DeepMind's WeatherNext Cyclones adds a day of warning to tropical storm forecasts

A day of extra warning

On August 6, 2026, Google DeepMind published research in Nature introducing WeatherNext Cyclones, an AI model that forecasts a tropical cyclone's track, intensity and wind structure in a single system — and then open-sourced the whole thing. According to a dev.to writeup of the work, the model delivered roughly a full extra day of reliable warning compared with leading operational systems when benchmarked against cyclones from 2023 through 2025.

One model instead of two

Cyclone forecasting has traditionally forced meteorologists to run two separate modeling approaches. A storm's track is driven by large-scale atmospheric currents, which coarse global models capture well. A storm's intensity depends on fine-grained thermodynamic processes near the core, which historically required specialized, high-resolution local models. Running both in parallel is expensive and does not always produce a consistent picture of the storm.

WeatherNext Cyclones collapses the two problems into one autoregressive model. It was trained end-to-end on nearly 20 terabytes of global atmospheric data alongside the IBTrACS historical database, which covers close to 5,000 past storms.

Cheap ensembles via Functional Generative Networks

At the core of the system is a technique called Functional Generative Networks, which replaces the diffusion-based approach DeepMind used in its earlier GenCast model. FGNs let WeatherNext generate large ensembles cheaply — up to 1,000 simulated scenarios per forecast — helping forecasters gauge the probability of rare but dangerous events such as sudden rapid intensification.

The model operates at a resolution roughly a hundred times coarser than conventional physics-based systems, yet still matches or beats them on accuracy. As the dev.to report notes, DeepMind's own researchers describe that result as not yet fully understood. The efficiency gain is concrete: a full 15-day forecast can be generated in under a minute on a single TPU.

The benchmark numbers

At the 5-day mark, the model's average track error was around 230 kilometers, compared with roughly 370 kilometers for the ECMWF ensemble system and 335 kilometers for GenCast. A 3-day WeatherNext forecast proved about as accurate as earlier systems achieved at 2 days.

In simulated blends with official consensus forecasts, WeatherNext Cyclones improved track accuracy by about 28 percent and intensity accuracy by about 6 percent — a jump DeepMind's team compared to roughly a decade's worth of typical progress in traditional forecasting.

Tested during a live hurricane season

This was more than a benchmarking exercise. During the 2025 Atlantic hurricane season, the U.S. National Hurricane Center used the model operationally to help forecast Hurricane Melissa's rapid intensification and its landfall in Jamaica — an early signal, per dev.to, that AI forecasting tools are moving from research demos into real forecasting workflows.

Open, not proprietary

DeepMind released the code and pretrained weights for three variants — WeatherNext 2, WeatherNext Cyclones, and a lightweight WeatherNext 2-mini that runs in a free Google Colab notebook — under an Apache 2.0 license on GitHub. Live forecasts are also available through the Weather Lab platform inside Google Earth AI.

The work was developed in collaboration with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office. DeepMind frames WeatherNext as a tool to support forecasters rather than replace them; official storm warnings still come from national weather services.

Why it matters

Tropical cyclones have caused over 700,000 deaths and an estimated $1.4 trillion in economic damage worldwide over the past five decades. Against that backdrop, an extra 24 hours of warning is not a marginal number — it maps directly onto evacuation windows, supply chain decisions and, ultimately, lives.

WeatherNext is also a clear example of where AI is having its most tangible scientific impact today: incremental, compounding gains on a well-defined, high-stakes physical prediction problem rather than speculative future capability. And because the code and weights are public under a permissive license, it sets a precedent for how AI weather research can be shared openly instead of being locked behind proprietary forecasting products.