DeepMind AI Predicts Cyclones a Full Day Before the Best Physical Models

Google DeepMind has released WeatherNext, an AI model that predicts tropical cyclones 24 hours ahead of the best existing systems. The authors describe this as the equivalent of ten years of traditional meteorological progress in a single step. During the 2025 hurricane season, the model helped the U.S. National Hurricane Center give Jamaica timely warning of the rapid intensification of Hurricane Melissa.

The model generates 1,000 probabilistic storm scenarios in one minute — a task that previously required hours of supercomputer computation. DeepMind has open-sourced everything: the code, weights, and methodology — any national weather service in the world can now achieve U.S.-level accuracy without years of development. This is the first time an AI forecast has outperformed physical models not on a benchmark, but in a live hurricane season.