AI model from DeepMind extends cyclone prediction lead time by a full day
DeepMind has developed an artificial intelligence model that can forecast cyclones three days in advance with the same accuracy that previous models achieved only two days ahead. This extra day of warning could significantly improve preparedness and reduce casualties from severe storms.
The WeatherNext system operates on a grid with cells measuring 28 square kilometres, relying on roughly 20 terabytes of atmospheric data and records from 5,000 past storms. Running on Google's custom chips, it produces a 15-day outlook in under a minute, whereas physics-based simulations typically require days of supercomputer time.
This development extends DeepMind's sequence of weather advances, following earlier tools for short-term rain prediction and ten-day forecasts. A meteorologist not involved in the work notes that the pace of change is so swift that academic publications trail the current state of the art by about a year and a half.
An extra day of accurate cyclone warning could significantly improve evacuation planning and resource allocation for coastal communities and emergency services. This extended lead time may reduce casualties and property damage by allowing residents in vulnerable areas to move to safety earlier. Additionally, faster and cheaper forecasting could make advanced warning systems more accessible to developing nations that lack massive supercomputing infrastructure, potentially broadening global protection against severe storms.