SEEDS: Emulation of Weather Forecast Ensembles with Diffusion Models

This is not an official Google product

Introduction

This repository contains the gridding library and a demo colab to use the SEEDS models in the paper https://arxiv.org/abs/2306.14066.

Model checkpoints and data

The associated data to reproduce all the results in the paper can be found in the GCS bucket gs://gresearch/seeds. In particular,

• model_checkpoints contains the 22 models in the paper in the Tensorflow SavedModel format. See the demo colab for its usage.
• data
  • gefs_forecast_2022_cubedsphere.zarr: The GEFS forecasts for 2022 regridded and processed to be used as the model input. See the demo colab for the complete processing procedure from operational GEFS forecasts.
  • era5_2022_cubedsphere.zarr: The ERA5 reanalysis for 2022 regridded as the label for evaluation.
  • climatology_cubedsphere.zarr: The static climatology as part of the model input.