Details

This NGC asset is a FourCastNet model checkpoint package. Model checkpoint package refers to the set of artifacts needed to run inference using pre-trained model which includes the model checkpoint, set of sample inputs, inference script.

Architecture

FourCastNet, short for Fourier Forecasting Neural Network, is a global data-driven weather forecasting model that provides accurate short to medium-range global predictions. FourCastNet accurately forecasts high-resolution, fast-timescale variables such as the surface wind speed, precipitation, and atmospheric water vapor. FourCastNet uses the the Adaptive Fourier Neural Operator (AFNO) archiecture. This particular neural network architecture is appealing as it is specifically designed for high-resolution inputs and synthesizes several key recent advances in DL into one model. Refer to the reference paper for additional details.

Training

This model was trained on a comprehensive dataset called ERA5, that ECMWF provides publicly, which consists of hourly estimates of several atmospheric variables at a latitude and longitude resolution of 0.25 degrees. This checkpoint is trained on 26 atmospheric variables from ERA5, using snapshots spaced 6 hours apart at 0000 hrs, 0600 hrs, 1200 hrs and 1800 hrs. The full training dataset consists of data from the year 1979 to 2015 (both included).

How to use?

A minimal inference script is provided in this model checkpoint package to get you started easily.

To run inference on this checkpoint, you can follow the below steps (all the files needed are included in the zip file):

Launch Modulus docker container

docker run --rm --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 --runtime nvidia -v ${PWD}:/examples -it nvcr.io/nvidia/modulus/modulus:23.09

Download this checkpoint zip file, and unzip it

wget 'https://api.ngc.nvidia.com/v2/models/nvidia/modulus/modulus_fcn/versions/v0.1/files/fcn.zip'
unzip fcn.zip

Run the simple inference

cd fcn/
python simple_inference.py

You can also use the model checkpoint from this model checkpoint package to evalue it with the earth2mip framework.

References

• FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

• Adaptive Fourier Neural Operators: Efficient Token Mixers for Transformers

• The ERA5 global reanalysis