NCEP Implementation of FourCastNet model using GDAS data as ICs

We use ECMWF's ai_models_fourcastnetv2 plugin to run the model. FoureCastNet v2-small applies Spherical Fourier Neural Operators (SFNOs) as neural network architecture.

Prerequisites

The following packages are needed:

• numpy
• boto3
• xarray
• pygrib
• torch
• ai-models-fourcastnetv2
• iris
• iris_grib
• eccodes
• cf_units

Run the model

Download pre-computed stats

wget https://get.ecmwf.int/repository/test-data/ai-models/fourcastnetv2/small/weights.tar

wget https://get.ecmwf.int/repository/test-data/ai-models/fourcastnetv2/small/global_means.npy

wget https://get.ecmwf.int/repository/test-data/ai-models/fourcastnetv2/small/global_stds.npy

Get input data (initial conditions) from GDAS product

The script ncep/gdas.py can be used to prepare input data, simply use:

python gdas.py YYYYMMDDHH -s <s3 or nomads> -m <wgrib2 or pygrib> -k <yes or no>

file input_YYYYMMDDHH.npy will be created and saved in the current firectory.

Run FourCastNetv2 in inference mode

Run inference using

python inference.py YYYYMMDDHH -w </path/to/weightsandstats> -i </path/to/input/input_YYYYMMDDHH.npy> -o </path/to/output/> -l <forecast-hours>

Output

The forecast results will be saved in GRIB2 format.

Products:

The FourCastNetGFS model runs 4 times a day at 00Z, 06Z, 12Z, and 18Z cycles. The horizontal resolution is on 0.25 degree lat-lon grid. The vertical resolutions are 13 pressure levels.

The 13 pressure levels include: 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, and 1000 hPa.

The model output fields are:

• 3D fields on pressure levels: temperature
  U and V component of wind
  geopotential height
  relative humidity

• 2D surface fields: 10-m U and V components of wind
  100-m U and V components of wind
  2-m temperature
  mean sea-level pressure
  Precipitable water in the entire atmosphere as single column

The near real-time forecast outputs along with inputs are available on AWS.

For each cycle, the dataset contains input files to feed into FourCastNet found in the directory:

    'fourcastnetgfs.yyyymmdd/hh/input_yyyymmddhh.npy'

and 10-day forecast results for the current cycle found in the following directories:

    'fourcastnetgfs.yyyymmdd/hh'