African rainfall: AI models are about 5% ahead of the physical forecast, but lose on dry terrain

GraphCast, GenCast and the Functional Generative Network are compared to the IFS model after calibration, on several satellite datasets.

Learned weather forecasting models cost much less to run than physical models — a decisive argument for regions where the computing infrastructure is lacking. It remains to be seen what they are worth where the stakes are most concrete. A team led by Shruti Nath, with Docko Sow and Koomi Toussaint Amoussouvi, submitted an assessment of their potential operational usefulness for rains over Africa on September 24, 2026.

Three learned systems — GraphCast, GenCast, and the Functional Generative Network — are compared to the IFS physical model. To make the comparison fair, the outputs are post-processed by isotonic distributional regression and then scored at the continuous ranked probability score against multiple precipitation satellite datasets.

The learned systems outweigh all observation games by about 5% in media, but the average masks a geography: they dominate in humid regions, while the physical model retains the advantage on dry and high ground. All of them retain a higher competence than climatology over most seasons and up to long time frames. GraphCast achieves a calibrated competence comparable to ensemble approaches, and the Functional Generative Network widens the gap on the most distant deadlines. The authors emphasize that spatial resolution and ensemble methods remain decisive.