A Novel Approach Combining Mechanism and Machine Learning to Unravel Oxidant-Driven Formation Regimes of α-Pinene-Derived Oxygenated Organic Molecules (OOMs)

Abstract. Monoterpene oxidation produces oxygenated organic molecules (OOMs) that act as secondary organic aerosol precursors, but the roles of ozone (O₃) and the hydroxyl radical (OH) in their ambient formation are hard to separate because the two oxidants co-vary. We fitted an ensemble of tree-based regression models to one month of online mass-spectrometric measurements at a coastal background site in eastern China and ranked the predictors of four OOM classes defined by hydrogen content (H14–H20) using Shapley additive explanations. Within the fitted models, ozone accounts for 70.8–79.4 % of the relative predictive importance and OH for 20.6–29.2 %. The balance shifts along the series, the class retaining the hydrogen count of the parent monoterpene depending most strongly on ozone and the most saturated class most strongly on OH. The ozone response is not uniform: below about 62 ppb (62.0–63.6 ppb across the four classes) the models extract little ozone signal, whereas above this level ozone becomes a strong positive predictor for all four classes. Three independent checks support this pattern (Appendix B). Both the weights and the transition level derive from a single campaign and are site-specific; the analysis can be repeated wherever these variables are measured.