FourCastNet
#4 v Věda a bio AICaltech · 5× · naposledy 27. 8. 2026
84
Momentum
FourCastNet is a data-driven AI weather model developed by researchers at NVIDIA Corporation together with Anima Anandkumar (Caltech). It uses Adaptive Fourier Neural Operators (or in version 2/3 Spherical Fourier Neural Operators) to forecast global atmospheric variables such as wind speed, temperature, precipitation, and water vapor based on the ERA5 dataset. The model achieves accuracy comparable to physics-based models like ECMWF IFS, but delivers forecasts orders of magnitude faster. FourCastNet is provided by NVIDIA as Open Source code (NVlabs/GitHub), as model checkpoints on NGC/Hugging Face, and as a hostable NIM microservice (Earth-2).
Vývoj momenta
01.06.30.08.
Vlastnosti
| Deployment Model | Self-hosted via NIM container (cloud, data center, workstation) or NVIDIA-hosted API on build.nvidia.com |
| Use Case Scope | Short- to medium-range global weather forecasting, extreme weather detection (cyclones, atmospheric rivers), subseasonal climate modeling |
| Integrations | Integrable via NVIDIA Earth2Studio, Earth2MIP, PhysicsNeMo framework, and as an ai-models plugin for ECMWF |
| License | FourCastNet 3 (Hugging Face) ready for commercial/non-commercial use; Makani training code under Apache License 2.0; torch-harmonics under BSD-3-Clause |
| Platform | NVIDIA Earth-2 platform, available via NGC, Hugging Face, GitHub, and as a NIM microservice on build.nvidia.com |
| Price | Model weights/code free (open source); NIM API usage via build.nvidia.com with free trial credits, production use requires NVIDIA AI Enterprise license |
| Release Date | February 22, 2022 (original preprint); FourCastNet 3 in July 2025 |