Overlooked possibility of a collapsed Atlantic Meridional Overturning Circulation in warming climate

Current climate models suffer from biases (1), and therefore, it is critically important to assess the potential impact of the model biases on future climate projections. One vital player for climate change is the Atlantic Meridional Overturning Circulation (AMOC). With a warm, northward near-surface flow and a colder, southward return flow at depth (2), the AMOC carries oceanic heat northward and contributes moderate climate to the U.K. and northwest Europe (3). There is evidence that the AMOC has slowed down since the early 20th century, although this long-term declining trend of AMOC strength is subject to great uncertainty (3). Under future global warming, the AMOC is predicted to further weaken (1, 4–8), but the degree of the change is uncertain. Most climate models predict a moderate slowdown but not a complete shutdown of the AMOC (8). These different AMOC responses essentially depend on the circulation stability (1–3). Recent studies (9–14) point to a serious bias in AMOC stability in current climate models. Observational analyses (9–16) suggest an unstable modern AMOC (with multiple equilibria), meaning that the AMOC may switch between “on” and “off” modes in the future, as it did in the past (17–20). However, climate models show a common bias toward a stable AMOC (with a single equilibrium) (21, 22). This bias in AMOC stability casts serious doubt on the projection of future AMOC change.

To test the bias impact on future AMOC change, we conduct parallel experiments based on the same climate model but of two versions: the present-day control run (CTL) of the National Center for Atmospheric Research (NCAR) Community Climate System Model version 3 (CCSM3) (23, 24) and a bias adjusted run (ADJ) (14) via a surface flux adjustment (see Materials and Methods for details). We further rely on an indicator $$\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}$$ (25, 26) to analyze the stability of the AMOC, which is defined as the difference of the AMOC-induced freshwater transports across the southern $$({\mathit{M}}_{\text{ovS}}^{*})$$ and northern $$({\mathit{M}}_{\text{ovN}}^{*})$$ boundaries of the Atlantic (see Materials and Methods for details). Previously, $${\mathit{M}}_{\text{ovS}}^{*}$$ has been used as an AMOC stability indicator (7, 10–13, 27), but it is not an accurate measure of AMOC stability (26) because it neglects the freshwater transport between the Atlantic and the Arctic $$({\mathit{M}}_{\text{ovN}}^{*})$$. For example, Weaver et al. (7) show that few models in the Coupled Model Intercomparison Project Phase 5 (CMIP5) (28) exhibit a rapid collapse of future AMOC, although $${\mathit{M}}_{\text{ovS}}^{*}$$ classifies 40% of them as being in an unstable regime. By contrast, $$\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}$$ accurately denotes the freshwater transport induced by the AMOC and therefore can better represent a basin-wide salt-advection feedback and the AMOC stability (26). In particular, a divergence of the AMOC-induced freshwater transport $$(\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}<0)$$ indicates an unstable AMOC in response to buoyancy perturbation owing to a positive feedback with salinity advection. We suppose that an energetic modern AMOC induces a freshwater divergence. Under global warming conditions, an initial buoyancy perturbation in the North Atlantic weakens the AMOC and hence the associated freshwater divergence. The reduced freshwater divergence may then lead to an accumulation of freshwater in the North Atlantic, further amplifying the initial freshwater perturbation and resulting in a collapse of the AMOC.

Figure 1A shows modern AMOC stability inferred from $$\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}$$ in observations (reanalysis data) (see Materials and Methods for details) and model simulations (table S1). The observations suggest a freshwater divergence induced by a vigorous AMOC $$(\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}<0)$$ and thus an unstable AMOC, whereas climate models (without flux adjustment, see Materials and Methods for details) simulate a freshwater convergence $$(\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}>0)$$ and therefore a stable AMOC. This bias in AMOC stability $$(\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*})$$ is mostly caused by a biased $${\mathit{M}}_{\text{ovS}}^{*}$$ because $${\mathit{M}}_{\text{ovN}}^{*}$$ is consistently negative in both observations and models, as associated with a freshwater import from the Arctic (Fig. 1B). At the southern boundary, observations suggest a strong freshwater export owing to a saltier northward flow (Fig. 2B) of surface and thermocline waters (<500 m) than the deep southward flow (Fig. 2B) of the North Atlantic Deep Water (NADW) (1500 to 4000 m), whereas climate models show a freshwater import or a diminished transport primarily due to a fresh bias in salinity above 500 m (Fig. 2A).

The flux adjustment enables a correction of AMOC stability. By eliminating the salinity bias in the upper levels of the South Atlantic (Fig. 2A), the ADJ simulates a strong freshwater export across the southern boundary, an Atlantic freshwater divergence, and therefore an unstable AMOC consistent with observations (Fig. 1, A and B). Meanwhile, the flux adjustment only slightly modulates the magnitude of the AMOC. The AMOC volume transport remains similar in both CTL and ADJ (fig. S1, A and C) and close to observations (29).

We conduct two parallel doubling CO₂ experiments (the CTLCO₂ and ADJCO₂) based on two versions of the CCSM3 (see Materials and Methods for details). The atmospheric CO₂ concentration is instantaneously doubled (at year 201) from the present-day level and then remains constant thereafter. This warming scenario is between the Representative Concentration Pathways (RCP) 4.5 and RCP6.0 scenarios from the Intergovernmental Panel on Climate Change Fifth Assessment Report (1). The two models show similar climate responses during the first three decades. The AMOC strength reduces (Fig. 1C), and the Arctic sea ice shrinks (fig. S2B). The surface warming pattern is consistent with observations, which is characterized as a warming minimum (or slight cooling) to the south of Greenland, intensified warming over Eurasia and North America, and enhanced warming in the tropical Pacific, with a warming minimum in the southeastern Pacific (fig. S3). Nevertheless, predictions by the two models diverge markedly after the first 50 years. In the CTLCO₂, the AMOC slightly and slowly recovers (Fig. 1C), which is consistent with projections by most CMIP5 models (6, 7). In the ADJCO₂, by contrast, the AMOC continues to decelerate and eventually collapses at year 500 (Fig. 1C). These distinct AMOC behaviors between models originate from the AMOC stability. During years 251 to 500, the initial weakening of the AMOC in the ADJCO₂ (CTLCO₂) causes an anomalous freshwater divergence (convergence) in the Atlantic (Fig. 3C), primarily through the transport change across the southern boundary (Fig. 3, A and B). This freshens (salinifies) the Atlantic (Fig. 3D), inhibits (promotes) the deep convection and NADW formation, and finally leads to a collapse (partial recovery) of the AMOC. It is worth mentioning that $$\mathrm{\Delta }{\mathit{M}}_{\text{ov}}^{*}$$ in the ADJCO₂ eventually switches from negative to positive, implying that the evolving AMOC acts to accumulate freshwater in the Atlantic and, thus, the final collapsed state is steady.

The distinct AMOC changes between the ADJCO₂ and CTLCO₂ cause distinct long-term climate responses. For example, during years 251 to 500, in the ADJCO₂ (CTLCO₂), sea ice area increases (decreases) in the Northern Hemisphere (fig. S2B). We further examine the centurial mean difference between the ADJCO₂ (CTLCO₂; years 501 to 600) and the ADJ (CTL; years 101 to 200). The ADJCO₂ predicts a large cooling over the northern North Atlantic and neighboring areas (Fig. 4C). Surface air temperatures over the U.K., Iceland, and northwest Europe can drop to greater than 7°C during boreal winter. Associated with this cooling, sea ice expands over the Greenland-Iceland-Norwegian (GIN) seas and to the south of Greenland (fig. S4C). Compared with the CTLCO₂, the ADJCO₂ predicts a general cooling in the Northern Hemisphere polar region to 45°N (Fig. 4E), and sea ice retreat (fig. S4D) around Antarctica. The global mean surface air temperature in the ADJCO₂ is ~0.4°C less than that in the CTLCO₂ (fig. S2A).

The effect of flux adjustment on climate change extends into the tropics. Unlike the CTLCO₂, the ADJCO₂ simulates a stronger (weaker) surface warming south (north) of the equator in the Atlantic (Fig. 4C), as caused by a severe decline in the AMOC and associated heat transport. The warming pattern effectively displaces the Altantic Inter-tropical Convergence Zone (ITCZ) southward and generates a rainfall dipole astride the equator: enhanced precipitation over the tropical South Atlantic and northeastern Brazil and reduced precipitation over the tropical North Atlantic and Central America (Fig. 4D). This ITCZ shift corresponds to a robust southward migration of the Hadley cell (fig. S5). Overall in the tropics, the difference in rainfall change between the ADJCO₂ and CTLCO₂ runs closely follows the patterns of SST difference (Fig. 4, E and F) (30).

We have used a bias correction approach to illustrate the impact of the AMOC stability bias on future climate projections. The corrections put the AMOC in an unstable regime and lead to an AMOC collapse 300 years after a CO₂ doubling. With the AMOC shutdown, the bias-adjusted model predicts a distinct climate change from most CMIP5 models: a prominent cooling over the northern North Atlantic and neighboring areas, a remarkable sea ice expansion over the GIN seas and to the south of Greenland, and a significant southward rain-belt migration over the tropical Atlantic. Our results suggest that prevailing predictions by the CMIP5 models underestimate the model uncertainty in AMOC response and regional climate change around the North Atlantic.

The same model simulates an AMOC collapse with a large 1-sverdrup (1 sverdrup = 10⁶ m³/s) freshwater pulse (14) that mimics major ice-sheet discharges into the North Atlantic in paleoclimate observations. The AMOC collapse is primarily caused by a halinely induced reduction of surface buoyancy (see fig. S6F and Materials and Methods for details) associated with an extreme freshening in the northern North Atlantic and GIN seas (fig. S6L). On the other hand, the current study uses a more realistic setting, that is, the CO₂ increase for future global warming. The AMOC shutdown under global warming is primarily caused by a thermally induced buoyancy reduction (fig. S6B). The resultant buoyancy change (fig. S6A) is about one order of magnitude smaller (~0.1 sverdrup) than that in the hosing experiment (fig. S6D). Another important difference between the global warming and hosing scenarios is the response of atmospheric moisture transport from the Atlantic to the Pacific across Central America. Under global warming, the moisture transport intensifies because of atmospheric moisture increase, a negative feedback that increases Atlantic salinity (31) and stabilizes the AMOC (32–34). However, this mechanism is generally absent in the hosing experiments (34).

We thank T. Delworth for helpful discussions. We thank two anonymous reviewers for their valuable comments that helped improve the manuscript. Funding: This work was supported by the NSF, NSFC(41130105), the U.S. Department of Energy, and the Ministry of Science and Technology of the People’s Republic of China (2012CB955201). Author contributions: W.L. and Z.L. conceived and designed the experiments. W.L. and J.Z. performed the experiments. W.L. and S.-P.X. analyzed the data. W.L., S.-P.X., and Z.L. wrote the manuscript. Competing interests: The authors declare that they have no competing interests. Data and materials availability: The ERSST.v4 data are publicly available at https://www.ncdc.noaa.gov/data-access/marineocean-data/extended-reconstructed-sea-surface-temperature-ersst-v4. The MIMOC data are publicly available at www.pmel.noaa.gov/mimoc/. The GISTEMP data are publicly available at http://data.giss.nasa.gov/gistemp/. The C-GLORS data are publicly available at www.cmcc.it/c-glors/. The ECCO-v4 data are publicly available at http://www.ecco-group.org. The ORAS4 data are publicly available at https://climatedataguide.ucar.edu/climate-data/oras4-ecmwf-oceanreanalysis-and-derived-ocean-heat-content%20and%20Integrated%20Climate%20Data%20Center%E2%80%93ICDC. The SODA 2.2.4 data are publicly available at https://iridl.ldeo.columbia.edu/SOURCES/.CARTON-GIESE/.SODA/.v2p2p4/index.html?Set-Language=en. The CMIP5 data are publicly available through the portal of the U.S. Earth System Grid Center for Enabling Technologies on the page http://pcmdi9.llnl.gov/. All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors.