Diffusion Distillation for Efficient Weather Ensembles
Machine Learning
Summary
The authors worked on making weather forecasting faster by teaching a simpler model to imitate a complex diffusion model that normally takes many steps to make predictions. They created a new way to train this simpler model so it learns directly from both the complex model's outputs and real weather data. Their experiments showed that this simpler model still predicts weather well, even for extreme events like typhoons, but with much less computing effort. Overall, the simpler model worked as well or better than the original complex one on important tests.
Authors
Yiming Yang, Valentin Brekke, James Briant, Serge Guillas
Abstract
Diffusion models generate skillful weather ensembles but require costly iterative sampling. We introduce a supervised energy-distance distillation method that compresses a multi-step diffusion teacher into a single-step student by aligning student forecasts with teacher samples and ground-truth observations. Experiments on global forecasting and typhoon-track prediction show that our student outperforms existing distillation methods and preserves skill for extreme events. It matches or surpasses the teacher across key metrics using only one neural function evaluation per autoregressive step.