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.