Multi-Modal Foundation · ecmwf
AIFS Single 2.0
A data-driven medium-range weather model by ECMWF providing deterministic 15-day global atmospheric and wave forecasts.
What AIFS Single 2.0 is
AIFS Single 2.0 (Artificial Intelligence Forecasting System Single version 2) is a data-driven global weather forecasting model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). Built on the open-source Anemoi framework, it generates deterministic medium-range forecasts on an operational basis, producing a single 15-day global forecast at six-hour intervals four times daily.
The model addresses the computational demands of global numerical weather prediction (NWP) by replacing or augmenting physical simulation pipelines with a machine-learned system. Relative to previous releases, version 2 incorporates an operational surface ocean wave component, land-surface snow representation, an extended stratospheric reach, and revised vertical velocity handling.
For atmospheric, oceanic, and environmental researchers, the system outputs global prognostic and diagnostic fields across 14 vertical pressure levels and the Earth's surface at approximately 31-kilometre horizontal resolution.
How it works
AIFS Single 2.0 employs an encoder-processor-decoder architecture featuring graph neural network (GNN) encoders and decoders linked by a sliding window transformer processor. The model is trained autoregressively to step atmospheric and surface states forward by 6 hours, taking atmospheric representations at time t-6h and t0 to predict conditions at t+6h.
The training pipeline involves pre-training on historical ERA5 reanalysis data spanning 1979 to 2022 across 260,000 steps, followed by 7,900 steps of fine-tuning on ECMWF operational analysis and prototype IFS Cycle 50r1 esuite analysis data spanning 2018 to 2024. Native O1280 (roughly 0.1°) fields are interpolated to an N320 grid (roughly 0.25° or ~31 km) for model processing.
Key facts
- Architecture
- Graph neural network encoder-decoder with a sliding window transformer processor
- Horizontal resolution
- ~31 km (N320 grid)
- Vertical resolution
- 14 pressure levels (10 to 1000 hPa)
- Forecast step
- 6 hours
- Forecast horizon
- 15 days
- Pre-training data
- ERA5 reanalysis (1979–2022)
- Fine-tuning data
- ECMWF operational analysis and IFS 50r1 esuite analysis (2018–2024)
- Checkpoint size
- 994 MB
- Training hardware
- 16 GH200 GPUs (120 GB)
What you can use it for
- Global medium-range weather forecasting
- Simulating deterministic atmospheric trajectories out to 15 days at 6-hour time steps using operational analyses or reanalysis initial states.
- Ocean wave and sea-state prediction
- Forecasting significant wave height, mean wave period, mean wave direction, wave drag coefficients, and period-banded wave heights across six specific period intervals from 10 to 30 seconds.
- Stratospheric dynamics research
- Analysing upper-atmosphere wind, geopotential, and temperature fields using the expanded vertical grid extending up to the 10 hPa pressure level.
- Surface hydrology and snow monitoring
- Tracking land-surface changes, soil temperature, volumetric soil moisture across two layers, runoff, precipitation, and fractional snow cover.
Run it here
We list this model but cannot run it yet: loads, then fails at inference — see FAILS_AT_INFERENCE.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Integrates a dedicated 11-variable wave component for data-driven marine wave predictions.
- Includes expanded stratospheric vertical resolution reaching 10 hPa.
- Operates at an approximately 31-kilometre horizontal resolution with a compact 994 MB checkpoint.
- Improves vertical velocity representations by transitioning parameter W from a prognostic to a diagnostic output.
Watch out for
- Provides deterministic, single-trajectory forecasts without explicit ensemble-based uncertainty spread.
- Fine-tuning includes proprietary IFS 50r1 esuite data that is not publicly accessible.
- Vertical velocities are provided purely diagnostically rather than being stepped forward prognostically.
Inputs and outputs
You provide
Graph
Node and edge data in the format this model expects — check its model card.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · cc-by-4.0
The cc-by-4.0 licence carries terms worth reading before you depend on it.
Questions
Can I use AIFS Single 2.0 commercially?
Check the licence before commercial use. The cc-by-4.0 licence carries terms worth reading before you depend on it.
What vertical pressure levels does the model cover?
The model spans 14 pressure levels: 10, 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, and 1000 hPa.
How are inputs and forecast intervals structured?
The model takes atmospheric states from two preceding time points (t-6h and t0) and outputs predictions for t+6h, which can be iteratively rolled out to a 15-day forecast.
Which wave variables are available?
The model predicts significant wave height, mean wave period, mean wave direction, coefficient of drag with waves, and partitioned significant wave heights for periods spanning 10–12 s, 12–14 s, 14–17 s, 17–21 s, 21–25 s, and 25–30 s.
Which software tools are used to run inference?
Inference is executed using the open-source anemoi-inference package and configured via YAML files within the Anemoi software framework.
How to cite it
Lang, S., Alexe, M., Chantry, M., Dramsch, J., Pinault, F., Raoult, B., ... & Rabier, F. (2024). AIFS-ECMWF's data-driven forecasting system. arXiv preprint arXiv:2406.01465.
Written from AIFS Single 2.0's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.