Annali di Chimica, 96, 2006, by Societ Chimica Italiana

(), Corresponding Authors; E-mail: Fax: (++39)-0432-558803; e-mail: pierluigi.polese@uniud.it;

Fax: (++39)-0432-558803; e-mail: marilena.tolazzi@uniud.it.



A NOVEL MULTIPURPOSE EXCEL TOOL FOR EQUILIBRIUM
SPECIATION BASED ON NEWTON-RAPHSON METHOD AND ON A
HYBRID GENETIC ALGORITHM


Silvia DEL PIERO, Andrea MELCHIOR, Pierluigi POLESE(), Roberto
PORTANOVA, Marilena TOLAZZI()


Dipartimento di Scienze e Tecnologie Chimiche, Universit di Udine, Via
Cotonificio 108, I-33100 Udine, Italy


Summary - A new algorithm for simulation of chemical equilibria is developed,
based on classical Newton-Raphson method applied to mass balance. This tool,
named EST (Equilibrium Speciation Tool), is improved by using a robust Genetic
Algorithm. In addition, EST works by using Excel spreadsheets and therefore offers
the innovation of a great simplicity and versatility. In fact, it allows the users to
simulate, or to obtain from experimental data, desired chemical-physical parameters
as well as to interact with other available or freely created Excel tools. The reliability
of this utility is here proved by comparison with some published data by other
authors, concerning both complicated homogeneous and heterogeneous equilibria. In
addiction its flexibility is tested computing thermodynamic parameters by using
experimental calorimetric data referred to the complex formation of cobalt(II) with a
macrocyclic ligand. A brief review and comparison of the relative robustness and
quickness of main numerical methods are also reported.


INTRODUCTION

The simulation of ionic and non-ionic speciation is of great importance in many fields: for example,
bio-inorganic- chemists are interested in the speciation of fluids such as blood and urine, geo- and
marine- chemists in studying the distribution of species in sea, estuarine, lake, river and ground
water, whereas analytical- chemists often match with general speciation problems and need
implementation of analytical methods. Also metallurgists are interested to study the distribution of
species involved in corrosion and environmental chemists to the speciation of pollutants in aquatic
systems. Speciation routines have therefore undergone great development and grown up as one of
the core components of most computer programs devoted to the determination of solution
equilibrium constants. In addition, more and more interest lies in obtaining simulated chemical-
physical parameters, such as spectroscopic and thermodynamic quantities.
In spite of the large availability of some excellent softwares devoted essentially to three main areas,
i.e. chemical equilibria simulations,
1-9
thermodynamical parameters determination,
10-21
geochemical
modelling,
22-35
the need of flexible and easy to use tools is still present.
DEL PIERO and coworkers

For example, the Equilibria Package
36
has been recently proposed, which provides various Java
Classes (an object-oriented language
37
) offering the additional opportunity to be used for solving
equilibria problems in a Web based interface. Unfortunately the Web browsers till now available is
not able to provide the manipulating and graphing ability which are, on the other hand, typical of
spreadsheet software. Even the recent version 3.0 of CHEMEQL
34
, and JCHESS a graphical
interface for CHESS
29
which are based on JAVA applets, works essentially as a traditional
programs.
Some other developed applications
38-40
use spreadsheets or macro formulas present in one of the
most diffuse spreadsheet software such as Microsoft Excel, but they are generally limited to face
specific, simple systems. On the other hand, we have no news of a more systematic approach to
speciation by using Excel, with the exception of Equilibrium Expert. This is a Microsoft Excel Add-
in for multiple equilibria simulations and parameter estimations
41
, but is essentially devoted to
protein binding, thus not being so suitable to solve multicomponent/multiphase systems.
In this contest a new application, working by using Microsoft Excel and named EST (Equilibrium
Speciation Tool), is here presented together with a brief review of main aspects of simulation of
chemical equilibria in solutions.
EST is based on classical stoichiometric mass balance equations which allow the simulation of the
free concentrations of the species involved in multiple equilibria: its great innovation consists in the
flexibility of this software as the input and output data can be manipulated very freely by the users,
taking advantage of the versatility of Excel. So people are allowed to build graphics but also to
obtain chemical-physical parameters from both simulated and experimental data. Last but not least,
EST is also able to interact with other available (Solver
34,42
, SolverStat
43
) or freely created Excel
tools depending on what the users require.
The reliability of this utility is here tested by comparison with some published data by other
authors, concerning both complicated homogeneous
44
and heterogeneous equilibria
45
, its flexibility
by computing thermodynamic parameters using previous collected experimental calorimetric data
46

referred to the complex formation of cobalt(II) with a macrocyclic ligand.
The writing of EST, also stimulated a comparison with main numerical methods for the resolution
of non linear equations systems, in order to evaluate their relative robustness and quickness: the
detailed results are also reported.


SIMULATION METHODS

There are two general approaches to simulate complex equilibria systems: (i) Gibbs free-energy
minimisation and (ii) equilibrium constant method. The former is based on the minimisation of the
Gibbs energy of the considered system using stoichiometries based on explicit chemical reactions
(stoichiometric algorithms) or approaches in which no explicit stoichiometric analysis of the system
are used (nonstoichiometric algorithms)
8
.
The nonstoichiometric methods are more suitable for multiphase complex systems and preferred in
the geochemical, and chemical-engineering applications
47-49
as the resolution of large systems, even
in non standard conditions, is relatively easier.
Approach (II) is more direct and used by most chemists which are familiar in dealing with ionic
equilibria. It is based on the solution of a set of equilibrium conditions satisfying stoichiometric
mass balance equations.
To solve these equations various approaches are used
50
, the most common being (a) Newton
Raphson (NR) resolution of the resulting nonlinear system of equations
3,8,51
and (b) iterative
resolution of one equation at time
1,4,6
. Method (a) ensures a rapid quadratic convergence but requires
An Excel Tool for Equilibrium Speciation
3
initial estimates close enough to the true solution, (b) is generally slower. Both methods (a) and (b)
can diverge in particular chemical systems. To bypass these inconveniences, a variety of tricks and
procedures has been proposed in the course of the years and the research of new methods still
continues.
For example Mioshi and do Lago
52
proposed an algorithm based on extent of reaction instead of
mass balance. Brassard and Bodurtha
53
studied various minimisation routines implemented in
MATLAB (Quasi-Newton, Gauss-Newton, Levemberg-Marquardt (LM), NR and the Nelder Mead
Simplex) and proposed an algorithm based on LM method using a cubic polynomial line search
and, in the small number of cases when convergence fails, a robust Simplex method. Colonna and
D'Angola
54
have developed a stoichiometric algorithm based on a modified hierarchical approach in
which one reaction at time was solved. The authors checked it successfully on a single phase
problem and claimed that it can be easily extended to multiphase equilibria. Other general or
particular approaches to chemical equilibria have been also developed as a result of the great
interest for new tools in this field
55-60
.


EST MATHEMATICAL EQUILIBRIA EXPRESSIONS

In a solution elemental component and derived species can be individuated. At equilibrium, in a
system with m components and n species, for a formation reaction the mass-action laws and mass-
balance may be stated as follows:


[ ]
1
ij
m
a
j j i
i
C A
=
=


(1)

1
[ ]
i
n
A i ij j
j
T A a C
=
= +

(2)

where i denotes a component and j a species;
j
is the formation constant of the j species, a
ij
is the
stoichiometric coefficient of the component i in the species j, C
j
is the concentration of a derived
species j, [A
i
] the concentration of free component i and T
Ai
is the total concentration of component
i. Equation (2) can rewritten as:


1
[ ] 0
i
n
T i ij j A
i
j
f A a C T
=

= + =

(3)

The terms in brackets represent the computed total concentration, while T
Ai
is the analytical total
concentration. The equations coming from (2) individuate a nonlinear system of m equations with n
unknown variables (the concentration of free components) that can be solved numerically. The
mentioned treatment can be extended to different type of equilibria.
In the case of Solubility Equilibrium it can be written:


[ ]
1
ik
m
a
k i
i
Ksp A
=
=

(4)

where the index k denotes a partially soluble species and Ksp
k
the correspondent Solubility Product
Constant. Then we can write:

DEL PIERO and coworkers


[ ]
1
0
ik
m
a
sp i k
k
i
f A Ksp
=

= =

(5)
and this expression must be added to the nonlinear system when a precipitate is really formed, that
is when the computed Solubility Product is equal or higher than the correspondent Ksp. In this case,
to the mass-balance expression, a term X
k
must be added which is related to the amount (expressed
as concentration) of the insoluble species formed:


1 1
[ ] 0
i
n s
T i ij j ik k A
i
j k
f A a C a X T
= =

= + + =




(6)

where s is the number of partially soluble species. A similar approach has been used in HYSS
8
.
Similarly for a Partition Equilibria, the Distribution Constant (also called Partition Ratio) KD
l

between the solution and a liquid or gaseous phase, can be written as :

[ ]
[ ]
1
il
l
l m
a
i
i
S
KD
A
=
=

(7)
[ ]
[ ]
1
0
il
l
D l m
l
a
i
i
S
f KD
A
=
= =

(8)

S
l
is the species l formed in the second phase,


1 1 1
[ ] 0
i
n s d
T i ij j ik k il l A
i
j k l
f A a C a X a Y T
= = =

= + + + =




(9)

d the number of Partition Equilibria and


[ ]
0
l
l l
V
Y S
V
=

(10)

is the concentration of the species l reported to the volume V
0
of the solution and V
l
the volume of
the second phase. Using the three forms of equilibria defined above in (1), (4), (7) and equations
(5), (8) and (9) we can manipulate most of the typical equilibria problems.

Non Ideality
EST can account for non ideality of the solutions at relative high concentrations. With the aim of
making EST of general application, the simple and well known semi-empirical relationship of
Davies
61
, derived from Debye-Hckel theory, has been used in EST. This relationship is probably
the most universally applicable estimator for the activity coefficients and can be used to model
solutions up to 0.1-0.5 mol kg
-1

62
. EST users can modify the two parameters of the Davies equations
in order to mimic the analysed systems more and more accurately.

Acid-Base and Redox Reactions
EST can work with pK
a
, pK
b
and Redox Potential, which are internally converted to Formation
Constant. In aqueous solutions the electron activity is used to compute equilibrium expression
63
, but
the free concentration of electron is set to zero.
An Excel Tool for Equilibrium Speciation
5

RESOLUTION OF NONLINEAR EQUILIBRIUM EQUATIONS

The numerical core of a speciation program is the resolution of nonlinear systems with u unknown
variables defined by equations (5), (8) and (9): one of the oldest methods used to achieve this goal
is the Newton Raphson one whose limits and opportunities are well known
3,8,51,64
. As reported in
Newton Raphson Method section below, this method occasionally fails to converge. In this case
more robust numerical methods must be used, which, on the other hand, present the disadvantage to
be slower, as discussed in section More Robust Methods.

Newton Raphson Method
The NR method is an iterative procedure based on the Taylor expansion of the objective function
around its root. It requires the computation of the Jacobian matrix J, that is the matrix of partial
derivatives of each functions f, as definite above, with respect to each unknown variables x. In
matrix notation we have:

( + ) + = F x x F(x) J x
(11)
= J x F(x)
(12)

where F is the vector of functions f. Equation (12) can be easily solved to obtain the shift vector x
by LU decomposition. The computed corrections are then added to the solution vector:


( 1) ( )
+
h h +
= x x x (13)

where (h) is an indicator of the iteration. Equations (12) and (13) are applied repeatedly until
convergence of functions F and variables x is reached (see
65
for more computational details).
This method presents the advantage of rapid convergence in few iterations if the starting values of
the variables are close to the true solution. Nevertheless it is subject to some problems: when the
initial guess is poor, the numerical method may fail to converge by exhausting its maximum
prescribed number of iterations or, worse, returning to a local minimum; when J is singular or near
singular, resolution of (12) is impossible and causes numerical overflow. In addition, this method is
especially vulnerable to near zero gradient system and to highly nonlinear system with more than
five equations.
In order to partially overcome these troubles, various procedures can be applied.
Damping: used to improve convergence capacity with bad initial estimates. This iterative
procedure
1
, represented in (14), has given excellent results and has been differently
implemented
5,6,51,64
:


1/
Anal
( 1) ( )
Calc
[ ] [ ]
i
i
g
A h h
i i
A
T
A A
T
+

=



(14)

where Anal refers to Analytical Total Concentration, Calc is the Calculated Total Concentration
and g was computed as the average of the stoichiometric coefficient of Equation (2), weighted for
species concentrations. The above iterative procedure was applied until the highest ratio of
Calculated vs. Analytical Total Concentration is lower than 10 and the lowest is higher than 0.1.
Scaling: used to overcome singularity problems, and applied to matrix J and to vector F according
to equations:

DEL PIERO and coworkers


( )
1/ 2
*
ab ab aa bb
j j j j

=

(15)

* * 1/ 2
a a aa
f f j

=

(16)

where
*
ab
j and
*
a
f are the elements of the scaled matrix with dimensions u x u and vector
dimensions u, respectively
3,51
.
Convergence Forcer: another procedure used if initial estimates are not good. Divergence can be
prevented viewing the process as a minimisation of the function g(x):


2
1
( ) (1/ 2) =(1/ 2) ( )
u
i
i
g x f x
=
=

T
F F

(17)

As the Newton step is in a descent direction for g(x), a good strategy is simply to find the value of a
stepsize controller with a minimum located in the domain (0,1]. Equation (13) can be rewritten as:


( 1) ( )
+
h h

+
= x x x (18)

Although various approaches have been used to determine such as Hartley Method
3
and Cubic
Interpolation
65
, the following procedure is proposed in EST:


2
min
( , 0) ( , 1)
=
= -2 ( , 0)
- (2 )
g x g x
g x

= = + =

(19)

which has been developed assuming a parabolic trend with a minimum of zero in the domain [0,1].
This formula presents the advantage that the evaluation of g(x) is needed only at the full NR step for
=1.
Step Limiter: as the NR procedure can determine negative concentrations (when x < -x in (13)) it is
necessary to restrict this step. This can be done using the approximation EXP(-) ~ (1 - ), where
is a small number. Equation (13) can be rewritten as suggested in
3
:


( )
( +1) ( )
x x EXP
h h
x =

(20)

This ensures that x
(h+1)
0.
Computation of Jacobian: the calculation of the Jacobian in EST has been implemented both
algebraically and numerically by finite difference approximation.

More Robust Methods
Occasionally, despite the improvements described above, sometimes NR method fails, essentially
becouse the initial estimates are too much different from the true values. For example, as far as a
typical titration is concerned, this may occur at the start point, at the first addition, at an equivalence
point, or when a phase appears or disappears. In these cases, a more robust approach to the
resolution of equilibria equations is needed.
Three common robust non-derivative methods have been here analysed: Simplex, Simulated
Annealing and Genetic Algorithm. Our main aim was therefore to select a search method with good
ability to escape from difficult local minima and zero-gradient region. In order to do that, Equation
(17) has been therefore minimised.
An Excel Tool for Equilibrium Speciation
7
Simplex Method
The original Simplex Method was proposed by Spendley et al. in 1962
66
and enhanced in 1965 by
Nelder and Mead
67
with a simple but powerful modification to the original method.
This Nelder and Mead Downhill Simplex (DS) is a classical very powerful local descent algorithm,
making no use of the objective function derivatives. This procedure derives its name from the
geometrical figure which moves across the function surface in search of the minimum: a simplex
is a geometrical figure defined by one more point than the number of the unknown parameters in
the function. For example, if the function contains three unknown parameters, the simplex will
consist of four vertices in a three-dimensional space. Each point represents a set of parameter values
and is associated with a vector in the parameter space.
Through a sequence of elementary geometric transformations (reflection, contraction, expansion
and multi-contraction), the initial simplex moves, expands or contracts. To select the appropriate
transformation, the method simply uses the values of the function to be optimised at the vertices of
the simplex considered. After each transformation, the current worst vertex is replaced by a better
one.
At the beginning of the algorithm, only the point of the simplex, where the objective function is
worst, is moved and another point, image of the worst point, is generated. This operation is the
reection. If the reected point is the best among all the points, the method expands the simplex in
this direction, otherwise, if it is at least better than the worst, the algorithm performs again the
reection with the new worst point. The contraction step is performed when the worst point is at
least as good as the reected point, in such a way that the simplex adjusts itself to surround the
optimum. If the worst point is better than the contracted point, the multi-contraction is performed.
At each step the generated point is checked to be not outside the allowed reduced solution space
65,68
.
It is not very efficient in terms of the number of function evaluations that it requires but, despite its
slowness, it is considered robust enough and may frequently be the election method to be used for
problems whose computational burden are small.
In EST the Amoeba
65
version of DS is used, with some added improvements. In particular, the
routine is restarted at the last minimum, until the same values are three times obtained.
Simulated Annealing
The first idea of Simulated Annealing (SA) was introduced by Metropolis et al. in 1953
69
but the
application to deterministic optimisation problems comes from the work of Kirkpatrick et al. in
1983
70
. SA is a stochastic search method analogous to the physical annealing process where a
molten compound is cooled gradually so that a minimal energy state is achieved. If the system is
cooled too fast, defects in the molecule configuration will be frozen and the system will be left in a
metastable state, i.e. in a local energy minimum (e.g. a crystal with several defects in its structure).
Thermal equilibrium at a given temperature is characterised by a Boltzmann distribution function of
the energy states. Under these conditions, even at a low temperature, although with a small
probability, a transition may occur from a low to a high energy level. Such transitions are assumed
to be responsible of the ability of SA to escape from local minimum energy states.
The Metropolis algorithm was the first one proposed to simulate this process. Starting from a high
energy state, corresponding to a particular system temperature T
(i)
, a series of new energy states E
(h)
,
with (h) (i), are stochastically generated. Each new system configuration is accepted if E
(h+1)
E
(h)
.
Otherwise, and by analogy with the Boltzmann distribution for energy states at thermal equilibrium,
E
h
will be accepted as an improved state with a probability P determined by:


( )
( +1) ( ) ( )
P=EXP -(E E ) / T
h h i
B
K

(21)

DEL PIERO and coworkers

where T
(i)
is the current system temperature and K
B
is Boltzmann's constant. At high temperatures
this probability is close to one, i.e. most energy transitions are permissible. As the system
temperature decreases, the probability of accepting a higher energy state as being an improved
energy state approaches zero, and it is assumed that thermal equilibrium is reached at each
temperature. From an optimisation point of view, SA explores the key feature of the physical
annealing process of generating transitions to higher energy states, applying to the new states an
acceptance/rejection probability criterion which should naturally become more and more stringent
with the progress of the procedure. For solving a particular problem with a SA algorithm, the
following steps are thus necessary:
(i) Definition of an objective function to be minimised.
(ii) Adoption of an annealing cooling schedule, i.e. the initial temperature, the number of
configurations generated at each temperature and a method to decrease the temperature itself must
be specified.
(iii) At each temperature in the cooling schedule, stochastic generation of the alternative
combinations, centred on the currently accepted state must be made.
(iv) Criteria for the acceptance or rejection of the alternative combinations, against the currently
accepted state at that temperature, must be adopted.
Three SA algorithms have been analysed with the proposal to select the best to be used in EST:
SIMMAN
71
, the hybrid SA-Simplex Amebsa
65
and a code developed in our laboratory (EST-SA).
The latter is based on Gaussian distribution with adaptive variance for the randomisation process
and a variable length exponential cooling schedule
72,73
. The variance for the randomisation process
and the exponential cooling schedule are driven by the rejection ratio with the Markov chain length
set to four times the number of parameters
72
.
Genetic Algorithms
Genetic Algorithms (GA) are a group of mathematical techniques which were initially designed to
simulate the behaviour of biologically based adaptive systems. Despite GA were not initially
developed as an optimisation technique in itself, it can be modified to produce a powerful optimiser
based on parallel search technique
74,75
.
Introduced by Holland in 1975
76
, it consists of an iterative procedure which mimics the evolutionary
process in nature. There are numerous variant of GA, but the classical architecture includes the
three components Selection, Crossover and Mutation.
It starts with a randomly selected population of potential solutions: each individual is characterised
by a chromosome composed by genes in which the parameters are encoded. The initial population,
by means of Genetic Operators, undergoes a simulated evolution, which copies the natural selective
reproduction scheme, survival of the fittest. The Fitness of the individual is estimated by the
objective function, which plays the role of an environment. This Fitness is then used to direct the
application of the operations, which produces a new population (a New Generation). The new
population is formed by selecting chromosomes with a probability relative to their Fitness
(Selection). As this reproduction rule can lead to a decrease of chromosome diversity in the next
few generations, the two genetic operators, Mutation and Crossover are introduced to
counterbalance this phenomenon. Mutation works on the single genes level by changing at random
single gene of individuals, with the probability of this event usually being very small. Crossover,
the recombination operator of GA, is the main variation operator which hopefully recombines
useful genes from different individuals. The Crossover Rate indicates the probability per individual
to undergo recombination. Then with the two parent individuals, selected (at random) from the
population, crossover forms two new offspring individuals. The new population of chromosomes is
again individually evaluated and transformed into a subsequent generation. This process continues
until convergence within a population is achieved.
An Excel Tool for Equilibrium Speciation
9
EST-GA has been developed from the skeleton of Genetik
77
of which the general scheme is hold
with a floating point representation. From this program, our GA maintains also the Elitist Strategy
(the preservation of the best solution during the application of genetic operators), the Tournament
Selection and the Whole Linear Crossover. The routine incorporates numerous suggestions from the
works of Brunetti
78
and Nikitas
79,80
.
EST-GA has been improved using a Gaussian distribution for the Generation of New Individuals;
the standard deviation of the genes in the new individuals is controlled by hits in the last generation:
if a success is gained in objective function reduction, the standard deviation is doubled, otherwise is
halved. This enhances the capability of exploration and convergence. The Mutation Probability is
controlled by the variance of population; this ensures an adequate element of random search while
mutation becomes more productive and crossover less productive, as the population converges.
Finally the whole population is replaced (with the exception of the best individual) if there is no
improvement for a predetermined number of generations. The termination also occurs when there is
no improvement after a fixed number of cycles. Main genetic parameters in EST-GA are Population
of 50 individual, Crossover Probability of 0.85 and Mutation Probability of 0.01-0.1.
To improve convergence capacity, all SA and the GA have been hybridised with Amoeba so the
results of the former method give the seed of the latter one. In addition, for the sake of
completeness, the relative performance of LM and NR methods heve been analysed by
implementing the routines Mrqmin and Newt
65
as VBA code and by using a simplified version of
EST-NR algorithm.
Finally, some other common used optimisation algorithms, such as the Powells Quadratically
Convergent Direction Set Method in Multidimensions, the Conjugate Gradient Fletcher-Reeves-
Polak-Ribiere Minimisation (FRPR), and the Broyden-Fletcher-Goldfarb-Shanno variant of the
Quasi-Newton Davidon-Fletcher-Powell Minimisation (BFGS)
65
, have been taken into account but,
as their performances were found to be not satisfactory, they were dropped out.
Other algorithms, such as the MINPACK libraries or TENSOLVE
57
have been analysed, but not
considered suitable for EST application, as they are based on very complicated globalisation
strategies
81
: on the contrary, our target was the building of a very flexible tool without black boxes.

Results of Tests on Computational Methods
Tests on the reliability of analysed methods have been run on a large number of minimisation
problems: among them, we only remember those carried out on functions such as Rosenbrock
68
, on
sets of data from the nonlinear regression of Statistical Reference Dataset
s(http://www.nist.gov/itl/div898/strd) of National Institute of Standards and Technology (NIST),
such as the MGH10, and on a typical equilibrium problem here reported as ML. In TABLE 1 we
report the results of only the last two representative tests.
MGH10 is a nonlinear regression with a high level of difficulty. ML is a simple example of
formation of chemical complex 1:1 with log
ML
= 10, total component concentration of each
component 1 mol dm
-3
, starting values in interval 10
-10
- 1 mol dm
-3
with uniform random probability
in the logarithmic range. This simple example underlines the problem of good initial guesses. The
table reports the relative execution time, the successful ratio (defined as the number of located on
solution with relative error less of 1%), the average (avLRE) and the standard deviation (sdLRE) of
Log Relative Error (LRE):


10
q-c
LRE=-log
c





(22)

DEL PIERO and coworkers

where q is the estimated value and c the correct value
82
. The definition of successful rate used here
is not the classical one, but it is useful because our objective is to obtain a procedure that can be
used as seed for the NR routine.


Examining TABLE 1 the first observation is that the deterministic methods (e.g. NR and LM) can
be faster than stochastic algorithms (e.g. GA and SA), which are, on the other hand, more
computationally expensive. NR appears more reliable then LM. This could be a surprise as the latter
is considered more robust: we should remember that a modified NR algorithm is compared in this
case and so the differences can be ascribed to the supplementary trick introduced, which is able to
increase the speed and prevent numerical failures.
Both GA and SA show problems of premature convergence and slow finishing (namely they locate
promising region but was unable to converge or converge to local minima), but GA appears more
efficient. In literature the relative performances of these two algorithms are controversial: it is likely
that SA works better in some cases and worse in other
74
. From our experience, SA results elegant
and offers the opportunity to be easily implemented, but the tuning of parameters is problem-
dependent. GA is slightly more complex but performances are more robust to tuning. Either
Amebsa or SIMANN performances are slightly worse than those obtained by EST-SA.
Amoeba shows an intermediate behaviour between deterministic and stochastic methods. In fact it
is a robust but local optimiser: it is known that this method can be trapped in local minima and
plateau regions
83
. During the improvement of EST, we have however observed that in the handling
of speciation problem, Amoeba can be frequently faster than the robust EST-GA when some
parameters are near the optimum.
From data in TABLE 1 it emerges that when Amoeba is associated to SA or GA, the performances
improve notably as the search is confined to most promising regions: yet EST-GA associated to the
Simplex, works better. The best algorithms, emerged from the analysis of TABLE 1, are again
applied to ML, when the starting values are in the range 10
-20
- 1 mol dm
-3
for 1000 runs, and the
results reported in TABLE 2. This more severe constraint causes the EST-NR method to fail in
about the 40% of trials whereas the hybridised EST-GA/Amoeba presents full success. It can be
also noted that this method is very time consuming. In the reality the simple EST-GA alone can be a
TABLE 1 - Results of computational tests (for 100 runs) on MGH10 (starting values as
reported in NIST web site) and on ML (starting values in the range 10
-10
- 1
mol dm
-3
).

MGH10 ML
Average Relative
Time
Successful
Ratio
avLRE sdLRE Average Relative
Time
Successfu
Ratio
avLRE sdLRE
Optimisation
Routines

Amoeba 1.00 0.92 6.75 2.32 1.00 0.95 4.91 1.79
Amebsa 0.073 0.00 -1.52 1.17 0.045 0.00 -0.67 1.16
Amebsa + Amoeba 0.64 0.98 7.21 1.44 1.39 1.00 5.30 0
SIMANN 1.85 0.00 -2.28 0.61 0.028 0.00 -0.77 1.02
SIMANN + Amoeba 2.79 0.96 7.09 1.79 1.31 1.00 5.30 0
EST-SA 0.71 0.00 -1.65 0.89 0.082 0.00 -0.49 0.91
EST-SA + Amoeba 1.10 0.99 7.30 1.09 1.50 1.00 5.30 0.00
EST-GA 0.43 0.00 -0.74 1.44 0.12 1.00 5.30 0.03
EST-GA + Amoeba 0.67 1.00 7.48 0.87 1.27 1.00 5.30 0.00
Deterministic
Methods

Mrqmin 0.94 0.10 -1.70 4.95 0.0026 0.66 3.44 2.56
Newt - - - - 0.0015 0.96 2.88 1.00
EST-NR - - - - 0.0014 1.00 5.30 0.00

An Excel Tool for Equilibrium Speciation
11
good seed for the NR method and in few difficult cases only, the use of the slower hybridised EST-
GA/Amoeba is needed.



At this point the strategy to build a robust and quick speciation program is clear: first our quickly
deterministic method (EST-NR) is run: if it fails, EST-GA is used, followed by NR to obtain an
accurate solution. If the work is still unsuccessful, the hybridised EST-GA/Amoeba, followed by
NR, is run.


THE INTERFACE

To write EST, Microsoft Excel environment has been used. It offers a friendly user interface,
flexible data manipulation, built-in mathematical functions and instantaneous plotting of data. The
Solver add-in included in Excel is, in addition, ideally suited to fit data with non-linear function.
More over, Excel contains Visual Basic for Applications (VBA), where user-defined functions or
programs can be created by combining mathematical expressions, built-in Microsoft Excel
functions and Visual Basic codes. In this highly flexible environment it is easy to concentrate on the
numerical aspects, without caring about the interfaces. Therefore, VBA has been used to code EST.
An alternative would it be to create EST as DLL (Dynamic Link Library): this method would have
the advantage of the speed (VBA is relatively slow), but it would prevent the users to be able to
modify the code in accordance with their needs.
After the activation, EST introduces a pull-menu through which it is possible to create a new
worksheet suitably formatted, or recalculate a worksheet already existing. An example is visualised
in FIG 1.
Two methods are available: Titration and Distribution. In the worksheet it is possible to distinguish
four areas: in the first the user introduces the model (constants and stoichiometric coefficients). In
the second, the data relative to the volumes of the vessel and the burette and the respective
concentrations are inserted (when Distribution is considered instead of the concentration in the
burette, the pX of the master species is inserted). Additional data as the number of points or options,
such as the calculation of the ionic strength or the volume of additional phases, are introduced as
Excel Notes. At this point the calculation runs and EST writes in the below cells (the third section)
the data of the total concentrations point for point, and on the right, the values of the calculated
concentrations cells (the fourth section).
It is also possible to write manually the data of the total concentration in the third cells. An
automatic procedure is finally available to activate EST calculations when changes in specific cells
TABLE 2 - Best alghoritms (from TABLE 1) applied to ML in
the case of more severe constraints (1000runs, starting
values in the range 10
-20
- 1 mol dm
-3
). Amoeba is
reported as a reference.

ML
Optimisation Routines Average Relative
Time
Successful
Ratio
avLRE sdLRE
Amoeba 1.00 0.79 3.66 3.39
EST-NR 0.00024 0.60 3.20 2.60
EST-GA 0.025 0.16 -0.06 2.01
EST-GA + Amoeba 1.41 1.00 5.30 0.00
DEL PIERO and coworkers

of worksheet are introduced for example by other Excel formulas, VBA or Add-in applications
(such us changing in the values of constants or concentrations).







FIG.1 - Layout of Excel after calorimetric data treatment for Co(II)-2,1 system (Application
3). In the frame the total molar enthalpy changes, h
v
, as a function of added volume
of ligand for Co(II)-2,1 system in dmso are reported. ( ) 0.0288, ( ) 0.0230, ( )
0.00863 mol dm
-3
in Co
2+
.



DATA PROCEEDING BY EST APPLICATION

Of course the reliability of EST has been proved by comparing the results obtained by EST with
published data.

Application 1
The first case analysed by using EST comes from the published data of Bianchi and Garcia-
Espaa
44
. This work has been chosen as a source for some main reasons: a part from the fact that
calculations have been run with another program (HYSS
8
), the system studied is somewhat
complicated and the calculated concentrations are available as supplementary materials. In
particular, in this nice paper the authors present a simple method for the analysis of thermodynamic
parameters for the system ATP (adenosine triphosphate) / 1,4,7,10,13,16 Hexaaza-Cyclooctadecane
(L1) / 1,4,7,13-Tetramethyl-1,4,7,10,13,16 Hexaaza-Cyclooctadecane (L2) which is characterised
by four components and twenty equilibria.
Results of our calculations are absolutely coincident with published data
44
. FIG. 2 shows the
distribution diagrams of the species formed as a function of pH respectively referred as individual
species (FIG. 2a) and as the cumulative amounts of ATP bound to L1 and L2 (FIG. 2b).
An Excel Tool for Equilibrium Speciation
13



4 6 8 10
pH
0
0.0002
0.0004
0.0006
0.0008
0.001
C
o
n
c
e
n
t
r
a
t
i
o
n

(
m
o
l

d
m
-
3
)
H
3
L1
3+
ATP
4-
H
3
L2
3+
H
5
L1ATP
+
H
4
ATPL1
H
2
L2
2+
L2
HL2
+
HL1
+
H
2
L1
2+
H
5
ATPL2
+
H
4
L1
4+
H
4
ATPL2
3+
H
3
ATPL1
-
H
3
ATPL2
-
HATP
3-
(a)

4 6 8 10
pH
0.0000
0.0002
0.0004
0.0006
0.0008
0.0010
(ATP)
Tot
(L1ATP)
Tot
(L2ATP)
Tot
(b)





FIG.2 - Distribution of: (a) the individual species, (b) cumulative amounts of ATP bound to L1
and L2, for the system ATP/L1/L2 (Application 1).


The comparison with published data and figures proof the numerical reliability of EST referred to
homogeneous chemical equilibria.

Application 2
EST is here run on published data by Brown and Ebinger
45
where the authors examine four
precipitation problems which are solved to face the use of numerical equilibrium codes. This study
emphasises concentrated solutions, assumes both ideal and non-ideal solutions, and employs
different databases and different activity coefficient relationships. The study uses the EQ3/6
numerical speciation code
25
. Results show satisfactory agreement between the solubility products
calculated from free-energy relationships and those calculated from concentrations and activity
coefficients.
The third precipitation problem analyzed in the paper
45
was solved with EST, that is the
simultaneous precipitation of calcium sulphate and calcium hydroxide by mixing solutions of
calcium chloride, sodium sulphate, and sodium hydroxide. This problem involves two precipitation
equilibria and is therefore very suitable to test the relative computational algorithm for
heterogeneous systems. The EST calculated concentrations of free and solid species are reported in
TABLE 3 together with previous comparable data
45
. The differences with published data are only
present in the last significative figure thus proving the reliability of EST also in the field of
heterogeneous equilibria.
For non-ideal solution the comparison was not significant as the concentrations considered by the
authors were over the Davies equation range validity.

Application 3
The third case studied by means of EST, takes advantage of same our data concerning the complex
formation of Co(II) with mixed N/O ligands in dimethylsulfoxide (dmso)
46
. In this previous work
the formation constants of CoLj complexes were obtained by means of UV-Vis spectrophotometric
titrations, using Cd(II) as a competitive ion and analysing the data with the Hyperquad program
18
.
DEL PIERO and coworkers

The enthalpy changes were obtained directly from calorimetric titrations by means of the computer
program Letagrop Kalle
84
, using the formation constants spectroscopically determined.




In some cases, such as the cobalt(II)-1,4,10-trioxa-7,13-diaza-cyclopentadecane (2,1) system,
chosen as an example, both the stability constants and the enthalpy changes could be obtained from
the minimisation of only calorimetric data
85
. This is here done by applying EST to the available
calorimetric titration points (for experimental details see Reference
46
). It should be evidenced that
this important third application allows to test its flexibility in the manipulation of input and output
data: therefore it represents a useful tool to be largely applied in every field where a signal can be
correlated to a concentration value by a suitable mathematical function. In addition, the good
interface with Solver and other Excel Add-in tools, such as Solverstat
43
, is also tested.
144 total observations from three titrations were available in the metal concentration range 9-30
mmol dm
-3
. To get our computational target we prepared an Excel sheet opportunely formatted. The
first nine columns forms the speciation region and have a structure identical to other typical
speciation simulations. The next seven columns form the regression area and are formatted for
successive convenience (in order to fit the format requested by the regression statistical utility
SolverStat
43
). The first nine columns contain respectively: the weights (set all to one), the values of
the independent variable (the added titrant volume), the experimental dependent variable (the
experimental enthalpy changes), the calculated values of the dependent variable (the simulated
enthalpy changes) as Excel formula, the squared differences between calculated and experimental
enthalpy changes. In addition, a cell containing the residual sum of squares (RSS) between
calculated and experimental enthalpy changes (above the squared differences) and the initial values
of parameters (K and H) to be minimised, have been entered in the last two columns.
The simulated enthalpy changes can be computed as:
TABLE 3 - Computational results on simultaneous precipitation of calcium sulphate and
calcium hydroxide by mixing solutions of CaCl
2
, Na
2
SO
4
and NaOH. In the
simulated experiment, three one-litre water solutions are mixed at 25C. Before
mixing, one solution contains 500 g of dissolved CaCl
2
, the second contains 100 g
of dissolved NaOH, and the third contains 500 g of dissolved Na
2
SO
4
.


EST EQ3/6
Ca(OH)
2(s) (
g dm
-3
)
28.6 28.6
CaSO
4
2H
2
O
(s)
(g dm
-3
)
192.0 192.0
PH 12.8 12.8
[Ca
2+
] (mol dm
-3
)
0.00121 0.00123
[OH
-
] (mol dm
-3
)
0.062 0.061
[SO
4
2-
] (mol dm
-3
)
0.058 0.058
Ca
2+


OH
-


SO
4
2-


pKsp Ca(OH)
2
5.33
pKsp CaSO
4
2H
2
O 4.15
An Excel Tool for Equilibrium Speciation
15


0
m l M
[M L ] T
i j
v j
j
h H

(23)

where h
vi
is the total molar enthalpy change per mole of metal ion for point i, T
M
the total
concentration of metal, [M
m
L
l
]
j
the concentration and H
j
the molar enthalpy of the j species: this
equation was introduced in the worksheet as Excel formula, linked to the cells containing the
respective values. With this method the simulation of a large number of chemical or physical
quantity correlated to species concentration is therefore possible (such as the absorbance,
conductance, potential, vapour tension and a lot of others).
The minimised values for the formation constants and the molar enthalpy changes can be obtained
using Solver
86
, the powerful optimisation tool available in Excel. This tool was used setting the
residual sum of squares cell as Solver target cell subject to minimisation changing the cells
containing the searched parameters. At this stage the problem was how to run EST in order to
answer to the changes imposed to formation constants by Solver. This can be done using the Excel
Events functions
42
i.e. the capacity of Excel to executing VBA code when a particular event occurs.
The code for this operation is reported in Appendix. This is of particular interest as can be adjusted
to recall EST from VBA users applications. It must be pointed out that in EST an automatic
procedure to get all this is present. Using these tricks, we were able to obtain optimised parameters
from our calorimetric data. The screen of working sheet after optimisation is reported in FIG. 1.
together with the experimental and calculated calorimetric curves (frame).
A great limit of Excel Solver is that it does not provide any regression statistic nor standard
deviation of estimated parameters, which are, on the other hand, essential for the validation of
results. Therefore the regression statistics previously presented in SolverStat
43
have been used to
analyze the goodness of results. In particular the results of regression with Solver and the statistical
analysis with SolverStat, using the method of Regression Triplet
87
, are reported in TABLE 4. In the
table the results for two different analysed models, Model 1, where only the ML complex is taken
into account and Model 2, where even a hypothetical complex ML
2
is considered. The meanings of
almost all the statistical parameters reported are discussed in the SolverStat paper. Here we present
only a brief discussion on results.
The Regression Triplet applied to Model 1, shows that only the Runs of Signs Test fails, being this
failure likely due to (i) the fact that the data come from three groups of data (titrations) and (ii) the
existence of a slight instrumental sinusoidal noise.
When Model 2 was analysed, also the Shapiro-Wilk, the Jarque-Bera Statistics and the F Levene
test fail, evidencing the inadequacy of the model. As no improvement of Model 2 was possible,
Model 1 was chosen as the best. This is in agreement with the results of the original paper. The two
data are reported in TABLE 5.
DEL PIERO and coworkers


TABLE 4 - Regression analysis, parameter values and their errors as found by Excel Solver-
SolverStat for Model 1 and Model 2 in Application 3.




TABLE 5 - Thermodynamic parameters for Co(II)-2,1 system original and
calculated by applying EST (Application 3). The errors quoted
correspond to three standard deviations


Statistical
Parameters

Rejection
Values
a

Model 1 Rejection
Values
a

Model 2
Data Quality Passed Passed
Model Quality
ANOVA F value >3.91
8.3710
5

>2.67
3.3110
5

Runs-of-Signs Test <51 18 <59 10
Regression Method
Quality

Shapiro Wilk Statistic >1.64 1.05 >1.64 3.35
Jarque-Bera Statistic >5.99 2.64 >5.99 7.39
F Value for Levene Test >3.91 1.09 >3.91 8.06
Model Selection
RSS 3.54 2.95
R 0.999915 0.999929
95% CI for R 0.999882 - 0.999939 0.999901 - 0.999949
R
2

0.999830 0.999859
R
2
Adj

0.999829 0.999986
PRESS 3.62 3.11
R
2
Prediction

0.999827 0.999851
AIC -3.68 -3.83

Calculated Model
Parameters

t
Critical
Value
s
b

t
Calculated
t
Critical
Value
s
b

t
Calculated

Log KF
ML
>1.98 3.75 0.03 139.2 >1.98 3.63 0.03 110.0
H
ML
(kJ mol
-1
)
>1.98 14.80 0.03 506.1 >1.98 15.16 0.08 199.3
Log KF
ML2
>1.98 >1.98 4.8 0.2 22.4
H
ML2
(kJ mol
-1
)
>1.98 >1.98 13.1 0.6 23.1


a
Critical Value selected for probability 5% and computed by SolverStat:
b
Asymptotic

Standard Deviation for calculated parameters
Calculated
Parameters
Original paper This work

Log KF
ML
3.7(0.2) 3.75(0.08)
-H
ML
(kJ mol
-1
)
14.8(0.2) 14.80(0.09)


An Excel Tool for Equilibrium Speciation
17
CONCLUSIONS

The proposed powerful EST add-in, with the computational improvements here described,
represents an innovative approach to the field of simulation of multicomponent/multiphase
equilibria. It takes advantage of Microsoft Excel spreadsheet flexibility in order to obtain chemical-
physical parameters both from simulated and experimental data, thus being of great help in the field
of analytical chemistry. The great interest of EST lies also in its easiness and, particularly, in its
ability to interact with other statistical, general purpose Excel tools, such as Solver and Solverstat,
as clearly evidenced in this work: of course, only limited cases are analyzed here by using EST,
being its range of application much wider than what can be proposed in a single paper.
The computational methods developed in EST are shown to be very reliable also when applied to
complicated homogeneous or heterogenous equilibrium systems.
EST can be downloaded, free of charge, from the web site http://www.freewebs.com/solverstat/est/est.htm.

Received September 16
th
, 2005

Acknowledgements This work has been supported by the Ministero dell'Istruzione,
dell'Universit e della Ricerca (MIUR, Rome) within the program COFIN
2002.

APPENDIX

In the Excel macro here reported, the following points should be noted:
1) a link is created between the worksheet and the EST program (see the Microsoft on-line
documentation);
2) the Equilibrium Data are in the A1:D2 range of the worksheet;
3) the total concentrations are introduced in the C7:D150 range;
4) the output of EST starts from H7 cell;
5) only the W5 cell is monitored as far as changes occur and contains the Log K value which is
subject to minimization by Solver;
6) to obtain a working sheet a reference to the EST workbook must be created (see the Excel online
help).
Private Sub Worksheet_Change(ByVal Target As Excel.Range)
' Variables declaration. Names used are informative
Dim ChangedRange As String
Dim InputEquilibriumDataRange As String
Dim InputStepConcRange As String
Dim FirstOutputCell As String
Dim AutomaticRange As String
Dim TotalVolumeRange As String
Dim Intersection As Range
Dim LabelFlag as Boolean

LabelFlag = False
Application.ScreenUpdating = False

' Changed range passed by Excel as argument
DEL PIERO and coworkers

ChangedRange = Target.Address

' Set the EST working ranges
AutomaticRange = "W5"
InputEquilibriumDataRange = "A1:D2"
InputStepConcRange = "C7:F150"
FirstOutputCell = "H7"
TotalVolumeRange = ""

' Test if the controlled range was changed
Set Intersection = Application.Intersect(Range(AutomaticRange),
_
Range(ChangedRange))

' If the controlled range was unchanged then end the macro
execution
If Intersection Is Nothing Then
Exit Sub
End If

Set Intersection = Nothing

' Prevents problem with calculation method of Excel
If Application.Calculation <> xlCalculationAutomatic Then
Application.Calculation = xlCalculationAutomatic
End If

' This statement call the Speciation routine of EST workbook
CallSpeciation InputEquilibriumDataRange, InputStepConcRange, _
FirstOutputCell, TotalVolumeRange, LabelFlag

End Sub


REFERENCES

1) D.D. Perrin, I.G. Sayce, Talanta, 14, 833-842 (1967).
2) N. Ingri, W. Kakolowicz, L.G. Silln, B. Warnqvist, Talanta, 14, 1261-1286 (1967).
3) I. Ting-Po, G.H. Nancollas, Anal. Chem., 44, 1940-1950 (1972).
4) G. Ginzburg, Talanta, 23, 149-152 (1976).
5) P.M. May, P.W. Linder, D.R. Williams, Dalton Trans., 588-595 (1977).
6) V.S. Tripathi, Talanta, 33, 1015-1020 (1986).
7) V.W.H. Leung, B.W. Darvell, A.P.C. Chan., Talanta, 35, 713-718 (1988).
8) L. Alderighi, P. Gans, A. Ienco, D. Peters, A. Sabatini, A. Vacca, Coord. Chem. Rev., 184, 311-
318 (1999).
9) K.J. Powell, L.D. Pettit, R.M. Town, K.I. Popov, University Chemistry Education, 4, 9-13
(2000), http://www.acadsoft.co.uk/soleq/soleq.htm.
10) N. Ingri, L.G. Silln, Ark. Kemi, 23, 97 (1964).
11) R. Ekelund, L.G. Silln, Wahlberg O., Acta Chem. Scand., 24, 3073 (1970).
An Excel Tool for Equilibrium Speciation
19
12) R.J. Motekaitis, A.E. Martell, Can. J. Chem., 60, 168-173 (1982).
13) P. Gans, A. Sabatini, A. Vacca, Dalton Trans., 1195-1200 (1985).
14) J. Havel, M. Meloun, Talanta, 33, 525 (1986).
15) C. De Stefano, P. Mineo, C. Rigano, S. Sammartano, Ann. Chim. (Rome), 83, 243-277 (1993).
16) M. Kyvala, I. Lukes, Program for the Determination of Equilibrium Constants from
Potentiometric, Spectrophotometric, NMR, and Other Data, CHEMOMETRICS'95 - 4th
International Chemometrics Conference of the Czech Chemical Society, Pardubice, Czech
Republic (1995).
17) J. Barbosa, D. Barrn, J.L. Beltrn, V. Sanz-Nebot, Anal. Chim. Acta, 317, 75-81 (1995).
18) P. Gans, A. Sabatini, A. Vacca, Talanta, 43, 1739-1753 (1996),
http://www.hyperquad.co.uk/hq2000.htm.
19) N. Ingri, I. Andersson, L. Petterson, A. Yagasaki, L. Andersson, K. Holmstrm, Acta Chem.
Scand., 50, 717-734 (1996).
20) C.L. Araujo, A. Ibaez, G.N. Ledesma, G.M. Escandar, A.C. Olivieri, Comput. Chem., 22, 161-
168 (1998).
21) R. Cazallas, M.J. Citores, N. Etxebarria, L.A. Fernndez, J.M. Madariaga, Talanta, 41, 1637-
1644 (1994).
22) J.C. Westall, J.L. Zachary, F.M.N. Morel, MINEQL: A Computer Program for the Calculation
of Chemical Equilibrium Composition of Aqueous Systems. Tech. Note no. 18. EPA Grant
no. R-803738., Massachusetts Institute of Technology, Cambridge, MA, (1976).
23) J.W. Ball, D.K. Nordstrom, User's Manual for WATEQ4F, U.S.Geological Survey Report 91-
183, (1991), http://wwwbrr.cr.usgs.gov/projects/GWC_chemtherm/software.htm.
24) J.D. Allison, D.S. Brown, K.J. Novo-Gradac, MINTEQA2/PRODEFA2, a Geochemical
Assessment Model for Environmental Systems: Version 3.0 User's Manual.,
U.S.Environmental Protection Agency, Washington D.C., (1991),
http://www.lwr.kth.se/English/OurSoftware/vminteq/index.htm.
25) T.J. Wolery, EQ3/6, A Software Package For Geochemical Modelling of Aqueous Systems:
Package Overview and Installation Guide (Version 7.0), (1992).
26) D.L. Parkhurst, D.C. Thorstenson, L.N. Plummer, PHREEQE-a Computerized Program for
Geochemical Calculations, U.S. Geological Survey, (1995).
27) D.L. Parkhurst, C.A.J. Appelo, User's guide to PHREEQC (Version 2) A computer program for
speciation, batch-reaction, one-dimensional transport, and inverse geochemical
calculations, U.S.Geological Survey Water-Resources Investigations Report 99-4259,
(1999), http://wwwbrr.cr.usgs.gov/projects/GWC_coupled/phreeqc/index.html.
28) C.M. Bethke, Geochemical Reaction Modeling, Oxford University Press, New York, (1996),
http://www.rockware.com/catalog/pages/gwb.html.
29) J. van der Lee, L. De Windt, CHESS Tutorial and Cookbook. User's Guide Nr. LHM/RD/99/05,
CIG-cole des Mines de Paris, Fontainebleau, France, (2000), http://chess.ensmp.fr.
30) D.A. Kulik, Am. J. Sci., 302, 227-279 (2002), http://les.web.psi.ch/Software/GEMS-PSI/.
31) R.P.T. Janssen, W. Verweij, Water Res., 37, 1320-1350 (2003).
32) G. Eriksson, Anal. Chim. Acta, 112, 375-383 (1979),
http://www.chem.umu.se/dep/inorgchem/samarbeta/WinSGW_eng.stm.
33) W.R. Smith, R.W. Missen, Chemical Reaction Equilibrium Analysis: Theory and Algorithms,
Krieger Publishing, Malabar, Florida, (1991), http://www.mathtrek.com/.
34) B. Mller, chemEQL, version 2.0. User's Guide to Application, Swiss Federal Institute of
Environmental Sciences and Technology (EAWAG), Dbendorf, Switzerland, (1996),
http://www.eawag.ch/research/surf/forschung/chemeql.html.
35) P.M. May, Chem. Commun., 1265-1266 (2000).
36) D.N. Blauch, J. Chem. Inf. Comput. Sci., 42, 143-146 (2002).
DEL PIERO and coworkers

37) L. Liwu, Java: Data Structures and Programming, Springer, New York, (1998).
38) E.J. Billo, Excel for Chemists, A Comprehensive Guide, Wiley-VCH, New York, (1997).
39) N. Maleki, B. Haghighi , A. Safavi, Microchem. J., 62, 229-236 (1999).
40) R. De Levie, Excel in Analytical Chemistry, University Press, Cambridge, (2001).
41) O. Raguin, A. Gruaz-Gruion, J. Barbet, Anal. Biochem., 310, 1-14 (2002).
42) J. Walkenbach, Excel 2003 Bible, Wiley Publishing, Indianapolis, (2003).
43) C. Comuzzi, P. Polese, A. Melchior, R. Portanova, M. Tolazzi, Talanta, 59, 67-80 (2003),
http://www.freewebs.com/solverstat/.
44) A. Bianchi, E. Garcia-Espaa, J. Chem. Educ., 76, 1727-1732 (1999).
45) L.F. Brown, M.H. Ebinger, Comput. Chem., 22, 419-427 (1998).
46) S. Del Piero, A. Melchior, P. Polese, R. Portanova, M. Tolazzi, Dalton Trans., 1358-1365
(2004).
47) D.V. Nichita, S. Gomez, E. Luna, Comput. Chem. Eng., 26, 1703-1724 (2002).
48) P. Vonka, J. Leitner, Collect. Czech. Chem. Commun., 65, 1443-1454 (2000).
49) E.L. Cheluget, R.W. Missen, W.R. Smith, J. Phys. Chem., 91, 2428-2432 (1987).
50) D.J. Leggett, Talanta, 24, 535-542 (1977).
51) A. De Robertis, C. De Stefano, S. Sammartano, Anal. Chim. Acta, 191, 385-398 (1986).
52) R.N. Mioshi, C.L. do Lago, Anal. Chim. Acta, 334, 271-278 (1996).
53) P. Brassard, P. Bodurtha, Comput. Geosci., 26, 277-291 (2000).
54) G. Colonna, A. D'Angola, Comput. Phys. Commun., 163, 177-190 (2004).
55) E. Marengo, M.C. Gennaro, Analusis, 20, 345-350 (1992).
56) P.J. Vaughan, Soil Sci. Soc. Am. J., 66, 474-478 (2002).
57) A. Holstad, Comput. Geosci., 3, 229-257 (1999).
58) Y.K. Rao, Mar. Chem., 70, 61-78 (2000).
59) R. De Levie, J. Chem. Educ., 76, 987-991 (1999).
60) R. De Levie, J. Electroanal. Chem., 323, 347-355 (1992).
61) C.W. Davies, Ion Association, Butterworths, Washington, (1962).
62) J.M. Casas, F. Alvarez, L. Cifuentes, Chem. Eng. Sci., 55, 6223-6234 (2000).
63) D.W. King, J. Chem. Educ., 79, 1135-1140 (2002).
64) R. Tauler, E. Casassas, Anal. Chim. Acta, 206, 189-202 (1988).
65) W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery, Numerical Recipes in C: The Art
of Scientific Computing, Cambridge University Press, Cambridge, (1992).
66) W.H. Spendley, G.R. Hext, F.R. Himsworth, Technometrics, 4, 441-461 (1962).
67) J.A. Nelder, R. Mead, Comput. J., 7, 308-315 (1965).
68) R. Chelouah, P. Siarry, Eur. J. Oper. Res., 148, 335-348 (2003).
69) N. Metropolis, M.N. Rosenbluth, A.W. Rosenbluth, A.H. Teller, E. Teller, J. Chem. Phys., 21,
1087-1092 (1953).
70) S. Kirkpatrick, C.D. Gellat Jr, M.P. Vecchi, Science, 220, 671-680 (1983).
71) W.L. Goffe, Studies in Nonlinear Dynamics and Econometrics, 1, 169-176 (1996).
72) V. Stark, Thesis: Implementation Of Simulated Annealing Optimization Method for APLAC
Circuit Simulator, Helsinki University of Technology, (1996).
73) M.F. Cardoso, R.L. Salcedo, S.F. Azevedo, Comput. Chem. Eng., 20, 1065-1080 (1996).
74) A.P. Alves da Silva, Learning Nonliner Models, 1, 45-60 (2002).
75) R.W. Wehrens, L.M.C. Buydens, Trac-Trends Anal. Chem., 17, 193-203 (1998).
76) J.H. Holland, Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann
Arbor, (1975).
77) N. Turkkan, Discrete Optimization of Structures Using a Floating-Point Genetic Algorithm,
Annual Conference of the Canadian Society for Civil Engineering, Moncton, N.B., Canada,
June 4-7 (2003), http://www.umoncton.ca/turk/logic.htm.
An Excel Tool for Equilibrium Speciation
21
78) A. Brunetti, Comput. Phys. Commun., 124, 204-211 (2000).
79) P. Nikitas, A. Pappa-Louisi, A. Papageorgiou, A. Zitrou, J. Chromatogr. ,A, 942, 93-105
(2002).
80) P. Nikitas, A. Papageorgiou, Comput. Phys. Commun., 141, 225-229 (2001).
81) E. Bodon, L. Luksan, E. Spedicato, Numerical Performance of ABS Codes for Nonlinear Least
Squares, Tech. Rep. DMSIA 27/2001, Universit degli Studi di Bergamo, Bergamo, (2001).
82) B.D. McCullough, Am. Stat., 52, 358-366 (1998).
83) M. Maeder, Y.M. Neuhold, G. Puxty, Chemometrics Intell. Lab. Syst., 70, 193-203 (2004).
84) R. Arnek, Ark. Kemi, 32, 81 (1970).
85) J.K. Grime, Analytical Solution Calorimetry, John Wiley & Sons, New York, (1985).
86) S. Wash, D. Diammond, Talanta, 42, 561-572 (1995).
87) M. Meloun, J. Militky, K. Kupka, R.G. Brereton, Talanta, 27, 721-740 (2002).