# Google DeepMind’s WeatherNext Cyclone Model Shows About a Day of Forecast-Skill Gains > Source: > Published: 2026-08-09 19:05:34.344103+00:00 # Google DeepMind’s WeatherNext Cyclone Model Shows About a Day of Forecast-Skill Gains - WeatherNext Cyclones delivered an average lead-time advantage of a day or more for track, intensity and wind-radius forecasts in evaluations covering tropical cyclones from 2023 through 2025. [[1]](https://www.nature.com/articles/s41586-026-10953-2) - The model used inputs orders of magnitude coarser than regional systems, suggesting high resolution is not always necessary for strong intensity forecasts. [[1]](https://www.nature.com/articles/s41586-026-10953-2) - NOAA’s 2025 verification report says the National Hurricane Center used AI models in real-time operations for the first time, but some systems were not consistently available for routine forecasting. [[2]](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf) - Google has released WeatherNext Cyclones code and pretrained weights, while warning that the software is experimental and does not replace official alerts or warnings. [[3]](https://github.com/google-deepmind/weathernext) Google DeepMind’s WeatherNext Cyclones model produced an average lead-time advantage of a day or more over leading operational tropical-cyclone models, according to a Nature study published August 6, 2026. The evaluation covered storms from 2023 through 2025 and measured forecasts of track, intensity and wind radii—the distance from a storm’s center to specified wind thresholds. [[1]](https://www.nature.com/articles/s41586-026-10953-2) The result is a model evaluation, not evidence that communities routinely receive an extra day of warning. WeatherNext Cyclones generated additional guidance for human forecasters, while official watches, warnings and public forecasts remained the responsibility of national meteorological agencies. [[2]](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf)[[3]](https://github.com/google-deepmind/weathernext) ## A global model aimed at the storm’s full profile WeatherNext Cyclones, or WN-C, was trained on global atmospheric analysis data and a worldwide historical tropical-cyclone database. It produces ensemble forecasts: multiple plausible future scenarios rather than one deterministic track. The paper says the system can scale to as many as 1,000 ensemble members, compared with conventional ensembles of roughly 50, improving coverage of low-probability outcomes. [[1]](https://www.nature.com/articles/s41586-026-10953-2) The model’s atmospheric inputs were orders of magnitude coarser than those used by regional hurricane models. The paper therefore challenges the assumption that high spatial resolution is a strict prerequisite for accurate intensity forecasting. It does not show that resolution is unimportant: operational forecasts also depend on initial conditions, observations, data assimilation, storm tracking and specialist regional guidance. [[1]](https://www.nature.com/articles/s41586-026-10953-2)[[2]](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf) The current National Hurricane Center model summary lists Google DeepMind’s ensemble mean at about 28 kilometers, with 50 members, and identifies track, intensity and wind radii as its forecast parameters. The center’s table also notes that public access to some listed AI models is restricted by data-provider agreements, a reminder that “open” access can differ between code, model weights, input data and operational feeds. [[4]](https://www.nhc.noaa.gov/modelsummary.shtml) ## Hurricane Melissa was the clearest operational case NOAA’s 2025 verification report says that season was the first in which the National Hurricane Center incorporated AI-based models into real-time operations. It calls Google DeepMind’s model, listed as GDMI, useful, but adds that several AI systems remained under development and were not consistently available on a timely basis for routine forecasting. [[2]](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf) Hurricane Melissa provided the strongest individual example. The NHC report says its track forecast four days before landfall passed over western Jamaica and missed the eventual path by about 11 nautical miles. It also says the center provided almost three days of advance notice that Melissa would make landfall in Jamaica as a Category 5 hurricane. [[2]](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf) Google’s retrospective account says WeatherNext predicted Melissa’s rapid intensification and Jamaican landfall five days in advance, with 80% confidence at that point and near 100% confidence three days before landfall. Those figures are Google’s description of the model’s contribution, not an independently isolated estimate of how much the model changed the NHC’s final forecast. The NHC report verifies the center’s overall performance rather than assigning the result entirely to WeatherNext. [[5]](https://deepmind.google/blog/how-weathernext-helped-the-national-hurricane-center-better-predict-hurricane-melissas-historic-landfall-in-jamaica/) ## The release is public, but not a plug-and-play forecast service Google’s WeatherNext repository now includes code and pretrained weights for WeatherNext 2 and WeatherNext Cyclones, including checkpoints used to reproduce the paper’s evaluations. The repository describes the software as research code, warns that its interface may change and says the models are not officially supported Google products. [[3]](https://github.com/google-deepmind/weathernext) The release does not remove the practical barriers to independent use. Full training requires large weather datasets, including ECMWF’s ERA5 data, whose terms are separate from Google’s software licenses. The repository says the larger models require substantial accelerator memory, while a smaller cyclone version is intended for lower-resource testing. [[3]](https://github.com/google-deepmind/weathernext) That distinction matters for operational adoption. A model can be available for researchers to run while the data, hardware, update cycle and quality-control procedures needed for a national forecast center remain specialized. NOAA’s own verification material describes AI as additional guidance used alongside physics-based models, satellites, hurricane-hunter observations and human forecasters. [[2]](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf) The Nature result is therefore best read as evidence of meaningful forecast-skill progress, not as proof that AI has replaced conventional hurricane prediction. Its next test is repeated operational performance across more storms, basins and forecast cycles—including cases in which the model’s ensemble spread is wide or its intensity guidance fails. ## Companies mentioned ## Further sources [[1] Ferran Alet et al., “Operational Tropical Cyclone Forecasting with AI,” Nature,… ↗](https://www.nature.com/articles/s41586-026-10953-2) [[2] National Hurricane Center, “2025 NHC Verification Report Preview.” The report c… ↗](https://www.nhc.noaa.gov/pdf/NHC_Verification_Report_2025_Preview.pdf) [[3] Google DeepMind WeatherNext GitHub repository. The repository lists WeatherNext… ↗](https://github.com/google-deepmind/weathernext) [[4] National Hurricane Center, “Model Summary.” The page lists the Google DeepMind … ↗](https://www.nhc.noaa.gov/modelsummary.shtml) [[5] Google DeepMind, “How WeatherNext helped the National Hurricane Center better p… ↗](https://deepmind.google/blog/how-weathernext-helped-the-national-hurricane-center-better-predict-hurricane-melissas-historic-landfall-in-jamaica/) The stories that matter, in one email. Free — unsubscribe anytime.