WeatherNext: AI model achieves breakthrough in forecasting cyclones

What it claims

A single model, WeatherNext Cyclones, forecasts tropical cyclone track, intensity and wind structure together, as published in Nature. Its three-day forecasts match the accuracy prior models reached at two days, which DeepMind equates to roughly a decade of meteorological progress. The usual approach splits this work in two: track is steered by large-scale circulation and handled by coarse global models, while intensity depends on fine-scale processes near the core and is handled by high-resolution regional models. This model does both at 28×28 km, which the page says is 100× coarser than traditional models. The authors say they do not fully understand why it works at this resolution.

How it works

• Co-training on two data sources: about 20 TB of global atmospheric data, plus the IBTrACS historical storm database, which covers nearly 5,000 storms. Training is end-to-end.
• Ensembles from Functional Generative Networks (FGNs). Ensemble size grew from 50 members last season, which matched global physics models, to 1,000 members this year. The larger ensemble is meant to capture rare tail events such as rapid intensification.
• Speed: one 15-day forecast takes under a minute on a TPU. The page states this per forecast. It does not say how long a full 1,000-member ensemble takes, so the existing summary's "1,000-member ensemble in under a minute" goes further than the page.

Evidence

• The page gives only headline results: three-day accuracy equals the earlier two-day accuracy, averaged across track, intensity and structure. Per-metric errors and the named baselines are in the Nature paper, not this post.
• Operational use: during the 2025 season the model helped the National Hurricane Center (NHC) forecast Hurricane Melissa's rapid intensification and its landfall in Jamaica, and the NHC issued an advance warning.
• Collaborators were the NHC, CIRA and the UK Met Office.

How to use it

Three models are released, with code and weights:

To see forecasts without running anything, Weather Lab shows cyclone tracks and global fields such as temperature, precipitation and wind. Both Weather Lab and the models are part of Google Earth AI.

Limits

• No licence is named for the code or weights. "Freely available" is the only wording, so check the repository terms before any operational or commercial use.
• The page gives no repository link, install command or hardware needs beyond TPU and Colab.
• The good results at coarse resolution are explicitly unexplained, an open research question.
• WeatherNext 2 and 2-mini are not the models evaluated in the paper. Their cyclone skill relative to WeatherNext Cyclones is not reported here.
• The page states that WeatherNext is not an official forecast source; local meteorological agencies remain the authority.