Assessing capability of Copernicus Atmosphere Monitoring Service to forecast PM2.5 and PM10 hourly concentrations in a European air quality hotspot

The accuracy of Copernicus Atmosphere Monitoring Service (CAMS) European forecasts of PM 2.5 and PM 10 hourly concentrations was assessed against hourly observations collected from low-cost stations during the 2022–2023 heating season in the Padana Plain (Italy). The intercomparison of all 11 air quality models integrated into the CAMS framework returned root mean square error (RMSE) values ranging 20.3–37.5 (PM 2.5 ) and 22.2–37.8 μg/m 3 (PM 10 concentrations), while hourly variation of observations was poorly captured ( r = 0.16–0.41 and 0.25–0.47, respectively). Agreeing with prior research, CAMS models exhibited a marked daily variability in forecasting particulate matter (PM) observations, with the largest discrepancies occurring during the early morning and evening hours. PM 2.5 observations were best predicted by the CHIMERE model, while PM 10 observations by the MINNI model. CAMS Ensemble returned the best r values among all models, while, since all (or the majority of) models over-predicted the observations, it failed to best fit their magnitude, returning mean bias of +8.1 for PM 2.5 and +4.0 μg/m 3 for PM 10 concentrations. This study demonstrated that further efforts are still needed to improve the performance of CAMS models in estimating PM concentrations. However, rather than acting on model final output, e.g. by implementing bias-correction techniques, a more robust strategy could be to act upstream, i.e. by adjusting the settings of the individual CAMS models. The latter could include a more region-specific characterisation of the emission input data to avoid unrealistic overweighting of anthropogenic emissions, increasing the number of surface stations used for PM concentration assimilation, or adjusting PM chemical composition. • CAMS-predicted PM 2.5 and PM 10 1-h concentrations assessed vs. low-cost observations. • CAMS Ensemble returned the best estimate in terms of correlation, but not of bias. • CHIMERE and MINNI were the best models in predicting PM 2.5 and PM 10 observations. • Further efforts still needed to improve CAMS models in predicting PM concentrations. • CAMS-predicted PM 2.5 concentrations were a scaled value of PM 10 concentrations.