Google has made the code and model weights behind its WeatherNext AI weather prediction system publicly available on GitHub, allowing scientists around the world to build on the technology. The move follows a study published in the journal Nature detailing how the model works and what it can do.
According to Engadget, WeatherNext was developed with contributions from the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, the UK Met Office and other weather agencies worldwide. The model was trained on nearly 20 terabytes of global atmospheric data, along with historical records from the International Best Track Archive for Climate Stewardship.
Predicting tropical cyclones has long presented a specific technical challenge. A storm's track across the ocean is best analyzed using broad global atmospheric models, while its intensity depends on thermodynamic processes deep inside the storm's core that require specialized local models to assess. Historically, those two tasks have required separate systems. WeatherNext handles both with a single model.
"We can now generate a single 15-day forecast in less than a minute on a TPU, empowering forecasters to quickly evaluate the probability distribution of potentially devastating tail-risks," the WeatherNext researchers wrote.
Speed matters in hurricane forecasting. The faster a model can generate a forecast, the more time emergency managers have to issue warnings, order evacuations and position resources. A 15-day window also gives forecasters more lead time than traditional models typically provide.
Google introduced the second generation of WeatherNext last year. The research teams have also worked separately on using AI to predict flash floods. By releasing WeatherNext as open source, Google is making it possible for other research groups and weather agencies to study the model, test it against their own data and potentially improve on it.
