Details

This NGC asset is a collection of diagnostic models built into Earth-2 MIP. Diagnostic models are a class of models that post-process the outputs of other weather/climate models into new quantities of interest.

Models

Precipitation AFNO

The FourCastNet diagnostic model which predicts total precipitation from 20 atmospheric variables. The total precipitation, sourced from the ERA5 re-analysis dataset, represents the accumulated liquid and frozen water that falls to the Earth’s surface through rainfall and snow. It is defined in units of length as the depth of water that would accumulate if spread evenly over a unit grid box of the model. This model uses the Adaptive Fourier Neural Operator architecture, trained for 25 epochs. For additional information on this model, Bill Collin's National Academies lecture provides an excellent overview of this model.

ClimateNet

This is a segmentation model trained on ClimateNet: a community-sourced human-expert-labeled curated dataset that captures tropical cyclones (TCs) and atmospheric rivers (ARs) in high-resolution climate model output. The model provided is a convolutional neural network trained to predict both TCs and ARs from a small set of atmospheric variables. The ClimateNet diagnostic model can be used to achieve various objectives such as: calculate a variety of TC and AR statistics at a fine-grained level; applied to different climate scenarios and different datasets; and rapidly analyzing large amounts of climate model output.

How to use

These diagnostic models are designed to be used using Earth-2 MIP package. Please refer to Earth-2 MIP documentation.

References

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

• Adaptive Fourier Neural Operators: Efficient Token Mixers for Transformers

• ClimateNet: an expert-labeled open dataset and deep learning architecture for enabling high-precision analyses of extreme weather

• The ERA5 global reanalysis