Google AI Introduces SEEDS: A Generative AI Model that Advances Medium-Range Weather Forecasting

Google AI introduced SEEDS to handle the challenge of generating accurate and computationally efficient ensemble weather forecasts. Current methods depend on physics-based simulation, which is computationally intensive and limits the scale of forecast ensembles, particularly for rare and extreme weather events. The unpredictable nature of weather makes it crucial to quantify uncertainty in forecasts, especially as climate change increases the demand for reliable weather information.

Traditionally, weather forecasts are generated using physics-based models that simulate the atmosphere’s behavior. However, these models are computationally expensive, limiting the scale of forecast ensembles and hindering the accurate characterization of maximum events. To address this, the Google researchers propose SEEDS, a generative AI model based on denoising diffusion probabilistic models. SEEDS efficiently generates large ensembles of weather forecasts at a fraction of the fee of traditional methods, enabling higher quantification of uncertainty and more accurate prediction of maximum events.

SEEDS leverages generative AI technology to provide ensemble forecasts that match or exceed the skill metrics of physics-based ensembles. It can efficiently generate ensembles conditioned on just one or two forecasts from an operational numerical weather prediction system. The generated ensembles accurately capture spatial covariance and correlations between atmospheric variables, providing more realistic forecasts. Moreover, SEEDS significantly reduces computational costs in comparison with traditional methods, with a throughput of 256 ensemble members per 3 minutes on Google Cloud TPUv3-32 instances. This scalability enables the generation of huge ensembles vital for assessing the likelihood of rare but high-impact weather events.

The performance of SEEDS is demonstrated through comparisons with operational weather prediction systems and Gaussian models. SEEDS outperforms Gaussian models in capturing spatial correlations and accurately predicting extreme weather events. For example, through the 2022 European heat waves, SEEDS generated forecasts with spatial structures much like operational forecasts, whereas Gaussian models did not capture cross-field correlations. Additionally, SEEDS provides higher statistical coverage of maximum events, enabling the quantification of their probability and sampling of weather regimes under which they might occur.

In conclusion, the paper presents SEEDS as a promising solution to the challenges of ensemble weather forecasting. By leveraging generative AI technology, SEEDS enables the efficient generation of huge ensembles that accurately quantify uncertainty and predict extreme events. This state-of-the-art model could completely change operational numerical weather prediction, giving people making decisions in lots of areas, from emergency management to energy trading, essential data.

Check out the Paper and Google Blog. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our Telegram Channel, Discord Channel, and LinkedIn Group.

Don’t Forget to affix our 39k+ ML SubReddit