DeepMind’s AI Models Show Progress in Tropical Cyclone Forecasting

Recent developments in artificial intelligence have introduced new tools for tropical cyclone forecasting, with Google DeepMind and Google Research launching Weather Lab as a public demonstration of their experimental AI-based prediction model.

The AI model showcased in Weather Lab is designed to generate multiple forecast scenarios for cyclone formation, track, intensity, size, and shape, extending up to 15 days into the future. These forecasts are produced as ensembles, typically including 50 different possible outcomes for each event.

Internal assessments using data from the National Hurricane Center for the 2023–2024 seasons in the North Atlantic and East Pacific indicated that the AI model’s five-day track forecasts were, on average, 140 kilometers closer to the actual cyclone locations than the ECMWF’s ENS model. This performance equaled the ENS model’s accuracy at 3.5 days, effectively providing an additional 1.5 days of lead time for forecasters.

What the numbers show

• 5-day AI track forecasts averaged 140 km closer to actual cyclone locations than ECMWF ENS
• AI model provides forecasts up to 15 days ahead with 50 ensemble scenarios
• Forecasts are generated up to eight times faster than traditional systems

In addition to improved track accuracy, the AI model outperformed NOAA’s HAFS regional high-resolution physics-based model in average intensity error and achieved similar results for predicting cyclone size and wind radii. The model’s five-day forecast accuracy was also reported to match the 3.5-day performance of traditional models, which translates to approximately 36 extra hours of warning time for communities.

One case study involved the AI model WeatherNext, which predicted Hurricane Melissa’s rapid intensification from Category 1 to Category 5 and its landfall in Jamaica five days in advance. The model’s confidence in this prediction increased from 80% to nearly 100% as the event approached, marking a first for the National Hurricane Center in predicting such rapid intensification at this lead time.

According to Google DeepMind, WeatherNext’s early prediction enabled the National Hurricane Center to issue advanced warnings, giving residents in Jamaica more time to prepare and evacuate. This five-day ahead forecast of rapid intensification was noted as a milestone in the ability to predict both cyclone track and intensity changes simultaneously.

Despite these advances, the AI models remain experimental and are intended to support, rather than replace, official forecasts from meteorological agencies. Weather Lab is currently being validated in collaboration with organizations such as the U.S. National Hurricane Center, CIRA at Colorado State University, the UK Met Office, University of Tokyo, and Weathernews Inc., which are providing feedback on the system’s performance.

Following its demonstrated improvements in cyclone forecasting, the WeatherNext model was later open-sourced to facilitate further research and validation. DeepMind has stated that its AI model can produce forecasts up to eight times faster than traditional methods, supporting the ongoing evaluation of AI’s role in weather prediction.

* This article is based on publicly available information at the time of writing.