Professional Documents
Culture Documents
1 s2.0 S0169433216312971 Main
1 s2.0 S0169433216312971 Main
a r t i c l e i n f o a b s t r a c t
Article history: Peak fitting is an essential part of X-ray photoelectron spectroscopy (XPS) narrow scan analysis, and the
Received 31 January 2016 Literature contains both good and bad examples of peak fitting. A common cause of poor peak fitting
Received in revised form 9 June 2016 is the inclusion of too many fit parameters, often without a sound chemical and/or physical basis for
Accepted 10 June 2016
them, and/or the failure to reasonably constrain them. Under these conditions, fit parameters are often
Available online 15 June 2016
correlated, and therefore lacking in statistical meaning. Here we introduce the uniqueness plot as a
simple graphical tool for identifying bad peak fits in XPS, i.e., fit parameter correlation. These plots are
Keywords:
widely used in spectroscopic ellipsometry. We illustrate uniqueness plots with two data sets: a C 1s
X-ray photoelectron spectroscopy
XPS
narrow scan from ozone-treated carbon nanotube forests and an Si 2p narrow scan from an air-oxidized
Uniqueness plot silicon wafer. For each fit, we consider different numbers of parameters and constraints on them. As
Correlation expected, the uniqueness plots are parabolic when fewer fit parameters and/or more constraints are
Fit parameters applied. However, they fan out and eventually become horizontal lines as more unconstrained parameters
Spectroscopic ellipsometry are included in the fits. Uniqueness plots are generated by plotting the chi squared (2 ) value for a fit vs.
a systematically varied value of a parameter in the fit. The Abbe criterion is also considered as a figure of
merit for uniqueness plots in the Supporting Information. We recommend that uniqueness plots be used
by XPS practitioners for identifying inappropriate peak fits.
© 2016 Elsevier B.V. All rights reserved.
1. Introduction There are a number of possible pitfalls in XPS peak fitting. These
include (i) adding too many unconstrained peaks or fit parameters,
X-ray photoelectron spectroscopy (XPS) is a widely used surface often without a reasonable chemical or physical basis for them,
characterization technique; it is currently mentioned in approxi- (ii) using inappropriate backgrounds, (iii) failing to show the sum-
mately 10,000 publications per year [1–3]. XPS uses peak positions mation of ones fit components, and (iv) failing to demonstrate the
and peak shapes to provide the identities and chemical states of statistical quality of one’s fit with chi squared (2 ), the reduced
the elements at surfaces. Because both the natural line widths of chi squared (2* ), the Abbe criterion, etc [11,12]. When too many
XPS signals and their chemical shifts are on the order of 1 eV, peak parameters can vary (float) in a fit, they are often correlated, which
fitting is an essential part of XPS data analysis [4,5]. However, peak means they lack statistical meaning. In XPS, the degree of cor-
fitting can be a subjective and challenging exercise [6]. For exam- relation can be identified through the Hessian matrix, H, where
ple, Wepasnick and coworkers fit a C 1s narrow scan from oxidized parameters are correlated if the mixed partial derivatives of 2 (the
carbon nanotubes (CNTs) using parameters from two previously off-diagonal elements of this matrix) are non-zero [11]. We recently
published fits [7–9]. In both cases, the overall fits were good. How- introduced the equivalent width and autocorrelation width as less
ever, the fraction of the carboxyl, C(III) [10], signal in the fits differed biased figures of merit for XPS narrow scans [13,14]. These width
drastically: ca. 6% vs. 11.0%. functions are sensitive to chemical changes in materials.
Spectroscopic ellipsometry (SE) data analysis does not involve
peak fitting per se. However, SE and XPS data analysis are simi-
lar in that both generally involve fitting multiple components and
∗ Corresponding author at: Department of Chemistry and Biochemistry, C-306,
parameters. In both cases, correlation occurs when a large num-
Benson Bldg., Brigham Young University, Provo, UT, 84602.
E-mail address: mrlinford@chem.byu.edu (M.R. Linford). ber of fit parameters are employed. In essence, ‘correlation’ means
http://dx.doi.org/10.1016/j.apsusc.2016.06.070
0169-4332/© 2016 Elsevier B.V. All rights reserved.
156 B. Singh et al. / Applied Surface Science 387 (2016) 155–162
of 500 m × 500 m, resolution: 3 (nominal pass energy 100 eV), 3. Results and discussion
number of scans: 20, and step size of 0.065 eV. The CNT spectrum
was referenced to the main C 1s peak in the narrow scan, which To show the usefulness of uniqueness plots in XPS, we peak
was taken at 284.6 eV, and the silicon spectrum was referenced to fitted narrow scans in a constrained and unconstrained fashion.
the O 1s peak taken at 532.0 eV. This latter signal appeared as a As expected, when the number of unconstrained fit parameters
single, shoulderless peak. Peak fitting was performed in our instru- increases, the uniqueness of the fits decreases.
ment software (‘ESCA Data Analysis’ version V5.01.02), and fits
were optimized based on a minimization of the 2* value. Shirley 3.1. Ozone-treated CNT sample (varying C-C peak width)
type backgrounds [29] were used for all the fits.
SE data were collected from ca. 200–1700 nm with an M-2000 A reasonable peak fit to a C 1s narrow scan of ozone-treated
instrument from the J.A. Woollam Co., Lincoln, NE. For Fig. 1, the CNTs was obtained by introducing peaks that represent: (i) carbon
data were range selected from 300 to 400 nm. The time necessary bonded to carbon and hydrogen (C-C), where this peak was initially
for the software to generate a uniqueness plot was excessively long placed at ca. 284.6 eV, but then allowed to float – it had a vari-
when a wider wavelength window was used with an unconstrained able position of x eV, (ii) carbon bonded to one oxygen atom (C-O)
model. at x + 1.5 eV, (iii) carbonyl carbon (C O) at x + 2.7 eV, and (iv) car-
boxyl carbon (O C O) at x + 3.9 eV. These peaks can be designated
by their oxidation states: C(0), C(I), C(II), and C(III), respectively
2.1. Theory of error estimation tools [13]. As appropriate for an aromatic material, a shake-up peak at
x + 6.5 eV was also included. All peak widths were constrained to the
The error estimation tools employed herein are the reduced chi same value, except the shake-up peak, which had twice this width.
squared (2* ) and Abbe criterion values. Hesse has described these The Voigt function, a convolution of a Gaussian and a Lorentzian,
error estimation tools in detail [11,12] which help determine the is the ideal function for modeling XPS narrow scans [5]. However,
quality of a fit vis-à-vis the difference between the experimental pseudo-Voigt functions (Gaussian:Lorentzian (G:L) sum or product
data and the modeled (fit) envelope. The fit envelope is the sum of functions) are also commonly used in XPS peak fitting. Of the two,
the individual fit components in a model. the G:L product function is less effective [5,30]. The only option in
The chi squared value is defined as follows: our software at BYU is the G:L sum function. All the peaks mod-
eled herein had symmetric G:L sum functions and the same G:L
N 2
M (i) − S i, p ratio, unless otherwise specified. The heights of the peaks floated.
2 =
p (1) Of course all these assignments are semi-empirical. Fig. 2a shows
M (i)
i=1 the optimal fit obtained through this approach, which gave a C-C
peak width of 1.37 eV and a G:L ratio of 93%.
) refer to the values of the experimental data
Where, M(i) and S(i, p In a uniqueness plot, one fit parameter is systematically fixed to
and fit at a given point, respectively. The vector, p , contains the fit different values about its optimal value, while the other adjustable
parameters used in the model. The difference between the experi- parameters vary to obtain the best possible fit. This is illustrated
mental data and fit envelope is squared at each data point and then graphically in Fig. 2a–d. Fig. 2a shows the width of the C-C peak
normalized to the value of experimental data point. The reduced at its optimal value of 1.37 eV in a fit. In Fig. 2c–d, this peak width
chi squared value is defined as: was fixed to other values and a new fit was performed. For these
latter fits, the only parameter that floated was the G:L ratio of the
2 p C-C peak – obviously one would not expect changes in this ratio
=
2∗ p (2) to compensate for the large changes in the C-C peak width shown
N−P
in the figure. The lack of fit in Fig. 2b–d, is shown in the residuals,
where N is the number of data points and P is the number of fit which clearly increase as the system diverges from the optimal C-C
parameters employed. width. This system appears to be unique.
Obviously, a lower value of 2 or 2* implies a better fit. How- A more quantitative way of describing the quality of the fits
ever, as Sherwood clarified [4], 2 is of greatest value when it is in Fig. 2 is through 2* and/or Abbe criterion values in unique-
reported for a series of peak fits on the same spectrum, or when ness plots. Fig. 3 shows a plot of 2* vs. C-C peak width (the black
applied to similar spectra. squares) for the series of fits that is partially shown in Fig. 2. Here,
) − M(i).
The residuals, R(i), of a fit are given by S(i,p the C-C peak widths were systematically varied from 0.8 to 2.0 eV.
The Abbe criterion is defined as: The parabolic shape of this uniqueness plot (the ‘Approach 1’ data)
is consistent with uniqueness, and therefore statistical meaning,
N−1 2
1 [R (i + 1) − R (i)] to the C-C peak width and model used in Fig. 2. As expected, the
i=1
Abbe =
2
N 2
(3) minimum error (2* value) is obtained at the optimal value of the
i=1 [R (i)] C-C peak width. Uniqueness plots based on the Abbe criterion are
shown in the Supporting Information. They tend to be inversely
Where R(i + 1) and R(i) refer to the residuals at the (i + 1)th and ith related to the plots based on reduced chi squared values.
data points, respectively. An Abbe criterion of zero indicates that To further show the use of uniqueness plots in XPS, the con-
the fit envelope consistently misses the experimental peak enve- straints on the fit/approach in Fig. 2 were increasingly relaxed (see
lope by the same amount, i.e., R(i) = R(i + 1). An Abbe criterion value Table 1). The resulting uniqueness plots are presented in Fig. 3.
of two is obtained for anti-correlated residuals. The Abbe criterion Here, the general result is that as the constraints are removed, the
is unity when the residuals are statistically distributed as random curves become increasingly flat, fanning out and showing less cur-
noise. This is the ideal case. vature, while the minimum 2* value drops somewhat. Finally, with
Of course it is vital to use common sense and a priori knowl- no constraints (Approach 8), a horizontal line is obtained (dark
edge when peak fitting XPS narrow scans [4,12,26], e.g., one should green hexagons). This flat line, like the one in Fig. 1, indicates
not ignore earlier measurements on similar systems. For example, that the C-C peak width can be varied over a wide range of val-
while it is possible you will find unknown species by XPS when you ues, and the other fit parameters can compensate for any change
oxidize silicon, it is unlikely. in it, always yielding the same, low error. For this unconstrained
158 B. Singh et al. / Applied Surface Science 387 (2016) 155–162
Fig. 2. (a) Peak fitting of a C 1s narrow scan of ozone-treated CNTs obtained using the most constrained approach in Table 1 (Approach 1) showing the optimal C-C peak
width value of 1.37 eV. (b–d) Poorer peak fits, also from Approach 1, for constrained C-C peak widths of 1.00, 1.20, and 1.90 eV, respectively. Black traces at the top of each
panel are residuals to the fits.
Table 1
Fitting approaches used to generate the eight uniqueness plots in Fig. 3. The number of fixed parameters decreases proceeding down the table. Fits were to the C 1s spectrum
of ozone-treated CNTs shown in Figs. 2 and 4, and the corresponding uniqueness plots are based on varying the width of the C-C peak.
1 All peak positions, peak widths, peak heights Gauss% of C-C peak
and Gauss% (except C-C Gauss%)
2 All peak positions, peak widths, and peak Gauss% of all peaks
heights
3 All peak positions, peak widths, and peak Gauss% of all peaks, and peak height of C-C
heights (except C-C peak) peak
4 All peak positions, and peak widths Gauss% and peak heights of all peaks
5 All peak positions, and peak widths (except Gauss%, peak heights of all peaks, and peak
shake-up peak) width of shake-up peak
6 All peak positions Gauss%, peak heights, and peak widths of all
peaks
7 All peak positions (except C-C peak) Gauss%, peak heights, and peak widths of all
peaks, peak position of C-C peak
8 Nothing All Gauss%, peak heights, all peak positions,
and all peak widths
Fig. 4. (a) Peak fitting of a C 1s narrow scan of ozone-treated CNTs obtained using the least constrained approach in Table 1 (Approach 8) showing the optimal peak width
value of 1.37 eV. (b–d) Equally unreasonable peak fits for constrained C-C peak widths of 1.00, 1.20, and 1.90 eV, respectively. Black traces at the top of each panel are residuals
to the fits. (For interpretation of the references to colour in the text, the reader is referred to the web version of this article.)
Table 2
Approaches used to generate the eight uniqueness plots in Fig. 5. The number of fixed parameters decreased from Approach 1 to Approach 8. All fits were to the C 1s spectrum
of ozone-treated CNTs shown in Figs. 2 and 4, and are based on changing and fixing the position of the C-O peak.
1 All peak positions, peak widths, peak heights Gauss% of C-O peak
and Gauss% (except C-O Gauss%)
2 All peak positions, peak widths, and peak Gauss% of all peaks
heights
3 All peak positions, peak widths, and peak Gauss% of all peaks, and peak height of C-C
heights (except C-C peak) peak
4 All peak positions, and peak widths Gauss% and peak heights of all peaks
5 All peak positions, and peak widths (except Gauss%, peak heights of all peaks, and peak
C-C) width of C-C peak
6 All peak positions Gauss%, peak heights, and peak widths of all
peaks
7 All peak positions (except shake-up peak) Gauss%, peak heights, and peak widths of all
peaks, peak position of shake-up peak
8 Nothing All Gauss%, peak heights, all peak positions,
and all peak widths
in increments of 0.05 eV, and the 2* and Abbe criterion values
were calculated at each of these values. As expected, the most con-
strained approach yielded the curve (parabola) with the highest
curvature. As the constraints were relaxed, the uniqueness plots
showed less curvature, fanning out. The unconstrained approach
in Table 2 (Approach 8) showed the lowest 2* value with very
low curvature. These results again illustrate that in unconstrained
models the fit parameters lack statistical meaning.
Table 3
Approaches used to generate the six uniqueness plots in Fig. 7. The number of fixed parameters decreased from Approach 1 to Approach 6. All fits were to an Si 2p spectrum
from a sample of air-oxidized silicon (see Fig. 6), and are based on varying the width of the SiO2 peak.
1 All peak positions, peak widths, and peak Gauss% of all peaks
heights
2 All peak positions, peak widths, and peak Gauss% of all peaks and peak height of SiO2
heights (except SiO2 peak) peak
3 All peak positions, and peak widths Gauss% and peak heights of all peaks
4 All peak positions, and peak widths (except Si Gauss% and peak heights of all peaks, and peak
peak) width of Si peak
5 Peak position of Si peak Gauss% and peak heights of all peaks, peak
width of Si peak, and peak position of SiO2 peak
6 Nothing All Gauss%, peak heights, all peak positions and
peak widths
Fig. 6. Optimal fit to an Si 2p narrow scan using two components: a bulk Si and an Fig. 8. Optimal fit to an Si 2p narrow scan using five components: a bulk Si peak and
oxidized Si peak. four oxidized Si peaks denoted Six+ , with x = 1–4. This fitting approach is adapted
from Himpsel et al. [17].
Table 4
Approaches used to generate the six uniqueness plots in Fig. 9. The number of fixed parameters decreased from Approach 1 to Approach 6. All fits were to the Si 2p spectrum
from a sample of air-oxidized silicon (see Figs. 6 and 8), and are based on varying the width of the Si4+ peak.
1 All peak positions, peak widths, and peak Gauss% of all peaks
heights
2 All peak positions, peak widths, and peak Gauss% of all peaks and peak height of Si4+ peak
heights (except Si4+ peak height)
3 All peak positions, and peak widths Gauss% and peak heights of all peaks
4 All peak positions All Gauss%, peak heights, and peak widths
5 All peak positions, except Si3+ peak Gauss%, peak heights, and peak widths of all
peaks, and peak position of Si3+ peak
6 Nothing All Gauss%, peak heights, all peak positions and
peak widths
not remove from the analyst the need to read and understand the
related literature, seek expert advice, use restraint/common sense
in designing a model, etc. We recommend that uniqueness plots be
used in XPS peak fitting efforts.
Acknowledgements
[27] D.S. Jensen, S.S. Kanyal, N. Madaan, M.A. Vail, A.E. Dadson, M.H. Engelhard, [33] T. Marten, W. Olovsson, S.I. Simak, I.A. Abrikosov, Phys. Rev. B 72 (5) (2005)
M.R. Linford, Surf. Sci. Spectra 20 (1) (2013) 36–42. 054210.
[28] S.S. Kanyal, D.S. Jensen, Z. Zhu, M.R. Linford, Surf. Sci. Spectra 22 (2) (2015) [34] F.J. Himpsel, F.R. McFeely, A. Taleb-Ibrahimi, J.A. Yarmoff, G. Hollinger, Phys.
1–6. Rev. B 38 (9) (1988) 6084–6096.
[29] D.A. Shirley, Phys. Rev. B 5 (12) (1972) 4709–4714. [35] A. Herrera-Gomez, F.-S. Aguirre-Tostado, P. Pianetta, J. Vac. Sci. Technol. A 34
[30] R. Hesse, P. Streubel, R. Szargan, Surf. Interface Anal. 39 (5) (2007) 381–391. (2) (2016) 020601.
[31] R.J. Cole, P. Weightman, J. Phys.: Condens. Matter 10 (25) (1998) 5679.
[32] W. Olovsson, T. Marten, E. Holmström, B. Johansson, I.A. Abrikosov, J. Electron
Spectrosc. 178–179 (0) (2010) 88–99.