Here, we introduce the Artificial Intelligence Forecasting System (AIFS), a data driven forecast
model developed by the European Centre for Medium-Range Weather Forecasts (ECMWF).
The release of AIFS Single v1.1 represents a slight modification to the AIFS model. Version 1.1
supersedes the existing operational version, 1.0 AIFS-single.
The new version, 1.1, brings minor changes to the v1.0 model. These changes mainly correspond to the removal of
spurious rainfall points caused by incorrect soil moisture loss weighting during training of the v1.0 model.
AIFS 10 days Forecast
AIFS produces highly skilled forecasts for upper-air variables, surface weather parameters and
tropical cyclone tracks. AIFS Single is run four times daily alongside ECMWF’s physics-based NWP model and forecasts
are available to the public under ECMWF’s open data policy (https://www.ecmwf.int/en/forecasts/datasets/open-data).
Note that due to the non-determinism of GPUs, users will be unable to exactly reproduce an official AIFS forecast
when running AIFS Single themselves.
For more information on this update please see the confluence page
## Data Details
Data parameters
New parameters
More detailed information about the new parameters introduced with AIFS Single v1.0 is provided in the table below.
Short Name
Name
Units
Component Type
Lev.Type
ssrd
Surface short-wave (solar) radiation downwards
\(J m^{-2}\)
AIFS
sfc
strd
Surface long-wave (thermal) radiation downwards
\(J m^{-2}\)
AIFS
sfc
lcc
Low cloud cover
\((0 - 1)\)
AIFS
sfc
mcc
Medium cloud cover
\((0 - 1)\)
AIFS
sfc
hcc
High cloud cover
\((0 - 1)\)
AIFS
sfc
sf
Snowfall water equivalent
\(kg m^{-2}\)
AIFS
sfc
tcc
Total cloud cover
\((0 - 1)\)
AIFS
sfc
100u
100 metre U wind component
\(m s^{-1}\)
AIFS
sfc
100v
100 metre V wind component
\(m s^{-1}\)
AIFS
sfc
rowe
Runoff water equivalent (surface plus subsurface)
\(kg m^{-2}\)
AIFS
sfc
vsw
Volumetric soil moisture
\(m^3 m^{-3}\)
AIFS
sol
sot
Soil temperature
\(K\)
AIFS
sol
Changes to existing parameters
There are no changes to existing parameters already introduced with AIFS Single v0.2.1.
Note
Regarding precipitation units, it's worth noting that AIFS model was trained on \(m^{3}/m^{2}\) and will therefore produce precip in that units.
If one wants to retrieve precipitation from Open data, the units will be \(mm\).
#### Discontinued parameters
No parameters have been discontinued with regards to the previous version of AIFS Single v0.2.1.
Model Details
Model Description
AIFS is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor,
and is trained on ECMWF’s ERA5 re-analysis and ECMWF’s operational numerical weather prediction (NWP) analyses.
Encoder graph
Decoder graph
It has a flexible and modular design and supports several levels of parallelism to enable training on
high resolution input data. AIFS forecast skill is assessed by comparing its forecasts to NWP analyses
and direct observational data.
There are no changes in resolution compared to previous version AIFS Single v0.2.1.
Component
Horizontal Resolution [kms]
Vertical Resolution [levels]
Atmosphere
AIFS-single v1.1
~ 31
13
Model Sources
Repository:Anemoi is an open-source framework for
creating machine learning (ML) weather forecasting systems, which ECMWF and a range of national meteorological
services across Europe have co-developed.
To generate a new forecast using AIFS, you can use anemoi-inference. In the following notebook, a
step-by-step workflow is specified to run the AIFS using the HuggingFace model:
🚨 Note the aifs_single_v1.1.ckpt checkpoint just contains the model’s weights.
That file does not contain any information about the optimizer states, lr-scheduler states, etc.
How to train AIFS Single v1.1
To train this model you can use the configuration files included in this repository and the following Anemoi packages:
AIFS is trained to produce 6-hour forecasts. It receives as input a representation of the atmospheric states
at \(t_{−6h}\), \(t_{0}\), and then forecasts the state at time \(t_{+6h}\).
AIFS 2m Temperature
The full list of input and output fields is shown below:
Field
Level type
Input/Output
Geopotential, horizontal and vertical wind components, specific humidity, temperature
Surface pressure, mean sea-level pressure, skin temperature, 2 m temperature, 2 m dewpoint temperature, 10 m horizontal wind components, total column water
Surface
Both
Soil moisture and soil temperature (layers 1 & 2)
Surface
Both
100m horizontal wind components, solar radiation (Surface short-wave (solar) radiation downwards and Surface long-wave (thermal) radiation downwards), cloud variables (tcc, hcc, mcc, lcc), runoff and snow fall
Surface
Output
Total precipitation, convective precipitation
Surface
Output
Land-sea mask, orography, standard deviation of sub-grid orography, slope of sub-scale orography, insolation, latitude/longitude, time of day/day of year
Surface
Input
Input and output states are normalised to unit variance and zero mean for each level. Some of
the forcing variables, like orography, are min-max normalised.
Training Procedure
Based on the different experiments we have made - the final training recipe for AIFS Single v1.1 has deviated slightly
from the one used for AIFS Single v0.2.1 since we found that we could get a well trained model by skipping the ERA5
rollout and directly doing the rollout on the operational-analysis (extended) dataset. When we say 'extended' we refer
to the fact that for AIFS Single v0.2.1 we used just operational-analysis data from 2019 to 2021, while in this new
release we have done the fine-tunning from 2016 to 2022.
The other important change in the fine-tuning stage is that for AIFS Single v0.2.1 after the 6hr model training the
optimiser was not restarted (ie. rollout was done with the minimal lr of \(3 × 10^{-7}\)). For this release we have seen
that restarting the optimiser for the rollout improves the model's performance. For the operational-fine tuning rollout
stage, the learning rate cycle is restarted, gradually decreasing to the minimum value at the end of rollout.
Pre-training: It was performed on ERA5 for the years 1979 to 2022 with a cosine learning rate (LR) schedule and a
total of 260,000 steps. The LR is increased from 0 to \(10^{-4}\) during the first 1000 steps, then it is annealed to a
minimum of \(3 × 10^{-7}\). The local learning rate used for this stage is \(3.125 × 10^{-5}\).
Fine-tuning: The pre-training is then followed by rollout on operational real-time IFS NWP analyses for the years
2016 to 2022, this time with a local learning rate of \(8 × 10^{−7}\), which is decreased to \(3 × 10^{−7}\). Rollout steps
increase per epoch. In this second stage the warm up period of the optimiser is 100 steps to account for shorter length
of this stage. Optimizer step are equal to 7900 ( 12 epoch with ~630 steps per epoch).
As in the previous version of aifs-single for fine-tuning and initialisation of the model during inference, IFS fields
are interpolated from their native O1280 resolution (approximately \(0.1°\)) down to N320 (approximately \(0.25°\)).
Training Hyperparameters
Optimizer: We use AdamW (Loshchilov and Hutter [2019]) with the \(β\)-coefficients set to 0.9 and 0.95.
Loss function: The loss function is an area-weighted mean squared error (MSE) between the target atmospheric state
and prediction.
Loss scaling: A loss scaling is applied for each output variable. The scaling was chosen empirically such that
all prognostic variables have roughly equal contributions to the loss, with the exception of the vertical velocities,
for which the weight was reduced. The loss weights also decrease linearly with height, which means that levels in
the upper atmosphere (e.g., 50 hPa) contribute relatively little to the total loss value.
Speeds, Sizes, Times
Data parallelism is used for training, with a batch size of 16. One model instance is split across four 40GB A100
GPUs within one node. Training is done using mixed precision (Micikevicius et al. [2018]), and the entire process
takes about one week, with 64 GPUs in total. The checkpoint size is 1.19 GB and as mentioned above, it does not include the optimizer
state.
Evaluation
AIFS is evaluated against ECMWF IFS (Integrated Forecast System) for 2022. The results of such evaluation are summarized in
the scorecard below that compares different forecast skill measures across a range of
variables. For verification, each system is compared against the operational ECMWF analysis from which the forecasts
are initialised. In addition, the forecasts are compared against radiosonde observations of geopotential, temperature
and windspeed, and SYNOP observations of 2 m temperature, 10 m wind and 24 h total precipitation. The definition
of the metrics, such as ACC (ccaf), RMSE (rmsef) and forecast activity (standard deviation of forecast anomaly,
sdaf) can be found in e.g Ben Bouallegue et al. ` [2024]. No significant changes in skill wer found in the v1.1 model fix.
### AIFS Single v1.1 vs AIFS Single v1.0 (2024)
Scorecard comparing forecast scores of AIFS versus IFS (2024)
### AIFS Single v1.1 vs IFS (2024)
Scorecard comparing forecast scores of AIFS versus IFS (2024)
Forecasts are initialised on 00 and 12 UTC. The scorecard show relative score changes as function of lead time (day 1 to 10) for northern extra-tropics (n.hem),
southern extra-tropics (s.hem), tropics and Europe. Blue colours mark score improvements and red colours score
degradations. Purple colours indicate an increased in standard deviation of forecast anomaly, while green colours
indicate a reduction. Framed rectangles indicate 95% significance level. Variables are geopotential (z), temperature
(t), wind speed (ff), mean sea level pressure (msl), 2 m temperature (2t), 10 m wind speed (10ff) and 24 hr total
precipitation (tp). Numbers behind variable abbreviations indicate variables on pressure levels (e.g., 500 hPa), and
suffix indicates verification against IFS NWP analyses (an) or radiosonde and SYNOP observations (ob). Scores
shown are anomaly correlation (ccaf), SEEPS (seeps, for precipitation), RMSE (rmsef) and standard deviation of
forecast anomaly (sdaf, see text for more explanation).
Known limitations
This version of AIFS shares certain limitations with some of the other data-driven weather forecast models that are trained with a weighted MSE loss, such as blurring of the forecast fields at longer lead times.
AIFS exhibits reduced forecast skill in the stratosphere, partially due to a low model top.
AIFS currently provides reduced intensity of some high-impact systems such as tropical cyclones.
We acknowledge PRACE for awarding us access to Leonardo, CINECA, Italy. In particular, this version of the AIFS has been trained
on 64 A100 GPUs (40GB).
Software
The model was developed and trained using the AnemoI framework.
AnemoI is a framework for developing machine learning weather forecasting models. It comprises of components or packages
for preparing training datasets, conducting ML model training and a registry for datasets and trained models. AnemoI
provides tools for operational inference, including interfacing to verification software. As a framework it seeks to
handle many of the complexities that meteorological organisations will share, allowing them to easily train models from
existing recipes but with their own data.
Citation
If you use this model in your work, please cite it as follows:
BibTeX:
@article{lang2024aifs,
title={AIFS-ECMWF's data-driven forecasting system},
author={Lang, Simon and Alexe, Mihai and Chantry, Matthew and Dramsch, Jesper and Pinault, Florian and Raoult, Baudouin and Clare, Mariana CA and Lessig, Christian and Maier-Gerber, Michael and Magnusson, Linus and others},
journal={arXiv preprint arXiv:2406.01465},
year={2024}
}
APA:
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.
