WeatherNext AI Model Open-Sourced for Cyclone Forecasting

What changed

Google DeepMind, in collaboration with Google Research and international weather agencies, has open-sourced its WeatherNext AI models, specifically WeatherNext 2 and WeatherNext Cyclones. These models represent a breakthrough in cyclone forecasting, achieving state-of-the-art accuracy in predicting a cyclone's track, intensity, and wind structure. The key advancement is an improvement in predictive accuracy equivalent to an extra day of warning time; three-day forecasts from WeatherNext are comparable to two-day forecasts from prior models. This leap in performance is attributed to a novel approach that combines global weather dynamics with expert-curated historical cyclone observations, trained end-to-end on approximately 20 terabytes of global atmospheric data and the IBTrACS database. Unlike traditional models that often require a trade-off between global track prediction and localized intensity modeling, WeatherNext functions as a single AI model. It utilizes Functional Generative Networks (FGNs) to efficiently generate large ensembles of predictions, up to 1,000 members, which capture inherent weather uncertainties and rare but consequential scenarios like rapid intensification. Notably, WeatherNext achieves high accuracy even with relatively low-resolution input data (28x28km), a departure from methods that relied on very high spatial resolution for intensity forecasts. A compact version, WeatherNext 2-mini, is also available and can run on a single TPU.

Why it matters for builders

By open-sourcing WeatherNext, Google DeepMind is providing AI builders and researchers with powerful, state-of-the-art models for weather forecasting. This release allows developers to integrate advanced cyclone prediction capabilities into their own applications, research projects, or operational systems. The availability of model weights and code facilitates experimentation, adaptation, and the development of specialized, localized weather models, accelerating progress in meteorological AI and its application for public safety and resilience.

Practical impact

AI builders can now access and experiment with the WeatherNext Cyclones and WeatherNext 2 models. The code and model weights are freely available, enabling direct use in research or development. A compact version, WeatherNext 2-mini, is accessible via a public Colab notebook, offering a low-barrier entry point for testing. Furthermore, the Weather Lab platform has been updated to visualize these predictions, including temperature, precipitation, and wind speed, alongside cyclone tracks. This provides a practical environment for understanding and showcasing the capabilities of the open-sourced models. Builders are encouraged to explore these resources to enhance weather forecasting tools, contribute to climate resilience efforts, or develop new applications leveraging AI for natural disaster preparedness.

Caveats and source limits

The source material indicates that while WeatherNext achieves high accuracy at lower resolutions, the exact mechanisms behind this performance remain an open research question. The full implications and potential limitations of using these models at different scales or in diverse meteorological conditions are not detailed. Pricing for commercial use or specific hardware requirements beyond the mention of TPUs for ensemble generation are not provided. The source also notes that for official weather forecasts and warnings, users should refer to their local meteorological agency or national weather service.