How Google's DeepMind tool is 'more quickly' forecasting hurricane behavior (www.theguardian.com)

When Tropical Storm Melissa began rapidly intensifying, National Hurricane Center forecaster Philippe Papin cited Google DeepMind’s new hurricane model—released in June—as a primary reason for his unusually bold forecast that the storm would strengthen to Category 4–5 and turn toward Jamaica. Melissa did strengthen to Category 5 on landfall, and across 13 Atlantic storms this season DeepMind’s model outperformed traditional physics-based models and human track forecasts, often producing ensemble-based guidance (Papin noted roughly 40/50 ensemble members showing Cat 5) much faster than legacy systems. Technically, DeepMind is a machine‑learning model that detects patterns physics-based models can miss, delivering forecasts in minutes on modest hardware instead of hours on supercomputers. That speed and lower compute cost make it attractive for operational use, but the model is effectively a black box: top-line outputs are public in real time, yet internal methods remain largely hidden, prompting calls from forecasters for more interpretability and diagnostics. DeepMind has occasional intensity misses (e.g., Hurricane Erin, Typhoon Kalmaegi), but its strong season highlights a shift toward AI-driven forecasting—parallel efforts by government and startups aim to extend these gains into sub‑seasonal outlooks, tornado/flash‑flood warnings, and improved observational networks.