Fig. 4. PS of observed Sahelian precipitation (solid black curve) and the residual of observations and the ALL MMM (dotted-dashed black curve) and associated 95% confidence intervals (grey shading), compared to the average PS by model of piC simulations (brown to turquoise). Mean piC PS are colored by the average yearly piC precipitation by model, where brown simulations are drier than observed, and turquoise simulations are wetter than observed.
We must conclude that no linear combination of the simulated forced signal (which correlates poorly with observations) and simulated IV (which has insufficient low-frequency variance) in coupled CMIP6 simulations can explain observed Sahel variability during the 20th century. Thus, model deficiency cannot be blamed solely on the simulation of climate feedbacks: the CMIP6 ensemble displays a fundamental inability to simulate the observed fast and slow Sahelian precipitation responses to forcing, observed low-frequency IV, or both. To identify the proximate cause of this failure, in the next three sections we examine each causal path component identified in Figure 1.
b. AMIP simulations: the Response to SST, Atmospheric Internal Variability, and the Fast Response to Forcing (\(\overrightarrow{t}\),\(\overrightarrow{a}\), and \(\overrightarrow{f}\))
To isolate the effect of SST on the Sahel (\(\overrightarrow{t}\)), we examine precipitation in the CMIP6 amip-piForcing simulations, which force atmosphere-only models with the observed SST history (containing both internal, \(\overrightarrow{o}\), and forced,\(\overrightarrow{s}\), oceanic variability) and constant preindustrial external radiative forcing (no \(\overrightarrow{f}\)). The MMM of simulated Sahel precipitation filters out atmospheric IV (\(\overrightarrow{a}\)), leaving the precipitation response to the entire observed SST field. It is displayed in Figure 5a (orange) and compared to observations (black) on the same ordinates. Overall, the performance of the amip-piF MMM is much better than that of the coupled simulations: it achieves a high correlation (r = 0.60) and a low sRMSE (0.81, see orange curves in Figure 3). The good match with observations is achieved mostly at low frequencies: though it doesn’t accurately capture many interannual episodes—notably including the precipitation minimum in 1984—the MMM appears to capture the magnitude of low-frequency variability, even including wetting in the 50s and early 60s, which is missing from the coupled MMM. This can be seen more quantitatively by spectral analysis. In Figure 6a, the PS of the amip-piF MMM (dashed orange curve) and its 95% confidence interval (orange shaded areas), are compared to those of observations (black). Unlike previous generations of AMIP experiments (e.g. Scaife et al. 2009), the PS of the simulated MMM is roughly consistent with observations.