Baseline emulators

Three baseline emulators are developed to demonstrate various potential approaches to tackling the machine learning problem this dataset provides. These are performed using the Earth System Emulator (Watson-Parris et al., 2021) to provide a simple interface for non-ML experts and permit sampling the emulators for potential use in detection and attribution workflows (as discussed in the Section 5). The three emulators all perform skillfully, as summarised in Table 2 and Figure 4 and discussed in more detail in each of the following subsections. The emulators also show broadly similar biases, particularly for precipitation where they all slightly underestimate increases (decreases) in tropical (subtropical) rainfall in the western Pacific. This might suggest that these particular changes are driven by different climate forcers or longer time-scale changes than modelled in this study. A direct comparison of the emulator predictions and NorESM is shown in Figure A3.
Table 2: The average root mean square error (RMSE) of the different baseline emulators for the years 2050-2100 against the ClimateBench task of estimating key climate variables under future scenario SSP245. Another state-of-the-art model (UKESM1) and the average RMSE between NorESM ensemble members as an estimate of internal variability are included for comparison.