Enhancing Differentiation of Oxygenated Organic Aerosol: A Machine Learning Approach to Distinguish Local and Transboundary Pollution

Abstract

Accurate source apportionment of particulate matter (PM), especially of organic aerosol (OA), is crucial for targeted mitigation efforts. Positive Matrix Factorization (PMF) is powerful in source attribution of primary OA (POA); however, it often struggles to differentiate sources of oxygenated OA (OOA) due to their similar chemical profiles. In this study, a support vector regression machine learning (ML) model was developed to enhance the OOA source apportionment in Dublin from 2016 to 2023. Rolling PMF analysis identified four POA factors and differentiated OOA into less- and more-oxidized (LO-OOA and MO-OOA), highlighting the significant role of the OOA (47-74% of total OA). The ML model further distinguished locally produced OOA (LO-OOA_local and MO-OOA_local) from transboundary transport OOA and exhibited robust performance across different pollution scenarios. The relative importance analysis revealed that LO-OOA_local was more impacted by fossil fuel emissions like hydrocarbon-like OA (20%) and coal (14%), whereas MO-OOA_local was most influenced by LO-OOA (17%), providing insights into their sources and formation mechanisms. During a mixed pollution episode, the results show that despite the significant contribution of transboundary transport, local heating emissions were more critical sources of OA, with local OA accounting for 68% of total OA and reaching 78% during heating hours. These findings highlight the ongoing need to reduce local emissions to achieve cleaner air in Dublin. The ML model's ability to quantitatively separate local and transboundary OOA offers invaluable insights for future air quality regulations.

Figure 1

(a) Mass profiles of six OA factors identified by rolling-PMF analysis including Peat, Wood, Coal, HOA, LO-OOA, and MO-OOA, averaged over all segmented analysis periods. (b) The average chemical composition of total PM₁ and OA during cold (October to March) and warm months (April to September) and (c) diurnal variations of OA factors in cold months (solid lines) and warm months (dashed lines) from 2016 to 2023 in Dublin.

Figure 2

Scatter plots comparing predictions from the model and rolling-PMF analysis for (a) LO-OOA, (b) MO-OOA, and (c) total OOA during selected local events. Panel (d) shows the scatter plot of OOA_TBT calculated from the model versus measured NO₃ filtered for transboundary transport (NO_3 TBT).

Figure 3

Time series of PM₁ species including OA, SO₄, NO₃, NH₄, Cl, and eBC, and OA factors including Peat, Wood, Coal, HOA, LO-OOA_local, LO-OOA_TBT, MO-OOA_local, and MO-OOA_TBT during selected (a) local- and (c) trans-boundary-dominated pollution episodes. Panel (b) presents the average chemical composition for (a) PM₁ (top) and OA composition from rolling PMF analysis (bottom left) and the model (bottom right). Panel (d) provides the same data for episode (c).

Figure 4

(a) Time series of PM₁ species including OA, SO₄, NO₃, NH₄, Cl, and eBC, and OA factors including Peat, Wood, Coal, HOA, LO-OOA_local, LO-OOA_TBT, MO-OOA_local, and MO-OOA_TBT during selected mixed pollution episode in March 2022. Panel (b) shows the average chemical composition of PM₁ (bottom) and OA from rolling PMF analysis (top left) and model (top right) during this episode.