FourCastNet3 (FCN3) is the latest AI global weather forecasting system from NVIDIA Earth-2. FCN3 offers an unprecedented combination of probabilistic skill, computational efficiency, spectral fidelity, ensemble calibration, and stability at subseasonal timescales. Its medium-range forecasting accuracy matches that of leading machine learning models, such as GenCast, and exceeds that of traditional numerical weather prediction systems, such as IFS-ENS. A single 60-day FCN3 rollout with 0.25 and 6-hourly resolution is computed in under four minutes on a single NVIDIA H100 Tensor Core GPU-an 8x speedup over GenCast and a 60x speedup over IFS-ENS.
It also has remarkable calibration and spectral fidelity, with ensemble members retaining realistic spectral properties even at extended lead times of 60 days. FCN3 demonstrates a significant leap towards data-driven weather prediction with large ensembles from medium-range to subseasonal timescales.
data-src=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-ensemble-2.gif alt=5 FourCastNet3 ensemble members. class=lazyload wp-image-103924/>Figure 1. 2-week rollout of 15 FourCastNet3 ensemble members, displaying surface wind speeds during this period
FCN3 architecture FourCastNet3 employs a fully convolutional, spherical neural operator architecture, based on spherical signal processing primitives (see Figure 2). Unlike FourCastNet2, which is based on the Spherical Fourier Neural Operator, FCN3 uses local spherical convolutions alongside spectral convolutions.
These convolutions are parameterized using Morlet wavelets and formulated in the framework of discrete-continuous group convolutions. This approach enables anisotropic, localized filters well-suited to localized atmospheric phenomena, while also guaranteeing computational efficiency through a custom implementation in NVIDIA CUDA.
data-src=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-png.webp alt=FCN3 architecture diagram. class=lazyload wp-image-103851 data-srcset=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-png.webp 1999w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-300x184-png.webp 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-625x383-png.webp 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-179x110-png.webp 179w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-768x470-png.webp 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-1536x941-png.webp 1536w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-645x395-png.webp 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-490x300-png.webp 490w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-147x90-png.webp 147w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-362x222-png.webp 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-180x110-png.webp 180w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-1024x627-png.webp 1024w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-architecture-882x540-png.webp 882w data-sizes=(max-width: 1999px) 100vw, 1999px />Figure 2. FCN3 is a neural operator for spherical signals that maps atmospheric and surface variables at the current time step to the next. Stochasticity is introduced through a hidden Markov model approach, which takes a spherical noise variable as conditioning input
FCN3 introduces stochasticity at every predictive step through a latent noise variable whose evolution is governed by a diffusion process on the sphere. This hidden-Markov formulation enables efficient one-step generation of ensemble members-a key advantage over diffusion model-based approaches. FCN3 is trained jointly as an ensemble, minimizing a composite loss function that combines the continuously ranked probability score (CRPS) in space and in the spectral domain. This approach ensures that FCN3 learns the correct spatial correlations in the underlying stochastic atmospheric processes.
Scaling ML models is often crucial to achieving competitive skill, but the effects of scale haven't been investigated in data-driven weather models. FCN3 is unusual in its computational ambition. To scale it, we introduce a novel paradigm for model-parallelism inspired by domain decomposition in traditional numerical weather modeling.
data-src=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-png.webp alt=FourCastNet3 probabilistic scores. class=lazyload wp-image-103852 data-srcset=https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-png.webp 1999w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-300x194-png.webp 300w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-625x404-png.webp 625w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-179x115-png.webp 179w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-768x496-png.webp 768w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-1536x992-png.webp 1536w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-645x417-png.webp 645w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-465x300-png.webp 465w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-139x90-png.webp 139w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-362x234-png.webp 362w, https://developer-blogs.nvidia.com/wp-content/uploads/2025/07/FCN3-scores-170x110-png.webp 170w, https://deve
