Google’s AI-driven weather forecasting has achieved remarkable accuracy, outpacing traditional physics-based models in both speed and efficiency. While Google’s explorations in AI weather predictions have largely been experimental until now, the company is transitioning these capabilities into a core feature of its products and services.

The latest AI model, WeatherNext 2, is a significant leap forward, offering predictions eight times faster than its predecessor with improved accuracy for 99.9% of variables such as temperature and wind. Remarkably, it can generate numerous potential scenarios from a single input, completing forecasts in under a minute with Google’s TPU chips, a task that would otherwise take hours on a supercomputer using traditional methods.

Conventional weather models are demanding on computational resources as they attempt to simulate the complex physics of the atmosphere. In contrast, AI models like WeatherNext 2 analyze historical weather data to identify patterns and anticipate future conditions.

Google has optimized its forecasting process with a new approach called a Functional Generative Network (FGN) in WeatherNext 2. Unlike earlier models that required multiple processing cycles for a single forecast, FGN introduces targeted randomness, allowing the model to produce a variety of potential outcomes in one efficient step.

Google was able to streamline its process by using a strategy it calls a Functional Generative Network (FGN) in WeatherNext 2. Older AI weather models still required repeated processing to generate one forecast. FGN is more efficient because it incorporates noise — or targeted randomness — into the model each time it’s provided an input so that WeatherNext 2 can generate many different possible outcomes in a single step.

The advancements allow WeatherNext 2 to make predictions up to 15 days in advance and generate hourly forecasts. Google’s banking on that appealing to enterprise customers as well as individual consumers.

“We found that energy, agriculture, transportation, logistics, and customers in many other industries are quite interested in these one-hour steps. It helps them make more precise decisions relating to things that affect their business,” Akib Uddin, a product manager at Google Research, said on the call.

Aside from adding WeatherNext 2 to Maps, Search, Gemini, and Pixel Weather, Google is also offering an early access program for customers interested in custom modeling. The forecast data is also available in Google Earth Engine for geospatial analysis and BigQuery for large-scale data analysis.