Multi-model comparison of carbon cycle predictability in initialized perfect-model simulations

Predicting carbon fluxes and atmospheric CO₂ can constrain the expected next-year atmospheric CO₂ growth rate and thereby allow to independently monitor total anthropogenic CO₂ emission rates. Several studies have established predictive skill in retrospective forecasts of carbon fluxes. These studies are usually backed by perfect-model simulations of single models showing the origins of predictive skill in carbon fluxes and atmospheric CO₂ concentration. Yet, a comprehensive multi-model comparison of perfect-model predictions, which can be valuable in explaining differences in retrospective predictions, is still lacking. Moreover, as of now, we don't have sufficient understanding of how well do the models predict their own integrated carbon cycles and how congruent this predictability is across models.

Here, we show the predictive skill of land and ocean carbon fluxes as well as atmospheric CO₂ concentration in seven Earth-System-Models. Our first results indicate predictive skill of globally aggregated carbon fluxes of 2±1 years and atmospheric CO₂ of 3±2 years. However, the regional patterns, hotspots and origins of predictive skill diverge among models. This heterogeneity explains the regional differences found in existing retrospective forecasts and backs the overall consistent predictability time-scales at global scale.