AAAR 35th Annual Conference October 17 - October 21, 2016 Oregon Convention Center Portland, Oregon, USA
Abstract View
Fourier-Transform Infrared Determination of Organic and Elemental Carbon in the Chemical Speciation Network by Partial Least-Squares: The Importance of Spectra Pretreament on Factor Interpretation
ANDREW WEAKLEY, Ann Dillner, Satoshi Takahama, University of California, Davis
Abstract Number: 137 Working Group: Carbonaceous Aerosols in the Atmosphere
Abstract Organic and elemental carbon (OC, EC) are major constituents in urban ambient fine aerosols (PM2.5) and quantified in the EPA’s Chemical Speciation Network (CSN). Presently, OC and EC are determined by subjecting PM2.5 collected on quartz substrates to destructive thermal analysis using thermal optimal reflectance (TOR). Recently, it was demonstrated that the quantity of TOR OC and EC is non-destructively determinable using a partial least-squares (PLS) regression with Fourier-transform infrared (FT-IR) spectra. Here, a PLS regression correlates the infrared absorption of PM2.5 on collocated polytetrafluoroethylene (PTFE) filters to their corresponding TOR OC and EC measurements. The optimal prediction of TOR OC involves second derivative transforming spectra to suppress PTFE scattering and amplify the signal-to-background ratio of organic functional groups. Next, the PLS model is trained to use only infrared absorption (indexed on specific wavenumbers) best correlated to the TOR OC measurement. Judiciously selecting wavenumbers for calibration using the backward Monte Carlo unimportant variable elimination (BMCUVE) algorithm resulted in an accurate and parsimonious (few-component) PLS model. Further analysis indicated that a single PLS component (factor) modeled the average extinction of mid-infrared radiation by OC functional groups. Furthermore, the FT-IR OC method matched the accuracy and precision of parallel TOR OC measurements when predicting independent testing samples not used for calibration. Determination of TOR EC by PLS followed a similar procedure, with method repeatability again demonstrated using test samples. We also provide a first demonstration of transferability of calibration models for TOR OC and EC prediction developed from CSN spectra to samples collected in the Federal Reference Method (FRM). Preliminary results indicate that appropriately pretreating FT-IR spectra prior calibration and prediction is vital to the successful, interpretable, repeatable, and potentially transferable prediction of TOR OC and EC by the FT-IR method.