On Thursday, Google DeepMind and Google Research unveiled a public preview of Weather Lab, an interactive platform dedicated to presenting the company’s artificial intelligence (AI) weather models along with forecasts derived from their outputs. The Mountain View-based tech company also introduced its latest experimental AI-driven model for predicting tropical cyclones, designed to forecast aspects such as formation, track, intensity, size, and shape with an advance lead time of up to 15 days. It is important to note that the scientific validation for this AI model is still pending.
Google Releases New AI Model to Predict Cyclones
In a blog post, DeepMind detailed the debut of the Weather Lab platform and explained the functionalities of its new cyclone-oriented AI model. The website features both real-time and historical cyclone predictions utilizing AI-generated models alongside physics-based forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF).
Google DeepMind mentioned that multiple AI models, including WeatherNext Graph, WeatherNext Gen, and the new cyclone model, operate in real-time to analyze meteorological data and generate forecasts. Moreover, Weather Lab hosts over two years of historical AI-generated predictions that researchers can access for model evaluation.
The platform also facilitates comparisons between forecasts produced by different AI models and traditional physics-based approaches. However, the company has stated that the website is intended for research purposes and should not be seen as a source for official weather warnings.
Regarding its AI cyclone model, Google has made a pre-print version of its associated research paper available, although it has yet to undergo peer review. For its scientific validation, Google is collaborating with the US National Hurricane Center (NHC).
DeepMind pointed out the conventional cyclone prediction method, which employs two distinct physics-based models: a global low-resolution model for forecasting tracks based on atmospheric steering currents and a regional high-resolution model for assessing intensity through analysis of turbulent processes near the cyclone’s core.
This new AI model aims to address the challenges posed by this dual-approach strategy by integrating both track and intensity predictions. According to the blog post, the model has been trained using a comprehensive “reanalysis dataset” that assembles past weather data from millions of global observations, alongside a specialized database detailing attributes such as track, intensity, size, and wind radii of nearly 5,000 observed cyclones over the last 45 years.
In a recent test, DeepMind indicated that the AI model was deployed in the North Atlantic and East Pacific basins during 2023-24. The model’s five-day cyclone track predictions were reportedly, on average, 140 kilometers closer to the actual location than predictions made by the ECMWF’s ENS model. Furthermore, internal assessments suggest that the performance of the cyclone model is at least comparable to that of established physics-based models.
