Disparities in particulate matter (PM<sub>10</sub>) origins and oxidative potential at a city scale (Grenoble, France) – Part 2: Sources of PM<sub>10</sub> oxidative potential using multiple linear regression analysis and the predictive applicability of multilayer perceptron neural network analysis

Abstract. The oxidative potential (OP) of particulate matter (PM) measures PM capability to potentially cause anti-oxidant imbalance. Due the wide range and complex mixture species in particulates, little is known about pollution sources most strongly contributing OP. A 1-year sampling PM10 (particles with an aerodynamic diameter below 10) was performed over different sites a medium-sized city (Grenoble, France). An enhanced fine-scale apportionment sources, based on chemical composition, using positive matrix factorization (PMF) method reported companion paper (Borlaza et al., 2020). OP assessed as ability generate reactive oxygen (ROS) three acellular assays: dithiothreitol (DTT), ascorbic acid (AA), 2,7-dichlorofluorescein (DCFH) assays. Using multiple linear regression (MLR), contributions identified by PMF were estimated. Conversely, since atmospheric processes are usually non-linear nature, artificial neural network (ANN) techniques, which employ models, could further improve estimates. Hence, multilayer perceptron analysis (MLP), ANN-based model, additionally used model PMF-resolved well. This study presents spatiotemporal variabilities activity influences season-specific site typology specific local features, assay sensitivity. Overall, both MLR MLP effectively captured evolution primary traffic biomass burning strongest drivers Grenoble basin. There also clear redistribution source-specific impacts when instead mass concentration, underlining importance redox for identification toxicity. Finally, generally offered improvements prediction, especially where synergistic and/or antagonistic effects between prominent, supporting value models account dynamics behind affecting PM10.