Machine-learning based wind correction in WRF-CMAQ modeling system and its application

Abstract

Accurate representation of surface wind fields is fundamental to realistic air-quality simulations in coupled meteorology–chemistry systems such as WRF–CMAQ. However, WRF systematically overestimates near-surface wind speeds, causing excessive dispersion and persistent underprediction of ozone and particulate matter. To address this limitation, this study developed an XGBoost-based machine-learning wind-correction model and incorporated the corrected 10 m winds into WRF–CMAQ through three pathways: static replacements at the WRF and MCIP stages, and a dynamic assimilation approach in which ML-corrected winds were supplied as pseudo-observations during WRF integration. The ML correction effectively reduced surface-wind overestimation across seasons, improving agreement with observations under diverse meteorological conditions. Yet its influence on the coupled system differed substantially by integration method. The static replacements improved near-surface winds and reduced local bias and RMSE, but their effects were largely confined to the lowest layers and resulted in only minor changes in pollutant fields. In contrast, the dynamic assimilation pathway propagated corrections vertically within the boundary layer while maintaining mass balance, modifying mixing height, heat fluxes, and the overall wind structure in a more coherent manner. These corrected meteorological fields were then used to drive CMAQ simulations. Unlike the static replacements—whose seasonal performance remained close to the baseline—the dynamic assimilation pathway produced meaningful changes in pollutant concentrations. ML-nudging improved primary air pollutants, enhanced O3 in all seasons except summer, and yielded superior PM2.5 performance except in spring, with further gains in high-concentration episodes. Case studies showed that ML-nudging more accurately reproduced the evolution and transport of concentration plumes and mitigated the typical underestimation during pollution events. Overall, the results highlight that the effectiveness of machine-learning wind-bias correction is strongly dependent on where and how it is integrated within the WRF–CMAQ system. Of the three approaches, the dynamic assimilation pathway offers a physically consistent and demonstrably impactful mechanism for embedding data-driven corrections into numerical models, effectively bridging statistical learning and process-based atmospheric modeling while delivering meaningful improvements in air-quality simulations.