Exploring the drivers of PM2.5 under different pollution durations using an advanced modeling framework in an industrial inland city of China.

Characterizing the dynamic evolution of fine particulate matter (PM2.5) across varying pollution persistence is vital for structural air quality management, yet the duration-dependent shifts in driving mechanisms remain poorly understood. This study investigated the complex driving mechanisms of PM2.5 pollution in Linyi, a major industrial and logistics hub in eastern China. We achieved this by integrating multisource data into a positive matrix factorization (PMF)-based machine learning (ML) framework, coupled with SHapley Additive exPlanations (SHAP). Pollution episodes were classified as one-day, two-day, and lasting three or more days. With increasing pollution duration, PM2.5 and its precursors increased stepwise, and the contribution of secondary inorganic aerosols rose from 37.6% on clean days to 47.1% during prolonged episodes. Sulfate formation was driven by humidity-dependent heterogeneous reactions, while nitrate formation was synergistically enhanced under high humidity and elevated atmospheric oxidizing capacity. Six statistically distinct covariance patterns were resolved by PMF and interpreted as potential source categories. SHAP analysis revealed that anthropogenic factors dominated the model predictions across all episodes, with secondary nitrate exhibiting a relative feature importance weight of approximately 38% during polluted periods, reflecting its high predictive sensitivity in explaining PM2.5 fluctuations. Dust inputs became relatively more prominent during two-day episodes, while firework combustion contributed substantially during prolonged episodes. Regional transport patterns shifted with pollution duration: one-day episodes were influenced by southeast and southwest inputs, two-day episodes by northward transport, and prolonged episodes by sustained short-distance transport from the southwest and east, with intermittent long-range contributions. Source region analysis further showed that potential source regions expanded and contribution intensities increased with longer durations. These findings underscore the importance of informing retrospective diagnosis and long-term strategy optimization for pollution episodes of varying durations.