Mistakes Encountered During Automatic Peak Identification of Minorand Trace Constituents in Electron-Excited Energy Dispersive X-RayMicroanalysisAQ1

DALE E. NEWBURY

National Institute of Standards and Technology, Gaithersburg, Maryland

Summary: Automated peak identification in electronbeam-excited X-ray microanalysis with energy dis-persive X-ray spectrometry has been shown to be sub-ject to occasional mistakes even on well-separated, high-intensity peaks arising from major constituents (arbi-trarily defined as a concentration, C, which exceeds amass fraction of 0.1). The peak identification problembecomes even more problematic for constituents presentat minor (0.01rCr0.1) and trace (Co0.01) levels.‘‘Problem elements’’ subject to misidentification as ma-jor constituents are even more vulnerable to mis-identification when present at low concentrations in theminor and trace ranges. Additional misidentificationsattributed to trace elements include minor X-ray familymembers associated with major constituents but notassigned properly, escape and coincidence peaks asso-ciated with major constituents, and false peaks owing tochance groupings of counts in spectra with poorcounting statistics. A strategy for robust identification ofminor and trace elements can be based on application ofautomatic peak identification with careful inspection ofthe results followed by multiple linear least-squares peakfitting with complete peak references to systematicallyremove each identified major element from the spec-trum before attempting to assign remaining peaks tominor and trace constituents. SCANNING 31: 1–11,2009. r 2009 Wiley Periodicals, Inc.

Key words: energy dispersive X-ray spectrometry,automatic peak identification, peak fitting, qualita-tive analysis, scanning electron microscopy, traceelements, X-ray microanalysis

Introduction

Electron-excited energy dispersive X-ray spectro-metry (EDS) performed in the scanning electronmicroscope (SEM) or other electron beam platformshas become established as a core compositionalcharacterization method for a broad range of physi-cal and biological sciences as well as many branchesof engineering and technology (Goldstein et al. 2003).Commercial SEM/EDS instrumentation has maturedwith the development of extensive software resourcesfor computer-controlled microscope operation andcomputer-assisted X-ray analysis. Advanced softwarecontrols EDS X-ray spectrum acquisition and spec-trum analysis, including automatic qualitative ana-lysis (X-ray peak identification) and automaticquantitative analysis that can be standards-based orthat does not require local standardization (‘‘stan-dardless analysis’’). Perusal of recent EDS advertis-ing shows an increasing tendency in software systemsto simplify the operation of this complex in-strumentation by minimizing the need for theoperator’s input to the analysis. Indeed, advertisingclaims are frequently found for ‘‘one button’’ analy-sis, whereby the entire process of operation is en-capsulated in an automatic software routine thatrequires the operator only to select the desired beamlocation or image region where the analysis is to beperformed and initiate the spectrum collection, withthe software doing the rest of the task, directlyleading to a final report of analysis entry for thatspecimen location that identifies the elemental con-stituents present and assigns concentration values.Although this article is written from the perspectiveand experience of the SEM-EDS analytical commu-nity, it has been noted by a colleague of the authorthat these considerations apply equally well to thetechnique of energy dispersive X-ray fluorescence(EDXRF) analysis AQ2(Sieber 2009).

Correctly identifying the elements present in thebeam-excited volume during the qualitative analysis

This will be your only chance to review these proofs. Please note that once your corrected article is posted online,

it is considered legally published, and CANNOT be removed from the Web site for further corrections.

Published online in Wiley InterScience (www.interscience.wiley.com)

DOI 10.1002/sca.20151

Received 1 April 2009; Accepted with revision 2 May 2009

Address for reprints: Dale E. Newbury, National Institute of

Standards and Technology, 100 Bureau Drive, Gaithersburg, MD 20899.

E-mail: dale.newbury@nist.gov

SCANNING VOL. 31, 1–11 (2009)& Wiley Periodicals, Inc.

procedure is obviously the required foundation, ifthe subsequent quantitative analysis is to make anysense. The skeptical analyst will carefully test au-tomated analysis systems, starting with the auto-matic peak identification. Two earlier articlesexamined the occurrence of mistakes that sometimesoccur during automatic peak identification withcommercial software for major constituents (arbi-trarily, a ‘‘major’’ constituent will be defined as anelement present at a mass fraction40.1; see furtherdefinitions below) under ‘‘conventional’’ electron-excited EDS microanalysis conditions (Newbury2005a) and under low-beam-energy microanalysisconditions (Newbury 2007). ‘‘Conventional’’ ana-lysis conditions are generally regarded as involvingselection of the incident beam energy in the range15–30 keV, a choice that provides excitation of atleast one readily measurable characteristic X-raypeak family (i.e. peaks from one atomic shell) for allelements of the periodic table with atomic numbergreater than or equal to 4 (beryllium). Those pre-vious studies considered only the simplest peakidentification problem, namely the identification ofelemental constituents present at sufficiently highconcentration to produce high-intensity character-istic peaks with a high peak-to-background (char-acteristic to continuum) ratio. Moreover, only thosespectrometric situations were considered wherethere were no significant interelement peak overlapsand no peaks below 0.9 keV photon energy, whichare subject to large self-absorption in the specimen.Despite the apparent simplicity of this challenge, thestudy of automatic peak identification under con-ventional analysis conditions found that mis-identifications of major constituent peaks occurredin approximately 3–5% of the cases tested(Newbury 2005a) when specimens contained ele-ments selected throughout the periodic table, butexcluding Be to Ne, which could only be analyzedwith low-photon-energy, K-shell peaks below0.9 keV. Moreover, the peak misidentifications weregenerally found to be systematic and independent ofthe number of counts accumulated. As an example,Br La (note that in this article Siegbahn peak no-tation will be used because it is still the dominantchoice in analytical software) was consistently mis-identified by some software systems as Al Ka in theEDS spectra of a wide range of bromine-containingcompounds such as reagent-grade potassium bro-mide (KBr) (Newbury 2005a).

When the beam excitation conditions were re-stricted to ‘‘low-beam-energy microanalysis,’’ where‘‘low’’ is defined as the incident beam energy of5 keV or less, the problem of peak misidentificationwas found to be further exacerbated (Newbury2007). When the well-defined peaks with photonenergies above 5 keV were excluded from the

spectrum by the choice of the beam energy, furthermistaken peak identifications were encounteredwhen only more poorly separated low-photon-en-ergy peaks were available for consideration.

The occurrence of a mistake in identifying theelement responsible for a peak during the qualitativeanalysis procedure has a catastrophic impact uponthe overall measurement process. Such egregiousmistakes are properly considered as ‘‘blunders,’’which are defined as ‘‘illegitimate measurement er-rors’’ (Bevington and Robinson 1992). That is, theoccurrence of a blunder during the qualitative ana-lysis procedure renders the entire measurement in-valid. Thus, a proper determination of an errorbudget to associate with the concentration assignedto the incorrect element is not possible or evenmeaningful. Clearly, an accurate compositionalmeasurement has no value if it is assigned to thewrong element! Such blunders are analytical‘‘showstoppers.’’ Performing a quantitative analysisthat determines a concentration for an element thatis not actually present in the specimen is obviouslynonsense. Moreover, these peak identification blun-ders were generally found to be systematic and notrandom occurrences. The same peak identificationfailures were observed with a particular softwaresystem in repeated tests of materials when the ele-ment of interest was present in different compounds.For example, when a software system misidentifiedthe Br La peak in KBr as Al Ka or Yb Ma, this peakwas consistently misidentified when different bro-mine-containing compounds were tested.

For this final article of this study, the problem ofpeak identification for minor and trace constituentsis considered. For the purposes of this discussion,the following arbitrary definitions will be used formass fraction (C) ranges:

‘‘major’’ constituent:C40.1 mass fraction (10wt%),‘‘minor’’ constituent: 0.01rCr0.1 mass fraction

(1–10 wt%),‘‘trace’’ constituent: Co0.01 mass fraction (1 wt%).

The impact of misidentification of a minor or atrace element at these concentration levels on theoverall value of an analysis may seem to be lessconsequential compared with misidentifying a majorconstituent. Nevertheless, for certain problems,minor and trace constituents have critical im-portance in controlling the overall behavior of asystem, such as the strength of a material or thecorrosion resistance. Also, minor and trace elementsmay be used as analytical ‘‘fingerprints’’ where thepresence of a particular set of elements is consideredas an indicator of a process or a location so that anaccurate identification procedure is vital if suchfingerprints are to give a robust result.

2 SCANNING VOL. 31, 8 (2009)

Experimental Details

Test materials were selected that were either pureelements or stoichiometric compounds or that con-tained known constituents present at minor and tracelevels. Material sources included NIST standard re-ference materials, reference materials and internalresearch materials such as the NIST microanalysisglasses (Table I). In addition, reagent-grade chemicalcompounds obtained from commercial sources wereutilized.

All X-ray spectrometric measurements were per-formed with an incident beam energy selected in therange 10–25 keV so as to provide efficient excitationof the photon energy range up to 12 keV dependingon the particular problem. EDS spectra were re-corded with conventional Si(Li) detectors as well asthe silicon drift detector (SDD). The commercialEDS software systems were calibrated for photonenergy following suggested procedures, e.g. using asreference peaks the pairs Cu L–Cu Ka or Al K–CuKa, or for some systems the Cu Ka peak only withthe zero-strobe peak. The amplifier time constant(variously referred to as ‘‘peaking time,’’ ‘‘shapingtime,’’ etc.) was selected to provide ‘‘best resolu-tion’’ performance, generally 135 eV (full width athalf peak maximum intensity (FWHM)) measuredat Mn Ka or better. The input count rate was se-lected so that system deadtime was generally lessthan 30% to minimize the effects of pulse pileup andpeak distortion. For most examples, test materialswere selected to yield X-ray spectra with well-separated peaks that did not suffer significantinterference from the major constituents of thespecimen.

When peak fitting or quantitative analyticalcalculations were required, the NIST software en-gines Desktop Spectrum AnalyzerAQ3 (DTSA: Fioriet al. 1991) and DTSA-II (Ritchie 2008) wereemployedAQ4 .

Results: Examples of Mistakes Observed in

Automatic Peak Identification of Minor and

Trace Constituents

‘‘Problem Elements’’ at Low Concentrations

Based upon previous experience (Newbury2005a, 2007), Table II contains AQ5an extensive, but notnecessarily complete, list of observed and suspected‘‘problem regions’’ with specific peak interpretationchallenges for photon energies o5 keV. A ‘‘problemregion’’ is a spectral region where two or moreelements produce primary characteristic peaks suf-ficiently close in energy, which both might bepossible solutions to an unknown peak detected byan automatic peak-finding algorithm in this region.In compiling this list, only the principal peak ineach possible family is considered. If the databaseof all possible X-ray peaks is included in con-structing such a table so that a minor memberof an X-ray family may be considered as a possiblesolution in competition with a major memberof a second X-ray family, then the list of possibilitieswill greatly expand. Such a situation has beendiscarded in constructing Table II on theassumption that if an element is represented in aspectral region of interest only by a minorfamily member, then it is reasonable to assume thatthe principal peak for that element will bediscovered first in the energy region where it occurs.As the identification process operates for thatelement, all related minor family membersshould be properly marked and thus removedfrom further consideration in the region of interest.However, it must be noted that mistakes of thistype were encountered in the previous studies wheresuch meticulous bookkeeping of X-ray peak familieswas not rigorously performed (Newbury 2005a,2007).

Many problem elements reported in Table IIhave been directly observed to produce blundersduring automatic peak identification of major con-stituents. It is therefore not surprising that theidentification problem is exacerbated when such anelement is lowered in concentration to the minorand trace constituent levels so that the peak in-tensity and the peak-to-background are reduced.Figure 1 shows an example of the misidentificationof Br L as Al K in an experimental polymermixture (30% polystyrene and 70% polymethyl-methacrylate) treated with decabromyldiphenylether as a fire retardant. Bromine is present in thebulk at a mass fraction of 0.023 and is thus a minorconstituent. A second example is shown in Figure 2,where the Zr L peak is misidentified as Pt M in aglass where the Zr is present at a mass fraction of0.074, also a minor constituent level.

TABLE I Composition of microanalysis Glasses K456 andK493 (mass fraction)

K456 (matrix) K493 (matrix1traces)

B 0.00044O 0.204 0.206Al 0.00106Si 0.134 0.130Ti 0.00192Fe 0.00224Zr 0.00363Ce 0.00554Ta 0.00721Pb 0.661 0.641

D. E. Newbury: Automatic peak identification 3

Misidentification of Minor Family Members

The complex X-ray families produced from theL- and M-shells include one or more peaks thatoccur at relatively low intensity compared with theprincipal peaks of the family, but with sufficientseparation from the principal peaks to be well re-solved with EDS, e.g. Ll, Lg, Mz, Mg and M2N4. If

the automatic peak identification program does notuse a complete database of X-ray peaks that in-cludes these minor family members, then it is pos-sible that these minor peaks will be assigned to adifferent and incorrect element. An example isshown in Figure 3 for a gold–copper alloy (NISTStandard Reference Material 482 Gold–CopperWires for Microanalysis; Au40–Cu60), where theautomatic peak identification procedure has prop-erly identified the Cu Ka, Cu Kb, Cu L, Au La andAu Mab peaks, but the minor family member

Fig 1. EDS spectrum of an experimental polymer mixture(30% polystyrene and 70% polymethylmethacrylate) treatedwith decabromyldiphenyl ether as a fire retardant. (Specimencourtesy of Takashi Kashiwagi, Building and Fire ResearchLaboratory, NIST.). The Br L family peak, where Br ispresent at a mass fraction of 0.023, is misidentified as Al Ka.

TABLE II ‘‘Problem regions’’ for automatic peak identification, photon energies from 200 eV to 5 keV

Below 1 keV0.390–0.395 keV N K (0.392); Sc La (0.395)0.510–0.525 keV O K (0.523); V La (0.511)0.670–0.710 keV F K (0.677); Mn La (0.636); Fe La (0.705)0.845–0.855 keV Ne Ka (0.848); Ni La (0.851)0.900–0.950 keV Cu La (0.928); Pr Ma (0.929)1– 2 keV1.00–1.05 keV Na Ka (1.041); Zn La (1.012)Pm Ma (1.032)1.20–1.30 keV Mg Ka (1.253); As La (1.282)Tb Ma (1.246)1.45–1.55 keV Al Ka (1.487); Br La (1.480)Yb Ma (1.521)1.69–1.80 keV Si Ka (1.740); Rb La (1.694); Sr La (1.806); Ta Ma (1.709)W Ma (1.774)2– 3 keV2.00–2.05 keV P Ka (2.013); Zr La (2.042)Pt Ma (2.048)2.10–2.20 keV Nb La (2.166)Au Ma (2.120); Hg Ma (2.191)2.28–2.35 keV S Ka (2.307); Mo La (2.293)Pb Ma 2.342)2.40–2.45 keV Tc La (2.424)Bi Ma (2.419)2.60–2.70 keV Cl Ka (2.621); Rh La (2.696)2.95–3.00 keV Ar Ka (2.956); Ag La (2.983)Th Ma (2.996)3– 4 keV3.10–3.20 keV Cd La (3.132); U Ma1 (3.170)3.25–3.35 keV K Ka (3.312); In La (3.285)U Mb (3.336)3.60–3.76 keV Ca Ka (3.691); Sb La (3.605); Te La (3.769)4– 5 keV4.05–4.15 keV Sc Ka (4.090); Xe La (4.111)4.45–4.55 keV Ti Ka (4.510); Ba La (4.467)4.84–4.95 keV Ti Kb (4.931); V Ka (4.952); Ce La (4.840); Pr La (5.034)

Fig 2. EDS spectrum of NIST Microanalysis Glass K1404,showing the misidentification of the Zr L peak family, whereZr is present at 0.074 mass fraction, as Pt Ma.

4 SCANNING VOL. 31, 8 (2009)

Au Mz is misidentified as Rb La. Similarmisidentifications of the minor family members ofthe Pt M family are seen for the spectrum of theintermetallic compound LaPt3Ni2 in Figure 4,where Pt Mz is misidentified as Lu La and Pt Mg ismisidentified as Mo La.

Misidentification of Major Peak Artifacts: Escape Peak

and Coincidence Peak(s)

High-intensity peaks in the EDS spectrum areaccompanied by two classes of artifact peaks: siliconescape peaks (of energy equal to the principal peakenergy—1.74 keV, resulting from the loss of an SiKab peak during the photon energy-to-charge con-version process) and coincidence peaks (e.g. A1A,when two A photons enter the active volume of thedetector in a time shorter than the time resolution of

the pulse pileup inspection circuit; if there are twolarge peaks A and B, then A1B will also be ob-served). Unless the automatic peak identificationprocedure includes the appropriate escape peak andcoincidence peak energies in its database for eachelement and considers the possible set of coincidencepeaks, these peaks of low relative intensity are likelyto be misidentified as minor or trace element peaks.

An example of the misidentification of the Siescape peak is shown in Figure 5, where the escapepeak from Ca Ka (3.690�1.7405 1.950 keV) inCaF2 is incorrectly interpreted to be Y La(1.922 keV). The escape peak is a fixed fraction ofthe parent peak, with the fraction decreasing as theenergy of the parent peak increases. For this parti-cular case in which the Ca Ka escape peak is mis-identified as Y La, the equivalent yttrium k-value(peak count rate from the unknown/pure elementpeak count rate) is 0.0049. When this Y k-value isevaluated with a ZAF-matrix correction (atomicnumber–absorption–fluorescence effects) quantifi-cation procedure, an apparent Y mass fraction of0.0090 in the host CaF2 would be determined.

Figure 6 shows the misidentification of a coin-cidence peak from Al Ka (1.48711.4875 2.974 keV)as Ag La (2.984 keV). The occurrence and the re-lative abundance of coincidence peaks depend onthe rate of arrival of photons into the detector. Thisrate can be controlled by operating at low systemdeadtime through the choice of the beam currentand/or the EDS solid angle. For deadtime o10%,coincidence peaks are generally of negligible in-tensity. However, for parent peak energies belowapproximately 2 keV, the charge pulse associatedwith the parent peak photon is close to the noiselimit of the pulse inspection function reducingits efficiency so that a small coincidence peak isstill likely to be present, as shown in Figure 6.

Fig 4. EDS spectrum of LaPt3Ni2 showing the misidentifi-cation of Pt Mz as Lu Ma and Pt Mg as Mo La.

Fig 3. EDS spectrum of NIST Standard Reference Material482 Gold–Copper Wires for Microanalysis Au40–Cu60showing the misidentification of the Au Mz peak as Rb La.

Fig 5. Si(Li) EDS spectrum of CaF2 showing misidentifica-tion of the Ca Ka escape peak as Y La.

D. E. Newbury: Automatic peak identification 5

Moreover, choosing a low deadtime conflicts withthe legitimate need to maximize the number ofcounts that can be obtained in the measured spec-trum within the time constraints imposed on theanalysis and thus improve detectability. Whenminor and trace constituents are of interest, a moretypical strategy would be to work at a higherdeadtime, e.g. 30–40%, to record a much greaternumber of photons per unit of clock time and toaccept the inevitable occurrence of coincidencepeaks by having a protocol in place to detect andrecognize these peaks as artifacts. (See further dis-cussion below.) For the particular case illustrated inFigure 6 involving a deadtime of approximately20%, the misidentification of the Al Ka coincidenceas Ag La would result in a k-value of 0.0054 and,after matrix corrections, an apparent Ag massfraction of 0.0077 in the host Al.

False Peaks

The automatic peak identification process in-volves first applying a ‘‘peak-finding algorithm’’ tothe spectrum. This process is sensitive to noise in theX-ray continuum that forms the spectral back-ground because random groupings of backgroundcounts can mimic a characteristic peak. It is com-mon practice to set a user-defined statisticalthreshold to define what constitutes a significantpeak above this random variation in the back-ground. If this threshold is not invoked or is set toolow (that is, too sensitive to the noise), then theeffect shown in Figure 7 may be encountered. In thisspectrum of high-purity iron with a thin (�7 nm)carbon coating in Figure 7(a), there are four validpeaks: Fe Ka, Fe Kb, Fe La and C K with a max-imum channel count of 290. However, six additionalpeaks for W M, P K, Pd L, Ag L, Ca Ka and La Lat apparent trace levels are also labeled. These false

peaks arise from random variations in the countingstatistics of the background. When this spectrumaccumulation is continued a few seconds longer sothat the maximum channel count rises to 317 inFigure 7(b), the apparent trace element solution issubstantially altered: only La L remains from theoriginal set of six trace elements, but now Zn L, AsL, Al K and Au M are reported. If this spectrumaccumulation is continued for a substantially longertime to a maximum channel count of 1,000, then allof these apparent trace element peaks, W M, P K,Pd L, Ag L, Ca Ka, La L, Zn L, As L, Al K andAu M, will be rejected by the peak identificationalgorithm. However, with this particular automaticpeak identification system, false peaks will be foundin virtually all low count spectra. The analyst shouldbe especially wary of an unstable trace peak solutionthat changes with the accumulation time (seecontinued discussion below).

Fig 7. (a) Si(Li) EDS spectrum of high-purity iron withnumerous trace constituent peaks identified. Note themaximum channel intensity of only 290 counts. (b) Spectrumtaken from the same area after the maximum channelintensity is increased to 317 counts. Note that most of the‘‘trace’’ element solution has changed.

Fig 6. Si(Li) EDS spectrum of high-purity aluminumshowing the Al Ka sum peak misidentified as Ag La.

6 SCANNING VOL. 31, 8 (2009)

Discussion

In the original study of mistakes in the automaticpeak identification of major constituents, it was esti-mated that 3–5% of peak identifications were wrongwhen sample compositions were broadly drawn fromthe elements of the periodic table accessible to SEM/EDS analysis, that is, ZZ4 (beryllium). These mis-taken peak identifications are not random. When aparticular automatic peak identification tool is foundto fail on a given element present as a major con-stituent, it will typically fail consistently on that ele-ment no matter how many counts are accumulated inthe spectrum. At the minor constituent level, the rateof such systematic mistakes increases. However, asillustrated in Figure 7, at the trace level all of the peakidentifications may sometimes be incorrect, but if theanalyst takes care to record a sufficiently high spectralcount, many and perhaps all of the false trace-levelpeaks will eventually disappear from the automaticpeak identification solution.

Considering the range of problems revealed in theexamples above, it is perhaps not surprising thatsome commercial EDS automatic peak identificationsystems are designed to avoid attempting to identifypeaks belonging to legitimate trace constituents, aswell as the lower end of the minor constituent con-centration range. That is, the user-selected statisticalthreshold parameter cannot be made sensitive en-ough to actually force the identification of low-re-lative-intensity peaks that might belong to trace-levelconstituents. Although this conservative approachwill at least avoid the extensive mistakes seen inFigure 7, such a software tool does not aid theanalyst who is actually in need of correct detectionand identification of trace constituent peaks.

As the minor/trace element identification pro-blem is effectively left as an exercise for the analyst,it is useful to have an appropriate analytical strategythat makes use of the spectrum-processing softwaretools typically available in the commercial compu-ter-assisted analysis software.

A suitable procedure to achieve robust minor/trace constituent peak identification consists of twostages: (1) collecting spectra that are statisticallyadequate for trace-level work and (2) systematicallyinterpreting the peaks from major constituents andremoving all spectral features associated with thesemajor constituents before attempting the possibletrace constituents.

Acquisition Conditions for Adequate Spectra

Beam energy: A beam energy E0 5 20 keV shouldbe selected to begin the analysis as this incident

energy will provide adequate overvoltage to excitepeaks from major constituents to a photon energyof at least 12 keV. Such a beam energy provides atleast one family of X-rays for all elements with ZZ4(beryllium), and increases the likelihood that morethan one family of X-rays will be excited for ele-ments with ZZ22 (titanium). However, it must berecognized that a second spectrum of the same re-gion of the specimen should be measured with a lowbeam energy, E0 5 5 keV, to ensure that low-pho-ton-energy elements in the range from 0.1 to 1 keVare not lost owing to specimen self-absorption. Thispossibility is illustrated in Figure 8(a) for siliconcarbide where at E0 5 20 keV the C K peak can beeasily overlooked owing to the strong absorptionresulting from the Si LIII absorption edge, which liesjust below the C K peak energy, despite the presenceof carbon in SiC at 0.50 atomic fraction5 0.30 mass

Fig 8. (a) (Top) Si(Li) EDS spectrum of SiC, E0 5 20 keV.Note that with the Si K peak channel determining the verticalexpansion, the C K peak is not visible. (b) (Bottom) Si(Li)EDS spectrum of SiC at E0 5 5 keV. The C K peak is nowrecognizable with the same vertical scale that contains the fullSi Ka peak.

D. E. Newbury: Automatic peak identification 7

fraction. At E0 5 5 keV in Figure 8(b), the C K peakis much more evident owing to the substantial re-duction in the primary electron range, which resultsin lower self-absorption by the specimen.

EDS resolution: To achieve the best possiblelimits of detection, choose an amplifier time con-stant that gives the best resolution. (Note: Theanalyst should be aware of the resolution vs. timeconstant performance of the EDS.) Resolution im-proves as the time constant is increased, but thethroughput decreases. The analyst should not au-tomatically choose the longest amplifier time con-stant available. The resolution may not significantlydeteriorate with a choice of time constant that is afactor of two shorter, but an improvement in theoutput count rate of a factor of two will be ob-tained.

Deadtime: Choose dose conditions (beam currentand EDS solid angle) to maintain a deadtime of notgreater than approximately 30%, but recognize thatthe most highly excited peaks are likely to producecoincidence peaks. If the analyst is careful to ac-count for these coincidence artifact peaks, then theincrease in counts obtained by operating at a higherdeadtime is worth the possible coincidence penalty.

Count time: Virtually every aspect of X-raymicroanalysis improves as the number of counts inthe spectrum increases, and robust identification ofminor and trace peaks demands high numbers ofcounts. The number of counts P needed to detectthe major peak of a particular element above thebremsstrahlung background can be estimated as

P43B1=2 ð1Þwhere B is the number of background counts in thepeak window (typically selected as 1.5 FWHM).Consider a particular example of detecting traceiron in a silicon–lead–oxygen glass. NIST Micro-analysis Glass K456 provides this matrix, whereasK493 provides the same matrix with several addi-tional elements (Al, Ti, Ce, Fe) present at trace le-vels, as listed in Table I. Together, these materialsenable examination of the spectrum with and with-out trace elements present. Table III contains the

total spectrum integral (0.1–20 keV, including allpeaks) for a silicon–lead–oxygen glass, the numberof background counts in the window for Fe Ka(6.284–6.516 keV), and the calculated concentrationlimit of detection for Fe from Equation (1), basedupon the measured Fe Ka intensity from a pureelement standard. To detect Fe at a mass fraction of0.002, Equation (1) predicts that a total spectrumcount of approximately 290,000 counts is needed.Figure 9(a) shows the spectrum of K493 containingthis total integral, but the trace Fe peak is difficultto detect by visual inspection. Increasing the dose toyield total spectrum integrals of 700,000 and1,450,000 counts, illustrated in Figures 9(b) and (c),suggests that for detection of trace peaks by visualinspection, the statistical criterion given byEquation (1) must be exceeded by at least a factor oftwo, and preferably five AQ6.

Locating and Interpreting the Minor and Trace Peaks

With adequate counting statistics achieved, theanalyst should make use of the automatic peakidentification solution of the major peaks to providea starting point. Each peak identification should beconsidered only as a suggestion, and the analystshould follow a systematic approach to confirm thepeak assignment:

(a) Choose a labeled high-intensity peak withphoton energy 45 keV.

(b) For a Ka identification, confirm the presenceof the corresponding Kb peak. Then, locate thecorresponding low-photon-energy L-shellpeak(s).

(c) Similarly, for a high-photon-energy La identifi-cation, confirm the presence of the full suite of Lfamily lines: Lb, Lg, Ll, etc. Then, locate thelow-photon-energy M-shell peak(s). After thetentative peak identification is confirmed or analternative is chosen, check for the escapepeak and sum peak(s) (which may involve

TABLE III NIST microanalysis Glass K456 (Si–Pb–O); E0 5 20 keV

Dose (nA s)Total integral(0.1–20 keV)

Integral Fe Ka window(6.284–6.516 keV)

Limit of detection(mass fraction)

10 145,127 1,160 0.0028020 290,373 2,329 0.0019350 728,155 5,820 0.00122100 1,454,542 11,613 0.000862200 2,907,573 23,633 0.000615500 7,269,268 58,705 0.0003881,000 14,534,623 118,222 0.0002752,000 29,065,175 235,441 0.000194

8 SCANNING VOL. 31, 8 (2009)

Fig 9. (a) SDD EDS spectrum of NIST Microanalysis Glass K493 with a total spectrum integral (0.1–20 keV) of 290,000 counts.(b) Dose increased to 720,000 counts. (c) Dose increased to 1,450,000 counts. (d) Dose increased to 14,500,000 counts.

D. E. Newbury: Automatic peak identification 9

combinations of different high-intensity peaks)associated with the highest-intensity member ofthe family.

(d) Strip all peaks from each family associated withthis element from the spectrum with the multi-ple linear least-squares (MLLS) peak fitting toolcommonly available in computer-aided analysissoftware. Be sure that the MLLS procedureuses a full peak reference that contains allfamily members, preferably from a spectrummeasured on the EDS system being used for theanalysis rather than a remote library spectrum(check by stripping peaks from a pure elementstandard to be sure all minor family membersare properly accounted for).

(e) Proceed down in photon energy to the low-energy region of the spectrum, recognizing thatthe decreased separation of the family membersand the relatively poor resolution of EDS meanthat eventually only one peak will be availablefor identification. The analyst must make useof every feature available, such as the asym-metry of the L and M family peaks owing tothe relative heights and separations of theLa–Lb and Ma–Mb peak pairs. Again thequality of the fit when the MLLS peakstripping is applied is the measure by whichthe automatic peak ID solution is to be judged,such as the Al K–Br L–Yb M recognitionproblem.

(f ) This MLLS stripping process has been appliedin Figure 10(a) to the spectrum of K493 (thicktrace) showing the residuals (fine trace) afterMLLS fitting and subtraction of the Si Ka,bPb M and Pb L families. Note the series ofpeaks that arise from coincidence of the O K,Si K and Pb M, which are found in both thematrix and the matrix1trace elements spectra.After all the peaks associated with the majorconstituents have been identified and strippedfrom the spectrum, the peaks remaining arelikely to belong to legitimate minor and traceelements. Again starting at the high-photon-energy end of the spectrum, use the KLMmarkers to select a candidate element for apeak. As can be seen in Figure 10, even withextremely large numbers of counts (14,500,000in this case), only a single K-shell peak is likelyto be detected when identifying trace elementsfor photon energies o8 keV. Above 3 keV, twoL-shell peaks (La and Lb1 in a ratio of 1 to0.25 to a ratio of 1 to 0.45 depending on theelement) may be found because of theirincreasing separation and relatively similarweights. Figure 10(b) shows the residuals afterMLLS stripping of both K456 (matrix) andK493 (matrix plus trace elements). Note the

appearance of the coincidence peaks in bothspectra. The trace elements Ti, Ce, Fe and Tacan be identified without peak stripping.However, identifying the Al peak requires peakstripping to be sure that all minor familymembers of Pb have been removed, whereasthe Zr L can only be revealed by stripping offthe major interference from Pb M.

From the sequence of spectra in Figure 9, theissue of recording adequate spectrum counts beforeattempting to perform trace element analysis is seento be of critical importance. Obtaining a totalspectrum integral of 200,000 counts with a conven-tional Si(Li) EDS, spectrometer technology that stilldominates the microanalysis field, operating at or atleast near optimum resolution with modest dead-time represents a considerable task. With athroughput of approximately 1 kHz to constrain thedeadtime, a live time of 200 s would be required, andif the five-fold criterion is applied, the time required

Fig 10. (a) SDD EDS spectrum of NIST MicroanalysisGlass K493 (thick trace) with a total spectrum integral(0.1–20 keV) of 14,500,000 counts and after MLLS peakstripping of the Si K, Pb M and Pb L families (thin trace).(b) Overlay of K456 (dot spectrum) and K493 (solid line)after MLLS fitting.

10 SCANNING VOL. 31, 8 (2009)

expands to 1,500 s to obtain 1,500,000 counts. Jud-ging by the appearance of EDS spectra reported inthe literature, many analysts seem to collect spectrafor far less time. The emergence of the SDD EDS,which for a given resolution has an output countrate improved by a factor of 10–100 compared withthe Si(Li), represents a great improvement for traceelement analysis because the required spectrum in-tegral can be obtained in time periods more con-sistent with the patience of analysts (Newbury2005b).

Finally, when applying EDS to the identificationof trace elements in the mass fraction range0.001–0.01, there is inevitably a lack of redundancyof information compared with the situation that isroutinely obtained during the identification of majorconstituents (i.e. finding multiple members within afamily and often two families; having peak shape asa guide for the single peak regime below 3 keVphoton energy), which means that the confidencewith which a trace element can be reported is in-evitably going to be lower. When it is critical toreport the presence of a trace element with the bestpossible confidence, these inevitable limitations im-posed upon EDS because of the poor resolutionmay force the analyst to make use of wavelengthdispersive X-ray spectrometry (WDS), which isgenerally capable of spectral resolution that is betterby a factor of ten or more (2–15 eV) than the EDSresolution. Such a resolution improvement leads tobetter peak-to-background, more accurate knowl-

edge of the peak channel energy, and detection ofmore family members, all of which contribute to amuch improved degree of confidence in trace ele-ment identification.

References

Bevington PR, Robinson DK: Data Reduction and ErrorAnalysis for the Physical Sciences 2nd ed., McGraw-Hill,New York (1992).

Fiori C, Swyt C, Myklebust R: Desktop Spectrum Analyzer(DTSA), a Comprehensive Software Engine for Electron-Excited X-Ray Spectrometry, National Institute ofStandards and Technology, Standard Reference DataProgram, Gaithersburg, MD (1991). Available from thefollowing World Wide Web site: www.cstl.nist.gov/div837/Division/outputs/software.htm. AQ7

Goldstein J, Newbury D, Joy D, Lyman C, Echlin P, et al.:Scanning Electron Microscopy and X-Ray Micro-analysis 3rd ed., Springer, New York (2003).

Newbury DE: Misidentification of major constituents byautomatic qualitative energy dispersive X-ray micro-analysis: a problem that threatens the credibility of theanalytical community. Microsc Microanal 11, 545–561(2005a).

Newbury DE: X-ray spectrometry and spectrum imagemapping at output count rates above 100 kHz with asilicon drift detector on a scanning electron microscope.Scanning 27, 227–239 (2005b).

Newbury DE: Mistakes encountered during automatic peakidentification in low beam energy X-ray microanalysis.Scanning 29, 137–151 (2007).

Ritchie N: Desktop Spectrum Analyzer—II (2008). Availablefrom the following World Wide Web site: http://www.cstl.nist.gov/div837/837.02/epq/dtsa2/index.html.

D. E. Newbury: Automatic peak identification 11

Author Queries Form

John Wiley

 

Queries and / or remarks

JOURNAL TITLE: SCA 03/06/2009

ARTICLE NO: 09-0006

Query No. Details required Author's response

AQ1Please confirm the suitability of the short title.

 

AQ2Ref. Sieber (2009) is not listed in the reference list. Please include the reference in the list and provide publication details.  

AQ3Please provide city, state code if the place is in USA or city, country if the place is not in USA for the manufacturer of NIST software engines DTSA and DTSA-II.  

AQ4Please confirm if the change of the bibliographic cross-reference Fiori et al., 1992, to Fiori et al., 1991, here as per the reference list is okay.  

AQ5Please confirm if the alignment of Table II is okay.

 

AQ6Please confirm if the term "700,000 counts" should be changed to "720,000 counts" in the sentence "Increasing the dose . . . . . . . and preferably five" as per the caption of Figure 9.  

AQ7Please provide date of access in References Fiori et al. (1991) and Ritchie (2008).