The company’s AI-enhanced forecasts turned out to be quite exact yet. AI-based weather models have also enabled faster and more effective forecasting than conventional physics-based models. So far, Google’s work in this area has been mostly experimental. Now these predictions become an advantage for Google products and services.
“We take it out of the lab and actually put it into the hands of users.”
“We’re taking it out of the lab and actually giving it back to users in more ways than before, sort of giving up the experimental designation because we have confidence that our predictions are really effective and quite useful,” Peter Battaglia, senior director of research and sustainability at Google DeepMind, said during a briefing with reporters.
The new AI model, WeatherNext 2, it can generate forecasts eight times faster than Google’s previous model, and it also predicts 99.9 percent of variables such as temperature and wind more accurately. WeatherNext 2 can generate hundreds of potential results from a given starting point. Calculating a prediction using one of Google’s TPU chips takes less than a minute, and the company says it typically takes several hours using physics-based models on a supercomputer.
These conventional models are computationally intensive because they essentially try to recreate the complicated physics of the atmosphere to make predictions. In turn, artificial intelligence models try to recognize patterns from historical weather data 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 multiple processing to generate a single forecast. FGN is more effective because it incorporates noise—that is, directed randomness—into the model every time it receives input, allowing WeatherNext 2 to generate many different possible outcomes in a single step.
With the improvements, WeatherNext 2 can create forecasts up to 15 days in advance and generate hourly forecasts. Google is betting that it will be attractive to both corporate clients and individual consumers.
“We found that energy, agriculture, transportation, logistics and customers from many other industries are quite interested in these one-hour steps. It helps them make more precise decisions about things that impact their business,” Akib Uddin, product manager at Google Research, said during the call.
In addition to adding WeatherNext 2 to Maps, Search, Gemini, and Pixel Weather, Google also offers early access program for customers interested in custom modeling. Forecast data is also available on Google Earth Engine for geospatial analyzes and BigQuery for large-scale data analysis.