Multi-Modal Foundation · ecmwf
AIFS ENS 1.0
A probabilistic data-driven weather forecasting model for upper-air, surface variables and cyclone tracks.
What AIFS ENS 1.0 is
AIFS ENS 1.0 is an ensemble-based probabilistic data-driven global weather forecasting model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). It serves as the initial release of the AIFS ensemble system and acts as the predecessor to AIFS ENS 2.0. The system produces forecasts for upper-air atmospheric variables, surface weather conditions, and tropical cyclone trajectories.
The model addresses the need for computationally efficient probabilistic forecasts alongside traditional physics-based numerical weather prediction (NWP) systems. ECMWF operates AIFS ENS four times daily alongside its operational physics-based NWP models. The model's forecasting performance is evaluated against NWP operational analyses and direct observational measurements.
How it works
AIFS ENS uses an encoder-processor-decoder architecture. The encoder and decoder are constructed using graph neural networks (GNNs), while the central processor employs a sliding window transformer trained with flash attention. The architecture is modular and supports multiple levels of parallelism to facilitate training on high-resolution meteorological inputs.
The model was trained on ECMWF's ERA5 reanalysis and operational numerical weather prediction analyses. Training was guided by minimising the Continuous Ranked Probability Score (CRPS) computed over a small ensemble, optimising the network specifically for probabilistic weather estimation.
Key facts
- Model architecture
- Graph neural network (encoder/decoder) with sliding window transformer (processor)
- Training objective
- Continuous Ranked Probability Score (CRPS) over a small ensemble
- Training data
- ERA5 reanalysis and ECMWF operational NWP analyses
- Horizontal resolution
- ~31 km
- Vertical resolution
- 13 levels
- Precipitation output unit
- m³/m²
- Default GPU memory requirement
- 38 GB
What you can use it for
- Global probabilistic weather prediction
- Generating ensemble forecasts for surface weather parameters and upper-air atmospheric variables across the globe at roughly 31 km horizontal resolution.
- Tropical cyclone tracking
- Forecasting and tracking the trajectory of tropical cyclones using data-driven ensemble rollouts.
- Precipitation estimation
- Producing quantitative precipitation forecasts natively in cubic metres per square metre.
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
- Delivers skilled probabilistic predictions for surface conditions, upper-air variables, and tropical cyclone tracks.
- Employs a modular graph-transformer architecture supporting distributed parallelism on high-resolution datasets.
- Optimised directly for probabilistic forecasting skill by training on Continuous Ranked Probability Score (CRPS) loss.
Watch out for
- Requires 38 GB of GPU memory by default for inference, which can only be reduced to 24 GB by chunking at the expense of throughput.
- Hardware non-determinism in GPUs prevents exact reproduction of official ECMWF operational forecasts during local execution.
- Requires specialized software and hardware dependencies, including Flash Attention.
- Checkpoint files contain model weights only, omitting optimizer and learning rate scheduler states.
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 1.0 commercially?
Check the licence before commercial use. The cc-by-4.0 licence carries terms worth reading before you depend on it.
What atmospheric resolution does AIFS ENS 1.0 cover?
The atmospheric component provides a horizontal resolution of approximately 31 km across 13 vertical levels.
What units are used for precipitation predictions?
The model was trained on and directly predicts precipitation in m³/m². By contrast, ECMWF Open Data serves precipitation values in millimetres (mm).
How can memory consumption be reduced during inference?
Memory demand can be decreased from 38 GB to 24 GB by setting ANEMOI_INFERENCE_NUM_CHUNKS to a multiple of 2 (such as 16), though this decreases prediction throughput.
Can I perfectly reproduce official ECMWF AIFS ENS forecasts locally?
No. Due to GPU non-determinism, independently executed runs will not perfectly match official ECMWF AIFS ENS forecast outputs.
How to cite it
Simon Lang, Mihai Alexe, Mariana C. A. Clare, Christopher Roberts, Rilwan Adewoyin, Zied Ben Bouallègue, Matthew Chantry, Jesper Dramsch, Peter D. Dueben, Sara Hahner, Pedro Maciel, Ana Prieto-Nemesio, Cathal O'Brien, Florian Pinault, Jan Polster, Baudouin Raoult, Steffen Tietsche, Martin Leutbecher. AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score. arXiv preprint arXiv:2412.15832, 2024.
Written from AIFS ENS 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.