Advancing Particulate Matter Chemical Composition Analysis: A Hybrid Machine-Learning Approach with UV-Vis Spectroscopy

Fine particulate matter (PM2.5) significantly impacts public health, necessitating detailed chemical composition analysis. Conventional PM2.5 chemical analysis methods, including filter-based sampling and laboratory analysis, are time-consuming and resource-intensive. This study introduces a hybrid machine-learning approach combining Support Vector Machine (SVM) and Partial Least Squares (PLS) regression to estimate PM2.5 chemical compositions using UV-Vis spectroscopy. The hybrid PLS-SVM model demonstrated superior performance across various PM2.5 species, achieving R-2 values exceeding 0.9 with minimal RMSE and MAE values. Specifically, the model accurately estimated concentrations of organic carbon (OC) (R-2 = 0.96, RMSE = 2.37 mu g/cm(2)) and elemental carbon (EC) (R-2 = 0.95, RMSE = 1.27 mu g/cm(2)), outperforming optical color space methods. The model also illustrated superior performance for sulfate and nitrate (R-2 similar to 0.98, RMSE similar to 0.45 mu g/m(3)) and other crustal metals (R-2 exceeding 0.9). The hybrid algorithm also allows us to identify essential spectral bands and their unique optical signatures for each PM2.5 component, enhancing the interpretability of the results. For instance, OC and EC exhibited significant contributions across the UV, visible, and IR spectral regions, with distinct differences in their spectral profiles. Crustal elements and PM2.5 mass showed higher contributions in the UV and visible regions, with unique peaks distinguishing different components. This detailed spectral profiling provides valuable insights into the optical properties and light-absorbing characteristics of PM2.5 species. The hybrid model offers a robust, non-destructive, and cost-effective approach for PM2.5 chemical characterizations, enhancing current PM monitoring techniques and offering significant potential for large-scale air pollution studies and mitigation strategies.