Analysis of laser-induced breakdown spectroscopy spectra: The case for extreme value statistics

Abstract

In most instances, laser-induced breakdown spectroscopy (LIBS) spectra are obtained through analog accumulation of multiple shots in the spectrometer CCD. The average acquired in the CCD at a given wavelength is assumed to be a good representation of the population mean, which in turn is implicitly regarded to be the best estimator for the central value of the distribution of the spectrum at the same wavelength. Multiple analog accumulated spectra are taken and then in turn averaged wavelength-by-wavelength to represent the final spectrum. In this paper, the statistics of single-shot and analog accumulated LIBS spectra of both solids and liquids were examined to evaluate whether the spectrum averaging approach is statistically defensible. At a given wavelength, LIBS spectra are typically drawn from a Frechet extreme value distribution, and hence the mean of an ensemble of LIBS spectra is not necessarily an optimal summary statistic. Under circumstances that are broadly general, the sample mean for LIBS data is statistically inconsistent and the central limit theorem does not apply. This result appears to be due to very high shot-to-shot plasma variability in which a very small number of spectra are high in intensity while the majority are very weak, yielding the extreme value form of the distribution. The extreme value behavior persists when individual shots are analog accumulated. An optimal estimator in a well-defined sense for the spectral average at a given wavelength follows from the maximum likelihood method for the extreme value distribution. Example spectra taken with both an Echelle and a Czerny–Turner spectrometer are processed with this scheme to create smooth, high signal-to-noise summary spectra. Plasma imaging was used in an attempt to visually understand the observed variability and to validate the use of extreme value statistics. The data processing approach presented in this paper is statistically reliable and should be used for accurate comparisons of LIBS spectra instead of arithmetic averaging on either complete or censored data sets.