New classes of climate model emulators to improve paleoclimate reconstructions

It is well known that the predictability of the climate varies over time and will depend on the initial conditions, especially when considering non-linear systems such as El Niño Southern Oscillation (ENSO). While recent decades have seen a few extreme ENSO events, proxy data reveal a large amplitude in tropical Pacific sea surface temperatures low frequency modulations over past millennia. To better interpret what is observed in proxies, a useful approach is to combine the climate information derived from natural archives with the physics of GCMs using paleoclimate data assimilation (PDA). Recently, efficient online ensemble-based data assimilation techniques have been developed relying on climate model emulators and the predictable components of the climate system. The skill of these ensemble forecasts is a key factor for the success of PDA especially when considering Particle Filters. Such predictability may however change according to the host-GCM, the emulator skills in capturing the host-GCM non-linear behaviours and the dimension of the problem. In this study, we explore these issues in a perfect model framework across PMIP3 and PMIP4 climate model simulations for the past millennium, relying on various types of architectures and climate model emulators. Our results indicate that relying on such a hierarchy of multi-model approaches provides a promising way to better quantify uncertainties and decipher the relative contribution from internal dynamics and external forcings embedded in proxy records, particularly regarding ENSO.