Comparison of different observational constraint methods to reduce the uncertainty in the AMOC at the end of the 21st century

Abstract

For a given greenhouse gas emission scenario, climate models are showing very significant differences for the future climate, due to processes complex to simulate. Among them, the Atlantic Meridional Overturning Circulation (AMOC) fate has been shown to be very uncertain, explaining large amount of the differences in climate projections in the North Atlantic region. Indeed, under the ssp2-4.5 scenario, CMIP6 models show an AMOC maximum at 26°N that goes from 18.1 ± 4.1 Sv (1 Sv=106 m3/s, ensemble mean of 30 models one standard deviation) during the period 1850-1900, to 11.6 ± 3,9 Sv in the last decade of 2100, with AMOC ranging from 4.3 to 21.3 Sv depending on the model. There is thus a clear need to improve estimates of the AMOC in the future. In this respect, methods called emergent or observational constraint (OC) have been recently developed. They combine climate models and observations, by finding and using an emergent relationship between a given climate variable in the future and observable predictors, to refine the best guess and uncertainty estimation of this climate variable. This study proposes to apply them to the case of AMOC projections. Both the choice of the predictor, and of the OC method can be key. What are the best choices to reduce most the uncertainty of the AMOC at the end of the 21st century? To answer such a question, this study compares two possible cases: it uses either only one predictor, the past AMOC, or a set of predictors, the sea surface temperature and salinity from various regions in the world, which are known to impact on-going and future fate of the AMOC. Moreover, this study compares five different OC methods. The uncertainty is evaluated for each couple of predictors and OC choice, considering cross-validation and observational errors. The best estimates of future AMOC under ssp2-4.5 scenario, constrained by the observed AMOC over the period 2004-2021, is 11,6 ± 2,5 Sv, using the linear regression method. When constrained by a larger set of predictors, it is 7,9 ± 2,1 Sv, using also the linear regression found here as the best OC method, with a Ridge regularization that limits overfitting. Thus, the future AMOC, in 2091-2100 compared to 1850-1900, is weakening by 56% when constrained by various recent observations that count for AMOC dynamics, in place of 36% when estimated using ensemble mean or AMOC observation only, as done in last IPCC reports. This result might have therefore considerable impacts on future adaptation plans within the North Atlantic regions.