“Accurate medium-range global weather forecasting with 3D neural networks,” by Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu and Qi Tian of Huawei Cloud, appeared in Nature in July 2023, with an earlier preprint posted in November 2022. It introduced Pangu-Weather, one of the first machine-learning weather models shown to outperform the European Centre for Medium-Range Weather Forecasts (ECMWF) on operational reanalysis data.
The model uses three-dimensional deep networks with what the authors call Earth-specific priors, plus a hierarchical strategy that combines forecasts at different time steps to reduce the errors that accumulate when a model is run forward many times. Trained on 39 years of global data, Pangu-Weather produced stronger deterministic forecasts than ECMWF’s system across all the variables tested, and generated a forecast in seconds rather than the hours a physics-based supercomputer run requires.
Pangu-Weather, alongside GraphCast and FourCastNet, was part of the wave of 2022-2023 results that convinced the meteorology community that data-driven forecasting was not a curiosity but a genuine challenger to the simulation methods built over half a century. For a general reader, it is a concrete case of an AI model matching and beating a mature scientific tool at a fraction of the cost.