Journal of Physics D: Applied Physics

You may also like

- Approximate inverse for a one-dimensional
Inverse estimation of the boundary condition in inverse heat conduction problem
P Jonas and A K Louis
three-dimensional heat conduction - Numerical study of the modeling error in
the online input estimation algorithm used
for inverse heat conduction problems
To cite this article: Ching-yu Yang and Cha'o-Kuang Chen 1997 J. Phys. D: Appl. Phys. 30 2209 (IHCPs)
S K Ali, M S Hamed and M F Lightstone

- SOTE: A Nonlinear Method for Magnetic

Topology Reconstruction in Space
Plasmas
View the article online for updates and enhancements. Y. Y. Liu, H. S. Fu, V. Olshevsky et al.

This content was downloaded from IP address 196.248.178.89 on 15/12/2023 at 00:13

J. Phys. D: Appl. Phys. 30 (1997) 2209–2216. Printed in the UK PII: S0022-3727(97)79716-X

Inverse estimation of the boundary

condition in three-dimensional heat
conduction
Ching-yu Yang† and Cha’o-Kuang Chen‡
† Department of Mold and Die Engineering, National Kaohsiung Institute of
Technology, Kaohsiung City, Taiwan 807, Republic of China
‡ Department of Mechanical Engineering, National Cheng Kung University,
Tainan City, Taiwan 700, Republic of China

Received 20 November 1996, in final form 7 May 1997

Abstract. An inverse method based on the symbolic approach is proposed to

determine the boundary condition in three-dimensional inverse heat conduction
problems. This method uses symbols to represent the unknown boundary condition
and then executes an implicit alternating-direction finite-difference method to
calculate the temperature distribution. The calculated results are expressed
explicitly as functions of the symbol of the undetermined boundary condition. Then,
a set of linear equations is constructed from the comparison between the output
symbolic temperature and the measured numerical temperature. Thus, the set of
linear equations is solved through the linear least-squares error method and leads
to the solution of the unknown boundary condition. Results from examples confirm
that the proposed method is applicable in solving the three-dimensional inverse
heat conduction problem. In the example problems, a temporally dependent and a
spatially–temporally dependent boundary condition are used to demonstrate the
procedure of the proposed method. In the first example, the temporally dependent
boundary conduction, the result shows that only one-point measurement is needed
to estimate the temporally dependent condition whether the measurement errors
are considered or not. In the second problem, the four- and eight-point
measurement methods are adopted to estimate the temporally–spatially dependent
boundary condition. The result shows that only the four-point measurement is
needed to estimate the surface temperature when measurement errors are
neglected. When measurement errors are considered, the eight-point
measurement is required in order to increase the congruence of the estimated
results to the exact solutions.

1. Introduction machine cutting, or inside a combustion chamber, or at the

outer surface of a reentry vehicle. Wide attention has been
The potential of the symbolic computation to improve called to the inverse problem, and most studies employ
the analysis of the engineering problems has been numerical methods [9–16] coupled with a nonlinear least-
recognized. The applications of symbolic computation squares formulation to determine the unknown conditions
in the field of analysis include solid mechanics, heat in the inverse problems. In the methods, the inverse
transfer, supersonic transition, gas kinetic, fluid, dynamics, problem is formulated from the finite-difference, the finite-
differential geometry, and control systems [1–4]. However, element, or the boundary element methods to calculate
there are only limited numbers of works that have adopted the responses of the system. In order to determine
symbolic computation into the field of inverse problems the unknown boundary condition, these methods have
[5–8]. often been combined with optimization algorithms such as
The inverse heat conduction problems deal with the the regularization technique, the sequential regularization
determination of the crucial parameters in analysis such as approach, and the conjugate gradient method coupled
the heat transfer coefficient, contact conductance, surface with the adjoint equation approach, the Newton–Raphson
heat flux, internal energy source, thermal properties, method, genetic algorithms, and the multi-dimensional
etc. They have been widely applied in many design simplex method.
and manufacturing problems especially when direct Based on the numerical approach, the determination of
measurements of the surface conditions are not possible. the boundary condition includes two phases: the process
For example, it would be difficult to measure the of the analysis and the process of the optimization. In
temperature or heat flux at the tool–work interface of a the analysis process, the boundary condition is assumed

0022-3727/97/152209+08$19.50 c 1997 IOP Publishing Ltd 2209

Ching-yu Yang and Cha’o-Kuang Chen

and then the temperature field of the differential governing 2. Approach to a three-dimensional heat
equation is solved through the numerical method. Solutions conduction problem based on the proposed
from the above process are integrated with the temperature method
measured at the interior point of the solid. Thus, a nonlinear
least-squares problem is formatted for the process of the 2.1. Problem statement
optimization. In the optimization process, an optimizer
is used to guide the exploring points systematically to Consider a block with constant thermal properties. The
approach the boundary condition. Then the newly guessed isothermal conditions are applied at the face of the block
boundary condition is substituted for the unknown boundary where x̄ = ȳ = z̄ = 0 and x̄ = ȳ = L̄. It is
condition in the following analysis process. As such, initially at a uniform temperature T̄0 and then suddenly
several iterations to solve the governing equation and a temperature function f¯(x̄, ȳ, t¯) is applied to the face
nonlinear optimization problem are needed to obtain the z̄ = L̄. A dimensionless mathematical formulation of the
undetermined boundary condition. Additionally, most three-dimensional heat conduction problem is presented as
inverse problems have only been used to deal with one- or follows:
two-dimensional problems but few have included the three- ∂ 2T ∂ 2T ∂ 2T ∂T
dimensional problems because those problems are more + + =
∂x 2 ∂y 2 ∂z 2 ∂t
sophisticated.
The purpose of this research is to demonstrate the use 0 < x, y, z < 1 t >0 (1)
of symbolic computation in the inverse problem to solve T (x, y, z, 0) = T0 = 1
the three-dimensional heat conduction problems. In the
process of the symbolic approach, the boundary condition 0 ≤ x, y, z ≤ 1 t =0 (2)
is represented as a series form first and the boundary T (0, y, z, t) = 1
condition can be assumed as a series form with the x = 1, 0 ≤ y ≤ 1, 0 ≤ z < 1 t >0 (3)
linear combination of the unknown coefficients. Then,
the direct symbolic finite-difference analysis is performed T (1, y, z, t) = 1
to find the temperature field in the solid. The calculated x = 0, 0 ≤ y ≤ 1, 0 ≤ z < 1 t >0 (4)
results of temperature distribution are expressed as symbols, T (x, 0, z, t) = 1
which describe the boundary condition. Because the
calculated temperature field is explicitly given as functions y = 0, 0 < x < 1, 0 ≤ z > 1 t >0 (5)
of the boundary condition, the inverse analysis can be T (x, 1, z, t) = 1
directly achieved by the resulting symbolic expressions and
measurement datum. As a result, the inverse problem y = 1, 0 < x < 1, 0 ≤ z < 1 t >0 (6)
becomes a set of equations with linear combination of T (x, y, 0, t) = 1
the unknown coefficients, which lead to a solution of
z = 0, 0 ≤ x, y ≤ 1 t >0 (7)
the inverse problem through the linear least-squares error
method. The unique feature of this approach is that the T (x, y, 1, t) = f (x, y, t)
inverse computation is in a linear domain and the nonlinear z = 1, 0 ≤ x, y, ≤ 1 t >0 (8)
optimization process used in the traditional approach can be
eliminated. It is not difficult to extend our analysis to more where the following dimensionless quantities are defined:
complicated boundary conditions and geometries without
major reprogramming in this research [17]. Furthermore, x̄ ȳ z̄ T̄
if a simple three-dimensional conduction problem can be x= y= z= T =
L̄ L̄ L̄ T̄0
solved analytically [18], the proposed inverse technique can
be used for a symbolic analytic solution as well as for a
symbolic finite-difference solution. k̄ k̄r t¯
k= t=
This paper includes three more sections. In the k̄r ρ̄ C̄p L̄2
second section, the features of using symbolic computation
in the inverse problems are stated. The characteristics ρ̄ C̄p is the heat capacity per unit volume, T̄0 and k̄r
of the three-dimensional problem are delineated and an refer to the nonzero reference temperature and thermal
implicit alternating-direction finite-difference method is conductivity, respectively. We assume k̄ = k̄r and
presented. Meanwhile, the process of the proposed method f (x, y, t) = f¯(x̄, ȳ, t¯)/T̄0 in the problem.
is illustrated. In the third section, two examples with The inverse problem is to identify the unknown
temporally dependent and temporally–spatially dependent boundary condition f (x, y, t) from the temperature
unknown boundary conditions are employed to demonstrate measurements taken at the grid points of the block. The
the process of the proposed method. A discussion of the proposed method is to formulate a set of equations from
analysed results is also presented in this section. Finally, the calculated symbolic temperature and the measured
the overall contribution and possible applications of this temperature. In order to find the symbolic temperature,
research to the field of inverse heat conduction problems an implicit alternating-direction finite-difference method is
conclude in the fourth section. adopted. It is further described in the following section.

2210
 Three-dimensional heat conduction
∗∗∗
Ti,j,k − Ti,j,k,l ∗ ∗∗ ∗∗∗
= δx2 Ti,j,k + δy2 Ti,j,k + δz2 Ti,j,k (12)
1t/2
Ti,j,k,l+1 − Ti,j,k,l ∗ ∗∗ ∗∗∗
= δx2 Ti,j,k + δy2 Ti,j,k + δz2 Ti,j,k (13)
1t
where
1
δx2 {.}i,j,k,l = [{.}i−1,j,k,l − 2{.}i,j,k,l + {.}i+1,j,k,l ] (14)
1x 2
1
δy2 {.}i,j,k,l = [{.}i,j −1,k,l − 2{.}i,j,k,l + {.}i,j +1,k,l ] (15)
1y 2
1
δz2 {.}i,j,k,l = [{.}i,j,k−1,l − 2{.}i,j,k,l + {.}i,j,k+1,l ] (16)
1z 2
and 1x, 1y, and 1z are the increments of spatial
coordinate and 1t is the increment of temporal domain, i is
the ith grid along the x-coordinate, j is the j th grid along
the y-coordinate, k is the kth grid along the z-coordinate, l
is the lth grid along the temporal coordinate, and Ti,j,k,l is
the temperature at the grid point (i, j, k, l).
∗∗∗
The third intermediate value Ti,j,k can be eliminated
between the equations (10)–(13) to produce an equivalent
∗
system (equations (10), (11), and (17)) in which Ti,j,k
∗∗
and Ti,j,k can be regarded as successive approximations to
Ti,j,k,l at the half time-step:
∗∗
Ti,j,k,l+1 − Ti,j,k ∗ ∗∗
= δx2 Ti,j,k + δy2 Ti,j,k + δz2 Ti,j,k,l+1 . (17)
1t/2

Figure 1. The process of the proposed symbolic approach Let 1x = 1y = 1z and γ = 1t/1x 2 , these
to determine boundary condition in three-dimensional heat equations (10), (11), and (17) take the following form:
conduction problem.  
∗ 1 ∗ ∗
−Ti−1,j,k +2 1+ Ti,j,k − Ti+1,j,k
γ
2.2. An implicit alternating-direction finite-difference  
method with symbolic boundary condition to solve the 2
= − 4 Ti,j,k,l + Ti,j −1,k,l + Ti,j +1,k,l + Ti,j,k−1,l
three-dimensional heat conduction problem γ
+ Ti,j,k+1,l (18)
Suppose that the unknown boundary condition f (x, y, t) is  
represented as the following linear forms with respect to ∗∗ 1 ∗∗ ∗∗
−Ti,j −1,k + 2 1 + Ti,j,k − Ti,j +1,k
am forms in a certain spatial and temporal domain: γ
∗ ∗ ∗
= Ti−1,j,k − 2Ti,j,k + Ti+1,j,k + Ti,j,k−1,l
X
m̄
 
f (x, y, t) = am φm (x, y, t) (9) 2
m=1 + − 2 Ti,j,k,l + Ti,j,k+1,l (19)
γ
 
where φm (x, y, t) can be any linearly independent function 2
in the problem domain. am is the undetermined coefficients. −Ti,j,k−1,l+1 + + 2 Ti,j,k,l+1 − Ti,j,k+1,l+1
γ
m̄ is a positive integer. ∗ ∗ ∗ ∗∗
= Ti−1,j,k − 2Ti,j,k + Ti+1,j,k + Ti,j −1,k
After the boundary condition is represented as the above  
form, an implicit alternating-direction finite-difference 2 ∗∗ ∗∗
+ − 2 Ti,j,k + Ti,j +1,k . (20)
modification of the Crank–Nicolson method given by γ
Ozisik [18] and Carnahan et al [19] is used to execute ∗
The intermediate value Ti,j,k is solved in equation (18)
the analysis process. This algorithm uses a system of and then it is used in equations (19) and (20) and Ti,j,k ∗∗
is
equations with a tridiagonal coefficient matrix, which solved in equation (19) and then it is used in equation (20).
affords a straightforward solution. The method employs Thus it leads to the solution Ti,j,k,l+1 at the end of the whole
four difference equations which use three intermediate
∗ ∗∗ ∗∗∗ time interval 1t. The stability of this procedure is proved to
values Ti,j,k , Ti,j,k , and Ti,j,k .
be unconditionally stable and converge with discretization
∗
Ti,j,k − Ti,j,k,l error O[1t 2 + 1x 2 ] in [19].
∗
= δx2 Ti,j,k + δy2 Ti,j,k,l + δz2 Ti,j,k,l (10) After equation (8) is substituted by equation (9) and
1t/2
equations (1)–(8) are discretized through the implicit
∗∗
Ti,j,k − Ti,j,k,l ∗ ∗∗ alternating-direction finite-difference method, the symbolic
= δx2 Ti,j,k + δy2 Ti,j,k + δz2 Ti,j,k,l (11)
1t/2 temperature can be solved directly.

2211
Ching-yu Yang and Cha’o-Kuang Chen

(a) (a)

(b) (b)
Figure 3. (a) Estimation of the boundary condition
f2 (x , y , t ) at x = 0.5 and y = 0.5 with four measured points.
(b) Estimation of the boundary condition f2 (x , y , t ) at
x = 0.5 and y = 0.5 with eight measured points.

Pm̄
condition is given symbolically (i.e. m=1 am φm (x, y, t)
for f (x, y, t)). Then, the symbolic temperatures Ti,j,k,l
(function of am ) are obtained by the symbolic solver,
which has been described in the previous paragraph. After
the symbolic temperature Ti,j,k,l is calculated, the next
step is to build a set of linear equations in which the
calculated temperature is equal to the measured temperature
i,j,k,l
(Ti,j,k,l = Tmeasurement ). Thus, this set of equations can
be solved through the linear least-squares error method.
A special feature of the proposed process is that the
(c) inverse computation is in a linear domain and the nonlinear
Figure 2. (a) Estimation of the boundary condition at optimization process used in the traditional approach can
surface z = 1 when the measured point is taken at x = 0.5, be eliminated.
y = 0.9, and z = 0.6. (b) Estimation of the boundary
condition at surface z = 1 when the measured point is
taken at x = 0.5, y = 0.9, and z = 0.7. (c) Estimation of the 3. Results and discussions
boundary condition at surface z = 1 when the measured
point is taken at x = 0.5, y = 0.9, and z = 0.8. In this section, problems defined from equations (1)–(8) are
used as examples to estimate the boundary condition, which
2.3. The method to determine boundary conditions includes the temporally dependent and the temporally–
The method to determine boundary conditions can be best spatially dependent forms. In example 1, the undetermined
illustrated by figure 1. In this process, the input boundary boundary condition is expressed by a time-varying function

2212
 Three-dimensional heat conduction

(a)

(b)
Figure 4. (a) The exact boundary condition f2 (x , y , t ) at t = 0.15. (b) Estimation of the boundary condition f2 (x , y , t ) at
t = 0.15 with four-point measurement when the measurement error is one per cent. (c) Estimation of the boundary condition
f2 (x , y , t ) at t = 0.15 with eight-point measurement when the measurement error is five per cent.

2213
Ching-yu Yang and Cha’o-Kuang Chen

(c)
Figure 4. (Continued)

Table 1. The comparison between the exact solutions and the estimated results without measurement errors in example 2.

Coefficients Exact solution 4-point measurement 8-point measurement

a1 1.000 000 000 000 000 0.999 999 999 999 994 0.999 999 999 999 999
a2 1.000 000 000 000 000 1.000 000 000 000 006 0.999 999 999 999 999
a3 1.000 000 000 000 000 0.999 999 999 999 448 1.000 000 000 000 004
a4 1.000 000 000 000 000 1.000 000 000 000 503 0.999 999 999 999 998
a5 1.000 000 000 000 000 0.999 999 999 999 998 1.000 000 000 000 003
a6 1.000 000 000 000 000 1.000 000 000 000 541 0.999 999 999 999 995
a7 1.000 000 000 000 000 0.999 999 999 999 504 0.999 999 999 999 999
a8 1.000 000 000 000 000 0.999 999 999 999 997 0.999 999 999 999 997
a9 1.000 000 000 000 000 1.000 000 000 000 003 1.000 000 000 000 000
a10 1.000 000 000 000 000 1.000 000 000 000 016 1.000 000 000 000 005

in which one-point measurement with different locations be shown in the following equation:
is taken. In example 2, a polynomial is used to
express the temporally–spatially dependent function in Tmeasurement = Texact + λTexact (21)
which four-point or eight-point measurements are taken
independently to estimate the boundary condition of the |λ| ≤ ω (22)
problem. The exact temperature and boundary condition where λ is the random error of measurement, and ω is
used in the following examples are preselected so that the bound of the λ. Texact in equation (21) is the exact
these functions can satisfy equations (1)–(8). The temperature and Tmeasurement is the measured temperature
accuracy of the proposed method is assessed by comparing at the grid points. The time domain in all cases is from
the estimated boundary condition with the preselected 0 to 0.25 (tf = 0.25) with 0.01 increment. As well, the
boundary condition. Meanwhile, the simulated temperature increment of spatial coordinates is 0.1. The results of the
measurement is generated from the preselected exact estimation of the unknown condition from the knowledge
temperature in each problem and it is presumed to have of the temperature at measurement points are examined.
measurement errors. In other words, the random errors of Detailed descriptions for the example problems are
measurement are added to the exact temperature. This can shown in the following.

2214
 Three-dimensional heat conduction

Example 1. The unknown boundary condition is In the second case, eight thermocouples are allocated at the
presented as the following time-varying function following coordinates:

fˆ1 (t) = 5 sin(4t) + 2 exp(4t) + 1280t 4 . (23) x = 0.1 y = 0.1 z = 0.8

Temperatures in equation (18) are calculated and measured x = 0.1 y = 0.5 z = 0.8
at the interior point (x, y, z) = (0.5, 0.9, 0.6), (0.5, 0.9, x = 0.1 y = 0.9 z = 0.8
0.7), or (0.5, 0.9, 0.8) in this example.
The estimated condition is presumed in the following x = 0.5 y = 0.9 z = 0.8
form: x = 0.5 y = 0.9 z = 0.8
 2  3  4
t t t t
f1 (t) = a1 + a2 + a3 + a4 + a5 x = 0.9 y = 0.1 z = 0.8
tf tf tf tf
 5  6 x = 0.9 y = 0.5 z = 0.8
t t
+ a6 + a7 . (24) x = 0.9 y = 0.9 z = 0.8.
tf tf
The estimated results from the measured point In example 2, the results without measurement
(x, y, z) = (0.5, 0.9, 0.6), (0.5, 0.9, 0.7), or (0.5, 0.9, 0.8) errors are tabulated in table 1. Both cases have
are shown in figures 2(a) to 2(c) respectively. The estimated good approximations when measurement errors are not
results have excellent approximations when measurement considered. The estimated results at spatial coordinates
errors are not considered. When measurement errors are x = 0.5 and y = 0.5 are shown in figure 3(a) and 3(b).
considered (i.e., ω = 5%), the approximations of the The exact solution at temporal coordinate t = 0.15 is shown
estimated solutions vary in different measured location. In in figure 4(a). The estimated results at temporal coordinate
general, the further the measurements are taken from the t = 0.15 are shown in figure 4(b) and 4(c). In general, large
unknown boundary, the further the estimated results deviate errors make the estimated results diverge from the exact
away from the exact solution. For example, compare the solution. For example, the results shown in figure 3(b)
estimated results measured at t = 0.01 with the true value have the maximum deviation from their exact solutions
of the boundary conduction. In figure 2(a), when the when the error is equal to 5%. Moreover, it appears that
measurement is taken at x = 0.5, y = 0.9, and z = 0.6, the the results in figure 3(a) have a similar trend as those in
deviation from the true value is 1.7880. In figure 2(b), when figure 3(b).
measured at x = 0.5, y = 0.9, and z = 0.7, the deviation From the results for both cases, it is evident that there
is 0.588 62. In figure 2(c), when measured at x = 0.5, is no difference in the two individual measured methods
y = 0.9, and z = 0.8, the deviation is 0.104 22. From when there are no measurement errors. However, the
the results, it is evident that the largest deviation is when difference becomes influential when measurement errors are
the measured point is at x = 0.5, y = 0.9, and z = 0.6 considered. Generally speaking, the deviations in the four-
and the smallest deviation is when the measured point is at point measurement are unacceptable when the error appears
x = 0.5, y = 0.9, and z = 0.8. (see figure 3(a) and 4(b)). In the eight-point measurement,
Example 2. The problem described by equations (1)–(8) even when the error is equal to 5%, the deviations are
with the given boundary condition fˆ2 (x, y, t) is a smaller than those in the four-point method at ω = 1%
polynomial form with spatial and temporal variables: (see figures 3(a), 3(b), 4(b), and 4(c)). It is concluded that
the deviations in the eight-point measurement give a better
 2
t t t t approximation for the result when measurement errors are
ˆ
f2 (x, y, t) = 1+ +x+y+ +x 2 +y 2 +x +y +xy.
tf tf tf tf included in the problem.
(25)
The unknown boundary condition f2 (x, y, t) is a poly-
nomial form with spatial and temporal variables and can 4. Conclusion
be approximated by the following form:
 2 The proposed symbolic approach has been introduced for
t t
f2 (x, y, t) = a1 + a2 + a3 x + a4 y + a5 solving the three-dimensional inverse conduction problems.
tf tf An implicit alternating-direction finite-difference method is
t t employed to solve the symbolic temperature field when
+a6 x 2 + a7 y 2 + a8 x + a9 y + a10 xy. (26)
tf tf the boundary condition is represented as symbols. A
In the first case, four measured points are allocated at the special feature of the proposed method is that the inverse
following coordinates: computation is in a linear domain and the nonlinear
optimization process used in the traditional approach can
x = 0.1 y = 0.1 z = 0.8 be eliminated. Two examples have been used to show the
usage of the proposed method. In the first example, the
x = 0.1 y = 0.9 z = 0.8 result shows that only one-point measurement is needed to
estimate the temporally dependent condition whether the
x = 0.9 y = 0.1 z = 0.8
measurement errors are considered or not. In the second
x = 0.9 y = 0.9 z = 0.8. problem, the result shows that only four-point measurement

2215
Ching-yu Yang and Cha’o-Kuang Chen

is needed to estimate the surface temperature when [7] Reuber M, Kornegay J and Melehy E 1987 Symbolic
measurement errors are neglected. When measurement processing as an aid to the solution of inverse design
errors are considered, eight-point measurement is required problems: application Proc. 2nd Int. Conf. on Inverse
Design Concepts and Optimization in Engineering
in order to increase the congruence of the estimated results Sciences ed G S Dulikravich (Penn State University)
to the exact solutions. Finally, the proposed method is pp 405–24
applicable to the other kinds of inverse problem such as [8] Yang Ching-yu 1997 A symbolic approach to estimate
initial estimation and source strength estimation in the two-sided boundary conditions in two-dimensional
three-dimensional heat conduction problems. conduction problems AIAA J. Thermophys. Heat Transfer
11 July/September issue
[9] Stolz G Jr 1960 Numerical solutions to an inverse problem
of heat conduction for simple shapes J. Heat Transfer 82
References 20–6
[10] Sparrow E M, Haji-Sheikh A and Lundgren T S 1964 The
inverse problem in transient heat conduction J. Appl.
[1] Noor A K (ed) 1990 Symbolic Computations and Their Mech. 86 369–75
Impact on Mechanics 1st edn (New York: American [11] Tikhonov A N and Arsenin V Y 1977 Solutions of
Society of Mechanical Engineering) Ill-posed Problems 1st edn (Washington, DC: Winston)
[2] Grossman R (ed) 1989 Symbolic Computation Applications [12] Beck J V, Blackwell B and St Clair C R 1985 Inverse Heat
and Scientific Computing 1st edn (Philadelphia, PA: Conduction—Ill Posed Problem 1st edn (New York:
Society for Industrial and Applied Mathematics) Wiley)
[3] Noor A K and Anderson C M 1979 Computerized [13] Jarny Y, Ozisik M N and Bardon J P 1991 A general
symbolic manipulation in structural mechanics of optimization method using adjoint equation for solving
progress and potential Comput. Struct. 10 multidimensional inverse heat conduction Int. J. Heat
95–118 Mass Transfer 34 2911–9
[4] Jenk R D (ed) 1976 Proc. 1976 ACM Symp. on Symbolic [14] Hensel E 1991 Inverse Theory and Applications for
and Algebraic Computation, SYMSAC’76 1st edn (New Engineers 1st edn (Englewood Cliffs, NJ: Prentice-Hall)
York: Association for Computing Machinery) [15] Morozov V A and Stessin M 1993 Regularization Methods
[5] Yoda K 1993 A method for solving inverse boundary-value for Ill-posed Problems 1st edn (Boca Raton, FL:
problems using symbolic computation Inverse Problems Chemical Rubber Company)
in Engineering Mechanics, Int. Union Theor. Appl. [16] Alifanov O M 1994 Inverse Heat Inverse Problems 1st edn
Mech. Symp. (Toyko, 1992) ed M Tanaka and H D Bui (Berlin: Springer)
(Berlin: Springer) pp 27–30 [17] Read W W and Broadbridge P 1996 Series Solutions for
[6] Reuber M, Kornegay J, Adams W and Barratt C 1987 Steady State Flow in Irregular Porous Domains
Symbolic processing as an aid to the solution of inverse (Transort in Porous Media 22) pp 195–214
design problems: theory Proc. 2nd Int. Conf. on Inverse [18] Ozisik M N 1968 Boundary Value Problems of Heat
Design Concepts and Optimization in Engineering Conduction (Scranton: International Textbook Co)
Sciences ed G S Dulikravich (Penn State University: [19] Carnahan B, Luther H A and Wilkes J O 1977 Applied
Penn State University Press) pp 387–404 Numerical Methods 1st edn (New York: Wiley)

2216