AI is beating the weatherman
A recent scientific report has highlighted how the AI-powered GraphCast system was better at predicting weather than meteorologists using conventional forecasting technologies.
The tool — developed by researchers at Google’s DeepMind — uses two most recent states of Earth’s weather, including several variables at the time of the test and six hours earlier, to predict what the weather will be like six hours later.
The scientists testing the tool found it had a 90% verification rate, better than the commonly used technologies supported by supercomputers.
Instead of the complex physics and fluid dynamics calculations, these supercomputers execute in numerical weather predictions (NWPs), it uses a machine-learning algorithm trained on historical weather patterns.
In addition to delivering a forecast faster, the scientists argued that machine learning-based weather prediction (MLWP) could improve forecast accuracy by capturing patterns in the data that are not easily represented in explicit equations.
GraphCast can also predict severe weather events, including tropical cyclones and extreme temperature waves over certain regions.
“Recently, MLWP has helped improve on NWP-based forecasting in regimes where traditional NWP is relatively weak, for example sub-seasonal heat wave prediction and precipitation nowcasting from radar images, where accurate equations and robust numerical methods are not as available,” the scientists said.
“Over the past several years, MLWP methods for medium-range forecasting trained on reanalysis data have been steadily advancing, facilitated by benchmarks such as WeatherBench.
The Wired reports that Google is considering ways to integrate GraphCast into its services.