Google study – Ars Technica

Google study – Ars Technica

Zoom in / Archive image of Tropical Storm Fiona as seen in a satellite image from 2022.

On Tuesday, the peer-reviewed journal Science published a study showing how Google DeepMind’s AI-based meteorological model called GraphCast significantly outperformed traditional weather forecasting methods in predicting global weather conditions up to 10 days in advance. This achievement indicates that predicting future weather may become much more accurate, the Washington Post and Financial Times reported.

In the study, GraphCast showed superior performance over the world’s leading conventional system, run by the European Center for Medium-Range Weather Forecasts (ECMWF). In a comprehensive evaluation, GraphCast outperformed the ECMWF system on 90 percent of 1,380 metrics, including temperature, pressure, wind speed and direction, and humidity at various levels of the atmosphere.

And GraphCast does all this quickly: “It forecasts hundreds of weather variables, over a 10-day period with an accuracy of 0.25 degrees globally, in less than one minute,” the authors write in the paper “Learning Skilled Medium-Term Global Weather Forecasting.”

This represents a significant advance in speed and accuracy for artificial intelligence in meteorology. Matthew Chantry, machine learning coordinator at ECMWF, acknowledged the rapid progress in an interview with the Financial Times, saying the AI ​​system in meteorology has advanced “much faster and impressively than we would have expected even two years ago.”

GraphCast uses what researchers call a “graph neural network” machine learning architecture, trained on more than four decades of historical ECMWF weather data. It processes current six-hour global atmospheric states, generating a 10-day forecast in about one minute on a Google TPU v4 cloud computer. Google’s machine learning method contrasts with traditional numerical weather prediction methods that rely on supercomputers to process equations based on atmospheric physics, consuming much more time and energy.

A selection of cool graphs from Google DeepMind research titled,
Zoom in / A selection of great graphs from a Google DeepMind search titled “Skillful learning for medium-term global weather forecasting.”

Google DeepMind

Chantry highlighted GraphCast’s efficiency to the Financial Times, estimating that it is about 1,000 times cheaper in terms of energy consumption than traditional methods. An example of its forecasting success is that Hurricane Lee was predicted to make landfall in Nova Scotia nine days earlier, three days earlier than traditional methods.

Despite the advances, GraphCast has limitations. It did not outperform conventional models in all scenarios, such as the sudden intensification of Hurricane Otis, which hit Acapulco with minimal warning on October 25. Also, due to technological limitations, global AI models are not yet able to generate predictions as detailed or detailed as they are. Traditional methods, making them more ideal for studying smaller-scale phenomena, according to The Washington Post. And they have transparency issues because meteorologists can’t yet look inside the “black box” of an AI model and know exactly why it’s making the predictions.

Ultimately, Google DeepMind researchers see their AI-based approach as complementary to existing weather forecasting techniques. “Our approach should not be considered a substitute for traditional weather forecasting methods, which have been developed for decades, rigorously tested in many real-world contexts, and offer many advantages that we have not yet explored,” they wrote.

Looking to the future, ECMWF plans to develop its own artificial intelligence model and explore its integration with a numerical weather prediction system. The UK Met Office, in partnership with the Alan Turing Institute, is also developing a graph neural network for weather forecasting to be integrated into future supercomputer infrastructure.

You may also like...

Leave a Reply

Your email address will not be published. Required fields are marked *