Work

An operational data-driven chemical-weather model for skillful multi-pollutant forecasting

Year: 2026
Type: preprint
Abstract: Numerical chemical-weather models provide guidance, yet forecasts remain challenged by uncertainties in meteorology, emissions, nonlinear chemistry and boundary inflow. Whether learned models can jointly evolve particles, reactive gases and weather for 5–10 days without lateral conditions remains unresolved. Here we present CWAI, an operational data-driven model advancing a 166-variable atmospheric state on a 15-km grid at 6-h intervals for 10 days. Across 62 initializations in 2020, CWAI reduced root-mean-square error for six pollutants by 15.8–25.8% relative to the Weather Research and Forecasting model coupled with Chemistry (WRF-Chem). Improvements persisted to day 10 and through PM 2.5 and O 3 profiles from 30 to 3,000 m; particulate and SO 2 advantages also persisted at the domain edge without lateral forcing. Across 60 forecasts in July 2025 and January 2026, gridded errors for PM 2.5 , PM 10 , O 3 , NO 2 and CO fell by 16.4–33.7% relative to China Meteorological Administration Chemistry–Weather Coupled Model (CMA-CW), while independent-station errors for all six fell by 35.1–67.8%. Since May 2026, CWAI has operated twice daily. Post-startup inference and output take 1.5 min and the active analysis-to-product chain takes ~ 70 min. CWAI shows that a learned joint atmospheric state can provide rapid operational guidance for multiple pollutants while retaining three-dimensional chemical and meteorological structure at medium range.
Language: English
Open Access status: green
OpenAlex ID: W7226003781
Cites: 26
Cited by:
Related to: 10