[Key Drivers of PM2.5 in Qingdao from 2013 to 2023 Using an Interpretable Machine Learning Model].

To elucidate the deceleration trend in PM2.5 concentration reduction and its formation mechanisms， this study analyzed PM2.5 concentration variations in Qingdao from 2013 to 2023， combined with factors such as air pollutants， meteorology， and emissions. A phased interpretable machine learning model was applied to predict the variations of PM2.5， investigate the formation mechanisms， and identify key drivers under evolving environmental policies. The results showed that PM2.5 concentration decreased from （56.3±43.66） μg·m-3 in 2013 to （30.2±24.50） μg·m-3 in 2023 in Qingdao， with a reduction of approximately 46.3%， primarily driven by the reduction in secondary aerosols. Notably， the fastest decline of PM2.5 occurred before 2017 at a rate of -4.33 μg·（m3·a）-1， mainly due to reduced formation of secondary sulfate from the end-of-pipe control in the industrial and power sectors. After 2017， the deceleration in the decline of PM2.5 concentrations stemmed from asynchronous reductions in sulfate， nitrate， and ammonium precursors， leading to enhanced secondary PM2.5 formation. The machine learning model indicated the enhanced sensitivity of PM2.5 to NO2 under low SO2 conditions， with the contribution of NO2 to PM2.5 concentrations increasing by approximately 6%， while that of SO2 became negligible. Additionally， meteorological factor contribution to PM2.5 concentrations increased by 5.1%. PM2.5 concentrations exhibited a seasonal variation in the order of winter > spring > autumn > summer. Influenced by increased primary emissions from the residential sector， winter PM2.5 concentrations showed the slowest decline over the years. Therefore， controlling residential sources can reduce primary PM2.5 emissions， and a coordinated multi-pollutant emission control strategy can effectively reduce secondary PM2.5 formation.