Europe’s experimental weather AI now forecasts directly from incoming observations

Several drifting buoys reported pressure more than ten hectopascals below the estimate from an earlier short-range forecast. Fed unchecked into an experimental weather AI, those observations produced small, spurious low-pressure features. A first-guess check rejected the anomalous readings and the false lows disappeared. The example, documented by the European Centre for Medium-Range Weather Forecasts on 3 November, reveals a practical part of its new real-time forecasting experiment: a model that starts directly from observations still needs carefully designed checks on what arrives.[1]

ECMWF's broader announcement is that its AI-DOP, short for Direct Observation Prediction, has moved from offline experiments into a real-time prototype connected to the centre's operational observation-processing infrastructure. Rather than starting from a conventional gridded atmospheric analysis generated by data assimilation, it takes in recent observations and predicts subsequent weather. The model is experimental; the blog does not announce a replacement for the operational Integrated Forecasting System or all existing AI forecast products.[1]

What it means to skip the conventional analysis

Standard numerical forecasting estimates an initial atmospheric state by combining satellite, station, ship, aircraft and other observations with an earlier short-range forecast. ECMWF uses four-dimensional variational data assimilation, known as 4D-Var, in this process. The resulting analysis supplies the starting condition for the next numerical simulation. Training an AI model on those analyses still makes it dependent, indirectly, on the older assimilation pipeline.[1]

AI-DOP tests a different route. The current system is built from an Artificial Intelligence Forecasting System architecture trained on gridded observational inputs rather than historical analysis fields. It passes through the latest 30 hours of observations in six-hour windows, using the sequence to infer an evolving atmospheric state. ECMWF is running two prototypes, on grids of approximately 100 kilometres and 30 kilometres respectively. Grid spacing describes the model representation; it does not guarantee predictions accurate to that number of kilometres.[1]

The observations include microwave and infrared sounders, scatterometer winds, aircraft and weather-station reports, ship and buoy measurements, radiosondes and atmospheric-motion vectors. These instruments observe different aspects of the atmosphere at different places and times. The neural network must combine their uneven coverage and learn relationships that normally are reconciled through an explicit analysis stage. A model that accepts less-preprocessed data may remove one substantial processing dependency while becoming more sensitive to changes in the observing system.[1]

The real-time engineering is the news

Historical tests can assume a stable dataset in which all records are already available and standardised. Real-time observations may arrive late, go missing, be noisy or come from instruments that are retired or replaced. ECMWF connected AI-DOP to its Scalable Acquisition and Pre-Processing System, the infrastructure that serves its existing analysis workflow. It operates on a ten-minute processing cycle. An ecFlow system prepares incoming records, runs the model and archives observations and forecasts for verification.[1]

That allows the organisation to launch a new checkpoint within a common processing chain and compare models more easily. It may also support faster refresh cycles in future; ECMWF says hourly or even half-hourly model runs are feasible in principle. This should not be confused with a promise that every model currently issues a verified high-resolution forecast every ten minutes. The ingestion rhythm, configured forecast schedule and actual operational service are different things.[1]

The organisation also describes a training mismatch. Some historical sensors are no longer flying; certain archived datasets are not present in the live processing chain. To reduce this mismatch, the final stage of model training uses only the most recent decade and excludes inputs unavailable in real time. Such adjustments make the model more representative of present conditions, but do not prevent unexpected future instrument outages.[1]

A storm forecast and an observation check

The team examined its forecast for Storm Boelo, which crossed the British Isles and north-western Europe on 24–25 October. For a forecast initialised on 22 October at 00:00 UTC and valid 72 hours later, it reported a realistic depiction of the storm's wind structure and a warm-air intrusion across western Europe, comparing the prototype with operational AI and physical forecasting systems and station observations. Those are encouraging case studies, not a systematic multi-season skill comparison. A striking success on one weather event cannot quantify reliability across different storm types and seasons.[1]

The buoy example probes a different part of the system. When pressure reports were more than ten hectopascals below a previous short forecast, the unfiltered model drew local false low-pressure features. With the check, they disappeared. The episode demonstrates that a weather AI trained directly on observations still needs meteorological judgement embodied in data validation. It also reveals a dependence: the previous forecast becomes part of the input quality-control process even if the primary prediction no longer needs a conventional atmospheric analysis.[1]

ECMWF is also deliberately removing classes of observations in controlled experiments to understand failure under instrument loss. Early findings suggest removing one microwave instrument can have limited effect, while removing an entire class of microwave temperature soundings causes larger errors at longer lead times. An operational forecaster must remain dependable when satellites fail, not just when a complete historical record is available.[1]

A real-time framework for comparison

The important question is whether direct-observation modelling improves speed or accuracy without sacrificing robustness, transparency and reproducibility. Answering it requires verification across many dates, seasons, geographic regions and observational disruptions, with equivalent baselines and computation costs. ECMWF has built a processing chain that lets researchers generate those comparisons against incoming weather. The prototype can therefore accumulate evidence under the changing conditions that an operational forecast service would face.[1]

The direct-observation approach changes where the atmosphere is reconstructed: the learned model takes on work conventionally performed in an explicit analysis stage. Connecting it to the live stream makes that idea testable under instrument outages, late arrivals and real storms. The buoy example shows why observation expertise remains central. It also shows the benefit of the prototype: a problem can be traced from incoming measurements through the forecast and tested against a specific correction.[1]

Report an error or suggest a correction