High-Time-Resolution Aerosol Chemistry and Machine-Learning Sensitivity Reveal a Highland Triad Mechanism Driving PM2.5 in Xining (Qinghai–Tibet Plateau)

Fine particulate matter (PM2.5) formation mechanisms in fragile highland ecosystems remain inadequately constrained, particularly regarding thermodynamic non-linearities (aerosol pH, liquid water content) and their interaction with geochemical modulation. Here, we present comprehensive year-long online measurements from Xining, Qinghai-Tibet Plateau, integrating hourly measurements of water-soluble ions, inorganic elements, and gaseous precursors with ISORROPIA-II thermodynamic modeling and ensemble machine learning. Median pH was 4.38 but exhibited two distinct pH regimes (14.8% pH < 3.0, 11.5% pH > 7.2), with acute acidification enhancing toxic metal solubility (Fe, Pb by 3-5×), and it posed distinct ecological risks. Our analysis reveals a distinct “highland mechanism triad” governing PM2.5 dynamics: (1) winter meteorological confinement amplifying dust-catalyzed sulfate formation (SOR = 0.68); (2) spring alkaline dust buffering (pH > 7.2) that titrates NH3 and suppresses nitrate formation (NOR < 0.10); and (3) summer photochemical oxidation constrained by chronic NH3 limitation within an oxidant-excess regime. Random Forest achieved optimal prediction for the chemically active inorganic fraction (RMSE = 6.63 μg/m3, R2 = 0.91) by learning regime-specific non-linearities, with local sensitivity analysis identifying Ca2+, SO42−, and Al as chemically sensitive drivers (S > 0.35) while revealing NH3’s seasonally variable influence (rank 15 in winter, significant in summer; S > 0.28), subsequently complemented by global SHAP analysis, which further revealed NO3− as the most robust predictor (ranking 1st–2nd) and captured NH3’s non-linear threshold effects (). Positive Matrix Factorization apportioned secondary aerosols (30.11%) within a unique alkaline–dust matrix. These findings demonstrate that highland PM2.5 inorganic chemistry operates through fundamentally different pathways than lowland photochemical haze, with acid-induced toxic metal activation providing a new target for ecological protection in this fragile ecosystem. Seasonally adaptive mitigation is required: concurrent SO2-NH3 control in winter, dust suppression infrastructure in spring, and agricultural NH3 capture in summer. This integrated framework provides a transferable methodology for air-quality management in alkaline dust-dominated, NH3-limited highland ecosystems (>2000 m).