Google DeepMind Launches AI Tool to Improve Tropical Cyclone Forecasts

Google DeepMind and Google Research have launched Weather Lab, a new platform that shares experimental AI-based weather models designed to improve tropical cyclone prediction. Developed in collaboration with the U.S. National Hurricane Center (NHC), this system aims to support more accurate and earlier warnings for cyclones—potentially helping protect lives and minimize damage as the 2025 hurricane season unfolds.

A High-Stakes Challenge for AI

Tropical cyclones—also known as hurricanes or typhoons—are among the most destructive weather events, responsible for $1.4 trillion in economic losses over the past 50 years. These rotating storms are notoriously difficult to forecast because of their sensitivity to small shifts in atmospheric conditions. Improved prediction could significantly enhance disaster preparedness, evacuation planning, and emergency response.

Weather Lab’s new cyclone model, based on stochastic neural networks, is trained to predict a storm’s formation, track, intensity, size, and wind radii. It generates 50 possible scenarios up to 15 days ahead, allowing forecasters to explore a wider range of outcomes than most traditional models.

Testing with Forecasters, Partnering with NHC

DeepMind’s cyclone model is being tested in partnership with the NHC, which manages cyclone forecasts for the Atlantic and East Pacific regions. Expert forecasters are now viewing AI-generated predictions alongside physics-based models to compare accuracy and usefulness in real time.

Internal evaluations show that the AI model performs as well as, and often better than, current leading physics-based models for both cyclone track and intensity. The model’s five-day track prediction was found to be, on average, 140 km closer to a storm’s actual path than forecasts from ENS—the top global ensemble model from the European Centre for Medium-Range Weather Forecasts (ECMWF). That performance matches what ENS typically achieves with 3.5-day forecasts, representing a 1.5-day gain in forecasting power.

On cyclone intensity, which has long been a challenge for AI models, DeepMind’s system also outperformed NOAA’s HAFS, a leading regional high-resolution physics-based model. Early tests show the AI model also holds its own in forecasting storm size and wind radii.

Inside Weather Lab

Weather Lab is a research-facing website that allows users to explore live and historical predictions from various AI and physics-based models. Models available include WeatherNext Graph, WeatherNext Gen, and DeepMind’s new cyclone model. The platform also includes over two years of backtested predictions and historical track data for researchers to evaluate.

Users can compare forecasts from AI and traditional models across global ocean basins—gaining insights into potential cyclone paths and intensities. While Weather Lab is not designed for official warnings, it gives meteorologists and disaster planners an advanced look at how emerging AI tools could support their work.

You can visit the Weather Lab here.

You can see graph depictions of the forecasting here.

A New Approach to Cyclone Modeling

Traditional forecasting relies on physics-based models that often trade off between scale and detail. Global models do well with storm tracking but struggle with intensity, while high-resolution regional models handle intensity better but lack full atmospheric context.

DeepMind’s approach trains a single AI system on both broad-scale atmospheric reanalysis data and a cyclone-specific dataset covering nearly 5,000 storms from the past 45 years. This combined method enables the model to predict both track and intensity with state-of-the-art accuracy—avoiding the typical trade-offs in resolution and scale.

Expert Validation and Global Collaboration

In addition to the NHC, DeepMind has been working with researchers at the Cooperative Institute for Research in the Atmosphere (CIRA) at Colorado State University. CIRA scientist Dr. Kate Musgrave praised the model’s performance, saying it had “comparable or greater skill than the best operational models” for both track and intensity.

The project also includes collaborations with the UK Met Office, the University of Tokyo, and Weathernews Inc. in Japan, among others, as part of a broader effort to improve global forecasting capacity.

What This Means

By developing an AI model that can accurately predict both cyclone tracks and intensities, Google DeepMind is tackling one of meteorology’s most stubborn problems. With a lead time of up to 15 days and improved performance over traditional models, this technology could significantly change how weather agencies prepare for and respond to tropical storms.

If validated through continued real-world use, tools like Weather Lab could offer earlier warnings, better resource planning, and stronger coordination during high-risk weather events—potentially saving lives in the process.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.