Logo image
Improving dust aerosol simulation over northern China: Synergy of updated numerical models and machine learning post-processing
Journal article   Peer reviewed

Improving dust aerosol simulation over northern China: Synergy of updated numerical models and machine learning post-processing

Tong Sha, Liangqing Li, Zipeng Dong, Qingcai Chen, Shuqi Yan, Huanxin Zhang, Jhoon Kim and Jun Wang
Atmospheric environment (1994), Vol.380, 122098
09/2026
DOI: 10.1016/j.atmosenv.2026.122098

View Online

Abstract

Frequent dust events in China pose significant challenges to the accurate numerical simulation of dust aerosols. Here, we present a novel framework that synergizes an updated physical parameterization in the Unified Inputs for WRF-Chem (UI-WRF-Chem) model with an ensemble machine learning (ML) post-processor, focusing on Shaanxi Province, a typical Northwest region often affected by dust storms. The key improvement involves a dynamically updated erodibility dataset derived from multi-source satellite observations, thereby better capturing the spatiotemporal heterogeneity of dust sources on a monthly scale. The numerical model provides 5-km resolution dust simulations, which are subsequently bias-corrected using an ensemble ML approach combining extreme gradient Boost (XGBoost), random forest (RF), and neural Net-enhanced LightGBM (NetGBM). All ML models are trained using ground-based dust concentrations calibrated against PM10 observations during spring (March-May) from 2018 to 2022, and evaluated with independent data in spring 2023 and PM10 as the assessment indicator. Results show that the ML post-processor provides more accurate PM10 simulations than the numerical model across the spring months, increasing correlation coefficient (R) by 50∼52% and reducing Normalized Standard Deviation (NSD) and Centered Root Mean Square Error (CRMSE) by 58∼78% and 48∼66%, respectively. The post-processing performance is region-dependent, substantial in northern Shaanxi but limited in the south due to lower dust contribution and uncalibrated non-dust components. Additionally, the ensemble ML model mitigates overestimation of dust-related PM10 by the UI-WRF-Chem model during dust episodes, and better captures dust transport processes. This study demonstrates that a robust framework for synergizing numerical modeling with ML approach can substantially enhance dust simulation accuracy in regions influenced by complex dust dynamics.

Details

Metrics

1 Record Views
Logo image