WeatherNext Cyclones forecasts track and intensity in one AI model

Google DeepMind on August 6 open-sourced WeatherNext Cyclones, an AI weather model that produces tropical-cyclone track, intensity and wind-structure forecasts in one system. The Nature paper accompanying the release reports that, in tests on storms from 2023 through 2025, the model provided an average lead-time advantage of at least a day over leading operational models across those measures.

The result is relevant beyond another weather-model benchmark. Cyclone forecasters have typically used different tools for two linked but technically different tasks: estimating where a storm will travel and estimating how powerful it will become. WeatherNext Cyclones is designed to provide probabilistic guidance for both, along with the reach of a storm's winds, from the same model.

How does WeatherNext Cyclones forecast track and intensity together?

Storm tracks are influenced by broad atmospheric currents, a problem traditionally suited to global models. Intensity depends more on localized processes near a cyclone's core, which has pushed forecasters toward specialized, higher-resolution regional models. Google DeepMind says WeatherNext Cyclones bridges that division by forecasting global weather patterns and cyclone characteristics together.

The system produces an ensemble, meaning it runs many possible weather and storm outcomes rather than presenting a single certain path. According to the Nature paper, it can generate up to 1,000 global weather and cyclone scenarios extending 15 days ahead. That output can be used to estimate the likelihood of different track, strength and wind-radius outcomes, including less common scenarios.

Google says the model was co-trained on nearly 20 terabytes of global atmospheric data and a historical cyclone database covering nearly 5,000 storms. The research was authored by teams from Google DeepMind and Google Research alongside contributors affiliated with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office.

What does the reported day of lead time mean?

Google characterizes the gain as a three-day forecast that can match the accuracy earlier systems offered at a two-day horizon. That is a description of the reported retrospective evaluation, not a guarantee for every storm or an official warning.

The Nature abstract says WeatherNext Cyclones' track, intensity and wind-radii predictions had an average advantage of a day or more over leading operational models on the 2023-2025 test set. It also reports that adding the model's predictions to a weighted consensus ensemble substantially improved skill. The paper positions the system as guidance for human forecasters.

Another notable finding is the resolution of the input data. The authors report results using inputs orders of magnitude coarser than those used by regional models. They say the finding suggests high spatial resolution is not a strict requirement for state-of-the-art intensity forecasting, rather than establishing that resolution no longer matters.

Google announced that WeatherNext Cyclones and WeatherNext 2 would be open-sourced. For forecasting organizations and researchers, the release makes the underlying model available for scrutiny and experimentation. The core evidence so far is a peer-reviewed evaluation; official cyclone forecasts and warnings remain a forecaster-led process that combines guidance from multiple models.

This story draws on original reporting from The Decoder.