Summary

  • ECMWF’s experimental AIFS-COMPO produces atmospheric-composition forecasts out to ten days, compared with five days from the operational CAMS global system.
  • Its starting state still comes from CAMS’s observation-assimilating IFS-COMPO. Public charts are available; the underlying AIFS-COMPO data is promised only after further validation.

A ten-day pollution chart looks like a clean break with yesterday’s forecasting machinery. Follow the first data point and the break disappears. ECMWF’s AIFS-COMPO release says the AI model starts from operational CAMS analyses. Those analyses are produced by IFS-COMPO, the physical and chemical system that assimilates near-real-time satellite observations. AIFS-COMPO makes forecasts; it does not yet perform that assimilation itself. The older engine still loads day zero.

That handoff matters to anyone building an alert service. Operational IFS-COMPO issues a five-day global forecast twice daily. The experimental AI system extends the horizon to ten days, jointly predicting weather and atmospheric composition on an approximately 80-kilometre grid at three-hour intervals. ECMWF’s authors’ study describes training on CAMS reanalysis and later operational analyses and forecasts, then testing against independent observations. It finds comparable or better results for several species, including aerosol optical depth and particulate matter. It also reports weaker short-lead and some ozone and total-column results against CAMS analysis. The ten-day ozone-hole case is a result from the 2024 test period, not a licence to treat every pollutant and location as equally well predicted.

There is a second handoff at the customer’s desk. CAMS announced ten-day charts of major pollutants; ECMWF says the underlying AIFS-COMPO data will be released later, after more validation. A forecaster can inspect a picture today. A developer seeking fields to ingest, rerun and audit cannot treat that picture as a dataset. The model also lacks explicit time-varying emissions as forcing inputs, forecasts fewer chemicals and no deposition fluxes, and runs at lower spatial resolution than the operational system. ECMWF is working to add fire emissions. For a smoke-response workflow, that is the input that can turn a fresh fire into a fresh forecast.

The next useful milestone is specific: a dated data release with access terms, alongside evidence that emission inputs and event-level validation work for the hazards users actually monitor. Until then, day ten is a valuable experimental window whose first frame and public delivery still belong to the existing system.