Paper No. 14
Presentation Time: 11:25 AM
APPROACHES TO CALIBRATION OF QUANTITATIVE ELEMENTAL ANALYSIS WITH LASER-INDUCED BREAKDOWN SPECTROSCOPY (LIBS)
DYAR, M. Darby1, CARMOSINO, Marco L.
2, SPEICHER, Elly A.
2, OZANNE, Marie V.
3, CLEGG, Sam M.
4 and WIENS, Roger C.
4, (1)Dept. of Astronomy, Mount Holyoke College, South Hadley, MA 01075, (2)Dept. of Astronomy, Mount Holyoke College, 50 College St, South Hadley, MA 01075, (3)Dept. of Chemistry, Mount Holyoke College, 50 College St, South Hadley, MA 01075, (4)Los Alamos National Laboratory, P.O. Box 1663, MS J565, Los Alamos, NM 87545, mdyar@mtholyoke.edu
Laser-induced breakdown spectroscopy (LIBS) is well-suited to both bench- and microscope-based laboratory studies as well as field work using backpack instruments. However, geological applications that require
quantitative elemental analyses have been limited by the magnitude of matrix effects, for which there are as yet no theoretical treatment. The classical approach to LIBS calibration that relates emission peak intensity/area to concentration produces only qualitative results. However, multivariate statistical techniques provide powerful strategies for quantitative analyses, analogous to the empirical matrix corrections used by Bence and Albee for the electron microprobe. Partial least-squares (PLS) techniques have been successful in geological studies where only small data sets of like samples are employed. For example, root mean square errors on a suite of sedimentary rock compositions compared with their average compositions in wt.% oxide are 51.89±1.88 for SiO
2, 13.54±0.57 for Al
2O
3, 1.10±0.37 for TiO
2, 7.68±1.21 for Fe
2O
3T, 5.10±1.00 for MgO, 0.24±0.07 for MnO, 3.87±0.64 for CaO, 2.07±0.46 for K
2O, 4.68±0.71 for Na
2O, and 0.20±0.06 for P
2O
5.
Wider application of PLS to LIBS of unknowns in field settings is limited by its dependence on matching the training set (used to calibrate the prediction model) to the unknowns. To this end, acquisition of a publicly-accessible calibration database of >2000 samples spanning a diverse range of rock types and compositions is in progress. Moreover, PLS uses all the channels of each spectrum, some of which contribute noise and error to the results. So we are pioneering the application of high-dimensional shrunken regression techniques that provide alternatives to PLS, combining the ability to shrink the number of input variables with automatic selection of similar training samples. These methods have the potential to reduce errors far beyond those for PLS and build more generalized models. For example, the Lasso method provides a sparser, more robust prediction model as an alternative to PLS with comparable errors.
