CWAI folds pollutants and weather into one learned state, beating WRF-Chem error at 10 days across six pollutants
Synopsis
Regional air-quality forecasting has relied on online-coupled numerical chemistry-weather models whose skill is sensitive to meteorological initial conditions, emission inventories and lateral boundary conditions. CWAI treats meteorology, particles and reactive gases as one joint atmospheric state, driven by two consecutive states, emissions and static surface descriptors through a hierarchical Swin Transformer, and rolls out 40 six-hour steps to a 10-day forecast. Against WRF-Chem over 62 January and July 2020 starts, CWAI had lower initialization-weighted RMSE for all six surface pollutants, from 15.8% for PM2.5 to 25.8% for NO2.
Interpretation
A regional chemical-weather forecast can now be advanced recursively to 10 days from a single learned joint atmospheric state, with particles, reactive gases and meteorology evolving together in one set of channels. Earlier regional learned systems addressed multi-pollutant or aerosol forecasting over shorter horizons, such as FuXi-Air at 72 hours and Zeeman extending regional atmospheric-chemistry forecasts to 5 days, while a joint 10-day national-scale evaluation including behavior near an unforced lateral boundary had not been established. Across all leads from 6 to 240 hours over 62 January and July 2020 starts, CWAI had lower initialization-weighted RMSE than the consistently configured WRF-Chem for all six surface pollutants, with 94% to 100% of starts favoring it across the six.
The model advances chemical weather as one whole state rather than regressing each pollutant separately. Each prediction step is conditioned on the recent tendency from two consecutive atmospheric states and on emissions and static surface descriptors that carry source and land-surface context. On a 15-km East Asian grid of 300 by 360 cells, 166 channels cover 11 variables at 14 model-height levels from 30 to 7,500 m plus 12 surface variables, and the same frozen model recurses 40 steps to 240 hours.
At independent national surface stations, CWAI's error reduction relative to contemporary CMA-CW is larger, but performance differs by species and season. Sixty starts in July 2025 and January 2026, with more than 1.6 million station-hour records, provide a verification target independent of either gridded forecast system. Station-verified RMSE reductions run from 35.3% for PM10 to 67.8% for CO, with 98% to 100% of starts favoring CWAI per pollutant; on matched Chinese-land grids over the same starts the SO2 interval crosses zero, and winter SO2 degrades by 34.7% at 126 to 228 hours.
Perspective
The evaluation covers six surface pollutants on a 15-km East Asian grid, a WRF-Chem comparison over 62 winter and summer 2020 starts, a CMA-CW comparison over 60 starts in July 2025 and January 2026, and national surface station observations from the same months; the vertical-structure check is against a gridded reference rather than independent vertical observations. The operational record covers 3,220 retained fields from 6 May to 29 July 2026 and describes issue continuity and latency, not prospective forecast skill.
Within 300 km of the boundary at day 10, O3 and NO2 errors rise by 5.2% and 14.9% respectively while CO is inconclusive; extreme-ozone RMSE improvement is only 4.8% and turns to negative 9.1% in January 2026; against observation persistence, intervals for PM2.5, SO2 and CO cross zero. The authors note that the initial states and training targets already contain the influence of their parent numerical system, so the rollout lacks externally supplied time-varying lateral forcing rather than being physically boundary independent.