GribStream Blog
ECMWF's AIFS Marine brings AI forecasting to the ocean and sea ice
ECMWF's experimental AIFS Marine jointly forecasts atmosphere, waves, surface ocean and sea ice, pointing toward an AI Earth-system model.
ECMWF is extending AIFS beyond atmospheric weather forecasting. Its research prototype AIFS Marine jointly predicts the atmosphere, ocean waves, sea ice and conditions at the ocean surface within a single machine-learning model.
Across nearly all marine variables evaluated, the researchers report improvements broadly equivalent to one additional day of medium-range forecast skill compared with their physics-based baselines. That is a result from a deterministic research evaluation, not a guarantee for every variable, location or forecast.
The work points toward an AI Earth-system forecasting model. AIFS Marine remains experimental, and ECMWF's October 12 announcement does not announce an operational implementation date.
One model across the atmosphere–ocean interface
Traditional Earth-system forecasting uses specialized atmosphere, ocean, wave and sea-ice models, with information exchanged through explicit coupling mechanisms. AIFS Marine instead places their state variables inside the same machine-learning model and learns relationships between them from historical data.
Those relationships matter: stronger winds generate larger waves, sea ice attenuates waves, and tropical cyclones can mix colder water toward the surface. Changes in ocean conditions can also affect the air above them.
The goal is more than a larger variable catalogue. Joint prediction aims to represent these interactions coherently. The approach does not explicitly resolve the deeper ocean or replace every physical process represented in a conventional coupled model.
What AIFS Marine predicts
Operational AIFS v2 already includes waves, alongside atmospheric and land variables. The Marine prototype extends that state with surface-ocean and sea-ice information. Its fields include:
- sea-surface temperature, salinity and height;
- surface ocean currents;
- sea-ice concentration, volume and velocity;
- significant wave height, mean wave period and wave direction.
This is a selection of the research variables, not an operational dissemination inventory. Training combines ERA5 atmospheric data, a dedicated wave hindcast and ECMWF's ORAS6 ocean reanalysis. Marine fields require adaptations for missing values over land, physical bounds and slower evolution than atmospheric variables. Preprocessing and training-loss weights help balance the components. Hahner, Zampieri et al. (2026).
Around one extra day of marine forecast skill
ECMWF reports broad improvements at medium-range lead times. In the wave evaluation, significant wave-height errors were about 10% lower against satellite observations. The paper's headline satellite comparison evaluates the AIFS Waves configuration; the full Marine configuration also benefits from explicit sea-ice information near the ice edge.
Sea-ice-edge forecasts improve in both hemispheres, especially in the Southern Ocean, and sea-surface-temperature forecasts improve systematically. The roughly one-day gain describes the evaluated marine variables under a deterministic framing. It does not mean that the complete AIFS forecast is universally one day better than IFS.
Sea-surface height is the main exception. It develops an increasingly negative bias with lead time. The researchers suggest that the long-term sea-level-rise signal in the training data may contribute and identify this as an area for further work. Research evaluation.
Does the model learn meaningful physical relationships?
Forecasts of Hurricanes Idalia and Franklin show strong winds alongside enhanced wave activity and cold wakes in sea-surface temperature. The response is consistent with storm-driven mixing that brings colder water toward the surface.
The researchers also tested unusual initial conditions. Isolated long-period swells propagated across ocean basins while new wind-driven waves developed. When all sea ice was removed from the initial state, it regenerated gradually in the seasonally appropriate hemisphere. Ice extent recovered faster than ice volume, as expected physically.
These experiments provide evidence for learned interactions across the components. They do not prove that every forecast will obey all physical constraints. Case studies and sensitivity experiments.
What happens to atmospheric forecasts?
Large-scale atmospheric skill remains broadly unchanged. Explicit sea ice improves near-surface temperatures around the ice edge, while an evolving ocean surface benefits tropical temperature forecasts.
There are also small degradations in some near-surface wind scores. ECMWF links these tradeoffs to consistency between training datasets and the balance of learning across components. Marine gains therefore need to be assessed alongside atmospheric performance as the system develops.
Research files are available; a live Marine feed is not announced
The distinction between a research result and a data-product launch matters. ECMWF has published supporting forecasts on Zenodo, including GRIB files for the AIFS Marine atmosphere, ocean and waves.
Those files support the paper's experiments. The repository explicitly describes them as non-operational forecasts for academic and scientific research, with restrictions on operational and commercial decision-making. They are not a continuously updated public forecast feed.
As of October 13, 2026, the cited announcement gives no operational implementation date or real-time Marine dissemination contract. Surface ocean and sea ice are the next components ECMWF aims to bring into AIFS. Further work includes the sea-surface-height bias, probabilistic forecasting and, at longer ranges, the deeper ocean.
What this means for GribStream users
GribStream already serves operational ECMWF products through AIFS Operational and AIFS Ensemble. AIFS Marine is not currently available through the GribStream API.
If these marine capabilities reach operational public output, they could broaden the usefulness of AIFS for offshore energy, shipping, sea-ice monitoring and tropical-cyclone analysis. Before offering them, we would need to verify the public feed, licensing, variables, forecast timing and coverage.
The research adds another dimension to AIFS's development, alongside the AIFS v2 expansion and the planned HourGlass hourly surface forecasts. We will watch for the step from experimental marine forecasts to operational output.