FourCastNet 3: A geometric approach to probabilisticmachine-learning weather forecasting at scale

FourCastNet 3 advances global weather modeling by implementing a scalable, geometric machinelearning (ML) approach to probabilistic ensemble forecasting. The approach is designed to respectspherical geometry and to accurately model the spatially correlated probabilistic nature of theproblem, resulting in stable spectra and realistic dynamics across multiple scales. FourCastNet 3delivers forecasting accuracy that surpasses leading conventional ensemble models and rivals the bestdiffusion-based methods, while producing forecasts 8 to 60 times faster than these approaches. Incontrast to other ML approaches, FourCastNet 3 demonstrates excellent probabilistic calibrationand retains realistic spectra, even at extended lead times of up to 60 days. All of these advancesare realized using a purely convolutional neural network architecture tailored for spherical geometry.Scalable and efficient large-scale training on 1024 GPUs and more is enabled by a novel trainingparadigm for combined model- and data-parallelism, inspired by domain decomposition methods inclassical numerical models. Additionally, FourCastNet 3 enables rapid inference on a single GPU,producing a 60-day global forecast at 0.25°, 6-hourly resolution in under 4 minutes. Its computationalefficiency, medium-range probabilistic skill, spectral fidelity, and rollout stability at subseasonaltimescales make it a strong candidate for improving meteorological forecasting and early warningsystems through large ensemble predictions.