Google updates its weather forecasts with a new AI model

So far, the company’s AI-enhanced forecasts have proven to be quite accurate. AI weather models have also made forecasts faster and more efficient than conventional physics-based models. Until now, Google’s work in this area has been mostly experimental. Now these predictions become a selling point for Google products and services.

“We’re taking it out of the lab and really putting it into the hands of people in more ways than before and sort of getting rid of the experimental type of designation because we believe our predictions are really very effective and very useful,” Peter Battaglia, senior director of research and sustainability at Google DeepMind, said in a briefing with reporters.

The new AI model, WeatherNext 2, can generate forecasts eight times faster than Google’s previous model and is also more accurate in predicting 99.9% of variables such as temperature or wind. WeatherNext 2 can generate hundreds of potential results from a particular starting point. It takes less than a minute using one of Google’s TPU chips to make a prediction, which the company says would typically take several hours using physics-based models on a supercomputer.

These conventional models are computationally intensive because they essentially attempt to recreate the complex physics of the atmosphere to produce forecasts. AI models, on the other hand, attempt to discern trends from historical weather data in order to predict future outcomes.

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 a forecast. FGN is more efficient because it incorporates noise – or targeted randomness – into the model every time it provides input so that WeatherNext 2 can generate many different possible outcomes in a single step.

The advancements allow WeatherNext 2 to forecast up to 15 days in advance and generate hourly forecasts. Google is counting on this to attract business customers as well as individual consumers.

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

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