Improving WRF-Chem PM2.5 predictions by combining data assimilation and deep-learning-based bias correction

In numerical model simulations, data assimilation (DA) on the initial conditions and bias correction (BC) of model outputs have been proven to be promising approaches to improving PM_2.5 (particulate matter with an aerodynamic equivalent diameter of ≤ 2.5 μm) predictions. This study compared the optimization effects of these two methods and developed a new scheme that combines DA and BC simultaneously. Four parallel experiments were conducted during winter 2019: a control experiment directly forecasted by WRF-Chem (experiment name: WRF-Chem); an experiment that assimilated in situ observations based on the GSI (Gridpoint Statistical Interpolation) system (WRF-Chem_DA); an experiment with deep-learning-based BC (WRF-Chem_BC); and an experiment considering the combination of DA on the initial conditions and BC (WRF-Chem_DA_BC). Statistically, the accuracy of PM_2.5 predictions could be optimized by both DA and BC for the first 24-h period, and WRF-Chem_BC performed better than WRF-Chem_DA in the initial field, especially in the period of 10–24 h, while the best performance was achieved by combining BC and DA. Throughout the initial 24-h period, compared with the control experiment, the results of WRF-Chem_DA_BC (WRF-Chem_DA, WRF-Chem_BC) showed an improvement in terms of root-mean-square error, with reduction proportions varying from 38.90 % to 48.86 % (18.88 % to 32.44 %, 30.10 % to 46.08 %). Besides having the best optimization effect over the whole domain, the combined method also performed well in different regions: during the forecasting period of 0–24 h, the RMSEs decreased from 32 % to 62 %, 39 % to 57 %, 28 % to 40 %, and 30 % to 49 % in the Beijing–Tianjin–Hebei, Yangtze River Delta, Central China, and Sichuan Basin urban agglomerations, respectively.