FourCastNet
#11 in Science & Bio AIAccelerated Understanding · 2× · last seen Aug 27, 2026
FourCastNet is a data-driven AI model by NVIDIA (part of NVIDIA Earth-2) for global weather and climate forecasting, based on Adaptive Fourier Neural Operator and Spherical Fourier Neural Operator architectures. It predicts atmospheric and surface variables such as wind speed, temperature, and pressure with 6-hour time steps while maintaining stable forecast quality over simulated periods of more than one year. The current version FourCastNet 3 delivers probabilistic ensemble forecasts that compete with leading physics-based weather models and is available via NVIDIA NIM as a microservice as well as through Hugging Face/NGC as a model. It is deployed commercially and non-commercially for meteorology, climate research, and early warning systems.
Features
| Deployment Model | Self-hosted via Docker container (NIM) on cloud, data center, or workstation; supports air-gapped deployment; requires NVIDIA GPU (e.g. H100, B200) with compute capability ≥ 8.0 |
| Use Case Scope | Short- to medium-range global weather forecasting, climate modeling, extreme weather early warning (cyclones, atmospheric rivers), meteorological and climate research |
| Integrations | Earth2Studio Python package for data ingestion (e.g. ERA5 via ARCO), HTTP API for inference, NGC CLI for model download, Docker/NVIDIA Container Toolkit |
| License | NIM container: NVIDIA AI Product Agreement; model usage: NVIDIA AI Foundation Models Community License; commercial and non-commercial use permitted |
| Platform | NVIDIA Earth-2 platform, available as an NVIDIA NIM microservice (Docker container), on NGC catalog, Hugging Face, and build.nvidia.com |
| Price | NVIDIA AI Enterprise license from $4,500/GPU/year; cloud usage approx. $1/GPU/hour; free prototyping via build.nvidia.com and NGC download |
| Release Date | First paper/preprint published February 22, 2022; FourCastNet NIM version 2.0.0 (FourCastNet 3) currently available |