ClimateBench is a benchmark dataset for climate model emulation inspired by WeatherBench. It consists of NorESM2 simulation outputs with associated forcing data processed in to a consistent format from a variety of experiments performed for CMIP6. Multiple ensemble members are included where available.
The processed training and validation data can be obtained from Zenodo: 10.5281/zenodo.5196512.
This benchmark dataset is currently being used for a hackathon during the NOAA AI workshop. Test data used for evaluation of these submissions will be released upon its conclusion.
| Model | TAS RMSE (2050 / 2100) [K] | DTR RMSE (2050 / 2100) [K] | Pr RMSE (2050 / 2100) [mm/day] | P90 RMSE (2050 / 2100) [mm/day] |
|---|---|---|---|---|
| Baseline GP | 0.32 / 0.41 | 0.14 / 0.15 | 0.42 / 0.62 | 1.29 / 1.82 |
| UNet | a / b | a / b | a /b | a /b |
| Random Forest | a / b | a/ b | a /b | a /b |
| UKESM | 1.71 / 2.70 | 1.17 / 1.34 | 0.59 / 0.82 | 1.77 / 2.48 |
| (Variability) | 0.53 / 0.59 | 0.21 / 0.22 | 0.74 / 0.86 | 2.26 / 2.55 |
The example scripts provided here require ESEm and a few other packages. It is recommended to first create a conda environment with iris::
$ conda install -c conda-forge iris
Then pip install the additional requirements:
$ pip install esem[gpflow,keras,scikit-learn] eofs