For the first time, artificial intelligence has successfully forecasted global temperatures up to 10 days ahead, surpassing conventional modeling methods.
A peer-reviewed publication released in the journal Science on Tuesday detailed the achievement of the GraphCast AI model developed by Google DeepMind. The researchers behind the model described it as a “turning point in climate prediction.”
A comprehensive evaluation demonstrated that GraphCast outperformed the European Centre for Medium-range Weather Prediction’s leading forecasting system for the three to ten-day forecast window.
In 90% of the 1,380 metrics evaluated—such as temperature, pressure, wind speed and direction, and humidity at different atmospheric levels—GraphCast exhibited superior performance compared to the ECMWF solution.
Matthew Chantry, the machine learning coordinator at ECMWF, acknowledged the rapid and remarkable progress of AI systems in meteorology, stating that advancements have exceeded expectations from just a couple of years ago.
Apart from utilizing its proprietary forecasting system, ECMWF, an international entity headquartered in Reading, UK, has been leveraging AI models from Huawei, Nvidia, and DeepMind for weather predictions.
Chantry concurred with DeepMind’s claim regarding the precision of their method, noting that GraphCast consistently outperformed other machine learning models like Pangu-Climate from Huawei and FourCoASTNet from Nvidia, as well as ECMWF’s in-house modeling system.
GraphCast leverages a graph neural network architecture that leverages over four decades of historical ECMWF data to comprehend the evolution and dissemination of weather patterns globally.
The forecasts generated by GraphCast are based on international climate observations from ECMWF and the atmospheric conditions from six hours prior. Running on a single Google TPU v4 cloud processor, GraphCast can produce instantaneous 10-day forecasts.
In contrast to the traditional quantitative weather prediction approach that relies on supercomputers and complex atmospheric physics equations, GraphCast’s data-driven methodology is more energy-efficient and faster.
Chantry highlighted the cost-effectiveness of GraphCast once trained, emphasizing a significant reduction in energy consumption, potentially up to 1,000 times less than conventional methods.
Researchers from Intel illustrated the effectiveness of GraphCast by accurately predicting the trajectory of Hurricane Lee in the northern Atlantic, allowing for better preparedness compared to traditional forecasting methods.
While GraphCast excelled in some predictions, such as Hurricane Lee, it struggled to outperform traditional models in forecasting events like the sudden intensification of Hurricane Otis off Mexico’s Pacific coast.
Looking ahead, ECMWF is exploring the integration of AI models into its numerical weather forecasting system to enhance predictions by incorporating a deeper understanding of atmospheric physics.
Similarly, the UK Met Office announced a collaboration with the Alan Turing Institute to develop a neural network for weather forecasting, aiming to combine AI capabilities with conventional models for more robust predictions, particularly in the face of climate change challenges.
