WeatherNext 2: Google’s updated model accelerated weather forecasts

Google has unveiled WeatherNext 2, a new AI weather forecast model that generates data eight times faster than the previous version and delivers four to six-hour forecasts per day with a horizon of up to 15 days.

Google has updated its weather forecasting system and introduced WeatherNext 2, which works in the Pixel Weather app, Google search, Gemini services for Android, and the Weather API of the Google Maps platform. In the coming weeks, WeatherNext 2 will take over the forecasts inside Google Maps as well.

The company notes that the new model generates a forecast in a minute by using the Tensor Processing Unit, Google’s specialized ASIC chip for machine learning tasks. The same amount of data on a traditional physics-based supercomputer model is calculated in about an hour.

WeatherNext 2 generates four six-hour forecasts per day and supports predictions for up to 15 days. Google says the model is more accurate than WeatherNext 1 at calculating 99% of weather parameters: temperature, precipitation, pressure, humidity, wind speed and direction.

WeatherNext 2 generates four forecasts six hours a day and supports predictions for days.

The new system is based on the Functional Generative Network method. It replaces the combination of Graph Neural Network and Conditional Diffusion used in WeatherNext 1. The previous version remains available for research applications, but the current model is already being rolled out to engineers and developers.

WeatherNext 2 takes into account the interconnectedness of weather variables described by equations and looks for stable, repeating patterns instead of classical physical modeling. This allows you to quickly create hundreds or thousands of forecasts, including a 15-day forecast that is generated in less than a minute.