Google DeepMind's WeatherNext AI Extends Cyclone Forecast Lead Time by 24 Hours, Now Open Source

Google DeepMind has unveiled WeatherNext Cyclones, an artificial intelligence model that significantly advances tropical cyclone forecasting by providing an average of an additional 24 hours of lead time for critical predictions. This breakthrough, detailed in a paper published in *Nature* on August 6, 2026, marks a substantial leap in meteorological capabilities, offering three-day cyclone forecasts with the accuracy previously achieved by earlier systems at only two days. The company has also made the model, alongside the broader WeatherNext 2 system, open-source, aiming to accelerate global research and disaster preparedness efforts.

1. Advancing Cyclone Prediction with AI

Tropical cyclones, encompassing hurricanes and typhoons, are among the most destructive weather phenomena globally, having caused over 700,000 deaths and an estimated $1.4 trillion in economic losses over the past five decades. Historically, accurate forecasting of these events has presented a complex challenge, with traditional numerical weather prediction (NWP) models requiring immense computational resources and time to simulate atmospheric dynamics. Google DeepMind's WeatherNext Cyclones model addresses this by leveraging AI to predict a cyclone's track, intensity, and wind structure with state-of-the-art accuracy. This enhancement translates to a "critical extra 24 hours" for communities and emergency services to prepare for an impending storm, an improvement that Google DeepMind characterizes as roughly equivalent to a decade's worth of meteorological progress.

2. Architectural Innovations and Extensive Training

WeatherNext Cyclones is built on a sophisticated AI architecture that utilizes Functional Generative Networks (FGNs). This design allows the model to produce ensemble forecasts, generating up to 1,000 possible scenarios for a given storm. This capability is crucial for identifying low-probability, high-impact events like rapid intensification, which are often challenging for deterministic models to capture.

The model's training regimen was extensive, drawing upon nearly 20 terabytes of global atmospheric data. It also incorporated the International Best Track Archive for Climate Stewardship (IBTrACS) database, which compiles records of nearly 5,000 historical storms. Despite the complexity of its training data, WeatherNext Cyclones can generate a complete 15-day forecast, covering track, intensity, and wind structure, in less than a minute on a single Tensor Processing Unit (TPU). This computational efficiency is a significant advantage over traditional physics-based models that demand more fine-grained inputs and processing power. Remarkably, the model operates effectively on a coarser input resolution of approximately 28x28 kilometers, or even 111x111 kilometers for its mini variant, a resolution a hundred times coarser than conventional models, yet maintains high accuracy. The exact reasons for this resilience to reduced resolution remain an "open research question" for the researchers.

3. Performance Benchmarks and Real-World Validation

The efficacy of WeatherNext Cyclones was rigorously evaluated against leading operational models using tropical cyclones from 2023 through 2025. The results demonstrated an average day or more of lead-time advantage for track, intensity, and wind-radius forecasts. This means a WeatherNext three-day forecast achieves the same accuracy that previous state-of-the-art systems could only deliver for two-day predictions.

The model's capabilities were also tested in real-world operational settings. During the 2025 Atlantic hurricane season, the National Hurricane Center (NHC) utilized WeatherNext Cyclones. A notable instance was its role in forecasting Hurricane Melissa, where the model accurately predicted the storm's rapid intensification and eventual landfall in Jamaica five days in advance. This successful prediction enabled the NHC to issue earlier warnings, providing critical time for on-the-ground teams to prepare and respond. The development of WeatherNext Cyclones was a collaborative effort between AI researchers and engineers at Google DeepMind and Google Research, along with expert forecasters from the NHC, the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office.

4. Open-Sourcing for Global Impact and Accessibility

To foster broader adoption and accelerate advancements in meteorological science, Google DeepMind has open-sourced the code and pretrained weights for WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini. These resources are available on GitHub under an Apache 2.0 license, allowing researchers, meteorological agencies, and developers worldwide to build upon this technology. The WeatherNext 2-mini variant is particularly accessible, as it can run on a single TPU within a free Google Colab notebook, effectively lowering the computational barrier for smaller agencies and academic institutions.

Beyond the open-source release, live outputs from WeatherNext are integrated into the refreshed Weather Lab interactive platform within Google Earth AI. Additionally, the broader WeatherNext 2 model, which offers faster and higher-resolution general weather forecasts, is now available for advanced research and geospatial analysis through Google Earth Engine and BigQuery, with early access via Google Cloud's Vertex AI.

5. Implications for Disaster Preparedness and Future Research

The extended lead time provided by WeatherNext Cyclones has profound implications for disaster preparedness. An additional 24 hours allows emergency managers more time to stage supplies, organize evacuations, and position response teams before a storm makes landfall, potentially saving lives and mitigating economic damage. This advancement represents a significant step towards building greater climate resilience worldwide.

The open-sourcing of the models is expected to galvanize further research and development in AI-driven weather forecasting. The unexplained phenomenon of WeatherNext Cyclones maintaining accuracy at a significantly coarser resolution than traditional models presents a compelling "open research question" that could lead to new insights in atmospheric modeling. By making this powerful AI technology broadly accessible, Google DeepMind aims to empower the global weather community to better predict a wide range of weather events and make more informed decisions to protect lives and infrastructure.

Frequently Asked Questions

What is WeatherNext Cyclones?

WeatherNext Cyclones is an artificial intelligence model developed by Google DeepMind and Google Research that achieves state-of-the-art accuracy in forecasting tropical cyclone track, intensity, and wind structure.

How much more lead time does WeatherNext Cyclones provide?

The model provides an average of at least one additional day (24 hours) of lead time for cyclone forecasts compared to leading operational models, effectively making three-day forecasts as accurate as previous two-day forecasts.

Has WeatherNext Cyclones been used in real-world scenarios?

Yes, it was used by the National Hurricane Center during the 2025 Atlantic hurricane season, notably helping to forecast Hurricane Melissa's rapid intensification and landfall in Jamaica.

Is the WeatherNext model available for public use or research?

Yes, Google DeepMind has open-sourced the code and model weights for WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini on GitHub under an Apache 2.0 license. Live outputs are also accessible via the Weather Lab interactive platform, and data from WeatherNext 2 is available through Google Earth Engine and BigQuery.

Sources

• Original X post by @GoogleDeepMind

* Google DeepMind says WeatherNext adds a day to cyclone forecast accuracy - MLQ.ai. (August 7, 2026). https://mlq.ai/google-deepmind-says-weathernext-adds-a-day-to-cyclone-forecast-accuracy/ * Google DeepMind Open Sources WeatherNext AI Cyclone Forecasting Model. (August 7, 2026). https://unite.ai/google-deepmind-open-sources-weathernext-ai-cyclone-forecasting-model/ * DeepMind opens WeatherNext cyclone forecasting model - Resultsense. (August 7, 2026). https://resultsense.ai/deepmind-opens-weathernext-cyclone-forecasting-model/ * WeatherNext: AI model achieves breakthrough in forecasting cyclones - Google DeepMind. (August 6, 2026). https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/ * WeatherNext 2: Our most advanced weather forecasting model - Google Blog. (November 17, 2025). https://blog.google/products/weather/weathernext-2-ai-weather-forecasting-model/ * Our WeatherNext 2 AI model demonstrated a massive leap forward in predicting cyclones. - Google Blog. (August 6, 2026). https://blog.google/technology/ai/weathernext-2-ai-model-cyclone-forecasting-breakthrough/ * Google's WeatherNext 2 Gains a Full Day of Cyclone Warning, Goes Open Source - Unite.AI. (August 6, 2026). https://unite.ai/googles-weathernext-2-gains-a-full-day-of-cyclone-warning-goes-open-source/ * WeatherNext 2 by Google DeepMind: AI Weather Forecasting Model Explained (2026). (March 23, 2026). https://www.aimodels.org/blog/weathernext-2-by-google-deepmind-ai-weather-forecasting-model-explained-2026 * Google DeepMind Says Its AI Can Now Predict Cyclones a Full Day Earlier - BigGo Finance. (August 7, 2026). https://www.biggo.finance/news/google-deepmind-says-its-ai-can-now-predict-cyclones-a-full-day-earlier-2026-08-07 * DeepMind opens WeatherNext, buys forecasters a day on cyclones | AI Weekly. (August 6, 2026). https://aiweekly.co/p/deepmind-opens-weathernext-buys-forecasters * WeatherNext 2 - Google DeepMind. https://deepmind.google/discover/blog/weathernext-2/