Multi-Modal Foundation · ecmwf
AIFS Single 0.2.1
A data-driven medium-range global weather forecasting model developed by ECMWF.
What AIFS Single 0.2.1 is
AIFS Single 0.2.1 is an early-release deterministic weather prediction model created by the European Centre for Medium-Range Weather Forecasts (ECMWF). It demonstrates data-driven medium-range global forecasting built on Anemoi, an open-source machine learning framework co-developed by ECMWF and European national meteorological services.
The model produces 6-hour atmospheric forecasts across upper-air and surface variables, which can be stepped forward auto-regressively to generate medium-range outlooks. It addresses the computational cost and formulation constraints of conventional physics-based numerical weather prediction by mapping historical atmospheric states directly to future states.
Forecast skill has been evaluated against ECMWF operational Integrated Forecast System (IFS) analyses and direct observational data, including radiosonde measurements and SYNOP ground weather reports.
How it works
AIFS employs an encoder-processor-decoder architecture. It uses a graph neural network (GNN) for the encoder and decoder stages to map between the geographical grid and internal latent representations, coupled with a sliding-window transformer as the processor. The network receives atmospheric states at t-6h and t0, alongside static or external surface forcings such as orography, land-sea masks, insolation, and temporal coordinates, and predicts the updated state at t+6h.
Training is divided into three stages: pre-training on ERA5 reanalysis data spanning 1979 to 2020 using an AdamW optimizer, followed by an initial fine-tuning rollout on ERA5 (1979–2018) stepping auto-regressively up to 72 hours (12 steps). A second fine-tuning stage performs rollout training on operational IFS numerical weather prediction analyses from 2019 and 2020. The loss function is an area-weighted mean squared error (MSE) across prognostic variables, scaled empirically across levels with weights that decrease linearly with height.
Key facts
- Model architecture
- GNN encoder-decoder with a sliding window transformer processor
- Forecast step
- 6 hours (accepts t-6h and t0 to predict t+6h)
- Primary training data
- ERA5 reanalysis (1979–2020) and operational IFS analysis (2019–2020)
- Checkpoint size
- 1.19 GB
- Framework
- Anemoi
What you can use it for
- Medium-range surface weather forecasting
- Generating multi-day deterministic forecasts for near-surface variables, including 2-metre temperature, dewpoint temperature, 10-metre wind components, mean sea-level pressure, and cumulative precipitation.
- Tropical cyclone track prediction
- Projecting the paths and trajectory movements of tropical cyclones over medium-range forecasting windows.
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
- Demonstrates high forecasting skill compared against operational IFS analyses, radiosonde soundings, and SYNOP observations.
- Employs a modular graph neural network and sliding-window transformer design that scales through data and model parallelism.
- Outputs comprehensive multi-level prognostic fields alongside surface precipitation (total and convective).
Watch out for
- Suffers from blurring of forecast fields at longer lead times, a characteristic linked to training with a mean squared error loss.
- Exhibits reduced forecast skill in the stratosphere due to loss weighting that decreases linearly with altitude.
- Underestimates the intensity of certain high-impact weather phenomena, including tropical cyclones.
- Inference code requires Flash Attention, which introduces specific hardware and software dependencies.
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 0.2.1 commercially?
Check the licence before commercial use. The cc-by-4.0 licence carries terms worth reading before you depend on it.
What inputs does AIFS Single 0.2.1 require to generate a forecast?
It requires global atmospheric states at two time steps: six hours prior (t-6h) and the initial condition time (t0).
How are forecasts run beyond 6 hours?
The model is evaluated auto-regressively, feeding its 6-hour forecast back as input for the next cycle to build multi-day lead-time predictions up to medium-range horizons.
How to cite it
Lang, S., Alexe, M., Chantry, M., Dramsch, J., Pinault, F., Raoult, B., Clare, M. C. A., Lessig, C., Maier-Gerber, M., Magnusson, L., et al. (2024). AIFS-ECMWF's data-driven forecasting system. arXiv preprint arXiv:2406.01465.
Written from AIFS Single 0.2.1'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.