Google Research's SEEDS Uses Diffusion Models to Generate Weather Forecast Ensembles

Google Research has published SEEDS (Scalable Ensemble Envelope Diffusion Sampler), a generative AI model that can efficiently generate ensembles of weather forecasts at a fraction of the cost of traditional physics-based methods. The work was published in Science Advances.

Weather is inherently unpredictable—small differences in initial conditions can lead to vastly different outcomes, a phenomenon known as the butterfly effect. To account for this, weather agencies generate ensembles of forecasts, each with slightly different starting conditions. But running these physics-based simulations requires massive supercomputers, limiting operational ensembles to just 10-50 members—too small to accurately predict rare but high-impact events like heatwaves or hurricanes.

"Generating these probabilistic forecasts is computationally costly," Google noted. "They require running highly complex numerical weather models on massive supercomputers multiple times."

SEEDS takes a different approach. It's a denoising diffusion probabilistic model—one of the same generative AI techniques behind image and video generation—that learns to generate plausible weather patterns. Given just one or two forecasts from an operational system, SEEDS can generate hundreds or thousands of ensemble members.

The results: SEEDS matches or exceeds physics-based ensembles on key metrics like rank histogram, RMSE, and continuous ranked probability score. Most importantly, it more accurately assigns probabilities to extreme weather events (±2σ and ±3σ). During the 2022 European heatwaves, SEEDS captured an extreme event that the operational U.S. ensemble missed entirely.

The computational cost is negligible compared to traditional methods—SEEDS can generate 256 ensemble members in just 3 minutes on Google Cloud TPUv3-32, versus hours on supercomputers.