Calibration approaches for the measurement of aerosol multielemental concentration using spark emission spectroscopy

A multivariate calibration approach, using partial least squares regression, has been developed for the measurement of aerosol elemental concentration. A training set consisting of 25 orthogonal aerosol samples with 9 factors (elements: Cr, Mn, Fe, Ni, Cu, Zn, Cd, Pb, and Ti) and 5 levels (elemental concentrations) was designed. Spectral information was obtained for each aerosol sample using aerosol spark emission spectroscopy (ASES) at a time resolution of 1 minute. Simultaneously, filter samples were collected for the determination of elemental concentration using an inductively coupled plasma mass spectrometry (ICP-MS) analysis. Two regression models, PLS1 and PLS2, were developed to predict mass concentration from spectral measurements. The prediction ability of the models improved substantially when only signature wavelengths were included instead of the entire spectrum. The PLS1 model with 45 selected spectral variables (PLS1-45 model) presented the lowest relative root mean square error of cross validation (RMSECV; 16-35%). The detection limits using the PLS1-45 model for the nine elements were in the range of 0.16-0.50 mg m(-3). The performance of both multivariate and univariate regression models was tested for an unknown sample of welding fume aerosol. The multivariate model did not provide significantly better prediction compared to the univariate model. In spite of the difference in the matrices of the calibration aerosol and the unknown test aerosol, the results from the PLS model show good agreement with those from filter measurements. The relative root mean square error of prediction (RMSEP) obtained from the PLS1-45 model was 13% for Cr, 23% for Fe, 22% for Mn and 12% for Ni. The study shows that in spite of lower spectral resolution and lack of sample preparation, reliable and robust measurements can be obtained using the proposed calibration method based on PLS regression.