The organization claims that its AI-driven weather system is swifter and more precise than it has ever been.
The organization claims that its AI-driven weather system is swifter and more precise than it has ever been.
Google unveiled a revamped AI weather model today that will be integrated into widely used applications such as Search, Gemini, and Pixel smartphones.
The company’s AI-powered forecasts have demonstrated notable accuracy thus far. AI weather systems have also been shown to make predictions more swiftly and efficiently compared to traditional physics-driven models. Until now, Google’s efforts in this domain have largely been experimental. Now, it is transforming those predictions into a marketable asset for Google products and services.
“We’re removing it from the laboratory and actively placing it in the hands of users in more ways than before, getting rid of the experimental label because we believe our forecasts are highly effective and truly beneficial,” expressed Peter Battaglia, senior director of research and sustainability at Google DeepMind, during a briefing with journalists.
The new AI model, WeatherNext 2, can produce forecasts eight times quicker than Google’s earlier system and is also superior in accuracy for predicting 99.9 percent of factors like temperature or wind. WeatherNext 2 is capable of generating hundreds of potential scenarios from a single starting point. Making a prediction takes under a minute with one of Google’s TPU chips, an operation that the company claims would ordinarily require several hours to achieve using traditional physics-based models on a supercomputer.
These conventional models are computation-heavy since they essentially aim to replicate the complex physics of the atmosphere to create forecasts. In contrast, AI models endeavor to extract patterns from historical weather data to forecast future results.
Google streamlined its approach by utilizing a method it refers to as a Functional Generative Network (FGN) in WeatherNext 2. Previous AI weather systems still required repetitive processing to generate a single forecast. FGN operates more efficiently as it introduces noise — or targeted randomness — into the model each time it receives an input, enabling WeatherNext 2 to produce multiple potential results in one go.
These improvements enable WeatherNext 2 to issue forecasts up to 15 days ahead and provide hourly predictions. Google anticipates that this will appeal to business clients as well as everyday consumers.
“We discovered that sectors such as energy, agriculture, transportation, logistics, and numerous other industries are quite keen on these hourly increments. It aids them in making more accurate decisions related to aspects influencing their businesses,” noted Akib Uddin, a product manager at Google Research, during the call.
In addition to integrating WeatherNext 2 into Maps, Search, Gemini, and Pixel Weather, Google is launching an early access program for clients interested in customized modeling. The forecast information is also accessible in Google Earth Engine for spatial analysis and BigQuery for extensive data analysis.
Google faces competition, naturally, in attempting to determine how to leverage generative AI in weather forecasting. The European Center for Medium-Range Weather Forecasts, Nvidia, Huawei, and various other firms are concurrently developing their own AI weather systems.
