Abstract
The Atlantic meridional overturning circulation (AMOC) is a major tipping element in the climate system and a future collapse would have severe impacts on the climate in the North Atlantic region. In recent years weakening in circulation has been reported, but assessments by the Intergovernmental Panel on Climate Change (IPCC), based on the Climate Model Intercomparison Project (CMIP) model simulations suggest that a full collapse is unlikely within the 21st century. Tipping to an undesired state in the climate is, however, a growing concern with increasing greenhouse gas concentrations. Predictions based on observations rely on detecting earlywarning signals, primarily an increase in variance (loss of resilience) and increased autocorrelation (critical slowing down), which have recently been reported for the AMOC. Here we provide statistical significance and datadriven estimators for the time of tipping. We estimate a collapse of the AMOC to occur around midcentury under the current scenario of future emissions.
Similar content being viewed by others
Introduction
A forthcoming collapse of the Atlantic meridional overturning circulation (AMOC) is a major concern as it is one of the most important tipping elements in Earth’s climate system^{1,2,3}. In recent years, model studies and paleoclimatic reconstructions indicate that the strongest abrupt climate fluctuations, the DansgaardOeschger events^{4}, are connected to the bimodal nature of the AMOC^{5,6}. Numerous climate model studies show a hysteresis behavior, where changing a control parameter, typically the freshwater input into the Northern Atlantic, makes the AMOC bifurcate through a set of codimension one saddlenode bifurcations^{7,8,9}. Stateoftheart Earthsystem models can reproduce such a scenario, but the intermodel spread is large and the critical threshold is poorly constrained^{10,11}. Based on the CMIP5 generation of models, the AR6 IPCC report quotes a collapse in the 21st century to be very unlikely (medium confidence)^{12}. Among CMIP6 models, there is a larger spread in the AMOC response to warming scenarios, thus an increased uncertainty in the assessment of a future collapse^{13}. There are, however, model biases toward overestimated stability of the AMOC, both from tuning to the historic climate record^{14}, poor representation of the deep water formation^{15}, salinity and glacial runoff^{16}.
When complex systems, such as the overturning circulation, undergo critical transitions by changing a control parameter λ through a critical value λ_{c}, a structural change in the dynamics happens. The previously statistically stable state ceases to exist and the system moves to a different statistically stable state. The system undergoes a bifurcation, which for λ sufficiently close to λ_{c} can happen in a limited number of ways rather independent from the details in the governing dynamics^{17}. Besides a decline of the AMOC before the critical transition, there are earlywarning signals (EWSs), statistical quantities, which also change before the tipping happens. These are critical slowing down (increased autocorrelation) and, from the FluctuationDissipation Theorem, increased variance in the signal^{18,19,20}. The latter is also termed “loss of resilience”, especially in the context of ecological collapse^{21}. The two EWSs are statistical equilibrium concepts. Thus, using them as actual predictors of a forthcoming transition relies on the assumption of quasistationary dynamics.
The AMOC has only been monitored continuously since 2004 through combined measurements from moored instruments, induced electrical currents in submarine cables and satellite surface measurements^{22}. Over the period 2004–2012, a decline in the AMOC has been observed, but longer records are necessary to assess the significance. For that, careful fingerprinting techniques have been applied to longer records of sea surface temperature (SST), which, backed by a survey of a large ensemble of climate model simulations, have found the SST in the Subpolar gyre (SG) region of the North Atlantic (area marked with a black contour in Fig. 1a) to contain an optimal fingerprint of the strength of the AMOC^{23,24,25}.
Figure 1b shows the SG and the GM SSTs obtained from the Hadley Centre Sea Ice and Sea Surface Temperature data set (HadISST)^{26}. Figure 1c shows the SG anomaly, and Fig. 1d shows the GM anomaly with a clear global warming trend in the last half of the record. The AMOC fingerprint for the period 1870–2020 is shown in Fig. 1e. This is the basis for the analysis. It has been reported^{11,27} that this and similar AMOC indices show significant trends in the mean, the variance and the autocorrelation, indicating early warning of a shutdown of the AMOC. However, a trend in the EWSs within a limited period of observation could be a random fluctuation within steadystate statistics. Thus, for a robust assessment of the shutdown, it is necessary to establish a statistical confidence level for the change above the natural fluctuations. This is not easily done given only one, the observed realization of the approach to the transition. Here we establish such a measure of the confidence for the variance and autocorrelation and demonstrate that variance is the more reliable of the two. A further contribution is an estimator of not only whether a transition is approaching but also the time when the critical transition is expected to occur. The strategy is to infer the evolution of the AMOC solely on observed changes in mean, variance and autocorrelation. The typical choice of control parameter is the flux of freshwater into the North Atlantic. River runoff, Greenland ice melt and export from the Arctic Ocean are not well constrained^{28}; thus, we do not assume the control parameter known. Boers^{27} assumes the global mean temperature T to represent the control parameter. Although T has increased since ~1920 (Fig. 1d), the increase is not quite linear with time. All we assume here is that the AMOC is in an equilibrium state prior to a change toward the transition. The simplest uninformed assumption is that the change is sufficiently slow and that the control parameter approaches the (unknown) critical value linearly with time. This assumption is confirmed by a close fit of the estimated model to the observed AMOC fingerprint. Although we make no explicit assumptions, the primary driver of climate change, the logarithm of the atmospheric CO_{2} concentration, does, in fact, increase close to linearly with time in the industrial period^{29}. Our results are robust without making specific assumptions regarding the driver of the AMOC.
In this work, we show that a transition of the AMOC is most likely to occur around 20252095 (95% confidence interval).
Results
Modeling and detecting the critical transition
Denote the observed AMOC fingerprint by x(t) (Fig. 1e). We model it by a stochastic process X_{t}, which, depending on a control parameter λ < 0, is at risk of undergoing a critical transition through a saddlenode bifurcation for λ = λ_{c} = 0. The system is initially in a statistically stable state, i.e., it follows some stationary distribution with constant λ = λ_{0}. We are uninformed about the dynamics governing the evolution of X_{t} but can assume effective dynamics, which, with λ sufficiently close to the critical value λ_{c} = 0, can be described by the stochastic differential equation (SDE):
where \(m=\mu \sqrt{\lambda/A}\) and μ is the stable fixed point of the drift, A is a time scale parameter, B_{t} is a Brownian motion and σ^{2} scales the variance. Disregarding the noise, this is the normal form of the codimension one saddlenode bifurcation^{17} (see “Methods”). The squareroot dependence of the stable state: \(\mu m \sim \sqrt{{\lambda }_{c}\lambda }\) is the main signature of a saddlenode bifurcation. It is observed for the AMOC shutdown in oceanonly models as well as in coupled models, see Fig. 2, in strong support of Eq. (1) for the AMOC.
At time t_{0}, λ(t) begins to change linearly toward λ_{c} = λ(t_{c}) = 0:
where Θ[t] is the Heaviside function and τ_{r} = t_{c} − t_{0} > 0 is the ramping time up to time t_{c}, where the transition eventually will occur. Time t_{c} is denoted the tipping time; however, an actual tipping can happen earlier than t_{c} due to a noiseinduced tipping. As the transition is approached, the risk of noiseinduced tipping (ntipping) prior to t_{c} is increasing and, at some point, making the EWSs irrelevant for predicting the tipping. The probability for ntipping can, in the small noise limit, be calculated in closed form, \(P(t,\lambda )=1\exp (t/{\tau }_{n}(\lambda ))\), with mean waiting time \({\tau }_{n}(\lambda )=(\pi /\sqrt{\lambda})\exp (8\lambda {}^{3/2}/3{\sigma }^{2})\) (see “Methods”).
The mean and variance are calculated from the observations as the control parameter λ(t) is possibly changing. These EWSs are inherently equilibrium concepts and statistical; thus, a time window, T_{w}, of a certain size is required for a reliable estimate. As the transition is approached, the differences between the EWSs and the preramping values of the variance and autocorrelation (baseline) increase; thus, a shorter window T_{w} is required for detecting a difference. Conversely, close to the transition critical slowing down decreases the number of independent points within a window, thus calling for a larger window for reliable detection. Within a short enough window, [t − T_{w}/2, t + T_{w}/2], we may assume λ(t) to be constant and the noise small enough so that the process (1) for given λ is well approximated by a linear SDE, the Ornstein–Uhlenbeck process^{30}. A Taylor expansion around the fixed point μ(λ) yields the approximation
where \(\mu (\lambda )=m+\sqrt{\lambda/A}\) and \(\alpha (\lambda )=2\sqrt{A\lambda}\) is the inverse correlation time. For fixed λ, the process is stationary, with mean μ, variance γ^{2} = σ^{2}/2α and onelag autocorrelation \(\rho=\exp (\alpha {{\Delta }}t)\) with step size Δt = 1 month. As λ(t) increases, α decreases, and thus variance and autocorrelation increase. From μ, γ^{2} and ρ the parameters of Eq. (1) are determined: \(\alpha=\log \rho /{{\Delta }}t\), σ^{2} = 2αγ^{2}, A = α/2(μ − m) and \(\lambda={({\sigma }^{2}/4{\gamma }^{2})}^{2}/A\). Closed form estimators for μ, γ^{2} and ρ are obtained from the observed time series within a running window by maximum likelihood estimation (MLE) (Supplementary text S1, see also ref. ^{31}).
The uncertainty is expressed through the variances of the estimators \({\hat{\gamma }}^{2}\) and \(\hat{\rho }\) obtained from the observations within a time window T_{w}. The hats indicate that they are estimators and thus stochastic variables with variances around the true values. Detection of an EWS at some chosen confidence level q (such as 95 or 99%) requires one of the estimates \({\hat{\gamma }}^{2}\) or \(\hat{\rho }\) for a given window to be statistically different from the baseline values \({\hat{\gamma }}_{0}^{2}\) or \({\hat{\rho }}_{0}\), which depend on the window size as well as how different the EWSs are from their baseline values.
Time scales in earlywarning signals
The detection of a forthcoming transition using statistical measures involves several time scales. The primary internal time scale is the autocorrelation time, t_{ac}, in the steady state. The ramping time τ_{r} over which the control parameter changes from the steady state value to the critical value sets an external time scale. For given α(λ) and qpercentile, the required time window T_{w}(q, α) to detect a change from baseline in EWSs at the given confidence level q is given in the closed form in the next section (Eq. (7) for variance and Eq. (8) for autocorrelation). The approach to the collapse and the involved time scales are schematically summarized in Fig. 3, while they are calculated in Fig. 4a, where the required window size T_{w} at the 95% confidence level is plotted as a function of λ for the variance (red curve) and autocorrelation (yellow curve). These are plotted together with the mean waiting time for ntipping, t_{noise}, (blue curve). With T_{w} = 50 years, increased variance can only be detected after the time when λ(t) ≈ −1.2 (crossing of red and reddashed curves). At that time, a window of approximately 75 years is required to detect an increase in autocorrelation, making variance the better EWS of the two. When λ ≈ −0.4, the mean waiting time for ntipping is smaller than the data window size. Thus, the increased variance can be used as a reliable EWS in the range −1.2 < λ(t) < −0.4, indicated by the green band. How timely an early warning this is depends on the speed at which λ(t) is changing from λ_{0} to λ_{c}, i.e., the ramping time τ_{r}. A set of 1000 realizations has been simulated with λ_{0} = −2.82 and τ_{r} = 140 years, indicated by the time labels on top of Fig. 4a. Ten of these realizations are shown in Fig. 4b on top of the stable and unstable branches of fixed points of model (1) (the bifurcation diagram). Figure 4c (d) shows the variance (autocorrelation) calculated from the realizations within a running 50year window (shown in Fig. 4c). The solid black line is the baseline value for λ = λ_{0}, while the solid blue line is the increasing value for λ = λ(t). The calculated 95% confidence level for the measurement of the EWS within the running window is shown by the dashed black and blue lines, respectively. The corresponding light blue curves are obtained numerically from the 1000 realizations. The green band in Fig. 4c corresponds to the green band in Fig. 4a and shows where early warning is possible in this case.
Statistics of earlywarning signals
The variances of the estimators are approximately (see Supplementary text S1).
where T_{w} = nΔt is the observation window.
The question is then how large T_{w} needs to be to detect a statistically significant increase compared to the estimated baseline values \({\hat{\gamma }}_{0}^{2}\) and \({\hat{\rho }}_{0}\). For a given estimate \({\hat{\gamma }}^{2}\), the estimated difference from the baseline variance is
and the estimated difference from the baseline autocorrelation is
Since the two EWSs, \({\hat{\gamma }}^{2}\) and \(\hat{\rho }\), are treated on an equal footing, in the following, we let \(\hat{\psi }\) denote either of the estimators (given explicitly in Supplementary text S1, Eqs. (S5) or (S6)). The standard error is \(s(\hat{\psi })={{{{{{{\rm{Var}}}}}}}}{(\hat{\psi })}^{1/2}\) (Eq. (4)) and \(\hat{{{\Delta }}}\) denotes either of the two estimated differences (5) or (6). The null hypothesis is that λ = λ_{0}, or equivalently α = α_{0}. The null distribution of \(\hat{\psi }\) is assumed to be Gaussian (confirmed by simulations). A quantile q from the standard Gaussian distribution expresses the acceptable uncertainty in measuring the statistical quantity ψ. We thus get that \(\hat{{{\Delta }}} \, < \, qs(\hat{\psi })\) at the qconfidence level (95%, 99% or similar) under the null hypothesis. To detect an EWS at the qconfidence level based on measuring ψ at time t, we require that \(\hat{{{\Delta }}}(t) \, > \, q(s(\hat{\psi }(t))+s({\hat{\psi }}_{0}))\), which, solved for T_{w} gives for variance:
and for autocorrelation,
Substituting \({\alpha }_{0}=2\sqrt{A{\lambda }_{0}}\) and \(\alpha (t)=2\sqrt{A\lambda (t)}\) provides the time window T_{w} needed to detect an EWS at time t with large probability. Eqs. (7) and (8) are illustrated in Fig. 4a (red and yellow curves), where it is seen that detecting a significant increase in variance requires a shorter data window than detecting a significant increase in autocorrelation. Two times \(s(\hat{\psi }(t))\) around the mean of the ramped variance and two times \(s({\hat{\psi }}_{0})\) around baseline values are illustrated in Fig. 4c, d (dashed lines). Once a trace leaves the baseline confidence interval, a statistically significant change is detected, and when the two dashed lines cross, 95% of the traces have detected an EWS (Eqs. (7) and (8)).
Predicting a forthcoming collapse of the AMOC
The AMOC fingerprint shown in Fig. 1e (replotted in Fig. 5a) shows an increased variance, γ^{2}, and autocorrelation, ρ, plotted in Fig. 5b, c as functions of the midpoint of a 50year running window, i.e., the EWS obtained in 2020 is assigned to the year 1995. The estimates leave the confidence band of the baseline values (pink area) around the year 1970. This is not the estimate of t_{0}, which happened earlier and is still to be estimated; it is the year where EWSs are statistically different from baseline values. The estimates after 1970 stay consistently above the upper limit of the confidence interval and show an increasing trend, and we thus conclude that the system is moving toward the tipping point with high probability.
To estimate the tipping time once it has been established that the variance and autocorrelation are increasing, we use two independent methods to check the robustness of our results: (1) Momentbased estimator that uses the variance and autocorrelation estimates within the running windows. (2) Approximate MLE directly on model (1)(2) with no running window. The advantage of the first method is that it has less model assumptions; however, it is sensitive to the choice of window size. The advantage of the second method is that it uses the information in the data more efficiently given model (1)(2) is approximately correct, it has no need for a running window and does not assume stationarity after time t_{0}. In general, MLE is statistically the preferred method of choice, giving the most accurate results with the lowest estimation variance.
The first method, the moment estimator of the tipping time obtains, within the running window, the parameters α(t) (Fig. 5d) and σ^{2} (Fig. 5e) of the linearized dynamics, Eq. (3), and thus also γ^{2}(t). Within the running window, the data are detrended before estimation by subtracting a linear regression fit in order not to falsely inflate the variance estimates caused by deviations from stationarity. Then we obtain Aλ(t) from σ^{2} and γ^{2}(t) (Fig. 5f) using that \(A\lambda (t)={({\sigma }^{2}/4{\gamma }^{2}(t))}^{2}\). This is consistent with a linear ramping of λ(t) beginning from a constant level λ_{0} at a time t_{0}. By sweeping t_{0} from 1910 to 1950 and T_{w} from 45 to 65 years, we obtain Aλ_{0} and τ_{r} from the least square error fit to the data. This shows a single minimum at t_{0} = 1924 and T_{w} = 55 years (Fig. 6e). Setting t_{0} = 1924, we obtain t_{c} from a linear fit (regressing λ on t) from the crossing of the xaxis (λ_{c} = 0). This is shown in Fig. 5f (red line). This yields −Aλ_{0} = 2.34 year^{−2} and τ_{r} = 133 years. Thus, the tipping time is estimated to be in the year 2057, shown in Fig. 5f. Since we have only obtained the combined quantity \(A\lambda={({\sigma }^{2}/4{\gamma }^{2})}^{2}\), we still need to determine A and m in Eq. (1). We do that from the best linear fit to the mean level \(\mu=m+\sqrt{\lambda/A}\) observing that \(\mu=m+\sqrt{A\lambda}(1/A)=m+({\sigma }^{2}/4{\gamma }^{2})(1/A)\). The estimates are shown by the red curves in Fig. 5a–f. The red dot in Fig. 5a is the tipping point, and the dashed line in Fig. 5b is the asymptote for the variance. With the parameter values completely determined, the confidence levels are calculated: The twostandard error levels around the baseline values of the EWS are shown by purple bands in Fig. 5b, c. Thus, both EWSs show increases beyond the twostandard error level from 1970 and onward.
The second method, the approximate MLE of the tipping time, is applied to model (1)–(2). The likelihood function is the product of transition densities between consecutive observations. However, the likelihood is not explicitly known for this model, and we therefore approximate the transition densities. From the data before time t_{0}, approximation (3) is used, where exact MLEs are available (Supplementary text S1). This provides estimates of the parameters λ_{0}, m as a function of parameter A, as well as the variance parameter σ^{2}. To estimate A and τ_{r}, the observations after time t_{0} are used. After time t_{0}, the linear approximation (3) is no longer valid because the dynamics are approaching the bifurcation point, and the nonlinear dynamics will be increasingly dominating. The likelihood function is the product of transition densities, which we approximate with a numerical scheme, the Strang splitting, which has shown to have desirable statistical properties for highly nonlinear models, where other schemes, such as the Euler–Maruyama approximation is too inaccurate^{32} (Supplementary text S2). Using t_{0} = 1924, the optimal fit is the same as the moment method, t_{c} = 2057, with a 95% confidence interval 2025–2095.
Confidence intervals for the estimate of the tipping time are obtained by bootstrap. The likelihood approach provides asymptotic confidence intervals; however, these assume that the likelihood is the true likelihood. To incorporate also the uncertainty due to the data generating mechanism (1) not being equal to the Ornstein–Uhlenbeck process (3) used in the likelihood, we chose to construct parametric bootstrap confidence intervals. This was obtained by simulating 1000 trajectories from the original model with the estimated parameters and repeating the estimation procedure on each data set. Empirical confidence intervals were then extracted from the 1000 parameter estimates. These were indeed larger than the asymptotic confidence intervals provided by the likelihood approach, however, not by much. Histograms of the bootstrapped estimates are shown in Fig. 6a–d. The histogram in Fig. 6a is the tipping year, repeated in yellow in 5f.
The mean of the bootstrapped estimates of the tipping time is 〈t_{c}〉 = 2050, and the 95% confidence interval is 2025–2095. The small discrepancy in the mean is probably due to the approximate model used for estimation being different from the datagenerating model (1), confirming that the linear model still provides valid estimates even if the true dynamics are unknown. To test the goodnessoffit, normal residuals (see “Methods”) were calculated for the data. These are plotted in Fig. 6f as a quantilequantile plot. If the model is correct, the points fall close to a straight line. The model is seen to fit the data well, further supporting the obtained estimates.
Discussion
We have provided a robust statistical analysis to quantify the uncertainty in observed EWSs for a forthcoming critical transition. The confidence depends on how rapidly the system is approaching the tipping point. With this, the significance of the observed EWSs for the AMOC has been established. This is a stronger result than just observing a significant trend in the EWS by, say, Kendall’s τ test^{27,33}. Here we calculate when the EWS are significantly above the natural variations. Furthermore, we have provided a method to not only determine whether a critical transition will happen but also an estimate of when it will happen. We predict with high confidence the tipping to happen as soon as midcentury (2025–2095 is a 95% confidence range). These results are under the assumption that the model is approximately correct, and we, of course, cannot rule out that other mechanisms are at play, and thus, the uncertainty is larger. However, we have reduced the analysis to have as few and sound assumptions as possible, and given the importance of the AMOC for the climate system, we ought not to ignore such clear indicators of an imminent collapse.
The hysteresis simulations gathered in the model intercomparison^{34} are equilibrium runs, for which a prediction of a future collapse does obviously not apply. Likewise, for the simulations specified in the CMIP6 experiment. It could though be relevant to evaluate our method on stateoftheart climate model simulations with linearly ramped external forcing and different ramping speeds in order to obtain the modelspecific confidence in early prediction of the collapse judged solely from the EWSs.
Though we have established firm statistical methods to evaluate the confidence in the observed EWS, we can at present not rule out the possibility that a collapse will only be partial and not lead to a full collapse of the AMOC as suggested by some models: Note in Fig. 2, the “MPM” in the top panel and the “MOM hor” in the bottom panel both seem to show only partial tipping prior to the tipping to the complete shutdown of the AMOC. This result is also found in a more recent ocean model^{35}. Furthermore, a high speed of ramping, i.e., a high speed at which the critical value of the control parameter is approached, could also increase the probability of tipping^{36}. This scenario is the case of rateinduced tipping^{37}. Even with these reservations, this is indeed a worrisome result, which should call for fast and effective measures to reduce global greenhouse gas emissions in order to avoid the steady change of the control parameter toward the collapse of the AMOC (i.e., reduce temperature increase and freshwater input through ice melting into the North Atlantic region). As a collapse of the AMOC has strong societal implications^{38}, it is important to monitor the flow and EWS from direct measurements^{39,40,41}.
Methods
To obtain the AMOC fingerprint, two steps are required: The seasonal cycle in the SST is governed by the surface radiation independent from the circulation and thus removed by considering the monthly anomalies, where the mean over the period of recording of the month is removed. Second, there is an ongoing positive linear trend in the SST related to global warming, which is also not related to circulation. This is compensated for by subtracting 2 × the global mean (GM) SST anomaly (small seasonal cycle removed). This differs slightly from ref. ^{23}, where 1 × the GM SST was subtracted. The “translation” from the proxy SST temperature and AMOC flow is 0.26 SV/K [ref. ^{23}, Fig. 3]. Here we have taken into account that the warming is not globally homogeneous: The warming in the SG region is larger than the global mean due to polar amplification. The way we have estimated this effect is by comparing the proxy with the AMOC estimates covering the period 1957–2004 from the socalled MOC_{z} reported in the review by ref. ^{42}. This shows a drop of 3 SV in that period. Minimizing the difference between the proxy SST_{SG}β SST_{GM} and this more direct measurement with respect to β, we get β = 1.95 ≈ 2 rather than β = 1 used by ref. ^{23}. The factor 2 is thus the optimal value for the polar amplification^{43} obtained by calibrating to recent direct measurements^{42}. The original and our calibrated proxies are shown in Fig. 7.
To check the robustness with respect to the AMOC fingerprint record, we repeated the analysis, subtracting 1x and 3x GM SST from the SG SST. Subtracting 3x GM SST only changes estimate and the confidence intervals by a few years, whereas subtracting 1x GM SST delays the tipping with 25 years, but the overall trend and conclusions do not change. The results are given in Table 1. In the reanalysis, we fixed t_{0} = 1924.
Estimator of the tipping time and model control
The AMOC fingerprint is assumed to be observations from a process X_{t} given as a solution to Eqs. (1) and (2), and we wish to estimate the parameters θ = (A, m, λ_{0}, τ_{r}, σ) from observations (x_{0}, x_{1}, …, x_{n}) before time t_{0} and observations (y_{0}, y_{1}, …, y_{n}) after time t_{0}. These equations cannot be explicitly solved, and the exact distribution of X_{t} is not explicitly known. A standard way to solve this is to approximate the transition density by a Gaussian distribution obtained by the Euler–Maruyama scheme. However, the estimators obtained from the Euler–Maruyama pseudolikelihood are known to be biased, especially in nonlinear models^{32}. We estimate by a twostep procedure, approximating the stationary distribution before time t_{0} by an Ornstein–Uhlenbeck process, of which exact maximum likelihood estimators are available (see Supplementary text S1), and using Strang splitting for the nonstationary and nonlinear part after time t_{0}, using methods proposed in ref. ^{32}, see Supplementary text S2 for details.
To test the model fit, uniform residuals, u_{i}, i = 1, …, n were calculated for the AMOC data using the estimated parameters from the MLE method as follows. The model assumes that observation x_{i} follows some distribution function \({F}_{i,\hat{\theta }}\) for the estimated parameter values \(\hat{\theta }\). If this is true, then \({u}_{i}={F}_{i,\hat{\theta }}({x}_{i})\) is uniformly distributed on (0, 1). Transforming these residuals back to a standard normal distribution provides standard normally distributed residuals if the model is true. Thus, a normal quantilequantile plot reveals the model fit. The points should fall close to a straight line. The reason for making the detour around the uniform residuals is twofold. First, since the data is not stationary, each observation follows its own distribution, and residuals cannot be directly combined. Second, since the model is stochastic, standard residuals are not welldefined, and observations should be evaluated according to their entire distribution, not only the distance to the mean.
Noiseinduced tipping
The drift term in Eq. (1) is the negative gradient of a potential, f(x, λ) = − ∂_{x}V(x, λ) = − (A(x−m)^{2} + λ) with V(x, λ) = A(x−m)^{3}/3 + (x − m)λ. For λ < 0, the drift has two fixed points, \(m\pm \sqrt{\lambda/A}\). The point \(m+\sqrt{\lambda/A}\) is a local minimum of the potential V(x, λ) and is stable, whereas \(m\sqrt{\lambda/A}\) is a local maximum and unstable. The system thus has two basins of attraction separated by \(m\sqrt{\lambda/A}\), with a drift toward either \(m+\sqrt{\lambda/A}\) or −∞ dependent on whether \({X}_{t} > m\sqrt{\lambda/A}\) or \({X}_{t} < m\sqrt{\lambda/A}\). We denote the two basins of attraction, the normal and the tipped state, respectively. When λ = 0, the normal state disappears, and the system undergoes a bifurcation and X_{t} will be drawn toward −∞.
Due to the noise, the process can escape into the tipped state by crossing over the potential barrier \({{\Delta }}(\lambda )=V(m\sqrt{\lambda/A},\,\lambda )\,\,V(m+\sqrt{\lambda/A},\,\lambda )=4\lambda {}^{3/2}/3{A}^{1/2}\). Assume X_{t} to be close to \(m+\sqrt{\lambda/A}\) at some time t, i.e., in the normal state. The escape time will asymptotically (for σ → 0) follow an exponential distribution such that
where P(t, λ) is the probability of observing an escape time shorter than t for a given value of λ. The mean noiseinduced escape time τ_{n}(λ) is^{44,45}:
Assume that the rate of change of λ(t) follows Eq. (2), then for τ_{r} < τ_{n}(λ), the waiting time for a random crossing is so long that a crossing will not happen before a bifurcationinduced transition happens (btipping). If τ_{r} > τ_{n}(λ), a noiseinduced tipping is expected before the bifurcation point is reached. Since τ_{n}(λ) decreases with increasing λ, at some point, the two time scales will end up matching.
Normal form of the saddlenode bifurcation
Consider the general dynamical equation
where x is a variable and λ is a (fixed) parameter. A point x_{0} with f(x_{0}, λ) = 0 is a fixed point or steady state. A fixed point is stable/unstable if \({\partial }_{x}f{(x,\lambda )}_{x={x}_{0}}\) is negative/positive; thus, the fixed point is attracting/repelling. If f(x, λ) is not a linear function of x, multiple steady states may exist. A saddlenode bifurcation occurs when changing the control parameter λ through a critical value λ_{c} a stable and an unstable fixed point merge and disappear. The situation is shown in the figure, where the blue surface is f(x, λ), while the gray (null) plane is f(x, λ) = 0. For a constant value of λ, the dynamics is determined by the black curve. The fixed points are determined by the intersection with the null plane (green); the point in the front is the stable fixed point, while the further point is the unstable fixed point. When changing λ toward λ_{c} = 0, the two fixed points merge at the saddlenode bifurcation (m, λ_{c}) (green). The normal form of the saddle node is obtained by expanding f(x, λ) to the lowest order around the point (m, λ_{c}), noting that f(m, λ_{c}) = 0, \({\partial }_{x}f{(x,\lambda )}_{(x,\lambda )=(m,{\lambda }_{c})}=0\) and \({\partial }_{\lambda }f{(x,\lambda )}_{(x,\lambda )=(m,{\lambda }_{c})} < 0\) (see Fig. 8):
where \(A=\frac{1}{2}{\partial }_{{xx}}f{(x,\,\lambda )}_{(x,\lambda )=(m,{\lambda }_{c})}\) and \(\tilde{\lambda }={\partial }_{\lambda }f{(x,\lambda )}_{(x,\lambda )=(m,{\lambda }_{c})}\times (\lambda {\lambda }_{c})\). This is the normal form for the saddlenode bifurcation. Thus, close to the bifurcation point, the stable steady state is
In order to see that this is indeed the case for the AMOC transition also in comprehensive climate models, Fig. 2 is adapted from the model intercomparison study^{34}. The steady state curves obtained are from simulations, with a very slowly changing control parameter (freshwater forcing). The top panel shows oceanonly models, while the bottom panel shows atmosphereocean models. The curves are, even away from the transition, surprisingly well fitted by Eq. (13). Note that for some models, the transition happens before the critical point, as should be expected from noiseinduced transitions. Note also that the data has been smoothed such that increasing variance close to the transition is not visible. This observation strongly supports the assumption of a saddlenode bifurcation, while it also shows that (m, λ_{c}) (black dots) are quite different between models, thus calling for reliable determination from observations.
Data availability
Data can be found in the following repository: https://doi.org/10.17894/ucph.b8f99b67d4e64a2eb51800bddeed323b.
Code availability
Computer code (Matlab and R) can be found in the following repository: https://doi.org/10.17894/ucph.b8f99b67d4e64a2eb51800bddeed323b.
References
Manabe, S. & Stouffer, R. J. Two stable equilibria of a coupled oceanatmosphere model. J. Clim. 1, 841–866 (1988).
Rahmstorf, S. Bifurcations of the Atlantic thermohaline circulation in response to changes in the hydrological cycle. Nature 378, 145–149 (1995).
Lenton, T. M. et al. Tipping elements in the Earth’s climate system. Proc. Natl Acad. Sci. USA 105, 1786–1793 (2008).
Dansgaard, W. et al. Evidence for general instability of past climate from a 250kyr icecore record. Nature 364, 218–220 (1993).
Vettoretti, G., Ditlevsen, P., Jochum, M. & Rasmussen, S. O. Atmospheric CO_{2} control of spontaneous millennialscale ice age climate oscillations. Nat. Geosci. 15, 300–306 (2022).
Ganopolski, A. & Rahmstorf, S. Rapid changes of glacial climate simulated in a coupled climate model. Nature 409, 153–158 (2001).
Wood, R. A., Rodríguez, J. M., Smith, R. S., Jackson, L. C. & Hawkins, E. Observable loworder dynamical controls on thresholds of the Atlantic meridional overturning circulation. Clim. Dyn. 53, 6815–6834 (2019).
Hawkins, E. et al. Bistability of the Atlantic overturning circulation in a global climate model and links to ocean freshwater transport. Geophys. Res. Lett. 38, L10605 (2011).
Weijer, W. et al. Stability of the Atlantic meridional overturning circulation: a review and synthesis. J. Geophys. Res. Oceans 124, 5336–5375 (2019).
Mecking, J., Drijfhout, S., Jackson, L. & Andrews, M. The effect of model bias on Atlantic freshwater transport and implications for AMOC bistability. Tellus A: Dyn. Meteorol. Oceanogr. 69, 1299910 (2017).
Rahmstorf, S. et al. Exceptional twentiethcentury slowdown in Atlantic Ocean overturning circulation. Nat. Clim. Change 5, 475–480 (2015).
MassonDelmotte, V. et al. IPCC, 2021: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (Cambridge University Press, 2021).
Gong, X., Liu, H., Wang, F. & Heuzé, C. Of Atlantic meridional overturning circulation in the CMIP6 project. Deep Sea Res. Part II: Top. Stud. Oceanogr. 206, 105193 (2022).
Hofmann, M. & Rahmstorf, S. On the stability of the Atlantic meridional overturning circulation. Proc. Natl Acad. Sci. USA 106, 20584–20589 (2009).
Heuzé, C. North Atlantic deep water formation and AMOC in CMIP5 models. Ocean Sci. 13, 609–622 (2017).
Liu, W., Xie, S.P., Liu, Z. & Zhu, J. Overlooked possibility of a collapsed Atlantic meridional overturning circulation in warming climate. Sci. Adv. 3, e1601666 (2017).
Guckenheimer, J. & Holmes, P. Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields (Springer, 1986).
Kubo, R. The fluctuationdissipation theorem. Rep. Prog. Phys. 29, 255–284 (1966).
Ditlevsen, P. D. & Johnsen, S. Tipping points: early warning and wishful thinking. Geophys. Res. Lett. 37, L19703 (2010).
Boulton, C., Allison, L. & Lenton, T. Early warning signals of Atlantic meridional overturning circulation collapse in a fully coupled climate model. Nat. Commun. 5, 5752 (2014).
Scheffer, M. et al. Earlywarning signals for critical transitions. Nature 461, 53–59 (2009).
Smeed, D. A. et al. Observed decline of the Atlantic meridional overturning circulation 20042012. Ocean Sci. 10, 29–38 (2014).
Caesar, L., Rahmstorf, S., Robinson, A., Feulner, G. & Saba, V. Observed fingerprint of a weakening Atlantic ocean overturning circulation. Nature 552, 191–196 (2018).
Jackson, L. C. & Wood, R. A. Fingerprints for early detection of changes in the AMOC. J. Clim. 33, 7027 – 7044 (2020).
Latif, M. Reconstructing, monitoring, and predicting multidecadalscale changes in the North Atlantic thermohaline circulation with sea surface temperature. J. Clim. 17, 1605–1614 (2004).
Rayner, N. A. et al. Global analyses of sea surface temperature, sea ice, and night marine air temperature since the late nineteenth century. J. Geophys. Res. 108, 4407 (2003).
Boers, N. Observationbased earlywarning signals for a collapse of the Atlantic meridional overturning circulation. Nat. Clim. Change 11, 680–688 (2021).
Yang, Q. et al. Recent increases in Arctic freshwater flux affects Labrador Sea convection and Atlantic overturning circulation. Nat. Commun. 7, 10525 (2016).
Keeling, C. D. et al. Atmospheric CO_{2} and ^{13}CO_{2} exchange with the terrestrial biosphere and oceans from 1978 to 2000: observations and carbon cycle implications. A History of Atmospheric CO_{2} and Its Effects on Plants, Animals, and Ecosystems (eds Ehleringer, J. R. et al.) (Springer, 2005).
Hasselmann, K. Stochastic climate models. Tellus 28, 473–485 (1976).
Ditlevsen, S., Cencerrado Rubio, A. & Lansky, P. Transient dynamics of Pearson diffusions facilitates estimation of rate parameters. Commun. Nonlinear Sci. Numer. Simul. 82, 105034 (2020).
Pilipovic, P., Samson, A. & Ditlevsen, S. Parameter estimation in nonlinear multivariate stochastic differential equations based on splitting schemes. Preprint at https://arxiv.org/abs/2211.11884 (2022).
Michel, S. L. L. et al. Early warning signal for a tipping point suggested by a millennial Atlantic multidecadal variability reconstruction. Nat. Commun. 13, 5176 (2022).
Rahmstorf, S. et al. Thermohaline circulation hysteresis: a model intercomparison. Geophys. Res. Lett. 32, L23605 (2005).
Lohmann, J., Dijkstra, H. A., Jochum, M., Lucarini, V. & Ditlevsen, P. D. Multistability and intermediate tipping of the Atlantic Ocean circulation. Preprint at https://arxiv.org/abs/2304.05664 (2023).
Lohmann, J. & Ditlevsen, P. Risk of tipping the overturning circulation due to increasing rates of ice melt. Proc. Natl Acad. Sci. USA 118, e2017989118 (2021).
Ashwin, P., Wieczorek, S., Vitolo, R. & Cox, P. Tipping points in open systems: bifurcation, noiseinduced and ratedependent examples in the climate system. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. 370, 1166–1184 (2012).
Kemp, L. et al. Climate endgame: exploring catastrophic climate change scenarios. Proc. Natl Acad. Sci. USA 119, e2108146119 (2022).
Baehr, J. et al. Timely detection of changes in the meridional overturning circulation at 26°N in the Atlantic. J. Clim. 20, 5827–5841 (2007).
Sévellec, F., Dijkstra, H. A., Drijfhout, S. S. & Germe, A. Dynamical attribution of oceanic prediction uncertainty in the North Atlantic: application to the design of optimal monitoring systems. Clim. Dyn. 51, 1517–1535 (2018).
AlexanderTurner, R., Ortega, P. & Robson, J. I. How robust are the surface temperature fingerprints of the Atlantic overturning meridional circulation on monthly time scales? Geophys. Res. Lett. 45, 3559–3567 (2018).
FrajkaWilliams, E. et al. Atlantic meridional overturning circulation: observed transport and variability. Front. Mar. Sci. 6, 260 (2019).
Holland, M. & Bitz, C. Polar amplification of climate change in coupled models. Clim. Dyn. 21, 221–232 (2003).
Berglund, N. Kramers’ law: validity, derivations and generalisations. Markov Process. Relat. Fields 19, 459–490 (2013).
Freidlin, M. & Wentzell, A. Random perturbations of dynamical systems. Grundlehren der mathematischen Wissenschaften (Springer, 1984).
Acknowledgements
This work has received funding under the project Tipping Points in the Earth System (TiPES) from the European Union’s Horizon 2020 research and innovation programme under grant agreement no. 820970. This is TiPES contribution #214. Furthermore, funding was provided by Novo Nordisk Foundation NNF20OC0062958; and European Union’s Horizon 2020 research and innovation program under the Marie SkłodowskaCurie grant agreement No 956107, “Economic Policy in Complex Environments (EPOC)”.
Author information
Authors and Affiliations
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Nature Communications thanks Chris Boulton and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Ditlevsen, P., Ditlevsen, S. Warning of a forthcoming collapse of the Atlantic meridional overturning circulation. Nat Commun 14, 4254 (2023). https://doi.org/10.1038/s4146702339810w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s4146702339810w
This article is cited by

Managing a low urban emissions world
npj Climate Action (2024)

Greenland’s glaciers are retreating everywhere and all at once
Nature (2024)

Impact of industrial versus biomass burning aerosols on the Atlantic Meridional Overturning Circulation
npj Climate and Atmospheric Science (2024)

Ubiquitous acceleration in Greenland Ice Sheet calving from 1985 to 2022
Nature (2024)

Reflecting on the Science of Climate Tipping Points to Inform and Assist Policy Making and Address the Risks they Pose to Society
Surveys in Geophysics (2024)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.