Initialized decadal predictions of the rapid warming of the North Atlantic Ocean in the mid 1990s


Corresponding author: J. I. Robson, NCAS-Climate, Department of Meteorology, University of Reading, Reading RG6 6BB, UK. (


[1] In the mid 1990s the North Atlantic subpolar gyre (SPG) warmed rapidly, with sea surface temperatures (SST) increasing by 1°C in just a few years. By examining initialized hindcasts made with the UK Met Office Decadal Prediction System (DePreSys), it is shown that the warming could have been predicted. Conversely, hindcasts that only consider changes in radiative forcings are not able to capture the rapid warming. Heat budget analysis shows that the success of the DePreSys hindcasts is due to the initialization of anomalously strong northward ocean heat transport. Furthermore, it is found that initializing a strong Atlantic circulation, and in particular a strong Atlantic Meridional Overturning Circulation, is key for successful predictions. Finally, we show that DePreSys is able to predict significant changes in SST and other surface climate variables related to the North Atlantic warming.

1. Introduction

[2] North Atlantic Sea Surface Temperature (SST) variability is thought to be an important driver of climate variability in North America and Europe [Sutton and Hodson, 2005; Knight et al., 2006]. In the mid 1990s the North Atlantic warmed significantly, with SSTs increasing by 1°C in the North Atlantic subpolar gyre (SPG) [Robson et al., 2012]. The warming of the SPG is thought to have had important effects on Greenland Ice sheets [Holland et al., 2008], ocean ecosystems [Hátún et al., 2009] and potentially the frequency of Hurricanes [Smith et al., 2010].

[3] Robson et al. [2012] and Yeager et al. [2012] showed that the 1990s North Atlantic warming can largely be attributed to increased oceanic meridional heat transport (MHT), associated with a strengthened circulation (especially the Atlantic Meridional Overturning circulation (AMOC)). The increase in the AMOC was largely forced by the persistent positive phase of the North Atlantic Oscillation (NAO) [Hurrell, 1995] in the late 1980s and early 1990s, which acted to cool the SPG and drive increased deep water formation [Robson et al., 2012]. In winter 1995/1996 the NAO was strongly negative, coinciding with the rapid change in SPG heat content, and has since remained largely neutral. However, although the negative NAO of 1995/1996 likely contributed to the timing, rapidity and magnitude of the warming (through circulation changes and reduced surface heat fluxes), Robson et al. [2012] suggest that the surge in MHT related to the preceding positive NAO was the primary cause.

[4] The 1990s rapid warming of the North Atlantic SPG is an excellent case study for assessing decadal prediction systems due to its large magnitude, and because the preconditioning of the ocean by the positive NAO suggests that the warming was potentially predictable [Robson et al., 2012]. Predicting the evolution of the climate on decadal timescales is now a major interest [Meehl et al., 2009], and predictions of the North Atlantic SPG appear to benefit significantly from initialisation [Robson, 2010; Smith et al., 2010]. However, to build confidence, understanding the improvement in skill is key. Robson [2010] showed that the UK Met Office's Decadal Prediction System (DePreSys) appeared to predict the 1990s warming, and that initializing the MHT was important. However, the predictions were found to be sensitive to spurious adjustments from the initial conditions [Robson, 2010]. Yeager et al. [2012]found that the warming was predictable when initializing a coupled model with states from a forced ocean and sea-ice integration, essentially due to the initialisation of increased MHT. This study builds on the work ofRobson [2010] and Yeager et al. [2012] by assessing the performance of a new version of the DePreSys prediction system for the 1990s warming event. In contrast to the system studied by Yeager et al. [2012], DePreSys assimilates actual ocean observations as anomalies. Lastly, this study also extends Yeager et al. [2012] by examining predictions of the climate impacts related to the warming.

2. DePreSys

[5] This study examines the Perturbed Physics Ensemble version of DePreSys [Smith et al., 2010], which uses 9 variants of the HadCM3 model (1 standard, 8 with perturbed parameters) to represent some model uncertainty in predictions. The parameter perturbations introduce radiative imbalances to the models, and so flux adjustments are used to maintain a realistic climatology for SST and surface salinity. Nine member ensemble hindcasts (1 member per model version) start every November from 1960–2005 after observed anomalies are assimilated into the ocean and atmosphere. More specifically, anomalies are computed from pre-calculated gridded data sets for the ocean (3D temperature and salinity from the Met Office Ocean analysis [Smith and Murphy, 2007]) and atmosphere (3D winds, 3D temperature and sea level pressure from ERA-40 [Uppala et al., 2005]) and are assimilated by relaxing the model to its own climatology plusthe observed anomalies. The ocean and atmosphere climatologies are defined as 1951–2006 and 1958–2001 respectively. Hindcasts are forced with (historical) anthropogenic and (projected) natural forcings. More specifically, total solar irradiance was assumed to follow the previous 11-year solar cycle, and initial volcanic aerosol was reduced to zero with an e-folding timescale of one year (i.e. no future eruptions). After 2000, SRES A1B is used. Importantly, this version of DePreSys has much reduced drift following initialization compared to the previous version used inRobson [2010](not shown) and bias corrections are not applied to predictions. A control experiment is also performed (NoAssim). NoAssim is identical to DePreSys (i.e. 9 member ensembles), except the initial conditions are taken from free-running transient simulations (using the same model variants as DePreSys), which were initialized in preindustrial conditions and forced with historical anthropogenic and natural forcings. A full description of DePreSys and NoAssim is found inSmith et al. [2010].

3. Results

3.1. Warming of the Subpolar Gyre

[6] Figure 1shows the 0–500 m SPG (60°W-10°W, 50°N-65°N) average temperature anomaly from the observations (black), and DePreSys hindcasts (red). Over the entire hindcast period, DePreSys is marginally more skillful than persistence (seeFigure 1f). However, DePreSys hindcasts starting between 1990–1996 clearly show an initial ensemble mean warming, similar to the observed warming of the SPG. However, the ensemble mean warming is earlier, and slower, than that seen in the observations. This early warming appears consistent with the role of the NAO in the warming [Robson et al., 2012]. Given the low predictability of the NAO, we would not expect DePreSys to predict the persistent positive NAO in the early 1990s, nor the negative NAO event in 1995/1996.

Figure 1.

(a) Evolution of subpolar gyre (60°W-10°W, 50°N-65°N) 0–500 m mean temperature anomalies (°C), from the Met Office ocean analysis (black), the DePreSys hindcast started in 1988 (red), and the NoAssim hindcasts (blue). All anomalies are relative to a 1951–2006 climatology. For DePreSys, the thick and thin red lines show the ensemble mean anomaly and 1σ ensemble spread respectively for each start date. For NoAssim, the thick line represents the average of all hindcast start dates, and the thin blue lines show the 1σ spread of the NoAssim anomalies. (b, c, d and e) The same as Figure 1a but for the 1990, 1992, 1994 and 1998 DePreSys hindcasts. (f) The anomaly correlation skill score as a function of lead time (solid line) for DePreSys (red), NoAssim (blue) and persistence (purple) predictions for the first year, and then rolling 3 year averages. Skill is evaluated for all hindcasts started between 1960–2005 against Met Office ocean observations. The thin dashed lines show the 5–95% confidence interval in which differences in skill are not significant [Smith et al. [2010].

[7] Averaging NoAssim hindcasts from different start dates together shows the transient simulation of the SPG, which warms monotonically over the period (see Figure 1). Therefore, NoAssim predicts the SPG heat content variability poorly (anomaly correlation of ∼0.2 for year one, see Figure 1f). However, the observations generally lie within the spread of the transient anomalies, except in the late 1980s and early 1990s when the SPG is cold. Thus, it is important to establish whether DePreSys is simply drifting back to the model's preferred state, or whether it captures the mechanisms that are believed to have operated in the real world.

[8] Figure 2 compares the spatial patterns of 0–500 m average temperature anomalies from hindcasts started in 1994, to those observed (the comparison is similar for all the hindcasts started in the early 1990s (not shown)). In 1995–1997, the observed SPG was cool, with warm anomalies in the Gulf Stream Extension (GSE, Figure 2a). By 1998–2000 warm anomalies reach across the GSE and into the eastern SPG (Figure 2b). In DePreSys, the anomalies are similar to those observed, but appear to be evolving more slowly (i.e. anomalies are absent from the north SPG in 1998–2000, Figure 2d). However, in NoAssim there are no warm anomalies moving north into the SPG (Figures 2e and 2f). Thus, Figure 2 suggests that the warming in DePreSys is consistent with the observed changes, whereas the warming in NoAssim is not.

Figure 2.

(a) The observed 1995–1997 mean of 0–500 m average temperature anomalies, relative to a 1951–2006 climatology, from the Met Office ocean analysis. (b) The same as Figure 2a but for the 1998–2000 mean. (c and d) The same as Figures 2a and 2b but for the November 1994 DePreSys hindcast. The black contours show where the anomalies are significant relative to the model's internal variability (taken from year 1 of the NoAssim hindcasts) where p ≤ 0.05 (dashed) and p ≤ 0.01 (dotted) based on a t-test. (e and f) The same as Figures 2c and 2d but for the November 1994 NoAssim hindcast.

3.2. Heat Budget and Ocean Heat Transports

[9] To quantify the role of ocean heat transports in the warming of the SPG in the hindcasts, a heat budget analysis similar to that in Robson et al. [2012] is performed. Figures 3a and 3b shows the anomalous contribution of ocean heat transport convergence (calculated from the seasonal mean temperature and meridional velocity, HO, black) and atmospheric heat loss (HA, orange) to SPG heat content change integrated over the first 3 years of each hindcast. When HO is greater than HA the heat content of the SPG increases (i.e. heat content change =HOHA).

Figure 3.

(a) Ensemble mean anomalies of the time-integrated column (surface-bottom) oceanic heat transport convergence (Ho, black) [J] and the atmospheric heat loss (Ha, orange) for the SPG region (50°N-65°N) for each DePreSys hindcast. Fluxes are integrated over the first 3 years and the x-axis represents the hindcasts start year. Anomalies are relative to the mean of year 1–3 flux integrals from NoAssim (1960–2005). WhenHo is larger than Hathe SPG heat content increases. Error bars show the 95% confidence interval based on the ensemble spread. (b) The same as Figure 3a, but for NoAssim hindcasts. (c) The latitude-time evolution of annual-mean MHT anomalies [PW] for the 1994 DePreSys hindcast. Anomalies are relative to the mean of all year 1 from NoAssim hindcasts. The black contours show significant anomalies compared to the models variability (represented by year 1 of all NoAssim hindcasts), where p ≤ 0.05 (dashed) and p ≤ 0.01(dotted) based on a t-test. (d) The same as Figure 3c but for the 1994 NoAssim hindcast. (e and f) The same as Figures 3c and 3d but for AMOC anomalies [Sv]. (g) The same as Figure 3c but for MHT associated with anomalies in the velocity ( inline image). (h) The same as Figure 3g but for anomalies in temperature ( inline image).

[10] DePreSys hindcasts started in the early 1990s are unusual in that consecutive hindcasts have increases in heat content. A decrease in HA is seen in the early 1990s, which is driven largely by a decrease in latent and sensible heat (not shown), consistent with cooler initial SSTs. However, for hindcasts started in 1994–1996, HO accounts for ∼70% of the change in heat content over the first 3 years. Therefore, the intialisation of ocean heat transport was key for DePreSys to predict the warming. In contrast, there is no increase in HO in NoAssim, and a small decreasing trend in HA over the whole hindcast period. Therefore, the slow monotonic warming of the NoAssim hindcasts (see Figure 1) appears largely due to decreased surface flux cooling.

[11] Examining DePreSys further, the increase in HO in the early/mid 1990s is dominated by changes in the overturning component of the ocean heat transport (calculated as in Bryan [1969], not shown). Figure 3also shows the anomalous AMOC (taken from the zonal-mean stream function at a depth of 1000 m) and MHT in the 1994 hindcasts (the results are consistent with other start dates in the early 1990s). The AMOC is initialized anomalously strong in DePreSys (Figure 3e), due to positive density anomalies along the deep western boundary [Robson, 2010]. DePreSys is also initialized with an anomalously strong MHT between 30°N-60°N (Figure 3c). Thus, the increase in HO is driven by anomalously strong MHT into the southern SPG, consistent with an increased AMOC (as similar to Robson et al. [2012] and Yeager et al. [2012]). In contrast, positive anomalies in MHT or AMOC are absent in NoAssim (Figures 3d and 3f).

[12] To examine the importance of anomalies in velocities ( inline image) compared to anomalies in temperatures ( inline image), the MHT is further broken down (as in Robson et al. [2012], where inline image and inline imagerepresents the 1961–2006 climatology of 3D ocean velocity and temperature, and ′ denotes grid-point anomalies relative to climatology). The decomposition shows that inline image dominates the increase in MHT in the first few years (Figure 3g), especially for the overturning component (not shown). After the 2nd year inline image contributes to the anomalous MHT into the SPG (Figure 3h). The vT′ term did not play a substantial role (not shown). Therefore, the initialisation of a strong ocean circulation, and in particular the AMOC, was key for the successful predictions of the rapid nature of the SPG warming, whereas the advection of the temperature anomalies was important to sustain the warming.

3.3. Climate Impacts

[13] To explore the predicted climate impact of the warming, DePreSys hindcasts that predict a warming (1991–1996 start dates, see Figure 3) are compared to earlier hindcasts (1960–1990). Comparing differences between hindcasts for the same lead times (i.e. years 1–4) removes the need to define a climatology, or to consider a mean bias correction. Also, comparing the difference between DePreSys and NoAssim hindcasts reveals the impact of initialization, and removes forced trends. Differences are tested for significance by using a two-sided t-test after the degrees-of-freedom (n, e.g. the total number of years 1–4 from both hindcast subsets, minus two) are reduced to take account of serial correlation of the NoAssim hindcasts. The serial correlation, α, is defined at each grid point as the mean of the individual correlations between consecutive NoAssim hindcasts for each lead time and model variant. The effective degrees-of-freedom (neff) is thus inline image [Zwiers and von Storch, 1995].

[14] Figure 4shows the March–November mean for SST, surface air temperature (SAT), Surface Level Pressure (SLP) and Precipitation (PRE) for DePreSys hindcasts that start in 1991–1996 (DJF is omitted because ocean-atmosphere interactions in this season are dominated by the atmosphere forcing the ocean [Visbeck et al., 1998]). Relative to NoAssim, DePreSys predicts significant warm SSTs in the GSE in years 1–4, with cool anomalies in the SPG (Figure 4a), but only weak anomalies in other fields (Figures 4b and 4c). However, for years 6–9, the impact of initialisation has grown (Figures 4g–4i). Significant warm SST anomalies extend across the Atlantic, and into the Tropical North Atlantic (Figure 4g). Significant anomalies are also seen in the wider climate. For example, warm SAT anomalies are present across North America and Northern Europe (Figure 4g) and low SLP anomalies are present in the Northern Tropical Atlantic (Figure 4h), with associated shift in the precipitation pattern (Figure 4i). Significantly reduced precipitation is also predicted over North America (Figure 4i). These anomalies in surface climate are similar to those associated with warm Atlantic SSTs [Sutton and Hodson, 2005; Knight et al., 2006]. Importantly, these anomalies are unusually large for this lead time (see Figure S1 in the auxiliary material), highlighting the unusual state of the Atlantic ocean in the early 1990s hindcasts.

Figure 4.

(a) The average March–November SST and SAT [°C] from years 1–4 of DePreSys hindcasts started between 1991–1996, minus the average of years 1–4 from DePreSys hindcasts started between 1960–1990, relative to the same difference in NoAssim. (b and c) The same as Figure 4a but for differences in SLP [hPa] and PRE [% of the mean] respectively. (d, e and f) The same as Figures 4a, 4b and 4c but for the 1987–1995 observed anomalies, relative to a 1961–1991 climatology. The observed SST and SAT has been linearly detrended with a trend fitted over 1951–2009 at each grid point. (g, h and i) The same as Figures 4a, 4b and 4c but for years 6–9. (j, k and l) The same as Figures 4d, 4e and 4f, but now for 1996–2005. The stippling shows where the differences are significant at the p ≤ 0.1 using a two-sided t-test.

[15] Figure 4 also shows the impact of the warming in the observations by comparing the 9 years before and after the warming for SST (from HadISST [Rayner et al., 2003]), SLP (from HadSLP2 [Allan and Ansell, 2006]), SAT and Precipitation (from CRU TS3.1 [Mitchell and Jones, 2005]). As previously discussed, the SPG warms more slowly in DePreSys than in the observations (Figure 1). Therefore, we use the hindcasts as an analogue for the observed warming, where forecast years 1–4 and 6–9 are analogous to an anomalously cool (i.e. years 1987–1995) and warm (i.e. years 1997–2005) SPG respectively. The predictions agree with observations in several aspects. After the warming, the SST increases over much of the Atlantic sector and SAT increases along the eastern North America and West Europe (compare Figures 4d and 4j) similarly to predictions. A significant reduction in SLP is also present between 10-40°N (compareFigures 4e and 4k), which is captured by DePreSys. Precipitation anomalies are noisy, but the observed negative anomaly over the central U.S.A (Figure 4l) is consistent with the DePreSys prediction for years 6–9 (Figure 4i).

4. Conclusions and Discussion

[16] By examining predictions using the Perturbed Physics Ensemble version of the UK Met Office Decadal Prediction System (DePreSys), it has been shown that

[17] 1. The rapid warming of the North Atlantic subpolar gyre in the mid 1990s could have been predicted in advance.

[18] 2. The success of DePreSys predictions, and especially their ability to capture the rapid nature of the change, was largely due to the initialisation of strong ocean heat transport and ocean dynamics. In particular, a strong Atlantic Meridional Overturning Circulation (AMOC) was key.

[19] 3. The predicted ensemble mean warming is not as rapid as observed, which is not surprising given the role of the negative NAO winter in 1995/1996 [Robson et al., 2012].

[20] 4. Associated with the warming of the North Atlantic, DePreSys also predicts significant changes in Sea Surface Temperatures (SST), Surface Air Temperatures (SAT), Sea Level Pressure (SLP) and Precipitation (PRE) in the wider Atlantic sector, which broadly agree with the observed changes, even over land in some regions.

[21] This study corroborates the evidence of Robson [2010], Robson et al. [2012] and Yeager et al. [2012], by showing that the North Atlantic warming was predictable years ahead. As in Yeager et al. [2012], the successful predictions are due to the initialization of a strong ocean heat transports. However, this study shows in addition that the initialization of the ocean circulation is important to predict the initial rapid nature of the warming (Figure 3g), and that the successful predictions of heat content change allow predictions of related climate variables, even over land. Therefore, these results provide further evidence that the ocean dynamics played a key role in the North Atlantic warming, and highlight that dynamical predictions of the North Atlantic are possible on multi-year timescales. Finally, this study gives encouragement to the development of skillful predictions of societal-relevant climate variables for up to a decade ahead.

[22] Although DePreSys is able to predict the main features of the observed warming, there are still some differences. For example, DePreSys does not predict the observed shift in the gyre boundaries [Sarafanov et al., 2008], (not shown). HadCM3 is not a perfect model, and the simulation of the important mechanisms could well be sensitive to the resolution of the model [e.g., Hodson and Sutton, 2012], the flux adjustments [Collins et al., 2006], or the initialisation [Robson, 2010]. Finally, the model could be missing important processes or feedbacks, that could have contributed to the observed changes (such as indirect aerosol forcing [Booth et al., 2012]). We suggest that the rapid warming of the North Atlantic will remain an important case study to validate and improve future models and decadal prediction systems.


[23] J.I.R. acknowledges financial support from the NERC funded VALOR project. R.T.S. was funded by NCAS. D.M.S. was supported by the UK Joint DECC/Defra Met Office Hadley Centre Climate Programme (GA01101) and the European Community's 7th framework programme (FP7/2007-2013) under THOR. We thank the anonymous reviewers for their comments.

[24] The Editor thanks the two anonymous reviewers for their assistance in evaluating this paper.