Conclusions

The application of machine learning to the prediction of future climate states has, perhaps justifiably due to the challenges laid out above, been cautious to date. Particular applications however, with carefully chosen training data and objectives, can provide fruitful avenues for research and open exciting opportunities for improvement over the current state-of-the-art. This paper introduces the ClimateBench dataset in order to galvanise existing research in this area, provide a standard objective with which to compare approaches and also introduce new researchers to the challenge of climate emulation. It provides a diverse set of training data with clear objectives and challenging target variables, some of which have been extensively studied (surface air temperature) and some which have been somewhat neglected (diurnal temperature range and precipitation).
Current impact assessments are often based on simple emulators, which are then scaled to match modelled patterns, but which are unable to predict non-linear responses in e.g. precipitation. A robust, trustworthy emulator which is able to provide such predictions could be immensely valuable in quantifying and understanding the changes and associated risks of different socio-economic pathways. Given the importance of faithfully and accurately reproducing the response of ESMs, we hope the challenge will also spur innovation in nascent physically informed ML techniques.
In order to meet these objectives, we have provided open, easy to access datasets and training notebooks which reproduce the results shown in this manuscript and demonstrate the use of the different baseline emulators. All software is open-source and readily available using commonly used package managers. We hope this dataset will provide a focus for climate and ML researchers to advance the field of climate model emulation and provide policy makers with the tools they require to make well informed decisions.