Figure 1. Global mean temperature anomalies from a 1850-1900
baseline for CMIP6 AOGCMs. The range between the ensemble maximum and
minimum temperature changes is shown by gray shading. Changes in
temperatures are forced by historical forcings during 1850-2014 and are
shown for the period 1915-2014. RMSEs are calculated over 1915-2014.
3.2 Roles of different forcings for near-surface temperature
change
In Figure 2 we focus on two AOGCMs with relatively large errors in their
emulations for the historical period (HadGEM3-GC31-LL and IPSL-CM6A-LR),
one AOGCM with relatively small errors (CanESM5), and one AOGCM whose
responses contrast with the other AOGCMs (NorESM2-LM).
Although EBM2 was calibrated using abrupt-4xCO2, errors predominantly
arise from emulation of the response to GHG forcing; in part because GHG
has the largest ERF. The EBM2 emulations overestimate the temperature
increase due to GHGs for HadGEM3-GC31-LL, IPSL-CM6A-LR and CanESM5 (even
though the CanESM5 historical fit is good). In contrast, the EBM2
emulation underestimates the temperature response to GHGs for
NorESM2-LM.
Emulation of the temperature response to aerosol forcing is the largest
source of error in one model (NorESM2-LM). For all models, errors
associated with aerosol forcing offset errors associated with GHG
forcing. This cancellation of errors gives a spurious impression of
better performance for the historical simulations. As shown for the
combined forcings (Figure 1), the step model produces closer emulations
of temperature for both GHG and aerosol forcings.
Emulation of the temperature response to natural forcings is a small
source of error for the eight AOGCMs and the emulations are mostly
within the spread of the AOGCM ensemble (Figures 2 and S2). Although
larger ensembles and longer simulations are required to robustly assess
the emulated response to volcanic forcing, thermal inertia of the EBM2
layers and allowance for rapid cloud adjustments within RFMIP ERFs will
contribute to closer emulations (Held et al. 2010; Gregory et al. 2016).