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, validation and test data can be obtained from Zenodo: 10.5281/zenodo.5196512.
A pre-print of the paper describing ClimateBench and the baseline models can be found here: https://www.essoar.org/doi/10.1002/essoar.10509765.2
The spatial, global and total NRMSE of the different baseline emulators for the years 2080-2100 against the ClimateBench task of estimating key climate variables under future scenario SSP245. The models are ranked in order of the mean of the total NRMSE across all tasks.
| ('tas', 'Spatial') | ('tas', 'Global') | ('tas', 'Total') | ('diurnal_temperature_range', 'Spatial') | ('diurnal_temperature_range', 'Global') | ('diurnal_temperature_range', 'Total') | ('pr', 'Spatial') | ('pr', 'Global') | ('pr', 'Total') | ('pr90', 'Spatial') | ('pr90', 'Global') | ('pr90', 'Total') | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Neural Network | 0.107294 | 0.0440271 | 0.327429 | 9.91735 | 1.37219 | 16.7783 | 2.1281 | 0.2093 | 3.1746 | 2.61022 | 0.345709 | 4.33876 |
| Gaussian Process | 0.109106 | 0.0738238 | 0.478225 | 9.20713 | 2.67495 | 22.5819 | 2.34092 | 0.341453 | 4.04818 | 2.5559 | 0.429154 | 4.70167 |
| Random Forest | 0.107574 | 0.0584057 | 0.399602 | 9.19503 | 2.65241 | 22.4571 | 2.52431 | 0.502126 | 5.03494 | 2.68209 | 0.543375 | 5.39896 |
The example scripts provided here require ESEm and a few other packages. It is recommended to first create a conda environment with iris or xarray::
$ conda install -c conda-forge iris
Then pip install the additional requirements:
$ pip install esem[gpflow,keras,scikit-learn] eofs