DeepMind unveils hurricane forecasting breakthrough in Nature

A significant advancement in weather prediction has been announced by Google DeepMind, which developed an artificial intelligence model capable of predicting cyclone intensity with unprecedented accuracy. The technology was recently utilised to forecast Hurricane Melissa five days before landfall, correctly identifying the system as a Category 5 hurricane despite initial models suggesting it would remain weak and impact Haiti instead. This accurate prediction allowed authorities in Jamaica to issue earlier warnings for flooding and landslides, providing communities vital time to prepare for the catastrophic event.

Researchers publishing their findings on Thursday in Nature state that the WeatherNext model provides forecasters with an average of one additional day of lead time compared to existing systems. Consequently, predictions made three days out by this new AI are as accurate as those produced two days ago by previous models. Mike Brennan, director of the US National Hurricane Center, emphasised the value of this extra window for organising evacuations and staging supplies. He noted that even a few hours can make a difference in time-sensitive decisions where wrong choices carry severe consequences.

Historically, extending forecast lead times by a day would have required a decade of work according to the research team. The development presented unique challenges because machine learning typically requires ample training data for future predictions, yet extreme weather events are rare occurrences. Ferran Alet, a research scientist at Google DeepMind and one of the paper’s lead authors, explained that while cyclone-specific data is limited, vast amounts of general weather data exist. To address this, the team trained the model to be proficient in both standard weather forecasting and specific cyclone prediction.

Kate Musgrave from the Cooperative Institute for Research in the Atmosphere highlighted that hurricanes operate at multiple spatial scales which complicates modelling. Predicting a storm’s track requires global-scale data regarding cold fronts and prevailing winds, whereas predicting intensity demands much smaller-scale information focused on local atmospheric and ocean conditions. Traditional global models often fail to capture these specific details required for intensity forecasting. While earlier AI systems performed well tracking storm paths, they struggled significantly with intensity prediction.

The ability to predict both track and intensity is critical because a change in strength can determine the difference between a weak storm and a major hurricane capable of rapid intensification overnight. Hurricane Melissa marked the first instance where the National Hurricane Centre predicted a Category 5 status when the system was still at Category 1 stage using this new technology. Before live deployment, researchers tested the model on retrospective data with results so promising that scientists were initially sceptical about its real-time performance.

Even the developers admit they do not fully understand how the AI produces such accurate predictions given it utilises much lower-resolution atmospheric data than traditional models require. Alet noted that when informing the scientific community of this coarse resolution, many were shocked to learn these inputs capture more signal regarding future events than previously believed. The model generates a range of potential scenarios rather than a single prediction, helping to account for butterfly effects where small deviations lead to larger changes downstream.

Forecasters can utilise these outputs alongside other models to inform predictions about storm development. Last year the system created 50 scenarios per storm but now generates 1,000 thanks to increased computing power. Brennan cautioned that while DeepMind’s model is a valuable new tool in forecasters’ toolbox, there are no guarantees it will be best for every season or storm and human expertise remains critical for translating forecasts into impact assessments.

Google DeepMind has announced plans to open-source the WeatherNext models used during hurricane season so researchers can utilise and improve upon them. Alet expressed excitement about scientific discovery through this collaboration, believing AI provides new tools to investigate fundamental laws of nature. The story originally appeared on wired.com.