In a groundbreaking development for numerical weather prediction, researchers have introduced SEEDS (Scalable Ensemble Emulation via Diffusion), a novel generative AI model that significantly accelerates the creation of weather ensembles. Published in a recent paper and detailed in a blog post authored by Lizao Li, Rob Carver, Ignacio Lopez-Gomez, Fei Sha, John Anderson, and Carla Bromberg, the model demonstrates the potential to transform how meteorologists assess extreme weather events and quantify climate risks.
SEEDS leverages the power of diffusion models, a type of generative AI, to produce ensemble forecasts that are comparable to those from the operational U.S. forecast system, but at a fraction of the computational cost. The model requires only two seeding forecasts from a physics-based model to generate hundreds or even thousands of additional forecasts. This hybrid approach enables a more efficient allocation of computational resources, potentially allowing for higher resolution models or more frequent forecasts.
One of the most compelling aspects of SEEDS is its ability to accurately capture rare and extreme weather events. During the extreme heat event in Lisbon on July 14, 2022, the U.S. operational ensemble failed to predict conditions as severe as those observed, with the event probability estimated to be below 1% based on a Gaussian kernel density estimate. In contrast, SEEDS, by generating 16,384-member ensembles, provided much better statistical coverage of the event, enabling the quantification of its probability and the sampling of weather regimes under which it occurred.
The implications of SEEDS extend beyond immediate forecasting improvements. The model's capability to generate large ensembles of climate projections positions it as a powerful tool for climate risk assessment, where accurate uncertainty quantification is essential. The research team envisions SEEDS as a catalyst for further advancements in operational numerical weather prediction, with applications that could reshape the field in the coming years.
"We believe that SEEDS represents just one of the many ways that AI will accelerate progress in operational numerical weather prediction in coming years," said Carla Bromberg, Program Lead for the project. "We hope this demonstration of the utility of generative AI for weather forecast emulation and post-processing will spur its application in research areas such as climate risk assessment."
The SEEDS project was supported by valuable contributions from colleagues at Google Research, including Leonardo Zepeda-Núñez, Zhong Yi Wan, Stephan Rasp, Stephan Hoyer, and Tapio Schneider, as well as technical program management and data coordination support from Tyler Russell, Alex Merose, and others. The team also thanks Tom Small for designing the animation that visually illustrates the model's capabilities.
This innovative approach to weather forecasting underscores the growing synergy between artificial intelligence and meteorology, offering a glimpse into a future where AI-driven models like SEEDS will be integral to understanding and predicting the complexities of our climate system.
Source: Google Research Blog
