Google DeepMind published a model on Nature that predicts trajectory, intensity, and wind structure of tropical cyclones better than classical meteorological methods. WeatherNext’s three-day forecasts match the accuracy that previous systems achieved at two days: an extra day of warning, which DeepMind compares to a decade of progress in meteorology.

Why it matters. AI is entering domains where classical physics worked alone, and it’s doing so with measurable results on a life-or-death problem. During the 2025 season the model helped the National Hurricane Center predict rapid intensification of Hurricane Melissa and its landfall in Jamaica, enabling early alerts. The pattern recurs in stories of applied AI that count: the model gives meteorologists more time to decide. DeepMind releases the weights of WeatherNext 2 and WeatherNext Cyclones as open source.

If you want to try it. The open models are on DeepMind’s blog. For those working in forecasting domains, the detail to study is the ensemble of 1,000 scenarios per cyclone that WeatherNext generates to support forecasters’ decisions.

In depth

Traditional weather forecasting solves fluid dynamics equations on supercomputers: take current data (pressure, temperature, wind) and calculate how it evolves over time. The finer the grid, the more computing power needed. Tropical cyclones are hard to forecast because their intensity can change in hours, and a 100 km error on trajectory completely changes who needs to evacuate. According to data cited in the paper, cyclones have caused over 700,000 deaths and 1.4 trillion dollars in economic losses in 50 years.

WeatherNext changes the approach. Instead of solving equations step by step, the model learns from decades of historical meteorological data (reanalysis, past forecasts, satellites) to recognize patterns leading to certain developments. Training happens on reanalysis data, coherent reconstructions of atmospheric state, and the model generates forecasts that compete with numerical systems.

The central result is on three-day forecasts. WeatherNext produces three-day forecasts with the same accuracy that previous systems achieved at two days. In practical terms: if you had 48 hours before to prepare evacuation, you now have 72. DeepMind compares this jump to a decade of progress in numerical meteorology.

What distinguishes WeatherNext is collaboration with forecasters. During the 2025 hurricane season, the team worked with the National Hurricane Center, CIRA, the UK Met Office, and meteorological agencies from various countries. Hurricane Melissa is the emblematic case: WeatherNext predicted rapid intensification and landfall in Jamaica, enabling early alert.

This year the system generates 1,000 possible scenarios for each cyclone. It’s the ensemble approach applied with AI: instead of a single trajectory, the model produces a range of possible trajectories with respective probabilities. For a forecaster, this means being able to communicate the probability of different trajectories: “there’s a 70% chance it hits this area, 30% it goes elsewhere.”

DeepMind made two models open source: WeatherNext 2 (general forecasting) and WeatherNext Cyclones (specialized on cyclones). The weights are downloadable from the blog.

What remains open. The paper is published on Nature, but the blog doesn’t report benchmark numbers on standard public datasets. Details are missing on how the model performs outside hurricane season, on extratropical cyclones, and how long the advantage holds over longer time scales (5-7 days). Real impact is documented for one season and one hurricane: more time is needed to understand if the advantage holds systematically.