What is AI Weather Forecasting?
TL;DR
Predicting weather with neural networks trained on historical data. Far faster and cheaper than traditional numerical weather prediction, with comparable or better accuracy. GraphCast and NVIDIA Earth-2 are leading examples.
AI Weather Forecasting: Definition & Explanation
AI weather forecasting trains neural networks on decades of reanalysis data (past atmospheric states, such as ERA5) to predict future weather from the current state. Traditional numerical weather prediction (NWP) turns atmospheric physics into equations solved over hours on supercomputers, but AI models generate forecasts up to 10 days out in minutes. With compute costs orders of magnitude lower and accuracy matching or exceeding traditional forecasts on many metrics, it has been a major shock to meteorology.\n\nLeading models include the graph-neural-network GraphCast (Google DeepMind) and its successor WeatherNext; NVIDIA Earth-2 (FourCastNet + CorrDiff) for fast, high-resolution GPU runs; the multi-Earth-system foundation model Microsoft Aurora; and the 3D-Earth Huawei Pangu-Weather. Commercially, Tomorrow.io provides business alerts and APIs, Atmo serves national agencies with high-resolution forecasts, Jua focuses on energy supply-demand, and Spire supplies satellite observation data.\n\n(★) AI models can err on extreme or rapidly evolving events not in their training data; for important decisions, check model agreement (ensembles) and combine with traditional forecasts. (★) For disaster-related decisions, always prioritize official warnings from meteorological authorities. (★) When using commercial APIs, verify resolution, update frequency and SLA for your region.