On July 5, 2023, Nature published “Accurate medium-range global weather forecasting with 3D neural networks,” describing Pangu-Weather, a deep-learning weather model built by researchers at Huawei Cloud. Huawei said it was the first AI model to forecast more accurately than the leading numerical weather prediction (NWP) systems for medium-range forecasts, while running far faster.
Pangu-Weather uses a 3D Earth-Specific Transformer (3DEST) architecture designed to handle the non-uniform, three-dimensional structure of atmospheric data, and it was trained on 43 years of ERA5 reanalysis (1979-2021) from the European Centre for Medium-Range Weather Forecasts. Once trained, it produces a global 24-hour forecast in about 1.4 seconds on a single GPU - roughly 10,000 times faster than traditional numerical models, which run on large supercomputers - while matching or beating their accuracy from one hour out to seven days.
Released alongside the closely related GraphCast from DeepMind, Pangu-Weather was part of a 2023 wave of papers that established machine-learning weather models as genuine competitors to decades of physics-based forecasting, and it pushed operational centers like ECMWF to adopt their own AI systems.