Report

Multiyear Prediction of Monthly Mean Atlantic Meridional Overturning Circulation at 26.5°N

See allHide authors and affiliations

Science  06 Jan 2012:
Vol. 335, Issue 6064, pp. 76-79
DOI: 10.1126/science.1210299

Abstract

Attempts to predict changes in Atlantic Meridional Overturning Circulation (AMOC) have yielded little success to date. Here, we demonstrate predictability for monthly mean AMOC strength at 26.5°N for up to 4 years in advance. This AMOC predictive skill arises predominantly from the basin-wide upper-mid-ocean geostrophic transport, which in turn can be predicted because we have skill in predicting the upper-ocean zonal density difference. Ensemble forecasts initialized between January 2008 and January 2011 indicate a stable AMOC at 26.5°N until at least 2014, despite a brief wind-induced weakening in 2010. Because AMOC influences many aspects of climate, our results establish AMOC as an important potential carrier of climate predictability.

Variations in Atlantic Meridional Overturning Circulation (AMOC) can substantially affect northward ocean heat transport and therefore European and North Atlantic climate (13). Through its influence on sea surface temperature (SST), AMOC is furthermore thought to influence climate phenomena such as Sahel droughts and North Atlantic hurricane activity (46). In the near term (interannual to decadal time scales), AMOC and other climate variations are influenced by the combination of anthropogenic forcing, natural forcing, and internal variability; near-term climate predictions must hence be started (initialized) from the present ocean state reflecting the present phase of internal variability. Multiyear climate predictions have, however, so far been limited to predicting surface temperature variations (710) and hurricane frequency (11); whether there is multiyear prediction skill for any element of the large-scale atmospheric or oceanic circulation has remained unclear. This paper demonstrates prediction skill for AMOC at 26.5°N for up to 4 years in advance.

Owing to a dearth of long-term AMOC observations, AMOC predictability has hitherto been addressed exclusively in a “perfect model” framework. Results of a model simulation were used as a surrogate for observations, and the degree of similarity between the surrogate observations and an initialized simulation was interpreted as an estimate of AMOC predictability. These studies have shown that AMOC strength is potentially predictable for up to a decade, with potential skill varying among different models (1215). Here, we take advantage of the first half-decade-long observed estimate of AMOC at 26.5°N [the RAPID-MOCHA (Meridional Overturning Circulation and Heatflux Array) projects (16, 17)] to quantify the predictive skill of initialized multiyear predictions performed with the ECHAM5/MPI-OM coupled climate model (18). In further contrast to previous work, we focus on the monthly varying AMOC instead of multiyear averages; this focus is warranted both because of potential applications such as hurricane predictions (11) and because there is clear model evidence that European climate is influenced by year-to-year changes in AMOC (3).

At multiyear to decadal time scales, the memory and, hence, the potential for predictability of the climate system are thought to reside in the ocean. The historical subsurface ocean data are limited in both space and time, and some have been recently shown to be systematically biased (19). Moreover, the current ocean reanalyses display a broad spectrum of AMOC mean state and variability (20, 21), although they use the same observational database. Therefore, we take an alternative route to ocean reanalysis and initialize the coupled model European Centre Hamburg Model 5 (ECHAM5)–Max Planck Institute Ocean Model (MPI-OM) from an ensemble of MPI-OM ocean-only runs that are forced with the atmospheric state of the National Centers for Environmental Prediction–National Center for Atmospheric Research reanalysis (22) for the period 1948 to 2010 (23). We thus ensure dynamical consistency of the forecast model and the initialized ocean state; moreover, forced experiments with MPI-OM have already been used successfully to reconstruct the observed variability of the Nordic Seas Overflows (24), a main contributor to North Atlantic Deep Water that feeds the lower limb of AMOC.

The skill of any prediction system is assessed by using it retrospectively; forecasts are made employing observations only up to some point in the past, and the time period after that point is then used to evaluate the quality of this retrospective forecast (“hindcast”). We have performed ensemble hindcasts starting in January of each year from 2004 to 2007 (23). AMOC strength in the hindcasts closely follows the observations for up to 4 years; the best resemblance to the phase of the observed variability is simulated by the hindcast ensemble mean (Fig. 1). Visual inspection already shows that the hindcast quality is not the same for every start year, a notion now made quantitative.

Fig. 1

Multiyear AMOC hindcasts compared with observations and noninitialized climate model simulations. AMOC strength is the zonally integrated northward flow at 26.5°N above 1000-m depth. The observed transports are from RAPID/MOCHA (red); the individual hindcasts and their ensemble mean are shown in gray and black, respectively. AMOC from the noninitialized climate simulations (ens20C) is shown in the upper plots in light blue for individual experiments and dark blue for the ensemble mean.

We measure the skill of the hindcasts primarily with the anomaly correlation coefficient (COR) between the observations and the individual hindcasts or their ensemble mean. We then compare COR against the benchmark skill of an ensemble of noninitialized model simulation (hereafter ens20C) and of the persistence, a commonly used statistical forecast (23). We consider lead times of 1 to 4 years within each of the 10-year hindcasts and find that, for many start dates, AMOC variations are predicted with significantly greater skill by the initialized hindcasts than by the persistence forecast for all lead times (Fig. 2). The ensemble-mean hindcasts initialized in each January of 2004 to 2006 have mostly high COR, ranging between 0.8 and 0.95 for year 1, 0.5 to 0.8 for year 2, around 0.3 to 0.8 for year 3, and 0.45 to 0.55 for year 4. In contrast, the start date January 2007 gives only low COR skill around 0.3. The ens20C ensemble mean obtains low COR skill of 0.4.

Fig. 2

COR skill scores of monthly mean AMOC for the ensemble mean hindcasts started in January 2004, January 2005, January 2006, and January 2007 are shown by black, violet, blue, and yellow filled symbols (respectively) connected by solid lines. COR skill scores of persistence forecast started in January 2005, January 2006, and January 2007 are shown by violet, blue, yellow open symbols (respectively) connected by dashed lines. A persistence forecast can not be issued at January 2004 due to the lack of observational data. Different symbols are used to indicate COR skill scores at various lead times: year 1, triangles; year 2, squares; year 3, circles; and year 4, diamonds. The 10% significance level is plotted as a dashed black line, and the averaged COR skill score of the ensemble mean noninitialized climate simulations is shown as a dashed gray line.

We diagnose skill improvement through initialization not only from higher COR but also from a tendency toward a more realistic amplitude of AMOC fluctuations (Fig. 1). We note that the hindcast ensemble mean underestimates the observed variability to a larger degree than do the individual hindcasts; conversely, the hindcast ensemble mean has considerably higher COR skill than individual hindcasts (fig. S1). Persistence forecasts, by contrast, do very well on AMOC amplitude but usually lose COR skill after 1 year (Fig. 2 and fig. S2).

The hindcast time series (Fig. 1) clearly show that the dominant AMOC signal that is captured better by the hindcasts than by ens20C arises from the seasonal cycle. The observed time series is too short to distinguish robustly between the mean seasonal cycle and monthly mean deviations thereof; the standard error of the seasonal AMOC is ±2 Sv (25), which is close to the typical monthly mean AMOC deviation from the mean seasonal cycle. Therefore, we cannot at present distinguish between predictability of climatological and anomalous seasonality.

Because it is AMOC seasonality that is most clearly improved through our initialization, one might argue that this particular skill increase is not difficult to achieve. We counter this challenge by noting that, unlike the seasonal cycle in temperature, the observed AMOC seasonal cycle, with a maximum in October and a minimum in April, has highly nontrivial causes (25, 26). The largest contribution to observed AMOC seasonality arises from the baroclinic upper-mid-ocean transport between the Bahamas and Africa, which in turn is dominated by density variations near the eastern boundary (25, 26). These seasonal density variations are coherent to depths of 1400 m (26), implying a substantial role of internal ocean dynamics in the seasonal cycle of AMOC. These observational facts both give credence to our claim of enhanced AMOC hindcast skill and suggest that AMOC predictability chiefly arises from the predictability of the upper-mid-ocean transport.

To investigate this suggestion, we first note that AMOC at 26.5°N can be estimated as the sum of three independently obtained transport components: upper-mid-ocean transport (MO), Ekman transport (EK), and Florida Current transport (FC) (16, 17, 27). MO is derived from thermal wind balance (23). We find that skilful hindcasts of MO can be made for up to 3 years (Fig. 3A), and we find much less sensitivity to start year; MO prediction is more robust than is AMOC prediction. MO skill improvement is also documented in the generally higher and more realistic amplitude of hindcast variability, compared to ens20C (figs. S3 and S4).

Fig. 3

COR skill scores of monthly mean MO (A) and upper ocean zonal density difference (Δρu) (B). COR skill scores for the ensemble mean hindcasts started in January 2004, January 2005, January 2006, and January 2007 are shown by black, violet, blue, and yellow filled symbols (respectively) connected by solid lines. COR skill scores of persistence forecast started in January 2005, January 2006, and January 2007 are shown by violet, blue, and yellow open symbols (respectively) connected by dashed lines. A persistence forecast can not be issued at January 2004 due to the lack of observational data. Different symbols are used to indicate COR skill scores at various lead times: year 1, triangles; year 2, squares; year 3, circles; and year 4, diamonds. The 10% significance level is plotted as a dashed black line, and the averaged COR skill score of the ensemble mean noninitialized climate simulations is shown as a dashed gray line.

In both observations and hindcasts, we see a very strong connection between MO and the zonal density difference across the basin, averaged over the depth range 200 m to 1000 m (Δρu) (fig. S5). We find in our hindcasts significant COR skill for Δρu for up to 3 years (Fig. 3B and fig. S5). This skill both explains the origin of the MO hindcast skill and makes plausible the greater robustness of MO hindcast skill for different start years when compared with AMOC hindcast skill. AMOC hindcast skill is also influenced by EK and FC, both of which are less predictable than MO (fig. S6 and SOM text). Hence, EK and FC can diminish the AMOC hindcast skill, as is the case for the hindcasts started in January 2007, or enhance it, as is the case for the hindcasts started in January 2004 and 2005 (fig. S6 and SOM text).

The observed seasonal cycles in AMOC and eastern-boundary density at 26.5°N are thought to be driven primarily by seasonal wind-stress-curl variations near the eastern boundary (25, 26). To find multiyear predictability for a wind-driven climate quantity might appear surprising, because we do not expect predictability for wind stress itself. We speculate, however, that the seasonally varying wind-stress curl near the eastern boundary might repeatedly imprint itself onto eastern-boundary density. Hence, knowledge of density at any given time would reflect forcing over a longer period in the past, and predictive skill for density would not imply predictive skill for wind stress.

With hindcast skill established, we are now in a position to produce AMOC forecasts. For each January from 2008 until 2011, we construct an ensemble of nine forecasts spanning 10 years. We consider all these simulations forecasts because the RAPID/MOCHA estimate is currently only available until the beginning of April 2009, and the period of overlap is very short. The existence of such a conceptual “gray zone”—where the ocean model forcing and, hence, our initialization as well as EK are known but MO is not—is inevitable for as long as in situ ocean interior measurements cannot be made in real time.

For all start years, the ensemble-mean forecasts until 2014 indicate a generally stable AMOC (Fig. 4). However, the forecast initialized in 2010 shows a pronounced AMOC minimum in March 2010 that arises from a minimum in EK (fig. S7), which in turn is induced by an extremely negative North Atlantic Oscillation in winter 2009–2010 (28). The real AMOC minimum in March 2010 may turn out to be even deeper than predicted, because our ensemble mean underpredicts AMOC amplitude (fig. S2). We are confident, however, that the AMOC minimum in March 2010 will be a short-lived phenomenon; our confidence is based on the insensitivity of our AMOC and MO forecasts to the start year.

Fig. 4

Multiyear predictions of AMOC transport. RAPID/MOCHA time series are shown in red; ensemble mean forecasts are shown in dark gray, light blue, dark blue, and green for the forecasts starting in January 2008, January 2009, January 2010, and January 2011, respectively. The pale shading represents the 95% confidence intervals of the nine-member forecast ensemble initialized in January 2008, January 2009, January 2010, and January 2011.

We cannot readily generalize our results for 26.5°N to other latitudes; recent studies reported a change in the character of AMOC fluctuations around 40°N, with a strong decadal component to the north and enhanced higher-frequency variability to the south (2931). However, for 26.5°N, we have established AMOC hindcast skill, we understand that this skill arises from the mid-ocean transport, and we confidently predict a stable AMOC at least until the end of 2014. Moreover, our findings demonstrate that skill in climate prediction arises not only from the large ocean thermal inertia but potentially also from the long time scales of internal ocean dynamics.

Supporting Online Material

www.sciencemag.org/cgi/content/full/335/6064/76/DC1

Materials and Methods

SOM Text

Figs. S1 to S7

References

References and Notes

  1. Materials and methods are available as supporting material on Science Online.
  2. Acknowledgments: This work was supported by the Bundesministerium für Bildung und Forschung North Atlantic project (D.M.) and the Deutsche Forschungsgemeinschaft–funded Cluster of Excellence CliSAP (J.B. and W.A.M.). We thank M. Giorgetta, U. Mikolajewicz, and B. Stevens for comments on the manuscript. All model simulations were performed at the German Climate Computing Centre (DKRZ). Data from the RAPID-WATCH MOC monitoring project are funded by the Natural Environment Research Council and are freely available from www.noc.soton.ac.uk/rapidmoc.
View Abstract

Navigate This Article