Google’s DeepMind has unveiled a weather prediction model that leverages artificial intelligence (AI) to outperform traditional methods for forecasts extending up to 15 days.
The new model, known as GenCast, excels in predicting extreme weather events, reported Financial Times.
GenCast employs probabilistic forecasting to assess multiple scenarios, enabling precise predictions ranging from wind energy production to the movement of tropical cyclones.
Financial Times' report cited findings from a research paper published in Nature on Wednesday, stating that the probabilistic capabilities of GenCast mark a significant milestone in the rapid advancement of AI-powered weather forecasting. This approach is increasingly being adopted by leading traditional forecasters.
Ilan Price, a research scientist at Google DeepMind, described it as a turning point in AI’s role in weather prediction. “Modern raw forecasts now originate from machine learning models,” he said. He added that GenCast could be integrated into operational weather forecasting systems, offering valuable insights to help decision-makers better understand and prepare for upcoming weather events.
The key advantage of GenCast over earlier machine learning models lies in its use of “ensemble” forecasts, which consider multiple outcomes—a technique employed in advanced traditional forecasting. GenCast is trained on four decades of data from the European Centre for Medium-Range Weather Forecasts (ECMWF).
In tests, the model outperformed ECMWF’s forecasts for 15-day predictions in 97.2% of 1,320 variables, such as temperature, wind speed, and humidity. These results represent an improvement in accuracy and scope over Google DeepMind’s previous model, GraphCast, unveiled last year.
GraphCast outperformed ECMWF in approximately 90% of metrics for 3-to-10-day forecasts.
AI-powered forecasting models are generally faster and more efficient than standard methods, which rely on substantial computational power to analyze equations derived from atmospheric physics. GenCast can generate forecasts in just eight minutes, compared to hours for traditional methods, using a fraction of the computational resources.
Researchers noted that GenCast could still be enhanced in areas like predicting the intensity of major storms. Its data accuracy can also be improved to match ECMWF’s upgrades implemented this year.
The ECMWF acknowledged that GenCast’s development represents a “significant milestone” in weather forecasting evolution. However, they emphasized that the innovative machine learning science behind GenCast still requires testing for extreme weather events.
AI Growing Role
The development of GenCast adds to the ongoing debate about AI’s growing role in forecasting. Many scientists favor a hybrid approach for certain purposes.
In July, Google introduced NeuralGCM, a model combining machine learning with traditional physics to achieve better results for long-term forecasting and climate trends.
Stephen Ramsdale, a leading AI forecasting expert, emphasized the importance of a hybrid approach, saying, “The greatest value comes from combining human evaluation with traditional physics-based models and AI-driven weather forecasting."
