Google DeepMind Researchers Introduce GenCast: Diffusion-based Ensemble Forecasting AI Model for Medium-Range Weather

You may have missed a big development in the ML weather forecasting revolution over the holidays: GenCast: Google DeepMind’s new generative model!  The importance of probabilistic weather forecasting cannot be overstated in various critical domains like flood forecasting, energy system planning, and transportation routing. Being able to accurately gauge the uncertainty in forecasts, especially concerning extreme events, is pivotal for making well-informed decisions that involve significant cost-benefit considerations and effective mitigation strategies.

Traditionally, the approach to probabilistic forecasting involves creating ensembles from physics-based models, which sample from a joint distribution over spatio-temporally coherent weather trajectories. However, this method can be computationally expensive. An appealing alternative is the use of machine learning (ML) forecast models to generate ensembles. Yet, the current cutting-edge ML forecast models for medium-range weather primarily focus on producing deterministic forecasts that minimize mean-squared error.

Despite the improved skill scores associated with these models, they face a limitation in terms of lacking physical consistency. This limitation becomes more pronounced at longer lead times, impacting their ability to characterize the joint distribution of weather events precisely. 

The paper introduces a novel machine learning-based approach for probabilistic weather forecasting known as GenCast. This innovative method generates global, 15-day ensemble forecasts that demonstrate superior accuracy compared to the leading operational ensemble forecast, namely the European Centre for Medium-range Weather Forecasts (ECMWF)’s ENS, all while requiring significantly less computation time. GenCast operates by implicitly modeling the joint probability distribution of the weather state over space and time. It works on a 1° latitude-longitude grid, utilizing 12-hour time steps, and represents six surface variables and six atmospheric variables at 13 vertical pressure levels.

The evaluation of GenCast’s forecasts shows that it keeps detailed patterns and consistency in weather predictions. Comparisons with ENS indicate that GenCast’s ensembles are just as reliable, if not more so. GenCast is efficient—it can create a 15-day forecast in about a minute using a Cloud TPU v4. This means generating a large number of forecasts (𝑁 ensemble members) in a short time is possible with multiple TPUs. This efficiency opens up the possibility of using much larger ensembles in the future. 

In a broader context, GenCast signifies a significant advancement in machine learning-based weather forecasting, demonstrating higher proficiency than the leading operational ensemble forecast at a 1° resolution. This development marks a pivotal step toward ushering in a new era of ensemble forecasting driven by machine learning, expanding its relevance and usefulness across a diverse array of domains. Moreover, as we look ahead, GenCast offers a glimpse into the potential of embracing machine learning to revolutionize our understanding and prediction of complex weather patterns, with far-reaching implications for various industries and decision-makers.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our 36k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel