Xps Wavelet

download Xps Wavelet

of 10

Transcript of Xps Wavelet

  • 7/29/2019 Xps Wavelet

    1/10

    SURFACE AND INTERFACE ANALYSISSurf. Interface Anal. 2004; 36: 7180Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/sia.1650

    Noise filtering and deconvolution of XPS data

    by wavelets and Fourier transform

    Catherine Charles,1

    Gervais Leclerc,2

    Pierre Louette,2

    Jean-Paul Rasson1

    and Jean-Jacques Pireaux2

    1 Department of Mathematics, Facult es Universitaires Notre-Dame de la Paix, B-5000 Namur, Belgium2 Department of Physics, LISE Laboratory, Facult es Universitaires Notre-Dame de la Paix, B-5000 Namur, Belgium

    Received 10 September 2002; Revised 15 October 2003; Accepted 15 October 2003

    In experimental sciences, the recorded data are often modelled as the noisy convolution product of an

    instrumental response with the true signal to find. Different models have been used for interpreting

    x-ray photoelectron spectroscopy (XPS) spectra. This article suggests a method of estimate the true XPS

    signal that relies upon the use of wavelets, which, because they exhibit simultaneous time and frequency

    localization, are well suited to signal analysis.

    First, a wavelet shrinkage algorithm is used to filter the noise. This is achieved by decomposing the

    noisy signal into an appropriate wavelet basis and then thresholding the wavelet coefficients that contain

    noise. This algorithm has a particular threshold related to frequency and time.

    Secondly, the broadening due to the instrumental response is eliminated through a deconvolution

    process similar to that developed in the previous paper in this series for the analysis of HREELS data. This

    step mainly rests on least-squares and on the existing relation between the Fourier transform, the wavelet

    transform and the convolution product. Copyright 2004 John Wiley & Sons, Ltd.

    KEYWORDS: XPS; wavelets; deconvolution; Poisson noise

    INTRODUCTIONWith reference to the two other papers in this series (this

    issue), wavelet analysis of XPS data must concentrate on

    different areas than applications to High-resolution elec-

    tron energy-loss spectroscopy (HREELS) data, for differ-

    ent reasons.

    First, a simple consideration of the fundamental differ-

    ences in the excitation processes reveals that in HREELS the

    elemental vibrational excitations have in general an intrinsic

    linewidth close to zero, at least at the scale of probe elec-

    trons that have an energy (110 eV for example) defined

    to 1 meV 8 cm1: Fourier transform infrared spectrashowing vibrational bands with a resolution of 0.1 cm

    1

    arenot uncommon, justifying the fact that a deconvolution pro-

    cedure of HREELS data should consider the true signal as a

    Dirac delta function. In XPS, in contrast, intrinsic core-level

    linewidths are in the range of 0.1 (metals) to 0.4 eV (carbon),

    while stand-alone spectrometers have intrinsic resolution

    (monochromatized x-ray source and analyser contributions)

    in the same energy range: peaks are naturally broad and

    should be approximated by finite Gaussian or Lorentzian

    shape (or a combination thereof).

    Correspondence to: Jean-Jacques Pireaux, Department of Physics,

    LISE Laboratory, Facultes Universitaires Notre-Dame de la Paix,B-5000 Namur, Belgium. E-mail: [email protected]/grant sponsor: Belgian FNRS.Contract/grant sponsor: Belgian Prime Ministers Office;Contract/grant number: PAI/UIAP(4/10).

    Secondly, the literature is rich in contributions onspectra acquisition and data handling in XPS. All the

    basic principles are illustrated in Sherwoods reference

    review,1 while some state-of-the-art concerns have been

    listed and commented more recently; as a result of an

    international workshop entitled X-Ray photoelectron Spec-

    troscopy: from Physics to Data, a compilation of recent

    developments in available handbooks and databases, in

    data processing software and standard test data is pre-

    sented and well documented.2 But, when processing XPS

    core-level spectra one should stay aware of the genuine

    algorithms hypotheses, and include easy-to-interpret sta-

    tistical diagnostics to judge objectively the quality of the

    regression.3

    The purpose of this contribution is therefore to explore

    what specific contribution(s) wavelet analysis could bring

    to XPS data: noise filtering and deconvolution (recovery

    of the true signal) will be tested on synthetic (theoret-

    ical) spectra and then applied to real spectra. As in the

    previous paper in this series (this issue) for the work

    on HREELS data processing, care will be taken to try

    to recover real peak intensities, a prerequisite to keep

    quantification through data analysis. Note that this con-

    tribution will not discuss issues related to the choice and

    use of routines (linear, polynomial, Shirley, Tougaard, etc.)

    to estimate background, or the algorithms to allow com-positional depth profile data to be deduced from angle-

    resolved XPS measurements, which are beyond the scope of

    this work.

    Copyright 2004 John Wiley & Sons, Ltd.

  • 7/29/2019 Xps Wavelet

    2/10

    72 C. Charles et al.

    WAVELET ANALYSIS MODEL AND METHOD

    As a mathematical model of the XPS spectrum, it is chosen

    to use

    tri Poi C aC bi with i D s li, i D 1 . . . n 1

    where X Po

    indicates that X is a Poisson randomvariable of parameter , s D sjlggjD1 is a Gaussian (orLorentzian or combination thereof) instrumental function

    (transmission function of the spectrometer), lgg 2 IN, l Dlj

    nlgg1jD1 refers to the true signal to be estimated and is

    modelled as a sum of different Lorentzians, a is a constant

    instrumental noise and b D biniD1 is a variable physicaland statistical noise (linear, Shirley, etc.); i is the channel

    number. Note that this model considers the true signal

    as a sum of Lorentzians but the method presented in this

    paper to estimate this signal is easy to generalize to a sum

    of Gaussians.

    The non-ideality of the spectra arises through several

    sources: the convolution of the electronic scattering functionwith the instrumental response, the constant noise, the vari-

    able background and the occurrence of Poisson-distributed

    random noise.

    In this paper a new method is proposed that filters the

    Poisson noise and reduces the broadening caused by the

    instrumental response. The suggested methods rely mostly

    upon wavelet signal processing. The choice of wavelets was

    justified in the first paper of this series (this issue) and

    also applied to HREELS spectra in the second paper of

    this series (this issue) the reader who wishes to gain a

    basic understanding of wavelet theory and wavelet signal

    processing is referred to these papers.In the proposed method, the noise is filtered using a

    wavelet shrinkage technique. Most publications concerned

    with the filtering of XPS spectra have focused so far on esti-

    mation of a signal to which an additive zero-mean normal

    noise is superposed. However, XPS is a technique in which

    the detection is performed by an electron counting device.

    Particle counting usually follows a Poisson probability dis-

    tribution. Consequently, the signal in each channel can be

    modelled by this distribution. The proposed approach is

    based on an idea by Donoho and Johnstone:4 6 decomposing

    the noisy signal in an appropriate wavelet basis and then

    thresholding the noisy coefficients. The DonohoJohnstone

    technique was developed for Gaussian noise. Recently,Kolaczyk7 has suggested a similar wavelet shrinkage tech-

    nique but for a specific case of Poisson-distributed noise. Our

    second paper in this series has introduced a wavelet shrink-

    age method related to the unconstrained Poisson process

    generalizing Kolaczyks algorithm. The same methodology

    will be applied to XPS spectra. The constant and variable

    noise are removed by subtraction.

    Once the filtering step is completed, the broadening due

    to Gaussian instrumental function is eliminated via a similar

    deconvolution process to that for HREELS. This will be

    explained later. This article concludes with an analysis of the

    results on the basis of the quadratic error, graphics and thetheory of XPS.

    As in the second paper concerning HREELS, most of the

    examples presented here use synthetic data.

    NOISE FILTERING

    Consider the model for an XPS spectrum described by

    Eqn. (1). The first step consists of removing the random

    (statistical) noise. The problem is then

    given the observations

    x D xin1iD0 where each xi Poi, estimate in1iD0 2

    This problem is solved by using the wavelet shrinkage

    method explained in the previous paper (this issue).

    To summarize briefly the calculation, the noise removal

    algorithm (LTFTIPSH) proceeds in three steps:

    (1) Decomposition of the observations in the Haar wavelet

    basis.

    (2) Thresholding at each scale and at each time of the

    coefficients relative to the details.

    (3) Application of the inverse Haar transform to the

    threshold coefficients.

    We prefer using the translation-invariant Haar wavelet

    transform to avoid a step estimator.

    Below is a presentation of some results (real experimental

    data) obtained so far with this LTFTIPSH algorithm. Figure 1

    is a survey spectrum of a clean copper foil, with quite

    good signal-to-noise ratio. Figure 2 reports a widescan of an

    organic layer with more noise and Fig. 3 illustrates a narrow-

    scan spectrum with (probably; see later) six core-level peaks.

    The advantages and efficiency of this algorithm are

    explained in detail in the second paper of this series. It is

    clear from the displayed data that the LTFTIPSH algorithm

    does not introduce any spurious information and operatesequally well in spectral areas with low and high intensity

    counts. Some comparisons with other filtering methods are

    made in Ref. 8.

    The second step consists of removing the constant and

    variable noise, as in conventional XPS data-handling codes.

    Given the mathematical model of Eqn. (1), a is estimated as

    the minimum oftr and b as linear of Shirley background, for

    example. Figure 4 presents such a result for the real C1s C K2pexperimental spectrum shown previously.

    Figure 5 illustrates the result of two filterings. The first

    filters the random noise using the LTFTIPSH algorithm and

    the second subtracts the constant and variable noise.

    LOCALIZED LEAST-SQUARES

    The previous section has explained how our XPS spectra

    are filtered. After filtering, the problem stated in Eqn. (1) is

    reduced to the following deconvolution problem

    given tri D s li 1 i n, estimate l D linlggC1iD1 3

    The reader is reminded of the following hypotheses:

    (1) tri corresponds to the experimental spectrum, i.e. the

    number of detected electrons versus binding energy.(2) s D silggiD1 is a Gaussian (or Lorentzian or combinationthereof) function featuring the spectrometer transmis-

    sion function (a Gaussian function is 1/p

    2ex2/22 ,

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    3/10

    Deconvolution of XPS data by wavelets 73

    -2000200400600800100012000

    1

    2

    3

    4

    5

    6

    7

    8 (a) (b)

    x 104

    COUNTS

    BINDING ENERGY (eV)

    -2000200400600800100012000

    1

    2

    3

    4

    5

    6

    7

    8x 104

    COUNTS

    BINDING ENERGY (eV)

    Figure 1. (a) Experimental XPS spectrum of copper foil after gentle ArC sputtering. (b) Noise filtering of (a) using theLTFTIPSH algorithm.

    0200400600800100012000

    0.5

    1

    1.5

    2

    2.5x 104 x 104

    BINDING ENERGY (eV)

    COUNTS

    0200400600800100012000

    0.5

    1

    1.5

    2

    2.5

    BINDING ENERGY (eV)

    COUNTS

    (a) (b)

    Figure 2. (a) Experimental XPS spectrum of a thin organic layer on silicon wafer. (b) Noise filtering of (a) using the

    LTFTIPSH algorithm.

    COU

    NTS

    BINDING ENERGY (eV)

    310 305 300 295 275290 285 280

    2500

    2000

    1500

    1000

    500

    0

    COU

    NTS

    BINDING ENERGY (eV)

    310 305 300 295 275290 285 280

    2500

    2000

    1500

    1000

    500

    0

    (a) (b)

    Figure 3. (a) Experimental XPS spectrum of carbon and potassium contamination on a dental titanium implant. (b) Noise filtering of

    (a) using the LTFTIPSH algorithm.

    where and are, respectively, the mean and the stan-

    dard deviation; a Lorentzian function is 1 2 C x 2 ,where and are, respectively, the mean and the stan-

    dard deviation).

    (3) l is a sum of Lorentzian functions, representing the

    true spectrum.

    This problem is similar to the HREELS problem, con-

    sequently it is proposed to resolve it by using the same

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    4/10

    74 C. Charles et al.

    COUN

    TS

    275

    BINDING ENERGY (eV)

    310 305 300 295 290 285 280

    2500

    2000

    1500

    1000

    500

    0

    Figure 4. The XPS spectrum of Fig. 3(a) and an estimation of

    its constant and variable noise.

    COUNTS

    275

    BINDING ENERGY (eV)

    310 305 300 295 290 285 280

    2500

    2000

    1500

    1000

    500

    0

    -500

    Figure 5. The XPS spectrum of Fig. 3(a) after noise filtering

    and background subtraction.

    strategy: localization of peaks by wavelets and estimation of

    their intensities by localized least-squares (LLS). This section

    focuses on the LLS procedure.

    Assume that the instrumental response s is known

    and that the true spectrum is one Lorentzian function

    or one ensemble of Lorentzian functions. Under this setof assumptions, it is not proposed to use the algorithm

    suggested in the HREELS case

    minl

    i

    [tri sli]2

    because it does not impose the Lorentzian character of

    the result. However, it would be interesting to take

    into account this particular hypothesis to obtain a better

    estimation. Moreover, this HREELS algorithm was based on

    the resolution of a minimization problem that owns a small

    number of variables (only some components ofl are differentto zero). However, there are a larger number of variables in

    the XPS case, because a Lorentzian is never zero. Thus, the

    previous algorithm risks being heavy on calculation time.

    Consequently, it is proposed to minimize

    min1,1,2,2,...

    j

    [trj s lj]2 where lt D

    i

    1

    i

    2i C t i24

    In this case, theparameters i, i of the different Lorentzians

    are the only variables and l is imposed to be one ensemble ofLorentzian functions. Therefore, if an initial point 0i ,

    0i is

    chosen for each Lorentzian sufficiently close to the solution,

    this method is expected to converge towards the global

    minimum that is the best solution to the problem.

    It is therefore necessary to choose a good initialization.

    The first idea has been to choose the inverse Fourier

    Transform of TR/S, where TR (S) is the Fourier transform

    of tr (s). If the deconvolution is realized by the Fourier

    transform, its main disadvantage is a possible division by

    zero. However, in the XPS case, the instrumental function s

    consists of a Gaussian (or Lorentzian or combination thereof)

    and it is known that its Fourier transform is never null.

    For example

    g1t D1p2

    expt2/2 ! G1 D exp2/2

    gt D1p2

    expt2/22 ! G D exp22/25

    However, even if this Fourier transform is theoretically never

    zero, in practice, if or is sufficiently large the Fourier

    transform of the Gaussian (Lorentzian) function becomes

    very small for large (see Eqn. (5)).In this case, thecomputer

    classifies it as zero. It is suggested bypassing this difficultyas follows. As in Eqn. (38) of the first paper in this series

    dtrj,k D

    l dsj,.

    k; ctrj,k D

    l csj,.

    k 6

    where dfj,k and c

    fj,k are the wavelet coefficients related to

    details and to a rough approximation, respectively, of the

    functionf at scalej and time k. These formulae mean that the

    convolution product is preserved by the wavelet transform.

    Therefore, iftr and s are decomposed in a wavelet basis and

    if the Fourier transform of dtrj,. is divided by the Fourier

    transform of dsj,., the Fourier transform of an estimator of

    l is obtained. Often, a quite good estimation is obtainedwith a good choice of wavelets. Between all the estimations

    obtained with all the wavelets, the one that has the smallest

    error is chosen and termed el. The error is defined as the

    quadratic error between the trace and the convolution of the

    estimated el and s. The parameters 0i and 0i relative to el are

    estimated by

    min1,1,2,2,...

    j

    i

    1

    i

    2i C xj i2 elj

    27

    The initial point of the LLS method for XPS is 0i , 0i iD1,... .

    Figure 6 illustrates the entire method. Figure 6(a) repre-sents a synthetic XPS spectrum generated by convoluting

    l, the true signal (sum of two Lorentzians of parameters

    1 D 3.54, 1 D 100, 2 D 4.55, 2 D 120), with s, the Gaussian

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    5/10

    Deconvolution of XPS data by wavelets 75

    -400 -200 0 200 400 600 8000

    0.1

    0.2

    0.3

    0.4

    0.5

    -400 -200 0 200 400 600 800-0.1

    -0.05

    0

    0.05

    0.1

    0.15

    0.2

    0.25

    0.3

    0.35

    0.4

    -400 -200 0 200 400 600 800-0.1

    -0.05

    0

    0.05

    0.1

    0.15

    0.2

    0.25

    0.3

    0.35

    0.4

    -400 -200 0 200 400 600 8000

    0.005

    0.01

    0.015

    0.02

    0.025

    0.03

    0.035

    0.04

    (a) (b)

    (d)(c)

    Figure 6. (a) Synthetic XPS spectrum (total envelope) and the true signal l (two discrete peaks) to estimate. (b) Estimation el

    (deconvolution by Fourier and wavelets) and the true signall. (c) The Lorentzians obtained with

    0

    i and 0

    i (solution of theoptimization problem (Eqn. (7)) and the true signal l. (d) Final result (solution of the optimization problem (Eqn. (4)) and the

    calculated true signal l.

    instrumental function D 10. The convolution productis termed tr. First tr and s are decomposed in a wavelet

    basis. Secondly, for each scale, the Fourier transform of the

    wavelet coefficients of tr are divided by the Fourier trans-

    form of the wavelet coefficients ofs; then the inverse Fourier

    transform of this division is calculated. The results are some

    estimations of l. Estimation el is that particular one giving

    the smallest error between the trace and the convolution of

    this estimation with s. In Fig. 6(b) it can be seen that thisestimation el recovers the two peak positions but also many

    ghost peaks. Thirdly, in order to impose the Lorentzian char-

    acter of the estimation of the true signal l, the parameters

    01 , 01,

    02 ,

    02 of the significant peaks of el are estimated

    by resolving the optimization problem (Eqn. (7)) Figure 6(c)

    illustrates the sum of the two Lorentzians obtained with 01 ,

    01, 02 ,

    02 and l. Finally, the optimization problem (Eqn. (4))

    is resolved with 01 , 01 ,

    02 ,

    02 as the initial point. The solution

    of this final problem or deconvolution result for this local-

    ized peak structure and the true signal l are represented in

    Fig. 6(d).

    The final result in Fig. 6(d) is satisfactory. The differencebetween l, thetrue signalto be calculated, andthe estimation

    is small. This result is better evaluated by comparing

    the positions and intensities of the peaks in Table 1. The

    Table 1. Parameters and estimated parameters of the

    Lorentzians of Fig. 6(d)

    Estimation of Estimation of

    100 3.54 100 3.57

    120 4.55 120 4.03

    positions are exactly recovered, the intensity of the first peakis well estimated and the intensity of the second peak is

    underestimated by some 10%.

    In order to analyse the efficiency of the method when

    two Lorentzians form only one peak in tr, the results when

    the distance between the two Lorentzians decreases while

    the intensity is kept equal and constant (Table 2) have to be

    analysed. As in HREELS, eval is defined as the FWHM of

    the instrumental function divided by the distance between

    the positions of the two Lorentzians. In the same way, inra

    is defined as the ratio of the smallest Lorentzian parameter

    to the largest Lorentzian parameter. The positions of the two

    Lorentzians are exactly recovered and the intensities are wellestimated even when the peaks are very close to each other.

    Note that the two Lorentzians visually form only one peak

    in tr in all these cases.

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    6/10

    76 C. Charles et al.

    Table 2. Parameters and estimated parameters of two

    Lorentzians with the same intensity but with a different

    interdistance

    eval

    Estimation

    of

    Estimation

    of

    2.35 100 3 100 3.13110 3 110 2.83

    1.567 100 3 100 3

    115 3 115 2.93

    1.175 100 3 100 2.98

    120 3 120 2.98

    Table 3. Parameters and estimated parameters of

    Lorentzians with the same position

    inra

    Estimation

    of

    Estimation

    of

    1/2 100 1 100 0.94

    120 2 120 1.28

    2/3 100 2 100 1.97

    120 3 120 2.53

    3/5 100 3 100 2.95

    120 5 120 3.97

    Table 3 analyses the results when the distance between

    the two Lorentzians is constant and the intensity of the two

    Lorentzians differs. The positions of the two Lorentzians are

    exactly recovered, the smallest intensity is well estimated

    but the largest intensity is not so well estimated. Note thatthe two Lorentzians visually form only one peak in tr for all

    these cases.

    DECONVOLUTION WITH WAVELETS

    The procedure for solving the deconvolution problem of

    Eqn. (1)consistsof applying theLLS approach to thedifferent

    features of the filtered spectrum. Detection of the position of

    the peaks and estimation of the instrumental function g are

    well explained and discussed in the two other papers of this

    series. Below is a presentation of some results of this new

    method applied to synthetic filtered spectra.

    In Fig. 7 is presented an illustration of the methodonce the position of the peaks in tr is detected and the

    instrumental function is estimated. In Fig. 7(a), a synthetic

    XPS spectrum has been generated by convoluting the true

    signal l (sum of five Lorentzians of different intensities) with

    Gaussian instrumental function s. The convolution product

    is termed tr. First tr and s are decomposed in a wavelet basis.

    Secondly, for each scale, the Fourier transform of the wavelet

    coefficients of tr is divided by the Fourier transform of the

    wavelet coefficients ofs and the inverse Fourier transform of

    this division is taken. These results are some estimations ofl.

    Estimation el is the one that gives the smallest error between

    the trace tr and the convolution of this estimation with s.Figure 7(b) presents the estimated el and the true signal l.

    Thirdly, in order to impose the Lorentzian character of the

    estimated el, the parameters 0i , 0i i of the significant peaks

    of el are calculated by resolving the optimization problem

    (Eqn. (7)) for each peak of tr. Figure 7(c) illustrates the sum

    of the Lorentzians obtained with 0i , 0i and l. Finally, for

    each peakthe optimization problem (Eqn. (4)) is resolved with

    0i , 0i as initial point. The solution of this final problem and

    the true signal l are represented in Fig. 7(d). This result is

    excellent as allthe intensities andpositions of theLorentziansare recovered.

    Figure 8 presents another synthetic spectrum obtained

    with five Lorentzians of different intensities convoluted with

    a Gaussian; the same strategy as for Fig. 7 has been applied

    to this spectrum. The true signal (the sum of the five

    Lorentzians) and their estimation are shown in Fig. 8(b).

    For better evaluation of the result, parameters and of

    the true Lorentzians and their estimation via the wavelet

    algorithm are presented in Table 4.

    With the graphics in Fig. 8 and the values in Table 4 it

    can be seen that the positions of the Lorentzians are well

    recovered and estimation of the true signal is satisfactory.

    Only Lorentzian 4 (that, with another Lorentzian, leads to

    one single peak in tr) has its intensity not so well estimated

    (eval D 1.175 and inra D 1/2).

    OTHER RESULTS

    Some results corresponding to experimental XPS spectra are

    reported in Figs 9 13. The noise filtering algorithm and the

    deconvolution method presented earlier have been applied.

    In Fig. 9, the same experimental spectrum as that shown

    and filtered in Figs 35 is presented; the estimated original

    XPS function (represented multiplied by a factor of 0.153)

    has been added. In order to evaluate the result obtained withthe new method, the result has been compared with data

    from a standard curve-fitting procedure. The curve-fitting

    method estimates the Lorentzians once the number of peaks,

    their positions, their widths and the type of noise are chosen.

    This comparison is reported in Table 5. The positions of the

    peaks are very similar and the intensity of the peaks differs

    slightly. Note that one additional peak is discovered with the

    wavelet-based code.

    The spectral analysis is completed in 12 min on a

    Pentium III biprocessor (900 MHz). In fact, thenoise filtering,

    the estimation of s and the estimation by Fourier and

    wavelets take

    1 min. Only theresolution of theoptimization

    problems is time consuming.Below are two other results of the new procedure applied

    to very noisy experimental XPS spectra: a single Ca 2p

    doublet (Figs 10 and 11) and a structured C 1s spectrum

    (Figs 12 and 13).

    Table 4. Parameters and estimated parameters of

    the Lorentzians of Fig. 8

    Estimation of Estimation of

    1 3 1 2.95

    50 2 50 1.74

    100 1 100 0.83120 2 119 1.14 x

    220 3 220 2.97

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    7/10

    Deconvolution of XPS data by wavelets 77

    -400 -200 0 200 400 600 800

    0

    0.1

    0.2

    0.3

    0.4

    0.5

    -400 -200 0 200 400 600 800-0.1

    -0.05

    0

    0.05

    0.1

    0.15

    0.2

    0.25

    0.3

    0.35

    0.4

    -400 -200 0 200 400 600 800-0.1

    -0.05

    0

    0.05

    0.1

    0.15

    0.2

    0.25

    0.3

    0.35

    0.4

    -400 -200 0 200 400 600 800-0.1

    -0.05

    0

    0.05

    0.1

    0.15

    0.2

    0.25

    0.3

    0.35

    0.4

    (a)

    (c) (d)

    (b)

    Figure 7. (a) Synthetic XPS spectrum and the true signal l to estimate. (b) Estimation el (deconvolution by Fourier and wavelets)

    and the true signal l. (c) The Lorentzians obtained with 0i

    and 0i

    (solution of the optimization problem (Eqn. (7))) and the true

    signal l. (d) Final result (solution of the optimization problem (Eqn. (4)) for each peak) and the true signal l.

    -400 -200 0 200 400 600 800-0.01

    0

    0.01

    0.02

    0.03

    0.04

    0.05

    0.06

    1 2 3 4 5

    -400 -200 0 200 400 600 800-0.01

    0

    0.01

    0.02

    0.03

    0.04

    0.05

    0.06(a) (b)

    Figure 8. (a) Synthetic XPS spectrum trand the true signal l to estimate. (b) The true signal land its estimation.

    Note that in Fig. 10 the estimated original XPS function

    is represented multiplied by a factor of 0.338. In order

    to appreciate the result, the estimated Lorentzians are

    reconvoluted with the calculated instrumental function, the

    constant andShirley noise added;the final resultis compared

    with the experimental spectrum without any further curve-

    fitting. It can be seen in Fig. 11 that the reconvolution isindeed very satisfactory.

    A similar type of very good result is obtained in Figs 12

    and 13. Note that in Fig. 12 the estimated original XPS

    function is represented multiplied by a factor of 0.272. In

    order to appreciate the result, the estimated Lorentzians

    are reconvoluted with the estimated instrumental function,

    the constant and Shirley noise added; the final result is

    compared with the experimental spectrum without any further

    curve-fitting (Fig. 13).

    The National Institute of Standards and Technology(NIST) provide well-characterized spectral data for assessing

    quality in computer-based data analysis procedures.911 The

    NIST data analysed here are simulations of XPS spectra.

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    8/10

    78 C. Charles et al.

    COUN

    TS

    275

    BINDING ENERGY (eV)

    280285290295300305310

    2500

    2000

    1500

    1000

    500

    0

    Figure 9. The XPS spectrum and its estimated original

    XPS function.

    Table 5. Parameters of the Lorentzians of Fig. 9

    estimated via the wavelet method (W) and

    curve-fitting (CF)

    W IntensityW CF IntensityCF

    286.25 2107.84 286.25 2107.84

    287.37 667.20 287.49 492.22

    289.29 352.45 289.28 305.68

    291.05 51.74

    293.93 1542.34 293.91 1531.50

    296.57 783.83 296.41 781.25

    COUNTS

    335

    BINDING ENERGY (eV)

    340345350355

    1300

    1200

    1100

    1000

    900

    800

    700

    600

    500

    400

    300

    Figure 10. The XPS Ca 2p spectrum and its estimated original

    XPS function.

    Because the XPS spectra are simulations, peak parameters

    are known from the models and are not estimates. The

    test data are designed to help the analyst to assess data

    analysis procedures from the errors in the analysts peak

    parameter determinations. Performance is measured bydeviations of the analysts parameter estimates from the true

    values for these parameters. The noise filtering algorithm

    and the deconvolution method are applied to 22 spectra

    COUN

    TS

    335

    BINDING ENERGY (eV)

    340345350355

    1300

    1200

    1100

    1000

    900

    800

    700

    600

    500

    400

    300

    Figure 11. The XPS Ca 2p spectrum and its estimation with

    reconvolution.

    COUNTS

    BINDING ENERGY (eV)

    270275280285290295300

    800

    700

    600

    500

    400

    300

    200

    100

    0

    Figure 12. The XPS C 1s spectrum and its estimated original

    XPS function.

    COUNTS

    BINDING ENERGY (eV)

    270275280285290295300

    800

    700

    600

    500

    400

    300

    200

    100

    0

    Figure 13. The XPS C 1s spectrum and its estimation with

    reconvolution.

    (18 doublets and 4 singlets). Following a factorial design,

    the doublet spectra present varying degrees of overlap

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    9/10

    Deconvolution of XPS data by wavelets 79

    between peaks (valley, shoulder, no shoulder), varying

    levels of relative intensity between peaks (Ii II iI) and

    varying levels of Poisson noise (low noise, high noise).

    The results are encouraging. Figure 14 presents the result

    of deconvolution and the reconvolution without any other

    curve-fitting for these 22 spectra. As the reader can see,

    the program recovers the 18 doublets and the 4 singlets. Theintensities are well estimatedexceptfor thespectra generated

    by two peaks of the same intensity (II). In these cases, the

    algorithm tends to the solution of one large peak and a very

    small peak.

    CONCLUSION

    This article has tried to show how wavelets can be used

    to estimate the original XPS function or the true signal.

    The first part of the method consists of filtering the noisy

    spectrum to eliminate the Poisson noise. It has be seen

    in the second paper of this series that the LTFTIPSH

    algorithm is a wavelet shrinkage technique specific to this

    kind of signal. Subtraction of the constant and variable

    noise is done traditionally. The second step consists of

    deconvoluting the filtered spectrum. It is proposed touse a method similar to the HREELS method presented

    in the second paper of this series. The positions of the

    Lorentzians are estimated via the wavelets. In order to

    deconvolute each peak, the localized least-squares method

    is used with parameters and of the Lorentzians as

    variables. Initialization of this optimization problem is done

    by means of combined wavelets and Fourier transform.

    The results for XPS spectra are satisfactory. The positions

    and the intensity of the peaks are well recovered. This

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    350

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    350

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    50

    100

    150

    200

    250

    300350

    280 285 290 2950

    50

    100

    150

    200

    250

    300350

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    350

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    350

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    350

    280 285 290 2950

    50

    100

    150

    200

    250

    300

    350

    Figure 14. Twenty-two XPS spectra: their estimated original XPS function and their reconvolution.

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180

  • 7/29/2019 Xps Wavelet

    10/10

    80 C. Charles et al.

    280 285 290 295

    280 285 290 295

    280 285 290 295

    280 285 290 295

    280 285 290 295 280 285 290 295 280 285 290 295

    280 285 290 295 280 285 290 295 280 285 290 295

    280 285 290 295

    280 285 290 295 280 285 290 295

    280 285 290 295

    280 285 290 295

    280 285 290 295

    0

    200

    400

    600

    800

    1000

    1200

    0

    200

    400

    600

    800

    1000

    1200

    0

    50

    100

    150

    200

    250

    300

    350

    0

    50

    100

    150

    200

    250

    300

    350

    0

    0

    50

    100

    150

    200

    250

    300

    0

    50

    100

    150

    200

    250

    300

    0

    200

    400

    600

    800

    1000

    1200

    0

    200

    400

    600

    800

    1000

    1200

    200

    400

    600

    800

    1000

    1200

    280 285 290 295 280 285 290 295 280 285 290 295

    200

    0

    400

    600

    800

    1000

    1200

    200

    0

    400600

    800

    1000

    1200

    200

    0

    400600

    800

    1000

    1200

    0

    50

    100

    150

    200

    250

    300

    350

    0

    50

    100

    150

    200

    250

    300

    350

    0

    50

    100

    150

    200

    250

    300

    0

    50

    100

    150

    200

    250

    300

    0

    50

    100

    150

    200

    250

    300

    0

    50

    100

    150

    200

    250

    300

    0

    200

    400

    600

    800

    1000

    1200

    280 285 290 2950

    200

    400

    600

    800

    1000

    1200

    Figure 14. (Continued).

    method possesses the advantage of having a fast algorithm

    (implemented in MatLab from WaveLab Toolbox available

    at http://playfair.stanford.edu/wavelab).In this series of three papers the potential, if not the

    real efficiency, of the wavelet shrinkage technique in the

    field of noise filtering of surface science experimental data

    has been demonstrated. It has also been illustrated that the

    wavelets are well adapted to the detection of singularities.

    However, because the convolution product is preserved

    by the wavelet transform, the use of wavelets is not yet

    the general solution to the deconvolution problem but

    the wavelet transform can be combined with the Fourier

    transform for some improvements, as was seen with the

    XPS data.

    AcknowledgementsC.C. is indebted to the Belgian FNRS for a Research Fellow grant.Part of the research at LISE is financed through the PAI/UIAP(4/10)

    programme of the Belgian Prime Ministers Office (Federal Servicesfor Scientific, Technical and Cultural Affairs).

    REFERENCES

    1. Sherwood PA. In Practical Surface Analysis (2nd edn), Briggs D,Seah M (eds). Wiley: Chichester, 1990; 555.

    2. Kover L. Surf. Interface Anal. 2000; 29: 671.3. Leclerc G, Pireaux JJ. J. Electron Spectrosc. 1995; 71: 141.4. Donoho DL, Johnstone IM. Biometrika 1994; 81: 425.5. Donoho DL, Johnstone IM, Kerkyacharian G, Picard D.J. R. Stat.

    Soc. Ser. B 1995; 57: 301.6. Donoho DL, Johnstone IM, Kerkyacharian G, Picard D. C. R.

    Acad. Sci. Paris (A) 1992; 315: 211.7. Kolaczyk ED. Biometrika 1996; 46: 352.8. Charles C, Rasson J-P. Computational Statistics and Data Analysis

    2003; 43: 139.

    9. http://www.acg.nist.gov/std10. Conny JM, Powell CJ, Currie LA. Surf. Interface Anal. 1998; 26:

    939.11. Conny JM, Powell CJ. Surf. Interface Anal. 2000; 29: 444.

    Copyright 2004 John Wiley & Sons, Ltd. Surf. Interface Anal. 2004; 36: 7180