Noise-induced tipping of Atlantic Meridional Overturning Circulation under climate mitigation scenarios

Abstract

Growing attention to decarbonization to limit global warming raises an important question of whether climate mitigation efforts could effectively prevent a potential collapse of the Atlantic Meridional Overturning Circulation (AMOC), a major tipping element in the climate system. Using CO₂ stabilization simulation with the Community Earth System Model version 1, we show a marked ensemble divergence in AMOC evolution under identical or comparable CO₂ forcing, towards either recovery or a multi-century collapsed state. This divergence arises from the accumulation of stochastic noise near the AMOC stability threshold, characterized by persistent high-pressure anomalies over the subpolar North Atlantic. In other words, internal atmospheric variability can push the AMOC across its stability threshold and trigger collapse, aided by the basin-wide salt-advection feedback. A modest delay in implementing mitigation further increases the likelihood of crossing this threshold, amplifying the risk of collapse. This study underscores high uncertainties in future climate projections in response to climate mitigation efforts and highlights the urgent need for strong and immediate international actions.

Introduction

The Atlantic Meridional Overturning Circulation (AMOC) is a critical component of Earth’s climate system¹, responsible for the global redistribution of oceanic heat, salt, and carbon. Its changes have substantial and far-reaching impacts on the global climate system, extending well beyond regional scales. Indeed, the influences of the AMOC on the climate have been observed and projected through the observations and future warming scenarios^2,3,4,5 (e.g., the North Atlantic warming hole). Paleo-proxy records reveal that the geological past witnessed even more dramatic AMOC-driven climate changes, the so-called Dansgaard-Oeschger events⁶. Seminal earlier studies have identified the existence of multiple equilibrium states within the AMOC⁷ and attributed the abrupt transitions between these equilibria to triggering the recurrent global-scale climate changes in the past^8,9,10.

These multiple lines of evidence from paleoclimate records and model-based studies highlight the multiple-equilibria nature of the AMOC, associated with various tipping points^11,12. In this context, a tipping point denotes a critical threshold at which a small additional perturbation can trigger substantial and self-perpetuating changes in the system’s future state¹². Such alterations ultimately drive the system into an alternative state, where positive feedbacks are strong enough to overcome the stabilizing negative feedbacks. Specifically, the basin-wide salt-advection feedback has been identified as the primary mechanism controlling the AMOC transition^7,13. Physical insights into the AMOC bistability have been advanced through a hierarchy of models, ranging from conceptual box models to oceanic eddy-permitting fully coupled global climate models (GCMs), in response to changing external forcing such as freshwater input over the North Atlantic and atmospheric carbon dioxide (CO₂) concentrations^{7,8,14,15,16,17,18,19,20,21,22}. These modeling studies have demonstrated a clear hysteresis behavior and tipping event of the AMOC.

The past serves as a mirror to the future; so, the paleo-proxy evidence of the past AMOC tipping supported by climate modeling raises the concern that the AMOC may cross the tipping point and collapse in a future warming world as the CO₂ concentrations continue to increase. While most state-of-the-art climate models project that a complete shutdown of the AMOC is unlikely within the 21st century²³, there are growing arguments that the current AMOC system may be on a self-perpetuating tipping course based on statistical early warning signals^24,25,26 and a diagnostic indicator of AMOC stability^{18,27,28,29,30,31}.

While there is an international consensus for decarbonization, this possibility of AMOC collapse represents an important source of uncertainty in future climate responses to anthropogenic mitigation efforts³². For example, if climate mitigation actions are implemented after the AMOC has already crossed its tipping point, it might not recover and continue to collapse despite the stabilization or reduction of CO₂ concentration, potentially pushing our climate system into uncharted future states. This underscores that when and how future mitigation actions are executed should be carefully planned, taking into account the tipping point of the AMOC. Analyzing this aspect could offer invaluable insights to decision-makers, helping them to weigh the benefits and risks of various climate mitigation options. However, this area still demands further exploration to understand its implications.

To address this question, here we evaluate the effectiveness of climate mitigation efforts initiated at different times in preventing an AMOC tipping, using large ensemble simulations of CO₂ stabilization with a state-of-the-art GCM. An overview of the experimental design is given in the following subsection; additional details can be found in the “Methods” section. Here we find that the AMOC can evolve toward either recovery or collapse under the same or comparable CO₂ forcing. We attribute this stochastic bifurcation to the accumulation of internal atmospheric perturbations near the AMOC stability threshold, with a spatial pattern characterized by persistent high-pressure anomalies over the subpolar North Atlantic. We further find that modest delays in implementing mitigation increase the probability of an AMOC collapse.

Results

CO₂ stabilization simulations using CESM1

We conducted three sets of experiments using the Community Earth System Model version 1 (CESM1). The first one, referred to as “PD”, is a present-day control simulation spanning 900 years with a constant atmospheric CO₂ concentration (367 ppm). The second, denoted as “1pctCO2”, is a typical global warming simulation with a 1% per year increase in CO₂ concentration over 140 years until it quadruples (black line in Fig. 1a, years 2001 to 2140). It consists of 28 ensemble members, branched from different initial conditions in the PD simulation. The final experiment, the central focus of this study, involves a climate mitigation scenario with the CO₂ concentration fixed at the same level as the starting point for 200 years. This experiment is branched from the 7 different CO₂ levels in the 1pctCO2 simulation (see colored horizontal parallels in Fig. 1a). Among them, the four blue-colored experiments with relatively lower CO₂ levels consist of 3 ensemble members each, while the other three experiments with relatively higher CO₂ levels (green, brown, and purple) have 28 ensemble members each (legends in Fig. 1a). Each experiment is labeled as “${{EXP}}_{{branched}\,{year}}$” and this naming will be used throughout the rest of this paper. Such CO₂ stabilization simulations, initiated from seven different global warming levels, allow us to explore the potential occurrence of AMOC tipping events under relaxed warming scenarios aligned with future climate mitigation actions implemented at different times.

Fig. 1: Design of CO₂ stabilization simulations and their Atlantic Meridional Overturning Circulation (AMOC) response.

a Temporal evolution of the prescribed CO₂ concentration forcing. The black line represents a CO₂ increase of 1% per year, referred to as 1pctCO2 experiment. The experiments, represented by horizontal lines in 7 different colors, are branched from the 7 different CO₂ levels of the 1pctCO2 (ranging from 2xCO₂ to 3xCO₂ relative to present-day level). Each experiment is integrated with constant CO₂ concentration for 200 years. The name of each experiment (${{EXP}}_{{branched}\,{year}}$) and the number of total ensemble members are given in the legend. b–d Temporal evolution of AMOC strength defined as the maximum of the meridional ocean stream function below 500 m at 45°N in each experiment. Thin and thick lines represent individual ensemble members and their ensemble mean in each experiment, respectively. All time series are based on annual means and smoothed by an 11-year running mean. Numbers in d represent the number of ensembles in which the AMOC collapses for each experiment.

AMOC collapse despite climate mitigation efforts

We first examine the temporal evolution of the AMOC strength defined as the maximum of the meridional ocean stream function below 500 m at $45^{\circ}$N in each CO₂ stabilization experiment. The AMOC continues to weaken in response to the increase in atmospheric CO₂ levels (black line in Fig. 1b), a consistent result among different state-of-the-art climate models²³. It is naturally expected that the AMOC will be gradually stabilized or potentially begin to recover once the anthropogenic increase in CO₂ concentrations ceases. Indeed, in the 4 blue-colored simulations (${{EXP}}_{2070}$, ${{EXP}}_{2080}$, ${{EXP}}_{2090}$, and ${{EXP}}_{2100}$), the AMOC gradually strengthens after an initial weakening for decades (Fig. 1b, c). The initial weakening of the AMOC persists for a couple of decades, due to the slow adjustment of the ocean circulation, especially in the deep ocean, toward its equilibrium. This is partly due to the continued increase in global mean temperature after CO₂ stabilization (Supplementary Fig. 1). As the forced CO₂ concentration is higher, the initial weakening phase of the AMOC tends to last longer, suggesting that the AMOC recovery takes longer as climate mitigation measures are delayed.

The AMOC recovery in all 4 blue-colored simulations implies that the AMOC has not yet passed its bifurcation tipping point with the earlier cessation of the anthropogenic CO₂ increases. However, the other 3 simulations (${{EXP}}_{2105}$, ${{EXP}}_{2110}$, and ${{EXP}}_{2115}$) show strikingly different outcomes. They exhibit significant disparities in the evolution of the AMOC and show a broader spread in their inter-ensemble distributions. Specifically, in the case of ${{EXP}}_{2105}$ (green line in Fig. 1d), most ensemble members exhibit initial weakening and subsequent recovery of the AMOC as in the 4 blue-colored simulations (Fig. 1c). Yet, one ensemble member shows a continued weakening of the AMOC, and this negative trend implies that the AMOC is transitioning into an alternative collapsed state. Interestingly, this bifurcation of the AMOC evolution among ensemble members is much more pronounced in the ${{EXP}}_{2110}$ (brown line in Fig. 1d), where the AMOC collapses in 17 ensemble members. Furthermore, with a 5-year delay in mitigation, the AMOC collapses in all 28 members of the ${{EXP}}_{2115}$ (purple line in Fig. 1d).

This suggests that, in this model, the bifurcation tipping point, defined as the threshold of external forcing (CO₂ concentration) for AMOC, may lie between 1097 ppm (${{EXP}}_{2110}$) and 1152 ppm (${{EXP}}_{2115}$). We caution, however, that this should not be misinterpreted as evidence that the AMOC can remain stable under such high CO₂ levels, thereby implying that higher emissions would be permissible. Such a high threshold may arise from a bias in AMOC stability in our model compared with observations (Fig. 2a); therefore, in reality, the tipping point may occur at much lower CO₂ concentrations. In other words, our results indicate that the AMOC tipping behavior is governed dynamically by the stability or strength of the AMOC system, rather than by CO₂ level itself. Further details on AMOC stability are provided in the following section.

Fig. 2: Atlantic Meridional Overturning Circulation (AMOC) stability indicator response.

a Temporal evolution of the annual mean freshwater transport convergence within the Atlantic by the AMOC ($\Delta {F}_{{ov}}$; black; see Methods) in the 1pctCO2 experiment (CO₂ increase of 1% per year; black line in Fig. 1a) and reanalysis data (ORAS5; blue and ECCOv4; skyblue). Solid lines and light shadings indicate the ensemble mean and full ensemble spread, respectively. b Temporal evolution of the $\Delta {F}_{{ov}}$ in the ${{EXP}}_{2105}$ (green), ${{EXP}}_{2110}$ (brown), and ${{EXP}}_{2115}$ (purple), respectively. Thin and thick lines represent the individual ensemble members and their ensemble mean in each experiment, respectively. All time series in b are based on annual means and smoothed by an 11-year running mean. The gray shading in b shows the $\pm 2\sigma$ range of the interannual variability of the $\Delta {F}_{{ov}}$ in the 900-year present-day control simulation (see Methods). A positive (negative) $\Delta {F}_{{ov}}$ indicates that the AMOC is in a stable (unstable) regime, characterized by a negative (positive) basin-wide salt-advection feedback.

At the same time, from a more practical perspective, despite only 10-year difference in the initiation of each experiment between ${{EXP}}_{2105}$ and ${{EXP}}_{2115}$, the future outcomes of the AMOC in these two simulations are markedly different. This reveals that a relatively small delay of global warming relaxation can abruptly increase the probability of the AMOC collapse near the bifurcation point. It is noteworthy that the declining rate of the AMOC in the ${{EXP}}_{2115}$ is comparable to that in the 1pctCO2 with increasing CO₂ concentrations (black line vs. purple line in Fig. 1b), even though there is no longer CO₂ increase. This highlights the critical role of the intrinsic dynamics of the AMOC, which persistently amplifies its initial weakening through positive salt-advection feedback, even under the stabilized external CO₂ forcing.

Moreover, we conducted an additional 250 years of integration for 3 ensemble members exhibiting AMOC collapse and 3 ensemble members showing AMOC recovery in ${{EXP}}_{2110}$, respectively (450 years in total; Supplementary Fig. 2). In the extended 250-year simulation, both the recovering and collapsing ensemble members reach their respective multi-century quasi-equilibrium states (~16 Sv for recovering vs ~3 Sv for collapsing members). Note that while these states persist for several centuries, it remains uncertain whether they represent truly stable equilibrium states. Given a longer integration period, the collapsed AMOC might either recover toward its original strength or weaken further, potentially transitioning into a reversed circulation state. However, from a practical standpoint, even a collapsed AMOC persisting for several centuries is sufficient to drive substantial and widespread climate impacts on human timescales. Therefore, we consider these states to be quasi-stable, as they persist for an extended period despite the inherent uncertainty regarding their ultimate fate.

To ensure the more robustness of our findings, we further analyzed the GISS-E2-1-G model, the only publicly available CMIP6 model that meets two key criteria: (1) integration of full shared socio-economic pathway (SSP) scenarios up to the year 2500, and (2) provision of multiple ensemble members. This model provides insights into how AMOC responses diverge under different future warming scenarios (Supplementary Fig. 3). In SSP1-2.6 (green), the AMOC consistently recovers across all ensemble members (#10). In SSP2-4.5 (brown), a bifurcated response is observed, with some ensemble members showing recovery (#13) while others exhibit collapse (#2)¹⁹. Under the high-emission SSP5-8.5 scenario (purple), the AMOC universally collapses across all ensemble members (#6). The three SSP scenarios follow distinct CO₂ forcing trajectories, complicating direct comparisons with our CESM1 mitigation experiments conducted at different times. However, the AMOC bifurcation response observed in the SSP2-4.5 scenario, which is comparable to the experimental protocol in this study, is particularly noteworthy. Comprehensively, these results underscore the robustness of AMOC tipping behaviors around its tipping point across different models and scenarios, further highlighting the critical importance of timely mitigation in preventing AMOC collapse.

AMOC shift towards unstable regime

To investigate the mechanism of this simulated AMOC tipping in a mitigated warming world, a simple and widely-used diagnostic indicator of the AMOC stability (denoted as $\Delta {F}_{{ov}}={F}_{{ov}{{\rm{S}}}}-{F}_{{ovN}}$) is analyzed (“Methods”). This indicator effectively represents the sign and strength of the basin-wide salt-advection feedback of the AMOC. The ${F}_{{ov}{{\rm{S}}}}$ and ${F}_{{ovN}}$ represent freshwater transport by the overturning circulation at the southern and northern Atlantic boundaries, respectively. The difference between them thus indicates freshwater convergence/divergence driven by the AMOC. A positive (negative) $\Delta {F}_{{ov}}$ indicates that the AMOC is in a stable (unstable) regime, characterized by a negative (positive) basin-wide salt-advection feedback in response to initial buoyancy forcing; the AMOC imports (exports) freshwater to (from) the Atlantic. Therefore, a negative $\Delta {F}_{{ov}}$ implies that an initial AMOC weakening could eventually trigger AMOC collapse through the self-amplifying salt-advection feedback. The $\Delta {F}_{{ov}}$ at the beginning of the CESM1 1pctCO2 simulation is positive (Fig. 2a), indicating a stable AMOC in the present-day climate.

With increasing CO₂ concentration, the $\Delta {F}_{{ov}}$ gradually decreases (Fig. 2a). The decreasing AMOC stability indicator under greenhouse warming is also shown in available observations as well as in the CMIP6 multi-model average under the SSP5-8.5 scenario^20,31. Even though the sign of $\Delta {F}_{{ov}}$ is positive until the end of the 1pctCO2 simulation (year 2140), but it approaches quite close to the neutral state ($\Delta {F}_{{ov}}\approx 0$). This suggests that the AMOC is gradually losing its barrier to tipping, moving into a more susceptible state to collapse. An intriguing point here is that the sign of $\Delta {F}_{{ov}}$ effectively constrains the bifurcation of the AMOC evolution shown in Fig. 1d. All runs exhibiting AMOC collapse in three CO₂ stabilization simulations have $\Delta {F}_{{ov}}$ crossing zero, whereas the others do not (Fig. 2b). In other words, when the sign of $\Delta {F}_{{ov}}$ becomes negative and further decreases, the AMOC continues to weaken through the relatively stronger self-amplifying process, confirming that the primary mechanism for AMOC tipping is indeed the powerful salt-advection feedback. Considering the usefulness of the $\Delta {F}_{{ov}}$ sign change in constraining the ensemble divergence of AMOC evolution, we hereafter use the point at which $\Delta {F}_{{ov}}$ reaches zero as the threshold of AMOC tipping in our simulations and analysis, marking the critical transition beyond which the AMOC continues to weaken. It is noteworthy that the $\Delta {F}_{{ov}}$ evolution in GISS-E2-1-G under the SSP2-4.5 scenario also effectively constrains the ensemble bifurcation of AMOC evolution (Supplementary Fig. 3c), providing further support for its diagnostic value across different model and scenario.

Stochastic bifurcation of the AMOC evolution

In general, there are three possible mechanisms for AMOC tipping^33,34. First, if the external forcing (e.g., freshwater input or CO₂ concentration) passes a certain bifurcation point, AMOC tipping can occur where a bistable system loses one of its stable states. This mechanism is referred to as bifurcation-induced tipping (B-tipping)^16,17,20. Second, stochastic noise can lead to the transition of the AMOC into its alternative state by helping it overcome tipping barriers even before the bifurcation point, a process known as noise-induced tipping (N-tipping)^19,35,36. Lastly, the third mechanism, rate-induced tipping (R-tipping), occurs when the rate of change in external forcing exceeds a critical threshold that the system can adapt to, causing the AMOC to transition to an alternative state^15,37.

These three mechanisms are not independent but are closely related, and their combined effects are important for understanding the tipping dynamics of the AMOC. For example, N-tipping of the AMOC is more likely to occur near the bifurcation point with a lower barrier to tipping, making the system more susceptible to collapse. This occurs as the basin of attraction flattens, reducing its curvature, thereby increasing its sensitivity to perturbations. The clear bifurcation of the AMOC evolution observed in ${{EXP}}_{2110}$ and the GISS-E2-1-G model under the SSP2-4.5 scenario (recall Fig. 1d and Supplementary Fig. 3b) is a good example of this. Moreover, these simulations can be considered an example of partial tipping, which indicates the system may undergo partial tipping before reaching full tipping, as described by ref. ³⁸.

To better understand this phenomenon, we next focus on the ${{EXP}}_{2110}$ simulation. Specifically, we explore the ensemble spread in ${{EXP}}_{2110}$ to understand the role of stochastic noise in determining the divergent AMOC behaviors in the ${{EXP}}_{2110}$. Figure 3a shows the temporal evolution of AMOC strength in the ${{EXP}}_{2110}$ as in Fig. 1d but categorized into 3 groups based on the future AMOC evolution. Three groups are named collapse (red), uncertain (gray), and recovery (blue) group, consisting of 17, 2, and 9 ensemble members, respectively. The temporal evolution of $\Delta {F}_{{ov}}$ in each ensemble member (Fig. 3b) is shown after sorting them in ascending order based on the last 50-year (years 2261–2310) average of the AMOC strength (Fig. 3a). As discussed in the previous section, all members in the collapse group cross the zero $\Delta {F}_{{ov}}$. Furthermore, the ensemble members with a stronger AMOC weakening cross the zero $\Delta {F}_{{ov}}$ more rapidly (blue dots in Fig. 3b).

Fig. 3: Stochastic bifurcation of the Atlantic Meridional Overturning Circulation (AMOC).

a Temporal evolution of the AMOC in the ${{EXP}}_{2110}$ categorized into 3 groups based on the last 50-year (2261–2310) average of AMOC strength. Three groups are named collapse, uncertain, and recovery group, and consist of 17, 2, and 9 ensemble members, respectively. Thin and thick lines represent all ensemble members and their mean in each group, respectively. The thick green line shows the ensemble mean. b Temporal evolution of the AMOC stability indicator ($\Delta {F}_{{ov}}$; see Methods) in the ${{EXP}}_{2110}$. The y-axis represents the ensemble rank, sorted in ascending order based on the last 50-year average of AMOC strength in (a). For example, the ensemble members with the weakest and strongest AMOC strength in the last 50 years are assigned rank 1 and 28 (bottom and top of the y-axis), respectively. c Same format as (b), but for each member’s deviation from the ensemble mean (i.e., stochastic noise, represented by the symbol ${{\boldsymbol{\delta }}}$). The physical quantity in (c) is the boreal winter (January-February-March; JFM) averaged surface density flux integrated over the Labrador-Irminger Sea (LIS, black contour in e). The blue dots in (b–d) indicate the timing of $\Delta {F}_{{ov}}=0$ for each ensemble within the collapse group. d Same as c, but accumulated over time (named as “cumulative stochastic noise” and represented by the symbol ${{{\boldsymbol{\delta }}}}^{{{\boldsymbol{cum}}}}$). e–g Collapse group mean of the cumulative stochastic noise until the timing of $\Delta {F}_{{ov}}=0$ in each member (17 blue dots in d). The physical quantities in e–g are the winter density flux, evaporation, and sea level pressure with 1000-hPa wind (sky-blue vectors), respectively. The black contours in g are climatological winter sea level pressure (zonal mean removed) of 0, −5, −10, and −15 hPa. Hatchings and the solid vectors indicate statistically significant response at the 95% confidence level.

Since all the ${{EXP}}_{2110}$ runs are forced by the identical external forcing, stochastic noise is responsible for the bifurcation of AMOC evolution (Fig. 3a) by controlling when and whether the $\Delta {F}_{{ov}}$ crosses the threshold. Atmospheric internal variability as a stochastic noise, characterized by a white noise spectrum, drives climate variability with a redder spectrum^39,40 via exchanges of properties such as heat, freshwater, and momentum at the air-sea interface. In this study, we consider the stochastic noise by calculating the deviation of each ensemble member from the ensemble mean, under the assumption that the ensemble mean reflects the externally forced response, while the deviations represent internal variability governed by a probabilistic distribution^19,41,42. By this definition, we check the stochastic noise of winter surface density flux (DF), including both heat and freshwater contributions (Fig. 3c), integrated over the Labrador-Irminger Sea (LIS, black contour in Fig. 3e), where oceanic convection is the most vigorous and closely tied to the AMOC. The negative (positive) winter DF in the LIS can directly decrease (increase) deep water formation in that region, hence the AMOC. Prior to reaching the threshold in the collapse group, the ensemble distribution of the stochastic noise of the LIS DF exhibits considerably chaotic behavior, making it challenging to discern any differences between the recovery and collapse groups (Fig. 3c).

Instead, we focus on the cumulative effect of the stochastic noise, as suggested by Hasselmann³⁹. The accumulation of stochastic noise indicates the time integration of these random internal processes. However, it should be noted that the ensemble deviation is no longer purely random after the tipping occurs, as the AMOC-driven mean state changes begin to dominate. Therefore, we focus on the time integration of stochastic noise from the start of the simulation up to the threshold ($\Delta {F}_{{ov}}=0$) in each ensemble member. This approach provides a reasonable estimate of the role of internal variability in shaping AMOC trajectories near the tipping threshold. The distribution of time-cumulative noise of the winter LIS DF indeed exhibits a distinctive difference between the recovery and collapse groups (Fig. 3d). The collapse group shows an overall negative signal before the threshold, indicating a larger cumulative influx of heat and/or freshwater into the ocean, which contributes to reducing the surface density. The negative DF deviation in the collapse group persistently weakens the AMOC, pushing it towards its threshold and ultimately resulting in the AMOC collapse via the stronger salt-advection feedback.

For a more detailed analysis, we examine the spatial pattern of the cumulative stochastic noise in the collapse group. The spatial patterns shown in Fig. 3e–g represent the 17-member collapse group average of the cumulative stochastic noise, calculated until the $\Delta {F}_{{ov}}$ reaches zero at each grid point in each ensemble member. The most pronounced signal of the cumulative DF is indeed located over the LIS (Fig. 3e). This is largely dominated by the thermal contribution, although the haline component also contributes to the negative deviation of DF (Supplementary Fig. 4). Specifically, the reduced evaporation in the collapse group (Fig. 3f) is crucial for the negative DF response in terms of both thermal and haline contributions: greater freshwater input into the ocean and lesser latent heat loss into the atmosphere (Supplementary Figs. 5 and 6). Based on our results, sustained monitoring of long-term cumulative evaporation, particularly over the LIS, could serve as a valuable early warning signal for AMOC collapse.

Furthermore, we found that the less evaporation is tightly linked to the reduced wind speed (Supplementary Fig. 7a) associated with a distinctive anomalous anticyclonic circulation (Fig. 3g) that weakens the semi-permanent Icelandic Low pressure system (black contours in Fig. 3g and shadings in Supplementary Fig. 7b). The cumulative atmospheric circulation forcing may result from multiple sources of internal variability, such as the North Atlantic Oscillation^43,44,45, teleconnections from the tropical Indian Ocean^46,47 or El Niño-Southern Oscillation^48,49,50. Disentangling the relative importance of these processes is beyond the scope of this study and could be conducted in an extended study.

Future climate uncertainty driven by AMOC tipping

Previous studies have shown that climate models simulate the far-reaching climate impacts of a substantially weakened or collapsed AMOC^4,51,52. In line with their findings, our result also reveals that the AMOC tipping causes the strong cooling in the subpolar gyre region in the Atlantic mainly by the declining northward oceanic heat transport within the Atlantic. This localized cooling subsequently extends to the entire Northern Hemisphere (NH) and also in the land area in the Southern Hemisphere (SH) (Fig. 4a). The substantial NH cooling results in an overall reduction in precipitation in the NH, pushing the intertropical convergence zone to the south (Fig. 4b). Moreover, it is observed that the distinct and continuous sea level rise (SLR) in the broad subpolar North Atlantic and Arctic (Fig. 4c, d) in the collapse group in response to the AMOC decline. The anthropogenic SLR is already known as irreversible⁵³, but curbing the increase in anthropogenic CO₂ can slowdown the ongoing SLR. However, our results imply that the AMOC collapse would significantly increase the irreversibility of regional SLR, particularly in the subpolar North Atlantic and Arctic regions.

Fig. 4: Climate impacts of the Atlantic Meridional Overturning Circulation (AMOC) tipping.

a–c Difference in surface air temperature (SAT), precipitation, and total sea level averaged over the last 50 years of the ${{EXP}}_{2110}$ between the collapse group (17) and the recovery group (9) defined in Fig. 4a. d As in Fig. 4a, but for the sea level area-averaged over the subpolar north Atlantic-to-Arctic Ocean (area marked in sky-blue in inset panel). The sea level anomaly shown in d is relative to the present-day (year 2000) levels. Hatchings indicate statistically insignificant response at the 95% confidence level.

These climate changes related to the AMOC collapse have been well documented in many previous studies, often based on comparisons between different external forcing scenarios (e.g., freshwater hosing on vs. off)^4,51 or between different periods within a given scenario (e.g., future vs. pre-industrial)⁵². In contrast, the global-scale climate impacts, demonstrated in this study (Fig. 4), arise under identical or nearly identical external forcing scenarios—either from stochastic fluctuations near the AMOC tipping point (${{EXP}}_{2110}$) or from small differences in the timing of mitigation forcing (${{EXP}}_{2105}$ vs. ${{EXP}}_{2115}$). This highlights that substantial uncertainty in future climate trajectories, especially across the NH (Fig. 5), can emerge not only from the magnitude of external forcing but also from internal variability and noise-induced tipping behavior. For example, using a metric, surface air temperature area-averaged over the NH, we display the distributions every 20 years in ${{EXP}}_{2105}$, ${{EXP}}_{2110}$, and ${{EXP}}_{2115}$ (Fig. 5c–e). While their initial states are almost similar, their final destinations are quite different, indicating a highly uncertain future climate under a mitigation scenario. Our results, therefore, underscore a distinct aspect compared with earlier studies: even under comparable levels of CO₂ forcing, noise-induced tipping of the AMOC can generate large and persistent spreads in regional climate outcomes. To ensure more reliable projections under the increasing likelihood of decarbonization, mitigation strategies must thus be carefully designed with consideration of the AMOC threshold and the possibility of tipping.

Fig. 5: Future climate uncertainty driven by the Atlantic Meridional Overturning Circulation (AMOC) tipping.

a, b The standard deviation of last 50-year averaged annual mean land surface air temperature (SAT) taken across 28 ensembles in the ${{EXP}}_{2105}$, and ${{EXP}}_{2110}$, respectively. c–e Ensemble distribution of the area-averaged SAT over the Northern Hemisphere time-averaged every 20 years in the ${{EXP}}_{2105}$, ${{EXP}}_{2110}$, and ${{EXP}}_{2115}$. The black dots in c–e indicate the ensemble mean.

Discussion

In this study, we found that the AMOC remains at risk of collapse despite climate mitigation efforts. Specifically, if carbon stabilization—where atmospheric CO₂ concentration stops increasing—is achieved only after the AMOC has crossed its bifurcation tipping point, the AMOC can tip to its alternative collapsed state due to the self-perpetuating salt-advection feedback, even in the presence of mitigated climate forcing (Fig. 1d). Moreover, even if carbon stabilization achieved before reaching the bifurcation tipping point, stochastic noise can still push the AMOC beyond its threshold, leading to tip (Fig. 3). Implementing stronger mitigation actions—such as carbon neutrality or negative CO₂ emissions—could allow AMOC recovery even if it temporarily crosses the tipping point. While such aggressive mitigation efforts may come with higher short-term costs, they could ultimately prevent far greater socio-economic and climatic consequences, making them a necessary investment for long-term global stability.

Our model exhibits a positive AMOC stability indicator at the beginning of the simulation (Fig. 2a) like other GCMs from CMIP5³⁰ and 6^31,54, while some studies suggested that it is already negative in the present-day observations^18,30,54 (Fig. 2a). Due to this positive bias, the present-day AMOC in our model is stable, thus the threshold of the AMOC tipping is significantly distant from the present-day. The negative AMOC stability indicator in the real world suggests that the current AMOC system may already be in an unstable regime, or at least, such a threshold might be much closer in the future than predicted by cutting-edge models due to their significant bias. This raises an alarm possibility: the risk of AMOC collapse is far more severe than thought, and it may be already too late to prevent the forthcoming AMOC collapse. Therefore, we again urgently advocate for the implementation of more robust climate mitigation policies, that is, the net-zero or negative CO₂ emissions.

In addition to the positive bias observed in AMOC stability indicator (Fig. 2a), considerable diversity exists in ∆${F}_{{ov}}$ during the historical period across different climate models^30,31,54. This diversity likely contributes to substantial uncertainty regarding the AMOC threshold across models under identical external forcing conditions. Additionally, model-dependent sensitivity of the AMOC to radiative forcing, combined with intrinsic ocean dynamics, further amplifies these uncertainties. Consequently, we refrain from overinterpreting our findings, as the timing of the AMOC tipping point is model-dependent given identical scenario. Rather than focusing on the exact timing, we first emphasize the possibility that the future fate of the AMOC under a mitigated warming scenario is highly sensitive around the tipping point of the AMOC to the internal climate variability and even slight delays in mitigation actions.

In our CESM1 ensemble experiments, the sign change of ∆${F}_{{ov}}$ consistently coincides with the ensemble divergence between recovery and collapse trajectories (Fig. 2b), suggesting that it remains a useful diagnostic of the relative tendency of AMOC evolution under transient forcing. However, we note that the ∆${F}_{{ov}}$ indicator was originally formulated in equilibrium contexts to characterize AMOC stability regimes¹³, and it may not necessarily indicate a tipping point in transient cases. We therefore acknowledge that ∆${F}_{{ov}}$ < 0 should not be overinterpreted as a definitive indicator of instability in transient scenarios. Instead, in this study, we use the ∆${F}_{{ov}}$ to illustrate relative shifts in stability regimes and to provide dynamical context for the ensemble divergence. While the timing of sign change of ∆${F}_{{ov}}$ aligns well with the tipping behavior in our experiments (Figs. 2 and 3), its application and interpretation in transient scenarios should be regarded with caution and need to be better understood through further investigations across models and theoretical frameworks.

Methods

CESM1 CO₂ stabilization simulations

We perform three sets of experiments using the CESM1, which consists of the atmosphere (Community Atmospheric Model version 5), ocean (Parallel Ocean Program version 2), ice (Community Ice Code version 4), and land models (Community Land Model version 4). The atmosphere and land models operate with ~1° resolution, with 30 vertical levels. The ocean and ice models use a nominal 1° resolution (the meridional resolution is ~1/3° near the equator), with 60 vertical levels. 1. Present-day control simulation (PD): The first experiment, referred to as “PD”, is a control simulation representative of present-day conditions. This experiment spans 900 years with a constant atmospheric CO₂ concentration of 367 ppm, providing a stable baseline for comparison and serving as the initial condition for subsequent experiments. 2. Global warming simulation (1pctCO2): The second experiment, denoted as “1pctCO2”, simulates a typical global warming scenario in which atmospheric CO₂ concentration increases by 1% per year over a period of 140 years until it quadruples relative to the initial value (black line in Fig. 1a, corresponding to model years 2001 to 2140). This experiment comprises 28 ensemble members, each initialized from different macro states within the PD simulation to capture variability arising from internal climate dynamics. More specifically, after discarding the first 600 years of the PD run as model spin-up, we selected initial conditions from the subsequent 300 years according to distinct phases of leading multi-decadal climate variability modes. Six ensemble members were initialized during maximum positive phases of the Atlantic Multidecadal Oscillation, and six during maximum negative phases. Similarly, eight ensemble members were drawn from maximum positive phases of the Pacific Decadal Oscillation, and eight from maximum negative phases. This macro-perturbation design ensures that the ensemble adequately samples uncertainties linked to multi-decadal variability in both the Atlantic and Pacific basins, thereby providing a diverse set of initial conditions for assessing climate evolution, including AMOC under external forcing. This setup has the advantage that all ensemble members are generated from the same climate model under identical external forcing. Consequently, the ensemble spread arises solely from differences in the initial conditions, allowing us to isolate the role of internal climate variability without conflating uncertainties arising from model structure or physics. It is noteworthy that the initial states at the time of branching do not determine the tipping behavior in our simulations, confirming that the ensemble divergence arises from the accumulation of stochastic internal variability during the integration rather than from differences in the initial conditions themselves, which is a central finding of this study. 3. CO₂ stabilization simulation(${{{EXP}}}_{{{branched}}\,{{year}}}$): The third and primary focus of this study is a climate mitigation experiment. In this setup, the atmospheric CO₂ concentration is stabilized at specific levels for an additional 200 years. These stabilization levels are selected based on seven distinct branching points from the 1pctCO2 simulation, representing different CO₂ concentrations at various stages of the warming trajectory (colored horizontal parallels in Fig. 1a). The four experiments with relatively lower CO₂ stabilization levels (blue-colored in Fig. 1a) each consist of 3 ensemble members. The remaining three experiments, associated with higher CO₂ stabilization levels (green, brown, and purple in Fig. 1a), each include 28 ensemble members, allowing for a robust exploration of variability under stronger forcing conditions. To facilitate clarity, each stabilization experiment is labeled as “${{EXP}}_{{branched}\,{year}}$” based on the year of branching from the 1pctCO2 simulation. This labeling convention is consistently used throughout the paper to refer to specific experiments.

AMOC stability indicator

The freshwater transport by the AMOC at a certain latitude ($\phi$) is calculated as

$${F}_{{ov}}\left(\phi \right)=-\frac{1}{{S}_{0}}\int\limits^{0}_{-D}\left[\int\limits^{E}_{W}{v}^{*}{dx}\right]\left[\left\langle S\right\rangle -{S}_{0}\right]{dz}$$

(1)

where ${S}_{0}=$ 34.8 psu is a reference salinity, ${v}^{*}=v-\hat{v}$ is the baroclinic meridional ocean velocity, $\hat{v}$ is the barotropic meridional ocean velocity (section average), and $\left\langle S\right\rangle$ indicate the zonal mean salinity. The AMOC stability indicator $(\Delta {F}_{{ov}})$ is measured as the difference of ${F}_{{ov}}$ across the southern and northern boundaries of the Atlantic as follows:

$$\Delta {F}_{{ov}}={F}_{{ovS}}-{F}_{{ovN}}$$

(2)

where ${F}_{{ovS}}$ denotes the freshwater transport at the southern boundary ($34^{\circ} S$) and ${F}_{{ovN}}$ indicates the transport at the northern boundary, which is the sum of the transport across the three sections between Arctic and Atlantic^18,55: the Canadian Arctic Archipelago $({F}_{{ovCAA}})$, the Fram Strait (${F}_{{ovFS}}$), and the Barents Sea Opening (${F}_{{ovBSO}}$).

$${F}_{{ovN}}={F}_{{ovCAA}}+{F}_{{ovFS}}+{F}_{{ovBSO}}$$

(3)

Note that the calculation of ${F}_{{ov},{CAA}}$, ${F}_{{ov},{FS}}$, and ${F}_{{ov},{BSO}}$ basically follows the Eq. (1), but with ${v}^{*}$ normal to each section. Integration and average are then performed along each section.

Surface density flux

As in previous studies^56,57, the surface density flux $({f}_{{density}})$ at the sea surface is composed of thermal and haline contributions, and it can be expressed as

$${f}_{{density}}=-\frac{\alpha }{{C}_{p}}Q-\beta \frac{{SSS}}{1-{SSS}}F$$

(4)

where the first and second terms are thermal and haline contribution, respectively. $\alpha$ and $\beta$ indicate thermal expansion and haline contraction coefficients, respectively. ${C}_{p}$, $Q$, ${SSS}$, and $F$ are the specific heat capacity of seawater, net surface heat flux, sea surface salinity, and net freshwater flux, respectively.