Fig. 2. Observed (black, left ordinates) and simulated (colored, right ordinates) Sahelian precipitation anomalies, forced with ALL (a, blue), AA (b, magenta), NAT (c, brown/red), and GHG (d, green). The CMIP6 MMMs are presented with solid curves surrounded by shaded areas demarking the bootstrapping confidence interval, while the CMIP5 MMMs are presented with dotted curves. The yellow shaded area is the confidence interval of randomized bootstrapped MMMs of CMIP6 piC simulations, and represents the magnitude of noise in the CMIP6 MMMs. Hemispherically asymmetric volcanic forcing from Haywood et al noted in panel (c). A negative sign denotes an eruption that cooled the northern hemisphere more than the southern hemisphere while a positive sign denotes the opposite, aligning with the sign of the expected Sahelian precipitation response to the eruption. Panel (a) additionally shows the CMIP6 ALL MMM when restricted to models, rather than institutions, that provide AA simulations (blue dashed curve), and a 20-year running mean of the sum of the AA, NAT, and GHG MMMs for CMIP5 (lavender dashed curve) and CMIP6 (burgundy dashed curve). The label shows the number of institutions used for each CMIP6 MMM (N), the correlation of the CMIP6 MMM with observations (r), and the standardized root mean squared error of the CMIP6 MMM with observations (sRMSE).
In the AA experiments (panel b), CMIP6 is anomalously wetter than CMIP5 in the 1970s and around 2000, but otherwise looks similar to CMIP5: precipitation declines in the mid-century and then recovers after the clean air acts, preceding the timing of observed variability by about 10 years. There are some differences in the NAT experiments between CMIP5 and CMIP6 (panel c), but the largest variations in both ensembles are interannual episodes that are clearly associated with volcanic eruptions. In the GHG experiments (panel d), CMIP6 shows anomalous wetting after 1970 that wasn’t present in CMIP5.
Similar changes can be seen in the ALL simulations (panel a): while CMIP5 reaches peak drought in 1982 – close to the observed precipitation minimum – CMIP6 dries very little and only until 1970, after which it displays an anomalously wetter climate than CMIP5 through the end of the century. But while the precipitation responses to different forcing agents appear to add linearly in CMIP5 (compare the lavender dashed curve to the blue dotted curve), the late century wetting in CMIP6 is larger than the sum of GHG and AA wetting (burgundy dashed curve; including NAT does not help.) This effect is robust to differences in model availability for the different sets of forcing agents (see figure caption and light blue dashed curve). Thus, in the ALL simulations, CMIP6 displays slightly less drying from AA compared to CMIP5, more wetting from GHG, and additional wetting after 1990 from a non-linear interaction between forcings.
As a result of these changes, the response to forcing in CMIP6 is a poor match to observations. Figure 3 displays the correlation (panel a, “r”) and sRMSE (panel b) between observations and simulated MMMs (dots) and bootstrapped MMMs (curves) from CMIP6 (ALL in blue, AA in magenta, NAT in brown, and GHG in green solid curves) and CMIP5 (ALL and AA in blue and magenta dotted-dashed curves; other simulations omitted for clarity) from 1901 to the end of the simulations (2003 for CMIP5 and 2014 for CMIP6). The dotted curves present the randomized bootstrapping distributions for the CMIP6 piC simulations, and the vertical dashed lines mark the one-sided p=0.05 significance level given by these distributions. Recall that correlation measures similarity in timing between simulations and observations where 1 is a perfect match, and sRMSE measures the amplitude of differences between the simulations and observations where 0 is a perfect match.