Revealing the sources of water-soluble PM_2.5 oxidative potential with explainable machine learning

A critical methodological gap persists in quantifying the individual contributions of organic compounds and transition metals to PM_2.5 oxidative potential (OP), as their redox coupling fundamentally undermine conventional regression approaches. This study developed an Extreme Gradient Boosting coupled with SHapley Additive exPlanations (XGBoost-SHAP) model, and investigated the key redox-active components and sources of the OP of water-soluble PM_2.5 in Beijing, based on dual-assay OP data including both DTT consumption and ·OH production. The dataset included 20 chemical features, 5 meteorological features, and 9 water-soluble organic compounds (WSOC) source features resolved by Positive Matrix Factorization (PMF). Among the analyzed chemical species, moderately hydrophilic WSOC emerged as the most influential factor for both DTT activity (DTT_v) and ·OH radical generation (OH_v). Moreover, while both primary and secondary WSOC sources contributed significantly to DTT_v, with respective contributions of 44.9 % and 55.1 %, secondary WSOC dominated OH_v, accounting for 84.6 % of its activity. Besides, simpler WSOC source factors demonstrated enhanced performance in summer simulations compared to the more comprehensive dataset including multiple chemical and meteorological features. By establishing a robust and reliable nonlinear regression approach, this study provides valuable insights into understanding the sources of water-soluble PM_2.5 oxidative potential and informs targeted strategies for mitigating the associated health risks.