Prediction of gas-particle partitioning of polycyclic aromatic hydrocarbons based on M5' model trees

Abstract

During the thermal combustion processes of carbon-enriched organic compounds, emission of poiycyciic aromatic hydrocarbons into ambient air occurs. Previous studies of atmospheric distribution of poiycyciic aromatic hydrocarbons showed low correlation between the experimental values and Junge-Pankow theoretical adsorption model, suggesting that other approaches should be used to describe the partitioning phenomena. The paper evaluates the applicability of multivariate piece-wise-linear M5' model-tree models to the problem of gas-particle partitioning. Experimental values of particle-associatedfraction, obtainedfor 129 ambient air samples collected at 24 background, urban, and industrial sites, were compared to the prediction results obtained using M5' and the Junge-Pankow model. The M5' approach proposed and models learned are able to achieve good correlation (correlation coefficient >0.9) for some low-molecular-weight compounds, when the target is to predict the concentration of gas phase based on the particle-associated phase. When converted to particle-bound fraction values, the results, for selected compounds, are superior to those obtained by Junge-Pankow model by several orders of magnitude, in terms of the prediction error.