Google DeepMind WeatherNext Adds 24 Hours to Cyclone Warnings

Google DeepMind has published research in Nature showing that its WeatherNext AI model sets a new standard for predicting tropical cyclones — storms responsible for more than 700,000 deaths and $1.4 trillion in economic losses over the past 50 years. The headline result: WeatherNext’s three-day forecasts are as accurate as what the best previous models could only achieve at two days, translating to a full 24-hour gain in effective warning time. DeepMind describes that margin as roughly equivalent to a decade of incremental meteorological progress.

Alongside the paper, the company is open-sourcing three model variants — WeatherNext Cyclones, WeatherNext 2 and the compact WeatherNext 2-mini — under an Apache 2.0 license. The smallest version can run on a single TPU inside a free Google Colab notebook, meaning students with no institutional computing budget can experiment with a state-of-the-art cyclone model today.

What Makes WeatherNext Different

Cyclone forecasting has long forced meteorologists into an uncomfortable trade-off. A storm’s track — where it travels — is governed by large-scale atmospheric currents that coarse global models handle well. Its intensity, however, is driven by fine-grained thermodynamic processes near the storm’s core, which traditionally required high-resolution local models to capture. The result was two separate modeling pipelines that didn’t always agree.

WeatherNext collapses both tasks into a single model. It was trained end-to-end on nearly 20 terabytes of global atmospheric data and the IBTrACS historical database, which covers nearly 5,000 storms. Benchmarked against top models on cyclones from 2023 and 2024, WeatherNext Cyclones outperformed competitors across track, intensity and wind structure simultaneously.

One technically surprising detail: the full model achieves its results using input data at only 28×28 km resolution — roughly 100 times coarser than what traditional intensity models require. A single 15-day forecast runs in under a minute on a TPU. DeepMind describes the resolution finding as an open research question, explicitly inviting the broader scientific community to help explain it.

“[I]t remains an open research question to fully understand how our models produce such accurate predictions at this resolution. We hope that, together with the research community, we can find out,” WeatherNext team at Google DeepMind wrote.

The ensemble approach also scaled significantly this year. In 2024, the system ran 50 forecast scenarios per storm. For the 2025 Atlantic hurricane season, that expanded to 1,000 scenarios per cyclone, powered by Functional Generative Networks that can rapidly generate probabilistic forecasts capturing rare events like rapid intensification. That capability was tested in the real world when the National Hurricane Center used WeatherNext to issue an advance warning for Hurricane Melissa’s rapid intensification and landfall in Jamaica — a forecast the NHC described as historic.

A Crowded but Still-Evolving Field

WeatherNext enters an increasingly competitive space. Microsoft published Aurora in Nature in May 2025, a general-purpose AI weather model pretrained on more than a million hours of geophysical data. ECMWF’s Artificial Intelligence Forecasting System has been operational since February 2025 and already outperforms physics-based models on tropical cyclone tracks. Microsoft also released a cyclone-focused fine-tune, Aurora 1.5, which achieved roughly one-third lower track error compared to the base Aurora model on 2024–2025 storms.

WeatherNext’s claim of a full 24-hour lead-time advantage is a notably larger margin than what competing systems have reported, though DeepMind’s primary source does not include a direct head-to-head comparison with ECMWF’s AIFS specifically on cyclone intensity. It’s also worth noting a persistent limitation across nearly all AI weather models, including WeatherNext: they tend to underperform traditional numerical weather prediction during record-breaking or historically unprecedented events, the very scenarios where accuracy matters most. That’s a known blind spot the field is still working through.

What does stand out in the WeatherNext release is Google’s open-source commitment. While some AI model providers have moved toward commercial licensing restrictions, Google is releasing full code and weights freely — a choice that has real consequences for who can use the technology.

Why This Matters for Students

For students in atmospheric science, climate science, or AI and machine learning, this release offers an unusually direct entry point. The WeatherNext 2-mini variant runs in a free Colab notebook — no university HPC cluster required. Google DeepMind’s Nature paper defines an explicit open research question about why high accuracy is possible at 28 km resolution, which is a ready-made thesis or capstone problem. And because WeatherNext was operationally deployed during the 2025 hurricane season and directly influenced an NHC forecast, working with it carries genuine portfolio weight.

There’s also a broader access angle that students interested in international development or humanitarian tech should consider. Running legacy systems like ECMWF HRES historically required supercomputer infrastructure costing hundreds of millions of dollars. A model of comparable or superior accuracy that runs on affordable cloud compute means national weather agencies in lower-income countries can now access world-class forecast quality for the first time. That’s a meaningful shift in global disaster preparedness — and an area where researchers and engineers are still needed.

For those considering careers in climate tech, AI-for-science, or disaster risk reduction, hands-on familiarity with a peer-reviewed, operationally deployed model is a concrete differentiator. The Colab notebook is the starting point.

Source: Google DeepMind