ECMWF Prototype Makes AI-DOP Forecasts Straight From Observations

Europe’s main weather centre is now producing forecasts from an AI model that has never seen a conventional picture of the atmosphere. The European Centre for Medium-Range Weather Forecasts (ECMWF) said on 3 November that a prototype now runs AI-DOP forecasts in real time, fed directly by live satellite, station, ship, buoy and aircraft observations, and delivered “within minutes of the latest observations arriving”.

That matters because almost every forecast today, physics-based or AI, starts from an analysis: a best estimate of the current atmosphere built by data assimilation, which blends observations with a short-range forecast. AI-DOP, short for Direct Observation Prediction, tries to skip that step entirely. ECMWF is careful to call this an experimental prototype, not an operational service, and it currently uses only a limited subset of the observations available.

How AI-DOP forecasts are built

According to ECMWF’s AIFS blog, the model is a version of the Artificial Intelligence Forecasting System (AIFS) trained only on gridded observations, using Anemoi, the open-source software that ECMWF develops with several European national weather services. It needs neither a real-time analysis from ECMWF’s physics-based Integrated Forecasting System (IFS) nor historical reanalysis for training.

The model steps through the latest 30 hours of observations in six-hour windows to build its starting point, then forecasts ahead. Two versions run every day: a coarse one at about 100 km resolution and a finer one at about 30 km, the same as the operational AIFS. Its inputs include microwave and infrared satellite soundings, scatterometer winds, GNSS radio occultation, surface reports and upper-air data from radiosondes and aircraft. Microwave imagers are due to be added soon.

The Annual Seminar 2026 presentation by Mihai Alexe and Eulalie Boucher sets out the idea plainly: where AIFS uses AI to emulate the forecast model, AI-DOP aims to emulate “the entire chain using only observations”.

From archive tests to live data

Until now the approach had been tested on past periods with carefully prepared data. The real-time prototype is wired into ECMWF’s existing observation-processing system, SAPP, which already feeds the operational data assimilation and runs on a ten-minute cycle. ECMWF says this means hourly or even half-hourly updates would be possible in principle. Swapping in a new model version simply means pointing the system at a new trained checkpoint.

Live data brings problems that archives hide. Some training datasets are not available in real time, and some instruments are no longer in orbit, so ECMWF fine-tunes the final model on the most recent ten years of data that match today’s observing system. It has added quality checks too: in one 1 November case, drifting buoys reporting pressure more than 10 hPa too low created false lows in the forecast until a “first-guess” check rejected them. Early tests that deliberately remove data show that losing all microwave temperature soundings causes substantially larger errors, particularly further ahead.

Storm Boelo gave an early read on AI-DOP forecasts

ECMWF compared its real-time model, which it calls AIFS-DOP, with the operational AIFS and IFS for Storm Boelo, which crossed the British Isles and north-western Europe on 24-25 October. Three days ahead, ECMWF says, AIFS-DOP captured the storm’s structure and winds across Ireland, the Irish Sea and northern Britain, and was not too windy along the North Sea, Danish and German coasts, unlike the other two. It said the model forecast the storm’s genesis, winds and mild air “at least as well as two operational systems”, while stressing that a single case is not a systematic evaluation.

Fewer inputs, more fragility

The striking point is not the storm result but what the prototype removes. Data assimilation is one of the most expensive and specialised parts of numerical weather prediction. If a model can go from raw observations to a usable forecast in minutes, faster and more frequent updates become realistic, which could help with fast-moving hazards. ECMWF’s open Anemoi tooling also means national weather services are not starting from scratch.

The trade-off is resilience. Traditional assimilation leans on a physics-based background forecast to fill gaps when an instrument fails. AI-DOP learns only from what it sees, so ECMWF’s own tests show it is more exposed when a group of observations disappears. That makes observing-system planning, satellite continuity and quality control more important, not less. The buoy example is a warning that a few bad readings can conjure weather that is not there.

This sits within a wider push to use AI across forecasting, from China’s FengQing ensemble running 51 forecasts out to 15 days to MIT and Microsoft work on forecasts two to six weeks out and Mila’s AI wildfire hazard surrogate. ECMWF’s version is the most radical, because it tests whether the analysis step is needed at all. One encouraging storm does not answer that; months of verified scores across seasons will.

ECMWF says it will keep monitoring real-time forecasts and run observing-system experiments routinely. The next marker is the addition of microwave imager data, and any published verification against IFS and AIFS over a full season.

Sources

• ECMWF: From observations to forecasts: bringing AI-DOP into real time (3 November 2026)
• ECMWF Annual Seminar 2026: Alexe and Boucher, End-to-end real time weather forecasting from observations (PDF)

Related stories

• MIT and Microsoft Researchers Use Machine Learning to Improve Weather Forecasts Two to Six Weeks Out