Abrupt cooling over the North Atlantic in modern climate models

Observations over the 20th century evidence no long-term warming in the subpolar North Atlantic (SPG). This region even experienced a rapid cooling around 1970, raising a debate over its potential reoccurrence. Here we assess the risk of future abrupt SPG cooling in 40 climate models from the fifth Coupled Model Intercomparison Project (CMIP5). Contrary to the long-term SPG warming trend evidenced by most of the models, 17.5% of the models (7/40) project a rapid SPG cooling, consistent with a collapse of the local deep-ocean convection. Uncertainty in projections is associated with the models’ varying capability in simulating the present-day SPG stratification, whose realistic reproduction appears a necessary condition for the onset of a convection collapse. This event occurs in 45.5% of the 11 models best able to simulate the observed SPG stratification. Thus, due to systematic model biases, the CMIP5 ensemble as a whole underestimates the chance of future abrupt SPG cooling, entailing crucial implications for observation and adaptation policy.


Supplementary
. Significance of the differences between the background stratification in non-abrupt models and in SPG convection collapse models. By assuming Gaussian distributions, curves represent probability density functions (PDF) for stratification indexes for non-abrupt sub-ensemble (red), for SPG convection collapse sub-ensemble (blue) and for all models ensemble (black). The intersection between PDF and dashed line at 0, i.e. stratification from observational data, indicates the likelihood that a specific subset of models is able to reproduce a SPG stratification matching that in observational data. Such a probability is higher for SPG convection collapse than for the non-abrupt models. However, this may be the consequence of the different sizes of the model ensembles, i.e. 7 members versus 29 members. To exclude sampling-flawed estimates we performed a statistical test based on Monte Carlo method. We standardized the sampling by choosing 10 4 random combinations of 7 nonabrupt models among the 1560780 possible combinations, i.e. C7,29=29!/(7!*22!), and calculated their PDF. We repeated this procedure 100 times finally finding that only 2.280.19% of the combinations of 7 non-abrupt models produce higher PDF values for null stratification index than in the SPG convection collapse ensemble, as elucidated by the interval bars on the left side of the panel. This test evidences that SPG convection collapse models reproduce a SPG background stratification that is closer to observational than that simulated by non-abrupt models, with significance higher than 95%. Note that all the SPG convection collapse models lie within this range. The low statistical relations found evidence that the SST response over the SPG cannot be constrained by the present-day AMOC. This analysis is limited to those models for which AMOC data were directly available.   Fig. 2a). These rapid changes take place in concurrence with a sudden decrease in sea surface density ( Supplementary Fig. 2b). The latter is fully driven by a rapid freshening of the surface ocean ( Supplementary Fig. 2c). The decrease in surface salinity reduces the MLD, leading to anomalies of O(1000 m), which obstruct the convective activity ( Supplementary Fig. 2d). Moreover, by defining a SPG index as the local minimum of the barotropic streamfunction within the reference region,

Model
i.e. the maximum cyclonic transport within the SPG, we find that a transition of the SPG circulation to a weaker mode precedes the SST drop by a few years ( Supplementary   Fig. 2f). A suspension of the SPG convection also affects the AMOC, which, however, does not dramatically decrease for these models (Supplementary Fig. 2e). A lead-lag analysis (not shown here) indicates that changes in surface salinity and MLD lead the rapid SPG cooling event rather than the AMOC.
On the other hand, CSIRO-Mk3-6-0, GFDL-ESM2M and MIROC5 feature strong highfrequency SST oscillations superimposed on a long-term weaker but persistent SST decrease ( Supplementary Fig. 3a). These models are characterized by intermittent suspension of the convective activity, modulated by multiple oscillations of the MLD over a long-term decreasing trend ( Supplementary Fig. 3d), which we suppose to be linked with similar responses of the SPG (although it can be shown only for CSIRO-Mk3-6-0 model due to lack of data for the others models.) This behaviour suggests a strong sensitivity to stochastic atmospheric conditions 4 , e.g. the NAO.
Overall, we propose the following mechanisms driving a convection collapse-induced abrupt SST cooling in the SPG. The rise in radiative forcing yields a general increase of surface temperature and an enhancement of the hydrological cycle, which in the net precipitation area of the NA translates into a freshening trend ( Supplementary Fig. 2c,   3c). Both warming and freshening contribute to a gradually decreasing sea surface density ( Supplementary Fig. 2b, 3b), which, in turn, diminishes the mixed layer depth in the convective areas ( Supplementary Fig. 2d, 3d). It is worth stressing that changes in Ekman-pumping due to a possible change in atmospheric winds are small ( Supplementary Fig. 2i, 3i) and cannot explain the sudden change in mixed-layer depth.
A thinner active layer in contact with the atmosphere decreases the heat capacity of the water column. As a consequence, the ocean cooling due to winter atmospheric conditions becomes less effective at depth, leading a reduction in upward heat flux ( Supplementary Fig. 2g, 3g). The oceanic heat loss is then confined to a shallower layer, resulting in a larger SST decrease during the winter that interferes with the background global warming trend. At the same time, less dense water at the core of the SPG reduces the density contrast with the surrounding lighter waters, thus slowing down the baroclinic cyclonic circulation ( Supplementary Fig. 2f, 3f). A weaker mode of the SPG transports less subtropical salty water into the Labrador and Irminger Sea, further reducing the surface density there ( Supplementary Fig. 2b, 3b). Moreover, a slower cyclonic circulation reduces the isopycnal outcropping in the centre of the SPG, which preconditions deep convection. Less convection in the SPG also causes the AMOC to slow down ( Supplementary Fig. 2c, 3c), meaning a decrease in northward transport of subtropical salty and warm water masses ( Supplementary Fig. 2h, 3h) into the NA. The interaction of these self-amplification processes of stratification pushes wide areas of the SPG across a threshold beyond which no deep convection is possible. The convection collapse therefore coincides with an SST drop, which locally overcompensates the warming contribution due to the rise in radiative forcing.
While this interplay of feedbacks with the stratification in the SPG applies to all SPG convection collapse models, the contribution of each single process to the local cooling is model dependent. Nevertheless, a common feature of these models is that a temperature drop over the SPG appears to be led by a sudden reduction of the local convective activity, but cannot be directly ascribed to a concurrent abrupt AMOC decline. The SPG convection and AMOC are strictly connected and therefore a reduction in AMOC plays an active role in these feedback mechanisms of SPG stratification. Moreover, a collapse of the convective activity does affect itself the AMOC strength. However, in SPG convection collapse models the AMOC weakening (and the associated northward heat transport reduction) is comparable to that in models showing a continuous SPG warming trend (Supplementary Fig. 6c, 6d, 6e). This is because, after the SPG convection collapse, deep-water formation is still sustained (and even reinforced in some models) in other locations, i.e. Nordic Seas and/or Faroe-Shetland Channel ( Supplementary Fig. 5). This explains, at least partially, why the AMOC does not strongly decrease in SPG convection collapse models, despite the interruption of deep-water formation in the Labrador/Irminger Sea. Thus, AMOC changes do not appear ultimately decisive in driving of the SST drop in the SPG convection collapse models. Rather, a rapid reduction of the upward heat flux from below (as a result of reduced convective mixing) appears to be the main cause of the abrupt SPG cooling in these models.

Abrupt cooling in AMOC disruption ensemble
For FGOALS-s2 and FIO-ESM, cooling concerns a more extended area covering the whole northern NA. We identified such a cooling due to an effective reduction of the northward heat transport caused by an almost ceased AMOC. The AMOC strongly slows down already during the historical period ( Supplementary Fig. 6b Fig. 5n, 5o).
Hence, both AMOC disruption models possess only one main site of deep convection, likely making the AMOC in these models overly sensitive to changes in local convective activity since is fed by such a unique sinking region. Thus, an initial decrease in the convection activity generates a local cooling and a progressive AMOC deceleration due to the positive salt advection feedback. This amplifies the cooling over the convection site, which, thereafter, intensifies and spreads gradually involving the whole northern NA. We can therefore infer that in FGOALS-s2 and FIO-ESM the abrupt events are mainly driven by self-amplification feedbacks involving the AMOC decline. However, the present-day AMOC strength simulated by these models significantly differs form the observed AMOC strength 8  It is worth stressing that also in the non-abrupt ensemble some models feature only one site of deep convection. However, they do not project an AMOC disruption, thus evidencing that an unrealistic simulation of a unique site of convection is not a sufficient condition for an overturning collapse to occur under RCP scenarios. Furthermore, models in the SPG convection collapse always feature two (or more) deep convection sites. Overall it can be concluded that models featuring only one deep-convection site are more susceptible to an abrupt AMOC collapse than models featuring more than one deep-convection site. Also, models correctly reproducing two deep-convection sites never simulate an AMOC collapse under RCP scenarios, but they are the only models that could simulate an isolated SPG convection collapse.