Académique Documents
Professionnel Documents
Culture Documents
Ph.D. THESIS
(Disertan prce)
Study Branch
DESIGN AND PROCESS ENGINEERING
Topic:
2006
Acknowledgments
All who know Miroslav Raudensk acknowledge his unique contribution. The topics which are
presented in this work could have hardly been conceived without his initial insight. I have been
fortunate to have the continuing encouragement of Jaroslav Horsk, Keith A. Woodbury, and Miloslav
Druckmller. I would also like to acknowledge the contribution of colleagues Petr Kotrbek,
Jonathan Woolley, and Antonn Dobk. Special thanks are due to my wife, my son, and my mother for
their great support.
ABSTRACT
In this work, complex inverse heat conduction problems are studied. Thermal numerical
models for design and control purposes in metallurgical industry require a precise description
of heat transfer phenomena. Comprehensive quantitative information on the heat transfer
phenomena is not available for cooling of hot moving surfaces. In this work, attention is
focused on the search for boundary conditions describing the heat transfer in engineering
applications of spray cooling of metal surfaces and cooling of molds during casting. First,
a finite difference method is extended to cover phase changes and temperature dependent
material properties. Behavior of the iteration process during numerical solution of heat
transfer partial differential equations is investigated and a stopping criterion based on the
geometric convergence of the computed results is proposed. Methods for discretization error
estimation are presented and tested for the simplification of the object, its boundary
conditions, and the time domain when the numerical approach is applied. These methods are
then used for discretization optimization. For solution of inverse tasks, such as determination
of boundary conditions, temperature dependent thermal material properties, and model
calibration, combined optimization methods or methods of artificial intelligence are used to
achieve a higher accuracy of results and to speed up the computational process. Finally,
a method for correction of measured pressure distribution of a spraying water nozzle is
presented to allow comparison of cooling intensity and real pressure distribution.
ii
Contents
CONTENTS
1 INTRODUCTION........................................................................................................1
2 PROBLEM FORMULATION AND GOALS.........................................................................3
3 PRESENT STATE OF KNOWLEDGE OVERVIEW............................................................6
3.1 Fundamentals of Heat Transfer Phenomena.....................................................................6
3.1.1 Heat Transfer............................................................................................................6
3.1.2 Boundary and Initial Conditions...............................................................................7
3.2 Analytical Solution of Heat Equation...............................................................................8
3.2.1 The Semi-Infinite Solid............................................................................................9
Constant boundary conditions......................................................................................................9
iii
Contents
3.5.3 B-Splines................................................................................................................52
4 PROBLEM SOLUTION..............................................................................................54
4.1 Numerical Heat Conduction Derivation of Discretization Equations.........................55
4.1.1 Cartesian Coordinates One-dimensional Heat Conduction.................................56
Temperature dependent heat conductivity..................................................................................57
Temperature dependent mass density and specific heat ............................................................58
Source- term linearization...........................................................................................................61
Set of equations for all internal grid points in the general form................................................61
iv
Contents
Contents
11 APPENDIX B CONVECTION..................................................................................V
12 APPENDIX C RADIATION...................................................................................VI
13 APPENDIX D SOLUTION OF LINEAR ALGEBRAIC EQUATIONS................................VIII
13.1 Gauss-Jordan elimination..........................................................................................VIII
13.2 LU decomposition......................................................................................................VIII
13.3 Tridiagonal systems of equations.................................................................................IX
13.4 Band diagonal systems of equations..............................................................................X
13.5 Iterative improvement of a solution of linear equations...............................................XI
13.6 Conjugate gradient method..........................................................................................XI
Contents
17 SUMMARY...........................................................................................................A
18 RESMEE............................................................................................................B
19 RESUM..............................................................................................................C
vii
Nomenclature
NOMENCLATURE
A............area, m2
A............matrix of coefficients
a.............coefficients in matrix A
i, j..........unit vectors
Bi...........Biot number
J/kg
L............lower triangular matrix
network
C............capacity, F
m............measured data
N............shape function
J/kg.K
cvol..........volume-averaged apparent thermal
capacity, J/m3.K
volume, W
out.........rate of energy transfer out of control
volume, W
W/m3, W/m3.K
q.............heat flux, W/m2
R............resistance,
r, , z.....cylindrical coordinates, m, rad
f.............N-dimensional function
RMS.......root-mean-square error
G............irradiation, W/m
T............temperature, K
T *..........measured temperature, K
T............average temperature, K
viii
Nomenclature
temperature, K
T............filtered
d.............solid phase
t.............time, s
i.............general index
u, v.........local co-ordinates
V............volume, m3
W...........weight matrixes of the feed-forward
neural network
w............weighting function
l.............liquid phase
x, y, z......rectangular coordinates, m
R............right side
x.............vector of unknowns
r.............row index
Z............impedance,
rad.........radiation
s.............surface condition
, , .....substitution parameters
sur.........surroundings
.............emissivity, 0..1
............safety coefficient
............position, m
Superscripts
m............time index
............unknown variable
i.............iteration index
Introduction
1 INTRODUCTION
Efficient, accurate and stable numerical methods for solving fluid flow problems, heat
and mass transfer processes, chemical reactions, and turbulent phenomena are of great
importance in many industrial applications. It is nowadays generally recognized that computer
analysis of complex problems may provide a cost-effective, quick and sufficiently reliable
method in many cases. Sometimes, the computational methods may also be an alternative or
a complement to experimental investigations. Despite the tremendous developments and
achievements in methodologies, computer capacity and range of applications during last few
decades, still research on many topics is needed.
Many industries and companies worldwide use commercially available so-called
CFD-codes (CFD = Computational Fluid Dynamics) for simulation of flow and heat transfer
topics in heat exchangers, for enhancement of heat transfer, turbulent combustion in gas
turbines and combustion engines, electronics cooling and gas turbine blade, etc. Among these
codes are: ANSYS, FLUENT, CFX, STAR-CD, PHONECIS, CFD2000, FIDAP, ADINA and
others. However, to successfully apply such codes and to interpret the computed results, it is
necessary to understand the fundamental methods of such computations.
Although computation of heat transfer has reached a certain level and it can be
significantly helpful in many engineering and industrial applications, still comprehensive
research is needed in, e.g., heat transfer, handling of complex geometries, modeling of
two-phase convective flow and turbulence modeling. In this work, attention is focused on the
search for boundary conditions describing the heat transfer in engineering applications of
spray cooling of metal surfaces (up to 1200 oC). The cooling in casting, descaling, cooling of
products in hot rolling, and cooling of rolls in rolling technology are the typical industrial
applications. Thermal numerical models for design and control purposes in metallurgical
industry require a precise description of heat transfer phenomena. Comprehensive quantitative
information regarding the heat transfer phenomena is not available for quenching of hot
moving surfaces.
An almost all-engineering analysis is based on a model that has nice smooth
mathematical properties. On the other hand, any measured data has some noise. Combining
measurement with an inverse analysis often results in an ill-conditioned problem.
Introduction
Mathematically, such problems are ill-posed and have to be overcome by developing new
numerical methods, introduction of new objective functionals into the optimal control
algorithm, new regularization techniques, new experimental procedures, etc.
The inverse problem is one type of an ill-conditioned or ill-posed problem. In a typical
direct problem, one is given a model, the initial conditions of the state variables, and the
forcing terms, and is asked to produce a solution. Solving the direct problem, an error in input
data is usually suppressed. In the general inverse problem, one is given a model and
measurements of some state variables, and is asked to estimate the initial conditions, the
forcing terms, and the rest of state variables. Unlike the direct problem, an error in input data
is extremely amplified in the inverse problem. The given results are often unusable, if any
results are obtained. The noise significantly decreases the stability of the computational
algorithm.
An improvement of computational algorithms and new designs of experimental
measurements are needed. The major attention is focused on solving inverse heat conduction
problems (IHCP), such as determination of the fast changing boundary conditions,
suppression of influence of an error in measured data, and calibration of computational
models. Along with these problems, an improvement in the self-adaptive design of
computational models and knowledge of accuracy of computed results using a numerical
method are necessary.
New approaches, such as usage of various optimization methods and data filtration
could be useful to overcome the above specified problems. Nowadays, an artificial
intelligence could play an important role. Approaches like feed-forward neural network or
genetic algorithm may be considered.
The contemporary revolution in computer technology is producing larger, faster, and
inexpensive computers with almost no limits in sight. This technology has made it possible
for engineers and scientists to construct more realistic mathematical models of physical
processes. An important area of engineering that will become even more significant in the
future is the combining of measurements with engineering models.
The main goal of this work is to improve the inverse heat conduction method for
applications in casting, descaling, cooling in hot rolling, etc. This includes stabilization of
computational methods and increase in precision of the computed results. Attention will also
be focused on error estimation of the computed results.
Along with the main goals, minor problems will have to be solved. For complex
geometries, there are no exact analytical solutions and thus numerical methods must be used.
The numerical methods simplify the reality e.g. the temperature profile is piecewise
linearized. This simplification leads to a computational error which increases for highly
transient problems. A method for an adaptively generated mesh of the computational model on
the basis of the estimated error will have to be developed to describe the real problem as
accurately as needed. During computation, time dependent boundary conditions are also
simplified. An automatic time-step refinement for unsteady problems will be needed to
achieve the desired accuracy. The material properties change with temperature. Typically, the
temperature increases with increasing thermal energy of the object and most of the
computational methods are based on this phenomenon. In some cases, there is a change in the
internal structure of the material called the phase change. During this change, the temperature
may remain constant with increasing thermal energy. A modification of the computational
method will have to be made to include phase changes.
To determine boundary conditions, an inverse task is used that requires some measured
data. The data are measured using a sensor but none of the sensors are exactly the same. The
position of the thermocouple may differ as well as surface contact conductance. To calibrate
the computational model, an experiment will have to be designed. To achieve the desired
accuracy of the computational model, an adaptively generated model and calibration of the
computational model using experimental data will also be needed. To compute boundary
conditions from the measured temperature, an inverse heat conduction problem has to be
solved. Some methods have already been developed but none of them is universal. They have
strict limitations of usage and thus they can be used only for a certain sort of problems. It will
be necessary to design a proper inverse heat conduction computational method to solve IHCP.
The input information for the IHCP is usually blurred along the time axis. To use an adequate
time interval, the solution method should be able to determine adequate time interval for the
inverse method.
Some applications also require description of other than thermal boundary conditions.
Knowledge of water spray impact is often used for comparison of thermal and non-thermal
boundary conditions. Usually, a water pressure impact field is measured. For such a kind of
measurement, a real finite-size pressure sensor is used. In some cases, the impact distribution
is very small or at least very narrow with respect to the dimensions of the measuring sensor.
The distribution is being averaged during the measurement. The shape of the obtained spray
distribution is larger and the measured maximum is lower than the real one. A method will be
needed to obtain a real distribution from the measured data.
The laboratory experiments in continuous casting, descaling, and cooling of rolls have
already started and experimental data, which can be used for testing the new methods, are
already available. However, an improvement in the design of experiments is also needed.
T1 > T2
T S T
T1
T2
Moving fluid T
q1
q
q
TS
q2
Surface, T1
Surface, T2
T
T
T
T
k
k
q=c
.
p
x
x
y y
z
z
t
(3.1.1)
1
T
1
T
T
T
kr
2
k
k
q=c
.
p
r r
r
z
t
r z
(3.1.2)
In heat transfer analysis, the ratio of the thermal conductivity to the heat capacity is
an important property termed the thermal diffusivity , which has units of m2/s:
=
k
.
cp
(3.1.3)
It measures the capability of a material to conduct thermal energy relatively to its capability to
store thermal energy. Materials with large will respond quickly to changes in their thermal
environment, while materials with small will respond more slowly, taking longer to reach
a new equilibrium condition.
3.1.2 Boundary and Initial Conditions
The solution of the heat equation depends on the physical conditions existing at the
boundaries of the medium and, if the situation is time dependent (unsteady), on conditions
existing in the medium at some initial time. The three kinds of boundary conditions, which
commonly occur in heat transfer, are summarized in Table 3.1. They are specified at the
surface x = 0 for a one-dimensional model.
The first condition corresponds to a situation for which the surface is maintained at
a known temperature (e.g. when the surface is in contact with boiling liquid or the surface
temperature is measured). It is commonly termed a Dirichlet condition, or a boundary of the
first kind. The second condition corresponds to the existence of a known heat flux at the
surface. This heat flux is related to the temperature gradient at the surface according to
Fouriers law. It is termed a Neumann condition, or a boundary condition of the second kind.
7
This condition can be measured using a special sensor usually equipped with two
thermocouples. A special case of this condition corresponds to the perfectly insulated, or
adiabatic, surface for which T / xx=0 =0. The boundary condition of the third kind
corresponds to the existence of convection cooling (or heating) at the surface and is obtained
from the surface energy balance. In addition, boundary condition of the third kind may be used
for radiation heat transfer, however, the h coefficient depends strongly on temperature. The
boundary condition of the third kind has a wide range of usage and is frequently used in our
analysis the water cooling intensity distribution described by h( T ).
Table 3.1 Boundary conditions for the heat diffusion equation at the surface (x=0)
1. Surface temperature
T 0, t =T S
TS
(3.1.4)
T(x, t)
x
2. Surface heat flux
a) Finite heat flux
T
k
=q
x x=0 s
qs
T(x, t)
(3.1.5)
x
b) Adiabatic or insulated surface
T
=0
x x=0
(3.1.6)
q s =0
T(x, t)
x
T(0, t)
k
T
=h [ T T 0, t ]
x x=0
(3.1.7)
T, h
T(x, t)
x
alternative is the one using a numerical technique. However, the analytical solution is very
useful for testing numerical methods and their accuracy. Hence, some of the most important
cases to be used for testing accuracy of the numerical models are reviewed here.
3.2.1 The Semi-Infinite Solid
A simple geometry for which analytical solution may be obtained is the semi-infinite
solid. It may also be used to approximate the transient response of a finite solid such a thick
slab. For this situation, it would be reasonable to use the approximation only for a short time
period, during which temperatures in the slab interior are not influenced by the change in
surface conditions. A constant starting temperature T 0 is assumed.
Constant boundary conditions
A sudden change in boundary conditions is assumed. Analytical solutions are
summarized as follows [1].
Constant surface temperature
T x , t T s
x
=erf
T 0 T s
2 t
q s t =
(3.2.1)
k T s T 0
(3.2.2)
T x , t =
2 qs
q x
x 2
x
exp
s erfc
T 0
4 t
k
2 t
(3.2.3)
or in the form of dimensionless temperature solution for a unit step increase in surface heat
flux, but valid only for t0.1, the solution is [4]
1
x2 2
T x , t =t x 2
3
2
k12 e k t cosk x
2
(3.2.4)
k =1
T x , t T 0
x
hx h 2 t
x
h t
=erfc
exp
erfc
2
T T 0
k
k
2 t
2 t
k
(3.2.5)
4 sin n
n t
T x , t T
=
cos nx
exp
T 0 T
L2
n=1 2 n sin 2 n
(3.2.6)
where the discrete values of n are positive roots of the transcendental equation
n tan n =
hL
.
k
(3.2.7)
Since the convection conditions are the same on both surfaces, the temperature distribution is
symmetrical about the mid-plane where x = 0.
10
T
=
t x
x
(3.3.1)
where the is the diffusion coefficient (thermal diffusivity, Eq. (3.1.3)). The equation
describes how T (x, t) propagates itself forward in time (see Figure 3.2). In other words,
Eq. (3.3.1) describes time evolution.
boundary
conditions
initial values
Figure 3.2 An initial value problem - initial values are given on
one time slice, and it is desired to advance the solution in time.
The goal of a numerical code should be to track time evolution with some desired accuracy. In
general, it is not possible to integrate forward in time stably using an initial value. The goal
of a numerical code is to converge to the correct solution. An important point is the stability of
the algorithm. Many reasonable-looking algorithms for initial value problems just do not
work they are numerically unstable. For example, the simplest problem on a 100 x 100
spatial grid would involve 10 000 unknowns. This would result in a 10 000 x 10 000 matrix
containing 108 elements.
3.3.1 Diffusive Initial Value Problems
The diffusion equation in one space dimension is Eq. (3.3.1). Let us assume 1,
otherwise Eq. (3.3.1) has a physically unstable solution [5]. A small disturbance leads to
increasing concentration of energy instead of to its dispersion. Consider first the case when
is constant. Then Eq. (3.3.1) can be differenced in the obvious way
T m1
T mj
T m 2 T mj T mj1
j
= j1
.
t
x2
(3.3.2)
11
This is the Forward Time Centered Space (FTCS) or explicit differencing scheme [5]. The
stability criterion is
2t
1 .
2
x
(3.3.3)
The physical interpretation of the restriction is that the maximum allowed time-step is, up to
a numerical factor, the diffusion time across a cell of width x .
The fully implicit differencing scheme is only first-order accurate in time for the scales
that we are interested in:
m1
T m1
T mj
T m1
T m1
j
j1 2 T j
j1
=
.
2
t
x
(3.3.4)
Schemes with this character are called fully implicit or backward time, by contrast with FTCS
(which is called explicit). To solve Eq. (3.3.4) one has to solve a set of simultaneous linear
equations at each time-step. Fortunately, this is a simple problem for one-dimension because
the system is tridiagonal. The scheme is unconditionally stable. The details of the small-scale
evolution from the initial conditions are obviously inaccurate for large t . But the correct
equilibrium solution is obtained. This is the characteristic feature of implicit methods.
Combining the stability of an implicit method with the accuracy of a method that is
second-order in both space and time yields to the average of the explicit and implicit FTCS
schemes which is the Crank-Nicholson differencing scheme
m1
m
m
m
T m1
T mj T m1
T m1
j
j1 2 T j
j1 T j1 2 T j T j1
=
t
2
x2
(3.3.5)
which is second-order accurate in time. Both the left and right-hand sides are centered at
time-step n+1/2. The Crank-Nicholson scheme should be stable for any t , but unfortunately
not in case of large t and temperature dependent thermal diffusivity .
t or m
x or j
(c) Crank-Nicholson
12
T
2 T 2 T
=
.
t
x2 y2
(3.3.6)
An explicit method (FTCS) can be generalized from the one-dimensional case. We implement
the Crank-Nicholson scheme in two dimensions
1 t 2 m1 2 m
m1
m
2
m1
2
m
T j , k =T j , k 2 x T j , k x T j , k y T j , k y T j , k
2
(3.3.7)
x= y ,
(3.3.8)
where
m1
m1
m1
2x T m1
j , k T j1, k 2 T j , k T j1, k
2
m
m
m
m
x T j , k T j1, k 2 T j , k T j1, k
m1
m1
m1 .
2y T m1
j , k T j , k 1 2 T j , k T j , k 1
2y T mj , k T mj , k 12 T mj , k T mj , k 1
(3.3.9)
A problem arises when solving the coupled linear equations. In one space dimension the
system was tridiagonal, which is no longer true, though the matrix is still very sparse. One
possibility is a slightly different way of generalizing the Crank-Nicholson algorithm. It is still
second-order accurate in time and space. It uses powerful concept of operator splitting or time
splitting and is called the alternating-direction implicit method (ADI). The idea is to divide
each time-step into two steps of size t / 2 . In each sub-step, a different dimension is treated
implicitly
1 t
2
2
T m1/
=T mj , k 2 2x T m1/
2y T mj , k ,
j ,k
j,k
2
(3.3.10)
13
1 t
m1/ 2
2
T m1
2 2x T m1/
2y T m1
j , k =T j , k
j ,k
j ,k .
2
(3.3.11)
The advantage of this method is that each sub-step requires only the solution of a simple
tridiagonal system. This method is usually stable but not always. It can exhibit unrealistic
behavior when time-step t exceeds a certain value. Hence, the diffusive problems are
usually best treated implicitly as will be described later. The implicit method is the only one
that allows us to use any value of time-step t without producing physically unrealistic
results [6]. It is true that, for smaller t , the Crank-Nicholson method can be more accurate
than fully implicit method. However, the guarantee of producing realistic results for any value
of time-step t is considered to be often more important.
3.3.3 Finite Difference Method (FDM)
The finite difference method is the oldest method and also the easiest one to apply to
problems with simple geometries.
The computational domain is covered by a grid. Taylor series expansion or polynomial
fitting is used to approximate the derivatives of the variables with respect to coordinates at
each grid point exact differential equation of the heat equation is reduced to an approximate
algebraic equation. Algebraic equations are achieved at each grid point and the resulting set of
equations may then be solved simultaneously for the temperature at each node. This method is
also called lumped capacitance because all of the control volume thermal capacitance is
lumped at the center node point for all interior control volumes. Further information is
available in [7].
For the numerical solution of an unsteady situation, time is discretized as well into the
time-steps t . There is a number of ways of deriving a discretization equation. Three already
mentioned methods are explicit, Crank-Nicholson, and fully implicit.
3.3.4 Finite Volume Method (FVM)
In the finite volume method [6, 8], the domain is subdivided into a number of so-called
control volumes of arbitrary and finite size as in Figure 3.4. The integral form of the
conversion equations is applied to each control volume. Conservation of energy for a solid
requires that the sum of the rate of heat flow across the control volume boundary and of the
time rate of change in energy within the control volume be equal to the rate at which energy is
14
produced within the control volume. In equation form, the conservation of energy can be
written as
T
= c
dV
qdA qdV
t
A
(3.3.12)
where A and V are the closed surface area and volume, respectively. The q parameter is the
heat flux vector leaving the control volume surface, q is the energy producing rate per unit
volume, and c mass density and specific heat.
Region of
interest
Control volume
boundary
dA=ndA
w(x)
1.0
x j1
xj
x j1
T
2 T
w x t x 2 dx=0
0
(3.3.13)
where L is the thickness of the slab and w(x) is a weighting function to be specified. There is
an infinite number of possible weighting functions. A weighting function that gives good
results should be chosen. A popular weighting function is the so-called tent function which is
shown in Figure 3.5:
x x j1
, x x x j
x j x j1 j1
x j1 x
w j x=
, x x x j1
x j1 x j j
w j x=0, x x j1 x j1 x .
w x=w j x=
(3.3.14)
Since the weighting function wj(x) is zero over a large portion of the domain 0 x L ,
Eq. (3.3.14) can be written as
x j1
w j x
x j1
T
2 T
dx=0 .
t
x2
(3.3.15)
Using the linear element temperature profile assumption and Eq. (3.3.14) for wj(x), the finite
element equation for an arbitrary interior node j can be written as
T j T j1
T T j1
x d 1
2
k j
c
T j1 T j
x
x
2 dt 3
3
x d 2
1
c
T j T j1 =0 ; j=2, 3 , ... , N 1.
2 dt 3
3
(3.3.16)
16
The only difference between FVM and FEM is the different weighting coefficients on the
thermal capacitance terms. The FVM gives greater weighting to the center temperature Tj. The
FEM equations for the surface nodes are
k
T 1T 2
x d 2
1
q t c
T 1 T 2 =0 ,
x
2 dt 3
3
q N t k
(3.3.17)
T N 1T N
x d 1
2
q t c
T T =0 .
x
2 dt 3 N 1 3 N
(3.3.18)
It can also be demonstrated that if the weighting function w(x) is chosen to be the Dirac delta
function
w j x= x x j =
1, x= x j
0, all other x
(3.3.19)
then the FEM results for interior nodes are identical to the FDM.
3.3.6 FDM, FVM, and FEM Similarity
All of the FDM, FVM, and FEM equations can be put in a similar form:
q t
d
T 1 T 2 =
T 1T 2
2
dt
c x ,
x
(3.3.20)
d
T j12 T j T j1 =
T j T j1
T j1T j ,
2
2
dt
x
x
(3.3.21)
q t
d
T N 1T N =
T N 1T N N
2
dt
c x
x
(3.3.22)
where and have the values listed in Table 3.2. Equations (3.3.20-3.3.22) are restricted to
temperature-independent thermal properties but the concepts can be extended to T-variable
cases.
Table 3.2 Values of the and of Eq. (3.3.20-3.3.22)
FDM
1/2
1/2
FVM
3/8
1/8
1/2
FEM
2/6
1/6
1/2
17
qdV = c t
V
dV .
(3.3.23)
Using the mean value theorem for integrals on the right and the divergence theorem on the left
gives
qn dA=V T c T ddtT
(3.3.24)
where the bar indicates a suitable average value over the region.
j =3
noc
2
c
y
x
noa
j =2
a
Figure 3.6 An irregularly shaped
domain and its discretization into
triangular elements.
j =1
Figure 3.7 A typical triangular
element.
The domain is divided into triangular elements. A sample domain discretization is shown in
Figure 3.6. In this discretization scheme, curved boundaries are approximated by
18
piecewise-linear curves. Then the centroids of the elements are connected with the midpoints
of the corresponding sides (see Figure 3.7): this creates polygonal control volumes around
each node in the calculation domain. An integral formulation can be obtained by applying the
conservation principle for energy to a control volume V1aoc
o
dT
(3.3.25)
where n is a unit outward vector normal to the differential area element dA. A linear
temperature variation is assumed throughout the triangular element. In each element heat flux
q can be expressed in terms of its components in the x and y directions
q=q x iq y j= k T
] [
T
T
i k T
j
x
y
(3.3.26)
where i, j are unit vectors in the x and y directions, respectively. The normal vectors noa and noc
needed to evaluate the integrals in Eq (3.3.25) are given by
noa =cos 1 icos 2 j=
n oc =cos1 icos2 j=
ya
x y
2
a
yc
x y
2
c
2
a
i
2
c
xa
x y
2
a
x c
x y
2
c
j ,
(3.3.27)
j .
(3.3.28)
2
a
2
c
Using the local x, y co-ordinate system shown in Figure 3.7, the integrals in Eq. (3.3.25) that
represent the transport of energy by conduction can be approximated as
o
[ y a , x a ]
qn dA=[q x , q y ]
a
c
2
a
[ y c , x c ]
qn dA=[q x , q y ]
o
x y
2
a
x y
2
c
2
c
x y
2
a
2
a
(3.3.29)
x y
2
c
(3.3.30)
2
c
Writing the heat balance equations for all control volumes, the system of N non-linear
ordinary equations is obtained whose solutions are the temperature changes in the central
nodes of the control volumes. The number of equations is equal to the number of the control
volumes N. The initial condition gives the starting values of the nodal temperatures.
19
Nodes
Figure 3.8 Finite-Volume unstructed-mesh vertex based discretization.
SCV
IP1
SCV
SS2
SS2
SS1
20
11 x 11
21 x 21
31 x 31
41 x 41
5.18
2.63
2.40
2.30
4.34
2.49
2.48
2.70
4.59
2.80
2.52
2.60
5.18
2.63
2.40
2.30
3.43
2.26
2.30
2.15
27500
FVUMVT
25000
22500
FVUMVQ
Time [s]
20000
17500
15000
12500
10000
FVUMCT
FVUMCQ
7500
5000
FVSM
2500
0
0
250
500
750
1000
1250
1500
1750
2000
Number of nodes
21
The material surface experiences a very high temperature gradient from the pulse. The
duration of the pulse is less than few picoseconds. Under these conditions, the finite speed of
heat propagation becomes important and a wave type propagation of heat, instead of diffusion,
has been proposed. The TLM method was first developed by Johns and Beurle [19] to treat
electromagnetic propagation. Transmission line matrix is a discrete numerical time-domain
modeling technique that can frequently be reduced to a solution of the diffusion equation.
The space is discretized into a network of finite cells with node points at the center of
each cell (see Figure 3.11). The following analogy of the electromagnetic propagation with
heat conduction is used [20]: in the equivalent TLM network, the resistors remain clustered
around the nodes and represent the thermal resistance of the material. The capacitor / inductor
combinations are replaced by impedance Z, which connect each node to its neighbors and
carry voltage pulses between the nodes in a finite time t. According to the fundamental
transmission line theory, the impedance is
Z=
L a t
L
.
=
=
C
C
a t
(3.3.31)
Sinking Boundary
Zb=0 Rf=-1
Absorbing
Boundary
Zb=Z Rf=0
R
Z
R
Z
R
Z
Zb
Rb
Radiative
Convective
Boundary
Insulating Boundary
Zb= Rf=1
Figure 3.11 Different TLM boundary conditions around one cell.
Since the capacitor models the heat capacity of the material, the TLM parameters R and Z can
be evaluated from
2 R=
1
,
k x
(3.3.32)
23
Z=
a t
3
c x
(3.3.33)
where the parameter a represents the number of link lines in each elemental volume.
Input
Time index
k=0
R
RZ
k=1
k=2
k=3
Spatial position
Figure 3.12 Space-time history during three iterations of one-dimensional
scatter of a single input.
A TLM solution is obtained by repeatedly considering delta voltage (temperature) pulses
to be applied simultaneously on all parts of all nodes. These pulses are scattered
instantaneously into pulses which, during the time-step t, travel along link transmission lines
to appear at neighboring nodes (see Figure 3.12). The TLM routine operates on the traveling,
scattering, and connecting of these pulses in the network. The transmission lines in the model
act as delay lines.
Transmission lines introduce a time delay, so that TLM-based models are explicitly
discretized in both space and time. Although this method is best known in electromagnetic
applications, the technique is also suitable for the numerical simulation of diffusion-based
phenomena such as heat transfer. However, this method is reasonable for problems where
high-power and short-pulse are used, such as pulsed lasers. The duration of the pulse is
usually less than few picoseconds.
3.3.10 Phase Change Implementation
Physical processes, such as solid/liquid and solid state transformations, involve phase
changes. The numerical treatment of this non-linear phenomenon involves many problems.
24
Methods for solving the phase change usually use a total enthalpy H, an apparent specific heat
coefficient cA, or a heat source q.
The nature of a solidification phase change can take many forms. The classification is
based on the matter in the phase change region. The most common cases follow:
a) Distinct: The phase change region consists of solid and liquid phases separated by a
smooth continuous front freezing of water or rapid solidification of pure metal.
b) Alloy: The phase change region has a crystalline structure consisting of grains and
solid/liquid interface has a complex shape most metal alloys.
c) Continuous: The liquid and solid phases are fully dispersed throughout the phase
change region and there is no distinct interface between the solid and liquid phase
polymers or glasses.
In a distinct phase change, the state is characterized by the position of the interface. In such
cases the class of the so called front tracking methods are usually used. However, in cases b)
and c) the models uses the phase fraction.
The phase change process can be described by a single enthalpy equation
H
g d H d s d g l H l s l =k T
t
(3.3.34)
where g is the phase volume fraction, s is the phase velocity, and subscript d and l refer to
solid and liquid phases, respectively [21]. The k is (in this case) a mixture conductivity defined
as
k =g d k d g l k l
(3.3.35)
H =g s d c d dT g l l c l dT l g l L
T ref
(3.3.36)
T ref
where Tref is an arbitrary reference temperature. To overcome the problem of the non-linear
(discontinuity) coefficient of a specific heat a non-linear source term is used. The term
H / t can be expanded as
gl
H
T
.
=c vol
H
t
t
t
(3.3.37)
Neglecting convection effects in Eq. (3.3.34) and substituting Eq. (3.3.37) results in
25
c vol
T
=k T q
t
(3.3.38)
where
q=
H
gl
.
t
(3.3.39)
Eq. (3.3.34) is non-linear and it contains two related but unknown variables H and T. It
is convenient to reformulate this equation in terms of a single unknown variable with
non-linear latent heat.
Apparent heat capacity
The apparent specific heat [22, 23] can be defined as
c A=
dg
dH
=c vol H l
dT
dT
(3.3.40)
where
c vol =g d d c d g l l c l
T
H = l c l d c d dT l L .
T ref
(3.3.41)
Neglecting convection effects and substituting into Eq. (3.3.34) yields apparent heat capacity
equation
cA
T
=k T .
t
(3.3.42)
Total enthalpy
From Eq. (3.3.36) it can be written
T = H /c vol H g l /c vol .
(3.3.43)
H
k
k
=
H
H gl .
t
c vol
c vol
(3.3.44)
26
Q
=
t
(3.3.45)
is
is the unknown variable, c is a specific heat,
where
is a diffusion coefficient and Q
a heat source [21]. The appropriate values are provided in the Table 3.4.
Table 3.4 - Possible values for the general form of the phase change equation
1. Basic equation
; c =1 ;
T
=H
=0 ; Q=k
2. Apparent heat capacity
=T
; c =c A ;
=k ; Q=0
3. Source
q=
=T
; c =c vol ;
H g l / t
=k ; Q=
4. Total enthalpy
=H
; c =1 ;
/c vol H g l
=k / c vol ; Q=k
Note:
H =g d H d g l H l
Mixture conductivity k =g d k d g l k l
Mixture volumetric specific heat c vol =g d d c d g l l c l
Mixture volumetric enthalpy
a M1 x 1a M2 x 2a MN x N =b M
(3.3.46)
where the N unknowns xc are related by M equations. The coefficients ar,c are known numbers
as well as the quantities br. The set of equations can be written in a matrix form as
Ax=b
(3.3.47)
27
where
A=
a 11
a 21
a 12
a 22
a 13
a 23
a M1 a M2 a M3
a1 N
a2 N
a MN
[] []
x1
x2
x= x 3
xN
b1
b= b 2 .
bM
(3.3.48)
If N = M then there are as many equations as unknowns, and there is a good chance of
obtaining a unique solution of xc set. Analytically, it can fail if one or more of the M equations
is a linear combination of the others. This kind of equations is called a singular one.
Numerically, at least two further things can go wrong:
Some of the equations may be so close to become linearly dependent. The roundoff
errors may yield them linearly dependent during the solution process. In this case,
the numerical procedure will fail.
Accumulated roundoff errors during the solution process can produce a false
solution. This problem arises when N is too large. The numerical procedure does
not fail algorithmically. However, it returns a set of xc that are wrong, as can be
discovered by direct substitution back into the original equations. It is difficult to
give any guidelines but there is a rough idea: Linear sets with N =2050 can be
solved in single precision (32 bit floating representations) without resorting to
sophisticated methods, if the equations are not close to singular. With double
precision (64 bits), this number can be extended to several hundreds.
(3.3.49)
but for huge A matrixes it might be an unstable and time-consuming task to compute
an inverse matrix. Thus sophisticated methods have been published to compute the unknown
vector x. The most often used methods are summarized in Appendix D.
3.3.12 Non-linear Set of Equations
A non-linear set of equations may have no solutions at all or it may have more than one
solution. In vector notation, we want to find one or more N-dimensional solution vectors x
28
f x =0
(3.3.50)
where f is the N-dimensional function. Its components are the equations to be solved
simultaneously. For smoothly varying functions and appropriate algorithm, it will converge
when the initial guess is good. Generally, the problem of finding solution vectors x is very
complicated.
In the heat transfer problems the equations are usually only slightly non-linear. The set
of non-linear equations can be usually solved by iterating the method for solving the linear set
of equations. We start by guessing or estimating the temperature at all grid points. Treating
this as the known temperature field, we calculate tentative values of the coefficients in the
discretization equations. Then we solve the equations to obtain a new temperature field.
Taking this field as a better estimate for temperature, we recalculate the coefficients and solve
the equations again. This process is repeated until the solutions are identical for all subsequent
iterations. This unchanging solution is called the converged solution. In practice, the iteration
process is terminated when the change in the solution is smaller than a prescribed small
number called convergence criterion. Although we proceed through a series of linear
equations, the final converged solution is the correct solution of the non-linear problem.
Under-relaxation
If values of coefficients change very rapidly from iteration to iteration, the iterating
method can diverge. In a highly non-linear problem, it is convenient to slow down the changes
in temperature from iteration to iteration. This process is called under-relaxation [24].
To slow down the changes in the of computed unknown variables x from iteration to
iteration, the value of xi+1 for the next iteration is a combination of values from this iteration xi
and from the previous iteration xi-1
x i1= x i 1 x i1
(3.3.51)
where i superscript is an index of iteration and the is the under-relaxation factor, which is
between 0 and 1. An optimum of the depends on the nature of the non-linearity, number of
unknown variables, and other factors. To find a suitable value, exploratory computations have
to be made. Although this is not very convenient, it is good to know, that the divergence can
be suppressed by using appropriate under-relaxation.
29
x1
L
Temperature sensor 1
convolution integral) which is restricted to linear problems [27]. Numerical procedures such
as finite differences [7, 28, 29] and finite elements [30] were also employed, due to their
inherent ability to treat non-linear problems. Exact solution techniques were proposed by
Burggaf [31], Imber and Khan [32], Langford [33], and others. Some techniques used Laplace
transforms but these are limited to linear cases [34]. The improvement in artificial intelligence
has brought new approaches, such as genetic algorithm [35] and neural networks [36, 37].
3.4.1 IHCP by Beck
One of the first approaches for solving IHCP using numerical methods was proposed by
Beck [11]. The main feature of Becks approach is sequential estimation of the time varying
boundary conditions. He demonstrated that function specification and regularization methods
could be implemented in a sequential manner and that they gave in some cases nearly the
same results as the whole domain estimation. Moreover the sequential approach is
computationally more efficient. Becks approach has been widely used to solve inverse heat
conduction problems to determine unknown boundary or material property information.
The method uses sequential estimation of the time varying boundary conditions and uses
future time steps data to stabilize the ill-posed problem. The heat transfer coefficient is found
after determining the heat flux at the surface using Eq (11.1). To determine the unknown
surface heat flux at the current time t m, the measured temperature responses T *i , m, are
m
compared with the computed T j from the forward solver (e.g. FDM, FVM, FEM, etc.), using
n f future times steps
nT
mn f
SSE=
T *i , f T fj 2 .
(3.4.1)
f =m1 ji ; i=1
Using the linear minimization theory, the value of the surface heat flux that minimizes
Eq. (3.4.1) is
mn f
nT
q m= f =m1
*, f
T i T j |q
ji ; i=1
mn f
nT
=0
(3.4.2)
f 2
i
f =m1 i=1
f
where T j |q =0 are the temperatures at the temperature sensor locations computed from the
m
forward solver using all the previously computed heat fluxes, but without the current one q m.
31
The i is the sensitivity of the ith temperature sensor at time t f to the heat flux pulse at time t m.
These sensitivity coefficients are mathematically the partial derivatives of the computed
temperature field to the heat flux pulse, but in this case they physically represent the rise in
temperature at the temperature sensor location for a unit heat flux at the surface. The
sensitivity coefficient of our interest is defined as
m
m
j
T j
q
(3.4.3)
Once the heat flux is found for the time t m, the corresponding surface temperature T 0
may be computed using the forward solver. When the surface heat flux q m and surface
temperature T 0m are known, the heat transfer coefficient is computed from
m
h =
q m
.
m
m
m1
T T 0 T 0 / 2
(3.4.4)
(3.4.5)
where
32
[
[
[] [ ] []
T m
T = T m1 , T i=
T m f 1
m
T1
q
i
m1
T 2 , q=
q
, q i=
q m f 1
T in
T*
] [
q1
i
q2 ,
q in
q*
1
1,1 i 1, n i
2
1
.
=
, i=
n , n i
n i
f f 1 1
q*
T * ,1
T*
q*
(3.4.6)
The sequential approach then temporary assumes that q is independent on time. Then using
Z =
I
11
where I =
1n
where n=n q
(3.4.7)
(3.4.8)
The matrix derivative of Eq. (3.4.8) with respect to q gives the estimated heat fluxes
m
m T
m 1
m T
*,m
m
q =[ Z Z ] Z T T |q=0 .
(3.4.9)
After it is obtained, m is increased by one and the procedure is repeated for the next time step.
3.4.2 Comparison of Inverse Heat Conduction Methods by Raynaud
Beck [11] has stabilized the IHCP using several future time temperatures with the
least-squares method while Hensel and Hills used smoothing functions on the input data. Beck
and Murio have combined this method with a regularization one [38]. Regularization method
[25] has been presented to reduce the sensitivity to the measurement error.
The quantitative comparison of methods is difficult. Raynaud [4] shows two test
problems and demonstrates how to compare the results given by different inverse algorithms.
The ideal inverse method would not be sensitive to the measurement error and thus the
minimum deterministic bias and minimum sensitivity to measurement errors are required. The
following four methods are compared by Raynaud: Method A DSouza [39]; Method B
Raynaud and Bransier [40]; Method C Weber [41]; Method D Beck [11]. They are
stabilized using different principles but all of them use future temperatures except Method A.
33
For the methods A, B, and C, the surface heat flux is calculated from the temperature
distribution at the surface node using the following energy balance
m
m1
m1
T T 2 x T 1 T 1
q = 1
x
2
2 t
m
(3.4.10)
n+6
n+5
Method B
n+4
n+3
t
n+2
Method C
n+1
n
Method A
n1
1
j1
j+1
x
known temperature
temperature to be calculated
Figure 3.14 Computation molecules.
DSouza Method A
This method uses the backward difference approximation for the time derivative and
central difference approximation for the space derivative
T mj12 T mj T mj1
x2
T mj T m1
j
=
.
t
(3.4.11)
The computation molecule is shown in Figure 3.14. It is evident that this method uses only
past and present temperatures to obtain the present temperature estimate at node j 1.After
calculating the temperature at node j 1, the temperature at node j 2 can be calculated in
an explicit way using Eq. (3.4.11). This continues to the surface temperature.
34
This method uses double approximation given by energy equation. The first energy
balance in the dimensionless form is
m
m1
m1
m1
T 1, j1 T 1, j T 1, j T 1, j1
T T 1, j
= x 1, j
x
x
t
(3.4.12)
(3.4.13)
The calculation in the time direction is performed for a given spatial node j before proceeding
to the next node. The computation molecule is shown in Figure 3.14. The estimation of the
surface temperature at time m involves the measured temperature at time m + j - 1 where j is
the node of the temperature sensor. At each step, T mj1 is approximated by the arithmetic
m
average of T 1, j1 and T 2, j1. Grid refining leads to an increasing number of future
temperatures and of steps required to reach the surface. More space steps lead to the more
crude approximation of the surface temperature.
Weber Method C
This method uses the hyperbolic form of the heat conduction equation
2
T T T
2
=
.
t x2
t
(3.4.14)
where is the nonnegative constant. Central differences in both time and space are used to
approximate Eq. (3.4.14)
T m1
2 T mj T m1
j
j
t2
T m1
T m1
T mj12 T mj T mj1
j
j
=
.
2 t
x2
(3.4.15)
The computation molecule is shown in Figure 3.14. The computation of the temperatures
proceeds toward the surface as for method B. The future temperature at time m + j - 1 is
involved in the estimation of the surface temperature at time m as in method B.
35
Beck Method D
This method is a time-marching technique. At each time step the surface flux is
computed using a least-squares minimization procedure for the simulated temperature
difference as described in the previous IHCP by Beck chapter.
Test cases
The one-dimensional linear heat conduction problem is chosen as a test case. The
temperature is known as a discrete function of time at position x under the surface. The
boundary at x = L is insulated. It is desired to predict the transient flux history at x = 0.
The purpose of the first test case is to estimate the deterministic bias of inverse methods.
A heat flux impulse is considered that is constant and equal to unity over just one time step
and zero at other times. The impulse heat flux is the smallest temporal fluctuation that can be
estimated, and any time variation of the surface flux can be represented by superposition of
such basic fluxes due to the principle of superposition because we are assuming a linear
problem. The bias is defined as the square root of the sum of the squares of the deviations of
estimated fluxes from the true ones
B=
qm q m
(3.4.16)
m=1
The purpose of the second test case is to study the sensitivity to measurement errors of
inverse algorithms. The test case considers input temperatures equal to zero, except at time t0
when the temperature is equal to unity. These input data can be considered as an error of
magnitude 1 at time t0. If the random errors are additive, uncorrelated, and have zero mean
and constant variance then the standard deviation of the heat flux components for linear
problems is
q=
q m 2 .
(3.4.17)
m=1
methods, when the stabilizing parameters are increased, the impulse is spread in time. As it
can be seen from the result tables 3.53.6 the best methods are B and D. Method B is a bit
more accurate and Method D is less sensitive to the measurement error.
0,005
0.01
0.05
Method A
1,85
1,65
1,43
Method B
0,79
0,81
0,8
Method C
0,76
0,84
1,03
Method D
0,82
0,84
0,91
0.005
0.01
0.05
Method A
35 832
33 02
78,2
Method B
573
63
5,4
Method C
883
103
10,8
Method D
442
53
3,1
37
good and were in agreement with the expectations done by the sensitivity study. All results
were obtained using a stabilizing parameter, the number of forward time steps, set to 3.
A larger error for a larger HTC was expected, since the sensitivity coefficients were smaller
for a larger HTC. The errors for noiseless data are listed in Table 3.7.
Table 3.7 Error summary for noiseless data [42]
HTC peak
Maximum error
RMS error
10
2.0 %
0.5 %
100
2.3 %
0.6 %
1000
4.5 %
2.4 %
Next, the Gaussian noise with a standard deviation of 0.5 C was added to the simulated
temperature history. The estimated HTC seemed good for the HTC peak, but worse for the
constant HTC (see Figure 3.15). For a larger stabilizing parameter, the loss of accuracy is
present near the peak in the curve. Table 3.8 gives a summary of the error obtained for the
cases with noisy data.
Table 3.8 Error summary for noisy data [42]
HTC peak
Stabilizing
parameter f
Maximum
error
RMS error
10
218 %
57 %
10
10
100 %
33 %
10
13
68 %
23 %
100
45 %
13 %
1000
56 %
15 %
1000
10
24 %
8.4 %
The used inverse algorithm gives very good results with errorless data. The error of
computed HTC increases with increasing magnitude of HTC. When errors are present in the
input data, a larger stabilizing parameter is required. Data errors are more problematic when
the response in the domain is small. The proposed inverse algorithm appears to give better
results when the HTC is changing with time.
38
A)
B)
Figure 3.15 Estimated heat transfer coefficients compared with exact values noisy data;
A) f = 7; B) f = 10; by Woodbury [42].
39
The stabilizing parameter, the number of forward time steps, was set to 2 for a time
shorter than 2.5 s and 12 for a longer time. The computed HTC history (see Figure 3.17
showed that during the important early time the HTC increased rapidly to a maximum value in
about 0.5 s after immersion. Then the HTC decreased with time with a distinct change of
slope at about 2 s. The results were very similar for ethylene glycol and water, only the
magnitude was different. The thermal field around the sphere is quite complicated due to
sudden immersion of a relatively large sphere. The HTC were compared with well-known
steady-state empirical equations for free convection (see Figure 3.18). The computed HTC
coefficients were about 100-120 % higher than the ones obtained using empirical equations
due to the sudden immersion. For later time, the obtained data match well the empirical ones.
40
measurement error was added to the input data. The fourth numerical experiment makes use
of the triangular shape of boundary conditions with a 10 % error in the input data.
The proposed method effectively reduces the average relative error in all four cases.
When a 10 % random measurement error was considered the error in the computed heat
source was reduced from 17.82 % to 2.65 % (see Figure 3.19). The disadvantage of this
method is the computation of large inverse matrices. The more the future time steps, the larger
the matrix. This approach can hardly be used for multidimensional models. The results will be
probably much worse for nonlinear problems.
Figure 3.19 The comparison of estimation results for sequential approaches in example 4,
10% random measurement error considered [44].
A two-dimensional IHCP
Woodbury [45] tested the multidimensional Becks approach on a two-dimensional
model using a numerical experiment. The unknown heat flux on the boundary is assumed as
a function of space and time. Functions are piecewise constant for space and time domains.
y, q
Pulse 2
Pulse 3
Pulse 1
Sensor 1
Insulation
Sensor 2
Sensor 3
x
During a numerical experiment, the object (see Figure 3.20) was exposed to known
(time and space dependent) heat flux (see Figure 3.21). The computed temperature histories of
three sensors inside the body were recorded (see Figure 3.22). A random noise was added to
the recorded values and these data were used as an input for the IHCP algorithm:
T data =T pure erand 1,1 .
Pulse 3
q, Heat flux
Pulse 1 Pulse 2
(3.4.18)
2 time
1
0
Figure 3.21 Time history of the tree heat flux pulses.
The inverse algorithm uses stabilizing parameter f = 1 for the case with noiseless data
(see Figure 3.23 for computed results) and the data with additional noise e=0.0017 (see
Figure 3.24). Very good results were obtained. For the case with more noisy data, e=0.017,
the results are much worse (see Figure 3.25). The maximum error of the estimated heat fluxes
was almost 100 % at the beginning of the experiment. It would be interesting to compare the
results computed using the 2D inverse algorithm and 1D inverse algorithm for each sensor
separately.
42
x2
1
exp
H t .
F x , t , , =
4t
4 t
(3.4.19)
For simplicity, he also assumed constant boundary elements, i.e. the temperature and the heat
flux constant over each element. A liner heat flux in the time domain on the boundary was
used in the numerical experiment.
Various methods have been tested for the solution of IHCP using BEM: direct method,
least square method, regularization method, and minimal energy method. The computed
results were compared with an analytical solution. It has been shown that the least square and
minimal energy methods give very inaccurate results at the end of the computed time interval
and are very unstable, while the regularization method gives good results also for noisy data.
The minimal energy method gives good results for a noisy data except at the end of the
computed time interval. Although the regularization method gives the best results, it is not
an easy task to determine the regularization parameter in practical computations. Similar
results were obtained by Han [16].
43
k T N i T k i
(3.4.20)
i=1
n
c T N i T ci .
i=1
(3.4.21)
= T T T *T
(3.4.22)
where T * is the vector of measurements, T are the values computed using the numerical
model, is the sensitivity matrix, which expresses the temperature dependence on the
unknown parameters, and the is a correction vector of the unknown parameters
[]
k1
k2
= k n .
c1
cn
(3.4.23)
(3.4.24)
c T =1sin T /10
(3.4.25)
44
within the temperature range 0T 5. During the numerical simulation with a time step of
0.001 s, temperature histories were recorded in five locations.
The results obtained from the algorithm vary with the number of nodes used for
approximation and with the additive noise. The number of nodes varies from 3 to 8. The
errors are the highest at the end nodes and lower in the middle for the temperature range (see
Figure 3.263.27). This is caused by the lack of the information on one side of the end nodes.
The results have shown that with four or five nodes the distribution of errors is approximately
uniform while for six or more nodes, the errors on the end nodes become significant, while the
errors of the interior nodes are negligible. The estimated temperature dependent functions of
the material properties were good, even in the case of 1% noise in the input data.
45
A numerical experiment was used to illustrate the accuracy of the approach. A triangular
and a step constant contact conductance are examined. Constant material properties of the
mold were assumed. One temperature sensor was located on the outer surface of the mold and
the second one was inside the mold. A noise was added to the measured data.
It has been found that the conjugate gradient method converges faster and is less
sensitive to the measurement error than the least square method. The disadvantage of this
method is that the function of the gap heat resistance is estimated at once for the whole time
domain. Thus, it can be hardly used for experiments with a lot of input data.
3.4.7 Methods Based on Artificial Intelligence
In the last ten years, a novel approach started to be used in IHCP. The new approach
uses artificial neural networks and a genetic algorithm to determine boundary conditions using
the temperature history from one or more internal sensors.
Artificial neural network
Radial basis function training algorithm
One of the first tests was done with artificial neural network (ANN) trained using the
radial basis function algorithm. The radial basis function algorithm is more efficient for
training than the back propagation method. The goal was to develop an alternative method for
approximation of the inverse function using ANN. Such networks are capable of
approximation of nonlinear functions. The ANN maps a vector of nin real-valued inputs onto
a vector of nout real-valued outputs (see Figure 3.28). The precise function computed by the net
is determined by the networks adjustable parameters its weight coefficients. These are
usually determined during training. Training a network involves adjusting some or all of the
networks weights to minimize the cost function. Once the training has been completed, the
networks ability to approximate the desired function can be tested by applying data not
included in the training.
The data used to train the ANN are usually computed by a forward solver. Dumek [49]
has tested ANN on a one-dimensional object (with internal temperature sensors) that was
subjected to time dependent boundary conditions. The ANN was trained using the radial basis
function algorithm and 300 forward solutions. The input vector contained 50 temperatures and
the output vector had 13 values of the desired HTC. The hidden layer had 30 neurons.
46
Input vector
1
1
1
...
2
2
...
4
nin
nh
...
Input layer
Hidden layer
nout
Output layer
Output vector
Figure 3.28 Structure of artificial neural network.
After completion of the training process, the ANN was tested using the data obtained
from numerical experiments where time dependent HTC were used. These consist of
sinusoidal, constant and triangular functions. The network generates a roughly sinusoidal
output. The results were a bit better for triangular HTC but very inaccurate for constant HTC.
The error of the computed boundary conditions varies from 10 % to 30 %. However, the
advantage of the described approach is the computation speed of HTC after the ANN was
trained.
Back propagation
Sablani [50] tested the trained ANN using the back propagation algorithm on
determination of Biot number
Bi=
hL
.
k
(3.4.26)
Numerical experiments of a cooled cube were realized linear and nonlinear problems.
A simple ANN was used for the estimation of Biot number from the temperatures recorded
inside the cube. The neural network models were capable in generalizing the behavior of both
linear and nonlinear problems. The error of the computed results was usually up to 10 %,
however, it increases up to 30% for very small Biot number (Bi < 0.1).
Cascade correlation
The inverse heat conduction task, treated by Raudensk [51], looks simultaneously for
boundary conditions and the time constant of a temperature internal sensor on the basis of the
47
knowledge of temperature measurement from that sensor. The recorded temperatures can be
seriously influenced by thermal resistance between the surface of the temperature sensor and
the measured material. The error increases when very fast dynamic processes are studied.
The one-dimensional heat conduction problem with constant material parameters was
studied. The cascade correlation learning algorithm was used. It has two special features: the
cascade architecture, and the learning rule. The cascade architecture is shown in Figure 3.29.
At the beginning, the system works with no hidden unit. When the error stops decreasing
hidden units are added one by one to the net during the training process. The main advantage
is that learning is much faster and the hidden units are added dynamically to the learning
process. No fixed topology has to be chosen at the beginning.
Output vector
Hidden neurons
Frozen weights
Input vector
Adjustable weights
+1
48
cascade correlation topology are much better than the ones computed using classical ANN
trained using radial basis function training or the back propagation algorithm.
Genetic algorithm
Genetic algorithms (GA) are random search strategies which use partially random
methods to evolve a solution toward the optimum or near-optimal solutions. In each step, the
GA deals with a group of solutions called population. The algorithm keeps the best found
solutions and it determines a new search direction in the space of possible solutions on their
base. In each step, the search strategy tries to generate a new generation of population
representing a better solution. The new population depends on the previous one. Using the
selection procedure, crossover and mutation operators, a new population is generated. The
procedure of generating a new population is repeated until an acceptable solution is found.
Because of the simple mathematical background of GA, they can be easily used in many fields
such as IHCP. However, it is usually more efficient to use the classical approach. Only if the
classical approach is unstable a GA may be a good choice.
The GA was tested by Raudensk [52] on a data obtained from a quenching experiment.
The obtained results were compared with the data computed using the classical Becks
approach. The accuracy of the obtained results are comparable with that obtained using ANN.
However, data from a real experiment were used and thus it would be very interesting to
compare temperature histories computed by the forward solver using the HTC obtained using
GA and Becks approach.
Liu [53] tested intergeneration-projection genetic-algorithm (IP-GA) for IHCP. This
inverse procedure should speed up the process of finding the desired parameters.
Identification of the heat convection constants of a typical microelectronic package was
performed as an example. The performance was compared with standard GA. It has been
found that the reduced-basis method combined with the IP-GA significantly speeds up the
solution process.
for inverse tasks, because the noise is gained. Various techniques are used to filter measured
data. Some use filtration in the time domain, some cut high frequencies in a frequency
domain, and others approximate the measured data using some function.
3.5.1 Filtration in Time Domain
One of the simplest and often used methods is averaging several values into one value.
This approach is inbuilt directly in measuring apparatuses when an integration measuring
method is used. However, this approach seriously limits the shortest possible time interval
between consecutive samples.
A similar approach can be used without stretching the time interval. One sample is
filtered using several neighboring samples multiplied by some weight coefficient
mt
=
1
mp t
p
(3.5.1)
where m are measured data, and p is a filter. An arbitrary filter can be chosen. However, the
most often used one is constant for a few neighboring samples, tent function, and Gaussian
filter defined as
p =e
2
2 2
(3.5.2)
50
M f = mt e
2i f t
dt ,
(3.5.3)
dt .
(3.5.4)
mt = M f e
2i f t
If t is measured in seconds, then f is in cycles per second. However, the equations work with
other units too.
With two functions mt and pt (where mt are the measured data and p t is a filter
function), and their corresponding Fourier transforms M f and P f , we can make
convolution of the two functions m p as
m p= mp t d
(3.5.5)
(3.5.6)
In other words, the Fourier transform of the convolution is just the product of the individual
Fourier transforms.
Fourier transform of discretely sampled data
In the most common situations, function mt is sampled at some time intervals. If the
time interval between consecutive samples is t, the sequence of sampled values is
mn =mn t
(3.5.7)
The time interval t is called the sampling rate. If t is measured in seconds, then the
sampling rate is the number of samples recorded per second.
For any sampling interval t, there is also a special frequency fc, called Nyquist critical
frequency [5], given by
f c=
1
.
2 t
(3.5.8)
The Nyquist critical frequency is important for two reasons. The first one is known as the
sampling theorem: A continuous function mt (sampled at an interval t) is the bandwidth
limited to frequencies smaller than fc. This theorem shows that the information content of
51
a bandwidth limited function is smaller than that of a general continuous function. It is bad
that all of the power spectral density that lies outside of the frequency range fc < f < fc is
spuriously moved into that range. This phenomenon is called aliasing. Any frequency
component outside of the frequency range (fc; fc) is aliased (falsely translated) into that range
(see Figure 3.30).
M( f )
aliased Fourier transform
1/ 2 t
1/ 2 t
b
T t = a j p j t
(3.5.9)
j=1
where pbj t are the known basis functions. An optimum number of parameters or basis
functions was chosen by using a visual approach. Loulou started with small number of
parameters and then he added one by one until the smooth profile become unstable. When he
52
observed that the smoothed curve is in agreement with the physical evolution of the profile
and when it become unstable he stopped increasing the number of parameters and chose the
last one as optimum.
53
Problem Solution
4 PROBLEM SOLUTION
The solution of the inverse heat conduction problem (IHCP) is an extensive task that
involves many sub-problems. All reviewed IHCP methods incorporate a forward solver of
heat conduction. To solve accurately IHCP, an accurate forward solver is needed. Any noise in
the input data is gained by the inverse task. Thus, the processing of measured data is obvious.
And, finally, the design of a measuring apparatus is important. An unsuitably placed thermal
sensor can significantly reduce the accuracy of the computed results.
Several methods have been considered for the forward solver of heat conduction. For
problems under solution in this work, the most suitable ones are FDM, FVM, and FEM. The
accuracy of these methods was tested as shown in Appendix E. It can be seen that the accuracy
of these methods is very similar. Even the simplest one, FDM, gives best results at nodes of
control volumes for very fast changes in boundary conditions. The equations of FVM and
FEM methods become much complicated for models with temperature dependent material
properties in comparison with FDM. Other methods were also considered. However, the BEM
is not suitable for multidimensional models with temperature dependent material properties.
The FV-UM can better fit within complex geometries. However, the computation is much
more complicated and time consuming. It is probably more efficient to use FDM and a finer
mesh. The TLM method is suitable mainly for extremely fast changes in boundary conditions,
such as a time interval shorter than few picoseconds.
Dealing problems in the metallurgical industry, the studied objects are usually made of
steel. Most of the common steel materials have temperature dependent material properties.
Even more, the phase change occurs not only during solidifying or melting but also at a lower
temperature, when the internal structure of the solid changes. Thus the forward solver must be
able to deal with latent heat.
As the finite methods build an equation for each control volume, the set of equations
should be built in a smart manner to allow as faster a solution as possible. In the case of
one-dimensional problem, the set of algebraic equations can be built as a tridiagonal system of
equations which can be solved very effectively and fast. However, moving into the
multidimensional heat conduction the system of equations becomes much more complicated
and the created matrices are sparse but neither tridiagonal nor band diagonal. The matrices
54
Problem Solution
become extremely large and only much slower algorithms than the tridiagonal one can be
used, such as LU decomposition.
To solve the IHCP, we first derive discretization equations for the forward solver. They
must handle the temperature dependent material properties, phase change, and effectively also
the multidimensional problem. Next, automatic refinement will be proposed in both time and
space domains. Effectively measured input data filtration will be designed and compared with
others that are usually used. Having prepared the forward solver and input data, we will treat
the IHCP.
The material properties are temperature-dependent but constant over the control
volume and related to the node of the control volume.
We first derive the equations in Cartesian coordinates. Next, the equations will be
extended to cylindrical coordinates. Treatment of the three kinds of boundary conditions
(surface temperature, surface heat flux, and convection) will be applied. The equations will
also be extended to handle the phase change.
55
Problem Solution
dx
dt
(4.1.1)
where
q=k
dT
.
dx
(4.1.2)
The time domain is discretized into the time steps where the index of the current time
step is m and the time step is defined as
T, Temperature
(4.1.3)
j-1
j+1
x, position
Node
xj
Control
volume
boundary
xj
2
q0
qJ
0
01
qJ+1
j-1
J-1
j+1
J+1
J+2
Problem Solution
The object domain is discretized into a number of control volumes (see Figure 4.1).
Some methods suggest to create the surface volume of zero thickness and others suggest the
use of surface volume of the half size of the interior volumes. We use zero thickness, because
we will not use constant size of the interior volumes. This enables us to create small volume
next to the zero-volume which gives us the possibility to compute surface temperature more
accurately.
For each control volume a conservation of energy is applied to derive a set of algebraic
equations. Integrating Eq. (4.1.1) over the j-th control volume, we can write
T mT m1
j
q mJ q mJ 1 x jq j = x j j T mj c j T mj j
t
(4.1.4)
where j T mj and c j T mj are the temperature-dependent mass density and specific heat,
m1
respectively, T j
is the temperature of the node in the previous time step and t is the
time step.
When there are temperature-dependent parameters in the equations (such as material
properties, heat source, or boundary conditions), the set of equations becomes non-linear. The
set of these non-linear equations is solved by iterating a method for the linear problem as will
be described later. For the first iteration, the initial guess T m of the temperature T m is equal
to T m1 . After first iteration, the T m is set to T i1, m , where i 1 is the previous iteration
index. During the iteration process, the T m becomes T m .
Temperature dependent heat conductivity
Because the conductivities kj-1 and kj of the neighboring volumes may be different, there
may be a discontinuity of the slope (dT/dx) at the control-volume boundary (see Figure 4.1).
The heat flux can be expressed on the J-th boundary at the time step m as
k j1 T mj1 m
k j T mj m
m
q =
T j1T J =
T J T mj
x j1
xj
2
2
m
J
(4.1.5)
57
Problem Solution
xj
x j1
2
2
m
qJ =
m
m
k j1 T j1 k j T j
1
T j1T j
m
(4.1.6)
k T k T
T = j1 j1 j j
x j1
xj
2
2
m
J
]
1
k j1 T j1 m
k T m
T j1 j j T j
.
x j1
xj
2
2
(4.1.7)
Whenever there is a need to get the temperature at a control-volume boundary (e.g. during
interpolation of temperature across the control volume), this equation should be used together
with the temperature at the node.
Temperature dependent mass density and specific heat
When the mass density or specific heat c becomes more dependent on the temperature,
the results are more inaccurate. Special care must be paid to time integration: if the time step
is too large, the latent heat of the phase change may be neglected [A9]. Figure 4.2 shows the
difference between the correct amount of energy stored and the amount of energy stored in the
case of neglected latent heat. In the case of a phase change, the properties are highly
dependent on the temperature.
c , H
Tf
H*
T f
58
Problem Solution
This problem can be eliminated by solving it using the enthalpy instead of specific heat
and mass density as was described in the Phase Change Implementation chapter. The
following approach was inspired by the Apparent Heat Capacity method. Enthalpy H
represents the energy stored and is defined as follows
T
H T = T c T dT .
(4.1.8)
T ref
dH T
.
dT
(4.1.9)
m
j
j T c j T
H j T mj H j T m1
j
T mj T m1
j
(4.1.10)
Substitution of this relation for j T mj c j T mj in Eq. (4.1.4) reduces the possibility of energy
loss. This substitution can be used when the denominator T mj T m1
is non-zero, otherwise
j
j T mj c j T mj should be used (e.g. for the first iteration when T m=T m1 ).
Enthalpy for piecewise linearized mass density and specific heat
The temperature dependent material properties are usually given in a table or the given
dependence can be easily piecewise linearized (see Figure 4.3). Thus it is very convenient to
,c
T w1 T w
T w1
59
Problem Solution
express the enthalpy H using piecewise linearized mass density and specific heat. Equations
T =a , w T b , w ,
(4.1.11)
c T =a c , w T b c , w
(4.1.12)
describe mass density and specific heat in one temperature interval w from temperature Tw1 to
Tw. Applying Eq. (4.1.114.1.12) in the enthalpy Eq. (4.1.8) gives us the equation for
computing enthalpy H within one wth temperature interval T w1 ;T w
T
H T = a , w T b , w a c , w T b c , w dT =a H , w T 3b H , w T 2 c H , w T d H , w H 0 (4.1.13)
T w1
where
1
a a ,
3 c
(4.1.14)
1
a b b a ,
2 ,w c ,w ,w c,w
(4.1.15)
c H , w =b , w b c , w ,
(4.1.16)
a H , w=
b H , w=
0;
H T w1 ;
for w=1
for w1
(4.1.17)
(4.1.18)
The H0 parameter is used for connecting two adjoining intervals and is equal to H(Tw1) that is
computed using the preceding temperature interval T w2 ;T w1 .
Phase change at constant temperature
The major problem in using the described method is the specification of an accurate
numerical approximation for the rapidly changing material properties an isothermal phase
change. Let us consider an isothermal phase change occurring at temperature Tm. To avoid
problems associated with the change in internal energy at a constant temperature, a small
artificial temperature region is introduced. Within this region, the heat conductivity can be
approximated by the linear piecewise continuous curve, and the product of mass density and
specific heat is constant and can be computed as
60
Problem Solution
T c T =
L
T w T w1
(4.1.19)
where L is the latent heat and the wth temperature interval is the introduced small artificial
temperature region.
Source-term linearization
When the source term q is temperature-dependent in Eq. (4.1.4), its value can be
obtained as
q j = q j T mj .
(4.1.20)
In some cases the source term can be expressed also using the unknown temperature
T mj as
a
m
b
m
m
q j = q j T j q j T j T j .
(4.1.21)
(4.1.22)
m
a
b
m
q j T j = q j q jT j .
(4.1.23)
The usage of Eq. (4.1.22) gives the solution faster, but this requires modification of the set of
linear equations. Eq. (4.1.23) is easier to implement, however, it can diverge as the q bj term is
increasing in absolute value and the under-relaxation must be used.
Set of equations for all internal grid points in the general form
The equations for all internal points of 1D model can be written in a general form
,m
,m
a j , j1T ij1
a j , jT ij, ma j , j1T ij1
=b j
(4.1.24)
where the j subscript is the node index from 1 to N 1. The i superscript is an index of
iteration and the m superscript is a time step. The solution of these equations give us the
temperatures for time step m. The a and b parameters can be obtained by substituting
Eqs. (4.1.6), (4.1.10), and (4.1.21) into Eq. (4.1.4). The result for the first iteration is
61
Problem Solution
a j , j1=2
x j1
m1
k j1 T j1
1
xj
m1
k j T j
m1
a j , j =a j , j1 a j , j1 x j q j T j
a j , j1=2
xj
m1
k j T j
m1
k j1 T j1
j T j
j T m1
c j T m1
j
j
,
t
1
x j1
m1
b j = x j q aj T m1
j
m1
c j T j
t
m1 .
T j
(4.1.25)
m
For all subsequent iterations T i1,
substitutes for T m1
in aj, j-1 and aj, j+1. The aj, j and bj
j
j
parameters are
m
a j , j =a j , j1a j , j1 x jq bj T i1,
,
j
i1, m
a
j
b j = x j q T
i1, m
j
H j T j
m1
H j T j
t
.
(4.1.26)
r dr
dt
(4.1.27)
where
q=rk
dT
.
dr
(4.1.28)
q q
m
J 1
T mj T m1
j
r j r jq j =r j r j j T c j T
t
m
j
m
j
(4.1.29)
62
T, Temperature
Problem Solution
j-1
j+1
r, radius
r j
Node
Control
volume
boundary
r j
2
qJ
q0= 0
0
01
2
2
qJ+1
j-1
3
J-1
J
rJ
j+1
J+1
J+2
rj
Figure 4.4 One-dimensional model for cylindrical geometry showing control
volume and its boundaries.
where r is the radius (distance from the center). For the cylindrical coordinates, the heat flux
q mJ of Eq. (4.1.29) can be evaluated as
r j1
r j
2
2
q mJ =r J
m
k j1 T j1 k j T mj
1
(4.1.30)
T mj1T mj
k T m k T m
T = j1 j1 j j
r j1
r j
2
2
m
J
]
1
k j1 T mj1 m
k T m
T j1 j j T mj
.
r j1
r j
2
2
(4.1.31)
63
Problem Solution
Equations for internal grid points in the general form for 1D cylindrical coordinates
We now use the same form of equations (Eq. 4.1.24) for all internal points as for the 1D
model in Cartesian coordinates. The a and b parameters can be obtained by substituting
Eqs. (4.1.30), (4.1.10), and (4.1.21) into Eq. (4.1.29). The result for the first iteration is
a j , j1=2 r J
r j1
m1
k j1 T j1
1
r j
m1
k j T j
m1
a j , j1=2 r J 1
r j
m1
k j T j
a
j
b j , j =r j r j q T
m1
j
j T m1
c j T m1
j
j
,
t
1
r j1
m1
k j1 T j1
j T m1
c j T m1
m1 .
j
j
T j
t
(4.1.32)
m
For all subsequent iterations T i1,
substitutes for T m1
in aj, j-1 and aj, j+1. The aj, j and bj,
j
j
parameters are
b
i1, m
a
j
b j =r j r j q T
i1, m
j
m
H j T i1,
H j T m1
.
j
j
(4.1.33)
64
Problem Solution
x N 1
x1
x1
2
x N -1
2
qs
qs
N-1
01
Ts
N-1
N N+1
Ts
(4.1.34)
(4.1.35)
(4.1.36)
where j = N.
Surface heat flux
Knowing the surface heat flux we can write the equation for the surface control volume
q s =k
T 0T 1
x1 .
2
(4.1.37)
(4.1.38)
65
Problem Solution
k N 1 T mN 1
; a j , j =a j , j1 ; b j =q s
a j , j1=
x N 1
2
(4.1.39)
where j = N.
Convection
The surface heat flux can be computed as
q s=h T s T .
(4.1.40)
Therefore, we can write the equation for the surface control volume
k 1 T 1m
h T 0 T =
T 1 T 0
.
x1
2
(4.1.41)
(4.1.42)
(4.1.43)
where j = N.
Radiation
The heat flux due to the radiation from the surface is expressed in Eq. (12.3). It is
convenient to express the radiation heat flux in the form
q rad =h T s T sur
(4.1.44)
(4.1.45)
Therefore, the radiation can be simply integrated into the equations for convection where
66
Problem Solution
h=hconv h .
(4.1.46)
set to T j
, where i 1 is the previous iteration index. This is repeated for all next
m
m
iterations. During the iteration process, T j becomes T j . The number of required iterations
(4.2.1)
/ 3D model /
,m
i ,m
a j , k , p1T ij,,mk , p1a j , k 1, pT ij,,mk 1, p a j1, k , pT ij1,
k , p a j , k , pT j , k , p
i ,m
i ,m
(4.2.2)
67
Problem Solution
1D heat conduction
A matrix
qJ
qJ+1
j1
J1
zeros
j+1
J+1
J+2
zeros
2D heat conduction
A matrix
K+2
zeros
j1, k+1
j, k+1
zeros
j+1, k+1
K+1
qj,K+1 q
J+1,k
qJ, k
j1, k
j, k
zeros
j+1, k
zeros
K
qj,K
j1, k1
K1
J1
y
j, k1
J
j+1, k1
J+1
J+2
A matrix
3D heat conduction
qj,K+1,p
qj,k,P+1
qJ,k,p
qJ+1,k,p
j,k,p
z qj,k,P
qj,K,p
x
Figure 4.6 Control volumes and A matrixes for 1D, 2D, and 3D heat conduction.
By arranging the set of equations into the matrix form, Eq. (3.3.47), we will get very
sparse matrixes A. Examples of these matrices are shown in Figure 4.6 for rectangular and
cubic geometries. Although A matrixes may seem to be band diagonal systems of equations
68
Problem Solution
they are not because the condition bandwidth << N is not fulfilled. This set of equations can
be effectively solved by Conjugate Gradient Method or LU Decomposition, but there is
a more efficient solution of this problem.
The key point to solution is the principle of superposition of heat conduction two or
three-dimensional conduction may be expressed as a product of one-dimensional conductions.
To solve our multi-dimensional problem, we can split it into separate one-dimensional
problems along the main axes (Cartesian, cylindrical, or spherical coordinates). This approach
is also called line-by-line method. We can compute Eq. (4.2.1) and Eq. (4.2.2) within two and
three iterations, respectively, as follows:
/ 2D model /
,m
i ,m
i ,m
,
m
m
a j1, kT ij1,
k a j , kT j , k a j1, kT j1, k =b j , k a j , k 1T j , k 1 a j , k 1T j , k 1
i1 i
(4.2.3)
a j , k 1T ij,,mk 1a j , kT ij,,mk a j , k 1T ij,,mk 1=b j , k a j1, kT mj1, k a j1, kT mj1, k .
(4.2.4)
/ 3D model /
,m
i ,m
i ,m
a j1, k , pT ij1,
k , p a j , k , pT j , k , p a j1, k , pT j1, k , p
(4.2.5)
We have three unknown temperatures on the left-hand side and all parameters known on the
right-hand side of the equations. The equation for the first iteration, Eq. (4.2.3), represents k
sets of tridiagonal systems equations one set for one line of nodes along X direction. Each
line of nodes along the X axis is solved as a 1D heat conduction. The next equation,
Eq. (4.2.4), represents j sets of tridiagonal systems equations one set for one line of nodes in
69
Problem Solution
the Y direction. A similar situation arises for the 3D model. The approach splits a huge set of
equations (especially for 3D model) into a number of relatively small tridiagonal systems of
equations which can be solved very effectively.
A smart choice of temperature T m can significantly reduce the number of required
iterations. So far, the temperature from the previous iteration T i1, m was used for temperature
T m . However, more up-to-date values can be used. We can use the values of the current
iteration of already computed lines. To avoid the new values are being every time on the same
side of the line (e.g. at the bottom and on the left side for the 2D model in Figure 4.6), it is
recommended to invert the order of the computed lines after each iteration.
4.2.1 Error Estimation of Iteration Algorithm
The previous chapter describes the iteration process for the solution of the non-linear
equations. The iterative solution process gives us approximate temperatures of the control
volume nodes. The temperatures may converge to the solution or diverge during the iteration
process. If divergence occurs, an under-relaxation should be considered although it is not very
common and it occurs only in some special cases with a heat source.
The maximum estimated error is usually computed as
m
i T ij, mT i1,
(4.2.6)
m
where T ij, m and T i1,
are the node temperatures computed from the last and next-to-last
j
iterations, respectively. This estimation is good for problems that converge fast. As you can
see in Figure 4.7, this estimation can be very inaccurate if the problem under solution
converges slowly [A9]. The estimated error after the ith iteration is i but the difference
between the last computed value and the value computed after an infinite number of iterations
i . This difference is the sum of the estimated errors from all next iterations that have not
is
been solved yet.
A better estimate of the convergence error can be determined using a method based on
i between
the observed geometric convergence of heat conduction problems. The difference
the last computed value and the value computed after an infinite number of iterations can be
estimated using the sum of a geometric series
70
Problem Solution
Temperature
i3, m
i2, m
T
i1, m
T i ,m
T
i1
i
i
i-3
i-2
i
number of iterations
i-1
Figure 4.7 Estimated and real errors after ith iteration.
i =i1i2= p1 p2 p3 p = p 1n = p 1
b
b1
b b b
n=1 b
(4.2.7)
b=
i1
T i2
j T j
i
T i1
j T j
p=T j T j
i1
i2, m
where T j
i1, m
, Tj
i ,m
, and T j
(4.2.8)
(4.2.9)
iterations.
The iteration process converges faster for a higher b parameter. For the cases where
b2, the estimated error is less than p and thus Eq. (4.2.6) can be used. On the other hand,
whenever the b parameter becomes closer to 1, the solution will converge slowly and
Eq. (4.2.7) should be used. This occurs mainly when highly conductive materials are used
inside materials with poor thermal conductivity.
Problem Solution
(B)
Boundary condition
Boundary condition
(A)
measured value
real profile
time
In most cases we do not know the real profile of boundary conditions, we only know the
values in certain measured time instants (see Figure 4.8). We usually use a piecewise
linearized profile between the measured values. Knowing the boundary conditions in certain
measured time instants, we can simplify them in various ways. Two methods are shown in
Figure 4.9. The B method is often used, but the A method is more effective.
We can estimate the error of simplified boundary conditions for case A as a sum of
partial errors
n
=
m=1
m m1
4
(4.3.1)
where the m is the measured boundary condition at time step m and n is the number of
computed time steps. The estimated error for case B is twice higher.
72
Problem Solution
(B)
time step
m+1
m+2
Boundary condition
Boundary condition
(A)
m+1
m+2
time
time
T, Temperature
computed, and based on this information the mesh refinement can be done.
x, position
(A)
(B)
control volume size
Figure 4.10 Mesh design.
73
Problem Solution
We will investigate the mesh discretization error by comparing the heat energy stored
within the control volume. The heat energy of the control volume can be expressed as
an enthalpy
x J 1
1
H =
x j x= x
c dT dx .
*
j
(4.3.2)
T ref
Because we are assuming constant material properties within one control volume, we can
simplify Eq. (4.3.2) as follows
*
*
H j =cT j .
(4.3.3)
This simplifies our search for the mesh discretization error. We can only investigate the
average temperature of the control volume and, if necessary, we can recalculate the average
temperature to heat energy using Eq. (4.3.3).
First, we must avoid a very rough mesh as shown in Figure 4.10, case A. Next, we can
assume two extreme conditions shown in Figure 4.11. Case A shows the situation shortly after
the rise in temperature of the control volume j 1 has risen. The temperature on the boundary
T J is very high compared to the rest of the temperatures within the control volume j. Case B
shows the situation later after the change. The thermal shock propagates deeper inside the
control volume. Because of the piecewise linearized profile, we only know the temperature at
the center of the node and we can estimate temperature on its boundaries. We do not know
anything about the real temperature profile between the centers of the control volumes.
However, we can express the two extreme conditions shown in Figure 4.11. As you can see in
case A, the real average temperature of the jth control volume T *j is very close to the
temperature of the jth node T j . On the other hand, in case B, the real average temperature is
much closer to the average temperature computed from the temperatures of the node and on
the boundaries of the control volume as follows
T j =
T J , j T j , J 1 T J T j / 2T j T J 1 / 2
.
=
2
2
(4.3.4)
If we take a closer look at FDM and FVM, we find that the FDM assumes the
temperature of the node T j for the average temperature of the control volume, while the
FVM assumes T j expressed in Eq. (4.3.4).
74
Problem Solution
(A)
(B)
T J , j T j , J 1
T j =
2
T *j
FVM
j
T FDM
j
Tj
T j , J 1=T j T J 1 / 2
Tj1
T, Temperature
T, Temperature
Tj1
Tj+1
TJ+1
TJ
T J , j =T J T j / 2
T j =
T FVM
j
FDM
T j
TJ+1
Tj
x, position
T J , j T j , J 1
2
*
T j
Tj+1
x, position
Figure 4.11 Temperature of a control volume in FDM, FVM, and real average
temperature of the control volume.
Assuming the real average temperature is between the T FDM
and T FVM
, we can
j
j
estimate the discretization error as the difference between the temperature at the center of the
node and the average temperature T j using Eq. (4.1.7) and (4.3.4) as
T =T FDM
T FVM
=T j T j=
j
j
1 k j T j x j1T j k j1 T j1 x jT j1 .
4
k j T j x j1k j1 T j1 x j
(4.3.5)
The same equation can be obtained for cylindrical coordinates using Eq. (4.1.31). The x is
only replaced with r.
4.3.3 Time Discretization Error
Dealing with the time discretization, we should compare the three main schemes for
temperature variation in time explicit, Crank-Nicholson, and fully implicit shown in
Figure 4.12. The explicit scheme essentially assumes that the old temperature lasts through the
entire time step. The fully implicit scheme supposes that the temperature suddenly drops at the
beginning of the time step and stays over the whole time step, while the Crank-Nicholson
scheme assumes a linear variation of the temperature during the time step.
Working only with the unconditionally stable fully implicit scheme, we will derive
a method for estimating the error caused by time discretization. As you can see in Figure 4.13,
none method describes correctly the variation of temperature in time shortly after a change in
the boundary condition.
75
Problem Solution
T, Temperature
Explicit
Crank-Nicholson
Fully Implicit
time step
m1
time
Figure 4.12 Variation of temperature in time for three different schemes.
The trick for estimating the error caused by time discretization is based on comparing
the error of the temperature (hatched area in Figure 4.13) for the time step and the double time
step. As you can see, for the later change in boundary condition, the hatched area for the
double time step is about twice larger than for two steps of the normal time step. Knowing
that the error (the hatched area) is half the size of that for half time step and having computed
our problem under solution using the double time step and normal time step we can estimate
the error caused by time discretization as
t T j =T //j T /j
(4.3.6)
//
where T j is the temperature computed using the double time step and T j is the temperature
computed using the normal time step.
Real tempearature
T, Temperature
Real tempearature
time
The situation is slightly different for the case shortly after the change in boundary
condition, because the hatched area for the double time step is not twice larger than that for
two steps of the normal time step, but slightly less. No ratio lower than 1.5 was observed
76
Problem Solution
when testing the method; the lowest values did not drop under 1.7. Knowing this behavior and
applying this ratio in Eq. (4.2.7) describing the sum of a geometric series, we can estimate the
error caused by time discretization using modified Eq. (4.3.6) as
t T j =T //j T /j
(4.3.7)
where the coefficient varies from 1 to 2, depending on the problem under solution. The
=2 should be used to make sure that the estimation of the maximum error caused by time
discretization is correct ( =2 for ratio 1.5 and =1 for ratio 2). For the infinite time step
(steady conditions), the implicit scheme describes the temperature variation in time almost
perfectly and there is no need to make such an estimation.
t*
trunc * /max 1
(4.4.1)
*
where t is a new time step, t * is time discretization of measured boundary condition,
max
is the sum of the partial error for the time step t * , is the maximum allowed value, and
the trunc function returns only the integer part of the expression in brackets.
77
Problem Solution
(B)
Boundary condition
Boundary condition
(A)
half time
step
m+1
m
time
Figure 4.14 Boundary condition refinement.
m+1
= / 2
m+2
time
control volumes. If the center node of a control volume corresponds to a temperature sensor
location, the volume should be divided into odd number of volumes so that the position of the
center node remains unchanged. If the T parameter is lower for two neighboring control
min
volumes than the limit value T , the control volumes can be merged. However, avoid
merging volumes corresponding to a temperature sensor location and volumes already
divided. The T parameters should be checked for all control volumes after each time step
and the computation along with mesh refinement should be repeated until no re-meshing is
necessary.
78
Problem Solution
computation. If the error is higher than the acceptable value t , the problem should be
computed using the half time step. This approach allows us to use repeatedly Eq. (4.3.7).
Temperature [C]
904.0
903.5
903.0
902.5
902.0
901.5
901.0
1.4
1.5
1.6
1.7
1.8
1.9
2.0
2.1
2.2 2.3
Time [s]
2.4
2.5
2.6
2.7
2.8
2.9
3.0
Problem Solution
1.0
p( t )
0.8
0.6
0.4
0.2
0.0
-3.0
-2.5
-2.0
-1.5
-1.0
-0.5
0.0
t
0.5
1.0
1.5
2.0
2.5
3.0
proposed
multi-level
filtration
procedure
was
designed
especially for
a thermocouple sensor built in a stainless steel plate 1.5 mm under the cooled surface.
However, it can also be modified for similar cases. The temperature was measured using
a 12-bits card at 100 and 300 Hz. An example of measured data is shown in Figure 4.15. The
algorithm uses convolution (Eq. (3.5.5) and Eq. (3.5.1) for discrete data) with the varying
Gaussian filter function
p t =e
t 2
2 krn/32
(4.5.1)
to filter the measured data, where krn represents the with of influenced area (see Figure 4.16).
First, several sets of filtered data are computed. Each set is computed using constant
filter Eq. (4.5.1). The krn varies from krnmin = f / 10 to krnmax = 2f / 10 with step krnstep = 1
which represents one sample. The filtration procedure starts at the beginning of the recorded
temperature history with the first measured sample for which the value from the set computed
using krn m=krn min = f /10 is used where the m superscript represents the current time step.
For the next value, the set computed using krn m=krn m12 (stronger filtration) is used but
only if the following condition is fulfilled
T *T step T T *T step
*m
T T step T m
m2
* m2
m1
* m1
T T
T step T T
T stepT m1T * m1T step T m2 T * m2T step
T T * mT step
* m2
m2
* m1
m1
T
T stepT T
T stepT T * m1T step T m1T * m2T stepT m2 . (4.5.2)
80
Problem Solution
915
905.0
Temperature [C]
Temperature [C]
904.5
914
913
912
911
904.0
903.5
Measured
Filtered
Corrected
903.0
902.5
902.0
901.5
910
0.9
1.0
1.1
Time [s]
1.2
1.3
901.0
1.4
1.5
1.6 1.7
Time [s]
1.8
Figure 4.17 Before the correction function is applied, the filtered data are below the
measured ones in the part where the data changes the slope.
If this condition is not fulfilled, smaller krn m=krn m1 is tested by Eq. (4.5.2). The krn is
decreasing
m
krn =krn
until
m1
the
condition
Eq. (4.5.2)
or
1.0
0.8
y( x )
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
1
[ T T * ]p t
(4.5.3)
81
Problem Solution
where p t is expressed in Eq. (4.5.1) and krn= f /10 is used. The correction function
Eq. (4.5.3) is modified by the non-linear correction function (see Figure 4.18) and then
applied as
1
d T
m
m
1 d T max
T = T cos
2
d T max
m
for m=0.. n .
(4.5.4)
The non-linear function shown in Figure 4.18 ensures that the correction function from
Eq. (4.5.3) is applied more intensively for a bigger deviation. When this correction procedure
is applied twice, the smoothed values are pushed back within the measured ones as can be
seen in Figure 4.17.
82
Problem Solution
x2
2 2
L for x
R for x
(4.6.1)
where the parameters and represent the maximum and minimum value of the HTC,
respectively. The parameter describes the shape in the x direction. The shape for x ,
where represents the nozzle position, is different from x . Hence the parameter is
divided into the two parameters L and R for the left and right sides, respectively.
The method is based on the fact that the optimum approximation function of HTC
minimizes the sum square error between the measured temperatures T *i and the computed
temperatures T i where the sum square error function is defined as
83
Problem Solution
SSE= T i T i .
*
(4.6.2)
This equation is used as a criterion function for the optimization method. The measured
*
temperatures T i at an interior location are known from the measurement and the computed
temperatures T i at the same interior location, but of the computational model, are computed
using tested boundary conditions.
Several methods can be used to determine unknown parameters of the approximation
function by minimizing the criterion function. In most cases, the approximation function has
more than one parameter to determine. The parameters must be thus determined using
methods which are able to compute multiple parameters at once. Usage of the downhill
simplex multidimensional optimization method and an artificial neural network will be
described in the following chapters.
2L
2R
T [C]
HTC [W/(m.K)]
Time [s]
Nozzle position
HTC
Surface temperature
Problem Solution
Original simplex
Maximum value
Step number
xm
x
x
Half simplex
3
3
2
xr
Minimum value
x
Centroid
xe
Reflection vertex
Figure 4.20 Vertex meaning within one step and moving simplex in 2D space.
The basic downhill simplex method was modified to allow optimization of a problem
with constrained parameters, and restart that helps overcome small local minimums is also
implemented [A9, A5]. The description of the modified downhill simplex is as follows:
1. Create a simplex object and compute the value for each vertex of the simplex.
2. Define (reflection), (contraction), and (expansion) parameters. Define the
maximum of optimization steps and maximum of restarts.
3. Define the limitation dimensions used in step 6. These limits can be different for
any parameter and should reflect the ability of the numerical method to compute
different results for two different input parameters. If the difference in input
parameters is smaller than the chosen limit, there might not be any difference in the
computed values or the difference is incorrect due to the rounding errors in the
numerical method. These limits help avoid simplex degradation.
85
Problem Solution
Start
Simplex nodes are defined as x1 ,L , x n +1 n
Find r where f ( x r ) = min 1 j n +1 f (x j )
r
Create empty vector s
Attach s
r
to vector s
1
n
n +1
j =1, j s
n n sr < 3
f ( x r ) > f ( x )
xe is within limits
and f ( x ) > f ( xe )
x x s
xe x s
f ( x m ) < f ( x )
x x s
x x s
Figure 4.21 One optimization step of the Downhill Simplex optimization method with
parameter limitation.
86
Problem Solution
4. Define the smallest allowed dimension of the entire simplex during the
optimization process. The smallest simplex should reflect the desired precision of
the optimization process.
5. Make a simplex movement using one optimization step as shown in Figure 4.21.
6. If any simplex dimension is smaller than the appropriate limit chosen in step 3, go
to step 8 (restart).
7. If the entire simplex is smaller than the smallest simplex defined in step 4, go to
step 8 (restart).
8. (This step stands for restart) If the number of restarts is higher than the defined
maximum or if the position of this minimum is very near to the previous restarting
minimums, you have found the optimum (minimum). If not, choose the node with
a minimum value and create a new simplex around this vertex. This simplex must
be bigger than the limits defined in steps 3 and 4. Restarts help overcome small
local minimums. Go to step 5.
In our case, the optimized values are the parameters of the approximation function, e. g.
the , , , L, and R parameters for Eq. (4.6.1). The downhill simplex method minimizes the
criterion function defined by Eq. (4.6.2) by setting the parameters of the approximation
function. The lower value of the criterion function indicates that the computed temperature
history matches the measured history better and thereby the HTC described by the
approximation function is more accurate.
Usage of feed-forward neural network
Let us imagine so called black box that can compute directly the desired boundary
conditions from the temperature history measured inside the investigated body. No such
general theory is available. However, we can create and train an artificial neural network for
our particular problem [A2]. Such a network will serve us as a black box where the input
vector is a measured temperature history, and the output vector consists of the approximation
function parameters of boundary conditions (see Figure 4.19).
Structure
The basic feed-forward neural network will be used for explanation (see Figure 4.22).
However, the cascade correlation can also be considered. A two-layer network [55] is used as
the main model. The number of neurons in the output layer is equal to the number of the
87
Problem Solution
requested parameters. The number of hidden neurons should reflect the complexity of the
problem under solution.
Training
The training data must be available before the start of the training. A set of various
vectors of approximation function parameters, which should cover the expected range of the
parameters to be computed, is prepared. The real measurement is simulated on a computer for
each vector in the prepared set. Different parameters and the computed temperature history
represent the requested training data. However, the training data should be standardized (to the
range from 0 to 1). Having prepared the training data, the feed-forward neural network is
trained using fast back-propagation, i. e. Levenberg-Marquardt functions.
INPUT
T(1)
HIDDEN LAYER
W 1(1,1)
OUTPUT LAYER
W 2(1,1)
b1(1)
T(2)
b1(2)
T(n)
W 1(n,3)
b2(1)
T(3)
OUTPUT
b2(2)
W 2(3,3)
b1(3)
hyperbolic tangent
sigmoid transfer function
linear transfer function
R
b2(3)
- weight matrixes
b1, b2
- bias vectors
Computation using the learned neural networks is fast and simple. Use the measured
temperature history as an input data, and the neural network gives you the desired parameters
of the approximation function. However, it has been proved that not all of the learned
networks can be used. Some of them return confused results. Verification should thus be
made, e. g. direct heat conduction computation using the computed approximation function
88
Problem Solution
and a comparison of the computed temperature history with the measured one. Note that the
learned neural network can be used only for that particular geometry for which the learning
process has been accomplished.
4.6.2 Computation of 2D Boundary Condition
In this section, an alternative computation of more than one boundary condition is
explained. The inverse method simultaneously computes two boundary conditions: the
thermal contact resistance (TCR) between an alloy casting and the silica mold, and the heat
transfer coefficient (HTC) between the mold and a bath [A9].
Problem description
The alloy-casting apparatus (see Figure 4.23 and 16.616.8 in Appendix G) consists of
a solidifying alloy, mold, and furnace with a molten metal bath. The furnace keeps the metal at
a prescribed temperature. This enables control of the speed of the solidification process. At the
beginning of the experiment the mold with liquid alloy is immersed into the tin bath. During
solidification, the temperature histories of four thermocouples are recorded. The thermocouple
T 1 is placed into the center of the casting. Two thermocouples, T 2 and T 3 , are placed into
the mold (see Figure 16.9) near the inside and outside surfaces of the mold, respectively. The
fourth thermocouple, T 4 , measures the temperature of the molten metal bath.
Mold
Alloy
T1 T2 T3
T4
Tin bath
Furnace
89
Problem Solution
considered as purely conductive, and as one object with spatially and temperature dependent
material properties. The gap between the alloy and the mold, with some thermal resistance and
no specific heat, enables us to simulate the effect of the TCR between an alloy casting and the
mold. During the computation, TCR is converted to the thermal conductivity k gap using
k gap
1
T =
T =h1 T
x gap
TCR
(4.6.3)
where x is the width of the space and T is a difference between the temperatures at the
sides of the gap. If the space is very small, it does not affect the model. In our model, we fix
the distance x gap and vary the conductivity k gap to determine the TCR. Using Eq. (4.6.3) the
TCR can be converted into the thermal conductivity or h1 . On the outside surface of the
mold, a boundary condition of the third kind is considered. The HTC between the mold and
bath is marked as h 2 . The temperature of the bath is considered as the ambient temperature.
Inverse task
A common solution of the inverse task uses Beck's sequential method [11] but any
optimization method can be used. In some cases, optimization methods can work better than
Beck's approach, especially for multiple heat fluxes with temperature-dependent material
properties where the problem under solution becomes nonlinear.
The downhill simplex method can be used to determine time-dependent TCR and HTC
( h1, h2 ). The measured temperature histories (temperature sensors T 1 , T 2 , T 3 , and T 4 )
are used as the input for the inverse task. The inverse task finds the boundary conditions by
minimizing the difference between the measured temperature histories and the temperature
histories computed using the direct heat conduction that uses the tested boundary conditions
and computational model of the problem under solution (a computational model of the
problem must be available to solve the direct heat conduction task). The optimal boundary
*,m
conditions minimize the sum square error between the measured temperatures T i and the
computed temperatures T im where the sum square error function is defined for this case as
nT t f
SSE= T *i , mT im 2
(4.6.4)
i=1 m=t
90
Problem Solution
where the nT is the number of compared temperature sensors, and f is the number of future
time steps. This equation is used as a criterion function for the modified downhill simplex
optimization method [A9]. One value of the boundary condition is computed for each time
step. This means that the described procedure is repeated for each time step.
The computation starts with searching boundary conditions for the first time step. The
optimization process starts with some initial values of the boundary conditions. During the
optimization, the boundary conditions are found by minimizing the criterion function. The
boundary conditions usually remain constant for f future time steps in the criterion function.
However, you may also use time-dependent functions, such as the linear one. Having found
the optimal boundary conditions for the first time step, the direct heat conduction computation
is performed for one time step and the search for boundary conditions is repeated for the next
time step until the boundary conditions for the whole time history are found.
4.6.3 Determination of Temperature Dependent Material Properties
The modified downhill simplex method [A9] can also be used for the determination of
thermal temperature-dependent material properties. The inverse task of this kind is explained
on the determination of thermal material properties of the mold form transient and steady state
measurements of the temperature histories [A5] (see Figures 4.244.25).
Computer
Data acquisition
Measured temperatures
Mold
Heater
A special annular mold was made (see Figure 16.10 in Appendix G), similar to that used
for casting. This silica mold consists of two different layers: soft smooth inner layer and rough
outer backup layer. Two extra thin temperature sensors were placed on the surface on the
91
Problem Solution
opposite sides of the cavity in the mold (see Figure 16.11) and two were placed on the outer
surface. Four more temperature sensors were built-in inside the mold as shown in Figure 4.25.
A cylindrical heater was placed inside the cavity of the mold and attached to the controlled
power source. Another thermocouple was placed on the surface of the heater to monitor the
temperature of the heater. All thermocouples and power were measured and information was
sent from the data acquisition system to the personal computer (see Figure 4.24).
At the beginning of the experiment, the mold and the heater were held at uniform room
temperature. Afterwards, the heater was heated up a little bit and the temperature responses
including the measured power for the heater were recorded into the computer. After reaching
steady conditions with the heated heater, the power of the heater was increased. This was
repeated several times to cover the whole investigated temperature range.
(4.6.5)
T 4 T 3
ln r m / r 4 ln r 3 / r m
2 L k2 L 2 L k1 L
(4.6.6)
q=
for backup layers and
q=
92
Problem Solution
for the smooth inner layer. The parameter q represents the heat rate generated by the heater.
Parameter L is the length of the mold. The indexes a and b represent any thermocouple T1, T2,
or T3 where the b thermocouple must be closer to the outer surface and different from the a
one. Parameters T and r represent the temperature and distance of the thermocouple from the
center of the mold, respectively. Similar equations can be obtained for the opposite side of the
mold where thermocouples T5, T6, T7, and T8 are placed.
There is some inaccuracy because we have to know the precise outgoing heat flux but
some unknown heat is lost on the top and bottom of the cylindrical mold. In addition, the
relative temperature for the computed conductivity is approximate and to compute the thermal
conductivity k2 we have to know conductivity k1 for the same relative temperature.
Computing simultaneously temperature-dependent thermal conductivity and specific heat
The basic idea for this algorithm is that the computational model must match the
experimental model. In other words, if the computational model matches the experimental
one, the computed temperature histories must be equal to the measured temperature histories.
The indicator SSE that represents the accuracy of the computational model is given by
8
*,m
m 2
SSE= T i T i .
i=1 m=1
This sum is computed over all eight temperature sensors and over whole
(4.6.7)
temperature
histories.
Two 1D models are used for computational models, for sides A and B. The boundary
conditions are very simple because we have measured all four surface temperatures. So we
know the surface temperatures on both sides for each model. On the other hand, the material
properties of the silica mold are unknown. And these have to be found in order to match the
computed temperature histories to the measured ones.
The temperature dependent material properties can be computed as piecewise linearized
ones. For each conductivity k1, k2, specific heat c1, and c2 several values are optimized. These
are conductivity and specific heat for node temperatures. The right material properties can be
found using an optimization method, e. g. the modified downhill simplex optimization
method [A9]. The optimization method minimizes the criterion function (given by Eq. 4.6.7)
by changing the material properties. When the minimum of the criterion function is reached
the desired material properties are found.
93
Problem Solution
To maximize the accuracy of the computed results by the inverse task, an accurate
computational model must be used. The calibration of the computational model is described in
the following. First, some experiments are performed to obtain data for the calibration. Next,
the model is calibrated using an optimization method [A12].
A-A
Shield
Insulation
Thermocouple wires
Gap
Main body
To measure temperature inside the body, special sensors with built-in K-thermocouples
are used as shown in Figure 4.26. The main body of the sensor is made of stainless austenitic
steel. A hole of 1.1 mm in diameter for a thermocouple is made from the side of the sensor.
The axis of the hole is 1 mm under the investigated surface and is perpendicular to the
94
Problem Solution
expected heat flux, so that the most important part of the inserted thermocouple lies in one
isotherm. Inside the sensor, a shielded ungrounded K-thermocouple is placed. The gap
between the sensor and the thermocouple is filled with ceramic material that can be exposed
to a high temperature. During measurements that serve for computing HTC, the sensors are
placed in the steel object on which the HTC is investigated.
Calibration experiment
None of the sensors is exactly the same. The position of the junction point inside the
shielded thermocouples differs. Also the hole inside the sensor is a bit bigger than the
thermocouple so that its position may differ. As the thickness of the material in the gap differs,
the heat resistance does, too. These are the main reasons why the calibration experiment has to
be done for each sensor.
The main idea is to perform an experiment we know all about except the heat resistance
between the sensor and built in thermocouple. The sensor at room temperature is exposed to
a high temperature. The measured temperature history must match the computed history in
case the computational model is correct.
Main body
of the sensor
Switch
DAQ
PC
Cross-section of
the shielded
thermocouple
Tested Sensor
High-conductive grease
High-conductive
Grease
Flat Thermocouple
Cu
Thermocouples
Heated
copper
rod
Thermocouple
Upper part
Shield
Gap
Insulation
Lower
part
High-conductive grease
Heated copper
Heater
95
Problem Solution
At the beginning of the experiment the copper rod is heated up to a uniform temperature
and the sensor is held at room temperature. A high conductive grease is applied onto the
surface of the copper rod to ensure a full contact between the sensor and copper rod.
Afterwards, the sensor is stuck to the heated copper rod and the switch indicates the time
when the contact occurred. A 0.05 mm thin pad placed in the grease ensures every time the
same thickness of the grease between the sensor and the heated copper rod.
Computational Model
A 2D axis symmetric model was used as shown in Figure 4.28. The model consists of
the heated copper rod, high conductive grease, and the tested sensor. The model includes the
shielded thermocouple with all its parts. The thermocouple must be taken into account
because the homogeneity of material is disturbed by the inserted thermocouple, and thus the
temperature profile is also disturbed.
Model Calibration
The 2D computational model described in the previous section is modified to match the
computed temperature histories with the measured ones. Supposing the axis of the hole of
1.1 mm in diameter is precisely 1 mm under the investigated surface, the following parameters
can vary:
Position of the thermocouple inside the hole of 1.1 mm in diameter, and the
thickness of the material in the gap between the sensor and thermocouple
Material properties (thermal conductivity, specific heat, and mass density) of the
material that fills the gap
Computational experiments have showen that knowing the specific heat and mass
density of the material that fills the gap, the remaining unknown parameters can be eliminated
by setting only two parameters thermal conductivity of the material that fills the gap in the
upper and lower parts. The dimensions of the model remain, only the material that fills the gap
is divided into the upper and lower part. Both parts can differ in thermal conductivity.
Changing independently the thermal conductivity allows compensation of the position error of
both the junction point and the thermocouple inside the hole. The modified downhill simplex
optimization method [A9] is used for finding the appropriate thermal conductivity of the
material that fills the gap in the upper and lower parts.
96
Problem Solution
Experiment
Initial
Conditions
Computational
Model
ku & kl
Computed
Temperature
History
Measured
Temperature
History
Criterion
Function
Downhill Simplex
optimization method
SSE= T * , mT m 2
(4.6.8)
m=1
where T and T* are the measured and computed temperatures of the thermocouple inside the
sensor, respectively.
Problem Solution
PC
DAQ
Nozzle
Impact
area
Pressure
Position
Moving Plate
Pressure Sensor
Measuring the pressure distribution (see Figure 16.12 in Appendix G), the nozzle sprays
on a moving plate (see Figure 4.30). This plate is equipped with a pressure sensor (see
Figure 16.13 in Appendix G) that may be of circular or rectangular shape. For a given nozzle
configuration, a pressure is measured as position dependent while the plate with the sensor is
moving under the spraying nozzle.
For nozzles with a very narrow spray spot and for a small distance of a nozzle from the
moving plate, the measured data doesnt represent a real pressure distribution. The values are
averaged over a large pressure sensor. The precision of the measured data depends on the
size ratio of a nozzle spray spot and the sensor. As the ratio becomes smaller the precision of
the measured data is getting worse.
Fourier transform
A physical process can be described either in the time domain or else in the frequency
domain as shown in Eqs. (3.5.33.5.4). If m is a function of position x (in meters), M will be
a function of inverse wavelength (cycles per meter). A two-dimensional Fast Fourier
Transformation (FFT) [5] is used to convert both the measured data and pressure sensor to the
frequency domain. The correction of the measured spray distribution is done in the frequency
domain and then an inverse two-dimensional FFT is used to convert the data back to the space
domain. This allows suppressing noise in the measured data that would be extremely
amplified.
98
Problem Solution
A large sensor averages real values and one measured value is equal to
m
=
1
m x , y dx dy
A A
(4.7.1)
where A is the surface of the sensor. The whole measured distribution can be described using
the following convolution equation
m p= m x , yp x X , y Y dx dy
(4.7.2)
where p is a filter function. This filter function describes how the sensor averages real values.
To obtain a real distribution from a measured one, a convolution equation can also be
used. In this case, the filter function is an inverse function to the sensor filter function. The
convolution, inverse function computation and noise reduction can be done more easily in the
frequency domain than in the space domain. The measured data and sensor filter function are
transformed from the space domain into the frequency domain using FFT.
As we have values transformed into the frequency domain we can easily compute the
inverse sensor filter using
P1 f x , f y =
1
P f x , f y
(4.7.3)
Having the inverse filter we can do the convolution, Eq (3.5.6), in the frequency domain
using this inverse filter and measured data to obtain a real pressure distribution (still in the
frequency domain). In our case, the convolution is described by
f x , f y P1 f x , f y
M f x , f y = M
(4.7.4)
are the measured data, P1 is the inverse sensor function and M represents
where M
a sharpened data. Transforming sharpened data from the frequency domain into the space
domain using inverse FFT we obtain a pressure distribution which should be very close to the
real pressure distribution. However, some noise is visible in the sharpened data (small waves).
Noise reduction and aliasing phenomenon
Noise is most significant at high frequencies and because we are working in the
frequency domain, noise can be suppressed. The sensor function consists of the main
frequency spectrum (in the middle) and higher harmonic frequencies. Cutting off the higher
99
Problem Solution
harmonic frequencies and making them equal to zero also in the inverse sensor filter, we get
a cut inverse sensor filter. Using this filter for convolution (Eq. 4.7.4) instead of the inverse
filter and transforming the sharpened data using inverse FFT, we get sharpened measured data
with suppressed noise. However, the computed maximum is still lower than the real one. This
is due to the aliasing effect that is described in the previous Data Filtration chapter.
The sampled continuous function that is not bandwidth limited to less than the Nyquist
critical frequency results in an incorrect frequency spectrum. Any frequency component that
lies outside of the range f c f f c is spuriously moved into that range (this phenomenon
is called aliasing). This effect is more significant for the sensor function because of sharp
edges of the sensor.
The aliasing can be avoided by using a low pass filter. The sensor function passes
through the Gaussian low pass filter in the space domain. Transforming this function into the
frequency domain, we get high frequencies equal to zero. This means there is no aliasing
effect. Using the cut smooth inverse sensor filter for convolution and transforming the
sharpened data using inverse FFT, we get sharpened measured data with a maximum which is
very close to the real maximum.
100
101
Figure 5.2 Error after the third iteration. Figure 5.3 Error after the fourth iteration.
Figure 5.2 and 5.3 show the error distribution after the third and fourth iteration,
respectively (zero error is assumed after an infinitive or very high number of iterations). The
error was computed for each node of the 2D model after each iteration and the maximum error
evolution in time is shown in Figure 5.4. The chart also shows the estimated error range
computed using Eqs. (4.2.74.2.9) and the simple error estimation computed using Eq. (4.2.6).
It is obvious that the simple error estimation cannot be used for slowly converging models.
102
Temperature
Upper range
294.06
Temperature [C]
Lower range
Simple upper range
Simple lower range
294.04
294.02
294.00
293.98
293.96
293.94
0
10
20
30
Iteration within one time step
40
50
Figure 5.4 Maximum error and its estimation after each iteration.
Although 44 and 145 iterations were required to reach the 0.0001 K accuracy and final
solution, respectively, the used line-by-line method with tridiagonal systems of equations is
still much faster than the full system of equations solved by LU decomposition. The full
system has n x n y =n=82 82=6 724 equations and the solution requires 1/3n3 operations.
Moreover, the system is nonlinear. Thus, the solution should be iterated at least three times to
know the accuracy of the computed solution. The total number of operations is
304 006 671 424 at least. In the case of the line-by-line method, the 6 724 equations have to
be also solved. However, these equations are arranged in n x n y =82 82=164 sets of
tridiagonal
equations.
The
solution
of
these
nonlinear
equations
takes
only
decomposition. Also the memory requirements are n times smaller for the line-by-line method
than for LU decomposition.
5.1.2 Implementation of Phase Change and Heat Generation
The purpose of this computation is to show the ability to deal with the implementation
of phase change [A10] and heat generation in a computational model of the heat conduction
problem.
Two-dimensional model of a heater unit
Let us imagine a heater unit as shown in Figure 5.5. The unit consists of a closed tube
placed in an insulation. A heater is situated in the center of the tube and the rest of the space in
103
the tube is filled with a medium that will be melted. The problem is considered in micro
gravity to avoid heat conduction due to the convection of the melted medium. The heater unit
is at an initial temperature of 0C. The phase change of the melting medium occurs at 100C.
At the beginning, the heater is switched on. The adiabatic surface conditions are used between
the insulation and unit. The model computes only one quarter of the problem because of the
symmetry. Thus, the adiabatic surface conditions are also used on the symmetry lines.
3D Scheme
Heater
Computational Model
Melted medium
Heater
Sensor
Shield
Insulation
Temperature [C]
200
150
100
50
0
0
20
40
During the computational experiment, a temperature history was recorded for the melted
medium at a distance of 15 mm from the center of the rotation symmetry. The recorded history
is shown in Figure 5.6. First, the whole system was heated and the temperature was smoothly
increasing. As the temperature of the melted medium reached the melting temperature, a first
edge appeared at a time of 60 s. The medium was melting and the increase in temperature was
104
slower due to the latent heat. Figure 5.7 shows the temperature distribution at a time of 100 s
after the start of the heating. The sharp edge in the temperature history represents the
liquid / solid interface. When all medium was melted, which occurred at a time of 130 s after
the start of the heating, the temperature increased more rapidly.
A phase change occurs also in solid states, e. g. when carbon steel is heated or cooled.
Tests of a phase change and heat generation are important for rolling applications when
carbon steel is cooled down from high temperatures and where heat is generated during
deformation of rolled products.
5.1.3 Verification of Discretization Error Estimation
Boundary Conditions and Time Discretization Error
To verify the method for the estimation of the time discretization error using Eq. (4.3.7)
a one-dimensional model was used. The model was assumed to be made of stainless steel. It
was 10 mm long for all tests except for the one with a constant heat transfer coefficient where
its length was 100 mm. One side of the model was assumed to be insulated while the other
one exposed to various boundary conditions. The starting temperature of the model was 0C.
A constant heat flux of 100 000 W/m2 was applied to one boundary. The step duration
was 1 s and 10 s for the first and second experiment, respectively. The temperature profiles
computed using a very fine time step are represented in Figures 5.85.9 by the solid line with
crosses. These temperature profiles were compared with those computed using a single time
step of 1 s and 10 s duration for the first and second experiments, respectively. The absolute
values of computed differences are represented in Figures 5.85.9 by the thick solid black
line. To verify the use of Eq. (4.3.7) for the estimation of the time discretization error, the
temperature profiles were also computed using half time steps. The estimated errors were
computed for =1 ;1.5 ; 2 and absolute values are shown in Figures 5.85.9. It is obvious
that the estimated errors computed using =2 and 1.5 well cover the error caused by time
discretization for the half time step. The only place where the computed error is higher than
the estimated one is where the error is very low. However, in most cases the important value is
the maximum error, which is correct.
105
16
1.6
1.4
Estimated error
12
1.2
1.5*Estimated error
2.0*Estimated error
10
Temperature
14
0.8
6
0.6
0.4
0.2
Temperature [C]
1.8
0
0
4
5
Position [mm]
10
Figure 5.8 Temperature profile for a constant heat flux of 100 000 W/m2 after 1 s.
3.5
50
Estimated error
2.5
2
1.5*Estimated error
2.0*Estimated error
40
Temperature
30
1.5
20
Temperature [C]
60
10
0.5
0
0
0
4
5
Position [mm]
10
Figure 5.9 Temperature profile for a constant heat flux of 100 000 W/m2 after 10 s.
The estimation algorithm was also verified for time dependent boundary conditions. The
sinusoid, triangular heat fluxes, and constant heat transfer coefficient were used as shown in
Figures 5.10, 5.12, and 5.14. The charts shows real boundary conditions and discretized ones
for normal time and half time steps. The computed temperature profiles, time discretization
errors, and estimated errors are shown in Figures 5.11, 5.13, and 5.15. It is obvious that the
estimated errors computed using =2 and 1.5 also well cover the error caused by the time
discretization as in the previous computational experiments. Moreover, =1 can be used for
the constant heat transfer coefficient.
The computational experiments have proved that Eq. (4.3.7) can be used for error
estimation caused by time discretization. The only place for which the equation should not be
used are those where the error is small (e.g. 10%) in comparison with the maximum one.
106
150 000
Real
Normal time step
100 000
Heat flux [W/m]
10
Time [mm]
12
14
16
18
20
1.8
1.6
Estimated error
1.4
1.5*Estimated error
2.0*Estimated error
1.2
Temperature
0.8
0.6
0.4
0.2
Temperature [C]
-1
0
4
5
Position [mm]
10
Figure 5.11 Temperature profile for heat flux q = 100 000*sin(t) after 20 s.
120 000
Real
Normal time step
100 000
Heat flux [W/m]
10
15
20
Time [mm]
25
30
35
40
55
1.6
Estimated error
53
1.4
1.5*Estimated error
2.0*Estimated error
52
Temperature
50
1.2
1
54
51
0.8
49
0.6
48
0.4
47
0.2
46
Temperature [C]
2
1.8
45
0
4
5
Position [mm]
10
Real
Normal time step
100 000
Heat flux [W/m]
100
200
300
400
500
Time [mm]
600
700
800
900
1000
Figure 5.14 Heat flux for constant heat transfer coefficient h = 1000 W/m2.K.
Error for normal time step
Error for half time step
1.2
1
0.8
100
90
Estimated error
80
1.5*Estimated error
2.0*Estimated error
70
Temperature
50
60
0.6
40
0.4
30
Temperature [C]
1.4
20
0.2
10
0
0
10
20
30
40
50
Position [mm]
60
70
80
90
0
100
Figure 5.15 Temperature profile for constant heat transfer coefficient h = 1000 W/m2.K
after 1000 s.
108
Real error
Max. estimated error
2.5
Temperature
2
45
40
35
30
25
1.5
20
15
Temperature [C]
10
0.5
0
0
10
20
30
40
50
Position [mm]
60
70
80
90
0
100
109
Real error
Max. estimated error
1.6
Temperature
140
120
1.4
100
1.2
80
1
0.8
60
0.6
40
Temperature [C]
2
1.8
0.4
20
0.2
0
0
10
20
30
40
50
Position [mm]
60
70
80
90
0
100
0.8
Temperature
500
450
400
0.7
350
0.6
300
0.5
250
0.4
200
0.3
150
0.2
100
0.1
50
0
0
10
20
30
40
50
Position [mm]
60
70
80
90
Temperature [C]
1
0.9
0
100
Real error
Max. estimated error
Temperature
3.5
160
140
120
100
3
2.5
80
60
1.5
40
Temperature [C]
5
4.5
20
0.5
0
0
10
20
30
40
50
Position [mm]
60
70
80
90
0
100
110
Real error
Max. estimated error
10
Temperature
1000
900
800
700
600
6
500
400
300
Temperature [C]
12
200
100
0
0
10
20
30
40
50
Position [mm]
60
70
80
90
0
100
111
where the slope is mild, see Figure 5.24. The stronger filter works better in these places but it
fails in places where the temperature history changes the slope rapidly as shown in
Figures 5.225.23. The proposed multi-level filtration algorithm works well in mild slopes as
well as in places where the temperature changes the slope rapidly.
925
Temperature [C]
Measured
920
915
Multi-level filtration
Gaussian s=0.03
Gaussian s=0.06
910
905
900
0
0.5
1.5
Time [s]
2.5
920
Measured
Low pass f=50
Low pass f=200
Multi-level filtration
Gaussian s=0.03
Temperature [C]
919.5
919
Gaussian s=0.06
918.5
918
917.5
917
0.8
0.85
0.9
0.95
1
Time [s]
1.05
1.1
1.15
1.2
112
860
Measured
Low pass f=1000
Multi-level filtration
859.5
Gaussian s=0.01
Gaussian s=0.02
Temperature [C]
859
858.5
858
857.5
857
11.3
11.35
11.4
11.45
Time [s]
11.5
11.55
11.6
859
Temperature [C]
858.5
858
857.5
Measured
Low pass f=1000
Multi-level filtration
Gaussian s=0.01
Gaussian s=0.02
857
856.5
9
9.1
9.2
9.3
9.4
9.5
9.6
9.7
9.8
9.9
10
Time [s]
113
sequential
a modified
downhill
simplex
optimization method and feed-forward neural network, are used for single boundary
conditions determination. Next, the optimization method is used in IHCPs where twodimensional boundary conditions are computed, temperature dependent material properties are
determined, and the computational model is calibrated.
5.3.1 Boundary Conditions Computation
In this chapter, the classical Beck's sequential approach, modified downhill simplex
optimization method and feed-forward neural network are used for the solution of IHCPs
where single boundary conditions are determined for unsteady problems.
Beck's approach using filtered data
The cooling intensity of a water nozzle was studied on a linear test bench (see
Figure 5.25 and Figure 16.1 in Appendix G). During the experiment a heated stainless steel
plate was moving repeatedly under a spraying nozzle, and the measured temperature inside the
steel plate was recorded using measuring frequencies of 100 Hz and 300 Hz (see
Figures 5.215.24). The internal structure of the temperature sensor (see Figure 16.3) in the
steel plate is shown in Figure 4.26 and Figure 16.5 in Appendix G.
Figure 5.25 Principal scheme of the linear test bench 1. cooling medium supply, 2.
pressure gauge, 3. nozzle, 4. moving deflector, 5. manifold, 6. test plate, 7. moving trolley,
8. data logger, 9. roller, 10. electric motor, 11. hauling steel wire rope, 12. girder.
114
Figure 5.26 Temperature profile of the optimized 2D model after 1s with applied heat flux
q = 100 000 W/m2 on a surface.
Before the IHCP was solved, the mesh of the 2D computational model (see Figure 4.28)
was optimized (see Figure 5.26). The estimated discretization error fields for heat conduction
in X and Y directions are shown in Figure 5.27.
The time dependent surface heat flux was computed from the measured temperature
histories using classical Becks sequential approach. The input data (temperature histories)
were used in several forms: raw data, data filtered using the Gaussian filter, and data filtered
using proposed multi-level filtration. The computed boundary conditions are shown in
Figures 5.28 and 5.29 for measuring frequencies of 100 Hz and 300 Hz, respectively. The
results computed using raw measured data are very noisy. The noise and measuring resolution
step were strongly magnified by the IHCP algorithm. The situation is getting better for data
filtered using the Gaussian filter in the time domain. The improvement is more significant for
0.008184
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
Error [C]
Figure 5.27 Estimated errors in X (left chart) and Y (right chart) directions caused by
space discretization after mesh optimization.
115
a measuring frequency of 300 Hz. However, the maximum peek is lower in comparison with
the raw and multi-level filtered data. In case of the 100 Hz, the problem with temperature
resolution step (see Figure 5.24) was not suppressed. Thus the computed results are still
oscillating for low heat fluxes (Figure 5.28). The results computed from the data filtered using
multi-level filtration are very good for low heat fluxes as well as for peaks. The oscillation of
the heat flux was well suppressed for low values and the peak was not smoothed. The
maximum and width of the peak is comparable with that computed using raw data.
Solution of the IHCP can be well improved when a sophisticated filtration is used for
the input data. An improvement is described in next chapter, where an alternative optimization
method instead of sequential Beck's approach was used for IHCP.
3500000
Measured
Gaussian s=0.01
3000000
Multi-level filtration
Heat flux [W/m]
2500000
2000000
1500000
1000000
500000
0
-500000
0
3500000
10
15
Time [s]
20
25
30
Measured
Gaussian s=0.01
3000000
Multi-level filtration
Heat flux [W/m]
2500000
2000000
1500000
1000000
500000
0
-500000
7.5
7.7
7.9
8.1
8.3
8.5
Time [s]
8.7
8.9
9.1
9.3
9.5
Figure 5.28 Heat flux history computed using 8 forward time steps
from data measured at 100 Hz.
116
Measured
Gaussian s=0.03
Multi-level filtration
700000
600000
500000
400000
300000
200000
100000
0
-100000
0
6
Time [s]
10
800000
12
Measured
Gaussian s=0.03
700000
Multi-level filtration
600000
500000
400000
300000
200000
2.1
2.2
2.3
2.4
2.5
Time [s]
2.6
2.7
200000
2.8
2.9
Measured
Gaussian s=0.03
150000
Multi-level filtration
100000
50000
0
-50000
-100000
5
5.2
5.4
5.6
5.8
6
Time [s]
6.2
6.4
6.6
6.8
Figure 5.29 Heat flux history computed using 25 forward time steps
from data measured at 300 Hz.
117
HTC inv. 5
HTC inv. 15
HTC [W/m.K]
HTC opt.
tim e [s]
The real measured data were taken to compare classical Becks sequential approach with
the combined approach. Three different computations were made (see Figure 5.30). The first
and second computation used classical Becks approach with 5 and 15 forward steps,
respectively. The third one combined the classical Becks approach with an optimization
method. The classical approach with 5 forward steps matched the measured temperature
history almost perfectly. The RMS error, defined as:
1
RMS =
n
m
2
T* mT computed
(5.3.1)
m=1
The feed-forward neural network (see Figure 4.22) was trained using fast
back-propagation (Levenberg-Marquardt functions). The and parameters were not
computed using the neural network but the position was taken from position information,
and was computed using Becks approach from the data far from the peak. The neural
network reached the sum-squared error goal (0.002) usually in less than 50 training epochs.
Computation using the learned neural networks has proved that not all of them can be used.
Some of them returned confused results. A computation experiment has shown that the
learned neural networks should fulfill the condition that the maximal absolute value of the
element of the matrixes W 1, b1, W 2, and b 2 is less than 1.7 (Eq. 5.3.2). The neural network
with the lowest maximum, within the group with equal sum-squared error goals, represented
the best one. In that case, the maximal value was 1.19.
max W 1, max , W 2, max , b1, max , b 2, max 1.7
W 1, max =maxw m , i;W 1=w m , i for m=1, .. n ; i=1, ..3
m ,i
(5.3.2)
Training of the neural network is a quite time consuming operation and the obtained
results are not very accurate. The trained network often depends on the random initial
matrixes W 1, b1, W 2, and b 2 . The computed results must be verified because some networks
give confused results for data not included in the training set.
5.3.2 Calibration of Computational Model of Measuring Sensor
In experiments described in the previous chapters, a special sensor (see Figure 4.26 and
Figure 16.3 in Appendix G) was used for measuring temperature inside the steel plate. As
mentioned in the Sensor Calibration chapter, the sensor geometry may vary and the
methodology of the computational model calibration was described in that chapter.
During the experiment, the sensor held at room temperature was exposed to a high
temperature (see Figure 16.4 for experimental apparatus). The sensor was stuck to the heated
120
copper rod. The recorded temperature histories are shown in Figure 5.31. The sensor was held
at room temperature and the copper was heated to a uniform temperature at the beginning. As
the sensor was stuck to the heated copper rod, the temperature of the high-conductive grease,
which was spread on the copper rod, suddenly dropped down and the sensor temperature
started to increase.
90
80
Sensor
70
T [C]
Copper1
60
Copper2
50
Grease
40
30
20
0
t [s]
Measured
55
50
T [C]
45
40
Inceasing
grease
thickness
35
30
25
20
0
0.5
1
t [s]
1.5
Figure 5.32 Perfectly overlapping measured and computed temperature histories for three
different thicknesses of high conductive grease and the same computational model of the
sensor (0.05 mm, 0.11 mm, and 0.17 mm)
121
Two computational experiments were made to show the importance of the usage of the
calibrated model (see Figure 4.28). The measured temperature history inside the sensor
(Tmeasured) and the computed temperature history with the calibrated model (Ttherm.) are shown in
Figure 5.33. Both curves are almost identical. Moreover, this figure shows the computed
temperature history for the model with no internal structure homogeneous steel (Tsteel). These
values were computed using the same boundary conditions as the ones computed with the
calibrated model. It is obvious that the response of such a sensor would be somewhat faster. In
addition, HTC were computed (see Figure 5.33) on the heated side (see Figure 4.28) of the
sensor during the experiment. The HTC were computed using both the calibrated model and
the homogeneous steel model, HTCtherm. and HTCsteel, respectively, from real measured
temperature history. The heated copper and grease were supposed as the surrounding medium.
The ambient temperature was the temperature of the copper. It is obvious that the computation
with the homogeneous steel model failed at the beginning because there is no reason why
HTC should be the lowest. HTC must be the highest at the beginning because the temperature
gradient in the heated copper rod is the lowest. Therefore, the heated copper rod is the most
intensive source of heat. As the temperature gradient increases in the heated copper, the HTC
decreases because the colder part of the copper is becoming a temperature resistor through
60
16000
55
14000
50
12000
45
10000
40
T measured
8000
35
T therm.
6000
30
T steel
4000
HTC therm.
25
HTC [W/m.K]
T [C]
2000
HTC steel
20
0
0
0.2
0.4
0.6
0.8
1.2
1.4
1.6
1.8
t [s]
Figure 5.33 Comparison of the results computed using two models one made of
homogeneous steel and a calibrated one with a built in thermocouple.
122
Al
Space
T1
mold
T2 T3
bath
Too
123
process lasted the first 2 minutes. The sharp rise of HTC indicates that the mold was
immersed into the molten bath 20 seconds after the start of the recording.
Figure 5.35 Temperature profile of model consisting of alloy and mold during casting.
After direct computation, the inverse task was made using a modified downhill simplex
and sequential Beck's approach. These two approaches are compared in Tables 5.1 and 5.2
where the time steps are 1 s and 0.1 s, respectively. A minimum of the required forward steps
is the number when the inverse tasks give us realistic results. The RMS error is defined as
RMS HTC =
1
n HTC *i mHTC im 2
2 i=1 m=1
(5.3.3)
where HTC *i m are the original heat transfer coefficients on both sides of the mold in the mth
time step that was used in direct computation, and HTC im are the heat transfer coefficients
computed using the inverse task. The Downhill Simplex optimization method was faster and
gives better results than Becks approach. This is because the problem under solution is
nonlinear, and Becks approach uses linear approximation. Comparing the results for time
steps 1 s and 0.1 s, it is obvious that the shorter time step gives better results. This is because
there is a fast change in boundary conditions when the mold is immersed into the tin bath. The
reason for why a shorter forward time is necessary for a smaller time step is that there are still
124
more forward time steps, and the rounding errors of the numerical part with several iterations
are averaged.
3000
2500
2000
1500
T [C]
600
500
1000
500
0
300
400
300
0
50
100
150
200
time [s]
250
T1
[W/m.K]
700
T2
T3
TCR
HTC
Table 5.1 Comparison of downhill simplex and Becks approach for the 1 s time step.
Downhill simplex
Becks approach
15 s
57 s
2
31 W/m .K
123 W/m2.K
67 s
123 s
2
161 W/m .K
180 W/m2.K
Table 5.2 Comparison of downhill simplex and Becks approach for the 0.1 s time step.
Downhill simplex
Becks approach
14
40
20 s
324 s
8.2 W/m2.K
21.6 W/m2.K
58 s
324 s
20.3 W/m2.K
21.6 W/m2.K
125
800
Unsteady conditions
2000
750
T1
1500
700
1000
650
500
600
0
20
40
60
80
P [W]
T [K]
2500
T2
T3
T4
Power
0
100
t [s]
Figure 5.37 Measured temperature histories and power for steady and
unsteady conditions.
126
1.00
k1 [W/m.K]
0.80
0.60
0.40
0.20
0.00
300
400
500
600
700
800
900
1000
T [K]
1.5
800
1
600
400
0.5
Cp [J/kg.K]
k [W/m.K]
k1
k2
Cp1
Cp2
200
0
0
300 400 500 600 700 800 900
T [K]
127
Having reduced the number of parameters, four values for k1, four values for k2, and two
values for c1 were computed. It has been found that the measured data do not contain enough
information to compute c2 as temperature-dependent. Hence specific heat c2 is assumed to be
constant. The results are shown in Figure 5.39. There are no values for k2 and c1 related to the
300 K temperature because any value gives the same value of the criteria equation. This
means that no information is involved in the measured data for this temperature. Comparing
the results computed only for steady conditions with the results obtained for unsteady
conditions, it is obvious that conductivity k1 is higher for steady conditions. This is due to the
improper assumption that the heat lost at the top and bottom of the mold can be neglected.
Comparing the results, the heat loss amounted to about 20 %.
Comparing the computed conductivity k1 of the backup mold with the values from
MAGMA database, the results differ by about 5% only. This can be due to the different
density of the mold. The thermal conductivity is also slightly increasing for a temperature
range of 600900 K. The computed mass density is by about 3.5 % lower than that in the
MAGMA database. The computed specific heat is by about 5 % lower than that in the
MAGMA database. This is probably due to the lower mass density of our mold. Looking at
the rapidly decreasing thermal conductivity k2, a similar effect can be seen for the Al2O3 in
database or for sapphire and polycrystalline aluminum oxide.
measured data and the inverse sensor. The computed result is shown in Figure 5.42b in the
frequency domain and in Figure 5.43a in the space domain. The computed pressure
distribution is close to the real pressure distribution, however, some noise is visible in the
sharpened data (small waves).
Noise can be partially suppressed as was described in the Data Focusing chapter. Cutting
off the higher harmonic frequencies and making them equal to zero also in the inverse sensor
filter, we get a cut inverse sensor filter (Figure 5.45a). Using this filter for convolution we
obtain sharpened measured data with suppressed noise. To suppress the aliasing phenomenon,
the measured data and the sensor were filtered using Gaussian low-pass filter in the space
domain. Transforming these functions into the frequency domain (see Figure 5.44), we get
high frequencies equal to zero. This means that the aliasing phenomenon was suppressed.
Using the cut smooth inverse sensor filter for convolution, we get sharpened measured data
(see Figure 5.45b) with a maximum of 98.9 MPa which is very close to the real maximum,
100 MPa. You can also notice that the noise was well suppressed, compared with the
computed result shown in Figure 5.43a.
Real measured data are shown in Figure 5.46a. The measurement was made for a high
pressure flat jet nozzle where the distance from the surface was 150 mm and water pressure
was 20 MPa. The measurement was made using a square sensor (3x3 mm). Some noise in
measured data is obvious. The maximum in the sharpened data (see Figure 5.46b) rose by
about 18 % (from 1534 to 1814 kPa).
Another measurement was made using a different measuring apparatus. A circular
sensor of 1.5 mm in diameter was used. Two overlapping flat nozzles were measured. The
measured data are almost noise free compared to the previous data measured using
a rectangular sensor (see Figure 5.47a). The measuring area was 30x160 mm. The maximum
of the sharpened data rose from 1527 to 1587 kPa (see Figure 5.47b). The increase is small
because a small pressure sensor was used.
129
(a)
(b)
Figure 5.40 An example of (a) simulated measured distribution and
(b) simulated real distribution using a circular sensor of 12 mm in diameter.
(a)
(b)
Figure 5.41 (a) Data measured in frequency domain;
(b) circular sensor in frequency domain.
(a)
(b)
Figure 5.42 (a) Inverse sensor function in frequency domain;
(b) convolution of measured data and inverse sensor function in frequency domain.
130
(a)
(b)
Figure 5.43 (a) sharpened measured data convolution of measured data and inverse
sensor function in space domain; (b) real pressure distribution in frequency domain.
(a)
(b)
Figure 5.44 (a) Circular sensor function in space domain filtered using low-pass filter;
(b) circular sensor function in frequency domain with removed high frequency.
(a)
(b)
Figure 5.45 (a) cut inverse sensor function in frequency domain;
(b) sharpened measured data.
131
(a)
(b)
Figure 5.46 (a) Real data measured using rectangular sensor (3x3 mm);
(b) sharpened data.
(a)
(b)
Figure 5.47 (a) Real data measured using circular sensor (1.5 mm in diameter);
(b) sharpened data
132
133
Conclusion
7 CONCLUSION
Many of inverse problems are often dismissed as ill-conditioned because the solution is
extremely difficult or almost impossible using concurrent methods. If any results are obtained,
they are inaccurate and the precision is unknown. Typical inverse problems studied in this
work include determination of boundary conditions in applications such as cooling in casting,
descaling, cooling of products in hot rolling, and cooling of rolls in rolling technology.
Because the studied inverse methods use direct computation of heat conduction, several
direct methods were compared. The suitable ones were finite difference method (FDM), finite
volume method (FVM), finite element method (FEM), and finite volume unstructured
methods (cell-centered and vertex-centered). The unstructured methods were not used in
inverse tasks because the demand of the inverse tasks for the computational power is high and
these methods are very slow. The FDM was found to be most convenient when testing the
FDM, FVM, and FEM for a direct solver of heat conduction. The speed was several times
higher when temperature dependent material properties were used and the accuracy of
computed results in nodes was the best for fast changes in boundary conditions. The general
FDM method was modified to treat phase change and temperature dependent material
properties.
Attention was also focused on accuracy of results computed by the direct heat
conduction method. Convergence of the iteration method for numerical solution of heat
transfer partial differential equations was tested. The convergence error was estimated using
a method based on the observed geometric convergence of heat conduction problems.
Methods for discretization error estimation were proposed and tested for the simplification of
the object, its boundary conditions, and time domain when the numerical approach is applied.
These methods were then used for discretization optimization.
New approaches and a special multi-level filtration for input data were developed for
solution of inverse heat conduction problems (IHCPs). The new approaches (optimization
method and neural network) were compared with classical Becks sequential approach. Usage
of the modified downhill simplex multidimensional optimization method was tested in
one-dimensional and several multi-dimensional inverse problems where the input data were
taken from real measurement. These inverse problems included determination of time
134
Conclusion
135
Literature
8 LITERATURE
[1]
Incropera, F. P.; DeWitt, D. P. Fundamentals of Heat and Mass Transfer. 4th ed. New
York: Wiley, 1996. ISBN 0-471-30460-3.
[2]
Kakac, S.; Yener, Y. Heat Conduction. New York: Hemisphere Publishing, 1985.
[3]
[4]
[5]
William, H. P.; Saul A. T.; William, T. V.; Brian, P. F. Numerical Recipes in C. 2nd ed.
1997. ISBN 0-521-43108-5.
[6]
[7]
[8]
[9]
Fletcher, C. A. Jr. Computational techniques for fluid dynamics. Berlin: Springer, 1991.
ISBN 0-387-18151-2.
[10] Reddy, J. N.; Gartling, D. K. The finite element method in heat transfer and fluid
dynamics. Boca Raton, FL: CRC Press, 1994. ISBN 0-849-39410-4.
[11] Beck, J. V.; Blackwell, B.; Charles, R. C. Inverse Heat Conduction: Ill-posed Problems.
New York: Wiley, 1985. ISBN 0-471-08319-4.
[12] Duda, P.; Taler, J Numerical method for the solution of non-linear two-dimensional
inverse heat conduction problem using unstructured meshes. Int. J. for Numerical
Methods in Engineering, 2000, Vol. 48, pp. 881-899.
[13] Chow, P.; Cross, M. An enthalpy Control-Volume-Unstructured-Mesh (CV-UM)
algorithm for solidification by conduction only. Int. J. for Numerical Methods in
Engineering, 1992, Vol. 35, pp. 1849-1870.
[14] Power, H.; Worbel, L. Boundary integral methods in fluid mechanics. UK: WIT Press
(Computational Mechanics Publications), 1995.
136
Literature
[15] Lesnic, D.; Elliott, L.; Ingham, D. B. Application of the boundary element method to
inverse heat conduction problems. Int. J. Heat Mass Transfer, 1996, Vol. 39, No. 7, pp.
1503-1517.
[16] Han, H.; Ingham, D. B.; Yuan, Y. The boundary elementh method for the solution of the
backward heat conduction equation. J. of Computational Physics, 1995, Vol. 116, pp.
292-299.
[17] Ochiai, Y.; Sekiya, T. Steady heat conduction analysis by improved multiple-reciprocity
boundary element method. Eng. Analysis with Boundary Elements, 1996, Vol. 18, pp.
111-117.
[18] Ochiai, Y. Two-dimensional unsteady heat conduction analysis with heat generation by
triple-reciprocity BEM. Int. J. for Numerical Methods in Engineering, 2001, Vol. 51, pp.
143-157.
[19] Johns, P. B.; Beurle, R. L. Numerical solution of two dimensional scattering problems
using a transmition line matrix. Proc. IEE, 1971, Vol. 119, pp. 1203-1208.
[20] Cogan, D.; Soulos, A. Inverse thermal modeling using TLM. Num. Heat Transfer, 1996,
Vol. 29, pp. 125-135.
[21] Voller, V. R.; Swaminathan, C. R. Fixed grid techniques for phase change problems: A
review. Int. J. for Numerical Methods in Engineering, 1990, Vol. 30, pp. 875-898.
[22] Song, R.; Dhatt, G.; Cheikh, A. B. Thermo-mechanical finite element model of casting
systems. Int. J. for Numerical Methods in Engineering, 1990, Vol. 30, pp. 579-599.
[23] Comini, G.; Giudice, S. D.; Saro, O. A conservative alghorithm for multidimensional
conduction phase change. Int. J. for Numerical Methods in Engineering, 1990, Vol. 30,
pp. 697-709.
[24] Patankar, S. V. Computation of Conduction and Duct Flow Heat Transfer. Innovative
Research, Inc., 1991.
[25] Tikhonov, A. N.; Arsenin, V. Y. Solution of Ill-Posed Problems. Washington, D.C.:
Winston, 1977. ISBN 0470991240.
[26] Mandrel, J. Use of the singular value decomposition in regression analysis. Am. Stat.,
1982, Vol. 36, pp. 15-24.
[27] Stloz, G. Jr. Numerical solutions to an inverse problem of heat conduction for simple
shapes. Int. J. Heat Transfer, 1960, Vol. 82, pp. 20-26.
[28] Beck, J. V. Nonlinear estimation applied to the nonlinear heat conduction problem. Int.
J. Heat and Mass Transfer, 1970, Vol. 13, pp. 703-716.
137
Literature
[29] Beck. J. V.; Litkouhi, B.; St. Clair, C. R. Jr. Efficient sequential solution of the nonlinear
inverse heat conduction problem. J. Numerical Heat Transfer, 1982, Vol. 5, pp. 275-286.
[30] Bass, B. R. Applications of the finite element to the inverse heat conduction problem
using Becks second method. J. Eng. Ind., 1980, Vol. 102, pp. 168-176.
[31] Burggraf, O. R. An exact solution of the inverse problem in heat conduction theory and
applications. Int. J. Heat Transfer, 1964, Vol. 86C, pp. 373-382.
[32] Imber, M.; Khan, J. Prediction of transient temperature distributions with embedded
thermocouples. Al AA J., 1972, Vol. 10, pp. 784-789.
[33] Lengford, D. New analytic solutions of the one-dimensional heat equation for
temperature and heat flow rate both prescribed at the same fixed boundary (with
applications to the phase change problem). Q. Appl. Math., 1967, Vol. 24, pp. 315-322.
[34] Grysa, K.; Cialkowski, M. J.; Kaminski, H. An inverse temperature filed problem of the
theory of thermal stresses. Nucl. Eng. Des., 1981, Vol. 64, pp. 169-184.
[35] Slma, L.; Raudensk, M.; Horsk, J.; Bezina, T.; Krejsa, J. Evaluation of quenching
test of rotating roll with unknown time constant of sensor using genetic algorithm. Int.
Conf. Mendel, Brno, 1996.
[36] Dumek, V.; Grove, T.; Raudensk, M.; Krejsa, J. Novel approaches to the IHCP: Neural
networks. In Int. symposium on inverse problems - Inverse problems in Engineering
Mechanics, Paris, 1994, pp. 411-416.
[37] Krejsa, J.; Slma, L.; Horsk, J.; Raudensk, M.; Ptikov, B. The comparison of
traditional and non-classical methods solving the inverse heat conduction problem. In
Int. Conf. Advanced Computational Methods in Heat Transfer, Udine, July 1996, pp.
451-460.
[38] Beck, J. V.; Murio, D. A. Combined function specification regularization procedure for
solution of inverse heat conduction problem. AIAA J., 1986, Vol. 24, pp. 180-185.
[39] D'Souza, N. Numerical solution of one-dimensional inverse transient heat conduction
by finite difference method. ASME Paper No. 75-WA/HT-81, 1975.
[40] Raynaud, M., Bransier, J. A new finite difference method for nonlinear inverse heat
conduction problem. Num. heat Transfer, 1986, Vol. 9, No. 1, pp. 27-42.
[41] Weber, C. F. Analysis and solution of the ill-posed inverse heat conduction problem. Int.
J. Heat Mass Transfer, 1981, Vol. 24, No. 11, pp. 1783-1792.
[42] Woodbury, K. A.; Ke, Q. An inverse algorithm for direct determiniation of heat transfer
coefficients. In 34th National Heat Transfer Conference, 2000, pp. 663-669.
138
Literature
139
Literature
[54] Loulou, T.; Artyukhin, E. A., Bardon, J. P. Estimation of thermal contact resistance
during the first stages of metal solidification process: I-experiment principle and
modelisation. Int. J. Heat Mass Transfer, 1999, Vol. 42, pp. 2119-2127.
[55] Demuth, H.; Beale, B. Neural Network Toolbox User's Guide. Ver. 3 MathWork Inc.,
1998.
140
Authors References
9 AUTHORS REFERENCES
[A1] Pohanka, M.; Raudensk, M.; Horsk, J.; Druckmller, M. How to Precisely Define
Computational Models of Heat Process with Experimental Setting under Marginal
Conditions (in Czech). In Inenrsk mechanika 99. Svratka (Czech Republic), 1999,
pp. 717722. ISBN 80-214-1323-9.
[A2] Pohanka, M.; Raudensk, M.; Horsk, J. Attainment of more precise parameters of
a mathematical model for cooling flat and cylindrical hot surfaces by nozzles. In
Advanced computational methods in heat transfer VI. Madrid: WIT Press, 2000, pp.
627635. ISBN 1-85312-818-X.
[A3] Pohanka, M.; Raudensk, M.; Horsk, J. Optimizing parameters of a mathematical
model for cooling hot surfaces by nozzles. In Engineering mechanics 2000. Svratka
(Czech Republic), 2000, pp. 267272. ISBN 80-86246-03-5.
[A4] Raudensk, M.; Horsk, J.; Pohanka, M. Optimal cooling of rolls in hot rolling.
Journal of material processing technology. 2002, Vol. 125126, pp. 700705.
[A5] Pohanka, M.; Woodbury, K. A.; Wolley, J. Obtaining temperature dependent thermal
properties of investment casting mold. In Proc. of the International Mechanical
Engineering Congress and Exposition. New Orleans (LA): ASME, November 2002,
2-20-5-1.
[A6] Raudensk, M.; Pohanka, M.; Horsk, J. Combined inverse heat conduction method
for highly transient processes. In Advanced computational methods in heat transfer VII,
Halkidiki: WIT Press, 2002, pp. 3542. ISBN 1-85312-9062.
[A7] Toovsk, J.; Pohanka, M.; Kotrbek, P. Stress analysis of scale layer for descaling
process. In 40th int. conf. experimental stress analysis. Praha, 2002, pp. 247252.
ISBN 80-01-02547-0.
[A8] Pohanka, M.; Raudensk, M. Determination of heat resistances between installed
thermocouple and body used for computing heat transfer coefficients. In Engineering
mechanics 2002, Svratka (Czech Republic), 2002, pp. 227228. ISBN 80-214-2109-6.
[A9] Pohanka, M.; Woodbury, K. A. A Downhill Simplex method for computation of
interfacial heat transfer coefficients in alloy casting. Inverse problems in engineering,
October 2003, Vol. 11, No. 5, pp. 409424.
141
Authors References
[A10] Raudensk, M.; Horsk, J.; Pohanka, M.; et al. Experimental Study of Parameters
Influencing Efficiency of Hydraulic Descaling. In 4th Int. Conf. Hydraulic Descaling.
London, 2003, pp. 2939.
[A11] Pohanka, M. Two-dimensional correction of data measured using a large pressure
sensor. In Computational methods and experimental measurements XI. Halkidiki: WIT
Press, 2003, pp. 587596. ISBN 1-85312-969-0.
[A12] Kotrbek, P.; Horsk, J.; Raudensk, M.; Pohanka, M. Influence of parameters of
hydraulic descaling on temperature losses and surface quality of rolled material. In
Metal Forming 2004. Krakw (Poland), 2004, pp. 367370. ISBN 3-937057-08-0.
[A13] Horsk, J.; Raudensk, M.; Pohanka, M. Experimental study of heat transfer in hot
rolling and continuous casting. In Material Science Forum. Switzerland: Trans Tech
Publication, 2005, Vols. 473474, pp. 347354. ISBN 0-87849-957-1.
[A14] Kotrbek, P.; Horsk, J.; Raudensk, M.; Pohanka, M. Experimental study of heat
transfer in hot rolling. In 26e Journes sidrurgiques internationales. Paris, 2005, pp.
4243. ISBN 2-911212-05-3.
142
dT
.
dx
(10.1.1)
T
T1
T(x)
qx
T2
L
(10.1.2)
T 2 T 1
T
.
=k
L
L
(10.1.3)
This equation provides a heat flux the rate of heat transfer per unit area. The heat transfer
rate by conduction, q x [W], through a plane wall of area A is then the product of the flux and
the area
q x =q xA .
(10.1.4)
q x
dx ,
x
(10.2.1)
q ydy = q y
q y
dy ,
y
(10.2.2)
q zdz = q z
q z
dz .
z
(10.2.3)
II
q ydy
q zdz
dz
E g , E st
q x
q xdx
dy
dx
q y
q z
Figure 10.2 Differential control volume for conduction analysis in Cartesian coordinates.
Within the medium there may also be an energy source term associated with the rate of
thermal energy generation. This term is expressed as
g = qdxdydz
(10.2.4)
where q is the rate at which energy is generated per unit volume of the medium. In addition,
there may occur changes in the amount of the internal thermal energy stored by the material in
the control volume. The energy storage may be expressed as
st =c p
T
dxdydz
t
(10.2.5)
where c p T / t is the time rate of the change in thermal energy of the medium per unit
volume.
On a rate basis, the general form of the conservation of the energy requirement is
in out g = st
(10.2.6)
p
x
y
z
t
(10.2.7)
The conduction heat rates may be evaluated from Fouriers law, Eq. (10.1.1),
III
q x =kdydz
T
,
x
(10.2.8)
q y =kdxdz
T
,
y
(10.2.9)
q z =kdxdy
T
z .
(10.2.10)
Substituting Eq. (10.1.410.1.3) into Eq. (10.1.2) and dividing out the dimensions of the
control volume dxdydz , we obtain
T
T
k
k
q=c
.
p
x
x
y y
z
z
t
(10.2.11)
This equation is the general form, in Cartesian coordinates, of the heat diffusion equation also
known as the heat equation. In the cylindrical coordinates, the heat diffusion equation [1] is
1
T
1
T
T
T
kr
2
k
k
q=c
.
p
r r
r
z
t
r z
(10.2.12)
IV
Appendix B Convection
11 APPENDIX B CONVECTION
The convection boundary condition is one of the most common cases in our problems
under solution and thus we will discuss it a bit more in detail. The convection heat transfer
mode needs some fluid and is comprised of two mechanisms. In addition to the energy
transfer due to the random molecular motion (diffusion), energy is also transferred by a bulk,
or macroscopic, motion of the fluid. Such a motion contributes to the heat transfer and the
total heat transfer is a superposition of energy transport by the random motion of molecules
and by the bulk motion of the fluid.
We speak of forced convection when the flow is caused by external forces, such as
spraying water or atmospheric winds. In contrast, for free (or natural) convection the flow is
induced by buoyancy forces that arise from density differences caused by temperature
variations in the fluid. In many cases, mixed forced and natural convection may exist.
However, there are convection processes for which there is a latent heat exchange. This latent
heat exchange is associated with the phase change e.g. between the liquid and vapor states of
the fluid (boiling and condensation).
Regardless of the particular nature of the convection heat transfer process, the
appropriate rate equation has the form
q=h T s T
(11.1)
where q is the convective heat flux. This expression is known as Newtons law of cooling, and
the proportionality constant h is termed the convection heat transfer coefficient (HTC). For
typical problems, typical values of HTC are given in Table 11.1.
Table 11.1 Typical values of the convection heat transfer coefficient [1]
Process
Free convection
Gases
Liquids
Forced convection
Gases
Liquids
Convection with phase change
Boiling or condensation
h [W/m2.K]
225
501000
25250
5020 000
2 500100 000
Appendix C Radiation
12 APPENDIX C RADIATION
The heat transfer due to the radiation becomes important for hot surfaces such as steel
plates in continues casting or descaling applications. Thermal radiation is the energy emitted
by the matter that is at a finite temperature [1]. The energy of the radiation field is transported
by electromagnetic waves (or, alternatively, photons) and the presence of material medium is
not required. The rate at which the energy is released per unit area is termed the surface
emissive power E. There is an upper limit to the emissive power that is prescribed by the
Stefan-Boltzmann law
E b = T 4s
(12.1)
where Ts is the absolute temperature [K] of the surface and is Stefan-Boltzmann constant.
Such a surface is called an ideal radiation or blackbody. The heat flux emitted by a real surface
is less than that of a black body and is given by
E b = T 4s
(12.2)
may be within the range 01. The radiation may be reflected or, in the case of
a semitransparent material, it may be transmitted. The irradiation may be approximated by
emission from a blackbody at Tsur (surrounding temperature), in which case G=T 4sur. The
net rate of radiation heat transfer from the surface is
q
q rad = =E b T s G= T 4s T 4sur .
A
(12.3)
There are many applications for which it is convenient to express the net radiation heat
exchange in the form
q rad =h A T s T sur
(12.4)
VI
Appendix C Radiation
h = T s T sur T s T sur .
2
(12.5)
The surfaces may also simultaneously transfer heat by convection to an adjoining gas. For
example, a 5 m long slab heated to a temperature 800oC moving at a rate of 5 m/s at room
temperature gives
q=q conv q rad = h T s T T s T sur
4
(12.6)
VII
13.2 LU decomposition
This method uses the fact that the matrix A can be written as a product of two matrices
LU = A
(13.2.1)
where L is the lower triangular (has elements only on the diagonal and below) and U is the
upper triangular (it has elements only on the diagonal and above). The decomposition is used
to solve the linear equations
Ax= LU x=LUx =b
(13.2.2)
Ly=b
(13.2.3)
Ux= y .
(13.2.4)
The advantage is that the solution of a triangular set of equations is quite trivial. Eq. (13.2.3)
can be solved by forward substitution and Eq. (13.2.4) can be solved by back-substitution. To
solve L and U for given A, the Crouts algorithm is used. The decomposition is in place and
does not require an additional memory. The pivoting is necessary for the stability of the
Crouts algorithm. Only partial pivoting (interchange of rows) is implemented efficiently. The
matrix A is not usually decomposed into LU form, but only a row-wise permutation of A is
used. This is enough to make the method stable.
The LU decomposition requires about 1/3M3 executions of the inner loops (each with
one multiply and one add). This is 3 times better than the Gauss-Jordan routine, and 1.5 times
better than a Gauss-Jordan routine which does not compute the inverse matrix.
(13.3.1)
for i=2NM 1 ,
a N , N 1 x N 1a N , N x N =b N ,
(13.3.2)
(13.3.3)
][ ] [ ]
a 1,1 a 1,2 0
0
0
0
x1
b1
a 2,1 a 2,2 a 2,3
0
0
0
x2
b2
= .
0
0
0 a N 1, N 2 a N 1, N 1 a N 1, N x N 1
b N 1
0
0
0
0
a N , N 1
aN ,N
xN
bN
(13.3.4)
(13.3.5)
where
IX
a 1,2
a 1,1
b1
Q 1=
.
a 1,1
P 1=
(13.3.6)
This relation is substituted into Eq. (13.3.2) for i = 2. The result is that x2 is expressed using
x3. As the substitution process continues, it expresses each xi-1 using xi
x i1=P i1 x i Q i1 .
(13.3.7)
(13.3.8)
(13.3.9)
(13.4.1)
Band diagonal matrices are usually stored in a compact form. It is not possible to store the LU
decomposition of a band diagonal matrix A as compactly as the compact form of A itself
because the decomposition produces additional nonzero values.
multiplied by the matrix A, it gives a product slightly different from the desired b
A x x =b b .
(13.5.1)
(13.5.2)
Eq. (13.5.1) can be solved by substituting Eq. (13.5.2) for b which gives
A x= A x x b
(13.5.3)
In this equation, the whole right-hand side is known. In the case of LU decomposition the A is
already available and the x can be easily and qickly computed by back-substitution.
1
xAxbx .
2
(13.6.1)
(13.6.2)
XI
is zero, which is equivalent to Eq. (3.3.47). To solve Eq. (3.3.47), an initial guess of x1 is
made for the solution. The residual is
r 1=b Ax 1
(13.6.3)
where the superscripts are iteration indexes and the initial values are
1
r =r
1
1
p =r
1
1
p =r .
(13.6.4)
(13.6.5)
where
i i
r r
i
i
p Ap
i1
i
i
i
r =r Ap
i1
i
i
T
i
r =r A p
i1 i1
r r
i
= i i
r r
i1
i1
i i
p =r p
i1
i1
i i
p =r p .
i =
(13.6.6)
When rm+1 = 0, the xm+1 is the solution of Eq. (3.3.47). There is no guarantee that the whole
procedure will not break down or become unstable for general A. The iterative process should
be halted when some appropriate error criterion is met.
XII
XIII
38.154
35
30
Temperature [C]
25
Fine model
Analytical
FDM
FVM
FEM
20
15
10
5
0
-3.986
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
12
Fine model
Analytical
FDM
FVM
FEM
10
8
6
4
2
0
-2
-4.407
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
Temperature [C]
57.04
Fine model
Analytical
FDM
FVM
FEM
40
30
20
10
-0.389
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
9.529
Fine model
Analytical
FDM
FVM
FEM
6
4
2
0
-1.865
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
Temperature [C]
152.969
Fine model
Analytical
FDM
FVM
FEM
100
50
-0.001
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
Fine model
Analytical
FDM
FVM
FEM
1
0
-0.65
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XV
q [W/m]
100000
95000
90000
86584.103
0.0
0.1
0.2
0.3
0.4
0.5
t [s]
0.6
0.7
0.8
0.9
1.0
Temperature [C]
20
Fine model
FDM
FVM
FEM
15
10
5
0
-2.988
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
8
Fine model
FDM
FVM
FEM
6
4
2
0
-3.383
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XVI
100000
95000
q [W/m]
90000
85000
80000
75000
70000
65852.238
0
5
t [s]
10
Temperature [C]
30
25
Fine model
Analytical
FDM
FVM
FEM
20
15
10
5
-0.301
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
1.5
Fine model
Analytical
FDM
FVM
FEM
1.0
0.5
0.0
-0.5
-1.0
-1.498
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XVII
100000
90000
q [W/m]
80000
70000
60000
50000
35149.688
0
10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
t [s]
Temperature [C]
50
Fine model
Analytical
FDM
FVM
FEM
40
30
20
10
-0.001
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
-0.1
Fine model
Analytical
FDM
FVM
FEM
-0.2
-0.3
-0.4
-0.5
-0.6
-0.7009
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XVIII
100000
q [W/m]
50000
0
-50000
-99999.992
0
5
t [s]
10
Temperature [C]
Fine model
FDM
FVM
FEM
-5
-10
-15.52
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
0
-5
Fine model
FDM
FVM
FEM
-10
-15
-20.663
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XIX
100000
q [W/m]
80000
60000
40000
20000
0
0
5
t [s]
10
Temperature [C]
14
12
Fine model
FDM
FVM
FEM
10
8
6
4
2
-0.535
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
-1
-2
Fine model
FDM
FVM
FEM
-3
-4
-5
-6
-7
-8.132
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XX
100000
q [W/m]
80000
60000
40000
20000
0
0
5
t [s]
10
16
14
Fine model
FDM
FVM
FEM
12
10
8
6
4
2
-1.069
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
dT [C]
-2
Fine model
FDM
FVM
FEM
-4
-6
-8
-10.69
0.00
0.01
0.02
0.03
0.04
0.05
x [m]
0.06
0.07
0.08
0.09
0.10
XXI
q=c
dx
dy
dt
(15.1.1)
where
q x =k
dT
,
dx
q y =k
dT
.
dy
(15.1.2)
The discretization of the 2D model is shown in Figure 4.6. Integrating Eq. (15.1.1) over
a control volume gives
m
y k q J , k q J 1, k x j q j , K q j , K 1 x j y kq j , k
T m T m1
j ,k
.
= x j y k j , k T mj , k c j , k T mj , k j , k
t
(15.1.3)
We can write Eqs. (15.1.3, 4.1.6, 4.1.10, and 4.1.21) in the form of Eqs. (4.2.3 and 4.2.4). The
a and b parameters are as follows:
a j , k 1=2 x j
a j1, k =2 y k
y k 1
k j , k 1 T m1
j , k 1
x j1
k j1, k T m1
j1, k
yk
1
k j , k T m1
j ,k
xj
k j , k T m1
j ,k
1
x j y k q
b
j ,k
m1
j ,k
m1
j , k T m1
j , k c j , k T j , k
,
t
XXII
a j1, k =2 y k
xj
m1
k j , k T j , k
a j , k 1=2 x j
yk
m1
k j , k T j , k
b j , k = x j y k q
1
x j1
m1
k j1, k T j1, k
y k 1
m1
k j , k 1 T j , k 1
1
m1
a
j ,k
m1
j ,k
m1
j , k T j , k c j , k T j , k m1 .
T j , k
t
(15.1.4)
m1
m
For all subsequent iterations T i1,
substitutes for T j , k in aj, k-1, aj-1, k, aj+1, k, and aj, k+1. The
j ,k
m
b j , k = x j y k q aj , k T i1,
j , k
m
m1
H j , k T i1,
j , k H j , k T j , k
.
T m1
j ,k
t
(15.1.5)
q=c
r dr r d
dt
(15.1.6)
where
q r =rk
q =
dT
,
dr
k dT
.
r d
(15.1.7)
m
j ,k
c j , k T
m
j ,k
T mj , k T m1
j ,k .
(15.1.8)
XXIII
r j
r j1
2
2
m
q J , k =r J
m
m
k j1, k T j1, k k j , k T j , k
1
T j1, k T j , k
m
(15.1.9)
k
k 1
1
2
2
m
q j ,K=
m
r j k j , k 1 T j , k 1 k j , k T mj , k
1
T j , k 1T j , k
m
(15.1.10)
We can write Eqs. (15.1.8, 15.1.9, 15.1.10, 4.1.10, and 4.1.21) in the form of Eqs. (4.2.3 and
4.2.4). The a and b parameters are as follows:
r j
k 1
k
a j , k 1=2
m1
r j k j , k 1 T j , k 1 k j , k T m1
j ,k
a j1, k =2 r J k
r j1
k j1, k T m1
j1, k
1
,
1
r j
k j , k T m1
j ,k
m1
m1
r j r j k q j , k T j , k
a j1, k =2 r J 1 k
r j
m1
k j , k T j , k
b j , k =r j r j k q
m1
j ,k
1
r j1
m1
k j1, k T j1, k
k 1
r j
k
a j , k 1=2
m1
r j k j , k T j , k k j , k 1 T m1
j , k 1
a
j,k
m1
j , k T j , k c j , k T j , k
,
t
1
m1
j , k T m1
j , k c j , k T j , k
.
T m1
j ,k
t
(15.1.11)
m1
m
For all subsequent iterations T i1,
substitutes for T j , k in aj, k-1, aj-1, k, aj+1, k, and aj, k+1. The
j ,k
aj, k and bj, k parameters are
m
a j , k =a j , k 1a j1, k a j1, k a j , k 1r j r j kq bj , k T i1,
j,k ,
m
b j , k =r j r j k q aj , k T i1,
j , k
m
m1
H j , k T i1,
j , k H j , k T j , k
.
T m1
j ,k
t
(15.1.12)
XXIV
q=c
dx
dy
dz
dt
(15.2.1)
where
q x =k
dT
,
dx
q y =k
dT
,
dy
q z =k
dT
.
dz
(15.2.2)
The discretization of the 3D model is shown in Figure 4.6. Integrating Eq. (15.2.1) over the
control volume gives
m
y k z p q J , k , p q J 1, k , p x j z p q j , K , p q j , K 1, p x j y k q j , k , P q j , k , P1
x j y k z pq j , k , p = x j y k z p j , k , p T
m
j ,k , p
c j , k , p T
m
j ,k , p
T mj , k , p T m1
j ,k , p
.
t
(15.2.3)
We can write Eqs. (15.2.3, 4.1.6, 4.1.10, and 4.1.21) in the form of Eq. (4.2.5). The a and b
parameters are as follows:
a j , k , p1=2 x j y k
a j , k 1, p =2 x j z p
a j1, k , p =2 y k z p
z p1
k j , k , p1 T m1
j , k , p1
y k 1
k j , k 1, p T m1
j , k 1, p
x j1
m1
k j1, k , p T j1, k , p
zp
1
k j , k , p T m1
j , k , p
yk
k j , k , p T m1
j , k , p
xj
m1
k j , k , p T j , k , p
1
1
XXV
m1
x j y k z p q j , k , p T j , k , p
a j1, k , p =2 y k z p
a j , k 1, p =2 x j z p
m1
j , k , p T m1
j , k , p c j , k , p T j , k , p
,
t
xj
m1
k j , k , p T j , k , p
yk
m1
k j , k , p T j , k , p
a j , k , p1=2 x j y k
zp
k j , k , p T m1
j , k , p
x j1
m1
k j1, k , p T j1, k , p
y k 1
m1
k j , k 1, p T j , k 1, p
z p1
1
1
1
k j , k , p1 T m1
j , k , p1
b j , k , p = x j y k z p q aj , k , p T m1
j , k , p
m1
j , k , p T m1
j , k , p c j , k , p T j , k , p
.
T m1
j ,k , p
t
(15.2.4)
m1
m
For all subsequent iterations T i1,
substitutes for T j , k , p in aj, k, p-1, aj, k-1, p, aj-1, k, p, aj+1, k, p,
j ,k , p
aj, k+1, p, and aj, k, p+1. The aj, k, p and bj, k, p parameters are
a j , k , p =a j , k , p1 a j , k 1, p a j1, k , p a j1, k , p a j , k 1, p a j , k , p1
b
m1
x j y k z pq j , k , p T j , k , p ,
b j , k , p = x j y k z p q
a
j,k , p
i1, m
j,k , p
m
m1
H j , k , p T i1,
j , k , p H j , k , p T j , k , p
.
T m1
j ,k , p
t
(15.2.5)
q=c
r dr r d dz
dt
(15.2.6)
where
q r =rk
q =
dT
,
dr
k dT
,
r d
dq z =k
dT
.
dz
(15.2.7)
XXVI
m1
T j , k , p T j , k , p
m
m
.
r j r j k z pq j , k , p =r j r j k z p j , k , p T j , k , p c j , k , p T j , k , p
t
(15.2.8)
We can write Eqs. (15.2.8, 15.1.9, 15.1.10, 4.1.6, 4.1.10, and 4.1.21) in the form of
Eq. (4.2.5). The a and b parameters are as follows:
z p1
a j , k , p1=2 r j r j k
m1
k j , k , p1 T j , k , p1
1
zp
m1
k j , k T j , k , p
r j z p
k 1
k
a j , k 1, p =2
m1
rj
k j , k 1, p T j , k 1, p k j , k , p T m1
j ,k , p
a j1, k , p =2 r J k z p
r j1
k j1, k , p T m1
j1, k , p
1
,
1
r j
k j , k , p T m1
j ,k , p
r j r j k z p q
b
j ,k , p
a j1, k , p =2 r J 1 k z p
m1
j ,k , p
m1
j , k , p T m1
j , k , p c j , k , p T j , k , p
,
t
r j
m1
k j , k , p T j , k , p
m1
k j1, k , p T j1, k , p
r j z p
k
k 1
a j , k 1, p =2
m1
rj
k j , k , p T j , k , p k j , k 1, p T m1
j , k 1, p
a j , k , p1=2 r j r j k
zp
k j , k , p T m1
j , k , p
b j , k , p =r j r j k z p q
a
j ,k , p
m1
j ,k , p
1
r j1
1
z p1
k j , k , p1 T m1
j , k , p1
1
m1
j , k , p T m1
m1
j , k , p c j , k , p T j , k , p
T j , k , p .
t
(15.2.9)
m
m1
For all subsequent iterations T i1,
j , k , p substitutes for T j , k , p in aj, k, p-1, aj, k-1, p, aj-1, k, p, aj+1, k, p,
aj, k+1, p, and aj, k, p+1. The aj, k, p and bj, k, p parameters are
XXVII
b j , k , p =r j r j k z p q
i1, m
a
j ,k , p
i1, m
j ,k , p
m1
H j , k , p T j , k , p H j , k , p T j , k , p m1
T j , k , p .
t
(15.2.10)
XXVIII
XXIX
XXX
XXXI
XXXII
XXXIII
17 SUMMARY
A computer analysis of complex problems may provide a cost-effective, quick and
sufficiently reliable method for many applications. Computational methods may also be
a complement to experimental investigations, especially for inverse tasks. Although computer
science development has reached a very high level and a wide range of applications is now
available, research on many topics is needed. Heat transfer computation can be significantly
helpful in many engineering and industrial applications but comprehensive research is still
needed. Thermal numerical models for design and control purposes in metallurgical industry
require a precise description of heat transfer phenomena. Comprehensive quantitative
information on the heat transfer phenomena is not available for quenching of hot moving
surfaces. In this work, attention is focused on the search for boundary conditions describing
the heat transfer in engineering applications of spray cooling of metal surfaces and cooling of
molds during casting. Determination of boundary conditions represents solving inverse heat
conduction problems using experimental data. Combining measurement with an inverse
analysis often results in an ill-posed problem. Such problems are extremely sensitive to
measurement errors. If any results are obtained, they are very inaccurate and the precision is
often unknown. This work is also aimed at estimation of errors caused by discretization in
both space and time domains. Inverse tasks, such as determination of boundary conditions and
temperature dependent thermal material properties, are computed using optimized models.
Combining optimization methods or methods of artificial intelligence with classical
approaches seems to be useful in solving such ill-posed problems. A higher accuracy is
achieved during solution of ill-posed inverse tasks.
18 RESMEE
Die Computeranalyse technischer Probleme kann eine finanzgnstige, schnelle und
genug zuverlssige Methode bei der Lsung verschiedener Probleme darstellen. Die
Berechnungsmethoden knnen auch ein Komplement der Experimentalforschung bilden,
insbesondere bei der Lsung der Inversionsaufgaben. Obwohl die Entwicklung in der
Computerwissenschaft bereits ein hohes Niveau erreichen hat und eine Reihe von
Programmsystemen existiert, ist es immerhin notwendig, Forschungen in vielen Richtungen
vorzunehmen. Die Berechnung der Intensitt der Wrmebertragung kann in vielen Ingenieurund Industrieapplikationen ntzlich sein, aber hufig ist es noch immer eine weitere
Forschung notwendig. Die fr den Entwurf und die Steuerung in der metallurgischen Industrie
angewandten
nummerischen
Beschreibung
der
temperaturabhngigen
Modelle
Bedingungen
der
der
Thermoprozesse
Wrmebertragung
Materialcharakteristiken.
erfordern
und
eine
Zusammenfassende
eine
genaue
Kenntnis
und
der
quantitative
Informationen, die die Bedingungen bei der Wrmebertragung zur Khlung heier, sich
bewegender Oberflchen betreffen, sind noch immer nicht verfgbar. In dieser Arbeit befat
man sich mit der Feststellung der Randbedingungen, die die Wrmebertragung in den
Ingenieurapplikationen bei der Khlung von Metalloberflchen durch Bespritzen und bei der
Khlung der Formen beim Abgieen beschreiben. Die Bestimmung der Randbedingungen
stellt eine Lsung der Wrmeleitunginversionsaufgabe mit der Ausnutzung von experimentell
gewonnenen Daten dar. Die Kombination einer Messung mit einer Inversionsaufgabe fhrt oft
zur Lsung eines nicht recht bedingten Problems. Die Berechnungen sind dann sehr
medatenfehlerempfindlich. Wenn irgendwelche Ergebnisse gewonnen werden, sind sie sehr
ungenau und oft ist ihre Przision nicht bekannt.Die Arbeit befat sich auch mit der
Berechnung des Maximalfehlers, entstanden durch die Diskretisierung eines kontinuierlichen
Problems sowohl in einer Raum-, als auch in einer Zeitdomne. Mit Hilfe optimalisierter
Modelle werden dann Inversionsaufgaben zum Gewinn von Randbedingungen und zur
Bestimmung temperaturabhngiger Materialeigenschaften berechnet. Die Kombination der
Optimalisierungs- und der klassischen Methoden, beziehungsweise die Ausnutzung der
Kunstintelligenzmethoden, die in der letzten Zeit immer mehr ausgenutzt werden, scheint bei
der Lsung nicht recht bedingter Probleme sehr ntzlich zu sein. So erreicht man hhere
Genauigkeiten bei der Lsung nicht recht bedingter Inversionsaufgaben.
B
19 RESUM
Potaov analza technickch problm me poskytovat finann vhodnou, rychlou
a dostaten spolehlivou metodu pi een rznch problm. Vpotov metody mohou bt
tak doplkem experimentlnho vzkumu, zejmna pi een inverznch loh. Akoliv vvoj
v potaov vd doshl ji velmi vysok rovn a existuje ada programovch systm, je
stle zapoteb provdt vzkum v mnoha smrech. Vpoet intenzity penosu tepla me bt
velmi uiten v mnoha inenrskch a prmyslovch aplikacch, ale asto je stle jet
zapoteb dal vzkum. Numerick modely tepelnch proces v metalurgickm prmyslu pro
nvrh a zen vyaduj pesn popis podmnek penosu tepla a znalost teplotn zvislch
materilovch charakteristik. Souhrnn a kvantitativn informace, kter se tkaj podmnek pi
penosu tepla pro ochlazovn horkch pohybujcch se povrch, jet stle nejsou k dispozici.
V tto prci je zamena pozornost na zjiovn okrajovch podmnek popisujcch penos
tepla v inenrskch aplikacch pi chlazen kovovch povrch ostikem a chlazen forem pi
odlvn. Uren okrajovch podmnek pedstavuje een inverzn lohy veden tepla
s vyuitm experimentln zskanch dat. Kombinace men s inverzn lohou vede asto
k een patn podmnnho problmu. Nsledn vpoty jsou pak velice citliv na chyby
v namench datech. Pokud jsou vbec njak vsledky zskny, jsou velmi nepesn a asto
nen znma chyba vsledku. Tato prce se tak zamuje na vpoet maximln chyby
vzniklou diskretizac spojitho problmu, a to jak v prostoru, tak v ase. Pomoc
optimalizovanch model jsou pak potny inverzn lohy pro zskn okrajovch podmnek
a pro uren teplotn zvislch materilovch vlastnost. Kombinace optimalizanch
a klasickch metod, ppadn vyuit metod uml inteligence, kter jsou v posledn dob
stle vce vyuvny, se zd bt velmi uiten pi een patn podmnnch problm.
Dosahuje se tak vych pesnost pi een patn podmnnch inverznch loh.