Improving aromatic water-contaminant detection with machine-learning classification and regression for simultaneous Absorbance-Transmission Excitation Emission Matrix (A-TEEM) spectroscopy

Optical detection of aromatic water-contaminants from petroleum or industrial spills is challenging due to background signals from natural and/or man-made components. Further, while target contaminants are regulated at microgram per liter (μg/L) levels, conventional Raman, FTIR and UV-VIS spectroscopy are generally limited to milligram per liter (mg/L) detection ranges. This study reports on patented A-TEEM spectroscopy which primarily uses fluorescence excitation emission matrix data that are corrected for inner-filter effects (IFE) to eliminate spectral distortion. IFE correction improves resolution of low concentration contaminants from higher concentration backgrounds. The multidimensional ATEEM dataset contains spectral information in the UV-VIS range for all chromophoric and fluorescent compounds in the sample matrix. Nevertheless, because the spectra of many compounds overlap or vary in intensity extracting qualitative and quantitative information generally requires multivariate analyses. Importantly, the UV-VIS and EEM data can be analyzed in a 'multi-block' format to leverage the resolution capacity of these simultaneously acquired independent data sets. We evaluated Benzene, Toluene, Ethylbenzene and Xylene (BTEX) as well as naphthalene in filtered (0.45 μm) raw surface water before drinking water treatment. We show that typical methods including Partial Least Squares (PLS) and Parallel Factor Analysis (PARAFAC) exhibit a variety of pitfalls that can confound accurate contaminant detection and quantification. We report that classification and regression using methods including Support Vector Machine (SVM) and especially XGradient Boost (XGB) algorithms can be more effectively validated to rapidly yield lower μg/L detection limits with potential to automate early-warning reporting.