Answers are grounded in this paper and cite their sources.
Air quality forecasting over China remains challenging because severe pollution episodes involve strong emissions, nonlinear chemistry, and regional transport. This study evaluates the Aurora foundation model for five-day retrospective forecasts of six surface pollutants—PM 2 . 5 , PM 10 , NO 2 , SO 2 , O 3 , and CO—over China from October 2023 to April 2024, with further analysis of an East China winter pollution episode using WRF-Chem simulation. Hourly observations from the China National Environmental Monitoring Centre (CNEMC) and China High Air Pollutants (CHAP) gridded products were used as complementary benchmarks. The forecasts were assessed using mean bias (MB), mean absolute error (MAE), root-mean-square error (RMSE), R 2 , spatial similarity (SS), and error density distributions. Aurora reproduced the major seasonal evolution and broad regional gradients, including winter increases in particulate matter and the springtime rise in O 3 . For PM 2 . 5 and PM 10 , regional RMSEs were 18.57–34.22 and 24.88–55.02 μg m −3 against CNEMC and CHAP, respectively. In North and Central China, SS reached 0.67–0.69 for PM 2 . 5 and 0.58–0.66 for PM 10 , indicating relatively robust representation of large scale particulate structures. Agreement with CHAP was generally stronger than with surface observations, showing that Aurora better captures spatially coherent concentration fields than local variability. However, the model systematically smoothed high concentration episodes, underestimated CO by about 93.66–241.54 μg m −3 across regions, and overestimated SO 2 by about 28.86–40.46 μg m −3 , while O 3 showed strong benchmark dependence. During the 6–12 February 2024 East China episode, Aurora captured the broad buildup to decay evolution and the main north–south particulate gradient, but underestimated abrupt PM maxima and localized hotspots. Relative to WRF-Chem, Aurora had lower RMSE for O 3 (34.4 versus 47.3 μg m −3 ) and NO 2 (20.1 versus 45.4 μg m −3 ), whereas WRF-Chem generated sharper process based variability. These results indicate that Aurora provides an efficient and scalable baseline for multi-pollutant air quality forecasting, but near surface correction, regional fine-tuning, and hybrid WRF-Chem support remain necessary for severe pollution episodes prediction.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: