Characteristics of the long-term variations in sulfate–nitrate–ammonium aerosols in a coastal city in Northern China—Interpretability analysis from a machine learning perspective

Despite emission reduction efforts, sulfate, nitrate, and ammonium (SNA) concentrations have not responded to precursor reduction as expected. In this study, the key drivers of long-term variations versus short-term pollution events were quantified by using a random forest (RF) model and the Shapley additive explanations (SHAP) method on the basis of atmospheric SNA concentrations in Qingdao from 2004 to 2023. The results validated the prediction accuracy of the RF + SHAP model for SNA trends. The rapid decrease in SO₄²⁻ was driven by SO₂ reductions, especially after 2013, at a rate of 8.0 %·yr⁻¹. In contrast, the NO₃⁻ concentration slightly increased overall before 2017 and then slowly decreased. The SHAP analysis results revealed that the increase in atmospheric oxidation capacity enhanced the secondary formation of NO₃⁻, thus offsetting the benefits of NOx emission reduction and marking a shift from SO₄²⁻-driven pollution to NO₃⁻-driven pollution. Corresponding to this shift, the NH₄⁺ concentration initially decreased before 2016, followed by a slight increase. Notably, during SNA pollution events, the key drivers shifted. Enhanced heterogeneous reaction processes, especially NH₃-driven gas–particle conversion, led to increased SNA formation. Our findings highlighted the changes in SNA formation mechanisms at different time scales, thereby advancing the understanding of the underlying atmospheric processes.