Researchers at Impa (Institute for Pure and Applied Mathematics) in Rio de Janeiro have developed an artificial intelligence system called Tupann, capable of predicting rainfall up to three hours in advance using satellite data, without relying on ground-based radar. The system generates maps every ten minutes, with a resolution of about two kilometers, delivering forecasts in less than three minutes after receiving the data.The model was awarded at the international ICLR conference, held in Rio in April 2025, and was developed by doctoral students Antônio Catão, Melvin Poveda, and Leonardo Voltarelli, under the supervision of mathematician Paulo Orenstein.
The technology aims to fill radar coverage gaps, which are concentrated in wealthy countries, according to the WMO. Tested in Manaus, Miami, and La Paz, Tupann showed competitive performance compared to international models. Its authors plan to expand it to the Amazon and Africa.
Among the risks is the generation of visually plausible but physically incorrect forecasts, known as hallucinations. To mitigate them, the model was trained with physical constraints.
Climatologist Carlos Nobre from USP warns that AI systems trained on historical data may fail when facing unprecedented climate events, stating: "Artificial intelligence will not be able to see events that have never happened."
