Multi-Modal Foundation · ecmwf
AIFS Single 1.1
A data-driven medium-range weather forecasting model producing 6-hour interval atmospheric and surface predictions.
What AIFS Single 1.1 is
AIFS Single 1.1 is a deterministic global weather forecasting model created by the European Centre for Medium-Range Weather Forecasts (ECMWF). It serves as an intermediate release built upon the open-source Anemoi framework, designed to forecast upper-air variables, surface weather conditions and tropical cyclone tracks. The model is integrated into operational workflows, running four times daily in parallel with ECMWF's physics-based Numerical Weather Prediction (NWP) systems.
This version modifies the earlier AIFS Single 1.0 release by adjusting soil moisture loss weighting during training, which eliminates spurious precipitation artefacts without significantly changing overall forecast skill. The model provides predictions on an approximate 31 km horizontal grid across 13 vertical pressure levels.
AIFS Single 1.1 allows meteorologists and atmospheric researchers to evaluate machine-learning-based forecast outputs alongside traditional numerical weather systems, facilitating comparative analyses and reproducible meteorological research.
How it works
AIFS adopts an encoder-processor-decoder architecture consisting of a graph neural network (GNN) encoder and decoder coupled with a sliding window transformer processor. The network receives atmospheric conditions from two preceding time points (t - 6h and t_0) to forecast the state at t + 6h. Prognostic atmospheric fields span 13 pressure levels between 1000 hPa and 50 hPa alongside various surface diagnostic and forcing parameters.
Training is divided into a pre-training phase using ECMWF's ERA5 reanalysis data spanning 1979 to 2022, followed by a fine-tuning rollout stage using operational Integrated Forecast System (IFS) analyses from 2016 to 2022. The model optimizes an area-weighted mean squared error (MSE) loss using the AdamW optimizer with Flash Attention. Loss weights decrease linearly with altitude and are empirically scaled across variables, with reduced weight applied to vertical velocity.
Key facts
- Model architecture
- GNN encoder/decoder with sliding window transformer processor
- Horizontal resolution
- ~31 km (N320 grid)
- Vertical resolution
- 13 pressure levels (50 to 1000 hPa)
- Forecast step
- 6 hours
- Training data
- ERA5 reanalysis (1979–2022) and ECMWF IFS operational analyses (2016–2022)
- Checkpoint file size
- 1.19 GB
What you can use it for
- Medium-range global weather forecasting
- Generate stepped 6-hour atmospheric forecasts across surface variables and pressure levels up to 10 days out using initial conditions from ECMWF open data.
- Tropical cyclone tracking
- Track the trajectory and progression of tropical cyclones using predicted wind, pressure and upper-air geopotential fields.
- Machine learning vs NWP comparison
- Benchmark machine learning forecast performance, skill scores and anomaly correlations against operational physics-based numerical weather prediction systems.
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
- Produces competitive forecast skill across surface parameters, upper-air states and tropical cyclone tracks relative to physics-based NWP models.
- Corrects spurious rainfall points present in version 1.0 caused by improper soil moisture loss weighting.
- Supports high-resolution training through data-parallel distribution and Flash Attention integration across multi-GPU setups.
- Natively integrates with the open-source Anemoi framework tools for data handling, training and operational inference.
Watch out for
- Subject to field blurring at longer lead times, a characteristic common to models trained with mean squared error objectives.
- Forecast skill is reduced in the stratosphere due in part to a relatively low model top at 50 hPa.
- Systematically underestimates the intensity of extreme high-impact phenomena such as tropical cyclones.
- User-executed model runs on GPUs are not bit-for-bit reproducible due to hardware non-determinism.
- Predicts precipitation in volumetric units (m³/m²) rather than traditional millimetres (mm).
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 1.1 commercially?
Check the licence before commercial use. The cc-by-4.0 licence carries terms worth reading before you depend on it.
What distinguishes AIFS Single 1.1 from AIFS Single 1.0?
Version 1.1 rectifies an issue in version 1.0 where incorrect weighting of soil moisture loss produced false rainfall points, while retaining identical resolution and broad forecast skill.
What inputs are required to generate a forecast?
The model requires atmospheric and surface states at two consecutive time steps 6 hours apart (t - 6h and t_0), including upper-air wind, temperature, humidity, geopotential and relevant surface variables derived from ECMWF analyses.
Why do precipitation forecasts differ in magnitude from standard observations?
AIFS is trained to produce precipitation in cubic metres per square metre (m³/m²). When retrieving ECMWF Open Data benchmarks, precipitation is usually converted to millimetres (mm).
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 1.1'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.