Google’s new hurricane model was breathtakingly good this season (arstechnica.com)

Google DeepMind’s Weather Lab — which only began releasing Atlantic cyclone track forecasts in June — dramatically outperformed traditional models this hurricane season, according to preliminary analysis by University of Miami researcher Brian McNoldy. Using mean position error across 13 named storms at forecast lead times from 0–120 hours, DeepMind’s model (GDMI) was the most accurate at nearly every forecast hour. At the 5‑day mark the difference was striking: GDMI’s average error was 165 nautical miles versus 360 nautical miles for the U.S. Global Forecast System (GFS/AVNI), and GDMI also beat the National Hurricane Center’s official forecasts and consensus products (TVCN, HCCA). This result is significant because it shows a machine‑learned forecasting system can surpass a decades‑old physics‑based, supercomputer‑driven operational model and even human forecasters’ consensus, suggesting AI can materially improve track guidance and early warning accuracy. Key technical context: the metric was mean track error (position) across the season’s storms, and the analysis is preliminary — official National Hurricane Center model comparisons are pending and the sample is one season (13 storms). If replicated, though, the finding points toward faster, potentially more accurate hybrid or AI-first operational forecasting, better resource allocation for warnings and evacuations, and renewed focus on integrating ML models into operational pipelines while rigorously validating robustness and intensity forecasting.