Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

ELSEVIER Applied Surface Science 93 (1996) 273-280

Savitzky and Golay differentiation in AES


I.S. Gilmore, M.P. Seah *
Division of Materials Metrology, National Physical Laboratory. Teddington, Middlesex, TWI I OLW, UK

Received 4 September 1995; accepted 5 October 1995

Abstract

For the analysis of Auger electron spectra it is popular and very useful to differentiate the acquired direct spectra
numerically by Savitzky and Golay’s method to provide peaks whose peak-to-peak height may be measured. To compare
data recorded at different channel separations, E (number of eV per channel), it is important to be- able to assign an energy
equivalent, E,, to the differentiating function which is independent of E. It is shown that the adoption of 2m& for the
differential width, where 2m + 1 is the number of points in the Savitzky and Golay function, leads to errors and
inconsistencies. It is recommended that (2m + 1)~ be identified with E,. To obtain E, values intermediate between those
possible with Savitzky and Golay’s method, a modified function is developed with added + (m + 1)tb terms. The strength of
these terms is between 0 and k(rn + 1)/N, where N is the normalisation constant, and the exact value is roughly in
proportion to the fractional excess of E, in the interval between the values permitted in Savitzky and Golay’s original
approach. An accurate assessment is made for the value to be used for E, values mid-way between those usually obtained
by Savitzky and Golay’s method.

1. Introduction are interesting features worth analysing in a spec-


trum but, for intensity and chemical state analysis,
In 1964 Savitzky and Golay [ 1I developed simple smoothing should be avoided before using a peak
convolutional smoothing and differentiation routines synthesis program. If one needs to define a peak area
which have found major use in many areas and, in and its position, any prior smoothing will not im-
particular, the electron spectroscopies used for sur- prove the quality of the fit and may well detract from
face chemical analysis. Many other smoothing rou- it [12].
tines exist, some as convolutional methods 12-51 Savitzky and Golay, in addition to smoothing also
others operating in Fourier space [3-101. For particu- considered spectral differentiation. Their differential
lar attributes some of these other methods are better, convolution function was simply the differential of
but the Savitzky and Golay method still remains a the smoothing function and the particular function
popular, simple, robust, easily used method of we are interested in here is their most basic first
smoothing [ 111. Smoothing is useful to see if there derivative quadratic function shown in Fig. 1.
For Auger electron spectroscopy it has been tradi-
tional to differentiate spectra, originally with analog
* Corresponding author. Tel.: + 44 I8 1 943 6634; Fax: + 44 sinusoidal modulation functions but, in the last
181 943 6453. decade, mainly by numerical differentiation of direct

0169-4332/96/$1_5.00 8 1996 Elsevier Science B.V. A11rights reserved


SSDI 0169-4332(95)00345-2
214 IS. Gilmore, M.P. Seah/Applied Surfuce
Science 93 (1996) 273-280

spectra. The numerical differentiation routine used in


most commercial software systems is the above Sav-
itzky and Golay first derivative quadratic function. It
is not the place here to argue the merits or otherwise
of differentiation but only to note that it is found
very useful and hence should be done consistently.
For the comparison of Auger electron spectral
intensities it is important that the differential func-
tion is used consistently from laboratory to labora-
tory. The issues here do not concern the optimal (a>
signal-to-noise ratio for minimal distortion since
spectral distortion in the differentiating process al-
ways occurs and can be beneficial in averaging out
small chemical state changes. As can be seen from
Fig. 1 the convolutional coefficients are always sym-
metrical about zero so that the number may be
characterised by 2m + 1 coefficients. This is clearly
repeatable without any problem if all the spectra
have the same energy increment, E, between the
channels. However, if they do not there are no clear
rules as to what to do to obtain equivalent results.
Problems occur particularly if the width of the differ-
ential function is comparable to the width of the (b)
peaks in the spectrum.
Fig. 2. A direct difference differential function without smoothing
In the present paper we shall address this problem (a) equal to one channel separation and (b) equal to the same
and the definition of the energy width, E,, of the energy separation as in (a) where the channel separation is one
differentiating function. tenth of that in (a).

2. The energy width of the differentiating function


number of channels times the channel separation.
The energy width of the differentiating function, Thus, the width of the differential function shown in
E,, is clearly the width of the function in terms of Fig. 2a, which gives the gradient between adjacent
points in the spectrum, must be one times the chan-
nel separation in eV. Note that Fig. 2a looks a bit
like a Savitzky and Golay function but is not part of
the true set. If, as in Fig. 2b we have data points at
ten times closer intervals and use a differential func-
tion with ten intervals of separation we would de-
scribe their energy widths as identical in these two
cases and indeed the convolution would be identical
I numerically.
For the true Savitzky and Golay convolutions,
however things are a little more complex. Fig. 3a
and Fig. 3b are the counter parts of Fig. 2a and Fig.
2b but for a 5 point and a 41 point Savitzky and
Fig. 1. Graphical presentation of the Savitzky and Golay first Golay differentiation, respectively. These look as
differential quadratic function. though they should be equivalent and, if the channel
IS. Gilmore, M.P. Seah /Applied Surface Science 93 (1996) 273-280 275

The effect on peak shapes, positions of minima


and distortion are detailed elsewhere [13]. In that
work it is also shown the dependence of the peak-
to-peak height h in the form of a plot of hW versus
ED/W. This plot was made for an arbitrarily small
channel separation, E, where E, = 2mc. The curve
on log axes rises linearly, reaches a peak at ED/W
= 2 and then falls again linearly.
In the work that follows we convolve the above
Gaussian with the Savitzky and Golay differential
function with normalisation coefficients adjusted so
that the convolution with a line of unity gradient
gives a constant differential of unit intensity. This
differs from Savitzky and Golay’s normalisation
which does not incorporate the effect of the channel
separation E. If we denote the intensity of the 2m + 1
coefficients as - m, - m + 1, - m + 2 , e-0, l,O, 1,
. . . ) m - 2, m - 1, m, all divided by a normalisation

Fig. 3. Savitzky and Golay differential functions of 5 and 41 point


with channels at 1 and 0.1 eV, respectively.

separations are 1 and 0.1 eV one may think to


describe their widths as 4 and 4.0 eV, respectively.
To study this equivalence we consider first the
differentiation of Gaussian peaks of unit area de- (4)
scribed by the function
At low values of W and large E, we are
1 -E2 convoluting a Gaussian with the
- -
‘= \/21;(+exp zr2 ’ (1) edge at In the this gives the
( 1 of the coefficient times the area under
where E is the spectral energy scale and the full the (unity> times E- ’ to allow for the
width at half maximum of the peak is given by W channel width.
where
6
w=2Ciz(+. h = 2m/Nc =
(2) (m + 1)(2m + 1)~~ *
W numerically ranges from 1.25 eV (intrinsic peak
If we work with large values of m we have
widths) to 26 eV (a rather broad peak, which may be 2mc= E, so that
at 2600 eV using an analyzer with a resolution of
1%). Additionally, in the measurement of intensity h = 12/E;. (5)
we are simply concerned with the peak-to-peak or To provide Eqs. (4) and (5) on a single plot we
peak-to-background signal and not with the detailed consider the function hWE,. Eqs. (4) and (5), re-
peak shape. Since we shall deal with Gaussian peaks spectively, become at high W/E,
the peak-to-background value is simply half the
peak-to-peak values that we shall calculate. hWE, = 2.6835( W/E,) -’ (6)
276 IS. Gilmore. M.P. Seah/Applied Surface Science 93 (1996) 273-280

and at low W/E, 10' I ’ .““‘1’, 1 1 ~~~-~~_


.. ,’

- a .. ;
hWE, = 12( W/E,,). (7) ‘,.:
,’ .
,’ .I
In Fig. 4a we show a series of plots with these axes ,’ -.
,’ ..
for E = 0.1 eV, for E, values in the range 4 to 40 ,’ ..
<’ ‘.
eV and for W values from 1.25 to 26 eV. Fig. 4b is .’ .,
,’ ..
the similar plot for E = 1 eV. At both high and low ,’ ‘.
2~0 fi ,’ 1
W/E, values the asymptotic behaviour is correct ,’
,’
but as E, reduces the curves are progressively shifted ,’
,’ ..
down and to the right at 45”. In these plots we have ,’ ..
,’ .._
assumed that E, is defined, as in our simple case ,’
,’
illustrated in Fig. 2, by the interval between the outer ,’ 4
,’
,’
I , , , , , , , I , , , , , , ,
10-l
lo-* 10" 10" 10'
WIED

Fig. 5. As for Fig. 4 but using E, = (2m + 1)~. The straight lines
are for Eqs. (6) and (7). respectively.

t 1

II_p‘-:
convolutional elements. Thus, in Fig. 4 for a 2m + 1
Savitzky and Golay function
E, = 2ms. (8)
Clearly, to shift the curves to the left and upwards at
45” in Fig. 4, the low value E, curves need a higher
effective value of E,. We therefore also consider
E,=(2m+l)~ (9)
and replot Fig. 4a and Fig. 4b in Fig. 5a and Fig. 5b.
Fig. 4. Plots of hWE, versus W/E, on log axes for calculations We now see that the lines all fall on a universal
at (a) E = 0.1 eV and (b) E = 1 eV. In these plots E, = 2m~. curve which is the same for Fig. 5a and Fig. 5b.
IS. Gilmore, M.P. Seah/Applied Surface Science 93 (1996) 273-280 277

Thus, if we use Eq. (9) to define E, the result will


be consistent, irrespective of the channel separation
E.
It should be noted that, in Fig. 4b and Fig. 5b, the
smaller W peaks are comparable with the channel
separation and so the results become significantly
affected by the precise peak position. Here the curves
are averages for peaks centred at 0, 0.25, OS, and
0.75 eV to reduce this effect. Additionally, in Fig. 5a
the two straight lines are the asymptotic approxima-
tions of Eqs. (6) and (7). It is clear that the fit is
excellent.
A further way of showing the importance of Eq.
(9) is to consider the ratio of h deduced from plots at 0 5 10 20 25 30

1 eV intervals to that for 0.1 eV intervals using Eqs.


(8) and (9) as a function of W. This is shown in Fig. 1.025 I I , , 1

6a and Fig. 6b for E, = 2ms and E, = (2m + l)~, (b)


respectively, for two values of E,. In Fig. 6a we
ratio a 5 point to a 41 point and a 7 point to a 61
point differential and label them 4 and 6 eV, respec-
tively, whereas in Fig. 6b we ratio a 5 point to a
function between 49 and 51 point and a 7 point to
one between 69 and 71 point differential functions,
and label them 5 and 7 eV, respectively. The precise
interpolation is given later in Eq. (18). It is clear that
the use of Eq. (9) leads to far greater consistency
than Eq. (8).
The universal plot of Fig. 5a is easy to generate
numerically but, for many purposes it is useful to 0 5 10 20 25 30
have an analytical form for this curve. An excellent tt

fit is given by Fig. 6. The ratio of peak-to-peak intensities at equivalent energy


differentiations at I eV and 0.1 eV energy intervals for (a) JZq. (8)
W and (b) Es. (9).
hWE, =
2.6835 E,
In the above, to obtain a 50 point function we
have averaged the 49 and 51 point functions. At such
a large number of points any such simple interpola-
tion is approximately valid but what should one do at
low numbers of points to find the equivalent energy
(10) differentiation? We discuss this below. There ap-
pears to be no analytical solution so we search for an
which is accurate to better than 0.7%. Using this adequate approximation.
relation it is easy to show that, for comparing peaks
at say 300 and 1000 eV using an analyzer of 0.5%
resolution, an error of over 20% occurs in the rela- 3. Intermediate energy widths
tive intensities if 4 or 6 eV are used instead of 5 eV.
For peaks at 300 and 2000 eV this error increases to At 1 eV energy intervals we may only obtain 5,7,
over 30%. 9 , . . . eV differentiations using Savitzky and Golay’s
278 IS. Gilmore. M.P. &ah/Applied Surface Science 93 11996) 273-280

coefficients 3 eV a naive to their 3.2. Normalisation


How do achieve intermediate There
are number of solutions although A second method is to simply consider that the
will be and one gives the h values normalisation factor N may progress smoothly as E,
not also the correct shapes. Here increases. From Eq. (3)
consider only
N(2m+ l)=m(m+l)(2m+l)~/3 (‘2)
3.1. Linear interpolation and hence
N(2m + 2Y + 1)
As a first step we consider interpolation, to 0.5
eV steps, between points for creating 4, 6, 8, . . . eV =(m+Y)(m+Y+l)(2m+2Y+l)&/3.
differentiations. If we now increase the Savitzky and
(13)
Golay width to include an extra interpolation point at
each end this will change a (2m + 1)~ differential For Y = f we find that
width into a (4m + 3)&/2 differential width, i.e. an
N(2m + 1;) - N(2m + 1)
increase by &/2 from, say, 5 to 5.5 eV or 7 to 7.5
eV. If we interpolate the spectral intensities simply = ( 16m2 + 20m + 5) e/32. (14)
and put intermediate point intensities at the arith-
We would expect this last term to be $ X 2(m +
metical average of adjacent points we may convolute
1)2&X,.
the new set of points at c/2 intervals by a regular
Thus
Savitzky and Golay (4m + 3) point differential. Al-
gebraically this is the same as convoluting our origi- X,($,m)=(2m+$+ [8(m+ I)]-‘)/(2m+2).
nal spectrum by a Savitzky and Golay-like function
at E intervals with coefficients (excluding the nor- (15)
malisation factor) of Thus, the factor given in Eq. (11) is very close to
-(2m+1)/8, -m, -m+l, . . . . l,O,l, .... that for Eq. (15) at the Y = a point and shows that
interpolation gives almost the same result as requir-
m - 1, m, (2m + 1)/8
ing the normalisation term to increase smoothly ac-
i.e. to go from 5 to 5.5 eV where m starts at 2 we cording to Eq. (31, even for non-integer values of m.
would go from a 5 point function to a 7 point For a general value of Y the smooth normalisa-
function where the outermost coefficient of 3 is tion would require
replaced by 5/8. Thus, to go 25% of the way from 5
to 7 eV differentiation the last coefficient will in- 2Y(m + l)*cXb(Y)
crease by 21% of its full value. The full value of this
=(m+Y)(m+Y+l)(2m+2Y+1)&/3
final coefficient is m + 1 so that, as m gets very
large, the outermost coefficient of (2m + 1)/8 ap- -m(m+ 1)(2m+ 1)&/3. (16)
proaches 25% of the full value. Hence
Linear interpolation thus gives the extra coeffi-
cient which would normally be m + 1 scaled by the (Y- 1)(6m+2Y+5)
X,( Y,m) = 1+ (17)
factor YX,(Y,m) for a fraction Y of the two channel 6(m+l)* ’
separation increase in differentiation where (for Y
= a> At large values of m, X, approaches unity for all Y.
This simply means that excluding the normalisation
XJ$,m) = (2m + 1)/(2m + 2). (11) term, at large m values, to achieve a differential that
Clearly X, approaches unity as m increases. We is equivalent to a 49 + 2Y point of differentiation
express the extra coefficient as (m + l)YX,(Y,m) one would use - 24 to +24 for the usual central
since, intuitively, one may expect it to be of the coefficients and the outermost coefficients would
order of (m + 1)Y. simply be +25Y.
I.S. Gilmore, M.P. Seah/Applied Surface Science 93 (1996) 273-280 219

3.3. Numerical tests

In order to verify the factor to be used for the 0.02 -


extra coefficient to the modified Savitzky and Golay
function, plots are made for even numbers of eV
differentiation with X,(f ,m) carefully adjusted. This $
al
o-

adjustment ensures that the data for 1 eV channel 3


E
separations fit the data in the universal plot of Fig. ‘Z
P -0.02 -
4a at 0.1 eV separations deduced for the unmodified I;.

Savitzky and Golay algorithm. This makes the modi-


fied Savitzky and Golay function transportable to -0.04 -

other channel separations. Fig. 7 shows the results


for these plots and Fig. 8 the resulting fractional
error. For this we have used a simple relation for
X,(f,m) as a function of the value of En or of the m
value for the basic Savitzky and Golay function to Fig. 8. The fractional error of the data of Fig. 7 from the universal
which the extra term is added. By trial and error it is function of Fig. 4a.
found that
X,(i,m) = 0.8m/(m + I), (18)
The errors of Fig. 8 are below 1% for W 2 1.5 eV. If
except for m = 1 where
we replace Eqs. (18) and (19) by X,(t,m) = m/(m
x&,1) = 0.375. (19) + 1) the fit of Fig. 7 deteriorates and the errors rise
Thus, the final coefficients excluding the normalisa- to a 9% deviation.
tion factor, for the f (m + 1)th terms, are given by In all of this work the normalisation factor is
(m + l>YX,(~,m> where Y is t. From Eq. (18) evaluated to ensure that a line of gradient unity gives
a differential value of unity.
(m + l)YXC(+,m) = 0.4m (20)
and from Eq. (19) for m = 1,
4. Conclusions
(m + l)YX,(+,l) = 0.375. (21)
A simple assignment of the energy width of the
Savitzky and Golay differential function is that E,
= 2ms. However, it is shown that this leads to a
number of inconsistencies and that E, = (2m + 1)~
is better, giving the same effect in spectra irrespec-
tive of the value of E. It is recommended that this
latter equation is used in assigning an energy width
to this numerical differentiation.
In order to derive differential intensities at inter-
mediate energies (e.g. even energies for E = 1 eV), a
modified Savitzky and Golay function is developed
where a fraction of the + (m + lhh channels is
added to the convolution at an intensity between 0
and f ( m + 1)/N, roughly in proportion to the frac-
tional value of E, between successive correct Sav-
itzky and Golay values. The precise fraction for
Fig. 7. As for Fig. 5 but using the modified Savitzky and Golay differentials at the mid-point between regular values
function with X,(i) as given in Eq. (16). is given by Eqs. (20) and (21).
280 IS. Gilmore, M.P.Seah/Applied Surface Science 93 (1996) 273-280

Acknowledgements [3] M.P. Seah, W.A. Dench, B. Gale and T.E. Groves, J. Phys. E
21 (1988) 351.
[4] P. Marchand and L. Marmet, Rev. Sci. Instr. 54 (1983) 1034.
This work forms part of the Valid Analytical
[5] B.C.H. Tmton, Meas. Sci. Technol. 3 (1992) 858.
Measurement programme of the National Measure- [6] G.K. Wertheim, Rev. Sci. Instr. 46 (1975) 1414.
ment System Policy Unit of the Department of Trade [7] L. Muehlhoff and H.M. Muehlhoff, J. Vat. Sci. Technol. A 4
and Industry. (1986) 1540.
[8] K. Piyakis and E. Sacher, Appl. Surf. Sci. 55 (1992) 159; 64
(1993) 361.
[9] M.P. Seah and P.J. Cumpson, Appl. Surf. Sci. 62 (1992) 195.
References [IO] R. Siuda, Appl. Surf. Sci. 81 (1994) 27.
[1 l] M.P. Seah and W.A. Dench, J. Electron Spectrosc. Rel.
[I] A. Savitzky and M.J.E. Golay, Anal. Chem. 36 (I%41 1627, Phen. 48 (1989) 43.
with corrections by J. Steiner, Y. Terrnonia and J. Deltour, [ 121 P.J. Cumpson and M.P. Seah, Surf. Interf. Anal. 18 (1992)
Anal. Chem. 44 (1972) 1906. 345.
[2] T.H. Edwards and P.D. Wilson, Appl. Spectrosc. 28 (1974) [ 131 M.P. Seah, M.T. Anthony and W.A. Dench, J. Phys. E 16
541. (1983) 848.

You might also like