Multi-Modal Foundation · ecmwf
AIFS ENS 2.0
A data-driven medium-range probabilistic weather forecasting model generating global ensemble forecasts.
What AIFS ENS 2.0 is
AIFS ENS 2.0 is an operational, data-driven ensemble forecasting model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). Implemented in May 2026 as the successor to version 1, it forms part of ECMWF's Artificial Intelligence Forecasting System (AIFS) alongside the deterministic AIFS Single 2.0. The model is deployed operationally to generate a 51-member, 15-day global forecast at six-hour intervals four times daily.
The system produces probabilistic predictions across the atmosphere, land surface, ocean waves, and tropical cyclone tracks. By serving as an open machine-learning counterpart to ECMWF's Integrated Forecasting System (IFS) ensemble, it provides working atmospheric scientists and meteorologists with an alternative computational approach to medium-range probabilistic weather prediction.
How it works
AIFS ENS 2.0 uses an encoder-processor-decoder architecture implemented on the open-source Anemoi framework. It combines a graph neural network (GNN) encoder and decoder with a sliding window transformer processor. The model takes atmospheric states at time t minus six hours and time t, alongside static and dynamic surface forcings, to forecast the atmospheric and surface state at time t plus six hours. Training incorporates a multi-scale loss formulation designed for proper score optimisation, replacing the earlier afCRPS loss, and enforces variable bounds to ensure physical consistency.
The model is pre-trained on ERA5 reanalysis data spanning 1979 to 2022 over 300,000 steps. It is subsequently fine-tuned over 7,900 steps using ECMWF operational analyses and IFS 50r1 esuite analysis data from 2018 to 2024. Input fields are interpolated from their native O1280 resolution (approximately 0.1 degrees) down to N320 resolution (approximately 0.25 degrees or 31 kilometres) with 14 vertical pressure levels extending up to 10 hPa.
Key facts
- Architecture
- GNN encoder-decoder with a sliding window transformer processor
- Horizontal resolution
- Approximately 31 km (N320 grid)
- Vertical resolution
- 14 pressure levels (10 to 1000 hPa)
- Forecast step
- 6-hourly
- Ensemble size
- 51 members
- Forecast lead time
- 15 days
- Pre-training data
- ERA5 reanalysis (1979–2022)
- Fine-tuning data
- ECMWF operational analysis and IFS 50r1 esuite analysis (2018–2024)
- Software framework
- Anemoi
- Checkpoint size
- 2.55 GB
What you can use it for
- Medium-range ensemble weather forecasting
- Generate 51-member global probabilistic forecasts up to 15 days ahead at 6-hourly intervals for atmospheric pressure levels and surface parameters.
- Ocean wave and sea-state prediction
- Forecast 11 wave variables including significant wave height, mean wave direction, mean wave period, drag coefficients, and wave energy across specific period bands.
- Tropical cyclone tracking
- Output probabilistic tropical cyclone track predictions in standardised BUFR format.
- Land surface and precipitation monitoring
- Predict diagnostic surface conditions such as fraction of snow cover, convective precipitation, soil temperature, and volumetric soil moisture.
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 ocean wave component with 11 distinct wave variables.
- Provides stratospheric coverage with the addition of pressure fields at 10 hPa.
- Treats vertical velocity as a diagnostic field to improve velocity fields.
- Applies variable bounding to prevent unphysical parameter values.
- Uses a multi-scale loss function for proper score optimisation across spatial scales.
Watch out for
- Operates at a spatial resolution of roughly 31 kilometres, requiring sub-grid phenomena to be parametrised or diagnosed.
- Relies on pre-interpolated N320 analysis inputs downscaled from native O1280 IFS grids.
- Requires high-memory GPU resources for training and ensemble execution, with each ensemble instance split across four 120GB GH200 GPUs.
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 ENS 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 is the difference between AIFS Single and AIFS ENS?
AIFS Single produces a deterministic forecast, whereas AIFS ENS generates a 51-member probabilistic ensemble forecast up to 15 days ahead.
Which new parameters were added in version 2?
Version 2 adds 14 new parameters, including 11 wave parameters such as significant wave height and mean wave direction, fraction of snow cover, convective precipitation, volumetric soil moisture, and tropical cyclone track outputs.
What vertical levels does the model output?
The model provides outputs on 14 pressure levels: 10, 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, and 1000 hPa.
How are forecasts executed using the model weights?
Inference is executed using the anemoi-inference package, which retrieves initial conditions from open data sources, loads the checkpoint, and steps the forecast forward.
How to cite it
Lang, S., Leutbecher, M., & Maciel, P. (2025). A multi-scale loss formulation for learning a probabilistic model with proper score optimisation. arXiv preprint arXiv:2506.10868.
Written from AIFS ENS 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.