Particulate nitrate is a dominant component of winter haze in East Asian megacities, yet its real-world formation mechanisms remain incompletely understood. We integrate high time-resolution aerosol composition measurements, explainable machine learning (ML), and conventional analyses to disentangle key drivers of wintertime nitrate production. While NO₂ availability is the primary control, our results reveal critical but underrepresented processes: (1) persistence of nitrate formation during late-morning relative humidity (RH) decline, sustained by metastable semi-liquid particles with retained liquid water that facilitate continuous gas-to-particle partitioning of photochemically produced HNO₃, and (2) temperature threshold effects, where subfreezing conditions suppress further nitrate formation primarily due to thermodynamic precursor saturation, compounded by potential diffusion limitations in highly viscous or solid phases. Contrary to common assumptions, boundary layer height contributes minimally to peak nitrate events. These findings demonstrate the need for air quality models to incorporate RH trends, aerosol phase transitions, and temperature-dependent reactivity to accurately predict nitrate episodes. The mechanistic framework presented here is transferable to other urban environments affected by secondary inorganic aerosols and offers new leverage points for mitigation strategies.