|
--- |
|
license: cc-by-sa-4.0 |
|
metrics: |
|
- mse |
|
pipeline_tag: graph-ml |
|
--- |
|
|
|
# AIFS Single - v0.2.1 |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
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). |
|
|
|
<div style="display: flex; justify-content: center;"> |
|
<img src="aifs_10days.gif" alt="AIFS 10 days Forecast" style="width: 50%;"/> |
|
</div> |
|
|
|
|
|
We show that AIFS produces highly skilled forecasts for upper-air variables, surface weather parameters and |
|
tropical cyclone tracks. AIFS 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) |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
<!-- Provide a longer summary of what this model is. --> |
|
|
|
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. |
|
|
|
<div style="display: flex; justify-content: center;"> |
|
<img src="encoder_graph.jpeg" alt="Encoder graph" style="width: 50%;"/> |
|
<img src="decoder_graph.jpeg" alt="Decoder graph" style="width: 50%;"/> |
|
</div> |
|
|
|
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. |
|
|
|
- **Developed by:** ECMWF |
|
- **Model type:** Encoder-processor-decoder model |
|
- **License:** CC BY-SA 4.0 |
|
|
|
### Model Sources |
|
|
|
<!-- Provide the basic links for the model. --> |
|
|
|
|
|
- **Repository:** [Anemoi](https://anemoi-docs.readthedocs.io/en/latest/index.html) |
|
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. |
|
- **Paper:** https://arxiv.org/pdf/2406.01465 |
|
|
|
## How to Get Started with the Model |
|
|
|
To be able to run AIFS to generate a new forecast, you can use ai-models https://github.com/ecmwf-lab/ai-models. |
|
`ai-models` command can be used to run different models, since in this case we are looking at using AIFS we need to specify |
|
`anemoi` as `model-name` and then pass the path to the checkpoint (`aifs_single_v0.2.1.ckpt`) and the initial conditions. |
|
You can find an example of a set of initial conditions in the GRIB file `example_20241107_12_n320.grib`. |
|
|
|
Use the code below to get started with the model. |
|
|
|
``` |
|
# 1st - create the conda environment |
|
export CONDA_ENV=aifs-env |
|
conda create -n ${CONDA_ENV} python=3.10 |
|
conda activate ${CONDA_ENV} |
|
|
|
pip install torch=2.4 |
|
pip install anemoi-inference[plugin] anemoi-models==0.2 |
|
pip install ninja |
|
pip install flash-attn --no-build-isolation |
|
|
|
# 2nd - Run ai-models to generate weather forecast |
|
|
|
ai-models anemoi --checkpoint aifs_single_v0.2.1.ckpt --file example_20241107_12_n320.grib |
|
``` |
|
|
|
🚨 **Note** we train AIFS using `flash_attention` (https://github.com/Dao-AILab/flash-attention). |
|
There are currently some issues when trying to install flash attention with the latest PyTorch version 2.5 and CUDA 12.4 (https://github.com/Dao-AILab/flash-attention/issues/1330). |
|
For that reason, we recommend you install PyTorch 2.4. |
|
Additonally the use of 'Flash Attention' package also imposes certain requirements in terms of software and hardware. Those can be found under #Installation and Features in https://github.com/Dao-AILab/flash-attention |
|
|
|
🚨 **Note** the `aifs_single_v0.2.1.ckpt` checkpoint just contains the model’s weights. |
|
That file does not contain any information about the optimizer states, lr-scheduler states, etc. |
|
|
|
**Note** By default, when running `ai-models` the model will be run for a 10-day lead time (240 hours). |
|
It is possible to extend or modify the lead time to for example 15 days by doing `ai-models --lead-time 360` |
|
Please refer to `ai-models` documentation for more information regarding defaults and available command line options. |
|
|
|
|
|
After running the `ai-models` command the output of the forecast should be written into `anemoi.grib` |
|
Below you can find an example to read that file and load it as numpy array or xarray. |
|
|
|
``` |
|
|
|
import earthkit.data as ekd |
|
import cartopy.crs as ccrs |
|
import numpy as np |
|
from plotting_utils import map_scatter |
|
|
|
source_filename='anemoi.grib' |
|
aifs_forecast = ekd.from_source('file',source_filename) |
|
|
|
# to load a variable as a numpy array |
|
aifs_forecast.sel(param="2t").to_numpy() |
|
|
|
# to load a variables as an xarray array |
|
aifs_forecast.sel(param="2t").to_xarray() |
|
|
|
aifs_forecast_2t_xr = aifs_forecast.sel(param="2t").to_xarray() |
|
|
|
fig, axes = plt.subplots(subplot_kw={'projection': ccrs.Robinson()}, figsize=(16,6)) |
|
|
|
map_scatter(data_v021_2t_xr.latitude, data_v021_2t_xr.longitude, data_v021_2t_xr.t2m[0,1,0,:], |
|
axes=axes, |
|
proj=ccrs.Robinson(), |
|
vmin=np.percentile(data_v021_2t_xr.t2m[0,1,0,:], 2), |
|
vmax=np.percentile(data_v021_2t_xr.t2m[0,1,0,:], 98), |
|
cmap="RdYlBu_r", |
|
mask_ocean=False, |
|
plot_title='2-Meter Temperature Forecast', |
|
cmap_label="Temperature (K)" |
|
) |
|
|
|
``` |
|
<div style="display: flex; justify-content: center;"> |
|
<img src="2t_aifs_v021.png" alt="AIFS 2m Temperature" style="width: 50%;"/> |
|
</div> |
|
|
|
|
|
|
|
## Training Details |
|
|
|
### Training Data |
|
|
|
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. --> |
|
|
|
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}\\). |
|
|
|
<div style="display: flex; justify-content: center;"> |
|
<img src="aifs_diagram.png" alt="AIFS 2m Temperature" style="width: 80%;"/> |
|
</div> |
|
|
|
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 | Pressure level: 50,100, 150, 200, 250,300, 400, 500, 600,700, 850, 925, 1000 | Both | |
|
| 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 | |
|
| 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 |
|
|
|
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. --> |
|
|
|
- **Pre-training**: It was performed on ERA5 for the years 1979 to 2020 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}\\). |
|
- **Fine-tuning I**: The pre-training is then followed by rollout on ERA5 for the years 1979 to 2018, this time with a LR |
|
of \\(6 × 10^{-7}\\). As in [Lam et al. [2023]](doi: 10.21957/slk503fs2i) we increase the |
|
rollout every 1000 training steps up to a maximum of 72 h (12 auto-regressive steps). |
|
- **Fine-tuning II**: Finally, to further improve forecast performance, we fine-tune the model on operational real-time IFS NWP |
|
analyses. This is done via another round of rollout training, this time using IFS operational analysis data |
|
from 2019 and 2020 |
|
|
|
|
|
#### 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 |
|
|
|
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. --> |
|
|
|
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 |
|
|
|
<!-- This section describes the evaluation protocols and provides the results. --> |
|
|
|
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]. |
|
|
|
<div style="display: flex; justify-content: center;"> |
|
<img src="aifs_v021_scorecard.png" alt="Scorecard comparing forecast scores of AIFS versus IFS (2022)" style="width: 80%;"/> |
|
</div> |
|
|
|
|
|
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). |
|
|
|
Additional evaluation analysis including tropycal cyclone performance or comparison against other popular data-driven models can be found in AIFS preprint (https://arxiv.org/pdf/2406.01465v1) section 4. |
|
|
|
# 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 forecast owing to the linear loss scaling with height |
|
- AIFS currently provides reduced intensity of some high-impact systems such as tropical cyclones. |
|
|
|
## Technical Specifications |
|
|
|
### Hardware |
|
|
|
<!-- {{ hardware_requirements | default("[More Information Needed]", true)}} --> |
|
|
|
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](https://anemoi-docs.readthedocs.io/en/latest/index.html). |
|
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 there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. --> |
|
|
|
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. |
|
``` |
|
|
|
|
|
## More Information |
|
|
|
[More Information Needed](https://arxiv.org/pdf/2406.01465) |
|
|
|
|
