FourCastNet

FourCastNet (Fourier Forecasting Neural Network) is NVIDIA's data-driven global weather model, built on an Adaptive Fourier Neural Operator backbone with a vision-transformer architecture. Trained on ERA5 reanalysis data, it forecasts atmospheric variables at 0.25-degree global resolution and generates a week-long forecast in under two seconds, orders of magnitude faster than the ECMWF's numerical IFS system, with comparable or better accuracy on several variables.

The original deterministic model and checkpoints (nvidia/fourcastnet1) are Apache 2.0 and usable via NVIDIA's PhysicsNeMo and Earth2Studio tooling. FourCastNet 3 (2025) later introduced a probabilistic ensemble formulation; this entry covers the original that kicked off the ML weather-forecasting wave.