OIRF-LEnKF v1.0: a novel data assimilation system by integrating incremental machine learning with a localized EnKF for enhanced PM 2.5 chemical component simulation and reanalysis

Assimilating observational data into numerical simulation is crucial for accurately estimating the spatiotemporal distribution of PM2.5 chemical components (NH4+, NO3-, SO42-, OC, and BC), which is beneficial to quantifying the impact of aerosols on the environment, climate change and human health. However, chemical transport model (CTM)-based data assimilation (DA) is computationally inefficient for large ensemble sizes and offers limited improvements in simulation skill, as it solely provides optimal initial conditions. This paper introduces an incrementally updatable machine learning-based data assimilation system (Optimized Incremental Random Forest coupled with Localized Ensemble Kalman Filter, OIRF-LEnKF v1.0) that achieves high efficiency and high quality in generating background and analysis fields for chemical components. Computational efficiency tests indicate that the total time consumed by OIRF-LEnKF v1.0 constitutes only 11.41 %–16.60 % of that of CTM-based DA, primarily because the simulation process requires only 0.13 %–0.20 % of the CTM computation time. Sensitivity tests demonstrate that the incremental learning during the simulation process enhances the percentage change of the Pearson correlation coefficient relative to its minimum value (ΔCORR) by 2.43 %–11.75 % and reduces the percentage change of the RMSE relative to its maximum value (ΔRMSE) by 32.55 %–40.36 %, compared to the stationary training mechanism. A 2-month DA experiment reveals that the RMSE values of chemical components after DA are less than 7.80 and 2.36 µg m−3 during the simulation and analysis processes, respectively, indicating reductions of at least 26.38 % and 68.99 % compared to values without DA. Notably, the RMSE values of our system during the simulation process exhibit a significant reduction of 33.16 %–90.10 % compared to those of the CTM-based DA, highlighting the superior simulation capability of our system. Furthermore, the spatial overestimation and underestimation of chemical components have been significantly mitigated following DA. Compared to multiple reanalysis datasets of inorganic salt aerosols (CORR: 0.56–0.89, RMSE: 2.55–8.52 µg m−3), the dataset generated by OIRF-LEnKF v1.0 (CORR: 0.97, RMSE: 1.12 µg m−3) demonstrates higher data quality.