FourCastNet 3
#12 v Věda a bio AIAccelerated Understanding · v3 · od Juli 2025 (arXiv-Paper veröffentlicht 16. Juli 2025; NGC-Modellversion 0.1.0 vom 29. Juli 2025) · 2× · naposledy 27. 8. 2026
FourCastNet 3 (FCN3) is a probabilistic AI model for global weather and climate forecasting developed by NVIDIA together with Lawrence Berkeley National Laboratory, UC Berkeley, and Caltech. It uses a purely convolutional architecture tailored for spherical geometry to generate ensemble forecasts with stable spectral fidelity for lead times up to 60 days. NVIDIA states the model surpasses conventional ensemble weather systems (e.g., IFS-ENS) in accuracy and is substantially faster than comparable diffusion-based models such as GenCast. It is distributed via training frameworks like Makani as well as Hugging Face, NGC, and NVIDIA NIM, targeting research, industry, and government users in weather and climate modeling.
Vlastnosti
| Deployment Model | Self-hosted via Docker/NIM container (nvcr.io/nim/nvidia/fourcastnet), locally on GPU workstations, data centers, or cloud; single-GPU inference possible |
| Use Case Scope | Medium-range and subseasonal weather forecasting and climate modeling for industry, academic, and government use |
| Integrations | Training framework Makani, signal-processing library torch-harmonics, deployment toolkit Earth2Studio, ERA5 training data (Copernicus/ECMWF) |
| License | Apache 2.0 License (model weights, Hugging Face); Makani training code also Apache 2.0 |
| Platform | NVIDIA Earth-2 platform; available via Hugging Face, NGC Catalog, Earth2Studio (Python), and NVIDIA NIM containers |
| Price | Model weights freely available via Hugging Face/NGC (commercial & non-commercial use); NIM hosting may require paid NVIDIA AI Enterprise license |
| Release Date | July 16, 2025 (arXiv preprint); NGC checkpoint v0.1.0 released July 29, 2025 |