Weather forecasting has greatly benefited from advancements in technology over the past century, but many believe there’s still room for improvement. Knowing what to expect can be the difference between life and death in instances ofsevere weather. For this reason, experts are searching for ways to make more accurate predictions, and AI is proving to be a helpful tool. Now, Google’s DeepMind — the company’s AI research laboratory — has developed a machine-learning model (MLM) that is demonstrating just how helpful it can be.
A study published inScienceon November 14 is shedding light on GraphCast, the MLM created by DeepMind to develop more reliable weather forecasts. To make GraphCast as accurate as possible, the MLM was trained to reanalyze previous weather data to make its predictions. Currently, many experts rely on numerical weather prediction (NWP) to create their forecasts. While there’s nothing inherently wrong with this, it is an undeniably slower approach than leveraging AI to analyze historical weather data and make similar predictions.
DeepMind started training GraphCast by using global weather data from between 1979 and 2017. This helped it understand variables, such as wind, temperature, and air pressure. In its current form, GraphCast now uses weather estimates from 6 hours ago and the “current” state of global weather to make predictions. In testing, DeepMind gave GraphCast global weather estimates from 2018 to create a forecast up to 10 days ahead. Researchers discovered that its predictions were more accurate than what came from the High RESolution forecasting system (HRES), an NWP variation. On top of that, it only took seconds to generate, compared to the hours required by the HRES. When researching GraphCast’s reliability in predicting severe weather, such as tropical cyclones, DeepMind saw promising results as well. However, researchers acknowledge that the MLM still needs more testing to determine how other metrics could impact varying results.
As Google’s AI research laboratory makes strides, the company is continuing to develop its approach to weather forecasting across the board. In June 2023, for example, itbegan leveraging datafrom the National Oceanic and Atmospheric Administration’s MRMS and HRRR sensor systems for Google Weather. With this information, users of the Weather app receive more accurate 12-hour predictions. The company also developed a feature called Nowcast, specifically created to hone in on severe weather. The data in Nowcast is up-to-the-minute, meaning users can rely on it for pressing weather information as it develops.
You might only think of the forecast as helpful for avoiding passing showers, but Google knows it can do more than inform you when to grab the umbrella. By generating more accurate predictions, MLMs like GraphCast might become some of the most valuable AI tools we have. Currently, Google Weather is available on recent Pixel devices, including thePixel 8 Proand thePixel Fold.
