TL;DR
- Google DeepMind published its WeatherNext results in Nature, its AI model for tropical cyclones.
- WeatherNext’s 3 day forecast matches the accuracy that previously only 2 day models could achieve.
- WeatherNext anticipated the rapid intensification of Hurricane Melissa (2025) and gave extra warning time for Jamaica.
- DeepMind is releasing the code for WeatherNext 2 and WeatherNext Cyclones to researchers worldwide.
- The model generates 1,000 scenarios per cyclone, twenty times more than the 50 from the previous year.
- It was trained on nearly 20 terabytes of atmospheric data plus the IBTrACS historical database, covering almost 5,000 storms.
- WeatherNext Cyclones works with 28×28 km grid cells, a hundred times coarser than traditional high resolution models.
- A complete 15 day forecast takes less than a minute to generate on a TPU.
What WeatherNext Is
⏱️ Lectura: 10 min
WeatherNext is the artificial intelligence model Google DeepMind uses to forecast global weather, and in particular tropical cyclones: hurricanes in the Atlantic, typhoons in the Pacific, and cyclones in the Indian Ocean. Unlike traditional physical models, which solve fluid dynamics equations on supercomputers for hours, WeatherNext is a neural network trained on decades of atmospheric data.
📑 En este artículo
The announcement, published on DeepMind’s official blog and backed by a peer reviewed paper in Nature, marks the first time a single model accurately predicts the trajectory, intensity, and wind structure of a cyclone at state of the art precision. Previously, those three variables required separate systems.
What Happened
The WeatherNext team combined researchers from Google DeepMind and Google Research with forecasters from the U.S. National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and weather agencies from other countries. The paper’s central finding: WeatherNext’s 3 day forecasts match the accuracy that previously only 2 day models could achieve. That leap, according to DeepMind, is equivalent to a decade of progress in meteorology.
The most cited case from the announcement is Hurricane Melissa. During the 2025 hurricane season, WeatherNext predicted the storm’s rapid intensification well in advance, along with its path toward Jamaica. That early warning gave the NHC time to issue a historic advisory, and gave teams on the ground extra time to prepare.
Beyond publishing the results, DeepMind decided to open source the code for WeatherNext 2 and WeatherNext Cyclones, the models it used during the hurricane season. The idea is for local meteorologists, universities, and agencies without a supercomputing budget to be able to run forecasts of similar quality.
Context and History
Tropical cyclones have caused more than 700,000 deaths and 1.4 trillion dollars in economic losses over the past 50 years, according to figures cited by DeepMind. For Latin America this isn’t an abstract figure: the Caribbean, Central America, and the Gulf of Mexico get hit every year by cyclones that directly affect Cuba, the Dominican Republic, Mexico, and much of Central America.
Predicting a cyclone has always forced a tradeoff. Trajectory depends on global atmospheric currents, which are better modeled with low resolution systems that cover the whole planet. Intensity, on the other hand, depends on highly local thermodynamic processes around the storm’s eye, which require ultra high resolution regional models like HWRF. Until now, no system handled both well at the same time.
WeatherNext didn’t come out of nowhere. It’s the latest generation of a family of climate models from DeepMind that had already shown a neural network could match physical systems in global forecasting. WeatherNext Cyclones is the first version of that family specialized in cyclones, trained to close the gap between trajectory and intensity in a single model.
Technical Details and Performance
WeatherNext Cyclones was trained end to end on two distinct data sources: nearly 20 terabytes of global atmospheric dynamics and the IBTrACS database, the historical archive documenting almost 5,000 storms. That joint training is what lets it learn both large scale atmospheric patterns and the specific behavior of each cyclone.
To generate its probabilistic forecasts, the model uses Functional Generative Networks (FGN), an architecture designed to efficiently produce many variants of the same forecast. Each variant represents a possible scenario: where the storm might turn, how much it might intensify, how wide its hurricane force wind radius might be.
Last year, the system generated 50 scenarios per run, on par with global physical models. This year it scaled up to 1,000 scenarios per cyclone, letting it capture rare but decisive events, like the kind of explosive intensification Melissa underwent. A full 15 day forecast, with all 1,000 members, is generated in under a minute on a TPU.
The most striking figure in the paper is resolution: WeatherNext Cyclones works with grid cells of just 28×28 km, a hundred times coarser than the resolution the meteorological community assumed was necessary to properly model a cyclone’s intensity. A lighter version, WeatherNext 2-mini, runs at 111×111 km and also performs well. Why such a coarse model predicts with this level of precision remains, according to the authors themselves, an open research question.
💭 Key point: a model with 28 km cells competing with much finer resolution physical systems challenges a decades old assumption in meteorology: that a cyclone’s intensity can only be captured at high local resolution.
flowchart TD
A["Global atmospheric data + IBTrACS"] --> B["WeatherNext Cyclones (FGN)"]
B --> C["Ensemble of 1,000 scenarios"]
C --> D["Trajectory, intensity, and wind"]
D --> E["National Hurricane Center (NHC)"]DeepMind evaluated the model against historical cyclones from 2023 and 2024, comparing its deterministic and probabilistic performance with the best operational reference models. On average, WeatherNext Cyclones beats them by more than 24 hours of lead time in both trajectory and intensity:
| Model | What It Measures | Resolution | WeatherNext Cyclones Advantage |
|---|---|---|---|
| ECMWF-ENS | Cyclone trajectory | Global, coarse | More than 24 extra hours of lead time on average |
| HWRF | Cyclone intensity | Regional, high resolution | Equivalent accuracy with 100x less resolution |
| WeatherNext 2-mini | Lightweight global forecast | 111×111 km | Good performance with a fraction of the compute |
How to Start Trying It
DeepMind released the code for WeatherNext 2 and WeatherNext Cyclones so researchers and weather agencies can integrate it into their own workflows. The technical starting point is the same as for any modern climate model: forecasts are distributed as data in Zarr or NetCDF format, designed to be read with xarray directly from cloud storage, without downloading the full file.
# Linux / macOS
python3 -m venv weathernext-env
source weathernext-env/bin/activate
pip install xarray zarr gcsfs numpy matplotlib
# Windows (PowerShell)
python -m venv weathernext-env
weathernext-env\Scripts\Activate.ps1
pip install xarray zarr gcsfs numpy matplotlibThis environment is enough to open any forecast output in Zarr format, whether from WeatherNext or another climate model published under the same standard, and plot it with matplotlib.
Once installed, the typical flow is to open the forecast dataset and pull out the variables of interest: trajectory, minimum pressure, and wind speed per ensemble member.
import xarray as xr
import numpy as np
# replace with the path or bucket documented in the official repository
forecast = xr.open_zarr("gs://YOUR_BUCKET/weathernext-cyclones/melissa_2025.zarr")
max_wind_speed = forecast["max_wind_speed"] # (member, time, lat, lon)
hurricane_threshold = 33 # m/s, category 1 on the Saffir-Simpson scale
# probability of hurricane force winds per cell, over the 1000 members
hurricane_prob = (max_wind_speed >= hurricane_threshold).mean(dim="member")
print(hurricane_prob.max().values)That last calculation is exactly the kind of map a forecaster uses: instead of a single trajectory line, a probability at each point on the map of experiencing hurricane category winds, calculated across the ensemble’s 1,000 scenarios.
💡 Tip: to verify you’re reading the full ensemble, check that the dataset’s member dimension has size 1000, not 1 or 50; if your download has fewer, you’re using a trimmed run.
Impact and Analysis for Latin America
WeatherNext being open matters especially for regions with smaller budgets for meteorological supercomputing. National weather services in Caribbean and Central American countries, which face cyclones every year and don’t always have their own clusters to run high resolution models, gain access to a system that previously only agencies like the NHC or the Met Office could run.
The Melissa case illustrates the point: the advantage wasn’t just accuracy, it was time. An extra day of warning, during a rapid intensification, is the difference between evacuating in time or not. For local governments coordinating shelters and emergency logistics, that additional window is worth more than a decimal point of precision in wind speed.
There’s an important limit worth clarifying: WeatherNext is a research and decision support tool, not a replacement for official advisories. The bulletins that determine evacuations still come from national agencies like the NHC, which integrate WeatherNext alongside other models and their own expert judgment.
⚠️ Heads up: running the model yourself doesn’t make you an official hurricane warning source; public advisories still depend on national weather services.
Beyond cyclones, DeepMind mentions other possible uses for the same family of models: renewable energy planning, which depends on wind and solar radiation forecasts, and anticipating extreme weather events in general. None of those uses is as proven in the real world as the cyclone case, but the technical infrastructure is the same.
What’s Next
DeepMind, the NHC, CIRA, and the Met Office plan to keep collaborating over the coming hurricane seasons, now with the model already released and subject to scrutiny from the scientific community. There’s also a question the researchers themselves acknowledge remains unresolved: why a model trained on such coarse resolution data predicts a cyclone’s intensity better than systems a hundred times finer. Answering that could change how the next generation of climate models is designed, not just DeepMind’s.
📖 Summary on Telegram: View summary
Try it yourself: install the environment above and open a real forecast from DeepMind’s official blog to see the 1,000 scenario ensemble in action.
Frequently Asked Questions
Does WeatherNext replace traditional weather models?
No, it complements them. Agencies like the NHC integrate WeatherNext alongside physical models such as ECMWF-ENS and HWRF, and keep human expert judgment in place for issuing official advisories.
Where can I read the full paper?
DeepMind published the results in a peer reviewed paper in Nature and an accessible summary on its official blog.
What’s the difference between WeatherNext Cyclones and WeatherNext 2?
WeatherNext 2 is the general purpose global weather forecasting model; WeatherNext Cyclones is the specialized variant for tropical cyclone trajectory, intensity, and wind structure, co-trained with the IBTrACS historical database.
How long does it take to generate a forecast?
A full 15 day forecast, with its 1,000 ensemble scenarios, is generated in under a minute on a TPU, according to DeepMind.
Does it work for hurricanes in the Caribbean and Mexico?
Yes. The model isn’t limited to a particular ocean basin: it predicts tropical cyclones in general, and the paper’s reference case (Melissa, 2025) was precisely a hurricane that hit the Caribbean.
Do I need my own TPU to use it?
Not necessarily. You can consume forecasts already generated by DeepMind using standard tools like xarray; training or running the full model from scratch does require accelerated compute.
References
- Google DeepMind: official announcement of WeatherNext and its cyclone forecasting results.
- Nature: journal where the peer reviewed paper with the full results was published.
- National Hurricane Center (NOAA): agency that used WeatherNext during the 2025 hurricane season, including the Melissa case.
- Google DeepMind on GitHub: repositories where the company publishes its open source climate models.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Imagen destacada: Foto de Brian McGowan en Unsplash
0 Comments