The Longer Time Goes [On], the Worse You're Going

DeepMind's WeatherNext AI model can predict cyclones three days ahead with accuracy that existing models only achieve two days out, potentially saving lives by improving evacuation decisions. The model generates 15-day forecasts in under a minute on Google's Tensor Processing Unit chips, compared with days of supercomputer time for physics-based models. It uses a coarser atmospheric grid (28 square kilometers per cell, 100 times less granular) but was trained on nearly 20 terabytes of global data and 5,000 historical storms, which DeepMind's Ferran Alet says was key to its accuracy.

In tests against Google's GenCast, the European Centre for Medium-Range Weather Forecasts' ENS, and NOAA's Hurricane Analysis and Forecast System, WeatherNext matched their two-day accuracy at three days for maximum wind speed and location error. "The longer time goes [on], the worse you're going to do, because weather is chaotic, Alet said. DeepMind is already building faster, more accurate models.

Hannah Cloke at the University of Reading called AI's impact on meteorology "one of the most exciting fields to work in, noting advances outpace journals by 18 months. However, she stressed the need for meteorological expertise to validate AI outputs. Tim Palmer at the University of Oxford cautioned that AI may fail on unprecedented events lacking training analogues and proposed testing models by withholding such events from training data.