Talama, Vol. 42, No. 4, pp.

561-572, 1995
Elsevier Science Ltd
Pergmon 0039-9140(95)01446-2 Printed in Great Britain
0039-9140/95 $9.50 + 0.00

NON-LINEAR CURVE FITTING USING MICROSOFT

EXCEL SOLVER

S. WALSH a n d D. DIAMOND*
School of Chemical Sciences, Dublin City University, Dublin 9, Ireland

(Received 8 September 1994; Revised 12 October 1994. Accepted 12 October 1994)

Sununary--Solver, an analysis tool incorporated into Microsoft Excel V 5.0 for Windows, has been
evaluated for solving non-linear equations. Test and experimental data sets have been processed, and the
results suggest that solver can be successfully used for modelling data obtained in many analytical
situations (e.g. chromatography and FIA peaks, fluorescence decays and ISE response characteristics).
The relatively simple user interface, and the fact that Excel is commonly bundled free with new PCs makes
it an ideal tool for those wishing to experiment with solving non-linear equations without having to
purchase and learn a completely new package. The dynamic display of the iterative search process enables
the user to monitor location of the optimum solution by the search algorithm. This, together with the
almost universal availability of Excel, makes solver an ideal vehicle for teaching the principles of iterative
non-linear curve fitting techniques. In addition, complete control of the modelling process lies with the
user, who must present the raw data and enter the equation of the model, in contrast to many commercial
packages bundled with instruments which perform these operations with a 'black-box' approach.

Recent trends in data generation, owing in part
to the computerization of instrumentation,
have led to larger and more complex data sets.
Traditional approaches to experimental data
processing are largely based on linearization
and/or graphical methods. However, this can
lead to problems where the model describing
the data is inherently non-linear, or where the
linearization process introduces data distortion
(e.g. in standard deviations of data which are
logarithmically related to the analyte concen-
tration as with ion-selective electrodes). Hence,
with the ready availability of the PCs in labora-
tories, there is increasing interest in applying
non-linear curve fitting techniques to exper-
imental data.
Computers provide an ideal tool to overcome
such problems. Although statistical programs
have been available for large main-frame com-
puters for many years, these were often cumber-
some to use and the computers were not readily
available. Many of these statistical packages
have now been converted for PCs, but their
application to solving analytical data processing
problems is still relatively rare as, although
scientifically correct and effective, these pack-
*Author to whom correspondence should be addressed.
561
ages tend to be engineer-orientated, not
especially user-friendly and expensive.
In contrast, many analysts are very familiar
with spreadsheets such as Microsoft Excel and
Lotus 1-2-3, and several books have appeared
recently describing their use as sophisticated
calculators for processing, displaying and inter-
preting scientific data, and as tools for teaching
calculations in chemistry) '2 The familiar Win-
dows user interface, ready availability (often
bundled free with new machines or on site-wide
licence) and access to desk-top publishing for
integrating text and graphics were the main
attractions. However, like Lotus 1-2-3, Excel
was primarily targeted at the business market,
and was frustratingly under-developed for scien-
tific applications, particularly in terms of dis-
play and data processing tools (e.g. non-linear
regression). The balance has been somewhat
addressed in recent years, and the appearance of
powerful analysis tools such as solver makes
Excel an important addition to the analyst's
armoury for processing experimental data.
Central to the successful application of solver
is the user's depth of understanding of the
problem, in contrast to many analytical data
processing packages, in particular those bun-
dled with instruments which often employ a
'black-box' approach. With solver, the user
562 S. WALSHand D. DIAMOND

m u s t p r e s e n t the r a w d a t a in a spreadsheet, where H = peak height above baseline,

enter the m o d e l a n d m o d e l p a r a m e t e r s cor-
rectly, a n d finally initiate the m o d e l building
process. G r a p h i c a l w i n d o w s enable the d y n a m -
ics o f the m o d e l b u i l d i n g process to be observed,
a feature which can be used to great effect in
teaching the principles o f d a t a modelling.
USING SOL VER
The following steps describe the general prin-
ciples of using solver which were used in all of
the theoretical and practical examples described
later. The data set to be modelled must first be
obtained. In the case of theoretical studies, the
data are generated by means of the equation of
interest. A suitable range of values can be
quickly obtained by means of the 'edit-fill-
down' command. Figures 1 and 2 show the
various procedures involved in finding the sol-
ution to a Gaussian peak generated using the
well known equation;
f(X) = H exp[-(~-~ "f)2 ] + B,
X = point on x-axis, ~ = distance along x-axis
to peak maximum, a = standard deviation of
the peak, B = baseline offset from zero.
Figure 1 shows the layout of a typical Excel
spreadsheet prior to using Solver using data
generated with equation 1. Column A contains
the x-axis data, and column C Gaussian data
generated via equation (1) using column A data
and known values for the parameters H, ~, a
and B, which have to be found by solver.
Column B shows data generated using equation
(1) with the initial starting values of H, 2, a and
B shown in cells G3 to G6. For convenience,
these parameters should be entered as named
variables in Excel using the 'insert-name'
command.
In order to find the solution to the best-fit
values of the Gaussian parameters, the user
must present a target cell which contains the
sum of squared residuals (SSR) between the
known data (column C) and the estimated data
(column B). The residuals are listed in column
(i) D and are obtained by subtracting equivalent
cells in columns B and C. Column E contains

Eile Edit ~iew Insert Format Iools Dsta ~ndow Help

c~t 3

time t Model Data residuals square i variables,

1 5.6172. 10000 4.6171 21.3179!:
2 " 5.:70051. . . . 110001 4.7005~ 22.0~3~ 4.00
3 ; 5.7916 1.0001i 4.7915 30.00
4 5.89O51 10001i 4.8904i 23.91641 .... $ =., 15.00
5 ...... ! 5.9974~ 1.0001~ 4.9973i 2z19728~ 5.00
6 6.1121 1.0002i 5.1120i 450.00
7 6.2346i 1.0002i 5 2344i
8 6.3644i 1.0003i 5.3642i
197 5.0000) ! qQOil ....... 3=09g01
198 50000': 1.0001i 3.9999i
199 #.ooooi 1.000oi

t =C200-D200
1. -H'EXP(-((A2OO-Z)^2/P))+B !
[
[2. J I
3. Named variables and
..............
I
their initial values
Fig. I. Typical layout of a solver problem in Excel. The data in column C are for the Gaussian peak (generated
via equation 1), and the model is entered in column B. Residual errors are in column D and their squares
are in column E. The named variables used by solver are in cells G3-G6 and the sum of squared residuals
(SSR) is in cell G9. Note that 2 is represented by z and a by S in the named variables list.
 Non-linear curve fitting
4- 100
),- 5
: <,, 200
i)- 40
Fig. 2. (a) Initial panel presented to the user when solver is activated via the 'tools' menu. (b) The 'options'
dialog box which opens when activated by the user. See text for further details.
the squared residuals, and cell G9 the sum of the
squared residuals (i.e. the sum of cells E2
to E200). This is the quantity which will be
minimized by solver.
Solver is activated via the tools menu. On
activation, the user is presented with the panel
shown in Fig. 2a, which in this case, is set up to
solve the Gaussian problem described above.
The user options available are:
Set target cell specifies the target cell which
can be maximized, minimized or set to a
certain value. In this case the target cell is G 9
and it is set to be minimised.
By changing cells specifies the cells which will
be varied by the search algorithm in order to
minimise the number in cell G9, i.e. cells G3
to G6 in this example.
Subject to the constraints: constraints can be
applied by the user to limit the search space
explored by the optimization algorithm. As is
usual in iterative search procedures based on
gradient-type algorithms, efficiency is best if
the search is initiated as near the global
563
0.01
solution of the problem, and the unknown
variables are restricted in value to realistic
ranges. This of course implies that the user
has considerable knowledge of the problem
prior to initiating the search.
Options displays the solver options dialogue
box (Fig. 2b) in which the user can vary more
features of the solution process. These are:
Max. time limits the time allowed for the
search process.
Iterations limits the number of iterations
during the search process.
Precision sets the precision of the search
process. This will determine the minimum
change in the target cell which will cause the
search process to stop. The best value for
precision varies according to how far from
the global optimum the initial search position
is, and how smooth the error surface is.
Tolerance represents a percentage of error
allowed in the optimal solution. A higher
tolerance would tend to speed up the solution
process but at the expense of accuracy in the
final solution.
564 S. WALSHand D. DIAMOND

Assume linear model will speed up the sol-
ution process but obviously should be used
only if the relationship is linear.
Estimates specifies the approach used to
obtain initial estimates of the variables.
Briefly, these are
Tangent uses linear extrapolation from a
tangent vector. That is from a tangent,
solver extrapolates in different directions
to identify which gives a minimum for the
target cell. This identifies the next direc-
tion of the search process.
Quadratic uses quadratic (i.e. non-linear)
extrapolation which can greatly improve
results in very non-linear problems at the
expense of speed at arriving at an answer.
Derivatives forward and central differen-
tiation options exist for estimates of partial
derivatives which give the gradient of the
search at that point.
Search determines which search algorithm
(Newton and Conjugate) is used at each
iteration. Both methods are dependent on the
calculation of gradient values in the error
surface at each stage in the iteration? '4
Figure 3 shows the results obtained with
solver for the Gaussian problem. Figure 3a
shows the initial position of the search, with the
estimated data being well displaced from the
required Gaussian curve. Figure 3b shows
the final positions of the known and estimated
data, illustrating the excellent fit obtained. The
values returned by solver for the model par-
ameters (H = 10, Y = 100, tr = 2 0 and B = 1)
are equal to the values used to generate the test
Gaussian curve.
Modelling two overlapping Gaussian peaks
A more difficult task is to model two over-
lapping Gaussian peaks, particularly where one
of the peaks appears only as a shoulder on a
more dominant peak, a situation which often
occurs in spectroscopy and chromatography.
Figure 4 illustrates such a situation which has
been successfully modelled using solver. In this
example, both the test data and the model data
were generated using the equation;
f(x):n I exp[---(S~ ~1)2.]
F-(x-~2)2I _
.2exp L j+.,
.

(2)
The Newton method typically requires more
memory than the conjugate search but re-
quires fewer iterations. The conjugate search
is useful if you have a complex problem 12.00•
(multiple model parameters, large data sets) 10.00 •
and memory usage is a concern. We have I 8,00.
found little difference in using either algor- g
ithm to date. 6.00.
4.00
Load/save model is used when more than one
solver model is to be used with the worksheet. '~ 2.00
Show iteration results gives the user a com- 0.00 i I ! I
plete update of all data after each complete 0 40 80 120 160 200
Data Index
iteration. If the estimated and known data
are graphed, the user can observe a dynamic
display of the search process as it proceeds 10.00 j ~ i
through each iteration. This is very useful in
observing visually the progress of the algor- I 8.00
ithm, but can obviously greatly slow down i .00
the process where a large number of iter-
4.00
ations are involved. If this option is not
exercised, the search process proceeds auto- 2.00
matically, reporting only the value of the o.00 I i I i
target cell after each iteration. 40 80 120 160 200
Use automatic scaling is useful in situations DaleI n d e x
Fig. 3. Results returned by solver for the fit to the Gaussian
where data may be distorted through gross data shown in Fig. 1. (a) The initial positions of the solver
differences in the magnitude of various par- data and the Gaussian peak. (b) The final fit to the Gaussian
ameters in the search algorithm. peak obtained using solver.
 Non-linear curve fitting
100
9O
'IgO
7O
4O
2O
10
0
20 40 60 SO 10O
Date Index
Fig. 4. Fit obtained with solver to two overlapping Gaussian
peaks where one of the peaks appears only as a shoulder on
a more dominant peak.
where H I = height of peak 1 above baseline,
/-/2=height of peak 2 above baseline,
~ = position of peak 1 maximum along x-axis,
5c2 = position of peak 2 maximum along x-axis,
tr I = standard deviation of peak 1, tr2 = stan-
dard deviation of peak 2, B = baseline offset
from zero on y-axis.
Once again, the search returns an exact sol-
ution to the problems, with all parameters being
successfully returned with the correct values
( H i = 8 0 , £1=52, tr I = 1 0 , /-/2=45, g2=72,
(r2 = 9 a n d B = 5 ) .
Examples like these are very useful for teach-
ing the principles of optimization and non-
linear least squares fitting to graduate students,
and in particular, in highlighting the need for
direction by the user of the process, as initial
conditions can easily be set which do not allow
the algorithm to locate the desired minimum.
M O D E L L I N G EXPERIMENTAL DATA
In this section, we demonstrate the appli-
cation of solver to modelling experimental data
sets which, unlike the theoretical tests discussed
previously, will not have exact solutions. How-
ever, the validity of the results can be inferred
by comparison with results obtained with other
optimization algorithms, through knowledge of
the system being studied or through examin-
ation of the error of the fit.
Example (1). Modelling chromatography peaks
Chromatography (and flow-injection) peaks
are characterized by a Gaussian-type shape
distorted by tailing which occurs on the falling
portion of the peak. Models such as the expo-
nentially modified Gaussian ~-s and the tanks-in-
series9a° have been developed in order to allow
this distortion of the standard Gaussian peak
shape to be described mathematically. This has
important applications in the area of data
565
storage (results can be described mathematically
rather than stored as relatively large ASCII
files), and in the analysis of peak purity (e.g. by
comparing the shape parameters of experimen-
tal peaks to that of a typical peak obtained with
the analyte under normal conditions).
Exponentially modified Gaussian model.~
129 The exponentially modified Gaussian function
(EMG) is the result of the convolution of a
Gaussian function and an exponential decay. In
Excel, this is achieved as follows. Y(n) rep-
resents an unconvoluted Gaussian data array of
n points calculated according to equation (1)
above. The difference AY(m) between a point
Y(m) and the previous point Y(m - 1) is easily
obtained by subtracting the appropriate cells.
The convoluted set of points, EMG(m) is de-
rived from the Y(m) array as follows:
EMG(m) = EMG(m - 1)
+ [Y(m) - EMG(m - l)], (3)
L
where
A = 1 - e x p ( - tt"2/~ W1, (4)
where • = time constant of the exponential
decay (where a time dependent function is being
modelled), I4I], W2 = weighting factors, with ~"2
normally set at unity and m is the data index in
the array.
On the spreadsheet, the Gaussian equation (1)
is first used to create a Gaussian data array over
the range of interest. The first point on this
array is then set equal to the first convoluted
point (EMG(1)). The second E M G point
(EMG(2)) can then be calculated via equations
(3) and (4) above, This procedure is repeated for
the entire Gaussian array using the 'edit-fill-
down' command in Excel.
Tanks-in-series model.9'I° In this approach,
the flow system is regarded as behaving as a
series of mixing chambers which serve to distort
the initial ideal square-wave concentration
profile of the sample plug as it travels to the
detector. The equation used in this instance is;
f(t'=H[(Ti(t/li)N-, ")
×((Nl~.)exp(--t/Ti)], (5)
where Ti = mean residence time of an element of
fluid in any one mixing tank (i). N = number
of tanks, t = time (x-axis index), H = scaling
factor.
566 S. WALSH and D. DIAMOND

Figure 5a-d illustrates the superior perform- best-fit obtained with the EMG and tanks-in-
ance of the EMG model in describing an exper- series models, respectively. Figure 5d compares
imental HPLC peak. Figure 5a shows the the residual errors obtained with each model
best-fit obtained with the Gaussian model using expressed as a percentage of the response maxi-
solver while Fig. 5b and c shows the equivalent mum, and while there is some structure in the

(a) 2500

.~ 2000
| -,, model ]
1500 0 HPLCdataJ

1000

~J 500

0
10 20 30 40
Data Index

2500
(b) ..................................................................................................................................

2000

iI=
i
1500

1000

500.

0
0 10 20 30 40

Data Index

(c) 2500

. • 2000
E
m,
!
1500 ° ata li
M
i
-~. 1000 i!
"6 !
•~P 500

0 LI~,..',..J.,,JU1;JWv, , - v~.rvc,,,uuiJUe,J,,.,e¢]
i
0 10 20 30 40
Data Index

(d) 8
6 -..O-., ~ enr E1Mv,
IG

~ 0

-4
-6
-8
5 !0 15 20 25 30 35 40
Data Index
Fig. 5. Comparison of (a) Gaussian, (b) exponentially modified Gaussian (EMG), (c) tanks-in-series
models to an experimental HPLC peak and (d) the residuals obtained for each fit (expressed as a % of
the peak maximum). The unit 'Bit-number' refers to the number returned by the I/O card on digitization
of the detector signal.
 Non-linear curve fitting
E M G residuals, the error is clearly much smaller
than that obtained with the other models. The
inability of the Gaussian model to cope with the
peak tail is demonstrated by the comparatively
large residual error for points 22-30 which is
much reduced in the E M G residuals.
Example (2). Modelling fluorescence decay
processes
The data set for the single exponential
modelled in this equation refers to fluorescent
emission decay lifetime measurements of
the compound: ruthenium-bis(2,2'-bipyridyl) (5-
isothiocyanate- 1,10-phenanthroline) [Ru(bpy)2
(NCSphen)] z+ using a Q-switched ND-YAG
laser. The samples were aerated and measure-
ments were carried out in 0.1M carbonate buffer
(pH 9.6). The compound absorbs at 455 nm and
is characterized by a fluorescence decay with a
single time constant.
The double exponential model has been
applied to fluorescent emission decay lifetime
measurements of the same compound after
attachment of bovine serum albumin via a
thio-urea linkage. 1~Once again the samples were
aerated and measurements carried out in 0.1M
carbonate buffer (pH 9.6). Binding of the
protein has little effect on the excitation and
absorbance wavelengths but lifetimes of fluor-
escent centres near the binding sites are affected,
leading to differing sets of fluorescent emissions
emanating from species bound in different
environments. The models used in this investi-
gation were general single and double exponen-
tial equations of the form:
f(t) = [A(1 - e x p ( - k t ) ] + B,
where A = pre-exponential factor, k --- rate
constant, (1/k = decay lifetime), t = time, B =
baseline offset and for the double exponential
f(t) = [A, (1 - e x p ( - k , t)]
+ [A2(1 - exp(-k2t)] + B,
where A,, A2 = pre-exponential factors, kl, k2 =
rate constants, t = time, B = baseline offset.
Data obtained from the instrumentation are
transformed to suit the above exponential
models via in-house software prior to
modelling.
Figure 6a and b illustrates the fit obtained
using the single exponential model (equation 6)
with the NCS compound, and the residuals of
the fit, respectively. Clearly the model par-
ameters returned by solver fit the data well, with
567
the error (expressed as a percentage of the
maximum response) never being greater than
+ 2%, which is quite acceptable given the noise
on the original signal, and the time base of the
experiment. The time constant obtained from
the fit (T = 450 nsec) is typical of this material.
Figure 6c and c shows fits obtained with the
protein bound ruthenium compound using the
double exponential model (equation 7). Again,
solver returns a good fit to the data, (z~ = 286
nsec, z2 = 833 nsec) with the error being within
about + 3 % outside the initial 200 ns of the
curve. While, fitting two exponentials to data
obtained in these experiments is almost certainly
an approximation such an approach can yield
important information on the environment of
the fluorescent centre through examination
of the time constants obtained, while the pre-
exponential factors may yield information on
the relative populations of centres in the differ-
ent environments." However, this is a matter of
considerable debate amongst specialists in the
area. For example, in these experiments it is
likely that the fluorescent centres exist in a
number of different environments which modify
the emission characteristics (e.g. the time con-
stants). The fit with the double exponential
model is reasonably good, and the time con-
stants obtained are in the range expected. How-
ever, the curvature in the residuals suggests that
there is some additional structure in the data.
This in turn indicates that the fluorescent
centres exist in two main environments but
this is probably not the complete picture. In
addition, careful examination of the residuals
suggests an oscillating structure above 1500
(6) nsec, which can also be seen in the original data.
It is likely that this is instrumental rather than
chemical in origin (e.g. environmental noise).
Example 3. Modelling ion-selective electrode
characteristics
ISEs can be characterized in terms of the well
(7)
known Nikolskii-Eisenman equation (8) below
in terms of the cell constant (E°), the slope (S),
and selectivity coefficients (g~ t) which describe
the effect of various interfering ions ( j ) on the
response of the electrode to the primary ion (i):
E=E °+slog a~+ K 0 aj , (8)
)
where a~ and aj are the activity of the primary
and interfering ions, respectively, and z~ and zj
are their charges. We have used a simplified
version of the above equation (neglecting the
 568 S. WALSH and D. DIAMOND
(a) 30000

20000
15000
~ 1 0000
5000 el
0 I I I
O.OOE+O0 5.00E+02 1.00E+03 1.50E+03 2.00E+03
Time /ne

(b) 2.52 l

1.5
1
0.5
0
-0.5
-1
-1.5
-2
0 500 1000 1500 2000
Time /hi

(c) 30o00

25000
i 2000O

15000

i 10O00

m 5000

0 I I I I I
500 1000 1500 2000 2500 3000
Tim • Ins

(d) 4 .]-.,..........o ..................................................................................................................................................

3
2 900 oo 00~ " ~ 0 0

-2 ~ '+~P . o ¢
-3 < ~ 0 0

-5 ~
0
-6 o i i J ,
500 1000 Tim1eSOJ)n, 2000 2500 30(:30
Fig. 6. (a) Fit obtained using a single exponential model for fluorescence decay obtained with ruthenium
bipyridyl isothiocyanate phenanthroline using a Q-switched N D - Y A G laser, U (b) residual errors obtained
with the single exponential model. (c) Fit obtained using the double exponential model for the protein
bound ruthenium bipyridyl isothiocyanate phenanthroline u and (d) the residual errors obtained with the
double exponential model.

charge factors on aj) extensively to enable errors the electrode characteristics (selectivity co-
on analytical signals arising from interferents to etficients, slopes, cell constants), which involved
be decoupled and corrected.': Approaches such extensive programme development. The meth-
as simplex optimization and genetic algor- odology basically involves generating a matrix
ithms 12J3 have been investigated for estimating of electrode responses to a series of calibration
 Non-linear curve fitting 569

solutions containing the ions of interest at simplex to fine-tune the position of the mini-
activity levels set according to a fractional mum), simplex (uses the simplex only) and
factorial calibration design (see Tables 1 and 2). solver, results obtained with solver.
Thus for a four electrode array, 32 solutions are The results show good agreement on the
required, eight of which are targeted at each values of the cell potential and slope for both
electrode (e.g. solutions 1-8, Table 1 are for the electrodes with all three approaches. Further-
ammonium electrode, solutions 9-16 for the more, the values for selectivity coefficient for the
sodium electrode, solutions 17-24 for the pot- main interferent are very close for both elec-
assium electrode and solutions 25-32 for the trodes (i.e. around 0.18 for potassium interfer-
calcium electrode). We typically model the ence on the ammonium electrode and 0.14 for
Nikolskii-Eisenman parameters for a particular the value of ammonium interference on the
electrode using the matrix of potentials obtained potassium electrode). Even the low-valued selec-
from the electrode array with the eight cali- tivity coefficients are modelled to around the
bration solutions for that electrode, and then same value in each case, which is surprising
estimate the validity of the model by examining given their relatively minor effect on the fit. The
the residuals between the predicted array poten- worst case is for the sodium interference on
tials (calculated via the model parameters and the potassium electrode (K~°+~Na÷), for which a
the known activities in Table 1) and the ob- value could not be obtained with solver. How-
served array potentials (Table 2). ever, examination of the values obtained with
Table 3 show a comparison of fits obtained the other methods show this to be the smallest
for the potassium and sodium ISEs by three coefficient with least effect on the fit. These
different approaches, genetic algorithm with sim- results show that solver can obtain good esti-
plex optimization (uses a genetic algorithm to mates of electrode parameters using the array-
locate the approximate position of the minimum FIA calibration design. However, the genetic
of the error surface, and then switches to a algorithm performs better and has the added
advantage of being an unsupervised search tech-
nique (i.e. requires no prior knowledge of the
Table 1. Matrix of ion-activities in calibration solutionst2J3 problem). Nevertheless, considering the ease
Activities (molar) with which solver can be used, we feel it per-
Solution NH,+ Na + K+ Ca 2+ formed creditably well in this particular task.
1 7.89E-03 7.10E-03 2.37E-03 3.18E-03
2 8.46E-03 2.54E-03 4.23E-04 4.68E-04
Example (4). Modelling of ion-selective electrode
3 8.42E-03 8.42E-04 4.21E-03 3.56E-04 (ISE) dynamic response inflow-injection analysis
4 8.37E-03 4.18E-04 5.02E-04 1.50E-03
5 8.57E-05 6.00E-03 3.43E-03 2.18E-04
We have also used solver to look at the
6 8.58E-05 4.29E-03 6.86E-04 1.10E-03 dynamic responses obtained with PVC mem-
7 8.34E-05 1.67E-04 2.50E-03 2.95E-03 brane electrodes in a flow-injection analysis
8 8.96E-05 2.69E-04 1.79E-04 1.30E-04
9 7.06E-03 7.84E-03 7.06E-03 2.73E-03
set-up similar to that used in the calibration
10 7.56E-04 8.40E-03 5.04E-03 3.03E-04 experiments. ~2One approach is to use a logistic/
11 3.32E-04 8.30E-03 5.81E-04 1.93E-03 sigmoid model to characterize the rising portion
12 2.54E-03 8.46E-03 7.61E-04 4.16E-04
13 4.21E-04 8.43E-05 1.69E-03 2.56E-03
of the ISE response peak. The equation used is:
14 5.08E-03 8.47E-05 5.93E-03 5.21E-04
15 5.77E-03 8.25E-05 3.30E-04 2.83E-03
16 1.78E-04 8.92E-05 2.68E-04 3.19E-04 E(t) = (1 + exp[b(t - c)]) ~ + d, (9)
[ a 1
17 6.26E-03 6.26E-03 7.83E-03 3.09E-03
18 7.41E-04 4.94E-04 8.23E-03 2.34E-03 where a = peak height (mY), b = slope co-
19 3.38E-03 6.76E-04 8.54E-03 3.61E-04 efficient (mV per decade change in [K+],
20 6.80E-04 2.55E-03 8.50E-03 2.11E-04
21 1.63E-03 7.34E-04 8.16E-05 4.07E-03
c = time from beginning of the peak to the
22 2.56E-04 3.42E-03 8.54E-05 1.61E-03 inflexion on the rise (sec), d = baseline offset
23 8.68E-04 4.34E-03 8.68E-05 5.15E-04 (mV), e = symmetry parameter for the sigmoid,
24 1.79E-04 3.58E-04 3.95E-05 1.93E-04
25 6.31E-03 5.52E-03 6.31E-04 3.98E-03
E(t)=electrode response at time t (mV),
26 3.96E-03 5.55E-04 6.34E-03 4.05E-03 t = time (sec).
27 2.44E-04 8.13E-04 3.25E-04 4.46E-03 This model can give some indication of the
28 5.69E-04 8.12E-05 8.12E-04 4.45E-03
29 5.13E-03 5.13E-03 5.98E-04 5.41E-05
rate of ion-uptake at the membrane surface as
30 2.62E-03 6.99E-04 2.62E-03 5.87E-05 the sample plug passes and enable comparisons
31 7.04E-04 1.76E-03 1.76E-03 6.05E-05 to be made for different experimental situations
32 5.37E-04 4.48E-04 8.95E-05 6.46E-05
(varying concentration of the primary ion, effect
570 S. WALS8 and D. DIAMOND

Table 2. Matrix o f electrodes responses to calibration solutions ~2't3

Measured potentials (m V)
Solution NH2- Na + K÷ Ca 2+
! 1.30E + 02 1.41E + 02 1.43E + 02 6.67E +01
2 1.32E + 0 2 1.18E+02 1.17E + 0 2 4.55E +01
3 1.32E + 02 1.08E + 02 1.54E + 02 4.01E +01
4 1.34E + 02 7.93E + 01 1.22E + 02 5.70E +01
5 8.32E + 01 1.49E + 02 1.50E + 02 3.37E +01
6 5.99E + 01 1.36E + 0 2 1.13E + 0 2 4.92E +Ol
7 7.82E + 01 7.86E + 01 1.41E + 02 5.86E +01
8 4.82E + 01 6.89E + 01 8.15E+01 2.31E +01
9 1.30E + 02 1.44E + 02 1.58E + 02 5.92E +01
10 9.89E + 0 1 1.43E + 02 1.54E + 0 2 4.16E +01
11 7.25E + 01 1.44E + 02 1.07E + 02 5.90E +01
12 1.12E + 02 1.46E + 02 1.16E + 02 3.49E +01
13 8.57E + 0 1 6.49E + 01 1.32E + 02 5.14E +01
14 1.31E + 0 2 8.49E + 01 1.57E + 02 4.00E +01
15 1.29E + 02 4.99E + 01 1.13E + 02 5.93E +O1
16 6.36E + 01 4.28E + 01 9.74E + 01 3.56E +01
17 1.31E + 02 1.42E + 02 1.65E + 02 6.95E +01
18 1.06E + 02 1.02E + 02 1.67E + 02 6.28E +O1
19 1.26E + 02 9.97E + 01 1.61E + 02 3.87E +O1
20 1.06E + 02 1.25E + 02 1.64E + 02 3.38E +01
21 1.02E + 02 8.55E + 0 1 9.27E + 01 6.15E +Ol
22 6.78E + 01 1.20E + 02 7.00E + 01 5.18E +Ol
23 9.21E + 01 1.26E + 02 8.07E + 01 4.30E +Ol
24 5.85E + 01 5.85E + 0 1 6.85E + 0 1 2.81E +01
25 1.31E + 02 1.29E + 02 1.22E + 02 6.38E +01
26 1.26E + 02 1.00E + 02 1.61E + 02 6.34E +01
27 6.57E + 01 8.83E + 01 9.36E + 01 6.36E +O1
28 8.11E+01 5.86E + 01 1.18E + 0 2 6.44E +01
29 1.24E + 02 1.48E + 02 1.19E + 02 2.06E +01
30 1.14E + 02 9.85E + 0 1 1.45E + 02 2.11E +Ol
31 9.07E + 01 1.19E + 0 2 1.34E + 0 2 2.20E +Ol
32 7.88E + 01 8.56E + 0 1 7.64E + 01 2.15E +Ol

of interferents, injection volume, flow rate, etc.). combination of a, b and c in terms of flow rate
The model parameters in turn can be used as and injection volume). Figure 7 shows fits ob-
inputs in a further optimization of the instru- tained to two valinomycin electrode responses
mental operating conditions (e.g. optimize a to potassium injections at differing flow rates

Table 3. Comparison o f electrode characteristics for valinomycin (K ÷) and

tetramethylacetate p-t-butylcalix[4]arene (Na ÷ ) based ion-selective electrodes
obtained by a (i) genetic algorithm followed by simplex (GA-Simplex), (ii) simplex
only, and (iii) solver, using the data in Tables I and 2
Electrode
parameter GA-simplex~3 Simplex 13 Solver
A m m o n i u m ISE
E ° (mV) 225.8 225.8 225.1
S (mV/decade) 45.4 45.5 45.2
K~,K, 0.185 0.185 0.182
K~°~,Na+ 9.36 X 10 -4 2.09 X 10 -3 5.49 X I0 4
K~°~,ca2+ 6.46 x 10 -3 4.19 x 10 -3 6.04 x 10 -3
SSR (8) 6.313 6.507 6.405
SSR (32) 270.2 228.29 223.3
Potassium ISE
E ° (mV) 270.2 270.5 270.3
S (mV/decade) 51.4 51.5 51.5
K~+tr~a,. 0.143 0.143 0.140
K~°+tNa+ < 1.0 × 10 -6 4.53 × 10 4 0.00
K~tca2+ 5.36 x 10 -3 5.32 x 10 -3 7.33 x 10 -3
SSR (8) 29.17 30.04 28.97
SSR (32) 497.7 496.3 481.8
The sum of the squared residuals (SSR) for each method is also shown. SSR (8)
refers to the squared residuals for the eight calibration solutions and SSR (32)
refers to the squared residuals over the whole situation set.
 Non-linear curve fitting 571

6O

4O

:~ 30

2O i J o model/0.5mL/min
lO ~ j J dataJl.OmL/min
0 model/O.5mLJmin
0 .. | I ...I !
0 5 10 Time/s 15 20 25
Fig. 7. Fit obtained with sigmoid model for the rising portion of FIA peaks obtained with a potassium
PVC membrane electrode at 1.0 and 0.5 mL/min flow rates (see Table 4); Carrier composition: 5 x 10-2M
MgC12, 10-6M KCI. Sample composition: 5 x 10 2M MgCI 2, 10-4M KCI; injection volume: 150/~1.

(1.0 and 0.5 ml/min). The model parameters
returned by solver for each flow rate are com-
pared in Table 4.
From these results we can deduce that in-
creasing the flow rate causes a slight reduction
in the peak height (a), an increase in the slope
of the rise (given by the increased magnitude of
b), a reduced time to the rise inflexion (c), and
a less symmetrical rise in terms of the sigmoid
model (larger value for e). Characterizing peaks
in this manner can be very useful for instrumen-
tal optimization purposes as mentioned above,
and for describing peak shapes in terms of a few
simple parameters. This characterization can be
useful for processing large numbers of peaks
and for identifying the possible presence of
impurities through the definition of a 'typical'
analyte peak as possessing these parameters
within certain limits.
CONCLUSIONS
Using examples from chromatography, spec-
troscopy and electrochemistry, we have demon-
strated the usefulness of solver as a tool for
mathematical modelling of experimental data.
Table 4. Model parameters for sigmoid model of FIA peak
rise obtained with valinomycin ISE in an FIA system at two
different flow rates (see Fig. 7)
Model parameters 1.0 ml/min
a (mV) 5.70E + 01
b (mV per decade -1.78E-01
change in [K+])
c (see) 4.84E + 00
d (mV) - 9 . 6 2 E + 01
e 7.49E + 00
Carrier composition: 5 × 10-2M MgC12, 10-6M KCI.
Sample composition: 5 × 10-2M MgCI 2, 10-4M KCI.
Injection volume: 150 #1.
While it is not on a par with custom-written
software packages for this purpose such as
Mathlab or Mathcad, or with more complex
search methods such as genetic algorithms, it is
nevertheless, quite a useful aid to data interpret-
ation and an excellent means for introducing
non-linear curve fitting into research and teach-
ing. Given the fact that Excel is readily avail-
able, and familiar to most research and
undergraduate students, it can be expected to be
quickly adopted by many workers for such
applications. However, the existence of solver
appears to be not widely known, a situation
which we hope this publication will redress in
analytical circles. The fact that the user must
import the experimental data, and arrange the
spreadsheet in a suitable manner for each prob-
lem implies a complete understanding of the
mathematics of the model. The ease with which
named variables may be varied in Excel enables
the user to experiment with equations and be-
come familiarized with their effect on the curve
obtained.
In conclusion, we hope the few examples
presented in this paper will stimulate other
workers unfamiliar with non-linear optimiz-
ation to experiment with this very useful Excel
add-on.
Acknowledgements--We wish to thank Maurice Burke
0.5 ml/min (DCU) for providing the chromatography data, Johannes
Vos (DCU) for the fluorescence data and useful discussions,
5.93E + 01 and F. J. S~.ez de Viteri and Margaret Hartnett (DCU) for
-I.03E-01 the ISE data and useful discussions.
9.59E + 00
- 1.14E + 00
4.16E + 00
REFERENCES
1. O. J. Parker and G. L. Brennan, Spreadsheet Chemistry.
Prentice Hall, Englewood Cliffs, N.J., 1991.
572 S. WALSHand D. DIAMOND
2. H. Freiser, Concepts and Calculations in Analytical
Chemistry---a Spreadsheet Approach. CRC Press, Boca
Raton, 1992.
3. W. H. Press, B. D. Flannery, S. A. Teukolsky and W. T.
Vetteding, Numerical Recipes. Cambridge University
Press, New York, 1990.
4. G. H. Hostetter, M. S. Santina and P. D'Carpio-
Montalvo, Analytical Numerical and Computational
Methods for Science and Engineering. Prentice Hall,
Englewood Cliffs, N.J., 1991.
5. E. Grushka, Anal. Chem., 1972, 44, 1733.
6. J. P. Foley and J. G. Dorsey, Anal. Chem., 1983, 55, 73.
7. R. Ddley, Anal. Chem., 1985, 57, 388.
8. A. Berthod, AnaL Chem., 1991, 63, 1879.
9. J. Ruzicka and E. H. Hansen, Flow Injection Analysis.
Wiley, New York, 1988.
10. O. Lee, G. A. Dumont, P. Tournier and A. P. Wade,
Anal. Chem., 1994, 66, 971.
11. E. Ryan, R. O'Kennedy, M. Feeny, J. Kelly and J. G.
Vos, Bioconjugate Chem., 1992, 3, 285.
12. F.J. Sfiez de Viteri and D. Diamond, Analyst, 1994, 119,
749.
13. M. Hartnett, Ph.D. Thesis, Dublin City University,
1994.