Google's AI weather models now run cyclone forecasts on a single chip

[CosmosWizard]

Google DeepMind published a Q&A explaining how its WeatherNext AI models have dramatically improved cyclone forecasting, condensing work that once required building sized supercomputers down to something that can run on a single TPU. Research scientist Ferran Alet explained that traditional physics based forecasting relied on simulating fluid dynamics using current weather data from global sensors, a process where the meteorological community historically gained about one day of forecast accuracy improvement per decade through building better models and adding more computing power and satellites

By training AI on 50 years of historical weather data instead, WeatherNext achieved that same decade of typical progress in a single generation leap, according to DeepMind. The improvement lets weather agencies extend public forecast horizons from five days toward seven, giving communities meaningfully more advance warning before a cyclone actually makes landfall. Alet pointed to Hurricane Melissa as a concrete example, where the storm initially appeared weak, but WeatherNext's high confidence predictions helped the National Hurricane Center issue a historic warning about its rapid intensification from a Category 1 storm into a Category 5, giving Jamaican authorities an additional day to make life saving emergency preparations

Google has open sourced both WeatherNext 2 and its dedicated cyclone models, along with an interactive Weather Lab website, specifically to let the broader global scientific community experiment freely and build on the underlying research. Curious what people think about this kind of open sourcing specifically for life safety critical AI tools, does making these models freely available meaningfully improve global disaster preparedness, or does the real bottleneck lie elsewhere in how individual countries actually act on the warnings they receive

Ask a Scientist: How do researchers use AI to predict a cyclone?blog.google