Automatic Identification of Emission Lines in Laser-Induced Plasma by

Letter pubs.acs.org/ac

Automatic Identification of Emission Lines in Laser-Induced Plasma by Correlation of Model and Experimental Spectra Timur A. Labutin,* Sergey M. Zaytsev, and Andrey M. Popov Lomonosov Moscow State University, Chemistry Department, Leninskie gory, 1 bld.3, Moscow, Russia, 119991 ABSTRACT: We have applied an algorithm to automatically identify emission lines in laser-induced breakdown spectrometry (LIBS). A Q-switched Nd:YAG laser at 355 nm was used to ablate a high-alloy stainless steel sample. The algorithm was implemented by three parts: simulation of the set of spectra corresponding to different temperature (T) and electron density (Ne), searching the best correlated pair of a model spectrum and an experimental one, and attributing the peaks with certain lines. In order to construct the model spectra, we used the parameters of atomic and ionic lines, levels, the mechanisms of the broadening of spectral lines, and the selected parameters of the spectrograph. The highest correlation coefficient between the model and the experimental spectrum was 0.943 for T = 0.675 eV and lg(Ne) = 16.7 cm−3. More than 40 emission lines were labeled automatically in the spectral region 393.34−413.04 nm.

L

So for CF-LIBS quantitative analysis, the spectral lines must be checked and assigned to a probable emitting element of the sample. The lack of well-established procedure of automatic assignment of lines in LIBS spectra is, therefore, one of the shortcomings of the CF-LIBS approach.12 Several attempts were made to simplify the identification of the lines in the spectra. Mateo et al.13 developed the software package SALIPS for the semiautomatic identification. The package implemented a module for the construction of model spectra based on the relative intensities of atomic lines from the NIST database.14 A broadening of the emission lines and their relative intensities were adjusted manually. Identification consisted in visual comparison of the model spectrum and the experimental one on the same plot. This procedure, however, requires the initial manual recognition of the most intense lines of the components and represents only a tool for the further visual identification of the lines. Ukwatta et al.15 suggested the use of the special metric for identification by assessing the similarity between the array of binary data (BLOB), which essentially represented the area of digital photography, and the image of a theoretical spectrum obtained with the use of NIST database. It was possible to identify pure substances, but for multicomponent samples it was not suitable. For spectrum identification, Amato et al.16 adopted the ranking algorithm for searching the text, who considered an element as a “document”, the peaks as “words”, and the sample as the “request”. The authors introduced a new characteristic, named “weight”, which was directly proportional to the peak intensity and inversely proportional to the number of peaks in the immediate vicinity of the peak. The “weighting” approach

aser plasma in laser-induced breakdown spectrometry (LIBS) is produced on the surface or in the inner volume of the sample and is also used as an atomizer and the excitation source. LIBS is one of the most promising and powerful techniques for the direct spectrochemical analysis of the objects of different origin. The variety of analyzed objects is extremely wide, from the solids, such as alloys and ores, to gas mixtures and fluid inclusions in biological and geological samples.1−3 The significant advances in laser source and detector technology achieved over the last years have opened new fields and industrial applications of LIBS. Considerable research effort demonstrated the potential of LIBS in a broad range of routine and specialized applications, which triggered the development of commercial LIBS technology.4,5 Noll presented state-of-the-art industrial applications of LIBS systems in his comprehensive book.6 In view of the significant potential of LIBS in many different fields, one must consider the use of LIBS instrumentation by nonexperienced users.7 The last means a complete automation of LIBS analysis, which is currently limited because of assignment of LIBS spectra. A reliable and rapid identification requires a great experience, knowledge, and qualification.1,8 It is also crucial for the calibration free-LIBS (CF-LIBS), proposed by Cuicci et al.,9 since one of the most important step is to construct a Boltzmann plot requiring a large set of identified lines. The coincidence of the temperature, calculated with the Boltzmann plot for individual components of the sample, commonly proves the existence of the local thermodynamic equilibrium (LTE) needed to implement CF-LIBS. The solution of the inverse task in modeling plasma processes by Gornushkin et al.10 also required the prior identification of the emission lines. In the work,11 identification for the procedure of concentrations determination was not necessary but it was needed for the selection of the lines of light elements (H to Al). © 2013 American Chemical Society

Received: November 9, 2012 Accepted: January 23, 2013 Published: January 23, 2013 1985

dx.doi.org/10.1021/ac303270q | Anal. Chem. 2013, 85, 1985−1990

Analytical Chemistry

Letter

mathematical expressions but provide the comments below with literature references for convenience. Step 1. Partition functions of neutral, singly-, doubly-, triply-, and 4-fold-ionized species of the major components (Fe, Cr, Ni, Mn, Si) in the steel sample were calculated at the value of T. Step 2. Abundance ratios of neutral and ionized species of each component were calculated by recurrent Saha equations combined with the equation of conservation of mass as described by Gornushkin et al.10 Stepwise ionization potentials were taken from ref 14, but the lowering factors of ionization potential had been neglected for simplicity of the calculation. If the fraction of species were less than 1%, the ionized form was dropped from further calculations. Step 3. Intensity of each spectral line over a wide spectral range (30 nm) was computed in accordance with Boltzmann’s distribution. Under assumed conditions, the calculated intensity can be equated with the full area of the spectral line. Including in the calculations all known lines of the species17 within this spectral range was a feature of the algorithm. Step 4. Stark component of the line width at a certain value of Ne was estimated from the Stark width of resonance lines of Mn I obtained recently by Srećković et al.18 It is equal to wS = 16 ± 3 pm at Ne = 6.6 × 1016 cm−3 and T = 47 000 K. Despite some of the discrepancy between this value of T and the typical plasma temperature (5 000−20 000 K), the value of wS seemed to be an adequate evaluation of the Stark parameter, at least, within the range of order of a magnitude. Because there are no data on Stark widths of other lines within a range from 388 to 418 nm in the literature, it was reasonable to use the value obtained for Mn I due to the similarity of iron, chromium, manganese, and nickel. Doppler and collisional broadening of each line was calculated. In the last case, we assumed that the main collision partner of atoms was atmospheric nitrogen N2 with a collision cross-section estimated by Gornushkin et al.19 as 66 Å2. Natural broadening was negligible with respect to other broadening mechanisms. Step 5. Convolution of Lorentz and Gauss profiles of each line produced a Voigt profile.20 The Lorentz width was the sum of the Stark and pressure widths while the Gauss width was equal to the Doppler width. In this step, the spectrum was a result of summation of the Voigt profiles of each line. Step 6. Convolution of the spectrum obtained in the previous step with a slit function of the spectrograph was performed (27 pm at 400 nm). A distance between two consecutive points in the model spectrum was equal to 1 pm. Step 7. Since the width of pixels is finite, each pixel of CCD plays a role of a light integrator. Therefore, a part of spectrum obtained after step 6 corresponding to a spectral range covering the pixel area was integrated. The spectral range covering the pixel area is equal to 14 pm. Afterward, a discretization of signal between 212 cells allowed the simulation of signal digitizing by ADC of CCD (12 bit). After discretization, the recovering spectrum resulted in a spectral profile with a step of 14 pm. We obtained a number of model spectra by varying T and Ne in the range, typical for nanosecond-laser produced plasma under atmospheric conditions1 since these parameters greatly influence the excitation conditions. As an example of a resulting model spectrum, in Figure 1 the spectral profiles obtained at different values of T and Ne are compared. Our assumption was that the good approximation would lead us to the right line identification. The use of correlation analysis for fitting the model spectrum with the experimental one was successfully implemented in the works.10,11 We have calculated Pearson’s

(principle), used for the automatic identification of spectra, is not considered reasonable enough, because if an element has a lot of other lines in the vicinity of the most intense line, the “weight” of the line is reduced. For example, the technique cannot identify the basic components like manganese or nickel in steel within the spectral range 300−450 nm. Since there is no automatic system for reliable and fast identification of emission lines in laser ablation of multicomponent samples until now, the aim of this work was to develop an algorithm and software for the automatic identification of emission lines in the spectra of laser plasma by comparing the experimental and synthetic spectra.

■

EXPERIMENTAL SECTION In the experimental setup, the radiation of the Nd:YAG laser “LOTIS-Tii” (Belarus) with a maximum pulse energy of 64 mJ at 355 nm and 10 ns pulse duration was focused perpendicular to the sample surface by a lens (16 cm focal length). The spot diameter was ∼150 μm (fluence ∼270 J/cm2). The plasma emission was collected by a two-lens condenser onto the slit of the Czerny-Turner 0.31-m spectrometer “ISA HR 320” (slit width 25 μm, reciprocal linear dispersion 1.08 nm/mm at 400 nm) with a decrease of an image of 2:1. The center of the plasma plume, the lenses, and the slit were aligned coaxially. The spectrometer was equipped with the ICCD camera “Nanogate-2 V” (“Nanoscan”, Russia) used as a detector. The main parameters of ICCD camera are a diameter of intensifier channel of 15.5 μm, an optical magnification of image transfer from an intensifier to CCD matrix of 2:1, and the size of the CCD pixel of 6.5 μm. We developed the specific software in the LabVIEW environment for camera control and preprocessing of experimental spectra. The latter consisted of three parts: (i) removing instrumental background of CCD; (ii) wavelength calibration with the use of known spectrum of a set of calibrated lamps; (iii) correction of the sensitivity photocathode with the use of the sensitivity curve. To verify the identification procedure, the standard sample of the high-alloy stainless steel (BAM, Germany) C2 was used as a test sample. The elemental composition of the sample is given in Table 1. Temporal parameters for spectrum recording were a 1500 ns delay and a 2000 ns gate. Table 1. Certified Contents of the Elements of the Used High-Alloy Steel Sample C2 elements

Ni

Mn

Cr

C

Si

Mo

contents, wt %

6.124

0.686

14.727

0.010

0.374

0.014

■

RESULTS AND DISCUSSION The simulation of the spectrum of laser-induced plasma was performed under the following conditions: (i) plasma should fulfill LTE conditions; (ii) plasma is homogeneous and optically thin; (iii) plasma composition is the same as the one of the sample. Input data used in the calculations is the content of the major components, plasma temperature T and its electron density Ne, spectroscopic parameters of transitions and levels (Einstein’s coefficients, excitation potentials, energy, and degeneracy) available in the NIST database14 and Kurucz’s database,17 as well as the above-mentioned parameters of the used spectrograph and detector. We assumed that the ablated mass was equal to 10 ng. A step-by-step algorithm of spectrum simulation is described below. We deliberately omit any 1986

dx.doi.org/10.1021/ac303270q | Anal. Chem. 2013, 85, 1985−1990

Analytical Chemistry

Letter

small discrepancy seems to be due to two reasons. First, they have used the fundamental harmonics of the Nd:YAG laser operated at a higher energy 160 mJ/pulse, which usually produced hotter and denser laser plasma. Second, they evaporated the steel sample into argon flow under atmospheric pressure. Aguilera and Aragón22 have compared plasma parameters for ablation in argon and air. The use of argon as the surrounding environment allowed the excitation of the laser plasma up to higher parameters, at least 1.5−2 times greater than when using air. Resuming these two reasons, the values of T and Ne obtained in the present work seemed to be considered as an adequate estimation of real values. It means that the algorithm provided for some informative values can be fruitful for rapid diagnostics of plasma. The procedure in C language performs the automatic identification of spectral lines. The best correlated pair of spectra was an input data of this procedure. Besides this pair, the nonprofiled model spectrum was additionally loaded into the program. The latter data presented a two column table (wavelength and intensity), and each row corresponded to an individual spectral line. Line intensities in this table are the result of the calculation in step 3 of the simulation algorithm (see above) at the values of T and Ne fulfilling the conditions for the best correlation between the model spectrum and the experimental one. There were 3111 lines of neutral and singly ionized forms of elements and their intensities at T = 0.675 eV and Ne = 5 × 1016 cm−3 in a text file describing the nonprofiled model spectrum. In the first step, all intense peaks should be filtered. With the term of “peak” we further understand the local maximum point in the model spectrum. We considered that the peak had a high intensity if the peak intensity Ij satisfied the following condition (Ij − IMIN)/(IMAX − IMIN) ≥ σ, where IMAX and IMIN are the maximum intensity of the spectrum and the minimal one, respectively, and σ is some parameter determined by the user. After exclusion of small peaks, the relevance of each line within the region of the peak should be estimated. Two boundaries, which were established at neighboring local minima (left and right with respect to the maximum), confined the region of each peak (see Figure 3). It is clear that many individual spectral lines within the region can make a contribution to the overall intensity of the peak. Therefore, part of the nonprofiled model spectrum within the region of the peak was considered to determine this contribution. Parameter αi = Ii/Imax, where Ii is the intensity of ith line and Imax is the maximum intensity among all lines within the processed peak, was a measure of this contribution. It should be stressed that αi is not a relative intensity available in the NIST database because it results from calculated intensities at certain values of T and Ne. If αi were larger than a certain value (α) determined by the user, then such an ith line, for example, Cr I 412.65 nm, was ascribed to the peak. The parameter α is intrinsically the relevance level above which we cannot neglect strong lines near the line with maximal intensity. Whether each peak could be matched to one or more emission lines it would depend on α. As a result of the procedure, we obtained a text file with the set of peaks of model spectrum containing the wavelength and intensity of peaks and one or more labels for each of them. As the lines in the model spectrum were assigned, it was possible to identify the experimental spectrum by comparing it with the model spectrum. It should be noted that if a peak in the spectra did not represent local maximum point, it could not be involved in the procedure automatically.

Figure 1. Model spectra obtained under different conditions given in the legend.

correlation coefficient (R) as a measure of mutual correlation between spectra. Figure 2 demonstrates an example of the

Figure 2. Best correlation of the experimental spectrum with the model one obtained at T = 0.675 eV and lg(Ne) = 16.7.

correlation between spectra at a maximal value of R. The coordinates {x; y} of each point on this diagram correspond to a pair of intensities {the measured intensity; calculated intensity} at the same wavelength. It is clear that divergence from the linearity on the correlation plot was observed at high intensities. We believe that one of the main reasons of this fact is the exclusion of the self-absorption of emission lines from the approximation used for spectra simulation, so the expected intensity is greater than the observed one. The correlation coefficients were calculated over 1386 points within the range 393.34−413.04 nm. First, we obtained the set of 105 model spectra for T = 0.5−2.5 eV (step 0.1 eV) and lg(Ne) = 16−18 (step 0.5), where Ne had a dimension of cm−3. The best value of R = 0.91 was reached at T = 0.7 eV and lg(Ne) = 17.0. Then for a more accurate determination of the optimal T and Ne, 99 spectra were calculated for T = 0.6−0.8 eV (step 0.025 eV) and lg(Ne) = 16.5−17.5 (step 0.1). In this case, the best R was 0.943 for T = 0.675 eV and lg(Ne) = 16.7 (Figure 2). The obtained plasma parameters are less than previously determined by Vrenegor et al.21 for high-alloy stainless steels by ∼1.2 times (8 000 K vs 10 000 K and 5 × 1016 cm−3 vs 6 ×1016 cm−3). This 1987

dx.doi.org/10.1021/ac303270q | Anal. Chem. 2013, 85, 1985−1990

Analytical Chemistry

Letter

Figure 3. Procedure of line identification in the best correlated model spectrum (right axis). The seven strongest lines in the nonprofiled spectrum (left axis) within the region of the peak are numbered. Since the intensity of line no.1 is sufficiently larger than others, it belongs to the Cr I line at 412.65 nm in this case.

Table 2. Identified Lines within the Spectral Range of 393.04−413.04 nm no of peak

1

2

3

4

5

6

7

8

9

10

11

wavelength, nm

393.581

394.148

395.116

398.389

399.111

Cr I

Cr I

Cr I

Cr I

Cr I

15

18

396.926 396.974 Fe I Cr I 19

397.666

Fe I

395.645 395.667 Fe I Fe I 17

396.369

element(s)

395.237 395.260 Cr I Fe I 16

20

21

22 403.075 Mn I

no. of peak

12

13

394.810 394.877 Fe I Fe I 14

wavelength, nm

399.739

400.145

400.524

400.971

401.248

401.453

401.715

402.187

402.500

element(s)

Fe I

Cr I

Fe I

Fe I

Cr I

Fe I

Fe I

Fe I

Cr I

Fe I

no. of peak

23

24

25

26

27

28

29

30

31

402.617 402.709 Cr I Cr I 32

wavelength, nm

403.306

403.908

404.136

404.581

405.524

405.876

406.359

407.174

407.663

407.981

element(s)

Mn I

Cr I

Mn I

Fe I

Fe I

Cr I

Fe I

406.693 406.798 Cr I Fe I 39

Fe I

Fe I

Cr I

no. of peak

34

35

36

37

38

wavelength, nm

408.449 408.530

410.749

410.980

411.137

411.854

element(s)

Fe I Fe I

Fe I

Fe I

Cr I

Fe I

412.180 412.182 412.251 Fe I Cr I Fe I

33

40

41

412.338

412.652

Cr I

Cr I

components, the relevance level α was equal to 0.7. This value was chosen so that the weakest revealed peak covered the lines with the expected intensity of at least a 2-fold higher than the noise level. The selected parameters allowed the automatic identification of 39 lines within the region 393.34−413.04 nm for the spectrum of the sample C2 (see Table 2). The best correlated pair of spectra is compared in Figure 4. Two peaks (nos. 25 and 30) were added manually to a set of the identified peaks because we observed them as a shoulder of the spectral profile. Also, the only line marked with an asterisk is visible in the experimental spectrum, and one cannot see it in the model spectrum. The reason for such discrepancy is that we included the certified matrix elements in the modeling algorithm only. Perhaps, the unidentified line belongs to the resonance ionic line of calcium Ca II 396.85 nm, the content of which were not

The value of σ was set that the peaks comparable with the noise level were excluded from the procedure of identification. We have estimated this parameter with the use of the following expression σ = 5s/ΔImax, where ΔImax and s are the range of intensity of experimental spectrum and the standard deviation of the background, respectively. We suggested that noises satisfied the normal distribution of errors. Therefore, a 5-fold value of s provides the more precise removal of small peaks. We equated s to the standard deviation of background within the spectral range 409.084−409.434 nm where there were no strong lines. In the present study, we have established the value of σ at the level of 0.01. The spectral interferences caused by overlapping the strong lines within the region of the peak should be considered in the term of the relevance level. To avoid excessive inclusion of the weak spectral lines of the matrix 1988

dx.doi.org/10.1021/ac303270q | Anal. Chem. 2013, 85, 1985−1990

Analytical Chemistry

Letter

Figure 4. Comparison of the experimental spectrum (left axis) with the best correlated model spectrum (right axis) within the range 393.34−413.04 nm. Identified lines are marked with numbers 1−41. The lines with a red number (shoulder of lines in the model spectrum) were identified manually by the use of the nonprofiled spectrum. The peak marked with an asterisk was not visible in the model spectrum and could not be identified automatically or manually.

and 12-02-31273) and the Russian Ministry of Education (Agreements 8359 and 8773).

certified. Another observation is that the strongest iron lines (nos. 26, 29, and 31) seemed to be self-absorbed. These lines are exactly responsible for the divergence from linearity observed in Figure 2 because of their high intensity.

■

■

(1) Cremers, D. A.; Radziemski, L. J. Handbook of Laser-Induced Breakdown Spectroscopy; John Wiley & Sons, Ltd.: Chichester, U.K., 2006. (2) Miziolek, A. W., Palleschi, V., Schechter, I., Eds.; Laser-Induced Breakdown Spectroscopy (LIBS): Fundamentals and Applications; Cambridge University Press: Cambridge, U.K., 2006; p 620. (3) Singh, J. P., Thakur, S. N., Eds. Laser-Induced Breakdown Spectroscopy; Elsevier Science B.V.: Amsterdam, The Netherlands, 2007. (4) Workman, J., Jr.; Lavine, B.; Chrisman, R.; Koch, M. Anal. Chem. 2011, 83, 4557−4578. (5) Bings, N. H.; Bogaerts, A.; Broekaert, J. A. C. Anal. Chem. 2008, 80, 4317−4347. (6) Noll, R. Laser-Induced Breakdown Spectroscopy: Fundamentals and Applicatons; Springer: Berlin, Germany, 2012. (7) Hahn, D. W.; Omenetto, N. Appl. Spectrosc. 2012, 66, 347−419. (8) Herrera, K.; Tognoni, E.; Omenetto, N.; Gornushkin, I. B.; Smith, B. W.; Winefordner, J. D. J. Anal. At. Spectrom. 2009, 24, 426− 438. (9) Ciucci, A.; Corsi, M.; Palleschi, V.; Rastelli, S.; Salvetti, A.; Tognoni, E. Appl. Spectrosc. 1999, 53, 960−964. (10) Gornushkin, I. B.; Kazakov, A. Y.; Omenetto, N.; Smith, B. W.; Winefordner, J. D. Spectrochim. Acta, Part B 2005, 60, 215−230. (11) Yaroshchyk, P.; Body, D.; Morrison, R. J. S.; Chadwick, B. L. Spectrochim. Acta, Part B 2006, 61, 200−209. (12) Tognoni, E.; Cristoforetti, G.; Legnaioli, S.; Palleschi, V. Spectrochim. Acta, Part B 2010, 65, 1−14. (13) Mateo, M., P.; Nicolas, G.; Pinon, V.; Alvarez, J., C.; Ramil, A.; Yanez, A.. Spectrochim. Acta, Part B 2005, 60, 1202−1210. (14) Kramida, A., Ralchenko, Yu., Reader, J., and NIST ASD Team. NIST Atomic Spectra Database, ver. 5.0; (online), http://physics.nist. gov/asd, 2012. (15) Ukwatta, G.; Samarabandu, J.; Hall, M. Machine Vision Appl. 2012, 23, 111−121. (16) Amato, G.; Cristoforetti, G.; Legnaioli, S.; Lorenzetti, G.; Palleschi, V.; Sorrentino, F.; Tognoni, E. Spectrochim. Acta, Part B 2010, 65, 664−670. (17) Kurucz, R. L.; Bell, B. 1995 Atomic Line Data, Kurucz CD-ROM no. 23.; Smithsonian Astrophysical Observatory: Cambridge, MA,

CONCLUSIONS The algorithm of automatic line identification developed in this study provides a simple, fast, and reliable procedure for the line assignment in emission spectra of a laser plume. The functionality of the algorithm was demonstrated by analyzing one the challenging spectra of the steel sample, showing numerous and very often overlapped lines. A spectral range from 393.34 to 413.04 nm was interrogated as an example and can be changed or enlarged if necessary. In the selected range, the suggested procedure recognized 41 peaks, including both single lines and groups of several lines. Preliminary plasma diagnostics can be performed after two-parameter fitting of simulated spectra to the experimental spectrum through the mutual correlation. The best one (R = 0.943) was achieved at T = 0.675 eV and Ne = 5 × 1016 cm−3, which are in good agreement with experimental values. The main peculiarity of this approach is to construct a model spectrum without experimental parameters of plasma (such as size, shape, etc.) typically used to restore its true profile. Our algorithm has proved itself as a powerful tool that gives some physical insight into the nature of lines in emission spectra, clearly distinguishing a lone line and a group of several overlapped lines.

■

REFERENCES

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS We thank Dr. Anastasia Drozdova for help in preparing the manuscript. We appreciate the financial support from the Russian Foundation for Basic Research (Grants 11-03-01187 1989

dx.doi.org/10.1021/ac303270q | Anal. Chem. 2013, 85, 1985−1990

Analytical Chemistry

Letter

http://www.cfa.harvard.edu/amp/ampdata/kurucz23/sekur.html (accessed November 2, 2012). (18) Srećković, A.; Nikolić, Z.; Bukvić, S.; Djeniže, S. J. Quant. Spectrosc. Radiat. Transfer 2007, 105, 536−541. (19) Gornushkin, I. B.; Anzano, J. M.; King, L. A.; Smith, B. W.; Omenetto, N.; Winefordner, J. D. Spectrochim. Acta, Part B 1999, 54, 491−503. (20) Demtröder, W. Emission and Absorption of Electromagnetic Radiation by Atoms/in Atoms. In Molecules and Photons; Springer: Berlin, Germany, 2010; pp 248−288. (21) Vrenegor, J.; Noll, R.; Sturm, V. Spectrochim. Acta, Part B 2005, 60, 1083−1091. (22) Aguilera, J. A.; Aragón, C. Appl. Phys,. Part A 1999, 69, S475− S478.

1990

dx.doi.org/10.1021/ac303270q | Anal. Chem. 2013, 85, 1985−1990