Predicting Extreme Weather Risk Three Weeks in Advance With FourCastNet

Опубликовано 3 июля 2023, 12:53

A small 50-member ensemble takes one hour to generate on a large supercomputer using traditional numerical simulations. But small ensembles often miss low-probability, high-impact events.

By running FourCastNet—an AI framework developed by researchers at NVIDIA, Caltech, and Lawrence Berkeley Lab—in NVIDIA Modulus, and leveraging FNO, we were able to generate 21-day global weather trajectories of 1,000 ensemble members in a tenth of the time it previously took to do a single trajectory, while using 1,000X less energy.

In this demo, we look at how a 1,000-member ensemble was able to predict the increased risk of an extreme 2018 heat wave in North Africa that broke all records.

arxiv.org/abs/2306.03838
arxiv.org/abs/2208.05419
arxiv.org/abs/2202.11214

#FourCastNet, #Physics-ML, #The-Berlin-Summit-For-EVE