Google opens WeatherNext, gaining an extra day to warn of cyclones

Google DeepMind and Google Research will open-source their WeatherNext cyclone forecasting model after a Nature paper showed it delivers roughly one additional day of predictive lead time over prior systems. The team's three-day cyclone forecasts now match the accuracy of older two-day forecasts, a jump the authors equate to "a decade's worth of meteorological progress."

Why an extra day matters

Tropical cyclones — hurricanes in the Atlantic, typhoons in the Pacific — have killed more than 700,000 people and caused $1.4 trillion in economic losses over the past 50 years, according to figures cited in the paper. Each additional hour of reliable warning changes evacuation logistics, shelter staffing, and grid preparation. The leap from a two-day to a three-day forecast at the same accuracy level compresses hours of uncertainty into actionable lead time.

The work brought together AI researchers at Google DeepMind and Google Research with operational forecasters at the US National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office.

How WeatherNext performed on Hurricane Melissa

The model is no longer a research curiosity. During the 2025 Atlantic hurricane season, WeatherNext helped the NHC forecast Hurricane Melissa, predicting the storm's rapid intensification and Jamaica landfall before either signal was certain. That prediction fed into an advance warning that gave response teams on the ground critical preparation time.

For the 2026 season, Google scaled its cyclone ensemble from 50 members last year to 1,000 scenarios per storm. Larger ensembles surface rare but consequential outcomes — rapid intensification events like Melissa's — that smaller samples can miss.

What makes the model different

Cyclone forecasting has long forced a trade-off. A storm's track is steered by planetary-scale atmospheric currents, best captured by coarse global models. A storm's intensity depends on fine-scale thermodynamic processes around its core, best captured by high-resolution regional models. WeatherNext is a single AI system that handles both.

The architecture rests on three pillars:

• Co-training on two data modalities. The model was trained end-to-end on roughly 20 terabytes of global atmospheric data and the IBTrACS historical cyclone database, which spans nearly 5,000 named storms.
• Functional Generative Networks (FGNs). The architecture produces large ensembles efficiently, capturing the uncertainty inherent to atmospheric prediction.
• Low-resolution inputs. Despite conventional wisdom that intensity forecasting demands fine grids, WeatherNext Cyclones operates at 28×28 km — about 100 times coarser than traditional regional models. A lighter variant, WeatherNext 2-mini, runs at 111×111 km and still performs well.

"This has surprised scientists," the team wrote of the resolution finding, "and it remains an open research question to fully understand how our models produce such accurate predictions at this resolution."

How fast does it run?

A single 15-day global forecast completes in under one minute on a Google TPU. The team first generated 50 forecasts at a time, matching the scale of physics-based ensembles; this year, they pushed that to 1,000 forecasts per cycle. Each member of the ensemble captures a plausible storm trajectory, allowing forecasters to read the probability distribution of tail risks — the rare scenarios where a Category 2 storm becomes a Category 5 overnight.

What is being open-sourced

Google is releasing three models alongside the paper:

• WeatherNext 2, the current operational version that replaced the original in October.
• WeatherNext Cyclones, the configuration that ran through the 2025 hurricane season and underwrote the Nature results.
• WeatherNext 2-mini, a compact variant that runs on a single TPU through a free public Colab notebook.

The release ships code and weights, enabling academic, operational, or nonprofit groups to fine-tune the model for regional or domain-specific work. Google is also refreshing its Weather Lab interface to show temperature, precipitation, and wind predictions alongside cyclone tracks. Both products sit inside Google Earth AI.

Where the limits are

The model is trained on historical observations, including the IBTrACS archive. Storms in a warmer climate may exceed the intensity distribution of that training set. Google does not position WeatherNext as a replacement for national weather services; the paper and product page direct readers to local meteorological agencies for official warnings.

What competitors and peers are doing

WeatherNext enters a field that already hosts AI weather systems from ECMWF, NVIDIA's FourCastNet family, Microsoft-backed research, and the deep-learning ensembles operated by insurers and energy traders. The differentiator Google emphasizes is co-training with the operational forecaster community and a public-weight release, rather than a closed API. Open weights let national agencies in lower-resource countries run the model in-house, provided they have the TPUs.

The road ahead

Google framed the open-source release as an invitation. The company wants research groups, meteorological agencies, and nonprofits to build on the weights, stress-test the resolution paradox, and extend the forecasts to heat waves, floods, and renewable-energy planning. If the 28×28 km finding holds up under outside scrutiny, it could reset how the field thinks about input resolution — pushing it down the list of model design priorities and elevating data quality and ensemble diversity instead.

The 2025 season gave Google a real-world case study in the form of Hurricane Melissa. The 2026 season — already running — will be the first where 1,000-member ensembles and open weights are both in play at the same time. Forecasters, emergency managers, and rival AI labs will now get to grade the work.