A comparison between three unmixing models for source apportionment of PM2.5 using alkanes in air from Southern Chile

Abstract

ABSTRACTFine particulate matter in the atmosphere, especially the fraction less than 2.5 µm in diameter (PM2.5), arises from several sources. Assessing the relative contributions from each source may be modeled through a mixing method, where the chemical signatures of known sources are mixed in a variety of proportions to provide the best explanation of the measured data. Alternatively, unmixing models determine what the chemical composition of the end members must have been in order to produce the observations. This study uses three different unmixing models with both a synthetic and a real-life (environmentally measured) alkane dataset from PM2.5 collected in five locations in Chile. Polytopic vector analysis (PVA), positive matrix factorization (PMF), and UNMIX modeling were used with ∼300 samples collected across 18 months. Using the synthetic data, both PVA and PMF were able to satisfactorily reconstruct the initial sources and their contribution to the samples, with PMF marginally more accurate than PVA. UNMIX was unable to complete this task with the synthetic data. With the real-life data, all three models produced numerical solutions that could be ascribed to sources that had similar chemical compositions and might represent diesel fuel, diesel particulate matter from combustion, and terrestrial matter, probably wood. Additionally, PVA and PMF produced a factor that could be ascribed to fuel oil used in domestic heating. Of the three models, PVA was the easiest to use; PMF was robust and readily available from the U.S. EPA but did require significantly more processing time, and UNMIX required considerable manipulation in order to produce solutions that might be related to chemical signatures and contributions.