JOURNAL OF COLLOID AND INTERFACE SCIENCE 196, 136–147 (1997)
ARTICLE NO. CS975214
Axisymmetric Drop Shape Analysis: Computational Methods
for the Measurement of Interfacial Properties from the Shape
and Dimensions of Pendant and Sessile Drops
O. I. del RıB o and A. W. Neumann 1
Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, Ontario M5S 3G8, Canada
Received September 27, 1996; accepted October 9, 1997
State-of-the-art axisymmetric drop shape analysis (ADSA)
techniques for the computation of interfacial tensions and contact
angles by fitting the Laplace equation of capillarity to the shape
and dimensions of pendant and sessile drops are presented. More
accurate, efficient, and reliable versions of the technique for the
measurement of contact angles from the volume and diameter of
sessile drops [axisymmetric drop shape analysis—diameter
(ADSA-D)] and for interfacial tension measurements from a se-
ries of arbitrary profile coordinates of sessile and pendant drops
[axisymmetric drop shape analysis—profile (ADSA-P)] have
been developed. Advanced numerical methods have been used to
improve the numerical stability and global convergence, for more
accurate results and a wider range of applicability of the methods.
A new technique called axisymmetric drop shape analysis—height
and diameter (ADSA-HD) has been developed to estimate interfa-
cial tensions from the height and diameter of sessile and pendant
drops. Numerical simulations using numerically generated drop
profiles were used to evaluate the accuracy and applicability of
the methods. q 1997 Academic Press
Key Words: interfacial tension; surface tension; contact angle;
drop shape analysis; pendant drop; sessile drop.
1. INTRODUCTION
In recent years, the use of automated computer software
for the measurement of interfacial properties of fluids, i.e.,
surface tensions g and contact angles uc , has not only im-
proved the accuracy of the measurements considerably (1–
4) but it has also permitted the study of phenomena that
were not possible or difficult to investigate in the past, such
as ultralow interfacial tensions (5), the relaxation of adsorp-
tion layers (6), and dynamic surface tensions (7). Using a
desktop computer, the time required for a single interfacial
tension measurement has been reduced from hours of dedi-
cated work to just a few seconds, allowing thousands of
time-dependent measurements to be performed, unattended,
in a single day.
1
To whom correspondence should be addressed.
0021-9797/97 $25.00
Copyright q 1997 by Academic Press
All rights of reproduction in any form reserved.
AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57
Numerous methodologies have been developed for the
measurement of interfacial tensions and contact angles,
as outlined in Refs. ( 8, 9 ) . Of these, axisymmetric drop
shape analysis ( ADSA ) methods are considered to be the
most powerful because of their accuracy, simplicity, and
versatility ( 3, 10 ) . They are also very suitable for auto-
mated computer implementation by means of digital im-
age analysis ( 2 ) .
ADSA methods are based on the numerical fit between
the shape of experimental drops (or bubbles) and the mathe-
matical model given by the classical Laplace equation of
capillarity. They are applicable to pendant and sessile drops
in a very wide range of surface tensions, for any fluid–liquid
system that can be represented by the Laplace equation of
capillarity. Early efforts in the analysis of axisymmetric
drops to determine surface tensions were those of Bashforth
and Adams (11) and Maze and Burnet (12), who developed
the numerical strategy on which modern methods are based,
i.e., the numerical integration of the Laplace equation and
the use of nonlinear least-squares optimization techniques.
The most powerful and widely used method to date is that
of Rotenberg et al. (1), called axisymmetric drop shape
analysis—profile (ADSA-P). Jennings and Pallas (10) de-
veloped a similar method, built on Rotenberg and col-
leagues’ work, in an attempt to make the method computa-
tionally more efficient, but they introduced numerical simpli-
fications that affect the accuracy of the method.
Another advantage of ADSA methods is that contact
angles, as well as volume and surface area, can be computed
simultaneously. This led to the development of axisymmetric
drop shape analysis—diameter (ADSA-D), which com-
putes contact angles based on the volume and diameter of
sessile drops with known surface tension (13, 14). ADSA-
D is particularly powerful for low contact angles (e.g., uc
õ 207 ) where other methods present difficulties.
Another popular method for estimating surface tensions
is the one developed by Malcolm and Elliott (15), which
uses a semiempirical equation to approximate surface ten-
136
coidas
 AXISYMMETRIC DROP SHAPE ANALYSIS
FIG. 1. Coordinate system used in the numerical solution of the Laplace
equation for axisymmetric liquid–fluid interfaces.
sion from the total height H and maximum diameter D of a
sessile drop or captive bubble with a 1807 contact angle.
Although this method is limited to a particular case, their
approach has been extended to sessile drops of any contact
angle (16) and can be further generalized to pendant drops.
This approach has been applied to measure low interfacial
tensions that can be generated in pulmonary surfactant by
compressed captive bubbles (17–20) and other systems with
ultralow interfacial tensions (5).
This paper describes recent developments in the ADSA-
P and ADSA-D methodologies and introduces a new tech-
nique, axisymmetric drop shape analysis—height and diam-
eter (ADSA-HD), to determine surface tension based on the
height and diameter of sessile and pendant drops. Efficient,
general-purpose numerical algorithms that make use of re-
cent developments in numerical analysis have been imple-
mented, and a numerical library of ADSA algorithms has
been created to facilitate implementation and further devel-
opment of the methodologies. Emphasis has been given to
the numerical accuracy, efficiency, flexibility, and stability
of the methods. The only assumptions made are that the
drops or bubbles are Laplacian and axisymmetric.
2. NUMERICAL SOLUTION FOR AXISYMMETRIC
LIQUID–FLUID INTERFACES (ALFI)
The classical Laplace equation of capillarity describes the
mechanical equilibrium conditions for two homogeneous
fluids separated by an interface. For axisymmetric interfaces
it can be written as the following system of ordinary differen-
tial equations as a function of the arc length s, as shown in
Fig. 1 (1):
dx
Å cos u, [1a]
ds
dz
Å sin u, [1b]
ds
du sin u
Å 2b / cz 0 , [1c]
ds x the Burlisch–Stoer scheme with adaptive stepsize control is
AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57
137
dV
Å px 2 sin u, [1d]
ds
dA
Å 2px, [1e]
ds
x(0) Å z(0) Å u(0) Å V (0) Å A(0) Å 0, [1f]
where b is the curvature at the origin of coordinates and c
Å ( Dr )g/ g is the capillary constant of the system. u is the
tangential angle which, for sessile drops, becomes the con-
tact angle uc at the three-phase contact line. Although the
surface area A and the volume V are not required to define
the Laplacian profile, they are included here because of their
importance and the fact that they can be integrated simulta-
neously with little computational overhead. It should also
be noted that since at the apex of the drop,
sin u
Åb at s Å 0,
x
then Eq. [1c] becomes
du
Åb at s Å 0, [1g]
ds
which avoids a division by zero, and there is no need to use
analytical approximations to initialize the integration as done
by Dimitrov et al. (21).
For given values of b and c, a unique shape of a Laplacian
axisymmetric fluid–liquid interface can be obtained by si-
multaneous integration of the above initial value problem.
However, there is no known analytical solution for this prob-
lem, except for very limited cases, and a numerical integra-
tion scheme must be used, as implemented in the axisymme-
tric liquid–fluid interfaces (ALFI) computer program de-
scribed below.
There exist several numerical methods to solve systems
of ordinary diferential equations for initial value problems
and considerable research is still devoted to this subject
(22). The three major types of methods most commonly
used are the Runge–Kutta, Burlisch–Stoer extrapolation,
and predictor–corrector methods (23). For simple equations
and when efficiency is of no concern, fixed stepsize imple-
mentations are often used, but they are not suitable when
accuracy and computational efficiency are important. Hart-
land and Hartley (24) used a fourth-order Runge–Kutta
method with truncation error control to solve the Laplace
equation (1). Rotenberg et al. (1) and Jennings and Pallas
(10) implemented a second-order implicit Euler method.
After testing and comparing other methods, it was found that
coidas
138 DEL RIO AND NEUMANN
computationally efficient but it can fail for large drops with
very low surface tension.
A more efficient and flexible method is the fifth- and
sixth-order Runge–Kutta–Verner pair, as implemented by
Hull, Enright, and Jackson in the DVERK FORTRAN nu-
merical routines (25, 26). Some important features of this
code are its defect-based error and stepsize control that
allows large integration steps with controlled accuracy, its
flexible design that allows the calling program to interrupt
and continue the integration after any step, and its capability
to produce continuous approximations to the solution on
each step using very accurate interpolations. These features
are important to improve the speed and accuracy of ADSA
computations, where numerous integrations are required, as
described later.
ALFI was written, implementing the Hull–Enright–Jack-
son numerical integration scheme, to generate Laplacian pro-
files of pendant and sessile drops of any size and surface
tension by numerically integrating the initial value problem
[1]. The following are some of the features of ALFI:
• The volume V and surface area A are computed simulta-
neously with the drop profile with negligible computational
overhead. These properties are useful in applications such
as the ADSA film balance (27).
• The integration can be stopped when any given values
of s, u, x, z, V, or A are reached, allowing the computation
of drop profiles of any specified contact angle, volume, sur-
face area, or size. The integration also terminates if u reaches
1807 (sessile drops) or becomes negative (pendant drops).
• The inflection point of pendant drops is accurately com-
puted, which is useful for testing and evaluating drop profile
methods.
• The origin of the coordinate system can be translated
and rotated arbitrarily, and the coordinates can be scaled in
the horizontal and vertical directions. This feature permits
the comparison between theoretical and experimental drop
profiles which have generally an arbitrary origin of coordi-
nates, can be vertically misaligned due to a vertical misalign-
ment of the video camera, and have an arbitrary magnifica-
tion.
• The profile coordinates can be randomly perturbed in
the normal direction, allowing the simulation of experimen-
tal errors, which can be used to evaluate ADSA methods.
ALFI has been found very useful not only in producing
tables and plots of Laplacian data as done by Hartland and
Hartley (24) and by Bashforth and Adams (11), but it is
also a very important tool to test and evaluate drop shape
analysis methods, as described later.
As mentioned before, ALFI generates complete Laplacian
profiles from values of b and c by integrating the initial
value problem [1]. The inverse process of determining b
and c (from which g and contact angle uc can be easily
AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57
computed) based on drop profile characteristics is a more
difficult task and forms the basis of the ADSA methods
described in the following sections.
3. CONTACT ANGLE MEASUREMENTS
FROM DIAMETER AND VOLUME
OF SESSILE DROPS (ADSA-D)
The ADSA-D methodology to compute contact angles uc
from the contact or maximum diameter D (usually measured
from a picture of the drop looking from above) and volume V
of sessile drops with known surface tension g was originally
developed by Skinner et al. (13) and Moy et al. (14). The
basic scheme of their implementation was to generate Lapla-
cian profiles by integration of [1] for different values of b,
comparing the measured and the computed volumes at the
given diameter, until convergence, employing a Newton iter-
ation. However, this approach, also known as single shoot-
ing, has numerical deficiencies and can fail if the initial
values are distant from the solution. In this section, a bound-
ary-value-problem approach to ADSA-D is described, which
is numerically more stable and easier to implement.
As in the original ADSA-D, there are two cases to con-
sider—(1) uc ¢ 907 and (2) uc õ 907 —which represent
two separate boundary-value problems since in the first case
the maximum diameter corresponds to the equatorial diame-
ter of the drop (at u Å 907 ) and in the second case the
maximum diameter corresponds to the three-phase contact
line (at u Å uc ) (see Fig. 2).
a. Contact Angle Greater than or Equal to 907
Rewriting equations [1a] – [1c] as functions of x, consid-
ering the curvature b as a new variable, and with the bound-
ary conditions as seen in Fig. 2a, the Laplace equation can
be written as the following boundary-value problem for uc
¢ 907:
du
dx
Å
1
cos u S 2b / cz 0
sin u
x D , [2a]
dz
Å tan u, [2b]
dx
db
Å 0, [2c]
dx
p
z(0) Å u(0) Å 0, u(R) Å [2d]
2
where R Å D/2 is the maximum (equatorial) radius. Problem
[2] completely defines the Laplacian shape; its solution
gives directly the profile shape for 0 ° x ° R, and the
constant value of the apex curvature b. The contact angle
coidas
 AXISYMMETRIC DROP SHAPE ANALYSIS
FIG. 2. Boundary conditions for ADSA-D boundary-value problems
[2] and [3]. (a) Contact angle greater than or equal to 907; (b) contact
angle less than 907.
uc can then be computed by integrating Eq. [1], for the
known values of b and g, past the maximum diameter, stop-
ping when the computed volume reaches the drop volume,
Vc . This problem thus requires a boundary-value problem
solver plus a numerical integration.
b. Contact Angle Less than 907
Similarly, as seen in Fig. 2b, the Laplace equation for
uc õ 907 can be written as the following boundary-value
problem:
139
additional numerical integration since the contact angle can
be obtained simply from uc Å u(R).
To initialize ADSA-D it is necessary to determine, for
given values of Vc and R, whether uc ¢ 907 (problem [2])
or uc õ 907 (problem [3]). On occasion, the user can give
this information as input, but in many cases, especially for
uc near 907, it is not known. The approach implemented in
the program is (1) if the user knows whether the drop is
wetting ( uc õ 907 ) or nonwetting ( uc ¢ 907 ), solve the
respective problem and exit; otherwise, (2) assume uc ¢ 907
and solve problem [2] for the given R, and compute volume
V90 at u Å 907 by numerically integrating [1]. If V90 ° Vc
the initial assumption was correct, compute uc , and exit;
otherwise, (4) solve problem [3] for uc õ 907.
The nonlinear problems [2] and [3] must be solved nu-
merically. Numerical methods to solve boundary-value prob-
lems can be divided into two main categories (28, 29):
initial value or shooting methods, and finite difference or
collocation methods. The original implementation of ADSA-
D used a variation of the single shooting method, which is
known to have stability drawbacks. Multiple shooting meth-
ods are more stable but, after some tests, they were found to
be inefficient for Laplacian problems, particularly for well-
deformed drop shapes. A more efficient approach is the use
of finite-difference methods with collocation formulas, as
implemented in the COLSYS FORTRAN library by Ascher
and Bader (28), which uses collocation at Gaussian points
to approximate the solution.
COLSYS, as well as any other nonlinear numerical
method, requires initialization with approximate values of
the solution. For the case of uc ¢ 907, the profile and the
curvature are initialized with an elliptical approximation,
and for uc õ 907 the solution is initialized with zero values
and the curvature with 1/R. It was found that these rough
approximations are sufficient for COLSYS to converge.
Nevertheless, a continuation algorithm (subdividing the
problem into several easier-to-solve steps) was implemented

S D
to guarantee convergence to a solution in case of an initial
du 1 sin u failure of COLSYS, using the capillary constant c as the
Å 2b / cz 0 , [3a]
dx cos u x continuation parameter.
dz
Å tan u, [3b] 4. SURFACE TENSION MEASUREMENTS FROM
dx HEIGHT AND DIAMETER OF SESSILE
dV AND PENDANT DROPS (ADSA-HD)
Å px 2 tan u, [3c]
dx
Several methods to approximate g from the height H and
db diameter D of sessile drops, based on semiempirical equa-
Å 0, [3d] tions and/or Laplacian tables, have been developed in the
dx
past (15, 16). However, these methods are restricted to spe-
z(0) Å u(0) Å V (0) Å 0, V ( R) Å Vc . [3e] cial cases (e.g., sessile drops with uc É 1807 ) and their
accuracy is limited for a certain range of height/diameter
Here Vc is the total volume of the drop. Problem [3] com- ratios (20), even when perfect numerical data are used.
pletely defines the Laplacian shape. There is no need for an In this section a boundary-value-problem approach to this

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

140 DEL RIO AND NEUMANN

case, problem [4] is a three-point boundary-value problem;

if H and D are measured at the same location, i.e., u1 Å u2 ,
it becomes a two-point boundary-value problem.
As in the ADSA-D implementation (see above), the COL-
SYS library can be used to solve problem [4], taking advan-
tage of its ability to handle multiple-point boundary condi-
tions. The algorithm in this case is initialized using an ellip-
tical fit that satisfies the boundary conditions [4e], an initial
curvature equal to the curvature of the fitted ellipse at the
origin, and the capillary parameter is initialized with a user-
supplied value, which can be a very rough estimate, as seen
in Section 9. A continuation algorithm has also been imple-
mented in case COLSYS fails initially.

5. SURFACE TENSION MEASUREMENTS BY

FIG. 3. Boundary conditions for ADSA-HD boundary-value prob-
lem [ 4 ] .
problem is presented, which is exact (to a specified numeri-
cal tolerance) and general: it is applicable to sessile drops
(wetting and nonwetting) and pendant drops, and it can be
used with any measured height and diameter, not necessarily
the equatorial diameter and maximum height.
The most general case, as shown in Fig. 3, can be defined
by the following boundary-value problem, written as a func-
tion of the tangential angle, with both the apex curvature b
and the capillary parameter c as unknown constants:
dx
du S
Å cos u 2b / cz 0
sin u
x D01
, [4a]
DROP PROFILE FITTING OF SESSILE
AND PENDANT DROPS (ADSA-P)
The ADSA-P methodology to determine interfacial prop-
erties by means of a numerical fit of several arbitrary drop
profile coordinates to the Laplace equation was originally
developed by Rotenberg et al. (1) using a four-parameter
nonlinear least-squares fit with a Newton optimization
scheme and incremental loading or parameter continuation.
This method has been used successfully to measure interfa-
cial tensions in a variety of systems for several years (4)
but it has presented convergence problems for some systems
with ultralow interfacial tension and very flat drop shapes
because of the limitations of Newton’s method when the
initial parameters are distant from the solution. A new ver-
sion of ADSA-P is presented in this section, implementing
additional optimization parameters and more efficient and

S D01
globally stable numerical methods. Also, the original optimi-
dz sin u zation parameters have been redefined, using the curvature
Å sin u 2b / cz 0 , [4b]
du x at the apex b instead of the radius of curvature (R0 Å b 01 )
and making the parameters independent of each other, which
db simplifies the algebra and the numerical analysis.
Å 0, [4c]
du Similarly to the original ADSA-P, the strategy used is to
dc construct and minimize an objective function E, defined as
Å 0, [4d] the sum of the weighted squared normal distances between
du
any N profile coordinates and the Laplacian profile (problem
x(0) Å z(0) Å 0, x( u1 ) Å R, z( u2 ) Å H. [4e] [1]), as seen in Fig. 4:
N
Here R Å D/2. Problem [4] completely defines the drop E Å ∑ wi ei , [5a]
shape; its solution determines the profile coordinates for any iÅ1
u, as well as the (constant) values of b and c from which
g and other drop properties, e.g., uc , V, and A, can be readily ei Å 12 d 2i Å 12[(xi 0 Xi ) 2 / (zi 0 Zi ) 2 ]. [5b]
computed by the numerical integration of [1]. The same
definition of the problem can be applied to pendant and Here, wi is a weighting factor, (Xi , Zi ) are the measured
sessile (wetting and nonwetting) drops. The only limitation drop coordinates, and (xi , zi ) are the Laplacian coordinates
is that, for pendant drops, the boundary conditions at u1 and closest to (Xi , Zi ). Currently, wi is set equal to 1.0 until
u2 must be measured before the inflection point ( uinf in Fig. more studies are available on the effect of weighting factors.
3) for the problem to have a unique solution. In the general By introducing the generally unknown origin (x0 , z0 ) and

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

 AXISYMMETRIC DROP SHAPE ANALYSIS
FIG. 4. Definition of error function parameters for ADSA-P optimiza-
tion problem [5].
angle of rotation of the system of coordinates a, and scaling
factors on both coordinates (Xs , Zs ), the individual error can
be written (dropping the subscript i) as
e Å 12 (e 2x / e 2z ), [5c]
ex Å x 0 x0 0 Xs X cos a / ZsZ sin a, [5d]
ez Å z 0 z0 0 Xs X sin a 0 ZsZ cos a. [5e]
The objective is therefore to compute the set of M optimiza-
tion parameters a that minimizes [5], where a Å [b c x0 z0
a Xs Zs ] T or any subset of it. This definition is more general
than the four parameters used by Rotenberg et al. and is
similar to the parameter set used by Jennings and Pallas
(10). It should be noted though that only one of the scaling
factors, Xs or Zs , can be optimized simultaneously with b
and c for the solution to be unique. Generally one of the
scaling factors is known from the experimental setup and
can be held constant while optimizing the other to correct for
the aspect ratio to calibrate the optical system. The rotational
angle a can also be optimized to correct for the rotational
misalignment of the camera for calibration purposes, but it
can be held constant once the system has been calibrated,
provided the camera has not been moved.
The optimization problem can be written as
N method (30).
min E(a) Å ∑ wi ei (a). [6]
a iÅ1
Problem [6] is a multidimensional nonlinear least-squares
problem that requires an iterative optimization procedure.
When the minimum has been found, the optimization param-
eters determine the Laplacian profile that best fits the given
profile, from which g and other properties can be readily
computed.
Evaluating E for a trial set of a, i.e., for each optimization
iteration, involves determining the minimum (normal) dis-
AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57
141
tance from the Laplacian curve to each experimental point.
This is done using a one-dimensional Newton–Raphson iter-
ation to solve, for each ith point,
de
f (s) Å Å ex cos u / ez sin u Å 0.
ds
This procedure requires several numerical evaluations of the
initial value problem [1] for each data point, and in previous
versions of ADSA-P it required several numerical integra-
tions of [1] (approximately 3N integrations) that made the
evaluation of E computationally expensive. With the imple-
mentation of the numerical integration package of Hull, En-
right, and Jackson, DVERK, only one integration of [1] is
necessary to evaluate E for any given set of the parameters
a, and any intermediate values are computed with the inex-
pensive and equally accurate function approximation routine
of DVERK, called INTRP (26).
In the work of Jennings and Pallas the residuals are com-
puted using a linear interpolation in an attempt to reduce
computational time. However, this approach introduces
larger errors on sections of the drop with larger curvature
and is not suitable for highly curved surfaces. The Newton–
Raphson approach implemented in ADSA-P is exact, to the
given numerical tolerance, for all curvatures and drop pro-
files.
There exist several numerical methods to solve optimiza-
tion problems. Among them, Newton’s method is well
known for its second-order convergence if the initial values
are very close to the solution, but it is unpredictable other-
wise, particularly for multidimensional problems. To over-
come this problem, several Newton-like algorithms have
been developed with more advanced convergence strategies.
The original ADSA-P used Newton’s method with incre-
mental loading to approach the solution, but this approach
is computationally expensive and its convergence is not
guaranteed. Jennings and Pallas (10) used a modified
Gauss–Newton method with restricted step. Another method
commonly used for nonlinear least-squares optimization is
the Levenberg–Marquardt algorithm which can be described
as a trust-region implementation of the Gauss–Newton
In this version of ADSA-P, a combination of the New-
ton and Levenberg – Marquardt methods were imple-
mented. Very often, as in the case of time-dependent stud-
ies, the results from a previous run can be used as initial
values and Newton’s method can be used to take advan-
tage of its fast convergence. The FORTRAN linear algebra
packages LAPACK and BLAS were used to solve the
system of equations to compute the Newton iteration
steps. The Newton optimization is aborted as soon as di-
vergence is detected or if the Hessian matrix is found to
be nonpositive-definite ( 27 ) .
coidas
142 DEL RIO AND NEUMANN

If good initial values are not available or if Newton’s

method fails, the Levenberg–Marquardt method is used.
This method takes advantage of numerical properties partic-
ular to least-squares problems, is known to be globally con-
vergent, and needs only first-order derivatives to build the
Jacobian matrix. The numerical implementation chosen here
is the Levenberg–Marquardt with derivative (LMDER) rou-
tines from the FORTRAN MINPACK library developed by
Moré and Wright (31).
Both the first derivatives needed to construct the Jacobian
matrix used by LMDER and the second derivatives needed
to construct the Hessian used by Newton’s method are com-
puted with the same degree of accuracy as the objective
function by simultaneous numerical integration of the deriv-
atives together with the integration of [1], as described in
Ref. (1). This involves the simultaneous numerical integra-
tion of a system of 11 first-order ordinary differential equa-
tions for the LMDER routines and 20 for the Newton
method.
Of the seven possible optimization parameters, the most
difficult one to initialize is the capillary parameter c. It can
be initialized with either a call to ADSA-HD, with results
FIG. 5. Sessile and pendant drop profiles generated with ALFI for
from a previous run (e.g., time-dependent studies), or with testing and evaluating ADSA program.
user-supplied values. The effect of this initial value on the
performance of the program is evaluated in Section 10. Good
initial values for b, x0 , and z0 can be found by a least-squares 7. TESTING AND EVALUATION OF ALFI
elliptical fit of several points near the drop apex, and the
rotational angle a and the scaling factors Xs and Zs are gener- The numerical profiles obtained with ALFI agree, within
ally known from the experimental setup. a specified numerical precision, with those published by
Hartland and Hartley (24) and with other numerical integra-
tion schemes tested; they are not reproduced here. By tracing
6. TESTING AND EVALUATION PROCEDURE the numerical integration steps, it was found that, as the
numerical integration routine DVERK automatically adjusts
The evaluation of ADSA methods requires that drop pro- the stepsize along the profile to control the estimated numeri-
files of known surface tension and apex curvature be given cal errors, small steps are usually needed near the drop apex
as input data, so the numerical results can be compared and increasingly larger steps (compared to the total arc
with the known values of the parameters. Since experimental length) afterward.
values of surface tension and curvature are generally uncer- Several drop profiles, both sessile and pendant, were gen-
tain or unknown, the evaluation of ADSA numerical meth- erated during the testing and evaluation of the programs.
ods is best done by means of ‘‘synthetic’’ or theoretical Figure 5 shows some selected profiles with a variety of drop
drops, obtained by numerical integration of the Laplace shapes, with the corresponding parameters shown in Table
equation. Another advantage of using synthetic drop profiles 1. The capillary parameter c Å 13.45 cm02 used for some
is that the sensitivity of the methods can be easily measured of the profiles corresponds to a value typical of a water–air
by perturbing the input data to simulate experimental errors, system. The computational performance of ALFI is of the
and comparing the results with the known exact solution. order of 0.05 s to produce 100 plotting points for each drop.
The following sections describe how the ALFI program A numerical tolerance of 10 08 was used to generate the drop
was used to generate synthetic profiles to test and evaluate profiles.
the ADSA methods presented above. The computational
times were measured on a Sun SPARCstation 10 (60-MHz 8. TESTING AND EVALUATION OF ADSA-D
SuperSPARC CPU) and are presented here only as refer-
ence. Comparative or even faster performance can be ob- ADSA-D requires as input c, R, and V of a sessile drop,
tained on other computer systems, e.g., Intel Pentium per- and returns the contact angle uc . Table 2 shows the results
sonal computers. of running ADSA-D for sessile drops S1–S5 with arbitrarily

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

 AXISYMMETRIC DROP SHAPE ANALYSIS 143
TABLE 1 which causes larger changes in contact angle for small per-
Parameters of Sessile (S) and Pendant (P) Drop Profiles, Gener- turbations of any parameter, especially for large values of
ated with ALFI, Used in the Testing and Evaluation of ADSA c. Note that drop S5 with uc Å 1807, which produced the
Methods.a largest errors, corresponds to a very large value of c. Also,
for drop S1 with uc Å 57 the maximum error is due to the
S1 S2 S3 S4 S5
very small volume of the drop, but with a more accurate
c (cm02) 13.45 13.45 13.45 1.0 1000.0 volume measurement the error would be negligible.
b (cm01) 0.734 0.03 10.0 1.0 1.0
P1 P2 P3 P4 P5 9. TESTING AND EVALUATION OF ADSA-HD

c (cm02) 013.45 013.45 013.45 01.00 01000.00 The only input required by ADSA-HD is coordinate infor-
b (cm01) 5.838 10.0 1.0 2.0 50.00 mation (x, z, u ) of one or two arbitrarily selected profile
a
points, from which c (therefore g ) is computed. See Eq.
c is the capillary constant (Dd)g/g and b is the curvature at the drop [4]. The points of maximum diameter and maximum height
apex.
are usually chosen since they are easier to measure, but the
method is not limited to any particular locations. In practice
selected contact angles. The input values were obtained from these coordinates can be accurately measured from a digital
the data generated by ALFI for the respective contact angles. image.
A numerical tolerance of 10 06 was used in the computations. Table 4 shows the results and performance of ADSA-HD
It can be seen that ADSA-D returns the correct contact angle for the sessile and pendant drops of Fig. 5 using arbitrarily
with negligible errors in all the cases. Table 2 also shows selected boundary points. D, H, u1 , and u2 are the input
the computational time t required to run ADSA-D. Its perfor- data, t is the computational time required to run the program,
mance is always satisfactory, with better performance for c is the output capillary constant, and cerror is the output error.
contact angles less than 907 (less than 0.5 s). In all the cases A tolerance of 10 06 was used, and the initial value of c was
tested, the method succeeded in the first call to the boundary always given with an error greater than or equal to 25% of
problem solver COLSYS; i.e., parameter continuation was the exact solution, e.g., cinit Å 0.1, 10, 500 for c Å 1, 13.45,
not required. 1000, respectively. Every run shown in the table converged
To evaluate the sensitivity of the method to experimental in the first iteration (i.e., no parameter continuation was
errors, each input parameter can be perturbed by a small required) in less than 0.5 s, except for pendant drop P3
amount d, comparing the resulting contact angles with the which required two continuation iterations, hence the longer
original one. In practice, V can be measured with a microme- computational time. It can be seen that the method is not
ter syringe with an accuracy of 10 05 cm3 (0.01 ml), D can very sensitive to the initial estimate of c. The program suc-
be measured using digital image analysis with an accuracy ceeded for all the sessile and pendant drops tested with rela-
better than 0.15% (approximately one pixel on a 640 1 480 tive numerical errors within the specified tolerance.
image), and c can be obtained with an accuracy better than To evaluate the sensitivity of the method to experimental
1.0%. These experimental errors would produce the contact errors, the input values of H and D were perturbed by 0.15%
angle errors shown in Table 3, where each row shows the
error obtained by perturbing each input parameter (keeping
TABLE 2
the others constant) and the last row is the maximum com-
Computational Results of ADSA-D for Exact Values
bined error obtained by perturbing all the input parameters.
of Input Dataa
The tests show that the method is not very sensitive to experi-
mental errors for small contact angles, for which the best c R V t uc uerror
results are obtained (e.g., an error of less than 0.57 for contact Drop (cm02) (cm) (cm3) (s) (deg) (deg)
angles less than 307 and less than 17 for contact angles less
than 907 ). For contact angles greater than 907 the method S1, 57 13.45 0.1161 0.00011 0.19 5 2.11e-04
S1, 1207 13.45 0.6562 0.4578 1.16 120 9.83e-06
is more sensitive and more accurate measurements are re- S2, 107 13.45 1.1152 0.1236 0.23 10 6.40e-08
quired for good results. S3, 807 13.45 0.0965 0.0014 0.38 80 6.71e-07
These results agree with the experimental results obtained S4, 307 1.00 0.4849 0.0486 0.30 30 5.33e-08
by Skinner et al. in the evaluation of the original ADSA-D S5, 1807 1000.00 0.1401 0.0032 1.50 180 õ1.0e-08
method (13). The effect of nonaxisymmetry of the drops S5, 457 1000.00 0.1299 0.00089 0.26 45 5.63e-05
was also addressed by Skinner et al. The larger errors for a
uc is the output contact angle and uerror is the numerical error. The error
contact angles greater than 907 are due to the faster angular values are expressed in standard scientific notation; e.g., 1.0e-05 is equal
change of the drop profiles below the maximum diameter, to 1.0 1 1005.

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

144 DEL RIO AND NEUMANN

TABLE 3
Maximum Contact Angle Errors (deg) from ADSA-D for Typical Experimental Errorsa

S1, 57 S1, 1207 S2, 107 S3, 807 S4, 307 S5, 1807 S5, 457

dc 0.0006 0.508 0.028 0.0082 0.0031 11.04 0.134

dD 0.022 0.520 0.036 0.187 0.122 9.816 0.156
dV 0.468 0.0033 0.0008 0.295 0.0056 9.498 0.492
Maximum 0.489 1.024 0.065 0.490 0.131 17.30 0.779

a
Each row represents the error obtained by perturbing the parameters c, D, and V, simulating typical experimental errors. The maximum error
corresponds to the worst-case scenario by perturbing c, D, and V simultaneously.

of the original values (approximately one pixel on a 640 1 are shown in Table 6. It can be seen that the numerical errors
480 digital image), and the resulting value of c was com- are negligible.
pared with the original one. These results are shown in Table As mentioned earlier, the ADSA-P optimization algorithm
5. It can be seen that the sensitivity of the method varies requires as input the initial values of the optimization param-
from drop to drop, with errors of less than 1% for some eters, which, except for the capillary constant c, can be
drops but higher errors in other cases. This behavior depends estimated using an elliptical least-squares fit. To evaluate
on the slope of the multidimensional topology of the given the global convergence and the sensitivity of the method to
problem and seems to be unpredictable, but it can be evalu- the initial value of c, a wide range of initial values were
ated by perturbing the input data by the estimated measure- given as input to the program and the results are shown in
ment errors as described above. Further tests show that, for Table 7 for both Newton and Levenberg–Marquardt optimi-
a given drop, the error becomes smaller when the input data zation methods. It can be seen that the Levenberg–Mar-
are measured at larger angles u1 and u2 ; e.g., the error ob- quardt method succeeds for any initial value of c without a
tained using data points near u Å 1807 for drop S4 is approxi- significant increase in computational time, while the Newton
mately 1%. method succeeds only if the initial value is very close to the
exact solution ( {10% in this particular case). However,
10. TESTING AND EVALUATION OF ADSA-P when successful, the Newton method is considerably faster
than the Levenberg–Marquardt method.
ADSA-P differs from the other ADSA methods in that it Figure 6 shows the computational performance of the
requires as input an arbitrary number N ú M of profile methods when different numbers of data points N are used.
coordinates, where M is the number of optimization parame- For the Newton optimizations, initial values of 0.9 c were
ters. To test the accuracy of the algorithm, synthetic drops used. It can be seen again that the Newton method is consid-
S1–S5 and P1–P5 were given as input to the program, using erably faster but it failed to converge for drops S1, S4, and
N Å 21 points and a numerical tolerance of 10 08 . The results S5, while the Levenberg–Marquardt method succeeded in
all the cases.
TABLE 4 TABLE 5
Computational Results of ADSA-HD for Exact Values Capillary Constant Errors (% of Exact Solution) from
of Input Dataa ADSA-HD for Typical Experimental Errorsa
D/2 H u1 u2 t c cerror S1 S2 S3 S4 S5
Drop (cm) (cm) (deg) (deg) (s) (cm02) (cm02)
dR 0.11 0.050 7.13 8.85 0.011
S1 0.6562 0.4578 90 120 0.35 13.45 õ1.0e-07 dH 0.41 0.35 7.42 9.16 0.31
S2 1.1152 0.0503 10 10 0.47 13.45 õ1.0e-07
Maximum 0.52 0.40 14.57 18.07 0.32
S3 0.0965 0.0793 80 80 0.35 13.45 4.0e-07
S4 0.4849 0.1278 30 30 0.35 1.0 õ1.0e-07 P1 P2 P3 P4 P5
S5 0.1401 0.0664 90 180 0.45 1000.0 9.8e-05
P1 0.1860 0.2705 90 100 0.50 013.45 1.0e-06 dR 0.71 1.01 5.29 6.95 1.11
P2 0.1024 0.2058 90 140 0.32 013.45 3.2e-06 dH 0.41 0.71 5.0 6.65 0.81
P3 0.1839 0.0166 10 10 1.33 013.45 4.5e-05 Maximum 1.12 1.71 10.33 13.62 1.92
P4 0.4461 0.2632 60 60 0.33 01.0 1.0e-06
a
P5 0.0217 0.0250 90 90 0.44 01000.0 4.4e-04 Each row represents the error obtained by perturbing the parameters R
and H, simulating typical experimental errors. The maximum error corre-
a
c is the output capillary constant and cerror the numerical error. sponds to the worst-case scenario of perturbing both R and H.

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

 AXISYMMETRIC DROP SHAPE ANALYSIS 145
TABLE 6
Numerical Results of ADSA-P for Sessile (S) and Pendant (P)
Drops Using N Å 21 Points

S1 S2 S3 S4 S5

c (cm02) 13.45 13.45 13.45 1.0 1000.0

cerror (cm02) 7.0e-08 7.6e-06 2.7e-04 8.1e-07 8.3e-05
P1 P2 P3 P4 P5

c (cm02) 013.45 013.45 013.45 01.0 1000.0

cerror (cm02) 2.4e-06 2.0e-06 1.3e-04 3.6e-06 3.6e-03

To simulate the effect of experimental errors on ADSA-

P, ALFI was used to randomly perturb the Laplacian profiles
with a maximum normal perturbation of approximately one
pixel of a 640 1 480 digital image. Sessile drops of various
contact angles and pendant drops from the apex to the inflec-
tion point were used, as seen in Table 8. The inflection
points of the pendant drops were used only for convenience;
ADSA-P is not limited by nor needs information about the
inflection point. It can be seen that the larger errors corre-
spond to sessile drops of low contact angle and pendant drop
P3, which does not have an equatorial diameter or a ‘‘neck.’’
Otherwise, the accuracy is better than 1% for sessile drops
with uc ú 907 and better than 0.5% for pendant drops with
a neck, particularly when using N ¢ 50 points.
11. DISCUSSION
Recent advances in ADSA techniques have been pre-
sented in this paper: new versions of the ADSA-D and
ADSA-P methods that overcome the main limitations of
earlier versions, e.g., numerical instability and failure to con-
verge for certain ultralow interfacial tensions and well-de-
formed drop shapes, and a new method called ADSA-HD
that computes surface tension from the height and diameter
of sessile and pendant drops. Advanced, general-purpose
numerical methods that are well known for their numerical
stability have been used, making the software more flexible
TABLE 7
Performance of ADSA-P for Initial Values
of c from 0.1c to 2.0ca
0.1c 0.5c 0.9c 1.1c
tN (s) — — 0.36 0.27
tLM (s) 1.30 1.30 1.30 1.22
cerror (cm02) 3.3e-06 2.4e-06 7.6e-07 9.7e-07
a
Drop P1 with N Å 21 points was used. tN and tLM are the computational
times required using the Newton and Levenberg–Marquardt optimization
methods, respectively.
FIG. 6. Computational performance of ADSA-P for sessile and pendant
drops using Newton and Levenberg–Marquardt optimization methods.
Computational times measured on a SPARCstation 10 computer (60 Mhz
SuperSPARC CPU). Similar results can be obtained on other computer
systems, including personal computers.
and easier to implement. No numerical simplifications were
made, to make the programs as accurate as possible. This
strategy has proven successful as can be seen by the wide
range of applicability of the methods, i.e., the wide range
of surface tensions, drop sizes, and variety of drop shapes
for which the programs succeed. Ultralow interfacial ten-
sions and well-deformed drop shapes are no longer a limita-
tion of the methods.
Numerical simulations, performed with the aid of a pro-
gram that generates theoretical drop profiles (ALFI), show
the applicability of the methods. Contact angles can be mea-
sured with an accuracy better than 17 with ADSA-D from
the volume and diameter of sessile drops, and an accuracy
of 0.17 is readily obtainable for drops with low contact
angles, e.g., uc õ 307. However, more accurate measure-
ments are required to obtain accurate contact angles for drops
near 1807 contact angle. ADSA-HD can be used to estimate
surface tension by measuring only height and diameter of a
pendant or sessile drop with an accuracy of about 2% for
drops with an equatorial diameter, i.e., sessile drops with uc
1.5c 2.0c ¢ 907 and pendant drops with a neck. ADSA-P can be used
for more accurate surface tension measurements, with which
— — an accuracy better than 0.5% is readily obtainable, with bet-
1.33 1.41 ter results for pendant drops with a neck and sessile drops
2.0e-06 1.9e-06 with uc ú 907.
Since the input required by ADSA programs consists of only
a few drop profile coordinate measurements, they can be used
with any kind of data acquisition, although digital image analy-

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

146 DEL RIO AND NEUMANN

TABLE 8
Computational Results of ADSA-P for Sessile (S) and Pendant (P) Drops after Perturbing Every Profile Coordinate by a Random
Normal Distance, with a Maximum Perturbation Equivalent to Approximately the Distance between Two Consecutive Pixels in a 640
1 480 Digital Image

N Å 21 N Å 51 N Å 101 N Å 501

c (cm02) cerror (%) c (cm02) cerror (%) c (cm02) cerror (%) c (cm02) cerror (%)

S1, 1207 13.419 0.22 13.415 0.26 13.407 0.32 13.408 0.31
S2, 107 13.428 0.16 13.733 2.11 13.700 1.86 13.517 0.50
S3, 807 14.371 6.84 13.667 1.61 13.679 1.70 13.426 0.18
S4, 307 1.146 14.6 1.035 3.54 0.953 4.73 1.020 2.03
S5, 1807 991.76 0.82 992.33 0.77 991.99 0.80 993.13 0.69
P1, 1107 013.432 0.14 013.433 0.13 013.428 0.16 013.425 0.19
P2, 1437 013.425 0.18 013.436 0.10 013.447 0.02 013.427 0.17
P3, 187 013.179 2.02 012.993 3.40 013.077 2.77 013.305 1.08
P4, 1277 00.995 0.52 00.996 0.35 00.997 0.27 00.998 0.21
P5, 1097 0997.03 0.30 0996.66 0.33 0997.93 0.21 0998.06 0.19

sis is recommended for more accurate results and complete ACKNOWLEDGMENTS

automation. The software was designed to be used as
standalone programs or as a numerical library that can be easily
integrated with other modules such as image acquisition, image
analysis, and volume and surface area control.
All the programs were written in C, with FORTRAN inter-
faces to make use of the third-party FORTRAN numerical
libraries mentioned throughout the paper. They can be com-
piled on any computer system with C and FORTRAN com-
pilers and can be obtained from the authors.
This work was supported by the Natural Science and Engineering
Research Council of Canada. The DVERK FORTRAN library was kindly
provided by W. H. Enright, Department of Computer Science, University
of Toronto. The COLSYS, MINPACK, LAPACK, and BLAS FORTRAN
libraries are available on the Internet at the Netlib Repository at the Univer-
sity of Tennessee, Knoxville, and the Oak Ridge National Laboratory
(http://www.netlib.org).
REFERENCES

1. Rotenberg, Y., Boruvka, L., and Neumann, A. W., J. Colloid Interface

APPENDIX: ABBREVIATIONS Sci. 93, 169 (1983).
2. Cheng, P., Li, D., Boruvka, L., Rotenberg, Y., and Neumann, A. W.,
Colloids Surf. 43, 151 (1990).
ADSA Axisymmetric drop shape analysis 3. Cheng, P., and Neumann, A. W., Colloids Surf. 62, 297 (1992).
4. Lahooti, S., del RıB o, O. I., Cheng, P., and Neumann, A. W., in ‘‘Applied
ADSA-P ADSA based on profile coordinates of pendant
Surface Thermodynamics’’ (A. W. Neumann and J. K. Spelt, Eds.),
and sessile drops Chap. 10, pp. 509–556. Marcel Dekker, New York, 1996.
ADSA-D ADSA based on the maximum diameter of ses- 5. Kwok, D. Y., Chiefalo, P., Khorshiddoust, B., Lahooti, S.,
sile drops Cabrerizo-VıB lchez, M. A., del RıB o, O. I., and Neumann, A. W., in ‘‘Sur-
ADSA-HD ADSA based on the height and diameter of factant Adsorption and Surface Solubilization’’ (R. Sharma, Ed.), ACS
Series 615, Chap. 24, pp. 374–386, Am. Chem. Soc., Washington, DC,
pendant and sessile drops
1995.
ALFI Axisymmetric liquid–fluid interfaces 6. Miller, R., Treppo, S., Voigt, A., Zingg, W., and Neumann, A. W.,
BLAS Linear algebra package of the University of Colloids Surf. 69, 209 (1993).
Tennessee 7. Kwok, D. Y., Cabrerizo-VıB lchez, M. A., Gomez, Y., Susnar, S. S., del
COLSYS Boundary-value problem solver by Ascher and RıB o, O. I., Vollhardt, D., Miller, R., and Neumann, A. W., in ‘‘Dynamic
Bader (28) Properties of Interfaces and Association Structures’’ (V. Pillai and
D. O. Shah, Eds.) Chap. 14, pp. 278–296. AOCS Press, Champaign,
DVERK Runge–Kutta–Verner initial value problem IL, 1996.
solver by Hull et al. (25) 8. Padday, J. F., in ‘‘Surface and Colloid Science’’ (E. Matijević, Ed.),
INTRP High-order interpolation routine of DVERK Wiley, New York, Vol. 1, p. 101. 1968.
LAPACK Linear algebra package of the University of 9. Neumann, A. W., and Good, R. J., in ‘‘Experimental Methods in Sur-
Tennessee face and Colloid Science’’ (R. J. Good and R. R. Stromberg, Eds.),
Vol. 11, pp. 31–91. Plenum, New York, 1979.
LMDER Levenberg–Marquardt with derivatives sub- 10. Jennings, J. W., Jr., and Pallas, N. R., Langmuir 4, 959 (1988).
routine of MINPACK 11. Bashforth, F., and Adams, J. C., ‘‘An Attempt to Test the Theory of
MINPACK Minimization package by Moré (31) Capillary Action.’’ Cambridge, 1892.

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas

 AXISYMMETRIC DROP SHAPE ANALYSIS
12. Maze, C., and Burnet, G., Surf. Sci. 13, 451 (1969).
13. Skinner, F. K., Rotenberg, Y., and Neumann, A. W., J. Colloid Inter-
face Sci. 130, 25 (1989).
14. Moy, E., Cheng, P., Policova, Z., Treppo, S., Kwok, D. Y., Mack,
D. R., Sherman, P. M., and Neumann, A. W., Colloids Surf. 58, 215
(1991).
15. Malcolm, J. D., and Elliot, D. D., Can. J. Chem. Eng. 58, 151 (1980).
16. Prokop, R. M., del RıB o, O. I., Niyakan, N., and Neumann, A. W., Can.
J. Chem. Eng. 74, 534 (1996).
17. Schürch, S., Bachofen, H., Goerke, J., and Possmayer, F., J. Appl.
Physiol. 67, 2389 (1989).
18. Schürch, S., Bachofen, H., Goerke, J., and Green, F., Biochim. Biophys.
Acta 1103, 127 (1992).
19. Schürch, S., Schürch, D., Curstedt, T., and Roberston, B., J. Appl.
Physiol. 77, 974 (1994).
20. Schoel, W. M., Schürch, S., and Goerke, J., Biochim. Biophys. Acta
1200, 281 (1994).
21. Dimitrov, A. S., Yoshimura, H., and Nagayama, K., Langmuir 11, 3937
(1995).
22. Hull, T. E., Enright, W. H., and Jackson, K. R., Appl. Numer. Math.
22, 225 (1996).
147
23. Lambert, J. D., ‘‘Numerical Methods for Ordinary Differential Sys-
tems: The Initial Value Problem.’’ Wiley, Chichester/New York, 1991.
24. Hartland, S., and Hartley, R. W., ‘‘Axisymmetric Fluid–Liquid Inter-
faces.’’ Elsevier, Amsterdam, 1976.
25. Hull, T., Enright, W. H., and Jackson, K. R., ‘‘User’s Guide for
DVERK: A Subroutine for Solving Non-stiff ODEs.’’ Tech. Rep. 100,
Department of Computer Science, University of Toronto, 1976.
26. Enright, W. H., Higham, D. J., Owren, B., and Sharp, P. W., ‘‘A Survey
of the Explicit Runge–Kutta Method.’’ Tech. Rep. 291, Department
of Computer Science, University of Toronto, 1995.
27. Kwok, D. Y., Vollhardt, D., Miller, R., Li, D., and Neumann, A. W.,
Colloids Surf. A Physicochem. Eng. Aspects 88, 51 (1994).
28. Ascher, U. M., Mattheij, R. M., and Russell, R. D., ‘‘Numerical Solu-
tion of Boundary Value Problems for Ordinary Differential Equations.’’
Prentice–Hall, Englewood Cliffs, NJ, 1988.
29. Enright, W. H., and Muir, P., ‘‘A Runge–Kutta Type Boundary Value
ODE Solver with Defect Control.’’ Tech. Rep. 267, Department of
Computer Science, University of Toronto, 1993.
30. Fletcher, R., ‘‘Practical Methods of Optimization.’’ Wiley, Chichester,
New York, 1987.
31. Moré, J. J., and Wright, S. J., ‘‘Optimization Software Guide.’’ SIAM,
Philadelphia, 1993.

AID JCIS 5214 / 6g38$$$$21 12-22-97 11:02:57 coidas