See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/304498692

Parameter estimation in heat conduction using a two dimensional inverse

analysis

Article in International Journal for Computational Methods in Engineering Science and Mechanics · June 2016

CITATIONS READS

3 366

2 authors:

Farzad Mohebbi Mathieu Sellier

Swansea University University of Canterbury
26 PUBLICATIONS 230 CITATIONS 165 PUBLICATIONS 1,565 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Self-propelled, Marangoni-driven droplets View project

Inertial and dimensional effects on the instability of a thin film View project

All content following this page was uploaded by Farzad Mohebbi on 23 March 2017.

The user has requested enhancement of the downloaded file.

 International Journal for Computational Methods in
Engineering Science and Mechanics

ISSN: 1550-2287 (Print) 1550-2295 (Online) Journal homepage: http://www.tandfonline.com/loi/ucme20

Parameter estimation in heat conduction using a

two dimensional inverse analysis

Farzad Mohebbi & Mathieu Sellier

To cite this article: Farzad Mohebbi & Mathieu Sellier (2016): Parameter estimation in heat
conduction using a two dimensional inverse analysis, International Journal for Computational
Methods in Engineering Science and Mechanics, DOI: 10.1080/15502287.2016.1204034

To link to this article: http://dx.doi.org/10.1080/15502287.2016.1204034

Accepted author version posted online: 27

Jun 2016.
Published online: 27 Jun 2016.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=ucme20

Download by: [Dr Farzad Mohebbi] Date: 27 June 2016, At: 15:26
Parameter estimation in heat conduction
using a two dimensional inverse analysis
Farzad Mohebbi*, Mathieu Sellier

Department of Mechanical Engineering, The University of Canterbury, Christchurch,

New Zealand

Abstract. This paper is concerned with a two dimensional inverse steady state heat conduction problem.

The aim of this study is to estimate the thermal conductivity, the heat transfer coefficient, and the heat
flux in irregular bodies (both separately and simultaneously) using a two dimensional inverse analysis.
The numerical procedure consists of an elliptic grid generation technique to generate a mesh over the
irregular body and solve for the heat conduction equation. This paper describes a novel sensitivity
analysis scheme to compute the sensitivity of the temperatures to variation of the thermal conductivity,
the heat transfer coefficient, and the heat flux. This sensitivity analysis scheme allows for the solution of
inverse problem without requiring solution of adjoint equation even for a large number of unknown
variables. The conjugate gradient method (CGM) is used to minimize the difference between the
computed temperature on part of the boundary and the simulated measured temperature distribution.
The obtained results reveal that the proposed algorithm is very accurate and efficient.

Keywords. Inverse Heat Transfer, Elliptic Grid Generation, Sensitivity Analysis, Finite Difference
Method, Conjugate Gradient Method

-------------------------------------------------------------------------------------------------------------------------------------------

*Corresponding author.

E-mail address: [email protected]

1
NOMENCLATURE:

d (k) direction of descent at iteration k

W
qɺ heat flux ( )
m2
W
h heat transfer coefficient ( )
m2 . C
Ja Jacobian matrix
J Jacobian of transformation
J objective function
W
k thermal conductivity of the solid body ( )
m. C
n outward drawn unit vector
T temperature (oC)
Tm measured outer surface temperature (oC)
T∞ ambient temperature (oC)
x, y Cartesian coordinates in the physical domain

Greek symbols
α, β, γ metric coefficients in 2-D elliptic grid generation
β (k) search step size at iteration k
Γ boundary
γ (k) conjugation coefficient at iteration k
Ω domain
ξ, η Cartesian coordinates in the computational domain

Subscripts
i grid index in ξ - direction
j grid index in η - direction
M number of grid points in the ξ - direction
N number of grid points in the η - direction

Superscript
k iteration number

2
 1. Introduction
Inverse problems play a significant role in almost every branch of science and
engineering. Inverse Heat Transfer Problems (IHTP) are a significant class of inverse
problems which have been greatly researched by engineers and mathematicians since the
beginning of the space program in 1957 [1, 2]. In a heat transfer problem, if the
boundary conditions, the thermo-physical properties, the geometrical configuration of
the heated body, and the applied heat flux are all known but the temperature
distribution is unknown, the heat transfer problem is referred to as the direct heat
transfer problem. Contrary to the direct heat transfer problem, the inverse heat transfer
problem is concerned with the determination of the boundary conditions, the thermo-
physical properties, the geometrical configuration of the heated body, and the applied
heat flux by knowing the temperature distribution on some part of the heat conducting
body boundary. Such problems are ill-posed. The difficulty in solving ill-posed problems
results from the fact that they are inherently unstable and very sensitive to noise and
perturbation. Therefore, the development of new numerical schemes, their numerical
implementation, and their efficiency improvement is of crucial importance.

Applications of IHTP include the determination of thermo-physical

properties such as the thermal conductivity and the heat transfer coefficient [3-18], the
estimation of heat flux [16, 19-23], and the estimation of the boundary shapes of
bodies [24-28], to name a few. Other parameters involved in heat transfer problems are
also estimated using an inverse analysis [29, 30]. For example, in transient conduction-
radiation heat transfer problems where the medium is participating, the extinction
coefficient and the scattering albedo are parameters that affect the temperature
distribution. In [30] a combination of the lattice Boltzmann method (to solve the
energy equation) and the finite volume method (to to compute the radiative
information) with the particle swarm optimization is used to recover the extinction

3
coefficient and the scattering albedo in a conducting-radiating planar participating
medium.

This paper deals with a two dimensional inverse steady state heat conduction
problem. The objective of this study is to estimate the thermal conductivity, the heat
transfer coefficient, and the heat flux in an irregular body. In the proposed numerical
methodology, an elliptic grid generation technique is used to generate a mesh over the
irregular body and solve for the heat conduction equation. A novel sensitivity analysis
scheme is formulated to compute the sensitivity of the temperatures to variation of the
thermal conductivity, heat transfer coefficient, and the heat flux. The conjugate
gradient method is employed to minimize the difference between the computed
temperature on part of the boundary and desired temperature. As will be shown, this
numerical methodology does not require the solution of an adjoint problem. An explicit
expression for the sensitivity coefficients is derived. This enables the computation of the
sensitivity coefficients in one single solve, a performance analogous to the adjoint
method, the current state-of-the art. The discretization in the computational domain is
based on the Finite Difference method, a method chosen for its simplicity and ease of
implementation.

The solution method introduced here is sufficiently general and can be employed for the
estimation of the thermal conductivity, the heat transfer coefficient, and the heat flux
applied on part of the boundary of a general two-dimensional region as long as the
general two-dimensional region can be mapped onto a regular computational domain.

2. Governing equation
The mathematical representation of the steady state heat conduction problem of interest
here can be expressed as below (see Fig. 1)

4
 ∇2T = 0 in physical domain Ω (1)

subject to the boundary conditions

∂T qɺ
= − on boundary surface Γ1 (2)
∂n k

∂T h
= − i (TΓ -T∞ ) on boundary surface Γi , i = 2, 3, 4 (3)
∂n k i i

h2,T¥
2
D
C
G2
body
(physical domain) h4 ,T¥
4

G3 W G4
h3,T¥
3

A G1 B
y
x
q&

Fig. 1 Physical domain (solid body) subjected to convective heat transfer on surfaces Γi , i = 2, 3, 4 and

heat flux qɺ on surface Γ1 . The thermal conductivity of the body is k .

In this study, the elliptic grid generation technique proposed by Thompson et al. [31,
32] is employed to discretize the physical domain and approximate the derivatives of the
field variable (temperature) by algebraic ones. This technique is based on solving a
system of elliptic partial differential equations to distribute nodes in the interior of the
physical domain by mapping the irregular physical domain from the x and y physical
plane (Fig. 1) onto the ξ and η computational plane (Fig. 2), which is a regular region
[33].

5
 C D
N

...
2

1
1 2 ... M
A B
h
x

Fig. 2 Computational domain.

More details on the implementation of the elliptic grid generation technique and
solution procedure for the steady state heat conduction equation can be found in [34].

3. The inverse analysis

First, the steady state heat conduction problem is solved using the known values for the
thermal conductivity of conducting body k , the heat transfer coefficient hi imposed on
the surface Γi ( i = 2, 3, 4 ), and the heat flux qɺ applied on the surface Γ1 to obtain the
temperature distribution on the outer surface Γ2 (Ti,N , i = 1, …, M ). Then the resulting

outer surface temperature distribution is used as the simulated measured temperatures

in the inverse analysis to recover the initially used values for three parameters: the
thermal conductivity, the heat transfer coefficient, and the heat flux. To fulfill the aim
of the inverse analysis of interest here, the square of the difference between the
temperature distribution of the outer surface Γ2 (obtained from the solution the direct
problem at each iteration) and the simulated measured temperature distribution of the
same surface ( Γ2 ) is to be minimized.

3.1 Estimation of the thermal conductivity k :

In more formal mathematical terms, the inverse problem of estimating the thermal
conductivity can be represented as

6
  
min  J := C TΓ − TΓ 2
: ∇2T = 0 in Ω, BCs in Eqs.(2)&(3) (4)
k in Ω 2 2m 

where C is a positive constant and can be considered as C = 10n , n = 0,1,2,... . The

parameter C is introduced to prevent a negative value for the parameters in the
iterative process. A negative value for, say, the thermal conductivity terminates the
iterative process.

The unknown variable here is the thermal conductivity of the heat conducting body k .
The value of the thermal conductivity has to be numerically optimized for the objective
function

M
J = C ∑ (Ti,N − T(i,N ) )2 (5)
m
i =1

to reach a minimum.

3.2 Sensitivity analysis ( k )

In gradient-based optimization, it is necessary to calculate the effect of the variation of
the unknown variable on the objective function. The computation of derivative of the
objective function with respect to the unknown variable is an integrated part of
gradient-based optimization, which is called the sensitivity analysis. Here we wish to
calculate the sensitivity of the objective function J defined by Eq. (5) to the thermal
conductivity value k . Therefore, we have

∂J M ∂Ti,N M ∂Ti,N
= 2C ∑ (Ti,N − T(i,N ) ) = 2∑ (Ti,N − T(i,N ) )(C ) (6)
∂k i =1
m ∂k i =1
m ∂k

∂Ti,N
The expression C in the above relation is called the sensitivity coefficient. The
∂k
sensitivity analysis can be performed by introducing small perturbations to the thermal
conductivity k . The direct problem may be solved using this perturbated value of

7
thermal conductivity to obtain the new temperatures Ti,N . Using these values for the

temperatures, one can evaluate the dependency of the temperatures Ti,N on the

perturbation of the thermal conductivity. The finite difference method may be used to
formulate these perturbations as follows

∂Ti,N Ti,N (k + εk ) − Ti,N (k )

= (7)
∂k εk

where ε is a very small number. The computation of the sensitivity coefficients using
this method requires an additional solution of direct heat transfer problem. The problem
of interest here is concerned with the estimation of only one variable ( k ). However, if
there are many, say N , unknown variables (for example, design variables in shape
optimization problems), the use of perturbation method requires N additional solution
of the direct problem which makes it of very high computational cost. In such a
situation, the adjoint method is preferable. However, the adjoint method comes with
mathematical complexity. In addition, the adjoint equation solution itself has a
computational cost comparable to the direct solution. In this study, we will introduce a
novel method to calculate the sensitivity coefficients based on the finite difference
method and transformation rules from the physical domain to the computational one.
As will be shown, it requires only one solution of the direct problem (at each
iteration) to compute all sensitivity coefficients, regardless of the number of
unknown variables.

From the direct solution of the heat conduction equation we have

qɺconduction |Γ = qɺconvection |Γ
2 2

∂T
−k = h2 (TΓ − T∞ ) (8)
∂n 2 2 2

8
 ∂T ∂T
We know that at a boundary surface in the physical domain is related to
∂n ∂ξ
∂T
and/or at the corresponding transformed boundary surface in the computational
∂η
domain. At surface Γ2 we have

∂T ∂T 1
|Γ = |Γ = (γTη − βTξ ) (9)
∂n 2 ∂η 2
J γ

where (from the elliptic grid generation method)

α = x η2 + yη2 (10)

β = x ξ x η + y ξy η (11)

γ = x ξ2 + yξ2 (12)

J = x ξy η − x ηy ξ (Jacobian of transformation) (13)

Therefore, Eq. (8) will be

 1 
−k 
 J γ
(γTη − βTξ ) 

2(
 = h T −T
i ,N ∞2 ) (14)

Using finite difference method, one can discretize the relations for Tη and Tξ at every

boundary surface with Neumann and Robin conditions. For example, for nodes on
surface Γ2 with i ∈ [2, M − 1] , it can be written

∂T T − Ti −1,N
= Tξ = i +1,N (15)
∂ξ 2

∂T 3Ti,N − 4Ti,N −1 + Ti,N −2

= Tη = (16)
∂η 2

9
(for i = 1 and i = M one-sided forward and one-sided backward relations may be
used). By substituting the relations (15) and (16) into Eq. (14) and then solving for Ti,N

using an algebraic equation solver such as Maple, a relationship is derived which allows
us to obtain the temperature distribution on the surface Γ2 . The expression for Ti,N ,

therefore, is

4k γTi,N −1 − k γTi,N −2 + k βTi +1,N − k βTi −1,N + 2h2J γT∞

2
Ti,N = (17)
3k γ + 2h2J γ

∂Ti,N
Therefore, the term in Eq. (6) is obtained by differentiating Eq. (17) with respect
∂k
to k as follows

∂Ti,N 2h2J γ (−4 γTi,N −1 + γTi,N −2 − βTi +1,N + βTi −1,N + 3γT∞ )
2
=− (18)
∂k 2
(3k γ + 2h2J γ )

The sensitivity matrix Ja can be explicitly written as

 ∂T 
 1,N 
 ∂k 
 
 ∂T2,N 
 
Ja = C  ∂k  (19)
 
 ⋮ 
 
 ∂TM ,N 
 
 ∂k  M ×1

3.3 The Conjugate Gradient Method (CGM)

CGM is one of the most powerful optimization methods. It is robust, reliable, and has a
high rate of convergence. In this method, the objective function given by Eq. (5) is

minimized by searching along the direction of descent d (k) using a search step size β (k) .

k (k +1) = k (k) − β(k)d (k) (20)

10
The direction of descent of the current iteration is obtained as a linear combination of

the direction of descent of the previous iteration and the gradient direction ∇J(k) . In

other words, it is a conjugation of the gradient direction of the current iteration, ∇J(k) ,
and the direction of descent of the previous iteration, d (k−1) , accordingly

d (k) = ∇J(k) + γ (k)d (k−1) (21)

where γ (k) is the conjugation coefficient. In this study, the Polak-Ribiere formula [35] is
used for the conjugation coefficient:

∇J (k)  T (∇J(k) − ∇J (k−1) ) ∇J (k)  T (∇J(k) − ∇J (k−1) )

   
γ (k) =  = (22)
∇J (k−1) 2 ∇J(k−1)  T ∇J(k−1)
 

By minimizing the objective function J with respect to the search step size β (k) , one
can find an expression for calculating the search step size as follows [2]

[ Ja(k)d (k) ]T [Ti,N − T(i,N ) ]

(k) m
β = (23)
(k) (k) T (k) (k)
[ Ja d ] [ Ja d ]

By the algorithm given below one can obtain the thermal conductivity of the heat
conducting solid k for the problem under consideration.

1. Specify the physical domain, the boundary conditions, and the simulated measured
outer surface temperature.

2. Generate the boundary fitted grid using the elliptic grid generation method.

3. Solve the direct problem of finding the temperature values at any grid points of the
physical domain using an initial thermal conductivity (initial guess).

11
4. Using Eq. (5), compute the objective function ( J (k) ).

5. If value of the objective function obtained in step 4 is less than the specified stopping
criterion, the optimization is finished. Otherwise, go to step 6.

6. Compute the sensitivity matrix Ja from Eq. (19).

7. Compute the gradient direction ∇J(k) from Eq. (6).

8. Compute the conjugation coefficient γ (k) either from Eq. (22). For k = 0 , set

γ (0) = 0 .

9. Compute the direction of descent d (k) from Eq. (21).

10. Compute the search step size β (k) from Eq. (23).

11. From Eq. (20), evaluate the new thermal conductivity k (k+1) .

12. Set the next iteration ( k = k + 1 ) and return to the step 2.

In Eq. (20), if the value of the expression β (k)d (k) is larger than the value of k (k) , then

the new value for the thermal conductivity, k (k+1) , becomes negative resulting in
termination of the iterative process. Thus the parameter C is introduced in Eq. (4) to

decrease the value of the expression β (k)d (k) and obtain a positive value for the thermal
conductivity in the next iteration, thereby continuing the iterative process. We start
every problem with an ordinary least square objective function for which n = 0 , C = 1 .

3.4 Estimation of the heat flux qɺ :

The inverse problem of estimating the heat flux can be expressed as

 
min  J := AC TΓ − TΓ 2
: ∇2T = 0 in Ω, BCs in Eqs.(2)&(3) (24)
qɺ at Γ 2 2m 
 1 
12
where A is a constant which will be defined later. The unknown variable here is the
heat flux imposed on the surface Γ1 , qɺ . The value of the heat flux has to be
numerically optimized for the objective function

M
J = AC ∑ (Ti,N − T(i,N ) )2 (25)
m
i =1

to reach a minimum.

3.5 Sensitivity analysis ( qɺ )

Here we wish to calculate the sensitivity of the objective function J defined by Eq. (25)
to the heat flux qɺ . Therefore, we have

∂J M ∂Ti,N
= 2A∑ (Ti,N − T(i,N ) )(C ) (26)
∂qɺ i =1
m ∂qɺ

∂Ti,N
To calculate the coefficient we can write (using the chain rule)
∂qɺ

∂Ti,N
∂Ti,N ∂k
= (27)
∂qɺ ∂qɺ
∂k

∂Ti,N
The term is calculated from Eq. (18). By considering the boundary condition
∂k
∂qɺ
given by Eq. (2), the term can be evaluated as
∂k

∂T  −1 
qɺ = −k = −k  (γTη − βTξ ) (28)
∂n  J γ  at Γ
at Γ1 1

Thus

13
 ∂qɺ γTη − βTξ
= (29)
∂k J γ at Γ1

At surface Γ1 , we can write, for example

Ti +1,1 − Ti −1,1
Tξ =
2

−3Ti,1 + 4Ti,2 − Ti,3

Tη =
2

Since the heat flux qɺ and the thermal conductivity k both are constant, it can be
γTη − βTξ
concluded from Eq. (28) that the term should be constant and is
J γ at Γ1

denoted by A . Therefore, from Eq. (29) we obtain

∂qɺ
= A = constant (30)
∂k

Substituting Eqs. (18) and (30) into Eq. (27) gives

2h2J γ (−4γTi,N −1 + γTi,N −2 − βTi +1,N + βTi −1,N + 3γT∞ )

2
−
∂Ti,N 2
(3k γ + 2h2J γ )
=
∂qɺ A

 
1  2h2J γ (−4 γTi,N −1 + γTi,N −2 − βTi +1,N + βTi −1,N + 3γT∞2 ) 
= ( ) −  (31)
A  (3k γ + 2h J γ )2 
 2 

Therefore, the gradient direction given by Eq. (26) will be

14
  
∂J M
C  2h2J γ (−4 γTi,N −1 + γTi,N −2 − βTi +1,N + βTi −1,N + 3γT∞2 ) 
= 2A∑ (Ti,N − T(i,N ) )( ) − 
∂qɺ i =1
m A  (3k γ + 2h J γ )2 
 2 

By simplifying, we get

M
 2h J γ (−4 γT + γTi,N −2 − βTi +1,N + βTi −1,N + 3γT∞ ) 
∂J  2 i,N −1 2 
= 2∑ (Ti,N − T(i,N ) )C −
∂qɺ i =1
m  (3k γ + 2h2J γ ) 2 
 
(32)

Consequently, the sensitivity coefficients can be considered as

2h2J γ (−4 γTi,N −1 + γTi ,N −2 − βTi +1,N + βTi −1,N + 3γT∞ )

−C 2
(33)
2
(3k γ + 2h2J γ )

which is the same as ones for the estimation of thermal conductivity. The optimization
algorithm for the estimation of the heat flux is the same as the one for the estimation of
the thermal conductivity except for the calculation of the objective function which
should be performed from Eq. (25). For calculating A , the variables J , γ , and β are
calculated from the grid generation step, and Tξ and Tβ are calculated from the direct

solution of heat conduction equation (all variables are calculated at surface Γ1 where
the heat flux qɺ is applied).

3.6 Estimation of the heat transfer coefficient h :

The inverse problem of estimating the heat transfer coefficient can be expressed as

 
min  J := C TΓ − TΓ 2
: ∇2T = 0 in Ω, BCs in Eqs.(2)&(3) (34)
h at Γ 2 2m 
 2 

15
The unknown variable is the heat transfer coefficient imposed on the surface Γ2 , h2 .
The value of the heat transfer coefficient has to be numerically optimized for the
objective function given by Eq. (5) to reach a minimum. All steps needed to estimate
the heat transfer coefficient are similar to the ones for estimating the thermal
conductivity except for sensitivity coefficients (Jacobian matrix). Thus only the
sensitivity analysis is given here to avoid repetition.

3.7 Sensitivity analysis ( h )

We wish to calculate the sensitivity of the objective function J defined by Eq. (5) to
the heat transfer coefficient h2 . Therefore, we have

∂J M ∂Ti,N M ∂Ti,N
= 2C ∑ (Ti,N − T(i,N ) ) = 2∑ (Ti,N − T(i,N ) )(C ) (35)
∂h2 i =1
m ∂h2 i =1
m ∂h2

∂Ti,N
The term can be obtained by differentiating Eq. (17) with respect to h2 as
∂h2
follows

∂Ti,N 2kJ γ (−4γTi,N −1 + γTi,N −2 − βTi +1,N + βTi −1,N + 3γT∞ )

2
= (36)
∂h2 2
(3k γ + 2h2J γ )

Thus the sensitivity matrix Ja can be explicitly written as

 ∂T 
 1,N 
 ∂h 
 2 
 ∂T 
 2,N 
 
Ja = C  ∂h2  (37)
 
 ⋮ 
 
 ∂TM ,N 
 
 ∂h 
 2  M ×1

16
Then the conjugate gradient method described previously can be utilized to estimate the
heat transfer coefficient h2 .

4. Results
Test Case: Numerical values of the coefficients involved in the test case are listed in
Table 1. Initially the direct problem is solved using the known values for the thermal
conductivity, the heat flux, and the heat transfer coefficient to obtain the temperature
distribution on the outer surface Γ2 (Fig. 3), Ti,N , which will be used as the simulated

measured temperatures T(i,N ) for the inverse analysis to recover the initial values of
m

the thermal conductivity, the heat flux, and the heat transfer coefficient separately.
W W
Using an inverse analysis, the known (desired) values of kd = 40( ) , qɺd = 1000( ) ,
m.C m2
W
and h2 = 10( ) are to be recovered separately by utilizing three different initial
d m2 .C
W W
guesses k initial = 2,100, 600( ) , qɺinitial = 100, 5000,25000( ) , and
m.C m2
W
h2 = 1, 30, 200( ) . The grid size is 50 × 30 .
initial m2 .C

W W W W W T∞ ( ° C) T∞ ( ° C) T∞ ( ° C)
qɺ( ) k( ) h2 ( ) h3 ( ) h4 ( ) 2 3 4
m2 m.C m2 .C m2 .C m2 .C
1000 40 10 10 10 35 35 35
Table 1 Data used for the test case

17
 T
30 183.719
180.000 55
170.000
160.000
25 150.000
grid size: 50x30 140.000
130.000

temperature
120.000 50
20 110.000
100.000
Y

90.000
80.000
15 70.000
60.000 45
50.000
39.767
10

5 40

-15 -10 -5 0 5 10 15 20 0 10 20 30 40 50
X grid nodes on outer surface

a) b)

Fig. 3 Temperature distribution in irregular physical domain (a) and on outer surface Γ2 (used as Tm for

inverse analysis) (b).

Before proceeding to the estimation of the parameters, the numerical procedure used for
the direct heat conduction should be validated and the independency of the results from
the grid size should be confirmed. As the heat conducting body is irregular, the
validation of results is carried out using another computational method (FEM: software
ANSYS) which is shown in Fig. 4 and the independency of the results from the grid size
is shown in Fig. 5.

18
 1 1
ELEMENTS NODAL SOLUTION
MAY 26 2016 MAY 26 2016
STEP=1
13:06:14 13:10:29
SUB =1
TIME=1
TEMP (AVG)
RSYS=0
SMN =38.8312
SMX =183.904

MX

Y MN

Y Z X

Z X

38.8312 71.0696 103.308 135.546 167.785

54.9504 87.1888 119.427 151.666 183.904

a) b)

Fig. 4 The validation of the implemented numerical method for the direct heat transfer problem using the
finite element analysis software ANSYS. The mesh used (a) and the temperature distribution (b).

T
T
30 30 183.956
182.441 180.000
180.000 170.000
170.000 160.000
160.000 150.000
25 25 grid size: 80x70
grid size: 20x20 150.000 140.000
140.000 130.000
130.000 120.000
120.000 110.000
20 20
110.000 100.000
Y
Y

100.000 90.000
90.000 80.000
80.000 70.000
15 15
70.000 60.000
60.000 50.000
50.000 39.206
40.174
10 10

5 5

-15 -10 -5 0 5 10 15 20 -15 -10 -5 0 5 10 15 20

X X

a) b)

Fig. 5 Three grid sizes 20 × 20 (a), 50 × 30 (Fig. 3), and 80 × 70 (b) are used to check the
independency of results from the grid size.

19
Estimation of the thermal conductivity:

700

600 40

35

30

thermal conductivity
500

objective function
25

20

400 15

10

5
300
0 5 10 15 20 25 30
iteration number

200

100

0
0 5 10 15 20 25 30
iteration number

Fig. 6 Thermal conductivity and objective function versus iteration number for initial thermal

conductivity kinitial = 2(W/ m .C)

1

2000 100

90
objective function

thermal conductivity

80
1500

70

60
1000
50

40
500 0 100 200 300 400 500 600
iteration number

0
0 100 200 300 400 500 600
iteration number

Fig. 7 Thermal conductivity and objective function versus iteration number for initial thermal
conductivity kinitial = 100(W/ m .C)
2

20
 140000

120000
600

100000 500

objective function

thermal conductivity
400

80000
300

60000 200

100

40000
0
0 500 1000 1500 2000
iteration number
20000

0 500 1000 1500 2000 2500

iteration number

Fig. 8 Thermal conductivity and objective function versus iteration number for initial thermal
conductivity kinitial = 600(W/ m.C)
3

Estimation of the heat flux:

200000

3500

3000

150000
objective function

2500
heat flux

2000

1500

100000 1000

500

0
0 10 20 30 40
iteration number
50000

0 10 20 30 40
iteration number

Fig. 9 Heat flux and objective function versus iteration number for initial heat flux qɺinitial = 100(W/ m2 )
1

21
 3E+07

5000
2.5E+07 4500

4000

3500

heat flux
objective function
2E+07
3000

2500

2000
1.5E+07
1500

1000

0 10 20 30 40 50
1E+07 iteration number

5E+06

0 10 20 30 40 50
iteration number

Fig. 10 Heat flux and objective function versus iteration number for initial heat flux

qɺinitial = 5000(W/ m2 )
2

5.5E+09

5E+09 25000

4.5E+09 20000
heat flux

4E+09 15000
objective function

10000
3.5E+09
5000
3E+09
0

2.5E+09 0 20 40 60 80
iteration number
100 120

2E+09

1.5E+09

1E+09

5E+08

0 20 40 60 80 100 120
iteration number

Fig. 11 Heat flux and objective function versus iteration number for initial heat flux

qɺinitial = 25000(W/ m2 )
3

22
Estimation of the heat transfer coefficient:

2E+06

10

1.5E+06

heat transfer coefficient

8

objective function
6

1E+06 4

0
500000 0 5 10
iteration number

2 4 6 8 10 12
iteration number

Fig. 12 Heat transfer coefficient and objective function versus iteration number for initial heat transfer

coefficient h2 = 1(W/ m 2 .C)

initial1

60000

50000 30
heat transfer coefficient

25
objective function

40000

20

30000
15

20000
10

2 4 6 8 10 12
iteration number
10000

2 4 6 8 10 12
iteration number

Fig. 13 Heat transfer coefficient and objective function versus iteration number for initial heat transfer

coefficient h2 = 30(W/ m 2 .C)

initial2

23
 1.4E+06

1.2E+06
200

1E+06

heat transfer coefficient

objective function
150

800000
100

600000 50

400000 0
0 50 100 150 200 250
iteration number

200000

0
0 50 100 150 200 250
iteration number

Fig. 14 Heat transfer coefficient and objective function versus iteration number for initial heat transfer

coefficient h2 = 200(W/ m 2 .C)

initial 3

Using the proposed numerical procedure, the parameters initially used in the given
steady state heat conduction problem are recovered accurately. At this stage, the
parameters are recovered separately. For example, when the thermal conductivity is
being recovered, the other two parameters ( qɺ and h2 ) are assumed known. Later, the
simultaneous estimation of the parameters will be considered. Fig. 6 to Fig. 8 show the
convergence history of the thermal conductivity and the minimization of the respective
objective function using three different initial guesses for the thermal conductivity. Fig.
9 to Fig. 11 represent the convergence history of the heat flux and the minimization of
the respective objective function and Fig. 12 to Fig. 14 illustrate the convergence history
of the heat transfer coefficient and the minimization of the respective objective function.
The summaries of the results are given in Table 2 to Table 4 which reveal that the
W
initially used values for the thermal conductivity, kd = 40( ) , the heat flux,
m.C
W W
qɺd = 1000( ) , and the heat transfer coefficient, h2 = 10( ) , are recovered
m2 d m2 .C
accurately and efficiently and the respective objective function values approach zero

24
thereby confirming the accuracy of the proposed method. The results are obtained by a
FORTRAN compiler and computations are run on a PC with Intel Core i5 and 6G

RAM. A tolerance of 10−8 is used in iterative loops to increase the accuracy of results.

Grid size Desired Initial Final value of Initial Minimum Computation time Percentage
value of (guess) value value of value of J of
C
k value of of k reduction
J
in
k
objective
function
5s (29 iterations)
50 × 30 40.00 2.00 40.00 1 616.63 3.78 × 10−6 100%

100.00 40.00 10 2286.28 1.34 × 10−11 20s (662 iterations) 100%

600.00 40.00 100 131928.21 1.04 × 10−9 2m:42s (2252 iterations) 100%

Table 2 Results for the estimation of the thermal conductivity.

Grid size Desired Initial Final value of Initial value of Minimum Computation time Percentage
value of (guess) value value of J of
C J
qɺ value of of qɺ reduction
in
qɺ
objective
function
8s (43 iterations)
50 × 30 1000.00 100 1000.00 1 27861.20 8.98 × 10−9 100%

−7 10s (47 iterations)

5000 1000.00 1 27517161.14 8.78 × 10 100%

−6 29s (107 iterations)

25000 1000.00 1 4953088782.28 2.71 × 10 100%

Table 3 Results for the estimation of the heat flux.

25
Grid size Desired Initial Final value of Initial Minimum Computation time Percentage
value of (guess) value value of J value of J of
C
value of of h2 reduction
h2
in
h2
objective
function

50 × 30 10.00 1.00 10.00 10 1857861.78 2.02 × 10−10 2s (12 iterations) 100%

30.00 10.00 10 55370.70 1.54 × 10−11 2s (12 iterations) 100%

200.00 10.00 100 1221367.83 3.06 × 10−10 20s (241 iterations) 100%

Table 4 Results for the estimation of the heat transfer coefficient

5. Effect of measurement error

For the case of no measurement error, the desired parameters were recovered accurately.
However, the temperature measurements contain errors which can be expressed as

Ti,meas
N
= Ti,exact
N
+ λTi,exact
N
(38)

where λ is the random error of measurement. Here three random errors of 1%, 2%, and

5% are considered in Test Case to investigate the effect of the temperature

measurement error on the estimation of the three parameters (the thermal conductivity,

the heat flux, and the heat transfer coefficient). The optimal values for the three

parameters and the respective errors and the minimum values for the objective function

are listed in Table 5 and the minimization of the objective function value for each

parameter estimation problem in the presence of the measurement errors of 1%, 2%, and

5% are shown in Fig. 15 to Fig. 17. For the test case considered in this study, the worst

error occurs in the estimation of the thermal conductivity. As can be seen in Table 5,

26
the error in the estimation of the thermal conductivity varies from 17.75% to 97.28%

corresponding to the temperature measurement errors of 1% and 5%, respectively. The

significant errors in the estimation of k can be related to the nature of the measurement

error definition (Eq. (38)). As can be seen in Table 5 (for the estimation of k ), the

minimum values of the objective function for the measurement errors of 1% and 5% are

obtained as Jmin =56.09 and Jmin =1588.17 , respectively. These imply that the more

optimal values for k cannot exist (the existence of a solution) to reduce completely the

mismatch between the estimated temperatures and the simulated measured ones.

Temperature Estimated Estimated Estimated Error in Error in Error in

measurement value of k value of qɺ value of h2 estimated estimated estimated
error value of k value of qɺ value of h2
0% 40.00 1000.00 10.00 0% 0% 0%
1% 47.10 1030.55 9.66 17.75% 3.06% 3.4%
( Jmax =1728.96 ( Jmax =647284.03 ( Jmax =75712.10
Jmin =56.09 , Jmin =30.88 , Jmin =6.18 ,
C = 10 ) C =1) C = 10 )
2% 54.45 1061.10 9.34 36.12% 6.11% 6.6%
( Jmax =2698.86 ( Jmax =607939.80 ( Jmax =69727.37
Jmin =231.90 , Jmin =123.53 , Jmin =24.71 ,
C = 10 ) C =1) C = 10 )
5% 78.91 1152.75 8.48 97.28% 15.28% 15.2%
( Jmax =7169.54 ( Jmax =497712.05 ( Jmax =53334.18
Jmin =1588.17 , Jmin =772.05 , Jmin =154.23 ,
C = 10 ) C =1) C = 10 )

Table 5 Estimation of the parameters and the respective errors.

27
 2000 2500

m.e.=2%

objective function
2000
m.e.: measurement error
1500

1000

500
1500

objective function
0
0 5 10 15 20
iteration number

7000

1000 m.e.=5%
6000

objective function
5000

4000
m.e.=1%
3000
500
2000

0 2 4 6 8 10 12
iteration number

0
5 10 15 20 25 30
iteration number

Fig. 15 The decrease of the objective function value in estimation of the thermal conductivity for the
measurement errors of 1%, 2%, and 5%.

800000
600000

m.e.=2%
500000

700000
objective function

400000
m.e.: measurement error
300000

600000 200000
objective function

100000

500000 0

0 20 40 60 80
iteration number

550000

400000 500000

450000
m.e.=5%
400000
objective function

350000
300000 300000

250000

200000

200000 150000

100000

50000

0
100000 m.e.=1% 0 20 40 60 80
iteration number

0 20 40 60 80 100
iteration number

Fig. 16 The decrease of the objective function value in estimation of the heat flux for the measurement
errors of 1%, 2%, and 5%.

28
 80000
70000

60000 m.e.=2%
m.e.: measurement error

objective function
50000

40000

30000
60000

objective function
20000

10000

0 2 4 6 8 10
iteration number

40000
50000
m.e.=5%
40000

objective function
30000

20000

20000
10000

m.e.=1% 0

1 2 3 4 5 6 7
iteration number

2 4 6 8 10 12 14 16
iteration number

Fig. 17 The decrease of the objective function value in estimation of the heat transfer coefficient for the
measurement errors of 1%, 2%, and 5%.

6. Simultaneous estimation
In this section, the simultaneous estimation of two parameters is considered. The two
parameters may be any combination of three parameters k , qɺ , and h2 . Here we
considerer only two parameters k and h2 . The simultaneous estimation can be
performed by implementing the sensitivity analysis presented earlier for each parameter
estimation problem in the iterative loops of the proposed algorithm. The convergence
history of the thermal conductivity and the heat transfer coefficient and the
minimization of the objective function value for two different initial guesses for the
parameters k and h2 are shown in Fig. 18 and Fig. 19. The estimated parameters are in
good agreement with the desired ones (Table 6). However, there is no guarantee that
the values for the two parameters used in the initial heat conduction problem (used for
determining the simulated measured temperatures) can be accurately recovered in the
inverse problem. The estimation of the two parameters depends significantly on the
29
initial guesses for them. There are many possibilities for the two parameters values
which result in the same (or close to) outer surface temperature distribution as the two
parameters values used in the initial heat conduction problem do. Therefore, the inverse
analysis results for the simultaneous estimation of the parameters are not unique (they
converge to a local optimum).

40
400000
35
thermal conductivity

350000 30

25
300000
objective function

20

0 50 100 150
250000 iteration number

200000 20

18
heat transfer coefficient

150000 16

14

100000
12

50000 10
0 50 100 150
iteration number

0
0 50 100 150
iteration number

Fig. 18 The convergence history of the thermal conductivity and the heat transfer coefficient and the
minimization of the objective function value.

30
 45

40

thermal conductivity
35

1E+06 30

25

20

15
800000

objective function
10
0 20 40 60 80
iteration number

600000
50

45

heat transfer coefficient

40

400000 35

30

25

20

200000 15

10
0 20 40 60 80
iteration number

0
0 20 40 60 80 100
iteration number

Fig. 19 The convergence history of the thermal conductivity and the heat transfer coefficient and the
minimization of the objective function value.

Initial Initial Final value Initial Final value Value Initial Minimum Reduction
guess (guess) of k (guess) of h2 of C value value of J in objective
number value of k value of h2 of J function
1 20.00 39.26 20.00 9.98 100 356861.87 5.56 ~100%
(desired (desired
value=40.00) value=10.00)
2 10.00 46.18 50.00 10.16 100 910442.93 343.72 99.96%
(desired (desired
value=40.00) value=10.00)

Table 6 The results for simultaneously estimating k and h2 .

31
 7. Conclusion
This paper presented a novel numerical procedure to estimate the thermal conductivity,
the heat transfer coefficient, and the heat flux in a two dimensional irregular conducting
body using an inverse steady state heat conduction analysis. The irregular body was
transformed into a regular computational domain to perform all computations related to
the direct and inverse heat conduction solution. To do this, an elliptic grid generation
scheme was used to generate a grid over the body. A novel sensitivity analysis scheme
was proposed to calculate sensitivity coefficients needed in a gradient based
optimization method. An explicit expression for the sensitivity coefficients was derived
which allows for the computation of the sensitivity coefficients in one single solve,
regardless of the number of unknown quantities. The conjugate gradient method was
used as a tool to minimize the objective function and recover the desired quantities. The
obtained results revealed that the proposed algorithm is very accurate and efficient.

8. References
[1] Beck J.V., Blackwell B., Clair C.R.S., Inverse Heat Conduction: Ill-Posed Problems,
Wiley, 1985.
[2] Özisi̧k M., Orlande H., Inverse heat transfer: fundamentals and applications, Taylor
& Francis 2000.
[3] Artyukhin E., Reconstruction of the thermal conductivity coefficient from the
solution of the nonlinear inverse problem, Journal of Engineering Physics and
Thermophysics, 41 (1981) 1054-1058.
[4] Alifanov O.M., Tryanin A.P., Determination of the coefficient of internal heat
exchange and the effective thermal conductivity of a porous solid on the basis of a
nonstationary experiment, Journal of Engineering Physics, 48 (1985) 356-365.
[5] Dantas L., Orlande H., A function estimation approach for determining temperature-
dependent thermophysical properties, Inverse Problems in Engineering, 3 (1996) 261-
279.
32
[6] Jurkowski T., Jarny Y., Delaunay D., Estimation of thermal conductivity of
thermoplastics under moulding conditions: an apparatus and an inverse algorithm,
International journal of heat and mass transfer, 40 (1997) 4169-4181.
[7] Yang C.-y., A linear inverse model for the temperature-dependent thermal
conductivity determination in one-dimensional problems, Applied Mathematical
Modelling, 22 (1998) 1-9.
[8] Divo E., Kassab A.J., Kapat J.S., Chyu M.-K., Retrieval of multidimensional heat
transfer coefficient distributions using an inverse BEM-based regularized algorithm:
numerical and experimental results, Engineering Analysis with Boundary Elements, 29
(2005) 150-160.
[9] Sawaf B., Ozisik M.N., Jarny Y., An inverse analysis to estimate linearly
temperature dependent thermal conductivity components and heat capacity of an
orthotropic medium, International Journal of Heat and Mass Transfer, 38 (1995) 3005-
3010.
[10] Liu F.-B., A hybrid method for the inverse heat transfer of estimating fluid thermal
conductivity and heat capacity, International Journal of Thermal Sciences, 50 (2011)
718-724.
[11] Zhang J., Delichatsios M.A., Determination of the convective heat transfer
coefficient in three-dimensional inverse heat conduction problems, Fire Safety Journal,
44 (2009) 681-690.
[12] Chen W.-L., Yang Y.-C., Lee H.-L., Inverse problem in determining convection heat
transfer coefficient of an annular fin, Energy Conversion and Management, 48 (2007)
1081-1088.
[13] Cui M., Gao X., Zhang J., A new approach for the estimation of temperature-
dependent thermal properties by solving transient inverse heat conduction problems,
International Journal of Thermal Sciences, 58 (2012) 113-119.
[14] Mierzwiczak M., Kołodziej J.A., The determination temperature-dependent thermal
conductivity as inverse steady heat conduction problem, International journal of heat
and mass transfer, 54 (2011) 790-796.
[15] Czél B., Gróf G., Inverse identification of temperature-dependent thermal
conductivity via genetic algorithm with cost function-based rearrangement of genes,
International Journal of Heat and Mass Transfer, 55 (2012) 4254-4263.

33
[16] Mohebbi F., Sellier M., Estimation of thermal conductivity, heat transfer
coefficient, and heat flux using a three dimensional inverse analysis, International
Journal of Thermal Sciences, 99 (2016) 258-270.
[17] Reddy B.K., Gnanasekaran N., Balaji C., Estimation of thermo-physical and
transport properties with Bayesian inference using transient liquid crystal thermography
experiments, Journal of Physics: Conference Series, 395 (2012) 012082.
[18] Singla R.K., Das R., Application of decomposition method and inverse prediction of
parameters in a moving fin, Energy Conversion and Management, 84 (2014) 268-281.
[19] Beck J.V., Surface heat flux determination using an integral method, Nuclear
Engineering and Design, 7 (1968) 170-178.
[20] Huang C.H., Wang S.P., A three-dimensional inverse heat conduction problem in
estimating surface heat flux by conjugate gradient method, International journal of heat
and mass transfer, 42 (1999) 3387-3403.
[21] Liu F.-B., Inverse estimation of wall heat flux by using particle swarm optimization
algorithm with Gaussian mutation, International Journal of Thermal Sciences, 54 (2012)
62-69.
[22] Yang Y.-C., Chen W.-L., A nonlinear inverse problem in estimating the heat flux of
the disc in a disc brake system, Applied Thermal Engineering, 31 (2011) 2439-2448.
[23] Moparthi A., Das R., Uppaluri R., Mishra S.C., Optimization of Heat Fluxes on the
Heater and the Design Surfaces of a Radiating-Conducting Medium, Numerical Heat
Transfer, Part A: Applications, 56 (2009) 846-860.
[24] Hsieh C.K., Kassab A.J., A general method for the solution of inverse heat
conduction problems with partially unknown system geometries, International Journal
of Heat and Mass Transfer, 29 (1986) 47-58.
[25] Dulikravich G., Martin T., Inverse shape and boundary condition problems and
optimization in heat conduction, Advances in Numerical Heat Transfer, 1 (1996) 381-
426.
[26] Mohebbi F., Sellier M., Optimal Shape Design in Heat Transfer Based on Body-
Fitted Grid Generation, International Journal for Computational Methods in
Engineering Science and Mechanics, 14 (2013) 227-243.

34
 [27] Mohebbi F., Sellier M., Three-Dimensional Optimal Shape Design in Heat Transfer
Based on Body-fitted Grid Generation, International Journal for Computational
Methods in Engineering Science and Mechanics, 14 (2013) 473-490.
[28] Sarvari S.M.H., Optimal Geometry Design of Radiative Enclosures Using the
Genetic Algorithm, Numerical Heat Transfer, Part A: Applications, 52 (2007) 127-143.
[29] Das R., Prasad D.K., Prediction of porosity and thermal diffusivity in a porous fin
using differential evolution algorithm, Swarm and Evolutionary Computation, 23 (2015)
27-39.
[30] Chopade R.P., Agnihotri E., Singh A.K., Kumar A., Uppaluri R., Mishra S.C.,
Mahanta P., Application of a Particle Swarm Algorithm for Parameter Retrieval in a
Transient Conduction-Radiation Problem, Numerical Heat Transfer, Part A:
Applications, 59 (2011) 672-692.
[31] Thompson J.F., Thames F.C., Mastin C.W., Automatic numerical generation of
body-fitted curvilinear coordinate system for field containing any number of arbitrary
two-dimensional bodies, Journal of Computational Physics, 15 (1974) 299-319.
[32] Thompson J., Warsi Z., Mastin C., Numerical grid generation: foundations and
applications, North-Holland, 1985.
[33] Özişik M., Finite difference methods in heat transfer, CRC Press, 1994.
[34] Mohebbi F., Optimal shape design based on body-fitted grid generation (PhD
thesis-link: http://hdl.handle.net/10092/9427), in, University of Canterbury 2014.
[35] Polak E., Ribiere G., Note sur la convergence de méthodes de directions conjuguées,
Revue Française d’Informatique et de Recherche Opérationnelle, 16 (1969) 35–43.

35

View publication stats