Multi-Modal Foundation · ecmwf
AIFS Single 1.0
A data-driven weather forecasting model by ECMWF for deterministic global predictions.
What AIFS Single 1.0 is
AIFS Single 1.0 is an operational, data-driven global weather forecasting model created by the European Centre for Medium-Range Weather Forecasts (ECMWF). Developed on the open-source Anemoi framework, it serves as the first fully operational release of the deterministic Artificial Intelligence Forecasting System (AIFS), replacing earlier experimental versions.
The model produces deterministic predictions for upper-air atmospheric parameters, surface weather variables, and tropical cyclone trajectories. It runs alongside ECMWF's traditional physics-based Integrated Forecast System (IFS) Numerical Weather Prediction (NWP) model, providing an alternative, highly efficient data-driven pipeline for global atmospheric forecasting.
How it works
AIFS employs an encoder-processor-decoder architecture consisting of a graph neural network (GNN) encoder and decoder paired with a sliding-window transformer processor. It takes atmospheric states at time steps t-6h and t0, along with static and temporal forcing fields such as orography and insolation, to forecast the atmospheric state at t+6h. Prognostic fields are evaluated at approximately 31 km horizontal resolution across 13 vertical pressure levels (ranging from 1000 hPa to 50 hPa).
Training is conducted in two stages using AdamW and an area-weighted mean squared error (MSE) loss scaled across vertical levels and prognostic variables. Pre-training is carried out over 260,000 steps on ERA5 reanalysis data spanning 1979 to 2022. The model is subsequently fine-tuned using multi-step autoregressive rollouts on ECMWF operational IFS NWP analysis data spanning 2016 to 2022, with fields interpolated from native O1280 resolution to N320.
Key facts
- Architecture
- GNN encoder/decoder with a sliding window transformer processor
- Horizontal resolution
- ~31 km (N320 grid)
- Vertical resolution
- 13 pressure levels (1000 to 50 hPa)
- Forecast step
- 6 hours
- Framework
- Anemoi
- Training data
- ECMWF ERA5 reanalysis (1979–2022) and IFS operational analyses (2016–2022)
- Model checkpoint size
- 1.19 GB
What you can use it for
- Global Deterministic Weather Prediction
- Forecast standard surface and upper-air atmospheric conditions, including geopotential, temperature, wind components, and humidity, at 6-hour intervals up to 10 days ahead.
- Surface Boundary and Hydrological Modelling
- Predict surface-level environmental conditions including soil moisture, soil temperature, runoff water equivalent, snowfall water equivalent, and total precipitation.
- Renewable Energy and Radiation Assessment
- Estimate 100-metre wind speed components as well as surface downwards short-wave (solar) and long-wave (thermal) radiation for energy meteorology applications.
- Tropical Cyclone Tracking
- Track the progression and movement of tropical cyclones using predicted global atmospheric circulation fields.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
AIFS Single 1.0
Strengths and limits
Good at
- Delivers high-skill deterministic forecasts for upper-air conditions, surface weather parameters, and tropical cyclone tracks.
- Incorporates a broad set of surface and land variables, including 100 m winds, soil moisture, soil temperature, and downward radiation.
- Designed within the modular Anemoi framework, facilitating multi-node, data-parallel training and inference.
Watch out for
- Tends to produce smoothed or blurred forecast fields at extended lead times due to the use of a weighted mean squared error loss.
- Exhibits reduced forecast skill in the stratosphere, partly caused by a relatively low model top at 50 hPa.
- Underestimates the intensity of certain high-impact weather phenomena, such as tropical cyclones.
- Outputs cannot exactly reproduce official operational forecasts when run by external users due to non-deterministic GPU computations.
Inputs and outputs
You provide
Forecast ahead
Measured from the most recent analysis ECMWF has published.
Field
You get back
The chosen field over the globe at the forecast time.
Analysis time, valid time and the field's range.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use AIFS Single 1.0 commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run AIFS Single 1.0 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on CPU only, typically in 12s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
What inputs are required to initialise a forecast?
The model requires two consecutive atmospheric states at t-6h and t0 containing the standard surface and 13 pressure-level upper-air variables, as well as static forcing fields such as orography, land-sea mask, and insolation.
What are the output units for precipitation?
AIFS is trained on and natively outputs precipitation in m³/m², whereas data downloaded directly from ECMWF Open Data is provided in mm.
Which software tools are needed to run inference?
Inference is performed using the anemoi-inference package alongside PyTorch and Flash Attention, which imposes specific GPU hardware and software driver requirements.
How to cite it
Lang, S., Alexe, M., Chantry, M., Dramsch, J., Pinault, Florian, 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 1.0'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.