What you need to know
- Google has unveiled its latest innovation, the WeatherNext 2 forecasting model, which leverages artificial intelligence to enhance weather predictions.
- This new model is now accessible to developers and researchers, boasting improvements over its predecessor, WeatherNext, in an impressive 99.9% of variables.
- WeatherNext 2 not only serves developers but also enriches user-facing services such as Google Search, Gemini, and the Pixel Weather app.
Weather data is integral to a variety of Google services, often in ways that may surprise many. The company’s proprietary weather forecasting capabilities enhance platforms like Search, Gemini, Maps, and, of course, the Pixel Weather app. Over the past few years, Google has harnessed cutting-edge AI technologies to refine the precision and efficiency of its weather forecasts. The result is the newly launched WeatherNext 2, which is already operational across multiple Google applications, as detailed in a recent blog post.
Dubbed the “most advanced and efficient forecasting model” by Google, WeatherNext 2 boasts an impressive speed, generating forecasts eight times faster than before, with a resolution time of up to one hour. Similar to its predecessor, it continues to produce four six-hour forecasts daily, extending lead times for future predictions up to 15 days.
According to Google, WeatherNext 2 can swiftly and accurately predict weather scenarios from a single starting point. Utilizing a Tensor Processing Unit (TPU), these predictions are completed in under a minute, a stark contrast to the hours required by traditional supercomputers employing physics-based models.
How Google is using AI to make weather forecasts
At the heart of WeatherNext 2 is a Functional Generative Network (FGN), a novel AI modeling technique that replaces the previous Graph Neural Network (GNN) and Conditional Diffusion model utilized in WeatherNext 1. Google describes the FGN as capable of injecting “noise” directly into the model architecture, ensuring that the forecasts produced remain both physically realistic and interconnected.
The rollout of WeatherNext 2 is extensive, making it available to developers, researchers, and users alike. It can be accessed through platforms such as Earth Engine and BigQuery, as well as Vertex AI as part of an early access program. Additionally, it enhances weather forecasts across Google Search, Gemini, the Pixel Weather app, and the Weather API within the Google Maps Platform. In the coming weeks, WeatherNext 2 will also manage weather data for Google Maps, further solidifying its role in the company’s suite of services.
