A day before landfall, emergency managers can still do concrete things: open shelters, stage rescue vehicles, suspend port operations, and move vulnerable residents first. That is the practical claim behind Google DeepMind’s WeatherNext Cyclones release on August 6.
The model was published with a Nature paper and open-sourced code and weights. Google DeepMind says its three-day tropical cyclone forecasts for track, intensity, and wind structure now match the accuracy earlier models delivered at two days, giving forecasters about 24 extra hours of useful lead time.
Track and intensity in one model
Cyclone forecasting has long split the problem. Large-scale circulation steers the path, while intensity depends on smaller-scale heat, moisture, and ocean-atmosphere processes near the core. WeatherNext Cyclones is trained across both global weather dynamics and expert-curated cyclone records.
The official post cites nearly 20TB of global atmospheric data and almost 5,000 historical storms from IBTrACS. The system can roll forecasts forward up to 15 days, then extract cyclone track, intensity, and wind-radius information.
“Time is really golden when it comes to those types of decisions.”
That quote from National Hurricane Center director Mike Brennan captures why a one-day improvement matters. Evacuations and supply staging are time-bound decisions, not abstract benchmark wins.
Low resolution is the surprise
The model’s unusual claim is that it predicts intensity using 28-by-28 km inputs, about 100 times coarser than traditional high-resolution approaches. A mini version still performs well at 111-by-111 km resolution. That suggests large-scale signals may carry more information about rapid intensification than researchers previously understood.
The caveat is just as important. Google’s repository says the models do not replace official alerts or warnings, and WIRED’s interviews stress that human forecasters still translate model output into impact-based decisions.
From 50 scenarios to 1,000
WeatherNext now generates 1,000 possible scenarios for each cyclone, up from 50 last year. Google says a 15-day forecast can run in under a minute on a single TPU, allowing forecasters to inspect tail risks rather than a single best track.
The 2025 Hurricane Melissa case is the operational signal. WIRED reports that the model forecast a Category 5 Jamaica landfall five days ahead with 80% confidence, helping the NHC issue earlier warnings. One storm does not settle the next season, but it shows the model has left pure backtesting.
Open source moves the test local
The repository lists WeatherNext 2, WeatherNext Cyclones checkpoints, and a smaller Cyclones Mini model, with code under Apache 2.0 and other materials under CC BY 4.0. For typhoon regions in East Asia, the next question is local validation across basins, coastlines, terrain, and warning systems.
The metrics to watch are independent basin errors, rapid-intensification hit rates, and whether forecasters using the model actually change evacuation or resource timing. A better forecast score is the first step; public action is the harder last mile.
Sources: Google DeepMind, Google Blog, Nature, WIRED, GitHub WeatherNext, CocoLoop; verification covers the release timing, Nature publication, three-day versus two-day forecast framing, 24-hour lead-time gain, 20TB atmospheric data, nearly 5,000 storms, 1,000 ensemble members, 28-by-28 km input resolution, 15-day forecasts, and open-source licensing.