i

In partial fulfillment of the requirement for the degree of

Master of Technology

in

Department of Mechanical Engineering

by

Gaurav Katendra
Roll No. 211ME3172


Under the guidance of
Dr. Suman Ghosh

Department of Mechanical Engineering
National Institute of Technology
Rourkela -769 008 (India)
June, 2013


ii





















iii

National Institute of Technology

Rourkela



CERTIFICATE

This is to certify that the report entitled PREDICTION OF COATING
THICKNESS THROUGH THERMAL ANALYSIS submitted to the
National Institute of Technology, Rourkela by Gaurav Katendra, Roll
No. 211ME3172 for the award of the Degree of Master of Technology in
Department of Mechanical Engineering with specialization in Thermal
Engineering is a record of bona fide work carried out by him under
my supervision and guidance.


Place: NIT Rourkela, Rourkela769008, Orissa. Date: June 03, 2013




Dr. Suman Ghosh
Assistant Professor
Department of Mechanical Engineering
NIT Rourkela, Rourkela769008, India



iv

v

ACKNOWLEDGEMENT


First and foremost I offer my sincerest gratitude and respect to my project supervisor,
Dr. SUMAN GHOSH in the Department of Mechanical Engineering, for their
invaluable guidance and suggestions to me during my study. I consider myself
extremely fortunate to have had the opportunity of associating myself with them for
one year. This thesis was made possible by their patience and persistence.
After the completion of this Thesis, I experience the feeling of achievement
and satisfaction. Looking into the past I realize how impossible it was for me to
succeed on my own. I wish to express my deep gratitude to all those who extended
their helping hands towards me in various ways during my short tenure at NIT
Rourkela.
I express my sincere thanks to Professor K.P. MAITY, HOD, Department
of Mechanical Engineering, and also the other staff members of Department of
Mechanical Engineering, NIT Rourkela for providing me the necessary facilities that
is required to conduct the experiment and complete my thesis.
Last but not the least I am especially indebted to my parents and friends for
their love, sacrifice, and support. They are my first teachers after I came to this world
and have set great examples for me about how to live, study, and work



Date: GAURAV KATENDRA
















vi


















vii

Contents
Description Page no
Title i
Certificate by Supervisor iii
Acknowledgement v
Contents vii
List of Figures ix
Abstract xi
Chapter1: Introduction and Literature Review 1
1.1Introduction 1
1.2 Literature Review 2
1.3 Gaps in the Literature 5
1.4 Aims and Objective 5
1.5 Organization of the Thesis 5

Chapter 2: Problem Statement 7
2.1 Problem descriptions 7
2.1.1 Forward problem 7
2.1.2 Prediction problem 8

Chapter 3: Methodology 9
3.1 Data Collection 9
3.1.1 Numerical calculation 11
3.1.1.1 Governing equation 12
3.1.1.2 Boundary condition 12
3.1.1.3 Grid pattern employed 12
3.1.1.4 Residual and convergence 13
3.2 Prediction Methodology 14
viii


Chapter 4: Results and Discussion 17
4.1 Grid Independence Test 17
4.2 Prediction of the Coating Thickness 22
4.2.1 Uniform thickness 24
4.2.1 Non-uniform thickness 27

Chapter 5: Concluding Remarks and Scope for Future
Work 37
5.1 Concluding Remarks 37
5.2 Scope of the Future Work 38

References 41



ix

LIST OF FIGURES

Figure No. Description Page No.
Figure 2.1

The schematic representation of square steel block with a
copper coating.
8

Figure 3.1 Schematic representation of data collection. 11

Figure 3.2 A schematic representation of the grids for a particular
coating thickness.
13
Figure 3.3 A schematic representation of NN 15


Figure 3.4 A schematic showing the proposed GA-NN methodology 16

Figure 4.1


Variation of temperature profile with start size of grid of
inner face.
18

Figure 4.2 Variation of maximum temperature at the bottom wall
with start size of the grid of inner face.
19


Figure 4.3 Variation of temperature profile with aspect ratio of inner
face.
20

Figure 4.4


Variation of maximum temperature at the bottom wall
with aspect ratio of inner face.
20

Figure 4.5 Variation of temperature profile with end size of grid. 21


Figure 4.6


Variation of maximum temperature at the bottom wall
with end size of grid for inner face
21

Figure 4.7 Variation of temperature profile at bottom with grid size
of outer face.
22

Figure 4.8 Variation of maximum temperature at the bottom wall
with the varying grid size of outer face.
22

Figure 4.9 Plot between original thickness during simulation and
predicted thickness.(uniform)
25


x

Figure 4.10 Comparison between the thickness during simulation and
the predicted thickness through the developed
methodology.(uniform)
27

Figure 4.11 Plot between original thickness during simulation and
predicted thickness.(nonuniform bottom)
28

Figure 4.12 Comparison between the thickness during simulation and
the predicted thickness through the developed
methodology.(nonuniform bottom)
29

Figure 4.13 Plot between original thickness during simulation and
predicted thickness.(nonuniform left)

30

Figure 4.14 Comparison between the thickness during simulation and
the predicted thickness through the developed
methodology. (nonuniform left)
31
Figure 4.15 Plot between original thickness during simulation and
predicted thickness. (nonuniform top)

32
Figure 4.16 Comparison between the thickness during simulation and
the predicted thickness through the developed
methodology. (nonuniform top)
33
Figure 4.17 Plot between original thickness during simulation and
predicted thickness. (nonuniform right)
34
Figure 4.18 Comparison between the thickness during simulation and
the predicted thickness through the developed
methodology. (nonuniform right)
35












xi

Abstract

An attempt has been made to develop a methodology to predict the uniform or non-
uniform thickness of a coating material on a homogeneous solid material. At first the
effects of variation of (uniform and non-uniform) coating thickness, on the heat
transfer (by conduction) through the coated body have been rigorously investigated. A
solid square steel block coated with copper (of various coating thicknesses) has been
considered as the sample problem. A constant heat flux has been maintained at the
bottom side where, the other three sides kept at constant temperature. The temperature
profile built at the bottom side of the square block is different for different (uniform
or non-uniform) coating thickness. Depending on the statistical parameters (Mean,
Standard Deviation, Skew-ness and Kurtosis) extracted from the temperature profile
(which has been found through CFD analysis), it may possible to predict the thickness
of the coating. Two cases namely: uniform varying thickness and non-uniform
varying thickness have been considered in the present problem. The training and test
data have been collected using 2D conduction analysis with Finite Volume Method
(FVM) in a fully automated way. Feed-Forward Neural Network (NN) with Back-
Propagation Algorithm is used to predict the coating thickness. The NN has been
modelled using C programming Language in LINUX operating system. A Genetic
Algorithm (GA) has also been coded (using C programming language in Linux
operating system) to optimize the performance of the NN. The above described
methodology based on thermal analysis can also be used in thermography or in other
prediction problem.

Keywords: Heat Conduction, FVM, Neural Network, Genetic Algorithm, Mean,
Standard Deviation, Skew-ness and Kurtosis.








xii
























1

Chapter 1:
Introduction and
Literature Review


1.1 Introduction
The steady-state heat conduction problems find several useful applications in the wide
era of developing technology. The multi-layered coated materials have numerous
industrial applications, such as transportation, aerospace, electronics and medical. In
such coated materials, the 2D or 3D analysis of heat conduction is one of the
furthermost significant corporeal phenomena which are needed to be considered.
Surface coating plays a very vital role in the manufacturing of a particular
product. The coatings provide surface protections against high temperature, wear,
corrosion, oxidation, and electrical conduction. The prediction of coating thickness is
very necessary for the safety of inner material (on which the coating has been done).
The study of thermal conduction across defective component (material with corroded
coating) plays a critical role in predicting the coating thickness for such material. By
knowing the temperature profile of any coated material it is possible to predict its
coating thickness.
The application of the above concept may be used in Thermography. In
medical science, Thermography is the technique that derives analytic suggestions
from vastly detailed and sensitive infrared images of the human body and it is
2

completely non-contact, involves no form of energy imparted onto or into the body.
Thermography has anticipatable applications in breast oncology, neurology,
orthopaedics, occupational medicine, pain management, vascular medicine/cardiology
and veterinary medicine. Thermography has a long history, although its use has
increased intensely with the commercial and industrial applications of the past fifty
years.
1.2 Literature Review
Several contributions of 2D heat conduction problem are available in scientific books
and journals. Firstly, a generalized fundamental solution with the estimation of
hardened layer dimensions in laser surface hardening processes with variations of
coating thickness has been proposed by Woo and Cho (1996). The heat conduction
across the interface between a thin isotropic thermal barrier coating and an anisotropic
substrate has been analysed by Shiah and Shi (2006) using the boundary element
method. Thin-film coating plays a prominent role on the manufacture of many
industrial devices. Coating can increase material performance due to the deposition
process. This paper proposes the estimation of hardness of titanium thin-film layers as
protective industrial tools by using multi-layer perceptron (MLP) neural network.
Based on the experimental data obtained during the process of chemical vapour
deposition (CVD) and physical vapour deposition (PVD), the optimization of the
coating variables for achieving the maximum hardness of titanium thin-film layers, is
performed. This optimization of coating variables for hardness of industrial tools by
using artificial neural networks has been demonstrated by Yazdi, Khorasani and Faraji
(2011). Prediction and control of surface roughness in CNC lathe is done by using
artificial neural network and this has been proposed by Karayal (2009). Effects of Ti-
and Zr-based interlayer coatings on hot-filament chemical vapour deposition of
3

diamond on high-speed steel has been analysed by Polini, Mantini, Braic, Amar,
Ahmed and Taylor (2006). Optimization of machinery variable by using micro
multivariable genetic algorithm has been proposed by Narimani and Rohani (2007).
The use of design of experiments to improve a neural network model in order to
predict the thickness of the chromium layer in a hard chromium plating process has
been explained by Lasheras, Vilan, Neito and Diaz (2010). A ferrite-core probe design
has several parameters, which includes the selection of ferrite material and the design
of the probe structure and core and in this article a total of five variables, which are
most influential on probe performance, were considered. A three-layer back-
propagation neural network model is used to correlate these variables to probe
performance using the data generated based on experimental observations and it is
used for predicting the coating thickness and performance of ferrite-core probe using
neural network approach which has been analysed by Baseri and Damirchi (2009).
Prediction of the coating thickness of wire coating extrusion processes using artificial
neural network (ANN) has been demonstrated by Cirak and Kozan (2009). Here, a
three layer back propagation artificial neutral network (ANN) model was used for the
description of wire coating thickness and on comparing the experimental data with the
ANN model prediction; it is found that the ANN model is capable of predicting the
coating thickness. The neural network model shows how the significant parameters
influencing thickness can be found and in this study, a back propagation neural
network model is developed to map the complex non-linear wire coating thickness
between process conditions. Integrated ANNGA for estimating the minimum value
for machining performance has been analysed by Zain, Haron and Sharif (2011).
Modelling of electrostatic fluidized bed (EFB) coating process using artificial neural
networks has been proposed by Barletta and Guarino (2006). Here, a verification
4

experimental plan was performed and a related analytical model was developed to
check the reliability of the neural network model with GA to predict the whole
coating thickness trends according to the operational parameters and a comparison
between the neural network model and an analytical model was also carried out.
Optimization of a photoresist coating process for photolithography in wafer
manufacture via a radial basis neural network has been explained by Shie and Yang
(2005). A grey-fuzzy Taguchi approach for optimizing multi-objective properties of
zirconium-containing diamond-like carbon coatings has been proposed by Yang and
Huang (2011). Regression and ANN models for estimating minimum value of
machining performance has been demonstrated by Zain, Haron, Qasem and Sharif
(2011). Two-dimensional optimization of material composition of functionally graded
materials using mesh less analyses and a genetic algorithm has been analysed by
Goupee and Vel (2004). A simulation technique for predicting the coating thickness
of thermal sprayed coatings has been proposed by Goedjen, Miller and Brindley
(1995). Here, a two dimensional finite difference simulation model has been
developed to predict the thickness of coatings deposited using thermal spray process
and the minimum coating thickness limit was predicted in order to determine the
application range of slot die coating. For this, a simulation technology is prepared to
predict the coating thickness limit in slot die coating by Yoshifumi and Makoto
(2010). Detection of defects in thermal barrier coatings by thermography analyses has
been done by Sakamoto, Kishi, Shobu, Tabaru, Tateyama and Akimune (2003).
Thermography is widely used medical. Thermography as a technique for
monitoring early age temperatures of hardening concrete has been analysed by
Azenha, Faria and Figueras (2011). The usage of IR thermography for the temperature
measurements inside an automobile cabin has been demonstrated by Korukcu and
5

Kilic. A review of thermography as promising non-invasive detection modality for
breast tumour has been proposed by Ng (2009).The study reveals the effect of heat
conduction on the thickness of coating material due to the change in temperature
profile.
1.3 Gaps in the Literature
As concluded from above, No theoretical models have employed GA-NN for
predicting the coating thickness according to the best of the authors knowledge.
Hardly any few of the models are perfectly able to correlate the experimental results
with those of modelled ones.
1.4 Aims and Objective
The current work examines the effect of coating thickness of material on heat
conduction due to the change in temperature profile of solid material. The temperature
variation on surface of the solid bar has been used here to predict the thickness of the
coating material. Then a methodology is proposed using GA-tuned neural Network to
predict the thickness of coating by knowing the temperature distribution on the
surface. The basic goal is to correlate the statistical parameters of the temperature
distribution on the surface with the thickness of coating material.
1.5 Organization of the Thesis
The present thesis has been organized in five chapters. The present chapter (that is,
Chapter 1) introduces the problems studied in the present dissertation. To define the
aims and objective of the present dissertation, an exhaustive literature survey is
carried out related to the present problem. In Chapter 2, the detailed description about
the problem formulation is given. The developed methodology to solve the defined
problem is described in Chapter 3. Chapter 4 deals with the obtained results and
6

consequence discussions. Chapter 5 focus on the important findings concluded from
the present study. In this chapter, the scope for future work related to the problem is
also been described.













7

Chapter 2:
Problem Statement


In the present chapter, the present problem is described in a detailed way endeavoured
in the current thesis.
2.1 Problem descriptions
Figure 1 shows the schematic representation of the problem consisting of a square
steel block with a copper coating on its surface. The dimensions of the square block
and its square coating are also shown in Figure 2.1. Each side of steel square block
represented by a and has been kept equal to 1.0 m. The thickness of copper coating
is denoted by b (b is varying case to case). All the locations or positions in the
present problem are considered with respect to a two-dimensional coordinate system
with its origin set at the centre of the block, where positive x and y axes are
considered along the right and upper directions, respectively.
2.1.1 Forward Problem
Now, the left, top and right sides of square are kept at constant temperature with wall
boundary condition. The bottom side has been given wall boundary condition with
constant heat flux. Once the geometry, material with these boundary conditions are
specified, the temperature distribution at the bottom side (bottom wall) can be
determined by solving two dimensional heat conduction problems numerically using
Finite Volume method. This establishes the forward heat transfer problem.
8

2.1.2 Prediction problem
After solving the forward problem, the objective is decided to predict the coating
thickness of the solid block from some known temperature distribution at the bottom
side (bottom wall) of the block.



Figure 2.1: The schematic representation of square steel block with a copper coating.






Square
steel
block
Copper coating
9

Chapter 3:
Methodology


Due to heat conduction by the constant heat flux used at the bottom wall, the heat
propagates towards the other three sides of the block. Now, the temperature
distribution within the block can be found by solving two-dimensional heat
conduction problem. For each coating thickness, there will be a distinct temperature
profile at the bottom wall. To solve the 2D heat transfer problem numerically, the two
commercial software (Gambit and Fluent) have been used. Gambit is used for
constructing geometry and meshing where Fluent is used for numerical calculation
(using Finite Volume method) purpose.
3.1 Data Collection
For a particular coating thickness (of copper) on the square steel block there will be a
particular temperature profile at the bottom wall. This particular coating thickness
with corresponding temperature profile (in terms of temperature values at some
positions along the bottom wall with a regular interval) constitutes one complete data.
The temperature profile is recorded in terms of temperatures at the different positions
across the bottom wall of square block. In the same way, a number of data have been
collected for the uniform and non-uniform varying thickness. This leads to the
forward problem. The data (temperature at different positions across the bottom wall
of square block) have been evaluated by using CFD software i.e. Gambit and Fluent.
10

In this way, a set of training and test data have been collected for both uniform and
non-uniform varying thickness. The network has been trained using the training data
and the test data is used to check the prediction capability of developed methodology.
For a specific coating thickness, Gambit is used to model the problem and then the
data has been kept in a mesh file. Now, Fluent has been used to solve the 2D heat
conduction problem taking that mesh file as input to achieve the temperature profile at
the bottom wall of the solid block. In this way, one complete set of data has been
collected. This process goes on until the complete data for all the different coating
thickness have been collected. For the whole data collection, Gambit and Fluent are to
be used simultaneously on the same task again and again, which is very problematic
to do manually. Here, for solution of this problem, Gambit and Fluent journal file are
used. The journal file of Gambit contains all the instructions required to create the
mesh file in a proper way and the Fluent journal file contains the instructions required
to solve the heat conduction problem. Now for a particular coating thickness, the
Gambit and Fluent operation occurred in a fully-automated way using the above
mentioned journal file respectively. At first, the coating thickness is calculated and
then, this information is given to the Gambit journal file which is then modified. Now,
Gambit is launched with this modified journal file and starts its operation to create the
mesh file. The mesh file has been exported after completion of its operation. Then,
Fluent starts its operation with Fluent journal file and this newly generated mesh file.
Figure 3.1 shows a schematic representation for data collection.

11


Figure 3.1: Schematic representation of data collection.

3.1.1 Numerical calculation
The data have been numerically calculated by using software i.e., Gambit and Fluent.
Using Gambit the problem has been modeled and discretized for a particular coating
thickness as shown in Figure 3.2, and then this information has been stored in mesh
file. The two-dimensional heat conduction problem is solved by using Fluent, where
the mesh file has been taken as input and simultaneously, the temperature profile
across the bottom wall have been achieved. Statistical analyses have been made to
express the temperature profiles in term of four parameters i.e., mean, standard
deviation, skew-ness and kurtosis. The required statistical analysis to evaluate these
statistical parameters (mean, standard deviation, skew-ness and kurtosis), coding has
been made in MATLAB. A particular coating thickness with corresponding set of
statistical parameters (mean, standard deviation, skew-ness and kurtosis) constitutes
Calculation of coating thickness
(m)

Gambit journal file modified

Launch Fluent with Fluent
Journal file and mesh file

Launch Gambit with journal file


Solution of 2D heat conduction
problem

Gambit creates mesh file

Fluent exports the temperature
profile at bottom

Read the export temperatures
and write down in a file

Write down the coating thickness
in a file

Start
No Yes
Written data Maximum
numbers of data

End
12

one data. This will be repeated till the data set for all the different coating thickness
with the square steel block have been collected.
3.1.1.1 Governing equation
Here, 2D steady state heat conduction with no heat generation is considered in order
to calculate the coating thickness of solid block. In 2D steady state heat conduction
with no heat generation the governing equation is
0
T T
k k
x x x y
| | c c c c | |
+ =
| |
c c c c
\ .
\ .
(3.1)

3.1.1.2 Boundary condition
The Wall boundary condition in Fluent has been used on the left, top, right and
bottom side of the square block (Figure 2.1). At the wall boundaries, the non-slip and
no penetration conditions have been applied, that is, at wall boundaries u = 0, v = 0.
Now, the left, top and right side of the square block are kept at a constant temperature.
But at the bottom side of the square a constant heat flux is maintained.
At left, top and right wall, Constant specified value T = (3.2)
At bottom wall Costant specified value
T
k
y
| | c
=
|
c
\ .
(3.3)

3.1.1.3 Grid pattern employed
For the inner face (steel), the triangular grid has been taken with a start size of 0.0025
m, aspect ratio 1.5 and end size of 0.02 m. For the outer face (copper coating),
uniform triangular grid has been taken with a grid size of 0.0015 m.
13


Figure 3.2: A schematic representation of the grids for a particular coating
thickness.

3.1.1.4 Residual and convergence
Residuals are used for checking the convergence of solution. For the discretization of
the governing equations, a first order upwind differencing scheme has been employed.
After discretization, the conservation equation for variable at a cell q can be
written as
q q nb nb
nb
a a b | | = +

(3.4)

Here, q is the centre coefficient,
nb
are the influence coefficient for the neighbouring
cells, and b is the constant part of the source term S
c
in S=S
c
+S
q
and of boundary
conditions, in equation (3.4)
q nb q
nb
a a S =

(3.5)
In equation (3.4), the imbalance summed over all the computational cells q is denoted
by the residual R

which is computed by Fluents pressure based solver as

14

_
nb nb q q
cells q nb
R a b a
|
| | = +

(3.6)
The above equation represents the unscaled residual. Since no scaling is employed,
it is very difficult to judge the convergence by examining the residuals defined by
equation (3.6). To scale the residual, a scaling factor (representative of flow rate of
through the domain) is used. The scaled residual is given as
_
_
nb nb q q
cells q nb
q q
cells q
a b a
R
a
|
| |
|
+
=

(3.7)
For most of the problems the scaled residual is an appropriate indicator of
convergence. Fluent is used for computing the scaled residuals. Once the scaled
residuals falls below the given value, it is presumed that the convergence is reached.
In this study, the value of 110
-6
was selected for the absolute criteria of continuity, x-
velocity and y-velocity. The absolute criterion of energy is 110
-9
.
3.2 Prediction Methodology
To construct the training and test data an exhaustive statistical analysis has been done
using MATLAB software. After getting the temperature distribution through CFD
analysis, four statistical parameters have been extracted which are used as input to
Neural Network. The statistical parameters which have been extracted are mean,
standard deviation, skew-ness, and kurtosis which are used as input to the Neural
Network. The coating thickness has been used as the output the Neural Network.
Figure 3.3 shows the schematic representation of Neural Network.

15


Figure 3.3: a schematic representation of NN



Four neurons have been used as input. Only one neuron has been used as output. Only
one hidden layer has been used. A GA tuned feed-forward neural network with back-
propagation algorithm has been used. The parameters optimized through GA are
learning rate, momentum constant, constant slope in the input layer, transfer function
coefficient in hidden layer, transfer function coefficient in output layer and the
number of neuron in hidden layer. The NN has been modelled using C programming
Language in LINUX operating system. Figure 3.4 shows the schematic representation
of the combined GA-NN approach.


OUTPUT
I
N
P
U
T
S

16


Figure 3.4: A schematic showing the proposed GA-tuned NN methodology
Yes
N
o
Generation Maximum Generation?
String Number Population Size?
End
Evaluation of control variables
from GA-string
Formation of Neural network

N
o
Yes
Mutation
Reproduction

Crossover
Generation = Generation + 1
Generation = 0
Star
t
Assignment of this RMS error as
the fitness value of the
corresponding string
String Number =
String Number + 1
Start back-propagation algorithm

Calculation of percentage error

No
No
17

Chapter 4:
Results and Discussion


This chapter contains the obtained results and corresponding discussions for the
selected problem using the developed methodology. Firstly, the detail about the grid
independence test has been mentioned. Then the results obtained through developed
methodologies for the prediction of uniform and non-uniform coating thickness have
been presented and discussed.
4.1 Grid Independence Test
The temperature profiles at bottom wall of the solid block have been generated
numerically using Finite Volume Method. Before producing results, a brief discussion
on grid independent test has been given. A triangular pave type grid has been adopted
here. The grid size has been varied keeping other parameters (size of block, thickness
of coating) fixed. Corresponding to each value of the grid size a mesh file has been
generated and consequently the temperature profile at bottom wall has been evaluated.
For selecting the grid size of inner face a parametric study of three parameters (start
size, aspect ratio and end size) has been carried out. Firstly, the start size is varied
keeping the value of other two parameters (aspect ratio and end size) fixed. Figure 4.1
shows the temperature profiles with varying start size of grid (other two parameters
kept constant) for the inner face. The maximum temperature at the bottom wall has
18

been evaluated corresponding to each start size of grid (keeping other two parameters
fixed) and a plot has been made between maximum temperature at bottom wall and
grid size as shown in Figure 4.2, to capture the variation of maximum temperature
with start size of grid. There is a range of start size where there is no significant
change in temperature profile (in Figure 4.1) or maximum temperature (in Figure 4.2).
Then, a value of grid size lying in this range has been chosen as start size of grid for
the inner face in Gambit for data (training and test) generation. Through this study,
the optimum start size has been selected and its equal to 0.0025 m.


Figure 4.1: Variation of temperature profile with start size of grid of inner face
(varying star size).



-0.51 -0.40 -0.30 -0.20 -0.10 0 0.10 0.20 0.30 0.40 0.51
400
500
600
700
800
900
1000
1100
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.0005 m
0.0010 m
0.0015 m
0.0020 m
0.0025 m
0.0030 m
0.0035 m
0.0040 m
0.0045 m
0.0050 m
0.0055 m
0.0060 m
0.0065 m
0.0070 m
0.0075 m
0.0080 m
0.0085 m
0.0090 m
0.0095 m
0.0100 m
19


Figure 4.2: Variation of maximum temperature at the bottom wall with start size of
the grid of inner face.


Now, the aspect ratio is varied keeping the other parameters (start size and end size)
fixed. The optimum value of aspect ratio has been selected in the same way as before
and it is equal to 1.5. Figure 4.3 shows the variation of temperature profile with the
varying aspect ratio (keeping other two parameters constant) for the inner face. Figure
4.4 shows the variation of maximum temperature at the bottom wall with aspect ratio.
At last, the end size is varied keeping the other parameters (start size and ratio) fixed.
Through this study, the optimum value of end size has been selected and its equal to
0.02 m. Figure 4.5 shows the variation of temperature profile with variation in end
size (other two parameters kept constant) for the inner face. Variation of maximum
temperature at bottom wall with the varying end size is shown in Figure 4.6. Variation
of temperature profile of bottom wall with uniform grid size of outer face is shown in
Figure 4.7. Figure 4.8 shows the variation of maximum temperature at bottom wall
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
x 10
-3
400
500
600
700
800
900
1000
1100
1200
1300
1400
1500
Start size (m)
M
a
x
i
m
u
m

t
e
m
p
e
r
a
t
u
r
e

a
t

t
h
e

b
o
t
t
o
m

w
a
l
l

(
K
)
20

with uniform grid size of outer face. Through this test, the uniform grid size for the
outer face has been decided as 0.0015 m.


Figure 4.3: Variation of temperature profile with aspect ratio of inner face.



Figure 4.4: Variation of maximum temperature at the bottom wall with aspect ratio of
inner face.
-0.51 -0.40 -0.30 -0.20 -0.10 0 0.10 0.20 0.30 0.40 0.51
400
500
600
700
800
900
1000
1100
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.0
1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
400
500
600
700
800
900
1000
1100
1200
1300
1400
1500
Ratio
M
a
x
i
m
u
m
T
e
m
p
e
r
a
t
u
r
e

(
K
)
21


Figure 4.5: Variation of temperature profile with end size of grid.







Figure 4.6: Variation of maximum temperature at the bottom wall with end size of
grid for inner face





-0.51 -0.40 -0.30 -0.20 -0.10 0 0.10 0.20 0.30 0.40 0.51
400
500
600
700
800
900
1000
1100
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.0025 m
0.0050 m
0.0075 m
0.0100 m
0.0125 m
0.0150 m
0.0175 m
0.0200 m
0.0225 m
0.0250 m
0.0275 m
0.0300 m
0.0325 m
0.0350 m
0.0375 m
0.0400 m
0.0025 0.005 0.0075 0.01 0.0125 0.015 0.0175 0.02 0.0225 0.025 0.0275 0.03 0.0325 0.035 0.0375 0.04
400
500
600
700
800
900
1000
1100
1200
1300
1400
1500
End size (m)
M
a
x
i
m
u
m

T
e
m
p
e
r
a
t
u
r
e

(
K
)
22


Figure 4.7: Variation of temperature profile at bottom with grid size of outer face.



Figure 4.8: Variation of maximum temperature at the bottom wall with the varying
grid size of outer face.


4.2 Prediction of Coating Thickness
Firstly the training data has been used to train the Neural-Network (NN) and then, the
test data has been used to predict the coating thickness of solid block. Here, a Genetic
-0.51 -0.40 -0.30 -0.20 -0.10 0 0.10 0.20 0.30 0.40 0.51
400
500
600
700
800
900
1000
1100
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.00025 m
0.00050 m
0.00075 m
0.00100 m
0.00125 m
0.00150 m
0.00175 m
0.00200 m
0.00225 m
0.00250 m
0.00275 m
0.00300 m
0.00025 0.0005 0.00075 0.001 0.00125 0.0015 0.00175 0.002 0.00225 0.0025 0.00275 0.003
400
500
600
700
800
900
1000
1100
1200
1300
1400
1500
Grid size (m)
M
a
x
i
m
u
m

T
e
m
p
e
r
a
t
u
r
e

(
K
)
23

Algorithm (GA) has also been coded to optimize the performance of Neural Network.
Then, a comparison has been made between the predicted thickness and the original
thickness.
In NN here, there are six parameters; i.e. learning rate (), momentum factor
(), slope of linear transfer function of input layer (
I
), transfer function coefficient of
hidden layer (
H
) and output layer (
O
) and lastly the number of neurons in the hidden
layer (N
H
). The neurons are assumed to have linear transfer function in the input
layer, whereas neurons of hidden and output layer are assumed to have log-sigmoid
transfer function. A GA is used to optimize these NN parameters. The value of
particular parameter for which the average training percentage error is minimum, has
been treated as optimum one. The average percentage error is defined as follows



1 1
1 1
100
z Z K
Ozk Ozk
z k
Ozk
T O
Z K T
=
= =
( | |


(
` |
( \ . )

(4.1)

where Z is the number training/test situations and K is the number of outputs, T
Ozk

indicates target output of k-th neuron corresponding to the z-th situation, O
Ozk
denotes
the network predicted value for the same. The performance of optimized NN has been
tested with the help of test data set.
Here, a brief out-line is presented about the working principle of GA. The GA
starts with the different set of NN parameters in the present case, which are generated
at random. The field size has been chosen as 12 for each of the variable. As there are
six variable, the chromosome becomes 72. Hence, a specific GA-string will look as
follows:

I

H

O
N
H


000010000100 000000000100 111001001000 010011001100 111110000001 100101110001
24

The first 12 bits, that is 000010000100 yields the decoded value (DV) as 02
11
+
02
10
+ 02
9
+ 02
8
+ 12
7
+ 02
6
+ 02
5
+ 02
4
+ 02
3
+ 12
2
+ 02
1
+ 02
0
=
132. The real value of can be determined following the linear mapping rule as given
below.


,max ,min
,min
2 1
I I
I I fieldsize
a a
a a DV

= +

(4.2)
Similarly, the real value of the other variables can also be calculated. After finding the
real values from the GA string the network is formed. Now, a BP algorithm is used to
optimize the weights of the network, so that it can predict the output with minimum
error. After this, the fitness values of all the members of population are calculated. In
the present problem, the fitness value is the sum of the percentage training and test
error in prediction. Two cases namely: uniform thickness and non-uniform thickness
have been considered in the present problem. To find an optimal set of GA parameter
(like population size, number of generation, crossover probability, mutation
probability, tournament size), a parametric study has been made. In this study, a
particular parameter has been varied from its minimum value to its allowable
maximum value keeping the other parameters fixed to find its optimal value. Once an
optimum value for a particular parameter has been found, it is keep fixed for
remaining part of study. A similar process is carried out, to find the optimal values of
other parameters.
4.2.1 Uniform thickness
Here, thickness of coating across each of the side of the solid block is equal. The
optimized fitness value (optimized through GA) got in this case is equal to 0.084285
and the corresponding optimized NN parameter values (optimized through GA) are
as follows: learning rate () = 0.033211, momentum factor () = 0.000977, slope of
25

linear transfer function of input layer (
I
) = 2.678388, transfer function coefficient of
hidden layer (
H
) = 1.499389, transfer function coefficient of output layer (
O
) =
4.846154 and lastly the number of neurons in the hidden layer (N
H
) = 19. Using this
optimum set of NN parameters, the predicted thickness corresponding to each
individual input set of the test cases has been calculated. Now, a comparison has been
made between the predicted and original thickness of coating. The Figure 4.9 shows
this comparison between original thickness during simulation and predicted thickness
for each individual test case. In the Figure, a line passing through all the points has
been drawn. A line making an angle 45
0
with x or y axis is also been drawn. The
deflection between these two lines showing the gap between the original and
predicted thickness.


Figure 4.9: Plot between original thickness during simulation and predicted thickness.

26

Using optimum set of NN parameters, comparison between temperature
profile with original and predicted thickness has also been made. Figure 4.10a, 4.10b
and 4.10c shows this type of 2 comparisons.




-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
1200
1300
1400
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.0020m(Original)
0.001587m(Predicted)
-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.0120m(Original)
0.011292m(Predicted)
a
b
27

Figure 4.10(a, b and c): Comparison between the thickness during simulation and the
predicted thickness through the developed methodology.

4.2.2 Non-Uniform thickness
In this case, thicknesses of coating for the sides of the homogeneous solid block are
not equal to each other. Thickness of coating on one particular side is varying where
on the other sides the coating thickness remain unchanged. The optimized fitness
value (optimized through GA) got in this case is equal to 0.24874 and the
corresponding optimized NN parameter values (optimized through GA) are as
follows: learning rate () = 0.362637, momentum factor () = 0.037851, slope of
linear transfer function of input layer (
I
) = 1.310623, transfer function coefficient of
hidden layer (
H
) = 3.300366, transfer function coefficient of output layer (
O
) =
4.144078 and lastly the number of neurons in the hidden layer (N
H
) = 23. Using this
optimum set of NN parameters, the predicted thickness corresponding to each
individual input set of the test cases has been calculated. Now, a comparison has been
-0.4 -0.2 0 0.2 0.4 0.6
400
450
500
550
600
650
700
750
Dimension along the bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.0320m(Original)
0.030148m(Predicted)
c
28

made between the predicted and original thickness of coating. Here, there are four
cases for the four sides (bottom, left, top and right) respectively.
4.2.2.1 Thickness varying on bottom side (keeping fixed on the other sides)
With present condition, the Figure 4.11 shows the comparison between original
thickness during simulation and predicted thickness for each individual test case. In
the Figure, a line passing through all the points has been drawn. A line making an
angle 45
0
with x or y axis is also been drawn. The deflection between these two lines
showing the gap between the original and predicted thickness.


Figure 4.11: Plot between original thickness during simulation and predicted
thickness.

29

Using optimum set of NN parameters, comparison between temperature
profile with original and predicted thickness has also been made. Figure 4.12a and
4.12b shows this type of 2 comparisons.





Figure 4.12 (a, b): Comparison between the thickness during simulation and the
predicted thickness through the developed methodology.
-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
1200
1300
1400
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.002m(Original)
0.002486m(Predicted)
-0.4 -0.2 0 0.2 0.4 0.6
400
450
500
550
600
650
700
750
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.032m(Original)
0.030092m(Predicted)
b
a
b
a
30

4.2.2.2 Thickness varying on left side (keeping fixed on the other sides)
With the present condition, the Figure 4.13 shows the comparison between original
thickness during simulation and predicted thickness for each individual test case with
the present condition.


Figure 4.13: Plot between original thickness during simulation and predicted
thickness.


Using optimum set of NN parameters, comparison between temperature
profile with original and predicted thickness has also been made. Figure 4.14a and
4.14b shows this type of 2 comparisons.


31




Figure 4.14(a, b): Comparison between the thickness during simulation and the
predicted thickness through the developed methodology.




-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.002m(Original)
0.012861m(Predicted)
-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.032m(Original)
0.012954m(Predicted)
a
b
32

4.2.2.3 Thickness varying on top side (keeping fixed on the other sides)
In the same way, for the present condition, the Figure 4.15 shows the comparison
between original thickness during simulation and predicted thickness for each
individual test case.


Figure 4.15: Plot between original thickness during simulation and predicted
thickness.


Using optimum set of NN parameters, comparison between temperature
profile with original and predicted thickness has also been made. Figure 4.16a and
4.16b shows this type of 2 comparisons.

33









Figure 4.16(a, b): Comparison between the thickness during simulation and the
predicted thickness through the developed methodology.


-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e


0.002m(Original)
0.012876m(Predicted)
-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.032m(Original)
0.012880m(Predicted)
a
b
34

4.2.2.4 Thickness varying on right side (keeping fixed on the other sides)
With this condition, Figure 4.17 shows this comparison between original thickness
during simulation and predicted thickness for each individual test case.


Figure 4.17: Plot between original thickness during simulation and predicted
thickness.


Using optimum set of NN parameters, comparison between temperature
profile with original and predicted thickness has also been made. Figure 4.18a and
4.18b shows this type of 2 comparisons.

35




Figure 4.18(a, b): Comparison between the thickness during simulation and the
predicted thickness through the developed methodology.





-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.002m(Original)
0.013318m(Predicted)
-0.4 -0.2 0 0.2 0.4 0.6
400
500
600
700
800
900
1000
1100
Dimension along bottom of solid bar (m)
T
e
m
p
e
r
a
t
u
r
e

(
K
)


0.032m(Original)
0.013934m(Predicted)
b
a
36


37

Chapter 5:
Concluding Remarks
and Scope for Future
Work


5.1 Concluding Remarks
Attempt has been made to develop a methodology for the prediction of coating
thickness by conduction analysis using C programming language and commercial
CFD software: Gambit and Fluent. The performance of the developed methodology
has been tested to predict the uniform or non-uniform thickness of a coating material
on a homogeneous solid body (solid square block) from some known temperature
values elsewhere at the bottom side (bottom wall) of block. The known temperature
values at several positions along the bottom wall has been used in terms of some
statistical parameter (for example: mean, standard deviation, skew-ness and kurtosis)
in the developed methodology. The temperature profiles at the bottom wall for
different coating thickness have been numerically evaluated by the forward
calculation using commercial CFD software: Gambit and Fluent in a fully automated
way by an indigenously developed journal file. Statistical parameters of the
temperature profile have been extracted using MATLAB to build the training and test
data. Here, Feed-Forward neural network (NN) with back-Propagation algorithm is
38

used to predict the coating thickness. The NN has been modelled using C
programming Language in LINUX operating system. A Genetic Algorithm (GA) has
also been designed (using C programming language in Linux operating system) to
optimize the performance of the NN by optimizing the values of NN parameters.
By seeing the temperature profile at bottom wall for the different coating
thickness, it can be concluded that the temperature profile at bottom wall is very much
effected by thickness of the coating. The thickness of coating can be predicted by
knowing the statistical parameters for temperature profile at the bottom wall.
The NN parameters have been optimized simultaneously using a local
optimizer namely BP-algorithm and a global optimizer like GA in the developed
methodology. That menas, local search power of BP-algorithm has been coupled with
the global search capability of GA and a hybrid optimization scheme has been
implemented and consequently, it has performed very well. The developed scheme is
seen to be an practically effective tool for prediction in heat transfer problems.
Morever, it can be said that one may apply the developed methodology for
other kinds of prediction problems in heat conduction or heat convection.
5.2 Scope for Future Work
The boundary conditions used to develop the methodlogy can change to get a more
robust methodology. For example one may use the constant heat flux boundary at two
different wall (where as, in the present study, constant heat flux boundary is used only
at a single wall) and constant temperature boundary conditions at the remaining two
wall (where as, in the present study, constant temperature boundary is used at three
sides) or so on. The boundary conditions may be changed in other ways also. Attemt
may be made in future to develop the methodology using experimental data through
an experimental investigation. The similar concept (as used here) may be used to
39

predict inportant parameters related to other conduction or convection heat transfer
problem. The proposed GA-NN approach used here can be modified and extended to
solve other different kind of problems in heat conduction or other prediction problem
related to another mode of heat transfer.












40














41

References


Azenha, M., Faria, R., Figueras, H., 2011. Thermography as a technique for
monitoring early age temperatures of hardening concrete. Construction and
Building Materials 25, 42324240
Barletta, M., Gisario, A., 2006. An application of neural network solutions to laser
assisted paint stripping process of hybrid epoxy-polyester coatings on
aluminum substrates. Surface & Coatings Technology 200, 66786689
Far, M.R., Absi, J., Mariaux, G., Smith, D.S., 2011. Crack propagation modeling on
the interfaces of thermal barrier coating system with different thickness of the
oxide layer and different interface morphologies. Materials and Design 32,
49614969
Karayel, D., 2008.Prediction and control of surface roughness in CNC lathe using
artificial neural network. Journal of material processing technology 209,
31253137
Korukcu, M.O., Kilic, M., 2009. The usage of IR thermography for the temperature
measurements inside an automobile cabin. International Communications in
Heat and Mass Transfer 36, 872877
Krishnapillai, M., Jones, R., Marshall, I.H., Bannister, M., Rajic, N., 2006. NDTE
using pulse thermography: Numerical modeling of composite subsurface
defects. Composite Structures 75, 241249
42

Lasheras, F.S., Vilan, J.A., Neito, P.J.G., Diaz, J.J.D.C., 2010. The use of design of
experiment to improve a neural network model in order to predict the
thickness of chromium layer in a hard chromium plating process.
Mathematical and Computer Modelling 52, 1169-1176
Levy, A., Dayan, A., David, M.B., Gannot, I., 2010. A new thermography-based
approach to early detection of cancer utilizing magnetic nanoparticles theory
simulation and in vitro validation. Nanomedicine: Nanotechnology, Biology,
and Medicine 6, 786796
Nesbitt, J.A.,2000. Thermal modeling of various thermal barrier coatings in a high
heat flux rocket engine. Surface and Coatings Technology 130, 141-151
Ng, E.Y.K., 2009. A review of thermography as promising non-invasive detection
modality for breast tumor. International Journal of Thermal Sciences 48, 849
859
Pandey, S., Hindoliya, D.A., Mod, R., 2012. Artificial neural networks for predicting
indoor temperature using roof passive cooling techniques in buildings in
different climatic conditions. Applied Soft Computing 12, 12141226
Polini, R., Mantini, F.P., Braic, M., Amar, M., Ahmed, W., Taylor, H., 2005. Effect of
Ti- and Zr-based interlayer coatings on the hot filament chemical vapour
deposition of diamond on high speed steel. Thin solid films 494, 116-122
Rashidi, A.M., Hayati, M., Rezaei, A., 2012 Application of artificial neural network
for prediction of the oxidation behavior of aluminized nano-crystalline nickel.
Materials and Design 42, 308316
43

Shie, J.R., Yang, Y.K., 2008. Optimizations of a photoresist coating process for
photolithography in wafer manufacture via a radial basis neural network: A
case study. Microelectronic Engineering 85, 16641670
Simon, L., Fernandes, M., 2004. Neural network-based prediction and optimization of
estradiol release from ethylenevinyl acetate membranes. Computers and
Chemical Engineering 28, 24072419
Tan, T.Z., Quek, C., Ng, G.S., Ng, E.Y.K., 2007. A novel cognitive interpretation of
breast cancer thermography with complementary learning fuzzy neural
memory structure. Expert Systems with Applications 33, 652666
Woo, H.G., Cho, H.S., 1997.Estimation of hardened layer dimensions in laser surface
hardening process with variation of coating thickness. Surface coating and
technology 102, 205-217
Xing, H., Zhao, H., Zhang, Y., 2012. Thermal modeling and analysis of a radiometer
for thermal control coatings in situ measurement. Acta Astronautica71, 99
108
Yang, Y.S., Huang, W., 2012. A grey-fuzzy Taguchi approach for optimizing multi-
objective properties of zirconium-containing diamond-like carbon coatings.
Expert Systems with Applications 39, 743750
Yazdi, M.R.S., Khorasani, A.M., Faraji, M., 2011.Optimization of coating variable
for hardness of industrial tools by using artificial neural networks. Expert
system with applicatoin 38, 1211612127
44

Zain, A.M., Haron, H., Qasem, S.N., Sharif, S., 2011. Regression and ANN models
for estimating minimum value of machining performance. Applied
mathematical modelling 36, 1477-1492
Zhang, Y.Y., Deng, H.X., Shi, H.J., Yu, H.C., Zhong, B., 2012. Failure characteristics
and life prediction for thermally cycled thermal barrier coatings. Surface &
Coatings Technology 206, 29772985