An inverse problem of parameter estimation for heat

and mass transfer in capillary porous media

L.B. Dantas
a
, H.R.B. Orlande
b
, R.M. Cotta
b,
*
a
UNIVAP, Av. Shishima Hifumi, 2911, S~ aao Jos ee dos Campos, SP 12244-000, Brazil
b
Department of Mechanical Engineering, EE/COPPE, Federal University of Rio de Janeiro, UFRJ, Cidade Universit aaria, Cx.
Postal 68503, Rio de Janeiro, RJ 21945-970, Brazil
Received 28 November 2001; received in revised form 30 September 2002
Abstract
This work deals with the solution of an inverse problem of parameter estimation involving heat and mass transfer in
capillary porous media, as described by the dimensionless linear Luikovs equations. The physical problem under
picture involves the drying of a moist porous one-dimensional medium. The main objective of this paper is to simul-
taneously estimate the dimensionless parameters appearing in the formulation of the physical problem by using
transient temperature and moisture content measurements taken inside the medium. The inverse problem is solved by
using the LevenbergMarquardt method of minimization of the least-squares norm with simulated measurements.
2002 Elsevier Science Ltd. All rights reserved.
Keywords: Luikovs equations; Capillary-porous media; Heat and mass transfer; Inverse problem; Parameter estimation; Levenberg
Marquardts method
1. Introduction
The phenomena of coupled heat and mass transfer in
capillary porous media has been drawing the attention
of research groups for a long time, because of its im-
portance in several practical applications that are still
quite relevant nowadays, such as drying and moisture
migration in soils and construction materials. For the
mathematical modeling of such phenomena, Luikov [1]
has proposed a model based on a system of coupled
diusion equations, which takes into account the eects
of the temperature gradient on the moisture migration.
The computation of temperature and moisture con-
tent elds in capillary porous media, from the knowl-
edge of initial and boundary conditions, as well as of the
thermophysical properties appearing in the formulation,
constitutes a direct problem of heat and mass transfer [1
3]. Such type of direct problem, based on Luikovs
theory, was solved analytically through dierent ap-
proaches, including the classical integral transform
technique [2]. Lobo et al. [4] found that the associated
eigenvalue problem, required for the solution with this
technique, had complex eigenvalues that were not ac-
counted for in previous works. The eects on the solu-
tion of including one pair of conjugate complex
eigenvalues were critically addressed by Lobo et al. [4],
while the inclusion of complex eigenvalues of higher
order were discussed in Ref. [5]. The use of the gener-
alized integral transform technique (GITT), with simpler
eigenvalue problems involving analytical eigenfunctions,
can avoid the calculation of complex eigenvalues for
Luikovs formulation. For more details on the use of
such hybrid numericalanalytical technique, the reader
is referred to [68]. We note that numerical techniques
have also been used for the solution of Luikovs system
of equations [911].
Appropriately formulated direct problems are
mathematically classied as well-posed, that is, their
solutions satisfy the requirements of existence, unique-
ness and stability with respect to the input data [1217].
On the other hand, the simultaneous estimation of
thermophysical and boundary condition parameters that
*
Corresponding author.
E-mail address: [email protected] (R.M. Cotta).
0017-9310/03/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved.
PII: S0017- 9310( 02) 00424- 6
International Journal of Heat and Mass Transfer 46 (2003) 15871598
www.elsevier.com/locate/ijhmt
appear in Luikovs formulation, by using temperature
and/or moisture content measurements taken in the
medium, is an inverse problem of coupled heat and mass
transfer [1217]. Generally, inverse problems are math-
ematically classied as ill-posed, that is, their solutions
may not satisfy the requirements of existence, unique-
ness and stability under small perturbations in the input
data [1217]. Despite the ill-posed character, the solu-
tion of an inverse problem can be obtained through its
reformulation in terms of a well-posed problem, such as
a minimization problem associated with some kind of
regularization (stabilization) technique. Dierent meth-
ods based on such an approach have been successfully
used in the past for the estimation of parameters and
functions, in linear and non-linear inverse problems
[1228].
Recently, several articles dealing with the solution of
inverse problems of coupled heat and mass transfer
appeared in the literature [1820,2328]. Boundary in-
verse problems were solved in Refs. [26,27] by using the
conjugate gradient method of function estimation. A
parameter estimation problem was solved in Ref. [28],
involving the identication of the initial condition, the
boundary mass transfer coecient and of two parame-
ters appearing in the denition of the mass diusion
coecient during drying. In [28], the energy conserva-
tion equation was written by using a lumped approach
and Ficks second law described the moisture diusion
process, with a diusion coecient dependent on tem-
perature and moisture content. Refs. [1820,2325] dealt
with the estimation of parameters appearing in Luikovs
formulation, by utilizing the LevenbergMarquardt
method of minimization of the least-squares norm with
temperature measurements. The estimation of moisture
diusivity as a function of temperature and moisture
content was examined by Kanevce et al. [23,24]. Later,
they examined the estimation of other parameters ap-
pearing in the formulation [25]. In Ref. [18], the Kos-
sovitch, Luikov and heat transfer Biot numbers were
simultaneously estimated by using only temperature
measurements in Luikovs linear dimensionless formu-
lation. The eects of lateral heat losses on the accuracy
of such estimated dimensionless numbers were examined
in [19] and, recently, the D-optimum experimental de-
sign [1214] was applied for the selection of the heating
procedure, in order to reduce the condence intervals of
the estimated parameters [20]. We note that the physical
problem considered in [2325,28] involved the drying of
thin innitely long samples of porous materials, heated
by forced convection on both sides. On the other hand,
Nomenclature
Bi
m
dimensionless mass transfer coecient de-
ned by Eq. (2h)
Bi
q
dimensionless heat transfer coecient de-
ned by Eq. (2g)
C measured moisture content
E vector with estimated values for temperature
and moisture content
I number of transient measurements taken
per sensor
J sensitivity coecients
J sensitivity matrix
Ko Kossovitch number dened by Eq. (2i)
L number of unknown parameters
Lu Luikov number dened by Eq. (2e)
M number of temperature sensors
M vector with temperature and moisture con-
tent measured data
N number of moisture content sensors
P vector of unknown parameters
Pn Posnov number dened by Eq. (2f)
Q dimensionless heat ux dened by Eq. (2c)
S weighted least-squares norm dened by Eq.
(3)
W weighting matrix given by Eq. (4c)
X dimensionless position dened by Eq. (2j)
Y measured temperature
Greek symbols
/ dimensionless moisture content dened by
Eq. (2b)
s dimensionless time dened by Eq. (2d)
e phase conversion factor
d thermogradient coecient
h dimensionless temperature dened by Eq.
(2a)
r standard deviation of the temperature mea-
surement errors
r

standard deviation of the moisture content

measurement errors
Subscripts
i refers to the measurement time s
i
l refers to the unknown parameters
n refers to the moisture content sensor num-
ber
m refers to the temperature sensor number
Superscripts
k iteration number
T transpose
1588 L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598
the physical problem considered in [1820] involved the
drying of one-dimensional samples of porous materials,
with one of the sides put into contact with a heater,
while the other was open to the surrounding air.
In this paper, we extend the analysis of our previous
works [1820] and examine the eects of utilizing mea-
surements of moisture content, in addition to tempera-
ture measurements, for estimating simultaneously all
parameters appearing in Luikovs linear dimensionless
formulation. The sensitivity coecients and the deter-
minant of the information matrix are examined in order
to design the experiment and to detect linear dependence
among the unknown parameters. The associated direct
problemis solved with the GITT [3,68]. The Levenberg
Marquardt method [12,13,1825] is applied as the mini-
mization procedure of the least-squares norm, in order to
estimate the parameters. Simulated temperature and
moisture content measurements containing random er-
rors are used in the inverse analysis, as described below.
2. Direct problem
The physical problem involves a one-dimensional
capillary porous medium, initially at uniform tempera-
ture and uniform moisture content. One of the bound-
aries, which is impervious to moisture transfer, is in
direct contact with a heater. The other boundary is in
contact with the dry surrounding air, thus resulting in a
convective boundary condition for both the temperature
and the moisture content, as illustrated in Fig. 1. The
linear system of equations proposed by Luikov [1], with
associated initial and boundary conditions, for the
modeling of such physical problem involving heat and
mass transfer in a capillary porous media, can be written
in dimensionless form as:
ohX; s
os

o
2
hX; s
oX
2
eKo
o/X; s
os
in 0 < X < 1; for s > 0 1a
o/X; s
os
Lu
o
2
/X; s
oX
2
LuPn
o
2
hX; s
oX
2
in 0 < X < 1; for s > 0 1b
hX; 0 0; /X; 0 0; for s 0;
in 0 < X < 1 1c; d
oh0; s
oX
Q;
o/0; s
oX
Pn
oh0; s
oX
0;
at X 0; for s > 0 1e; f
oh1; s
oX
Bi
q
h1; s Bi
q
1 eKoLuBi
m
1 /1; s;
at X 1; for s > 0 1g
o/1; s
oX
Bi

m
/1; s Bi

m
PnBi
q
h1; s 1;
at X 1; for s > 0 1h
where the following dimensionless groups were dened
hX; s
Tx; t T
0
T
s
T
0
; /X; s
u
0
ux; t
u
0
u

;
Q
ql
kT
s
T
0

; s
at
l
2
2ad
Lu
a
m
a
; Pn d
T
s
T
0
u
0
u

; Bi
q

hl
k
;
Bi
m

h
m
l
k
m
; Ko
ru
0
u

cT
s
T
0

; X
x
l
2ej
Bi

m
Bi
m
1 1 ePnKoLu 2k
The properties of the porous medium appearing
above include the thermal diusivity (a), the moisture
diusivity (a
m
), the thermal conductivity (k), the mois-
ture conductivity (k
m
) and the specic heat (c). Other
physical quantities appearing in the dimensionless
groups of Eqs. (2) are the heat transfer coecient (h),
the mass transfer coecient (h
m
), the thickness of porous
medium (l), the prescribed heat ux (q), the latent heat
of evaporation of water (r), the temperature of the sur-
rounding air (T
s
), the uniform initial temperature in the
medium (T
0
), the moisture content of the surrounding
air (u

), the uniform initial moisture content in the me-

dium (u
0
), the thermogradient coecient (d) and the
phase conversion factor (e). Lu, Pn and Ko denote the
Luikov, Posnov and Kossovitch numbers, respectively.
Problem (1) is referred to as a direct problem when
initial and boundary conditions, as well as all para-
meters appearing in the formulation, are known. The
objective of the direct problem is to determine the di-
mensionless temperature and moisture content elds,
hX; s and /X; s, respectively, in the capillary porous
media. The direct problem was solved here by applying
the GITT [3,68,1820]. The GITT is a powerful hybrid
numericalanalytical approach, which has been suc-
cessfully applied to obtain benchmark solutions for dif-
ferent classes of linear and non-linear diusion/
convection problems [3,68]. Such a technique, as ap-
plied to time dependent problems, includes the following
basics steps: (i) choose an auxiliary eigenvalue problem; Fig. 1. Geometry for the drying of a moist porous medium.
L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598 1589
(ii) develop the appropriate transform/inverse formulae
pair; (iii) integral transform the original problem by
substituting the inverse formula into non-transformable
terms or by using the integral balance approach; (iv)
solve the resulting coupled system of ordinary dieren-
tial equations in the time variable; and (v) apply the
inverse formula to the transformed eld in order to
obtain the solution for the original problem. For the
sake of brevity, we omit details of GITT as applied to
the present problem, but they can be readily found in
Refs. [3,68,1820].
3. Inverse problem
For the inverse problem of interest here, the parame-
ters Lu, Pn, Ko, e, Bi
q
and Bi
m
are regarded as unknown
quantities. For the estimation of such parameters, we
consider available the transient temperature measure-
ments Y
im
taken at the locations X
m
, m 1; . . . ; M, as well
as the moisture content measurements C
in
taken at the
locations X

n
, n 1; . . . ; N. The subscript i above refers to
the time when the measurements are taken, that is, t
i
, for
i 1; . . . ; I. We note that the measurements may contain
random errors, but all the other quantities appearing in
the formulation of the direct problem are supposed to be
known exactly. Also, most likely the standard deviation
of the temperature measurements is dierent fromthat of
the moisture content measurements.
Inverse problems are ill-posed [1217]. Several
methods of solution of inverse problems, such as the
one used here, involve their reformulation in terms of
well-posed minimization problems. We assume the mea-
surement errors to be additive, uncorrelated and nor-
mally distributed, with known standard deviation and
zero mean. Since the standard deviation of the mea-
surements is not constant, despite being assumed as
known, the solution of the present inverse problem is
obtained through the minimization of the weighted
least-squares norm in order to obtain minimum variance
estimates [12]. Such a norm can be written as
SP M EP
T
WM EP 3
where P
T
Bi
q
; Bi
m
; Lu; Pn; Ko; e denotes the vector of
unknown parameters and the superscript T above de-
notes transpose. The vector M EP
T
is given by
M EP
T

~
MM
1

~
EE
1
;
~
MM
2

~
EE
2
; . . . ;
~
MM
I

~
EE
I

4a
where
~
MM
i

~
EE
i
is a row vector containing the dier-
ences between measured and estimated temperatures
and moisture contents at the measurement positions X
m
,
m 1; . . . ; M, and X

n
, n 1; . . . ; N, respectively, at time
t
i
, that is,

~
MM
i

~
EE
i
Y
i1
h
i1
; Y
i2
h
i2
; . . . ; Y
iM
h
iM
; C
i1
/
i1
; C
i2
/
i2
; . . . ; C
iN
/
iN
4b
The estimated temperatures and moisture content are
obtained from the solution of the direct problem, Eqs.
(1), at the respective measurement locations, by using
estimates for the unknown parameters P
l
, l 1; . . . ; L.
The weighting matrix W is a diagonal matrix with
elements given by the inverse of the variances of the
measurements [12,13], that is,
W
1
r
2
11
0 0
.
.
.
.
.
.
.
.
.
.
.
.
1
r
2
1M
.
.
.
.
.
.
1
r
2
11
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
r
2
1N
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
r
2
I1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
r
2
IM
.
.
.
.
.
.
1
r
2
I1
.
.
.
.
.
.
.
.
.
.
.
.
0
1
r
2
IN
_

_
_

_
4c
4. Method of solution for the inverse problem
The present inverse problem of parameter estimation
is solved with the LevenbergMarquardt method of
minimization of the least-squares norm [12,13,1825].
Such a method was rst derived by Levenberg [21] in
1944, by modifying the ordinary least-squares norm.
Later, in 1963, Marquardt [22] derived basically the
same technique by using a dierent approach. Marqu-
ardts intention was to obtain a method that would tend
to the Gauss method in the neighborhood of the mini-
mum of the ordinary least-squares norm, and would
tend to the steepest descent method in the neighborhood
of the initial guess used for the iterative procedure [12].
The iterative procedure of the LevenbergMarquardt
method is given by:
P
k1
P
k
J
k

T
WJ
k
l
k
X
k

1
J
k

T
WM EP
k

5
where J
k
is the sensitivity matrix, l
k
is a positive scalar
named damping parameter, X
k
is a diagonal matrix and
the superscript k denotes the iteration number. The
purpose of the matrix term l
k
X
k
appearing in Eq. (5)
is to damp oscillations and instabilities due to the ill-
conditioned character of the problem, by making its
1590 L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598
components large as compared to those of J
T
WJ, if
necessary. The damping parameter is made large in the
beginning of the iterations, so that the matrix J
T
WJ is
not required to be non-singular and the Levenberg
Marquardt method tends to the Steepest Descent meth-
od, that is, a very small step is taken in the negative
gradient direction. The parameter l
k
is then gradually
reduced as the iteration procedure advances to the so-
lution of the parameter estimation problem and then
the LevenbergMarquardt method tends to the Gauss
method [12]. However, if the errors inherent to the
measured data are amplied, generating instabilities on
the solution as a result of the ill-posed character of the
problem, the damping parameter is automatically in-
creased. Such an automatic control of the damping para-
meter makes the LevenbergMarquardt method a quite
robust and stable estimation procedure, so that it does
not require the use of the discrepancy principle in the
stopping criterion to become stable, like the conjugate
gradient method [13,14]. Therefore, usual stopping cri-
teria can be used for the LevenbergMarquardt, such as
i SP
k1
< e
1
6a
ii kJ
k

T
Y TP
k
k < e
2
6b
iii kP
k1
P
k
k < e
3
6c
where e
1
, e
2
and e
3
are user prescribed tolerances and
k k is the vector Euclidean norm.
The criterion given by Eq. (6a) tests if the least-
squares normis suciently small, which is expected in the
neighborhood of the solution for the problem. Similarly,
Eq. (6b) checks if the norm of the gradient of SP is
suciently small, since it is expected to vanish at the
point where SP is minimum. The last criterion given by
Eq. (6c) results fromthe fact that changes in the vector of
parameters are very small when the method has con-
verged. Generally, these three stopping criteria need to be
tested and the iterative procedure of the Levenberg
Marquardt method is stopped if any of them is satised.
We note that, when the measurement errors start to cause
oscillations on the inverse problem solution and the
damping parameter l
k
is automatically increased by
the LevenbergMarquardt method, the increments on the
parameters become very small, so that the iterative pro-
cedure is stopped through the criterion given by Eq. (6c).
The sensitivity matrix is dened as
JP
oE
T
P
oP
_ _T

o
~
EE
T
1
oP
1
o
~
EE
T
1
oP
2
o
~
EE
T
1
oP
3

o
~
EE
T
1
oP
L
o
~
EE
T
2
oP
1
o
~
EE
T
2
oP
2
o
~
EE
T
2
oP
3

o
~
EE
T
2
oP
L
.
.
.
.
.
.
.
.
.
.
.
.
o
~
EE
T
I
oP
1
o
~
EE
T
I
oP
2
o
~
EE
T
I
oP
3

o
~
EE
T
I
oP
L
_

_
_

_
7a
where
o
~
EE
T
i
oP
l

oh
i1
oP
l
oh
i2
oP
l
.
.
.
oh
iM
oP
l
o/
i1
oP
l
o/
i2
oP
l
.
.
.
o/
iN
oP
l
_

_
_

_
; for i 1; . . . ; I and l 1; 2; . . . ; L
7b
The elements of the sensitivity matrix are the sensitivity
coecients. They are dened as the rst derivative of the
estimated quantities with respect to each of the un-
known parameters P
l
, l 1; . . . ; L. The sensitivity coef-
cients are required to be large in magnitude, so that the
estimated parameters are not very sensitive to mea-
surement errors. Also, the columns of the sensitivity
matrix are required to be linearly independent in order
to have the matrix J
T
WJ invertible. In problems in-
volving parameters with dierent orders of magnitude,
the sensitivity coecients with respect to the various
parameters may also dier by several orders of magni-
tude, creating diculties in their comparison and iden-
tication of linear dependence. These diculties can be
alleviated through the analysis of the normalized sensi-
tivity coecients, which are obtained by multiplying the
original sensitivity coecients by the parameters that
they are referred to. Note that the normalized sensitiv-
ity coecients have the units of the measured variable,
either temperature or moisture content; hence, they are
compared as having the magnitude of the measured
variable as a basis.
5. Statistical analysis
After the minimization of the least-squares norm gi-
ven by Eq. (3), a statistical analysis can be performed in
order to obtain condence intervals and a condence
region for the estimated parameters [12]. Condence in-
tervals at the 99% condence level are obtained as:
^
PP
l
2:576r
^
PP
l
6P
l
6
^
PP
l
2:576r
^
PP
l
; l 1; . . . ; L 8
where
^
PP
l
are the values estimated for the unknown pa-
rameters, P
l
, for l 1; . . . ; L, and r
^
PP
l
are the standard
deviations obtained from the covariance matrix of the
estimated parameters.
The condence region can be computed from [12]:

^
PP P
T
V
1

^
PP P 6v
2
L
9
L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598 1591
where v
2
L
is the chi-square distribution for L degrees of
freedom and for a given condence level. Also, V is the
covariance matrix of the estimated parameters, which is
given by [12]:
V J
T
WJ
1
10
We note that Eq. (10) is exact for linear estimation
problems and is approximately used for non-linear pa-
rameter estimation problems such as the one under
picture in this paper, where the sensitivity coecients are
functions of the unknown parameters.
6. Experimental design
The design of optimum experiments is of capital
importance in parameter and in function estimations. It
basically consists in examining a priori some kind of
measure of the accuracy of the estimated quantities in
order to choose experimental variables, such as the
number and location of sensors, experimental duration,
etc., so that minimum variance estimates are obtained
via the inverse analysis. Optimum experiments can be
designed by minimizing the hypervolume of the con-
dence region of the estimated parameters. The minimi-
zation of the hypervolume of the condence region given
by Eq. (9) can be obtained by maximizing the determi-
nant of V
1
, in the so-called D-optimum design [12
14,17].
For the design of optimum experiments, let us con-
sider the weighting matrix W as the identity matrix.
Therefore, the elements F
r;s
, r, s 1; . . . ; L, of the so-
called information matrix F J
T
J, are given by:
F
r;s
J
T
J
r;s

I
i1

M
m1
oh
im
oP
r
oh
im
oP
s
_ _
_

N
n1
o/
in
oP
r
o/
in
oP
s
_ _
_
;
for r; s 1; . . . ; L 11
where I is the number of transient measurements and L
is the number of unknown parameters.
We note that for non-linear estimation problems the
analyses of the sensitivity coecients and of the deter-
minant of F are not global, because these quantities are
functions of the unknown parameters. Therefore, a
priori estimated values for the parameters are required
for the design of optimum experiments. In practical
situations involving non-linear estimation problems, the
optimum design of the experiment is performed in an
iterative manner. Low-accuracy estimates for the un-
known parameters, which can be obtained from previ-
ous experiments or from literature data, are used in the
initial experimental design. As more accurate estimates
are obtained with new experiments, the experimental
design is improved and the accuracy of the estimated
parameters is further enhanced.
7. Results and discussion
For the design of optimum experiments and for the
solution of the present inverse problem of parameter
Table 1
Test-cases examined
Physical variables Wood Ceramics
Test-case 1 Test-case 2 Test-case 3 Test-case 4
l (m) 0.05 0.05 0.05 0.05
q
0
(kg/m
3
) 370 370 370 2000
h
q
(W/m
2
K) 22.5 1.6 22.5 17
h
m
(kg/m
2
s M) 2.5 10
6
6.0 10
8
2.5 10
6
1.6 10
5
k
q
(W/mK) 0.65 0.65 0.65 0.34
k
m
(kg/ms M) 2.2 10
8
2.2 10
8
2.2 10
8
2.4 10
7
d (M/K) 2.0 2.0 2.0 0.56
c
q
(J/kg K) 2500 2500 2500 607
c
m
(kg/kg M) 10
2
10
2
10
2
1.8 10
3
k (J/kg) 2.5 10
6
2.5 10
6
2.5 10
6
2.5 10
6
T
0
(C) 24 24 24 24
T
s
(C) 26 26 36 30
u
0
(M) 86 86 86 80
u

(M) 8 8 8 40
Bi
q
1.7 0.1 1.7 2.5
Bi
m
5.7 0.1 5.7 2.5
Lu 0.008 0.008 0.008 0.2
Pn 0.05 0.05 0.3 0.084
Ko 390 390 65 49
e 0.2 0.2 0.2 0.2
1592 L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598
estimation, we examine four test-cases of practical in-
terest, involving dierent materials and dierent
boundary conditions on the surface of the body in
contact with the surrounding air. Table 1 summarizes
the four test-cases examined in this work. Test-cases 13
deal with the drying of wood, while test-case 4 deals with
the drying of ceramics. The physical properties pre-
sented in Table 1 were obtained from Ref. [9] for wood
and from Ref. [11] for ceramics. Note in Table 1 that for
all test-cases the thickness of the sample was taken as
0.05 m and the initial temperature as 24 C.
In test-cases 13, dealing with the drying of wood, we
analyzed the eects of the air conditions on the deter-
minant of the information matrix. Test-case 1 involved
heat and mass transfer coecients at the boundary
X 1 of the order of those for forced convection
(h
q
22:5 W/m
2
C and h
m
2:5 10
6
kg/m
2
s M),
with the surrounding air slightly heated as compared to
the initial temperature for the medium (T
0
24 C and
T
s
26 C). Such air temperature was the same con-
sidered for test-case 2, but the heat and mass transfer
coecients for this test-case were of the order of those
for natural convection (h
q
1:6 W/m
2
C and h
m

6 10
8
kg/m
2
s M). For test-case 3, the heat and mass
transfer coecients utilized were the same as for test-
case 1, but the air temperature was taken as 36 C
instead of 26 C. Test-case 4 involved the drying of
ceramics, with heat and mass transfer coecients of the
order of those for forced convection (h
q
17 W/m
2
C
and h
m
1:6 10
5
kg/m
2
s M) with air slightly heated
(T
s
30 C).
Fig. 2a and b present the temperature and moisture
content variations with time, respectively, for dierent
positions along the medium, for test-case 1. We can
notice in Fig. 2a a large decrease in the temperature for
small times, as a result of the moisture vaporization. For
larger times, temperatures in the region become positive
when vaporization eects are less important and the
temperature distribution is governed by diusion.
The normalized sensitivity coecients for tempera-
ture and moisture content measurements are presented
in Figs. 3 and 4, respectively, at dierent locations inside
the medium, for test-case 1. Fig. 3 shows that the tem-
perature sensitivity coecients with respect to Bi
m
and
Lu are proportional, i.e., linearly dependent. The tem-
perature sensitivity coecients with respect to Pn and e
are practically null, while the others attain values com-
parable to the measured temperatures (see also Fig. 2a).
The temperature sensitivity coecients with respect to
Ko and Bi
q
are not linearly dependent with respect to the
others. Note in Fig. 3 that the temperature sensitivity
coecients are basically not aected by the sensor po-
sition. Let us now consider the analysis of the moisture
content sensitivity coecients presented in Fig. 4. Such a
gure shows that the moisture content sensitivity coef-
cients with respect to Ko, Bi
q
, Pn and e are practically
null for all measurement positions. The sensitivity co-
ecients with respect to Bi
m
and Lu attain magnitudes of
the same order of the measured moisture content (see
Fig. 2b), but they are linearly dependent, except for
X 0. An analysis of Figs. 3 and 4 reveals the diculty
involved in the simultaneous estimation of the dimen-
sionless parameters appearing in Luikovs linear for-
mulation. Both temperature and moisture content
measurements do not provide useful information for the
estimation of Pn and e, because their sensitivity coe-
cients are practically null. Also, there is a general ten-
dency of linear dependence for the sensitivity coecients
with respect to Bi
m
and Lu, independently of the type of
measurement considered. However, these sensitivity co-
ecients attain values of the order of the measured
variables and the constant of proportionality between
them is dierent for temperature and moisture content
0 15 30 45 60 75 90 105 120 135 150
Dimensionless time,

-5
-4
-3
-2
-1
0
1
2
D
i
m
e
n
s
i
o
n
l
e
s
s

t
e
m
p
e
r
a
t
u
r
e
,

X = 0 . 0
X = 0 . 2
X = 0 . 4
X = 0 . 6
X = 0 . 8
X = 1 . 0
(a)
0 15 30 45 60 75 90 105 120 135 150
Dimensionless time,

0.0
0.2
0.4
0.6
0.8
1.0
D
i
m
e
n
s
i
o
n
l
e
s
s

M
o
i
s
t
u
r
e

c
o
n
t
e
n
t
,

(b)
Fig. 2. Temperature (a) and moisture content (b) variations
inside the body for test-case 1.
L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598 1593
measurements; hence, the columns of the sensitivity
matrix do not become linearly dependent.
The determinant of the information matrix is pre-
sented in Fig. 5 for test-case 1, by considering dierent
numbers of temperature and moisture content sensors
available for the analysis. The moisture content sensors
were located inside the medium starting from the
boundary at X 0, because at this position the sensi-
tivity coecients with respect to Bi
m
and Lu are not
linearly dependent, as shown by Fig. 4. The temperature
sensors were located inside the medium starting from the
boundary at X 1, although the temperatures at dif-
ferent sensor locations are equally sensitive to variations
in the parameters, as shown by Fig. 3. For the results
presented in Fig. 5, we considered available one mea-
surement per sensor per unity of dimensionless time.
Also, the curves in Fig. 5 were based on six unknown
parameters, that is, P
T
Bi
q
; Bi
m
; Lu; Pn; Ko; e. We can
notice in Fig. 5, for each curve referent to the dierent
number of sensors, a large increase in the determinant of
the information matrix for small times and that the rate
of increase of the determinant is reduced for times
greater than 80. Therefore, a nal time equal to 80 ap-
pears to be suitable for test-case 1, because very little
improvement on the accuracy of the estimated parame-
ters is expected for larger times. We note that the de-
terminant increases when more sensors are utilized,
because more information is available for the inverse
analysis.
Fig. 6 presents the determinant of the information
matrix for test-case 2, where the heat and mass transfer
coecients at the boundary X 1 corresponded to
natural convection conditions. The frequency of mea-
surements and the number of unknown parameters are
the same as those considered for Fig. 5. A comparison of
Figs. 5 and 6 show a large reduction of the determinant
of the information matrix, when the heat and mass
transfer coecients are decreased. This is basically due
to the reduction observed in the magnitudes of the
moisture content sensitivity coecients and because of a
strong tendency towards linear dependence of the sen-
sitivity coecients for several parameters, when natural
convection boundary conditions are imposed. There-
fore, Figs. 5 and 6 show that forced convection
boundary conditions should be preferred for the body
surface open to the surrounding air at X 1.
(a) (b)
(c)
Fig. 3. Normalized temperature sensitivity coecients for test-case 1 for (a) X 0, (b) X 0:4 and (c) X 1.
1594 L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598
The determinant of the information matrix for test-
case 3 is presented in Fig. 7. Such as for test-case 2, the
frequency of measurements and the number of unknown
parameters are the same as those considered for Fig. 5.
We can notice in Fig. 7 that the increase of the tem-
perature of the surrounding air has the same eect of
reducing the determinant of the information matrix, as
the use of natural convection boundary conditions.
Therefore, a comparison of Figs. 57 reveals that forced
convection of slightly heated surrounding air should be
preferred for the boundary conditions at the open
boundary of the body at X 1, in order to improve the
accuracy of the estimated parameters.
It is interesting to note that similar results, with re-
spect to the boundary conditions and the surrounding
air temperature, were observed by Kanevce et al. [24] for
the estimation of the mass diusion coecient as a
function of temperature and moisture diusivity. The
use of larger heat and mass transfer coecients and
smaller temperatures of the surrounding air resulted in
larger values for the determinant of the information
matrix, despite the fact that the physical problem in-
volved the heating of the sample by convection on both
of its sides [24].
Let us now take into picture the analysis of test-case
4 involving the drying of ceramics. Based on the fore-
going analysis of test-cases 13, the boundary condition
at the open boundary of the body at X 1 for test-case
4 involved forced convection of slightly heated air, as
can be noticed in Table 1. Fig. 8 presents the determi-
nant of the information matrix for test-case 4, for six
unknown parameters, that is, P
T
Bi
q
; Bi
m
; Lu; Pn; Ko;
e and for a frequency of one measurement per sensor
per unity of dimensionless time. A comparison of Figs. 5
and 8 show larger values for the determinant of the
information matrix for ceramics than for wood. As a
result, more accurate estimates are expected for the es-
timation of the dimensionless parameters appearing in
Luikovs formulation for ceramics. The sensitivity co-
ecients for test-case 4 are not presented here for the
Fig. 4. Normalized moisture content sensitivity coecients for test-case 1 for (a) X 0, (b) X 0:4 and (c) X 1.
L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598 1595
sake of brevity, but qualitatively they are quite similar to
those for test-case 1, presented in Figs. 3 and 4. There-
fore, temperature and moisture content sensitivity co-
ecients with respect to Pn and e are also rather small
for test-case 4, so that the measurements do not provide
useful information for the estimation of such parame-
ters.
The foregoing analysis reveals that, for the cases
examined, only the parameters Bi
q
, Bi
m
, Lu and Ko can
be simultaneously estimated by using both temperature
and moisture content measurements. Therefore, only
such four parameters are considered as unknown
quantities for the inverse analysis presented below,
based on the LevenbergMarquardt method of mini-
mization of the least-squares norm, by using simulated
transient measurements. The parameters Pn and e are
then considered to be known exactly; but we note that
this is not a strong hypothesis, since their sensitivity
0 10 20 30 40 50 60 70 80
Dimensionless time,
1E-25
1E-24
1E-23
1E-22
1E-21
1E-20
1E-19
1E-18
1E-17
1E-16
1E-15
1E-14
1E-13
1E-12
1E-11
1E-10
1E-9
1E-8
1E-7
1E-6
1E-5
1E-4
1E-3
D
e
t
e
r
m
i
n
a
n
t
X
m
= 1.0 and X
n
= 0.0
X
m
= 1.0, 0.8 and X
n
= 0.0, 0.2
X
m
= 1.0, 0.8, 0.6 and X
n
= 0.0, 0.2, 0.4
X
m
= 1.0, 0.8, 0.6, 0.4 and X
n
= 0.0, 0.2, 0.4, 0.6
*
*
*
*
Fig. 7. Determinant of the information matrix for dierent
number of sensors for test-case 3.
0 2 4 6 8 10 12 14 16 18 20
Dimensionless time,
1E-15
1E-14
1E-13
1E-12
1E-11
1E-10
1E-9
1E-8
1E-7
1E-6
1E-5
1E-4
1E-3
1E-2
1E-1
1E+0
1E+1
1E+2
1E+3
1E+4
1E+5
1E+6
1E+7
D
e
t
e
r
m
i
n
a
n
t
X
m
= 1.0 and X
n
= 0.0
X
m
= 1.0, 0.8 and X
n
= 0.0, 0.2
X
m
= 1.0, 0.8, 0.6 and X
n
= 0.0, 0.2, 0.4
X
m
= 1.0, 0.8, 0.6, 0.4 and X
n
= 0.0, 0.2, 0.4, 0.6
*
*
*
*
Fig. 8. Determinant of the information matrix for dierent
number of sensors for test-case 4.
0 15 30 45 60 75 90 105 120 135 150
Dimensionless time,
1E-18
1E-17
1E-16
1E-15
1E-14
1E-13
1E-12
1E-11
1E-10
1E-9
1E-8
1E-7
1E-6
1E-5
1E-4
1E-3
1E-2
1E-1
1E+0
1E+1
1E+2
1E+3
1E+4
1E+5
1E+6
D
e
t
e
r
m
i
n
a
n
t
X
m
= 1.0 and X
n
= 0.0
X
m
= 1.0, 0.8 and X
n
= 0.0, 0.2
X
m
= 1.0, 0.8, 0.6 and X
n
= 0.0, 0.2, 0.4
X
m
= 1.0, 0.8, 0.6, 0.4 and X
n
= 0.0, 0.2, 0.4, 0.6
*
*
*
*
Fig. 5. Determinant of the information matrix for dierent
number of sensors for test-case 1.
0 10 20 30 40 50 60 70 80
Dimensionless time,
1E-25
1E-24
1E-23
1E-22
1E-21
1E-20
1E-19
1E-18
1E-17
1E-16
1E-15
1E-14
1E-13
1E-12
1E-11
1E-10
1E-9
1E-8
1E-7
1E-6
1E-5
1E-4
1E-3
1E-2
D
e
t
e
r
m
i
n
a
n
t
X
m
= 1.0 and X
n
= 0.0
X
m
= 1.0, 0.8 and X
n
= 0.0, 0.2
X
m
= 1.0, 0.8, 0.6 and X
n
= 0.0, 0.2, 0.4
X
m
= 1.0, 0.8, 0.6, 0.4 and X
n
= 0.0, 0.2, 0.4, 0.6
*
*
*
*
Fig. 6. Determinant of the information matrix for dierent
number of sensors for test-case 2.
1596 L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598
coecients are practically null and the measurements
are not aected by their values.
We present below in Table 2 the results obtained for
the estimation of Bi
q
, Bi
m
, Lu and Ko for test-case 4. For
the inverse analysis, we considered simulated tempera-
ture measurements to be available at X
m
0:9, 0.7, 0.5
and 0.0, and moisture content measurements at
X

n
0:1, 0.2, 0.3 and 0.4. Based on the analysis of Fig.
8, the nal dimensionless time was taken as 10, since the
determinant of the information matrix increases very
little for larger times. For the physical case under pic-
ture, such dimensionless time corresponds to 25 h.
Measurements were assumed available in a frequency of
one reading per sensor every 20 min. In order to examine
the accuracy of the LevenbergMarquardt method, as
applied to the present parameter estimation approach,
we used simulated measurements containing random
measurement errors with standard deviation r r

0
(errorless measurements), as well as r 0:01 Y
max
and
r

0:01 C
max
, where Y
max
and C
max
denote the maxi-
mum measured temperature and moisture content, re-
spectively. The initial guesses for the unknown
parameters, required for the iterative procedure of the
LevenbergMarquardt method, were taken as Bi
0
q
1:0,
Bi
0
m
5:0, Lu
0
0:02 and Ko
0
4:9. Table 2 shows that
the values were exactly recovered when errorless mea-
surements were used in the inverse analysis. Similarly,
quite accurate estimates were obtained when the mea-
surements with random errors were taken into account,
even for initial guesses as far as one order of magnitude
from the exact parameters.
8. Conclusions
In this paper we presented the solution for an inverse
problem of parameter estimation in a one-dimensional
capillary porous media, by using the dimensionless
Luikovs model for the heat and mass transfer process.
The associated direct problem was solved with the
GITT. The present parameter estimation problem was
solved by using the LevenbergMarquardt method of
minimization of the least-squares norm. Temperature
and moisture content measurements are assumed avail-
able for the inverse problem. Since temperature mea-
surements may contain errors with standard deviation
dierent from the moisture content measurements, the
weighted least-squares norm is used as the objective
function in order to obtain minimum variance estimates.
For the cases examined in this work, involving the
drying of wood and ceramics, an analysis of the sensi-
tivity coecients and of the determinant of the infor-
mation matrix shows that it is possible to estimate
simultaneously the Luikov number, the Kossovitch
number and the Biot numbers for heat and mass
transfer. The Posnov number and the phase change
conversion factor cannot be estimated, because their
sensitivity coecients are practically null. Results ob-
tained by using simulated measurements with random
errors illustrate the accuracy of the present approach for
the simultaneous identication of parameters in Lui-
kovs dimensionless linear formulation.
Acknowledgements
Useful discussions with Professors G. Kanevce and
L. Kanevce from St. Kliment Ohridski University,
Macedonia, are greatly appreciated. The computational
facilities of the Heat Engines Laboratory-PEM/COPPE
were donated by Renault-COPPE Plateau de Recherche.
This work was partially funded by PRONEX, CNPq
and FAPERJ.
References
[1] A.V. Luikov, Heat and Mass Transfer in Capillary-Porous
Bodies, Pergamon Press, Oxford, 1966.
[2] M.D. Mikhailov, M.N.

OOzisik, Unied Analysis and
Solutions of Heat and Mass Diusion, John Wiley, New
York, 1984.
[3] J.W. Ribeiro, P.D.C. Lobo, Drying in capillary porous
media, in: R.M. Cotta (Ed.), The Integral Transform
Method in Thermal and Fluids Science and Engineering,
Begell House Inc. Publishers, NY, 1998 (Chapter VII).
[4] P.D.C. Lobo, M.D. Mikhailov, M.N.

OOzisik, On the
complex eigenvalues of Luikov system of equations,
Drying Technol. 5 (2) (1987) 273286.
[5] S.M.S. Guigon, L.B. Dantas, F. Scofano Neto, R.M.
Cotta, Exact solution of Luikovs equations for drying in
capillary porous media, Hybrid Meth. Eng. 1 (1999) 365
387.
[6] J.W. Ribeiro, R.M. Cotta, M.D. Mikhailov, Integral
transform solution of Luikovs equations for heat and
Table 2
Results obtained for the estimation of parameters in test-case 4
Parameters Exact Estimated r 0 Estimated r 0:01M
max
Condence intervals
Bi
q
2.5 2.500 2.508 (2.505, 2.510)
Bi
m
2.5 2.500 2.487 (2.456, 2.519)
Lu 0.2 0.200 0.200 (0.146, 0.255)
Ko 49 49.000 49.123 (48.669, 49.578)
L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598 1597
mass transfer in capillary porous media, Int. J. Heat Mass
Transfer 35 (1993) 44674475.
[7] R.M. Cotta, Integral Transforms in Computational Heat
and Fluid Flow, CRC Press, Boca Raton, FL, 1993.
[8] J.W. Ribeiro, R.M. Cotta, On the solution of nonlinear
drying problems in capillary porous media through integral
transformation of Luikov equations, Int. J. Numer. Meth.
Eng. 38 (1995) 10011020.
[9] H.R. Thomas, R.W. Lewis, K. Morgan, An application of
the nite element method to the drying of timber, Wood
Fiber 11 (4) (1980) 237243.
[10] P.D.C. Lobo, M.A. Boles, M.N.

OOzisik, The eects of
geometry for multi-region heat and mass transfer governed
by the nonlinear Luikov system, in: Proceedings of the
International Symposium on Two-Phase Flow Modeling
and Experimentation, Rome, Italy, 1995, pp. 12431247.
[11] G. Comini, R.W. Lewis, A numerical solution of two
dimensional problems involving heat and mass transfer,
Int. J. Heat Mass Transfer 19 (1976) 13871392.
[12] J.V. Beck, K.J. Arnold, Parameter Estimation in Engi-
neering and Science, Wiley, New York, 1977.
[13] M.N.

OOzisik, H.R.B. Orlande, Inverse Heat Transfer:
Fundamentals and Applications, Taylor and Francis, New
York, 2000.
[14] O.M. Alifanov, Inverse Heat Transfer Problems, Springer-
Verlag, New York, 1994.
[15] G.S. Dulikravich, T.J. Martin, Inverse shape and boundary
condition problems and optimization in heat conduction,
in: W.J. Minkowycz, E.M. Sparrow (Eds.), Advances in
Numerical Heat Transfer, vol. 1, Taylor & Francis, New
York, 1996, pp. 381426 (Chapter 10).
[16] J.V. Beck, B. Blackwell, C.R. St. Clair, Inverse Heat
Conduction: Ill-posed Problems, Wiley Interscience, New
York, 1985.
[17] O.M. Alifanov, E. Artyukhin, A. Rumyantsev, Extreme
Methods for Solving Ill-Posed Problems with Applications
to Inverse Heat Transfer Problems, Begell House, New
York, 1995.
[18] L.B. Dantas, H.R.B. Orlande, R.M. Cotta, P.D.C. Lobo,
Parameter estimation in moist capillary porous media by
using temperature measurements, in: M. Tanaka, G.S.
Dulikravich (Eds.), Inverse Problems in Engineering Me-
chanics II, Elsevier, Oxford, 2001, pp. 5362.
[19] L.B. Dantas, H.R.B. Orlande, R.M. Cotta, Eects of
lateral heat losses on the parameter estimation problem in
moist capillary porous media, in: M. Tanaka, G.S.
Dulikravich (Eds.), Inverse Problems in Engineering Me-
chanics III, Elsevier, Oxford, 2002, pp. 1322.
[20] L.B. Dantas, H.R.B. Orlande, R.M. Cotta, Estimation of
dimensionless parameters of Luikovs system for heat and
mass transfer in capillary porous media, Int. J. Thermal
Sci. 41 (2001) 217227.
[21] K. Levenberg, A method for the solution of certain non-
linear problems in least squares, Quart. Appl. Math. 2
(1944) 164168.
[22] D.W. Marquardt, An algorithm for least squares estima-
tion of nonlinear parameters, J. Soc. Ind. Appl. Math. 11
(1963) 431441.
[23] G. Kanevce, L. Kanevce, G.S. Dulikravich, Moisture
diusivity estimation by temperature response of a drying
body, in: M. Tanaka, G.S. Dulikravich (Eds.), Inverse
Problems in Engineering Mechanics II, Elsevier, Oxford,
2000, pp. 6372.
[24] G. Kanevce, L. Kanevce, G.S. Dulikravich, Inuence of
boundary conditions on moisture diusivity estimation by
temperature response of a drying body, in: Proceedings of
34th ASME National Heat Transfer Conference, paper
NHTC2000-12296, Pittsburgh, Pennsylvania, August 20
22, 2000.
[25] G. Kanevce, L. Kanevce, G.S. Dulikravich, Simultaneous
estimation of thermophysical properties and heat and mass
transfer coecients of a drying body, in: M. Tanaka, G.S.
Dulikravich (Eds.), Inverse Problems in Engineering Me-
chanics III, Elsevier, Oxford, 2002, pp. 312.
[26] W.J. Chang, C.I. Weng, Inverse problem of coupled heat
and moisture transport for prediction of moisture distri-
butions in an annular cylinder, Int. J. Heat Mass Transfer
42 (1999) 26612672.
[27] C.-H. Huang, C.-Y. Yeh, An inverse problem in simulta-
neous estimating the biot numbers of heat and moisture
transfer for a porous material, Int. J. Heat Mass Transfer
45 (2002) 46434653.
[28] J.M. Frias, J.C. Oliveira, K. Schittkowski, Modeling and
parameter identication of a maltodextrin DE 12 drying
process in a convection oven, Appl. Math. Model. 25
(2001) 449462.
1598 L.B. Dantas et al. / International Journal of Heat and Mass Transfer 46 (2003) 15871598