FourCastNet (“Fourier Forecasting Neural Network”) was introduced in an arXiv paper submitted on February 22, 2022 by Jaideep Pathak and colleagues, largely from NVIDIA and academic collaborators. It was an early, influential data-driven global weather model and helped kick off the AI-weather wave that GraphCast and Pangu-Weather extended in 2023.
The model forecasts global weather at 0.25-degree resolution - the same grid spacing as leading operational systems - using Adaptive Fourier Neural Operators (AFNO) paired with a vision-transformer backbone. The Fourier neural operator learns in a way that is largely resolution-invariant and is well suited to the kind of partial-differential-equation dynamics that govern the atmosphere. FourCastNet accurately predicted fast, small-scale fields like surface wind speed, precipitation, and water vapor, which matters for wind energy and extreme-weather planning.
Its defining feature was speed: FourCastNet generated a week-long global forecast in well under a few seconds on a GPU, roughly five orders of magnitude faster than traditional numerical weather prediction, while approaching state-of-the-art accuracy on large-scale variables. That speed makes large ensembles - thousands of forecasts to quantify uncertainty - cheap, which is one of the hardest things to do with conventional models.