Particle.news

Download on the App Store

DeepMind’s AI Weather Lab Tops Short-Range Forecasts for Hurricane Erin

The experimental cyclone model is under NHC review, with any operational shift dependent on proven gains in the decisive three- to five-day window.

Image
Google's Weather Lab model beats NHC on Erin's track

Overview

  • Post-storm charts compiled by former NHC unit chief James Franklin show the GDMI model beat the NHC official track and several physics-based and hurricane-specific models through 72 hours.
  • The model also led on intensity forecasts for the first 72 hours, with particularly strong accuracy at the 48-hour lead time.
  • Erin’s rapid intensification to Category 5 provided the model’s first major real-time trial, with impacts reported in Bermuda, parts of the Caribbean, and along the U.S. East Coast.
  • DeepMind trained the system on a global reanalysis dataset and a specialized database of nearly 5,000 observed cyclones, taking a data-driven approach distinct from numerical weather prediction.
  • Google reports 2023–2024 backtesting with five-day tracks averaging about 140 km closer than the ECMWF ensemble, while experts note the small sample and the need to improve 3–5 day performance.