ClimateBench also includes a selection of more idealised simulations
which are intended to provide training data at the ‘corners’ of the
four-dimensional input space, again helping reduce the chances of
extrapolation in the resulting emulators (as demonstrated in Figure A1).
Two simulations that are commonly used to diagnose the equilibrium and
transient climate sensitivity are abrupt-4xCO2 and1pctCO2, respectively. As the name suggests, theabrupt-4xCO2 includes an abrupt quadrupling of CO2 over the
pre-industrial concentrations while all other forcing agents remain
unchanged. This level of concentration represents the high end of future
scenarios, broadly in line with ssp585 but with no contribution
from the other forcers, simplifying its interpretation. The abrupt
nature of the forcing also allows the timescale of the responses to be
determined which can be useful for emulators which account for this. The1pctCO2 simulation gradually increases the atmospheric
concentration of CO2 by 1% per year, again with other forcing agents
unchanged. Two simulations performed as part of the
Detection-Attribution Model Intercomparison Project (DAMIP; Gillett et
al., 2016) represent the historical period forced by only CO2 and other
long-lived greenhouse gases (hist-GHG ), or only anthropogenic
aerosol (hist-aer ). Again, these provide opportunities to train
emulators in regions of the input (emissions) space that are at the
limits of plausible future scenarios.
Finally, the piControl simulation provides a baseline simulation
with all forcings remaining unchanged from their pre-industrial values.
All target variables are calculated as a change against this climatology
to simplify the training and interpretation of the results. This long
(500 year) simulation also enables a robust estimation of internal
variability of the climate system for those emulators which are able to
represent it, as discussed further in Section 5.1.
The input data for these simulations is prescribed by the experimental
protocol and provided by the input4MIPS project
(https://esgf-node.llnl.gov/search/input4mips/),
which we collate and pre-process for ease of use. Specifically, we
extract the provided global mean emissions of CO2 and CH4 for each of
the realistic (historical, ScenarioMIP and DAMIP) experiments from the
checksum files provided by the Community Emissions Data System (CEDS)
dataset (Hoesly et al., 2018). We sum over each sector and each month in
order to derive annual total emissions and convert from Kg to Gt of CO2.
Some historical and future periods are only provided in 5-yearly
increments, so we linearly interpolate to yearly values for consistency.
The CO2 emissions are summed cumulatively since, for realistic
scenarios, a compensation between forcing efficiency and ocean uptake
means the temperature response to CO2 is approximately linear in the
cumulative emissions (Matthews and Caldeira 2008; Allen et al. 2009).
Figure 1 shows the global mean emissions of each of the forcing agents
under different future emissions scenarios, showing a wide range of
possible pathways.