Google updates Search, Gemini, and Pixel Weather with WeatherNext 2 forecasting model

What you need to know

• Google has a new WeatherNext 2 forecasting model that uses AI to predict weather events.
• WeatherNext 2 is already available for developers and researchers, and it surpasses WeatherNext in 99.9% of variables.
• The new weather model also supports user-facing services like Google Search, Gemini, and the Pixel Weather app.

Weather data plays a major role in more Google services than you may expect. Google's in-house weather forecasting contributes to Search, Gemini, Maps, and of course, the Pixel Weather app. The company spent the past few years using emerging AI technologies to improve the accuracy of weather forecasting, while being quicker and more efficient. Now, it has a new WeatherNext 2 forecasting model that already powers weather data in multiple Google apps and services, according to a blog post.

Google is calling it the company's "most advanced and efficient forecasting model." It's now eight-times faster at generating weather forecasts and has a resolution time of up to one hour. Like the WeatherNext 1 model it replaces, WeatherNext 2 generates four six-hour weather forecasts daily, with lead times of up to 15 days for future predictions.

While WeatherNext 1 models are remaining available for reference and research purposes, WeatherNext 2 is the future for Google weather forecasting. It surpasses the original on 99.9% of variables, according to Google, including temperature, wind speed and direction, precipitation, humidity, and pressure.

WeatherNext 2: Our most advanced weather forecasting model - YouTube

Watch On

WeatherNext 2 can quickly and efficiently predict weather scenarios from a single starting point, Google says. The predictions are completed in under a minute using a TPU, whereas it would typically take hours to complete on a supercomputer using physics-based models.

How Google is using AI to make weather forecasts

WeatherNext 2 uses a Functional Generative Network (FGN), which is a new AI modeling approach that replaces the Graph Neural Network (GNN) and Conditional Diffusion model used in WeatherNext 1. Google explains the FGN "injects 'noise' directly into the model architecture so the forecasts it generates remain physically realistic and interconnected."

WeatherNext 2 is widely rolling out for developers, researchers, and users. It's available in Earth Engine and BigQuery, as well as Vertex AI as part of an early access program. It's also powering upgraded weather forecasts in Search, Gemini, Pixel Weather and the Google Maps Platform's Weather API.

WeatherNext 2 will also handle Google Maps weather data in the coming weeks.