Atmospheric rivers: GraphCast and AIFS locate rainfall more accurately than conventional models, but get the amounts wrong

A global comparison covering a ten-day forecast period shows that AI-based weather models are better at pinpointing the location of heavy rainfall associated with atmospheric rivers, although they do not accurately estimate the volume of that rainfall.

Atmospheric rivers, these long corridors of water vapour, are the source of many of the most extreme precipitation on the planet. While AI-based weather models now compete with traditional numerical forecasting on major atmospheric variables, their reliability on these rain events remained poorly measured on a global scale.

Four models, ten days of expiry

Marina Vicens-Miquel, Amy McGovern and their co-authors compared the 24-hour rainfall totals predicted by two conventional US systems (GFS and its GEFS set) and two AI models (Google DeepMind's GraphCast and ECMWF's AIFS) from day one to day ten. The assessment, conducted with the Extreme Weather Bench framework, covers the entire world as well as North America, Europe, Australia and New Zealand, using IMERG satellite observations as a reference.

Good geography, bad volumes

GraphCast and AIFS better preserve the spatial structure of rainfall for up to ten days, especially for heavy precipitation and long time frames. However, this advantage does not translate into accurate cumulations: AI models tend to overestimate moderate to heavy rainfall and underestimate the most intense rainfall in the long term, while physical models drift towards a marked dry bias. The authors provide a reproducible testbed to track these deviations.