Diffusion Models on WeatherBench
WeatherBench is a benchmark dataset that explores the potential of Machine Learning methods for weather forecasting. WeatherBench is comprised of ERA5 reanalysis data and covers the entire globe. Various spatial resolutions are available, the time step is 1 hour. Authors compete to predict meteorological variables as well as possible 3 and 5 days into the future.
Diffusion Models are a recently popularised class of Machine Learning models and have proven especially effective at generating images. Particularly successful examples include Stable Diffusion and DALL-E 2. Diffusion Models can also be trained to generate output conditioned on input data such as text or other images.
We employed Diffusion Models for weather forecasting: we plan to give the model the current state of atmospheric variables as conditioning information and train it to predict realistic future states.
Specifically we:
– Explored the potential of Diffusion Models on the WeatherBench challenge – which has never been done before.
– Published code and trained models to make it easy to replicate and build on our results.
Mentors
- Jesper Dramsch
- Florian Pinault