Overview

• The U.S. National Hurricane Center has begun using DeepMind’s experimental AI model alongside traditional physics-based systems in its operational forecasting workflow.
• The model generates up to 50 scenarios predicting cyclone formation, track, intensity, size, and shape as far as fifteen days in advance.
• In internal tests, its five-day track predictions averaged 140 kilometers closer to actual storm locations than ECMWF’s leading ensemble and it outperformed NOAA’s HAFS on intensity forecasts.
• Google launched Weather Lab in public preview to display live and historical AI and physics-based cyclone predictions for expert evaluation.
• Trained on global reanalysis data and a specialized database of 5,000 observed cyclones, the system remains a research tool rather than an official source of warnings.