the performance of the ANN prediction the followings
are considered,
1) Air-water flow only
2) Circular tubes are having diameter ranging from 1–5 mm only
3) Horizontal orientation of tubes only
4) No normalization is done for the overall data for any analysis
5) MLPs, RBF and PCA are having only one hidden layer
6) Hyperbolic tangent function is used in all the hidden and output
layers of respective ANNs

117
7) The number of processing elements are varied from 1–25 only for all
cases
8) The amount of data for training, cross-validation and prediction (Table
7.4) are kept same for all the ANN analysis
9) Only one computer with same architecture is used for all analysis
7.3 Results and Performance
The total data is first randomized for five different samples to eliminate
sampling error. Then each of these five samples is analyzed (training, cross-validation
and final prediction) for all five different neural networks, i.e., two Multilayer
Perceptrons, Radial Basis Function, Support Vector Machine and Principal
Component Analysis network. The output is represented by 7 columns corresponding
to 7 different flow patterns. Then the output is translated from symbolic data to
numeric data, i.e., into 7 columns consisting of 0 and 1, where each input row
corresponding to a particular air velocity, water velocity and tube diameter is
represented in the output with 6 values of number 0 and one value of number 1. The
row having number 1 corresponds to that flow pattern.
A Neurosolution 5.07, on a computer with Intel Core i7 Processor (2.8 GHz),
Intel DP55KG Motherboard (1333 MHz), 16 GB of DDR3 RAM (1333 MHz), ATI
RADEON HD Graphics card of 1 GB DDR5 RAM with 64 Bit Windows 7 Ultimate
operating system are used for this analysis.
The total of 2332 data points have been used for ANN analysis. Out of this
1981 data points are used for training, 234 data points for cross-validation and 117
data for final prediction. Figs. 7.1–7.5 represent the data used for training, cross-
validation and final prediction respectively for Sample 1–5.

118
The numbers of processing elements are varied from 1 to 25 in a single hidden
layer. Table 7.2 present the minimum value of MSE for cross-validation for the 5
different networks. From the values of MSE in Table 7.2 it is clear that the desired
value of minimum MSE, i.e., 0.01 for cross-validation, which is also another criterion
to stop training has never been reached throughout the training and cross-validation
process by any network. The network with the particular number of processing
elements in the single hidden layer recorded the minimum value of cross-validation
MSE is further used for the final prediction. This is the procedure that was followed
for all the five different samples.
Table 7.3 presents the optimum number of processing elements in the hidden
layer for Multilayer Perceptrons, Radial Basis Function and Principal Component
Analysis network during training.
7.3.1 Backpropagation (BP)
For the training of MLP network with Backpropagation algorithm a maximum
of 32000 epochs are used. For the hidden layer of BP network the value of learning
rate α is 1 and that of momentum coefficient µ is 0.7. For the output layer of BP
network the value of learning rate α is 0.1 and that of momentum coefficient µ is 0.7.
However, training is stopped when we have observed no improvement in cross-
validation MSE for 10000 consecutive epochs. It takes a minimum of 16 hours 28
minutes to a maximum of 18 hours 32 minutes to train the five samples using MLP
network trained with BP algorithm. The 1st column of Tables 7.2 and 7.3 present the
minimum value of cross-validation MSE reached and the optimum number of
processing elements respectively.

119
Fig. 7.6 is the comparison between the minimum value of MSE for cross-
validation and the number of processing elements in the hidden layer for all the five
samples of the MLP network where the weights are updated using Backpropagation
algorithm.
Fig. 7.7 is the cross-validation curve for all the five samples of the MLP
network where the weights are updated using Backpropagation algorithm during
which the cross-validation MSE reaches its minimum value for all 5 samples.
7.3.2 Levenberg-Marquardt (LM)
For the training of MLP network with LM algorithm a maximum of 1000
epochs have been used. For the network with LM algorithm the initial value of λ is
0.01. However, training is stopped when no improvement in cross-validation MSE for
100 consecutive epochs is observed. The numbers of processing elements are varied
from 1 to 25 in a single hidden layer. It takes a minimum of 11 hour to a maximum of
12 hours 7 minutes to train the five samples using MLP network trained with LM
algorithm.
Fig. 7.8 presents the comparison between the minimum value of MSE for
cross-validation and the number of processing elements in the hidden layer for all the
five samples of the MLP network where the weights are updated using Levenberg-
Marquardt algorithm. The 2nd column of Tables 7.2 and 7.3 present the minimum
value of cross-validation MSE reached and the optimum number of processing
elements respectively.
Fig. 7.9 present the cross-validation curve for all the five samples of the MLP
network where the weights are updated using Levenberg-Marquardt algorithm during
which the cross-validation MSE reached its minimum value for all 5 samples.
120
7.3.3 Radial Basis Function (RBF)
For the training of RBF network, 1000 epochs are set for unsupervised
learning and maximum of 1000 epochs for supervised learning. The numbers of
clusters (k value) are kept 15. However, training is stopped when no improvement in
cross-validation MSE for 100 consecutive epochs during supervised learning phase
has been observed. For the supervised learning, Levenberg-Marquardt algorithm is
used for updating the weights. It takes a minimum of 77 hours 51 minutes to a
maximum of 89 hours 19 minutes to train the five samples using RBF network. The
3rd column of Tables 7.2 and 7.3 present the minimum value of cross-validation MSE
reaches and the optimum number of processing elements respectively.
five samples of the RBF network where the weights are updated using Levenberg-
Marquardt algorithm.
Fig. 7.11 presents the cross-validation curve for all the five samples for RBF
which the cross-validation MSE reaches its minimum value for all 5 samples at the
supervise learning phase.
7.3.4 Principal Component Analysis (PCA)
There are 3 principal components kept for this analysis. For training of PCA
network with three principal components, 1000 epochs have been used for
unsupervised learning and also 1000 epochs for supervised learning. However,
training is stopped when there is no improvement in cross-validation MSE for 100
consecutive epochs during supervised learning phase. For the supervised learning LM
algorithm is used for updating the weights. It takes a minimum of 21 hours 57 minutes
121
to a maximum of 31 hours 15 minutes to train the five samples using PCA network.
The 4th column of Tables 7.2 and 7.3 present the minimum value of cross-validation
MSE reaches and the optimum number of processing elements respectively.
five samples of the PCA network where the weights are updated using Levenberg-
Marquardt algorithm.
Fig. 7.13 presents the cross-validation curve for all the five samples for PCA
which the cross-validation MSE reaches its minimum value for all 5 samples at the
supervise learning phase.
7.3.5 Support Vector Machine (SVM)
For the training with SVM network 1000 epochs are used. However, training
is stopped when no improvement in cross-validation MSE for 100 consecutive epochs
is observed. It takes a minimum of 34 minutes to a maximum of 2 hours 15 minutes to
train the five samples using SVM network. Fig. 7.14 presents the cross-validation
curve for all the five samples for SVM network where the weights are updated using
Levenberg-Marquardt algorithm. The training for all the five curves overlap each
other as evident from the curve. The gradual decrease shows that training is good.
Near the end of the training, the upward bend of the curves and its abrupt end shows
the termination of training as the increase or no improvement of MSE continues for
100 epochs. The gradual decrease in the value of average MSE over 5 runs proved the
effectiveness of the training.

122
7.4 Comparison of ANN Presentation with experimental flow regime data
Table 7.4 shows the performance of the five different ANNs to predict the
flow patterns. Table 7.5 shows the comparison of the experimental to the predicted
output for the analysis done with the sample 3 using PCA network. The last column of
Table 7.5 shows that the network has predicted 106 data correctly. Similar types of
procedures are followed for the evaluation of the output of the other analysis and the
results are depicted in Table 7.4. The final predicted flow pattern result is identified to
be the number closest to 1. The predicted flow patterns are compared with the
respective experimental data.
The accuracy of the prediction is affected by the following reasons;
1) Overlapping of data points as indicated by the experimental observations
outside the region marked by the boundary lines.
2) Some data points that lay on the boundary lines of the different flow regimes
as observed by the researchers and represented in the graph.
From the mean calculated in Table 7.4, it is clear that the Radial Basis Function gives
slightly better predictability over the other networks. But the time may be a factor as it
is clearly seen that for the prediction using PCA network time is approximately 3
times lesser than that of the training with RBF network, which is an advantage where
the analysis is done in computer with lesser speed.
7.5 Conclusions
Experimental observation using air-water two-phase flows in 1 – 5 mm tubes
to identify the different flow regimes depicted in the literature have been collected.
Applicability of different neural network models, i.e., Multilayer Perceptrons, Radial
Basis Function, Support Vector Machine and Principal Component Analysis network
have been used to predict the flow pattern. The Radial Basis Function gave slightly
123
better predictability over the other networks if time is not considered an important
factor.
124
Data for Traning

Data for Cross-validation
Data for Final Prediction
0.005
Sample 1
0.004
0.003
Diameter (m)
0.002
1E-3 10
0.01
/s)
1 y (m
Sup 0.1 it
erfi
cial eloc
1 V
Air
velo 0.1 ater
w
city al
(m/ 10 rfici
s) pe
100 0.01 Su
Fig. 7.1 Graphical representation of data for sample 1

125
Data for Traning

0.005
Sample 2
0.004
0.003
Diameter (m)
0.002
1E-3 10
0.01
/s)
1 y (m
Sup 0.1 cit
erfi elo
cial 1 V
Air
velo 0.1 ater
w
city
(m/ 10 ial
s) erfic
p
100 0.01 Su

126
Data for Traning

0.005
Sample 3
0.004
0.003
Diameter (m)
0.002
1E-3 10
0.01
/s)
1 y (m
Sup 0.1 cit
erfi elo
cial 1 V
Air
velo 0.1 ater
w
city
(m/ 10 ial
s) erfic
p
100 0.01 Su

127
Data for Traning

0.005
Sample 4
0.004
0.003
Diameter (m)
0.002
1E-3 10
0.01
/s)
1 y (m
Sup 0.1 it
erfi
cial eloc
1 V
Air
velo 0.1 ater
w
city al
(m/ 10 rfici
s) pe
100 0.01 Su

128
Data for Traning

0.005
Sample 5
0.004
0.003
Diameter (m)
0.002
1E-3 10
0.01
/s)
1 y (m
Sup 0.1 it
erfi
cial eloc
1 V
Air
velo 0.1 ater
w
city al
(m/ 10 rfici
s) pe
100 0.01 Su

129
2.0x10-1
Network: Multilayer Perceptron
Algorithm: Backpropagation
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
Minimum value of MSE for cross-validation
1.5x10-1
1.0x10-1
5.0x10-2
5 10 15 20 25
Number of processing elements in hidden layer
Fig. 7.6 Variation of minimum value of cross-validation with the number of

backpropagation algorithm for all 5 samples
4x10-1
3.5x10-1
3x10-1
2.5x10-1
2x10-1
1.5x10-1
10-1

Algorithm: Backpropagation
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
5x10-2
1 10 100 1000 10000
Fig. 7.7 Cross-validation curve for MLP network trained with

BP algorithm for all 5 samples
130
2.0x10-1
Algorithm: Levenberg-Marquardt
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
1.5x10-1
1.0x10-1
5.0x10-2
5 10 15 20 25

processing elements in the hidden layer for MLP network trained with Levenberg-
Marquardt algorithm for all 5 samples
1

0.1 Algorithm: Levenberg-Marquardt
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
1 10 100
Number of epochs
Fig. 7.9 Cross-validation curve for MLP network trained with

LM algorithm for all 5 samples
131
1.5x10-1
Network: Radial Basis Function
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
1.0x10-1
5.0x10-2
5 10 15 20 25

processing elements in the hidden layer for RBF network trained with Levenberg-
Marquardt algorithm for all 5 samples
1
0.1
Network: Radial Basis Function

Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
1 10 100
Number of epochs
Fig. 7.11 Cross-validation curve for RBF network trained with

LM algorithm for all 5 samples
132
1.5x10-1
Network: Principal Component Analysis
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
1.0x10-1
5.0x10-2
5 10 15 20 25

processing elements in the hidden layer for PCA network trained with Levenberg-
Marquardt algorithm for all 5 Samples
1
0.1
Network: Principal Component

Analysis
Sample 1
Sample 2
Sample 3
Sample 4
Sample 5
0.01
1 10 100 1000
Number of epochs
Figure 7.13 Cross-validation curve for PCA network trained with LM algorithm for
all 5 samples
133
4x10-1
Network: Support Vector Machine
3.5x10-1 Sample 1
Sample 2
3x10-1 Sample 3
Sample 4
Sample 5
2.5x10-1
Average MSE for crossvalidation
2x10-1
1.5x10-1
10-1
1 10 100 1000
Number of epochs
Figure 7.14 Cross-validation curve for SVM network for all 5 samples
134
Table 7.1. Description of data collected from literature for Air-Water flow
Diameters Number of Flow regimes Flow Regimes

Authors
used (mm) data points (As per authors) (As per Present study)
Barnea et al. DB, EB, Sl, WA,
4 171 B, I, A, S,
(1983) SW, SS, A
Triplett et al.
1.097, 1.45 480 B, Sl, C, SA, A B, C, I, A
(1999)
Coleman and 1.3, 1.75, B, Di, A, WA, P,
295 B, Di, A, I,
Garimella (1999) 2.6 Sl
Yang and Shieh B, Di, A, SA, W,
1, 2, 3 355 B, Di, A, I, W
(2001) Sl, P
Chen et al. (2002) 1 125 C, BTS, Sl, A B, C, I, A
Ide et al. (2007) 1, 2.4, 4.9 179 B, I, A, B, I, A,
Venkatesan et al. 1.2, 1.7, B, DiB, SA, Sl, A,
509 B, I, A, S,
(2010) 2.6, 3.4 S, WA
Experiment 3, 4 218 I, A, Di, W I, A, Di, W
Table 7.2. Minimum value of MSE for cross-validation reached during training for
all 5 different ANNs
Minimum value of MSE for Cross-validation

Sample
MLP
Number RBF PCA SVM
BP LM
1 0.0823 0.0961 0.0561 0.0555 0.0929
2 0.0789 0.0913 0.0621 0.0559 0.0939
3 0.0662 0.0807 0.0448 0.0371 0.0791
4 0.0817 0.0877 0.0551 0.0547 0.0797
5 0.0887 0.0978 0.0611 0.0609 0.0908
Table 7.3 Optimum number of processing elements in the hidden layer during
training for 4 different ANNs
Sample MLP
RBF PCA
Number BP LM
1 24 25 20 25
2 23 21 22 24
3 25 20 23 22
4 24 18 24 24
5 24 20 23 25
Ul
5
4
3
2
1
0.455 0.366 0.414 0.0234 0.00803 1
m/s
Mean
Sample
Number
D
0.0034 0.001 0.0049 0.001 0.002 2
Input
m
Ug
2.27 6.2 0.7 0.648 54.1 3
m/s
90
94
93
93
91
BP
0 0 0 0 0
92.2
4 Wavy
0 0 0 0 0 5 Stratified
MLP
0 0 0 0 0 6 Dispersed
0 0 0 0 1 7 Annular
87
93
89
86
84
LM
87.8
1 1 1 1 0 8 Intermittent
0 0 0 0 0 9 Churn
Experimental output
0 0 0 0 0 10 Bubbly
98
-0.0545 -0.04687 -0.05556 -0.02299 0.070825

101
100
105
105
11 Wavy
RBF
5 different ANN used
101.8
117 data for Prediction
-0.05555 -0.02482 -0.05556 0.005204 -0.05556 12 Stratified

sample 3 using PCA network
-0.05366 -0.05555 -0.05556 -0.05556 -0.05556 13 Dispersed

0.236733 0.342923 -0.01613 0.08657 1.055556 14 Annular
97
98
101
106
104
PCA
101.2
Number of data predicted accurately
0.601992 0.855235 1.044874 0.939173 -0.05556 15 Intermittent

ANN output
-0.05436 -0.05554 -0.05554 -0.0555 -0.05555 16 Churn

-0.05556 -0.05556 -0.05556 -0.05556 -0.05556 17 Bubbly

93
93
98
92
100
95.2
SVM
Correct Correct Correct Correct Correct 18 Result

ANN analysis with 1981 data used for training, 234 data for cross-validation and
Table 7.5 Comparison of ANN predictions with experimental flow regime data for
Table 7.4 Comparison of ANN predictions with experimental flow regime data with
135
0.169 0.86 1.46 0.0602 0.0144 6.1 0.383 0.811 6.1 0.404 1.2 1
0.001 0.003 0.001 0.00145 0.002 0.0011 0.002 0.0026 0.0011 0.001 0.003 2
2.1 32.1 1.55 0.243 5.56 0.995 0.164 0.582 0.402 1.36 15.6 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 0 0 0 0 0 0 0 0 0 1 6
0 1 0 0 0 0 0 0 0 0 0 7
1 0 0 1 1 1 1 1 0 1 0 8
0 0 0 0 0 0 0 0 0 0 0 9
0 0 1 0 0 0 0 0 1 0 0 10
-0.05519 -0.05543 -0.05236 -0.05544 0.260565 -0.05405 -0.05547 -0.0555 -0.04296 -0.05396 -0.0319 11
-0.03046 -0.05556 0.003306 -0.05556 -0.05556 -0.05556 -0.05556 -0.05512 -0.05556 0.006418 -0.05556 12
-0.05556 0.342755 -0.05556 -0.05556 -0.05553 -0.05556 -0.05556 -0.05555 -0.05556 -0.05556 1.015296 13
-0.04976 0.290272 -0.05556 -0.05554 0.168903 -0.05556 -0.05556 -0.05554 -0.05556 -0.05551 0.072564 14
0.984979 -0.05556 0.369951 0.939087 0.148632 0.469522 1.050066 0.84248 -0.05534 0.886509 -0.05555 15
-0.05555 -0.05099 -0.02704 -0.05556 0.102766 0.007179 -0.05521 -0.0554 0.004972 -0.05552 -0.03409 16
-0.05556 -0.05556 0.736892 -0.05556 -0.05556 -0.04485 -0.05556 -0.05555 1.055475 -0.05556 -0.05556 17
Correct Wrong Correct Correct Wrong Correct Correct Correct Correct Correct Correct 18
136
0.368 1.31 0.46 0.0815 4.94 0.426 2.3 1.84 0.0307 0.316 0.348 1
0.0012 0.002 0.003 0.0011 0.0013 0.001 0.004 0.0026 0.0034 0.0026 0.00145 2
0.61 9.39 7.2 1.78 5.94 0.286 10.56 0.523 1.82 0.26 53.4 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 1 0 0 0 0 1 0 0 0 0 6
0 0 0 0 0 0 0 0 0 0 1 7
1 0 1 1 1 1 0 0 1 1 0 8
0 0 0 0 0 0 0 0 0 0 0 9
0 0 0 0 0 0 0 1 0 0 0 10
-0.05465 0.02219 0.017201 -0.05462 -0.04966 0.011756 -0.05556 -0.05553 -0.03411 -0.05551 -0.05556 11
-0.05556 -0.05556 -0.05545 -0.0498 -0.05556 -0.02594 -0.05556 -0.05511 -0.05554 -0.05536 -0.05556 12
-0.05556 0.765994 -0.04327 -0.05556 1.052468 -0.05556 1.054638 -0.055 -0.02899 -0.05556 -0.05556 13
-0.05556 -0.05381 0.302141 -0.02229 -0.05556 -0.0555 -0.05528 -0.05556 -0.00064 -0.05533 1.055523 14
0.863355 -0.05514 0.443665 1.054484 0.061507 1.04517 -0.05555 0.664873 0.717556 0.974111 -0.05556 15
-0.05556 -0.05502 -0.05538 -0.05464 -0.05556 -0.05484 -0.05525 -0.0553 -0.05299 -0.05553 -0.05111 16
-0.05556 1.055517 -0.05556 -0.05556 0.419432 -0.05556 -0.05556 0.510334 -0.05556 -0.05556 -0.05556 17
Correct Wrong Correct Correct Wrong Correct Correct Wrong Correct Correct Correct 18
137
7.17 0.0624 2.11 0.0602 0.42 0.0298 0.405 0.268 0.211 1.72 0.0737 1
0.0026 0.001 0.0026 0.00145 0.004 0.0017 0.003 0.0011 0.0049 0.001 0.002 2
2.95 1.12 7.89 15.4 3.76 18.3 0.0386 1.49 1.22 0.655 0.15 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
1 0 0 0 0 0 0 0 0 0 0 6
0 0 0 1 0 1 0 0 0 0 0 7
0 1 0 0 1 0 1 1 1 0 1 8
0 0 0 0 0 0 0 0 0 1 0 9
0 0 1 0 0 0 0 0 0 0 0 10
-0.05555 -0.05249 -0.02987 -0.04228 -0.05556 0.015058 -0.05529 -0.05465 -0.05556 -0.00444 -0.05543 11
-0.05556 -0.01627 -0.04273 -0.05556 -0.05556 -0.05556 -0.05555 -0.0431 -0.05556 -0.04755 -0.05556 12
0.959034 -0.05556 0.068475 -0.05556 -0.05555 -0.05555 -0.05541 -0.05556 -0.05556 -0.05547 -0.05556 13
-0.05556 -0.03576 -0.05555 1.055556 -0.05074 1.055556 -0.05556 -0.05529 0.355902 -0.05556 -0.05555 14
-0.05556 0.855545 -0.055 -0.05556 1.005602 -0.05556 1.030407 1.054869 1.046145 0.040205 1.05039 15
-0.05556 -0.05555 -0.05242 0.033854 -0.04532 -0.05555 -0.05556 -0.05482 -0.0555 0.665016 -0.051 16
0.90426 -0.05556 0.953015 -0.05556 -0.05556 -0.05552 -0.05556 -0.05556 -0.05556 0.027973 -0.05556 17
Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct 18
138
0.404 2.41 3.05 1.84 0.438 0.627 0.0164 0.0159 0.898 0.0306 0.184 1
0.001 0.0026 0.001 0.0034 0.00175 0.0034 0.003 0.004 0.002 0.002 0.0012 2
0.816 29 1.57 3.1 1.63 0.242 0.632 4.02 0.897 15 12.2 3

0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 1 0 0 0 0 0 0 0 0 0 6
0 0 0 0 0 0 0 1 0 1 1 7
1 0 0 0 1 1 1 0 0 0 0 8
0 0 0 0 0 0 0 0 0 0 0 9
0 0 1 1 0 0 0 0 1 0 0 10
-0.04034 -0.0554 -0.04926 -0.05549 -0.05459 -0.05544 -0.03867 -0.05544 -0.05518 0.209169 -0.03758 11
-0.0118 -0.05556 -0.04528 0.002498 -0.05556 -0.05556 -0.05465 0.209321 -0.05556 -0.05556 -0.05556 12
-0.05556 1.055418 -0.05556 0.090358 -0.05554 -0.05371 -0.05335 -0.05556 -0.03521 0.012346 -0.05556 13
-0.05525 -0.05556 -0.05556 -0.05556 -0.05548 -0.05552 -0.05554 0.197031 -0.05556 0.60756 1.055553 14
0.848764 -0.05556 0.135574 -0.05556 0.980789 0.922504 1.013223 0.049135 0.941861 -0.05552 -0.05556 15
-0.05512 -0.05556 0.087153 -0.0527 -0.05554 -0.05553 -0.05555 -0.02768 -0.05286 -0.05544 -0.05556 16
-0.05556 -0.05556 0.893455 1.055556 -0.05556 -0.05556 -0.05556 -0.05556 0.197452 -0.05555 -0.05556 17
Correct Correct Correct Correct Correct Correct Correct Wrong Wrong Correct Correct 18
139
0.27 4.03 2.36 2.81 0.481 0.75 0.226 0.41 0.24 2.7 0.0627 1
0.0011 0.0011 0.0026 0.00145 0.0012 0.004 0.002 0.002 0.0017 0.004 0.0034 2
0.773 0.394 3.96 0.853 15.2 0.56 5.36 27.7 15.2 18.52 1.51 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 0 0 0 0 0 0 0 0 1 0 6
0 0 0 0 1 0 0 1 1 0 0 7
1 0 0 1 0 1 1 0 0 0 1 8
0 0 0 0 0 0 0 0 0 0 0 9
0 1 1 0 0 0 0 0 0 0 0 10
-0.05095 -0.04177 -0.05095 -0.05553 -0.04188 -0.05556 -0.01087 -0.05555 -0.05018 -0.05556 -0.05118 11
-0.03163 -0.05555 0.108865 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 12
-0.05556 -0.05556 -0.02779 -0.05556 -0.05556 -0.05552 -0.0554 -0.05556 -0.05555 1.055385 -0.0175 13
-0.05554 -0.05556 -0.05556 -0.05556 1.055227 -0.05556 0.298658 1.049813 1.055556 -0.05511 -0.00852 14
1.049594 0.147148 -0.00461 0.924745 -0.05556 1.046516 0.714566 -0.05556 -0.05545 -0.05556 0.993105 15
-0.05421 -0.05001 -0.05322 -0.05537 -0.05556 -0.05535 -0.03235 -0.05556 -0.05555 -0.05556 -0.05407 16
-0.05556 1.015162 0.614567 0.026388 -0.05556 -0.05556 -0.05556 -0.05334 -0.05556 -0.05556 -0.05556 17
140
5.41 1.06 0.698 0.0237 0.481 0.187 0.403 0.612 0.364 2.02 0.0779 1
0.00175 0.0026 0.00145 0.001 0.0012 0.0034 0.0013 0.0011 0.0012 0.003 0.0026 2
1.28 5.23 3.02 41.3 6.12 0.301 2.95 0.0917 12.2 10.9 5.23 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 0 0 0 0 0 0 0 0 1 0 6
0 0 0 1 1 0 0 0 1 0 1 7
0 0 1 0 0 1 1 1 0 0 0 8
0 0 0 0 0 0 0 0 0 0 0 9
1 1 0 0 0 0 0 0 0 0 0 10
-0.05556 -0.04264 -0.05259 -0.02192 -0.02989 -0.05506 -0.05385 0.025738 -0.03336 -0.00741 -0.03876 11
-0.05556 -0.03337 -0.05556 -0.05549 -0.05556 -0.05556 -0.05556 -0.01246 -0.05556 -0.05555 0.016917 12
-0.04623 -0.02516 -0.05555 -0.05548 -0.05556 -0.01033 -0.05556 -0.05556 -0.05556 1.051999 -0.05555 13
-0.05556 -0.02217 -0.05555 0.901264 0.271959 -0.05527 -0.05555 -0.05556 1.055397 -0.05555 0.780626 14
0.255182 -0.0318 0.639807 -0.05556 -0.05555 1.0473 1.025309 0.540054 -0.05556 -0.05555 0.162313 15
-0.05556 0.031992 -0.03388 -0.05556 -0.05556 -0.05544 -0.05454 -0.01298 -0.05556 -0.05401 -0.0512 16
0.090607 0.935942 -0.05486 -0.05556 -0.05556 -0.05556 -0.05556 -0.05525 -0.05556 -0.05553 -0.05556 17
Wrong Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct 18
141
0.348 1.23 0.521 1.43 0.105 0.0204 0.00396 1.44 0.943 0.86 0.213 1
0.00145 0.00145 0.00145 0.001 0.0024 0.002 0.004 0.001 0.0034 0.001 0.001 2
0.469 2.53 0.0423 7 4.3 0.66 4.02 3.61 0.76 42.8 1.26 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 1 0 0 0 0 5
0 0 0 0 0 0 0 0 0 1 0 6
0 0 0 0 1 0 0 0 0 0 0 7
1 1 0 1 0 1 0 1 0 0 1 8
0 0 0 0 0 0 0 0 0 0 0 9
0 0 1 0 0 0 0 0 1 0 0 10
-0.05549 -0.05449 -0.05544 -0.03482 -0.04753 -0.05477 -0.05497 -0.05434 -0.05549 -0.05556 -0.05381 11
-0.05556 -0.05556 -0.05556 -0.03047 -0.05465 -0.05556 1.001979 -0.0272 -0.05547 -0.05556 -0.01307 12
-0.05556 -0.05555 -0.05556 -0.00723 -0.05556 -0.05556 -0.05556 -0.0553 -0.04577 1.055496 -0.05556 13
-0.05556 -0.05556 -0.05556 -0.05547 1.01561 -0.0555 0.093439 -0.05556 -0.05544 -0.05553 -0.05454 14
1.001494 0.825138 0.801744 0.701498 0.52225 1.038655 -0.04006 0.94669 0.088088 -0.05556 0.908248 15
-0.05555 0.018661 -0.05555 -0.03033 -0.00364 -0.04114 -0.02809 -0.03315 -0.05535 -0.05556 -0.05555 16
-0.05556 0.016728 -0.05556 0.207733 -0.05556 -0.05556 -0.05554 0.075364 1.030037 -0.05556 -0.05556 17
Correct Correct Wrong Correct Correct Correct Correct Correct Correct Correct Correct 18
142
0.6 0.226 0.63 0.402 1.29 0.521 0.182 1.22 0.404 3.3 0.507 1
0.0034 0.002 0.001 0.0026 0.002 0.00145 0.0012 0.0011 0.001 0.0011 0.0026 2
6.15 0.513 0.456 29.5 13.7 7.07 6.12 0.107 7.88 6.4 9.21 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 0 0 0 1 0 0 0 0 0 0 6
1 0 0 1 0 0 1 0 0 0 1 7
0 1 1 0 0 0 0 0 1 0 0 8
0 0 0 0 0 1 0 0 0 1 0 9
0 0 0 0 0 0 0 1 0 0 0 10
-0.05376 -0.05539 0.020944 -0.05555 -0.02999 -0.02311 -0.03619 0.04127 -0.04132 -0.02252 -0.0433 11
-0.05522 -0.05556 -0.02978 -0.05556 -0.05556 -0.05556 -0.05556 -0.02052 -0.04222 -0.02553 -0.04958 12
-0.05551 -0.05556 -0.05537 -0.05556 1.002468 -0.05556 -0.05556 -0.05556 -0.05553 -0.05556 -0.05555 13
1.049074 -0.05555 -0.05555 1.055555 -0.03016 0.41048 1.045099 -0.05556 0.464479 -0.05556 0.823185 14
-0.0531 1.051292 0.844081 -0.05556 -0.05554 0.146438 -0.05554 0.226413 0.746332 -0.05554 -0.02208 15
-0.02995 -0.05356 0.009983 -0.05346 -0.05552 -0.02474 -0.05556 -0.01112 -0.05554 1.053684 -0.05264 16
0.008354 -0.05556 -0.05534 -0.05556 1.055526 -0.05556 -0.05556 0.89285 -0.05556 -0.05555 -0.05556 17
Correct Correct Correct Correct Correct Wrong Correct Correct Correct Correct Correct 18
143
0.742 0.0306 1.9 0.206 0.42 0.0608 2.5 1.91 0.22 0.0395 1.24 1
0.0017 0.0012 0.0024 0.0011 0.004 0.00145 0.003 0.004 0.00175 0.0026 0.0034 2
3.04 1.82 0.0841 72.8 10.54 0.0822 16.7 0.0398 25.7 0.523 1.51 3
0 0 0 0 0 0 0 0 0 0 0 4
0 0 0 0 0 0 0 0 0 0 0 5
0 0 0 0 0 0 1 0 0 0 0 6
0 0 0 1 1 0 0 0 1 0 0 7
1 1 0 0 0 1 0 0 0 1 0 8
0 0 0 0 0 0 0 0 0 0 0 9
0 0 1 0 0 0 0 1 0 0 1 10
-0.0498 -0.05479 -0.05553 -0.05556 -0.05556 -0.05542 -0.02383 -0.05556 -0.05552 -0.05528 -0.05549 11
-0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05006 -0.04356 12
-0.05549 -0.05556 -0.05551 -0.05556 -0.05556 -0.05556 1.055549 0.17014 -0.05556 -0.05556 0.028049 13
-0.05432 -0.05372 -0.05556 1.055555 0.990807 -0.05555 -0.05556 -0.05556 1.055556 -0.05283 -0.05553 14
0.144656 0.941631 0.491399 -0.05556 0.053191 0.883437 -0.05556 0.017105 -0.05556 0.837489 -0.05486 15
-0.05373 -0.05556 -0.05525 -0.05556 -0.03146 -0.05556 -0.05554 -0.0549 -0.05555 -0.05547 -0.05161 16
0.011914 -0.05556 0.799618 -0.05556 -0.05556 -0.05556 -0.05556 0.83579 -0.05556 -0.05556 1.055556 17
144
0.0162 1.96 0.00643 0.75 0.0608 0.545 0.742 0.691 0.46 1.23 0.0307 1
0.002 0.0024 0.004 0.004 0.00145 0.002 0.0017 0.00145 0.003 0.0013 0.0034 2
39.8 0.136 15.5 23.87 0.721 1.09 0.61 1.61 8.3 5.94 1.05 3
1 0 0 0 0 0 0 0 0 0 0 4
0 0 1 0 0 0 0 0 0 0 0 5
0 0 0 0 0 0 0 0 0 0 0 6
0 0 0 1 0 0 0 0 1 0 0 7
0 0 0 0 1 1 1 1 0 1 1 8
0 0 0 0 0 0 0 0 0 0 0 9
0 1 0 0 0 0 0 0 0 0 0 10
0.791795 -0.05553 -0.04631 -0.05556 -0.05542 -0.05511 -0.05545 -0.05525 0.022917 -0.01619 -0.04558 11
-0.05556 -0.05556 1.054653 -0.05556 -0.05556 -0.05556 -0.05556 -0.05556 -0.05552 -0.05556 -0.05555 12
-0.05556 -0.05549 -0.05556 -0.05556 -0.05556 -0.05555 -0.05554 -0.05555 -0.0412 -0.05521 -0.00433 13
1.055556 -0.05556 0.252785 1.05278 -0.05516 -0.05555 -0.05556 -0.05556 0.405202 -0.05556 -0.05051 14
-0.05556 0.558549 -0.05556 -0.05556 0.976561 1.042185 0.842404 0.908248 0.291249 0.708032 0.956654 15
-0.05555 -0.05533 -0.0085 -0.05549 -0.05555 -0.05506 -0.05551 -0.05419 -0.05531 0.124366 -0.05481 16
-0.05555 0.750003 -0.05554 -0.05556 -0.05556 -0.05556 -0.04947 -0.05537 -0.05556 -0.02688 -0.05556 17
Wrong Correct Correct Correct Correct Correct Correct Correct Correct Correct Correct 18
145
0.907 0.0329 1
0.00145 0.0024 2
9.02 0.0891 3
0 0 4
0 0 5
0 0 6
0 0 7
0 1 8
1 0 9
0 0 10
0.01713 -0.05543 11
-0.05556 -0.05553 12
-0.05533 -0.05556 13
-0.05549 -0.05157 14
0.007539 0.899729 15
1.001559 -0.05419 16
0.009911 -0.05556 17
Correct Correct 18
146
CHAPTER 8
Conclusions
Conclusions
Conclusions
Modeling of hydrodynamics of liquid and gas-liquid flow through pipes and
piping components, bend, helical coils and flow regime prediction using artificial
neural network (ANN) have been reported. The data for this analysis is acquired from
literature and thesis from earlier research works are carried out in our laboratory. The
following ANN modeling have been incorporated in the thesis,
i. Prediction of pressure drop across the piping components is made for non-
Newtonian liquid flow using Backprogagation training algorithm in
hidden and output layer with the Eq. (2.1) as transfer function. The best
result has been obtained using number of nodes in the hidden layer as 8,
14, 9 and 20 for elbow, orifice, gate valve and globe valve respectively.
ii. Prediction of pressure drop across the pipes and piping components is
made for gas-non-Newtonian liquid flow using Backprogagation training
algorithm in hidden and output layer with the Eqs. (2.1 – 2.4) as transfer
functions. The best result has been obtained using transfer function and
number of nodes in the hidden layer as,
a) Transfer function 3, Eq. (2.3) and 25 - horizontal pipe
b) Transfer function 1, Eq. (2.1) and 14 – elbow
c) Transfer function 1, Eq. (2.1) and 15 – orifice
d) Transfer function 4, Eq. (2.4) and 13 – gate valve
e) Transfer function 2, Eq. (2.2) and 25 – globe valve
iii. Prediction of pressure drop across the 45o bend is made for gas-non-
Newtonian liquid flow using backprogagation training algorithm in hidden
and output layer with the Eqs. (2.1 – 2.4) as transfer functions. The best
148
Conclusions
result has been obtained using transfer function 1, Eq. (2.1) and the
number of nodes in the hidden layer 20.
iv. Prediction of pressure drop and holdup across the helical coil in horizontal
orientation are made for gas-non-Newtonian liquid flow using
backprogagation training algorithm in hidden and output layer with the
Eqs. (2.1 – 2.4) as transfer functions. The best result has been obtained
using transfer function and number of nodes in the hidden layer as,
a) Transfer function 1, Eq. (2.1) and 10 – two-phase pressure drop
b) Transfer function 1, Eq. (2.1) and 15 – gas holdup
v. Prediction of flow regimes are made for air-water flow through small
diameter (1 – 5 mm) pipeline using Backprogagation (BP), Levenberg-
Marquardt (LM), Radial Basis Function (RBF), Principal Component
Analysis (PCA) and Support Vector Machine (SVM) training algorithm in
hidden and output layer with the Eq. (2.1) as transfer functions. The
Radial Basis Function gives slightly better predictability over the other
networks.
Hence, it is clear that the predictive ability of ANN in the field of fluid dynamics
particularly hydrodynamic predictions of different systems are excellent and can be
used for design and optimum operating conditions for such systems.
Nomenclature
Nomenclature
D diameter, m
E error
e error matrix
F function, dimensionless
f friction factor, dimensionless
g acceleration due to gravity, m/s2
K′ consistency index (Nsn′/m2)
L length, m
N total number of data set
n′ flow behavior index, dimensionless
Npl ( )
liquid property group µeff4 g / ρlσ l3 dimensionless
o orifice
ΔP pressure drop, N/m
R cross-correlation coefficient (dimensionless)
Rc radius of curvature (m)
t epoch number
Q flow rate (m3/s)
U, V velocity (m s-1)
w connection weight
x experimental value
y predicted value
Greek letters
α ratio of the valve opening to the full opening of the valve, dimensionless
150
Nomenclature
αg gas holdup, dimensionless
αl liquid holdup, dimensionless
β gain, dimensionless
θ angle of the bend
λ combination coefficient, dimensionless
µ viscosity, Ns/m2
ρ density, kg/m3
σ standard deviation (dimensionless)
σs surface tension, N/m
Subscripts
b bend
c coil
f frictional
g gaseous state
l liquid state
t tube
eb elbow
eff effective
tp two-phase
ftp frictional two-phase
tplb two-phase based on liquid for bend

References
References
Adhikari, B., Jindal, V.K., Artificial neural networks: a new tool for prediction of
pressure drop of non-Newtonian fluid foods through tubes, J. Food Eng. 46(1),
43–51, 2000.
Akbar, M.K., Plummer, D.A., Ghiaasiaan, S.M., On gas–liquid two-phase flow

regimes in microchannels, Int. J. Multiphase Flow 29(5), 855–865, 2003.
Alizadehdakhel, A., Rahimi, M., Sanjari, J., Alsairafi, A.A., CFD and artificial neural
network modeling of two-phase flow pressure drop, Int. Commun. Heat Mass
Transfer 36(8), 850–856, 2009.
Alpsan, D., Towsey, M., Ozdamar, O., Tosi, A.C., Ghista, D.N., Efficiency of
modified backpropagation and optimization methods on a real-world medical
problem, Neural Net. 8(6), 945–962, 1995.
Amari, S., Wu, S., Improving support vector machine classifiers by modifying kernel
functions, Neural Net. 12(6), 783–789, 1999.
Androsiuk, J., Kulak, L., Sienicki, K., Neural network solution of the Schrödinger
equation for a two-dimensional harmonic oscillator, Chem. Phys., 173, 377–383,
1994.
Ansari, A.M., Sylvester, N.D., Sarica, C., Shoham, O., Brill, J.P., A comprehensive
mechanistic model for upward two phase flow in wellbores, SPE Production and
Facil. 9(9), 143–152, 1994.
Bai, B.F., Zhang, S.J., Zhao, L., Zhang, X.M., Guo, L.J., Online recognition of the
multiphase flow regime, Sci. China Ser E-Tech. Sci., 51(8), 1186–1194, 2008.
Baker, O., Simultaneous flow of oil and gas, Oil and Gas J. 53, 185–195, 1954.
Bandyopadhyay, T.K., Studies on non-Newtonian and gas-non-Newtonian liquid flow

through horizontal tube and piping components, Ph.D Thesis, University of
Calcutta, 2002.
Bandyopadhyay, T.K., Das, S.K., Non-Newtonian pseudoplastic liquid flow through

small diameter piping components, J. Petro. Sci. Eng. 55(1–2), 156–166, 2007.
Bandyopadhyay, T.K., Banerjee, T.K., Das, S.K., Gas-non-Newtonian liquid flow

through elbows, Chem. Eng. Comm. 182(1), 21–33, 2000.
Banerjee, T.K., Das S.K., Gas-non-Newtonian liquid flow through globe and gate
valves, Chem. Eng. Comm. 167(1), 133–146, 1998.
Barajas, A.M., Panton, R.L., The effect of contact angle on two-phase flow in
capillary tubes, Int. J. Multiphase Flow 19(2), 337–346, 1993.
152
References
Bar, N., Das, S.K., Comparative study of friction factor by prediction of frictional
pressure drop per unit length using empirical correlation and ANN for gas-non-
Newtonian liquid flow through 180° circular bend, Int. Rev. Chem. Engg. 3(6),
628–643, 2011.
Bar, N., Das, S.K., Gas-non-Newtonian liquid flow through horizontal pipe – gas
holdup and pressure drop prediction using multilayer perceptron, Am. J. Fluid
Dynamics 2(3), 7–16, 2012a.
Bar, N., Das, S.K., Gas holdup and frictional pressure drop prediction in vertical pipes
for gas-non-Newtonian flow using ANN, 2nd Int. Conf. Process Engineering and
Advanced Materials, (ICPEAM 2012), Malaysia, June 12 – 14, 2012b.
Bar, N., Das, S.K., Frictional pressure drop for gas-non-Newtonian liquid flow
through 90° and 135° circular bend: prediction using empirical correlation and
ANN, Int. J. Fluid Mech. Res. 39(5), 416–437, 2013a.
Bar, N., Das, S.K. Prediction of Flow Regime for Air-Water Flow in Circular Micro
Channels using ANN. Procedia Technol. 10, 242–252, 2013b.
Barnea, D., Luninski, Y., Taitel, Y., Flow pattern in horizontal and vertical two phase
flow in small diameter pipes, Can. J. Chem. Eng. 61(5), 617–620, 1983.
Barron, A.R., Universal approximation bounds for superpositions of a sigmoidal

function, IEEE Trans. Information Theory 39(3), 930–945, 1993.
Basheer, I.A., Hajmeer, M.J. Artificial neural networks: fundamentals, computing,

design, and application, J. Microbiol. Methods 43(1), 3–31, 2000.
Bell, K.J., Two-phase flow regime consideration in condenser and vapour design, Int.
Commun. Heat Mass Transfer 15(4), 429–448, 1988.
Berger, S.A., Talbot, L., Flow in curved pipes, Annu. Rev.Fluid Mech. 15, 461–512,
1983.
Biswas A.B., Studies on Two-Phase Gas-non-Newtonian liquid flow Through Helical

Coils, Ph.D Thesis, University of Calcutta, 2007.
Biswas, A.B., Das, S.K., Frictional pressure drop of air non-Newtonian liquid flow
through helical coils in horizontal orientation, Can. J. Chem. Engg. 85(2) 129–136,
2007.
Blanco, M., Maspoch, S., Villarroya, I., Peralta, X., Gonzalez, J.M., Torres, J.,
Determination of the physical properties of bitumens by use of near-infrared
spectroscopy with neural network. Joint modeling of linear and non-linear
parameters, Analyst. 126(3), 378–382, 2001.
Bonakdari, H., Baghalian, S., Nazari, F., Fazli, M., Numerical analysis and prediction
of the velocity field in curved open channel using artificial neural network and
153
References
genetic algorithm, Engg. Appl. Computational Fluid Mechanics 5(3), 384–396,

2011.
Boser, B., Guyon, I., Vapnik, V., A training algorithm for optimal margin classifiers,
5th Annual Workshop Computational Learning Theory, New York, ACM Press,
1992.
Broomhead, D., Lowe, D., Multivariable functional interpolation and adaptive

networks, Complex Systems, 2(3), 321–355, 1988.
Burges, C.J.C., A tutorial on support vector machines for pattern recognition, Data
Mining and Knowledge Discovery, 2(2), 1–47, 1998.
Caetano, C., Reis Jr., J.L., Amorim, J., Lemes, M.R., Pino, A.D., Using Neural
Networks to Solve Nonlinear Differential Equations in Atomic and Molecular
Physics, Int. J. Quantum Chem., 111(12), 2732–2740, 2011.
Cai, S., Toral, H., Qiu, J., Archer, J.S., Neural network based objective flow regime
identification in air-water two phase flow, Can. J. Chem. Eng., 72(3), 440–445,
1994.
Carol, B., Dickinson, B.D., Proc. Int. Joint Conf. Neural Net. vol. I, pp 607–611,
1989.
Chan, L-W., Efficacy of different learning algorithms of the back-propagation

network. IEEE Region 10th Conf. Computer and Communication Systems, Hong
Kong, September 24–27, vol. 1, pp 23–27, 1990.
Charalambous, C., Conjugate-gradient algorithm for efficient training of artificial

neural networks, IEEE Proc. G: Circuits Dev. Sys. 139, 301–310, 1992.
Chatterjee, C., Kang, Z., Roychowdhury, V.P., Algorithms for accelerated

convergence of adaptive PCA, IEEE Trans. Neural Net. 11(2), 338–355, 2000.
Chen, W.L., Twu, M.C., Pan, C., Gas-liquid two-phase flow in micro-channel, Int. J.
Multiphase Flow 28(7), 1235–1247, 2002.
Chen, Y., Jiao, T., McCall, T.W., Baichwal, A.R., Meyer, M.C., Comparison of four
artificial neural network software programs used to predict in vitro dissolution of
controlled-release tablets, Pharm. Technol. Eur. 8, 46–55, 2002.
Coleman, J.W., Garimella, S., Characteristics of two-phase patterns in small diameter

round and rectangular tubes, Int. J. Heat Mass Trans. 42(15), 2869–2881, 1999.
Cong, T., Su, G., Qiu, S., Tian, W., Applications of ANNs in flow and heat transfer
problems in engineering: A review work, Prog. Nucl. Energy 62, 54–71, 2013.
Cortes, C., Vapnik, V., Support-vector Networks, Machine Learning 20(3), 273–297,
1995.
154
References
Crowe, C.T., Multiphase Flow Handbook, CRC, 2006.
Cutts, D., Hoftun, J.S., Somburger, A., Johnson, C.R., Zeller, R.T., Neural Networks
for Event Filtering at D0, Comp. Phys. Commun. 57(1–3), 478–482, 1989.
Darsey, J.A., Noid, D.W., Upadhyay, D.H., Application of neural network computing
to the solution for the ground-state eigenenergy of two-dimensional harmonic
oscillators, Chem. Phys. Lett., 177(2), 189–194, 1991.
Das, S.K., Studies on two-phase gas-non-Newtonian liquid flow in horizontal, vertical

tubes and bends, Ph.D. thesis, Dept. Chem. Engg., Indian Institute of Technology,
Kharagpur, India, 1988.
Das, S.K., Water flow through helical coils in turbulent conditions. Ch. 13,
Multiphase Reactor and Polymerization system Hydrodynamics, Adv. In Engg.
Fluid Mech. Ser., Ed. N. P. Chermisinoff, Golf Publishing Co., pp 379–403,
1996a.
Das, S.K., Non-Newtonian liquid flow through globe and gate valves. Ch. 17,
Multiphase Reactor and Polymerization system Hydrodynamics, Adv. In Engg.
Fluid Mech. Ser., Ed. N. P. Chermisinoff, Golf Publishing Co., 487–505, 1996b.
Das, S.K., Bar, N., Hydrodynamics of Gas-non-Newtonian Liquid Flow and ANN
Predictibility, Lambart Academic Publishing. Saarbrücken, Germany, 2013.
Das, S.K., Biswas M.N., Mitra, A.K., Pressure loses in two-phase gas-non-Newtonian
liquid flow in horizontal tube, J. Pipelines 7, 307–325, 1989.
Das, S.K., Biswas M.N., Mitra, A.K., Holdup for two-phase flow gas-non-Newtonian
liquid mixtures in horizontal and vertical pipes, Can. J. Chem. Engg. 70(3), 431–
437, 1992.
Damianides, C.A., Westwater, J.W., Two-phase flow patterns in a compact heat

exchanger and in small tubes, Proc. 2nd UK National Conf. on Heat Trans.,
(Glasgow, 14–16 September, Mechanical Engineering Publications, London) vol.
II, pp. 1257–1268, 1988.
Damiantaras, K.I., Kung, S.Y., Principal Component Neural Networks, Wiley, New
York, 1996.
Desai, K.M., Vaidya, B.K., Singhal, R.S., Bhagwat, S.S., Use of an artificial neural
network in modeling yeast biomass and yield of β-glucan, Process Biochemistry
40, 1617–1626, 2005.
Dibike, Y.B., Solomatine, D.P., River Flow Forecasting Using Artificial Neural
Networks, Phys. Chem. Earth (B) 26(1), 1–27, 2001.
155
References
Dziubinski, M., A general correlation for the two-phase pressure drop in intermittent
flow of gas and non-newtonian liquid mixtures in a pipe, Trans. Chem. Eng. Res.
Des. 73, 528–533, 1995.
Fahlman, S.E., An empirical study of learning speed in backpropagation networks.

Technical Report CMU – CS – 88 – 162, Carnegie Mellon University, Pittsburgh,
PA (1988).
Feng, L.–H., Lu, J., The practical research on flood forescasting based on artificial
neural network. Expert System with Application 37(4), 2974–2977, 2010.
Fukano, T., Kariyasaki, A., Characteristics of gas–liquid two-phase flow in a

capillary, Nucl. Eng. Des. 141(1–2), 59–68, 1993.
Fukano, T., Kariyasaki, A., Kagawa, M., Flow patterns and pressure drop in
isothermal gas–liquid flow in a horizontal capillary tube, ANS Proc., Nat. Heat
Trans. Conf., ISBN 0-89448-149-5, 4, 153–161, 1989.
Ghanbarzadeh, S., Hanafizadeh, P., Saidi, M.H., Intelligent Imaged-Based Gas-Liquid

Two-Phase Flow Regime Recognition. ASME J. Fluids Eng., 134(6), 061302,
2012 (10 pages).
Ghiaasiaan, S.M., Abdel-Khalik, S.I., Two-phase flow in microchannels, Adv. Heat

Trans. 34, 145–254, 2001.
Glucksman, H., On improvement of a linear separation by extending the adaptive

process with a stricter criterion, IEEE Trans. Electronic Comp. EC-15(6), 941–
944, 1966.
Gorman, R.P., Sejnowski, T.J., Analysis of hidden units in a layered network trained
to classify sonar targets, Neural Netwo., 1(1), 75–89, 1988.
Govier, G.W., Aziz, K., The Flow of Complex Mixtures in Pipes, Van Nostran
Reinhold, New York City, 1972.
Hagen, G.H.L., Üeber die Bewegung des Wassers in engen cylidrischen Röhren,
Poggendorfs Ann. Phy. Chem., 46 423–442 (1839).
Hagan, M.T., Menhaj, M.B., Training feedforward networks with the Marquardt
algorithm, IEEE Trans. Neural Net. 5(6), 989–993, 1994.
Hanafizadeh, P., Ghanbarzadeh, S., Saidi, M.H., Visual technique for detection of
gas–liquid two-phase flow regime in the airlift pump, J. Petrol. Sci. Eng. 75(3–4),
327–335, 2011.
Hannan, J.M., Bishop, J.M.A., Comparison of fast training algorithms over two real
problems. IEE 5th Int. Conf. Artificial Neural Network, Conference Publication
No. 440, Cambridge, UK, July 7–9, 1–6, 1997.
156
References
Hassan, I., Vaillancourt, M., Pehlivan, K., Two-phase flow regime transitions in
microchannels: a comparative rxperimental study, Microscale Thermo-phys. Eng.
9, 165–182, 2005.
Haykin, S., Neural Networks: A comprehensive foundation. Macmillan, New York,

1994.
Hecht-Nielsen, R., Kolmogorov mapping neural network existence theorem, Proc. 1st
IEEE Int. Conf. Neural Net. 1, 11–14, 1987.
Hecht-Nielsen, R., Neurocomputing, Addison-Wesley, Reading, MA, 1990.
Henseler, J., Backpropagation. In: Braspenning, P.J. et al. (Eds.), Artificial Neural
Networks, An Introduction to ANN theory and practice. Lecture notes in
Computer Science, Springer, NY, 37–66, 1995.
Hernández, L., Juliá, J.E., Chiva, S., Paranjape, S., Ishii, M., Fast classification of
two-phase flow regimes based on conductivity signals and artificial neural
networks, Meas. Sci. Technol., 17, 1511–1521, 2006.
Hernández, L., Juliá, J.E., Ozar, B., Hibiki, T., Ishii, M., Flow Regime Identification
in Boiling Two-Phase Flow in a Vertical Annulus, ASME J. Fluids Eng. 133(9),
091304, 2011 (10 Pages).
Hestroni, G., Handbook of Multiphase Systems, Hemisphere Publishing. Corp.,

Washington, DC, 1982.
Himmelblau D.M., Application of artificial neural networks in chemical engineering.

Kor. J. Chem. Eng. 17(4), 373 – 392, 2000.
Ho, T.L., 3-D inversion of borehole-to-surface electrical data using a back-

propagation neural network, J. Appl. Geophys. 68(4), 489–499, 2009.
Hornik, K. Stinchcombe, M., White, H., Multilayer feedforward networks are

universal approximators, Neural Net. 2(5), 359–366, 1989.
Hoskins, J.C., Himmelblau, D.M., Artificial neural network models of knowledge

representation in chemical engineering, Comput. Chem. Eng. 12(9–10), 881–890,
1988.
Hosseini, H.G., Luo, D., Raynolds, K.J., The comparison of different feed forward
neural network architectures for ECG signal diagnosis, Med. Engg. Phys., 28,
372–378, 2006.
Hua, C., Wang, C., Geng, Y., Shi, T., Noninvasive Flow Regime Identification for
Wet Gas Flow Based on Flow-induced Vibration, Chin. J. Chem. Eng. 18(5), 795–
803, 2010.
157
References
Ide, H., Kariyasaki, A., Fukano, T., Fundamental data on gas – liquid two-phase flow
in microchannels, Int. J. Thermal Sci. 46(6), 519–530, 2007.
Jacobs, R.A., Increased rates of convergence through learning rate adaptation, Neural
Net. 1(4), 295–307, 1988.
Jana, A.K., Das, G., Das, P.K., Flow regime identification of two-phase liquid–liquid
upflow through vertical pipe, Chem. Eng. Sci. 61(5), 1500–1515, 2006.
Jones Jr., O.C., Zuber, N., The interrelation between void fraction fluctuations and
flow patterns in two-phase flow, Int. J. Multiphase Flow 2(3), 273–306, 1975.
Karmer, M.A., Nonlinear principal component analysis Using autoassociative neural

networks, AIChE J. 37(2), 233–243, 1991.
Karnin, E., A Simple procedure for pruning back-propagation trained neural

networks, IEEE Trans. Neural Net. 1(2), 239–242, 1990.
Karras, D.A., Perantonis, S.J., Comparison of learning algorithms for feedforward

networks in large scale networks and problems, Proc. Int. Joint Conf. Neural
Networks, Nagoya, Japan, October 25–29, 532–535, 1993.
Kaya, A.S., Sarica, C., Brill, J.P., Comprehensive mechanistic model of two-phase
flow in deviated wells, In: SPE Annual Technical Conference and Exhibition. SPE
56522, Houston, October 1999.
Kreitzer, P.J., Hanchak, M., Byrd, L., Horizontal Two Phase Flow Regime
Identification: Comparison of Pressure Signature, Electrical Capacitance
Tomography (ECT) and High Speed Visualization, Proc. ASME 2012
International Mechanical Engineering Congress and Exposition, Houston, Texas,
USA, 7, 1281–1291, 2012.
Kreitzer, P.J., Hanchak, M., Byrd, L., Flow Regime Identification of Horizontal Two
Phase Refrigerant R-134a Flow Using Neural Networks, Proc. ASME 2013
International Mechanical Engineering Congress and Exposition, San Diego,
California, USA, 7B, V07BT08A059, 2013.
Kumar, K.V., de Castro, M.M., Martinez-Escandell, M., Molina-Sabio, M.,

Rodriguez-Reinoso F., Neural network and principal component analysis for
modeling of hydrogen adsorption isotherms on KOH activated pitch based
carbons containing different heteroatoms, Chem. Eng. J. 159(1–3), 272–279,
2010.
Kuwahara, T., Vuyst, F.D., Yamaguchi, H., Flow regime classification in air–
magnetic fluid two-phase flow, J. Phys.: Condens. Matter, 20, 204141 (6pp),
2008.
158
References
Lahiri, S.K., Ghanta, K.C., Development of an artificial neural network correlation for
prediction of hold-up of slurry transport in pipelines, Chem. Eng. Sci. 63(6),
1497–1509, 2008.
Larachi, F., Belfares, L., Iliuta, I., Grandjean, B.P.A., Three-phase fluidization
macroscopic hydrodynamics revisited, Ind. Eng. Chem. Res. 40(3), 993–1008,
2001.
Lee, C.Y., Lee, S.Y., Influence of surface wettability on transition of two-phase flow
pattern in round mini-channels, Int. J. Multiphase Flow 34(7), 706–711, 2008.
Liu, J., Min, K., Han, C., Chang, K.S., Robust Nonlinear PLS based on neural
networks and application to composition estimator for high-purity distillation
columns, Kor. J. Chem. Eng. 17(2), 184–192, 2000.
Lopez, R., Oňate, E., A variation formulation for the multilayer perceptron. Proc.
ICAAN 2006, Lecture notes in Computer Science, 4132(1), Springer, Berlin, pp
159–168.
Lowe, D.C., Rezkallah, K.S., Flow regime identification in microgravity two-phase

flow using void fraction signals, Int. J. Multiphase Flow 25(3), 433–457, 1999.
Lynch, M., Patel, H., Abrahamse, A., Rajendran, A.R., Medsker, L, Neural Network
applications in Physics, In: Proc. Int. Joint Conf. on Neural Net., Washington
D.C., 15 – 19 July, vol. 3, 2054–2058, 2001.
Maier, H.R., Dundy, G.C., Undarstanding the behavior and optimizing the
performance of back-propagation neural networks: an empirical study, Environ.
Model. Softw. 13(2), 179–191, 1998.
Mandal, S.N., Das, S.K., Gas-Newtonian liquid flow through helical coils in
horizontal orientation, Can. J. Chem. Engg. 80(5), 979–983, 2002.
Mandal, S.N., Das, S.K., Gas-Newtonian liquid flow through helical coils in vertical
orientation, Ind. Engg. Chem. Res. 42(14), 3487–3494, 2003.
Mandhane, J.M., George, G.A., Aziz, K.A., Flow pattern map for gas-liquid flow in
horizontal pipes, Int. J. Multiphase Flow 1(4), 537–553, 1974.
Manickaraj, J., Balasubramanian, N., Estimation of the heat transfer coefficient in a

liquid–solid fluidized bed using an artificial neural network, Adv. Powder
Technol. 19(2), 119–130, 2008.
Marini, F., Artificial neural networks in foodstuff analyses: trends and perspectives. A
review, Anal. Chim. Acta 635(2), 121–131, 2009.
Matsui, G., Automatic identification of flow regimes in vertical two-phase flow using
differential pressure fluctuations, Nucl. Eng. Des. 95, 221–231, 1986.
159
References
Meneses, A.A.M., Giusti, A., de Almeida, A.P., Nogueira, L.P., Braz, D., Barroso,
R.C., deAlmeida, C.E., Automated segmentation of synchrotron radiation micro-
computed tomography biomedical images using Graph Cuts and neural networks,
Nucl. Instrum. Methods Phys. Res. A, 660, 121–129, 2011.
Mi, Y., Ishii, M., Tsoukalas, L.H., Vertical two phase flow identification using
advanced instrumentation and neural networks, Nucl. Eng. Des. 184(2–3), 409–
420, 1998.
Mi, Y., Ishii, M., Tsoukalas, L.H., Flow regime identification methodology with
neural networks and two-phase flow models, Nucl. Eng. Des. 204(1–3), 87–100,
2001.
Milano, M., Koumoutsakos, P., Neural Network Modeling for Near wall Turbulant
Flow, J. Comput. Phys., 182(1) 1–26, 2002.
Minai, A.A., Williams, R.D., Back-propagation heuristics: a study of the extended

delta-bar-delta algorithm. Proc. Int. Joint Conf. Neural Net., San Diego,
California, 17 – 21 June, vol. 1, 595–600, 1990.
Mishima, K., Hibiki, T., Nishihara, H., Some characteristics of air-water two-phase
flow in narrow rectangular ducts, Int. J. Multiphase Flow 19(1), 115–124, 1993.
Mishima, K., Hibiki, T., Nishihara, H., Some characteristics of air–water two-phase
flow in small diameter tubes, Proc. 2nd Int. Conf. Multiphase Flow, Tokyo, Japan,
4, April 3–7, 39–46, 1995.
Mishima, K., Hibiki, T., Some characteristics of air–water two-phase flow in small
diameter vertical tubes, Int. J. Multiphase Flow 22(4), 703–712, 1996.
Mӧller, M.F., A scaled conjugate gradient algorithm for fast supervised learning,
Neural Net. 6(4), 525–533, 1993.
Molga, E.J., Neural network approach to support modeling of chemical reactors:

problems, resolutions, criteria of application, Chem. Eng. Proc. 42(8–9), 675–695,
2003.
Montagu, G., Morris, J., Neural network contributions in biotechnology, TBITECH,

12(8), 312–323, 1994,.
Moody, J., Darken, C., Fast learning in networks of locally-tuned processing units,
Neural Comput. 1(2), 281–294, 1989.
Müller, U., Artificial intelligence—applications in high energy and nuclear physics,

Nucl. Instr. and Meth. Phys. Res., 502, 811–814, 2003.
Navier, C.L.M.H., Sur les lois de l’équilibre et du mouvement des corpes solides
élastique, Bull. Soc. Philomath., 75, 177–183 (1823).
160
References
Oh, H.J., Lee, S., Application of artificial neural network for gold-silver deposits
potential mapping: a case study of Korea, Nat. Resour. Res. 19(2), 103–124, 2010.
Oja, E.A., Simplified neuron model as a principal component analyzer, J. Math. Biol.
15(3), 267–273, 1982.
Oja, E.A., Principal components, minor components, and linear neural networks,
Neural Net. 5(6), 927–935, 1992.
Ozkan, G., Ucan, L., Ozkan, G., The prediction of SO2 removal using statistical
methods and artificial neural network, Neural Comput. Appl. 19(1), 67–75, 2010.
Paranjape, S., Chen, S.W., Hibiki, T., Ishii, M., Flow Regime Identification under
Adiabatic Upward Two-Phase Flow in a Vertical Rod Bundle Geometry, ASME J.
Fluids Eng. 133(9), 091302, 2011 (8 pages).
Pham, D.T., Ghanbarzadeh, A., Koc, E., Otri, S., Rahim, S., Zaidi, S., The bees
algorithm, a novel tool for complex optimization problems. Proc. 2nd Int. virtual
Conf. intelligent production machines and systems, IPROMS 2006.
Plumb, A.P., Rowe, R.C., York, P., Brown, M., Optimization of the predictive ability
of artificial neural network (ANN) models: A comparison of three ANN programs
and four classes of training algorithms, Eur. J Pharm. Sci. 25, 395–405, 2005.
Poggio, T., Girosi, F., Regularization algorithms for learning that are equivalent to
multilayer networks, Science 247(4945), 978–982, 1990.
Poiseuille, J.L., Recherches expérimentelles sur le mouvement des liquids dans les
tubes de très-petits diamètres, Comptes Rendus, 12, 112–115 (1841).
Ramesh, M.N., Kumar, M.A., Rao, S., Application of artificial neural networks to
investigate the drying of cooked rice, J. Food Proc. Engg. 19(3), 321–329, 1996.
Reed, R., Pruning algorithms – a survey, IEEE Trans. Neural Net. 4(5), 740–747,
1993.
Rezkallah, K.S., Weber number based flow-pattern maps for liquid–gas flows at
microgravity, Int. J. Multiphase Flow 22(6), 1265–1270, 1996.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., Learning internal representations by

error propagation, In: Parallel Distributed Processing, Eds. Rumelhart, D.E. and
McClelland, J.L, MIT Press. 1986.
Sablani, S.S., Shayya, W.H., Kachimov, A., Explicit calculation of the friction factor
in pipeline flow of Bingham plastic fluids: A neural network approach. Chem.
Eng. Sci. 58(1), 99–106, 2003.
Sablani, S.S., Shayya, W.H., Neural network based non-iterative calculation of the
friction factor for power law fluids, J. Food Engg. 57(4), 327–335, 2003.
161
References
Samanta, A.K., Banerjee, T.K., Das, S.K., Pressure loses in orifices for the flow of
gas-non-Newtonian liquids, Can. J. Chem. Engg. 77(3), 579–583, 1999.
Sando, T., Renatus, M., Sabanjo, J. Sapinhour L., Advantages and disadvantages of
different crash modeling techniques, J. Safety Res. 36(5), 485–487, 2005.
Sanger, T.D., Optimal unsupervised learning in a single-layer linear feedforward

neural network. Neural Net. 2(6), 459–473, 1989.
Scholkoff, R., Artificial Neural Networks. McGraw-Hill, New York, 1997.
Scholkopf, B., Burges, C.J.C., Smola, A.J., Advances in kernel methods-Support

vector learning, Cambridge, MA: MIT Press, 1999.
Serizawa, A., Feng, Z.P., Two-phase flow in microchannels, Proc. 4th Int. Conf.
Multiphase Flow New Orleans, LA, USA, May 27– June 1, 2001.
Shaikh, A., Al-Dahhan M., Development of an artificial neural network correlation

for prediction of overall gas holdup in bobble column reactors, Chem. Eng. Proc.
42(8–9), 599–610, 2003.
Sharma, H., Das, G., Samanta, A.N., ANN–based prediction of two-phase gas-liquid
flow patterns in a circular conduit, AIChE J. 52(9), 3018–3028, 2006.
Sidani, A., Sidani, T., A comprehensive study of the backpropagation algorithm and
modifications, IEEE Conference Record, Orlando, FL, USA. 29 – 31 March, 80–
84, 1994.
Šima, J., Orponen, P., General-purpose computation with neural networks: a survey of
complexity theoretic results, Neural Comput. 15(12), 2727–2778, 2003.
Singha, B., Bar, N., Das, S.K., The use of artificial neural networks (ANN) for
modeling of adsorption of Cr(VI) ions, Des. Wat. Treat., 52(1–3), 415–425, 2014.
Sola, J., Sevilla, J., Importance of input data normalization for the application of
neural networks to complex industrial problems, IEEE Trans. Nuclear Sci. 44(3),
1464–1468, 1997.
Spedding, P.L., Spence, D.R., Prediction of holdup in two-phase flow, Int. J Engg.
Fluid Mech. 1, 67–82, 1988.
Stokes, G.G., On the theories of internal friction of fluids in motion, Trans. Camb.
Phil. Soc. 8, 287–305 (1845).
Sugawara, M., Numerical solution of the Schrödinger equation by neural network and
genetic algorithm, Comput. Phys. Commun., 140, 366–380, 2001.
Sun, B., Zhang, H., Cheng, L., Zhao, Y., Flow Regime Identification of Gas–liquid
Two-Phase Flow Based on HHT, Chinese J. Chem. Eng. 14(1), 24–30, 2006.
162
References
Sunde, C., Avdic, S., Pázsit, I., Classification of two-phase flow regimes via image
analysis and a neuro-wavelet approach, Prog. Nucl. Energy, 46(3–4), 348–358,
2005.
Suo, M., Griffith, P., Two-phase flow in capillary tubes, J. Basic Eng. 86, 576–582,
1964.
Taitel, Y., Flow pattern transition in two phase flow, Keynote lecture, Proc. 9th Int.
Heat Trans. Conf., Jerusalem, Israel, 19-24 August, No. KN-14, pp. 237–254,
1990.
Taitel, Y., Dukler, A.E., A model for predicting flow regime transitions in horizontal
and near horizontal gas-liquid flow, AIChE J. 22(1), 47–55, 1976.
Tambouratzis, T., Pázsit, I., Non-invasive on-line two-phase flow regime

identification employing artificial neural networks, Ann. Nucl. Energy, 36, 464–
469, 2009.
Tambouratzis, T., Pàzsit, I., A general regression artificial neural network for two-
phase flow regime identification, Ann. Nucl. Energy, 37, 672–680, 2010.
Torkar, D., Novak, S., Novak, F., Apparent viscosity prediction of alumina-paraffin
suspensions using artificial neural networks, J. Mat. Proc. Tech. 203(1–3), 208–
215, 2008.
Trafalis, T.B., Oladunni, O., Papavassiliou, D.V., Two-phase flow regime

identification with a multiclassification Support Vector Machine (SVM) model,
Ind. Eng. Chem. Res. 44(12), 4414–4426, 2005.
Triplett, K.A., Ghiaasiaan, S.M., Adbel-Khalik, S.I., Sadowski, D.L., Gas–liquid two-
phase flow in microchannels. Part I: Two-phase flow patterns, Int. J. Multiphase
Flow 25(3), 377–394, 1999.
Tsoukalas, L.H., Ishii, M., Mi, Y., A neurofuzzy methodology for impedance-based
multiphase flow identification, Eng. Appl. Artif. Intell. 10(6), 545–555, 1997.
Übeyli, E.D., Güler, I., Multilayer perceptron neural networks to compute quasistatic
parameters of asymmetric coplanar waveguides, Neurocomputing 62, 349–365,
2004.
Udelhoven, T., Schutt, B., Capability of feedforward neural networks for a chemical
evaluation of sediments with diffuse reflectance spectroscopy, Chemom. Intell.
Lab. Syst. 51(1), 9–22, 2000.
Vapnik, V., Statistical learning theory, Wiley, New York, 1998.
Venkatesan, M., Das, S.K., Balakrishnan, A.R., Effect of tube diameter on two-phase
flow patterns in mini tubes, Can. J. Chem. Eng. 88(6), 936–944, 2010.
163
References
Wallis, G.B., One dimensional two-phase flow, McGrew-Hill Book Co. Inc., New
York, 1969.
Weisman, J., Duncan, D., Gibson, J., Crawford, T., Effects of fluid properties and
pipe diameter on two-phase flow patterns in horizontal lines, Int. J. Multiphase
Flow 5(6), 437–462, 1979.
Werbos, P.J., Beyond regression: new tools for prediction and analysis in the
behavioural sciences. PhD dissertation. Harvard University, Cambridge, 1974.
Widrow, B., Winter, R.G., Baxter, R.A., Layered Neural Nets for Pattern
Recognition, IEEE Trans. Acoustics, Speech and Signal Processing 36(7), 1109–
1118, 1988.
Wu, Guan-De, Lo, Shang-L., Effects of data normalization and inherent-factor on

decision of optimal coagulant dosage in water treatment by artificial neural
network, Expert Syst. Appl. 37(7), 4974–4983, 2010.
Xiao, J., Shoham, O., Brill, J., A comprehensive model for two-phase flow in
pipelines, In: The 65th SPE Annual Technical Conference and Exhibition, SPE
20631, New Orleans, Sep, 1990.
Xie, T., Ghiaasiaan, S.M., Karrila, S., Flow regime identification in gas/liquid/pulp
fibre slurry flows based on pressure fluctuations using artificial neural networks,
Ind. Eng. Chem. Res. 42(26), 7017–7024, 2003.
Xu, L.J., Xu, L.A., Gas/liquid two-phase flow regime identification by ultrasonic
tomography, Flow Meas. Instrum. 8(3–4), 145–155, 1998.
Yang, C.Y., Shieh, C.C., Flow pattern of air–water and two-phase R-134a in small
circular tubes, Int. J. Multiphase Flow 27(7), 1163–1177, 2001.
Yari, E., Ayoobi, A., Ghassemi, H., Applying the Artificial Neural Network to
Estimate the Drag Force for an Autonomous Underwater Vehicle. Open J. Fluid
Dyn. 4, 334–346, 2014.
Yetilmezsoy, K., Demirel, S., Artificial neural network (ANN) approach for modeling
of Pb(II) adsorption from aqueous by antep pistachio (Pistacia Vera L.) shells, J.
Haz. Mat. 153(3), 1288–1300, 2008.
Yildiz, C., Gultekin, S., Guney, K., Sagiroglu, S., Neural models for the resonant
frequency of electrically thin and thick circular microstrip antennas and the
characteristic parameters of asymmetric coplanar waveguides backed with a
conductor, Int. J. Electron. Commun. 56(6), 396–406, 2002.
Yunlong, Z., Chen, F., Sun, B., Identification method of gas-liquid two-phase flow
regime based on image multi-feature fusion and support vector machine, Chin. J.
Chem. Eng. 16(6), 832–840, 2008.
164
References
Zhao, L., Rezkallah, K.S., Gas–liquid flow patterns at microgravity conditions. Int. J.
Multiphase Flow 19(5), 751–763, 1993.
Zhao, T.S., Bi, Q.C., Co-current air–water two-phase flow patterns in vertical
triangular micro channels, Int. J. Multiphase Flow 27(5), 765–782, 2001.
Zhao, Z.C., Zhang, X.D., Wang, D., Xu, J.M., Calculation of thixotropic stress of
waxy crude oil by BP artificial neural network, Dalian Ligong Daxue
Xuebao/Journal of Dalian University of Technology 45(2), 181–185, 2005.
Zhang, H.Q., Wang, Q., Sarica, C., Brill, J.P., Unified Model for Gas–Liquid Pipe
Flow via Slug Dynamics–Part 1: Model Development, J. Energy Resour. Technol.
125(4), 266–273, 2003a.
Zhang, H.Q., Wang, Q., Sarica, C., Brill, J.P., Unified Model for Gas–Liquid Pipe
Flow via Slug Dynamics–Part 2: Model Validation, J. Energy Resour. Technol.
125(4), 274–283, 2003b.
Appendix
LIST OF BOOKS
1) S.K. Das and N. Bar, "Hydrodynamics of Gas-non-Newtonian Liquid Flow

and ANN Predictibility," Lambart Academic Publishing. Saarbrücken,
Germany, 2013. ( ISBN: 978–3–659–40779–6)
Other related papers published in International Journals/ Conference Proceedings
1) Bar, N. and Das, S.K. 2011. Comparative study of friction factor by prediction
of frictional pressure drop per unit length using empirical correlation and
ANN for gas-non-Newtonian liquid flow through 180° circular bend,
International Review of Chemical Engineering 3(6) 628 – 643. (ISSN: 2035 –
1755)
2) Bar, N. and Das, S.K. 2012. Gas-non-Newtonian Liquid Flow Through
Horizontal Pipe – Gas Holdup and Pressure Drop Prediction using Multilayer
Perceptron, American Journal of Fluid Dynamics, 2(3) 7 – 16. (ISSN: 2168 –
4715)
3) Bar, N. and Das, S.K. 2013. Frictional Pressure Drop for Gas-Non-Newtonian
Liquid Flow through 90° and 135° circular bend: Prediction Using Empirical
Correlation and ANN, International Journal of Fluid Mechanic Research
39(5) 416 – 437. (ISSN: 1064 – 2277)
4) Singha, B., Bar, N. and Das, S.K. 2014. The use of artificial neural networks
(ANN) for modeling of adsorption of Cr(VI) ions, Desalination and Water
Treatment, 52(1-3) 415 – 425. (ISSN: 1944 – 3986)
5) Bar, N. and Das, S.K. Gas Holdup and Frictional Pressure Drop Prediction in
Vertical Pipes for Gas-non-Newtonian Flow Using ANN. 2nd International
Conference on Process Engineering and Advanced Materials, (ICPEAM
2012), Malaysia, June 12 – 14, 2012. (ISBN: 978 – 983 – 2271 – 81 – 9)
6) Bar, N., and Das, S.K. 2013. Prediction of Flow Regime for Air-Water Flow
in Circular Micro Channels using ANN. Procedia Technology 10 242 – 252.
(ISSN: 2212 – 0173)
7) Bar, N., and Das, S.K. 2014. Modeling of Gas Holdup and Frictional Pressure
Drop Using ANN for Gas-non-Newtonian Flow in Vertical Pipes. Advanced
Materials Research 917 244 – 256. (ISSN: 1662 – 8985)
8) Mitra, T., Singha, B., Bar, N. and Das, S.K. 2014. Removal of Pb(II) ions
from aqueous solution using water hyacinth root by fixed-bed column and
ANN modeling. Journal of Hazardous Materials 273 94 – 103. (ISSN: 0304 –
3894)
9) Das, B., Ganguly, U.P., Bar, N. and Das, S.K. 2015. Holdup prediction in
inverse fluidization using non-Newtonian pseudoplastic liquids: Empirical
Correlation and ANN modeling. Powder Technology 273(C) 83 – 90. (ISSN:
0032 – 5910)
Apart from the above papers there are 4 more papers (oral/poster) presented in
national conferences and 2 paper where I was an author had been presented in
national conferences by others.

Final PHD Thesis of Nirjhar Bar 2016

Uploaded by

Copyright:

Available Formats

Final PHD Thesis of Nirjhar Bar 2016

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Final PHD Thesis of Nirjhar Bar 2016

Uploaded by

Copyright:

Available Formats

APPLICATIONS OF NEURAL

NETWORKS FOR THE PREDICTION

Networks for the Prediction of Hydrodynamic Parameters in Liquid and

Gas-liquid Flow through Pipelines”, submitted by Nirjhar Bar to the

Maulana Abul Kalam Azad University of Technology, West Bengal

(Formerly known as West Bengal University of Technology) for the

award of the degree of Doctor of Philosophy in Science (Physics), is a

completed the work truthfully and successfully to the best of our

Prof. Manindra Nath Biswas Prof. Sudip Kumar Das

It is with real pleasure that, I record my indebtedness to my Academic

those whose name I have forgotten to mention here.

Paper published/communicated in International Journals and Conference

The thesis entitled Applications of neural networks for the prediction of

hydrodynamic parameters in liquid and gas-liquid flow through pipelines is

hydrodynamics of single and multi-phase flow system.

Chapter - 4 Gas-non-Newtonian liquid flow through pipes and piping

7.1 Introduction 110

Chapter - 8 Conclusions 147

Figure No. Titles

2.1 Schematic diagram of artificial neural network

2.2 Flowchart representing training of the single hidden layer

3.13 Comparison of pressure drop across the elbow for testing

3.14 Comparison of pressure drop across the orifice for testing

3.17 Comparison of pressure drop across the elbow for prediction

3.18 Comparison of pressure drop across the orifice for prediction

4.1–4.5 Variation of the minimum value of MSE with the number of

4.22 Comparison of two-phase pressure drop across the elbow for

4.23 Comparison of two-phase pressure drop across the orifice for

5.1 Variation of the minimum value of MSE with the number of

5.4 Comparison of the two-phase frictional pressure drop

5.5 Comparison of the two-phase frictional pressure drop for final

6.1 Variation of the minimum value of MSE with the number of

6.2 Variation of the minimum value of MSE with the number of

6.7 Comparison of two-phase frictional pressure drop across the horizontal

6.9 Comparison of two-phase frictional pressure drop across the horizontal

7.1 Graphical representation of data for sample 1

7.2 Graphical representation of data for sample 2

7.3 Graphical representation of data for sample 3

7.4 Graphical representation of data for sample 4

7.5 Graphical representation of data for sample 5

7.6 Variation of minimum value of cross-validation with the number of

7.8 Variation of minimum value of cross-validation with the number of

7.9 Cross-validation curve for MLP network trained with LM algorithm

7.10 Variation of minimum value of cross-validation with the number of

7.12 Variation of minimum value of cross-validation with the number of

7.14 Cross-validation curve for SVM network for all 5 samples

Table No. Title

3.1 Description of the data used for neural network analysis

3.3 Optimum number of nodes for different systems

3.4 Performance of neural network for training for 5 runs

3.5 Performance of neural network for cross-validation for 5 runs

3.6 Performance of neural network for testing

3.7 Performance of the neural network for prediction

3.8 Comparison between experimental and ANN prediction for elbow

3.9 Comparison between experimental and ANN prediction for orifice

3.11 Comparison between experimental and ANN prediction for globe

4.1 Description of the data used for neural network analysis

4.3 Performance of best neural network for training over 5 runs

4.4 Performance of the neural network for cross-validation over 5 runs