Machine Learning-Augmented 3D-Variational Data Assimilation for Wintertime Diurnal Variability of Surface Ozone Over China

Abstract

Accurate surface ozone (O3) estimates are critical for mitigating air pollution and developing strategies. However, efforts of using data assimilation (DA) to improve O3 simulation have been hindered by limited observations. Here, we develop the first hourly, multispecies machine learning-augmented three-dimensional variational data assimilation system (hereafter ML-DA) for China. An optimized ML algorithm is employed to fuse multisource data, achieving seamless estimates of atmospheric composition. These estimates are incorporated into the WRFDA-Chem system as an augmented observation vector to produce analysis fields for updating the initial conditions. Observation DA improves the correlation coefficient (CORR) to 0.80 and reduces root mean square error (RMSE) by 65.84% compared to the no DA. ML-DA system achieves the highest accuracy (CORR = 0.89), with RMSE reductions of 74.75% and 26.07% compared to no DA and observation DA, respectively. The RMSEs of NO2, PM2.5, PM10, SO2, and CO are reduced by 27.14–41.03%. Incorporating the ML-estimated six pollutants at an hourly DA cycling frequency minimizes analysis errors, yielding consistent simulation improvements. High-resolution dataset reveals that although daytime O3 exceeds nighttime levels in winter, nighttime O3 increased by 79.49%, far exceeding the daytime increase of 49.87% during the lockdown, primarily driven by reduced NOx emissions weakening NO titration and adverse meteorological conditions in southern China. Reduced PM2.5 weakened the heterogeneous uptake and hydrolysis of N2O5, suppressing particulate nitrate formation and promoting nighttime O3 accumulation. The combined use of ML and DA maximize the accuracy of O3 and shows strong potential for extension to multispecies.