The surface and upper (0–500 m) layers of the subpolar North Atlantic (SPNA) have exhibited notable spatially coherent temperature anomalies across a broad range of timescales1,2,3,4,5. Extreme cold conditions characterised the middle years (2014–2016) of the past decade and have been linked to severe winter surface cooling2,6,7,8. In contrast, SPNA temperatures were particularly warm in the mid-1990s and have been related to changes in both the Atlantic Meridional Overturning Circulation (AMOC)9,10 and the wind-driven circulation11,12,13. At centennial timescales, the SPNA has experienced reduced warming relative to the rest of the globe, a feature referred to as the ‘warming hole’1. There is potential for confusion between the cold conditions arising from interannual variability, particularly in 2014–2015, with the signal from the warming hole as they exhibit similar anomalous patterns in sea surface temperature (SST)7. More generally the causes of cold and warm SPNA anomalies at interannual timescales are still not well understood and that is the focus of the present study. Our main goal is to understand the extent to which surface heat exchange and ocean heat transport (OHT) are responsible for year-to-year changes in mixed layer heat content and determine whether this balance has changed over the past century.

The balance between variations in ocean heat convergence and heat loss through the air-sea interface lies at the heart of the complex mechanism that determines SST and mixed layer heat content14,15,16,17. In our analysis, we first consider the sea surface, for which the long-standing Bjerknes conjecture states that atmospheric variability dominates air-sea heat flux and SST changes at interannual timescales, and that the ocean only plays a strong role at longer decadal timescales18,19. We test whether the interannual component of the conjecture can still be considered valid by analysing SST variability. Then we shift focus from the sea surface to the mixed layer and explore the relative roles of surface heat exchange and ocean heat convergence in setting its heat content. Variations in the energy made available from the ocean heat content (OHC) reservoir can significantly modify atmospheric dynamics and downstream European weather; including the dominant winter mode of variability20 and summer heatwaves2,5,21. We exploit observations and high-resolution-coupled climate model output to establish the strong role played by OHT variability in setting the mixed layer heat content at interannual timescales. Furthermore, we use a century-long SST based reconstruction to reveal that, over the past 60 years, interannual variability of SPNA mixed layer heat content is increasingly being caused by changes in ocean transport rather than surface heat exchange.


Surface temperature extremes

Ocean surface temperature reaches a minimum in March in the extratropical Northern Hemisphere. Thus, we have computed a measure of interannual variability by taking the difference (ΔSST) of March SST between successive years. Considering the well-observed epoch 1990–2019, the most extreme interannual changes occur from 1995–1996 (warming) and 2013–2014 (cooling). Between March 1995 and March 1996, a major warming event extends across much of the central and eastern SPNA (Fig. 1a). What mechanism drove this interannual change? The Bjerknes conjecture requires it to be due primarily to anomalous air-sea heat transfer. However, the observed 1995–1996 surface net heat flux anomaly is close to zero, \({Q}_{n9596}^{{\prime} }\) = 3 ± 2 Wm−2 (see Fig. 1 caption for full \({Q}_{n9596}^{{\prime} }\) definition). Thus, the event sits outside the Bjerknes paradigm for drivers of interannual variability indicating that variations in OHT must have played a major role in its generation as has been suggested by ocean model analysis10. These variations may involve both horizontal and vertical heat transport, and the vertical heat transports may be related to re-emergence of atmospherically forced anomalies from previous winters. In contrast, the 2013–2014 cooling event (Fig. 1b), occurs during a period of intense heat loss7, \({Q}_{n1314}^{{\prime} }\) = −20 ± 2 Wm−2. Given the extreme heat loss, this event is potentially consistent with the Bjerknes conjecture. To determine whether the interannual component of the conjecture can still be considered valid we carry out a correlation analysis of ΔSST with \({Q}_{n}^{{\prime} }\) using the set of March–March events within the period 1990–2019 (see Supplementary Note 1). The analysis shows that the sign of the correlation is positive over the SPNA. However, the ΔSST variance explained by \({Q}_{n}^{{\prime} }\) typically lies in the range 20–50% that implies that ocean processes (vertical mixing, variations in horizontal heat transport convergence) must also make a significant contribution to the SST variability19,22. Thus, the Bjerknes Conjecture does not hold at interannual timescales in the SPNA.

Fig. 1: Interannual North Atlantic surface temperature change.
figure 1

Difference in late winter SST (°C) between successive years a March 1996 minus March 1995; b March 2014 minus March 2013, source the HadISST55 dataset. Net heat flux anomalies (\({Q}_{n}^{{\prime} }\)) are shown by the arrows, positive is ocean heat gain from the atmosphere, negative is heat loss (Wm−2, averaging region 45–60 °N, 40–15 °W, white box, source the NCEP/NCAR28 atmospheric reanalysis). \({Q}_{n9596}^{{\prime} }\) is the monthly net air-sea heat flux anomaly (relative to 1981–2010) averaged over March 1995 to February 1996; likewise for \({Q}_{n1314}^{{\prime} }\) but for March 2013 to February 2014.

In addition to the 1995–1996 event, the variance results noted above suggest that changes in OHT, rather than surface heat flux, are important for interannual variations in ocean surface temperature on other occasions in the SPNA. For further insights, we extend our analysis by considering the ocean mixed layer heat content. It has been established that heat transport variability can take a leading role in other regions of the global ocean particularly in the Tropics and western boundary current regimes14,16,23,24. However, the balance of terms for the mixed layer heat budget in regions of strong water mass transformation like the SPNA25 remains unclear. In particular, the cause of specific year to year changes in mixed layer heat content, that we consider here, has not yet been established as earlier analyses14,16 have taken a climatologically fixed depth. These analyses provided valuable advances regarding the causes of variability in the fixed depth upper OHC. However, a mixed layer study needs to take variations in the layer depth from year to year into account and this is a central element of our approach (see Supplementary Note 2 regarding the importance of considering mixed layer depth variability). The SPNA is a region of intense interannual winter heat loss variability7 and thus one in which the mixed layer heat content may be expected to be largely set by the atmosphere. However, contrary to this expectation, in the next section, we show through analysis of mixed layer heat content variability that surprisingly the OHT plays as strong a role as the surface heat exchange.

Drivers of ocean heat content variability

To examine the role of the ocean further, it is necessary to move beyond SST and consider the heat budget of the mixed layer in the upper ocean which, by definition, has close to the same temperature as the surface26. We determine the change (∆OHC) in March mixed layer OHC between successive years, using a variable mixed layer depth approach, and compare with the integrated net heat flux anomaly (∫\({Q}_{n}^{{\prime} }\)) over the intervening period. The corresponding change (∆OHT) in the combined lateral and vertical OHT into the mixed layer is then determined by the difference of the two observed quantities (∆OHT = ∆OHC – ∫\({Q}_{n}^{{\prime} }\), see Methods). This term represents the combined contributions of a range of processes including diffusive mixing, advection and mixed layer depth variations.

This approach is applied separately to two regions covering the NE (55–65 °N, 40–10 °W) and SE (45–55 °N, 40–15 °W) SPNA. These regions were chosen to test whether the subsequent results are sensitive to which of the NE or SE SPNA is considered and allow for possible regional variations in the relative contributions from ∆OHT and ∫\({Q}_{n}^{{\prime} }\) (see Supplementary Note 3). In practice, the results obtained for each region are similar when considering the full set of yearly changes although different behaviour can occur in individual years. The UKMO EN427 dataset is used to determine ∆OHC and the NCEP/NCAR28 atmospheric reanalysis for ∫\({Q}_{n}^{{\prime} }\). Consistency with the SST variability is observed, as the strongest warming in the entire 30 year period 1990–2019 occurs between March 1995 and March 1996 in both regions (∆OHCNE = 2.6 ± 0.1 ZJ, ∆OHCSE = 2.1 ± 0.1 ZJ, 1 ZJ = 1021 J). Note that here, and subsequently, we use the terms ‘warming’/‘cooling’ for brevity to refer to an increase/decrease in OHC, which can be brought about by changes in both temperature and mixed layer volume (see Supplementary Note 4). The 1995–1996 warming cannot be explained by observations of the integrated surface net heat flux (∫\({Q}_{{nNE}}^{{\prime} }\) = 0.6 ± 0.1 ZJ, ∫\({Q}_{{nSE}}^{{\prime} }\) = −0.1 ± 0.1 ZJ). For both regions, ∆OHC is much larger than ∫\({Q}_{n}^{{\prime} }\) . Thus, the 1995–1996 warming must have been driven nearly entirely by increased OHT (∆OHTNE = 2.0 ± 0.2 ZJ, ∆OHTSE = 2.2 ± 0.1 ZJ).

To explore whether the ocean takes an active role in other years, we plot the variation of ∆OHT with ∫\({Q}_{n}^{{\prime} }\)for all individual March–March periods from 1990–2019 (Fig. 2). The March 1995–1996 warming is not the only occasion on which the ocean plays an active role. For the SE SPNA, |∆OHT| > |∫\({Q}_{n}^{{\prime} }\)| in 7 out of 18 years with strong events (|∆OHT| or |∫\({Q}_{n}^{{\prime} }\)| > 0.5 ZJ); similar results (8 out of 19 years) are obtained for the NE SPNA. Thus, the active ocean controls near-surface ocean temperature variability in about 40% of the strong events.

Fig. 2: Ocean transport and surface flux contributions to ocean heat content change.
figure 2

Interannual change in ocean heat transport (∆OHT) against net heat flux anomaly (∫\({Q}_{n}^{{\prime} }\)) for SE (black crosses) and NE (green circles) SPNA, all years 1991–2019. Labelled lines indicate ocean heat content change (∆OHC = ∆OHT + ∫\({Q}_{n}^{{\prime} }\)), units ZJ. Red (blue) arrows show increasing (decreasing) OHC. Shading indicates ocean transport dominance (grey, |∆OHT| > |∫\({Q}_{n}^{{\prime} }\)|) and surface flux dominance (white, |∫\({Q}_{n}^{{\prime} }\)| > |∆OHT|). One standard deviation error bars included for |∆OHT| or |∫\({Q}_{n}^{{\prime} }\)| > 0.5 ZJ. Symbol labels denote specific events, e.g., SE9596 is SE region March 1996 minus March 1995.

These occasions include notable cold (as well as warm) anomalies. In particular, from March 1994–1995, 76% of NE SPNA heat content change is due to the ocean (∆OHCNE = −1.7 ± 0.1 ZJ, ∆OHTNE = −1.3 ± 0.2 ZJ, ∫\({Q}_{{nNE}}^{{\prime} }\) = −0.4 ± 0.1 ZJ). The most extreme case of surface cooling driven change in OHC is March 2013–14 in the SE SPNA which is consistent with the intense heat loss known to have occurred during this winter7. When the full set of all year-on-year changes is considered, a more complex picture than the traditional one-way atmosphere-led control emerges. Instead, a two-way balance holds in which the ocean is frequently active in setting its own mixed layer heat content variability rather than passively responding to the atmosphere.

Additional insight into the balance between ocean and atmosphere-led control is provided by analysis of a 100-year climate simulation with the high-resolution HadGEM3.1 model (Fig. 3, see Methods for model details). The increase in resolution of this model compared to earlier models enables significantly better representation of Atlantic OHT and SPNA circulation29,30. Here, we find an active role for the ocean in the simulation consistent with the observation-based results discussed above. In particular, for the SE SPNA, |∆OHT| > |∫\({Q}_{n}^{{\prime} }\)| in 43 out of 72 years with strong events (|∆OHT| or |∫\({Q}_{n}^{{\prime} }\)| > 0.5 ZJ). Similar results are obtained for the NE SPNA (|∆OHT| > |∫\({Q}_{n}^{{\prime} }\)| in 39 out of 63 strong event years).

Fig. 3: Time series of modelled and observed ocean heat budget components.
figure 3

a Interannual variation of SE SPNA components from HadGEM3.1 simulation: ∆OHC (black), ∆OHT (red) and ∫\({Q}_{n}^{{\prime} }\) (blue). Grey shading indicates active ocean events (|∆OHT| > |∫\({Q}_{n}^{{\prime} }\)| and |∆OHT| > 0.5 ZJ). b Standard deviation (\(\sigma\)) of all individual March–March ∆OHC, ∫\({Q}_{n}^{{\prime}}\) and ∆OHT values for the full 100-year model simulation (filled squares) and individual decades (crosses). Values determined with the same method from observations for 1990–2019 (filled triangles and crosses). c as (a) but from observations, note the same approach is used to determine ∆OHT for both the model and observations (see Methods).

Furthermore, the simulation reveals that the degree to which OHC anomalies are set by OHT variability can change substantially between different decades. Specifically, ∆OHT is the lead term in only 2 years for model years (MY) 10–19 but in 7 years for MY 70–79 during which period OHT variability is particularly strong (Fig. 3a). For the century-long span of the run, the level of interannual variability in OHC as measured by the standard deviation of all individual March–March ∆OHC values is \({\sigma }_{\triangle {{{{{\rm{OHC}}}}}}}\)= 0.93 ZJ. However, for a given decade, \({\sigma }_{\triangle {{{{{\rm{OHC}}}}}}}\) varies from 0.69 to 1.38 ZJ, i.e., the level of interannual variability in OHC can vary by a factor of two depending on the decade under consideration (Fig. 3b).

The variability in model OHC is driven more strongly by changes in OHT (\({\sigma }_{\triangle {{{{{\rm{OHT}}}}}}}\)= 0.79 ZJ) than the surface flux (\({\sigma }_{\int {Q}_{n}^{{\prime} }}\) = 0.58 ZJ) reflecting the active role of the ocean in the model run. Standard deviations obtained using the same method from the shorter observational record are similar to those found with the model (Fig. 3b, triangles) as is the amplitude of variability in the observed time series (Fig. 3c, note any correspondence between the model and observed time series is coincidental as the model is a control run). Thus, both a heat conserving state-of-the-art coupled climate model and observation-based estimates are in agreement regarding drivers of SPNA heat content variability. Each reveals a strong role for the active ocean in regularly driving warm and cold OHC events.

Increasing role for the active ocean

We now explore whether SST responds differently to OHT and surface flux dominant conditions through a composite analysis of observed ∆SST for 1990–2019. The composite method selects on years with |∫\({Q}_{n}^{{\prime} }\)|>|∆OHT| for Fig. 4a (12 data points) and |∆OHT| > |∫\({Q}_{n}^{{\prime} }\)| for Fig. 4b (11 data points). In each case a threshold of 0.5 ZJ has to be exceeded for a given year to be selected (the results are not sensitive to this choice of threshold, see Supplementary Note 5). Both negative and positive events are included but with a scaling factor of −1 applied to the positive events as detailed in the Methodology. The effects of ∫\({Q}_{n}^{{\prime} }\) and ∆OHT dominant conditions on surface ocean temperature can clearly be seen (Fig. 4, note the figure shows the surface temperature change for increased heat loss in Fig. 4a and reduced OHT in Fig. 4b). The composite ∆SST field for the subset of years with ∫\({Q}_{n}^{{\prime} }\)dominant shows a broad region of anomalously cold surface temperature change over much of the basin north of 40 °N consistent with the pattern expected from cold air temperature anomalies associated with north-westerly airflow such as occurred in winter 2013–20146. In contrast, the composite ∆SST field for ∆OHT dominant conditions has a north-south split with cold (warm) ∆SST north (south) of 50 °N. This split is consistent with a reduction in northward heat transport and a consequent deficit (accumulation) of heat at higher (lower) latitudes.

Fig. 4: Observed ∆SST composite maps and pattern time series.
figure 4

Composite mean of ∆SST for a\({Q}_{n}^{{\prime}}\)dominant conditions, increased surface heat loss; b ∆OHT dominant conditions, reduced ocean heat transport. Stippling indicates 95% significance level. Boxes in a, b show the index regions for IQn (40–60 °N, 40–15 °W) and IOHT (50–65 °N, 40–20 °W and 32–50 °N, 40–20 °W). c Time series of IOHT (red) and IQn (blue), d Time series of 11-point centred moving average values for |IOHT|(red) and |IQn|(blue). Grey shading indicates prolonged periods of ∆OHT dominance. Blue and red shadings about solid lines in c, d indicate uncertainty due to the choice of boxes used to form the different indices (see Methods).

These results provide characteristic SST signatures of strong ∫\({Q}_{n}^{{\prime} }\)and ∆OHT within the modern era. In addition, they potentially enable historical SST data to be used to reconstruct the contributions of OHT and surface heat flux to interannual mixed layer heat content variability over the past 150 years. Such an approach is of particular value as SST observations are available back to 1870 via HadISST but reliable \({Q}_{n}\) and OHT observations are not.

To carry out the reconstruction, the contrasting ∆SST patterns have been used to define index values, IQn and IOHT, for ∫\({Q}_{n}^{{\prime} }\)and ∆OHT dominant conditions. IQn is the regional mean of ∆SST for the box spanning the cold eastern SPNA pattern in Fig. 4a normalised by its standard deviation for 1870–2019. IOHT is the corresponding north-south difference of normalised ∆SST for two smaller boxes centred on the poles of the ∆SST pattern associated with ∆OHT dominance (Fig. 4b). Time series of IQn and IOHT show strong interannual variability with some indication that the amplitude of the ∆OHT contribution increases and that of the ∫\({Q}_{n}^{{\prime} }\)contribution decreases in the second half of the record (Fig. 4c). We have quantified this using the mean magnitude of the 10 most extreme values for IQn and IOHT before and after 1960. In each case, the difference in values between the two periods exceeds the error: IQn (pre-1960) = 2.19 ± 0.15, IQn (post-1960) = 1.65 ± 0.10; IOHT (pre-1960) = 1.77 ± 0.14, IOHT (post-1960) = 2.12 ± 0.16.

To explore this further, we determine moving means of the absolute index values and these reveal notable variability in the relative strength of IQn and IOHT (Fig. 4d). From 1870 to about 1960, they tend to be similar in magnitude and strongly correlated (r = 0.72, significant at 95% level). However, after 1960 they become decorrelated (r = −0.15, not significant) and there are several periods, highlighted in grey, during which IOHT is noticeably larger than IQn. In these periods, the active ocean plays the primary role in setting interannual variability of the SPNA mixed layer heat content and the net heat flux contribution is reduced. Thus, the contribution of OHT variability relative to surface heat exchange has become more dominant in the past 60 years. Why is this the case? One possibility is that weakening of the AMOC31,32 has led to increased instability of the ocean circulation and the amount of heat it transports polewards. In this scenario, rather than staying predominantly in a relatively fixed strong state, the ocean has recently varied more frequently between strong and slightly weaker states at interannual timescales enabling greater ocean activity in setting SPNA OHC. We stress that this is a possible explanation and that it is important to remember that the change in contribution from the ocean may also include other processes, e.g., vertical mixing. Thus, it is not possible at this point to conclusively associate the change over the past 60 years with a weakening AMOC. In addition, a reduction in the overall severity of cold air outbreaks33 may have weakened the net heat loss in such events, thus making the mixed layer heat content less sensitive to the surface heat flux. Further research is needed to test these hypotheses, potentially using coupled model ensemble analysis.


Through a combined analysis of SPNA observations and a high-resolution climate model simulation, we have revealed an active role for ocean transports in generating both warm and cold interannual anomalies in mixed layer heat content. Our findings firmly establish that the ocean is frequently active in setting its own mixed layer heat content variability rather than passively responding to the atmosphere. The focus of this study has primarily been interannual variability but our results are also relevant to the ongoing debate about the role of the ocean at interdecadal timescales19,34,35,36,37,38,39. We have found evidence for variability in the extent of interannual ocean activity from decade to decade. This has the potential to confound attempts to identify robust decadal modes of variability and needs to be considered in such studies. Furthermore, from a reconstruction using SST observations since 1870, we infer that the level of ocean activity relative to surface heat flux has increased markedly over the past 60 years. This points to a possible role for a weakening AMOC (through increased instability of the ocean circulation and the amount of heat it transports polewards) that merits further investigation, particularly as a recent multi-proxy reconstruction indicates a rapid AMOC decline from about 196031.

Our results open the door to a clearer understanding of the impacts of variability in the North Atlantic ocean circulation on the coupled ocean-atmosphere system including European weather. Ocean surface and upper layer temperature anomalies have been shown to influence both winter cold events40,41,42 and summer heatwaves2,5 over Europe in model analyses. By integrating our new insights into such studies, it will be possible to examine whether an increasingly active ocean is now playing a greater role in determining European weather events by exerting a tighter control over the mixed layer heat content reservoir available to the atmosphere.


Heat transport calculation

For a given region, specifically the NE and SE boxes in our study, the change in the amount of heat transported laterally and vertically by the ocean (∆OHT) into the mixed layer is given by the residual:

$$\triangle {{{{{\rm{OHT}}}}}}=\iint\int {\rho} {{{{{\rm{c}}}}}}_{{{{{{\rm{p}}}}}}}\,({{{{{{\rm{T}}}}}}}_{{{{{{\rm{M}}}}}}2}-{{{{{{\rm{T}}}}}}}_{{{{{{\rm{M}}}}}}1}){{{{{\rm{dxdydz}}}}}}-\iint\int {Q}_{n}^{{\prime} }\,{{{{{\rm{dxdydt}}}}}}$$

with \({{{{{\rm{\rho }}}}}}\), the density of seawater (taken to be 1025 kg m−3), \({{{{{{\rm{c}}}}}}}_{{{{{{\rm{p}}}}}}}\), the specific heat capacity of seawater (3850 J kg−1 °C−1), \({{{{{{\rm{T}}}}}}}_{{{{{{\rm{M}}}}}}2}\), \({{{{{{\rm{T}}}}}}}_{{{{{{\rm{M}}}}}}1}\), the potential temperature in March of year 2 and year 1, respectively, \({Q}_{n}^{{\prime} }\) the net surface heat flux anomaly integrated over the intervening time period. The first term on the right hand side is integrated over the mixed layer depth H(x,y,t) where H is defined to be the first depth layer at which the potential temperature is more than 0.5 °C lower than the surface temperature. We stress that, unlike many previous studies, H is calculated for each individual month rather than following a climatological monthly definition and thus takes interannual variability into account.

The approach detailed above is taken for both the observation and model ∆OHT calculations. For the observations, the \({Q}_{n}^{{\prime} }\) integral, second term, is calculated by summing values of the monthly net air-sea heat flux anomaly, relative to the 1981–2010 seasonal cycle, from March of year 1 to February of year 2 inclusive. For the model, the \({Q}_{n}^{{\prime} }\)integral is calculated in the same way relative to a seasonal cycle determined from years 31–60 of the 100-year run. Note, earlier studies have considered the combined effects of the net heat flux and wind-driven Ekman transport in a single term16,23. However, we treat the net heat flux in isolation and the Ekman contribution is implicitly included within the ∆OHT term. We have explored whether the observed ∆OHT variations can be related to the local wind stress by calculating the convergence of the observed Ekman volume transport into both the SE and NE boxes and correlating with the ∆OHT values for each box from 1991–2019. For the NE box, there is a weak positive correlation (r = 0.37, significant at 95%), while the SE box value is close to zero (r = −0.14, insignificant at 95%). Although weak these results suggest a possible regional variation with the Ekman contribution becoming important towards the NE of the SPNA. We plan to explore this in subsequent research targeted at the relationship between the wind-driven transport, the air-sea heat flux and ∆OHT.

Definition of ∆OHT or ∫\({{{{{{\boldsymbol{Q}}}}}}}_{{{{{{\boldsymbol{n}}}}}}}^{{\prime} }\) dominance

In our analysis, we define whether a given year is ∆OHT dominant according to the requirement that |∆OHT| > |∫\({Q}_{n}^{{\prime} }\)|. We have also investigated whether using a more stringent requirement that one term needs to be substantially larger than the other impacts our key conclusion that ocean transports are dominant for nearly a half of the cases considered. Specifically, if we require that ∆OHT is larger in magnitude than ∫\({Q}_{n}^{{\prime} }\) by 0.5 ZJ as the condition for dominance then we find that 6 out of 11 years have ∆OHT dominant for the SE region (5 out of 11 for NE). This compares with 7 out of 18 years for the SE region (8 out of 19 for NE) with the simple inequality approach used previously. Thus, including the requirement that one term needs to be substantially larger than the other slightly increases the proportion of ocean transport dominant events such that ∆OHT is dominant for half of the cases considered.

Climate model specification

The 100-year climate simulation has been produced using the HadGEM3-GC3.1 model (Met Office Hadley Centre Global Coupled model General Circulation 3.1) which has an ocean resolution of 1/12° and an N512 (approximately 25 km) atmosphere43. HadGEM3-GC3.1 incorporates the GA/GL7.1 global atmosphere–land configuration that uses a regular latitude–longitude grid and has 85 levels extending to 85 km44. The global ocean component is GO645, which uses the Nucleus for European Modelling of the Ocean model at vn3.6, having a tripolar grid, with 75 ocean levels (and top level thickness of 1 m). The sea ice model configuration is GSI8.146, which uses the CICE5.1 model. The model is initialised using a combination of atmospheric reanalysis and ocean/ice output from a lower resolution model simulation consistent with the CMIP6 HighResMIP protocol47. Full details of the model configuration and spin up are provided by48. By using a coupled model we avoid having to prescribe the surface heat exchange that is necessary in forced ocean models; such models have shown discrepancies in their representation of the eastern SPNA heat budget49.

Error estimation and composites

Error estimates for ∆OHC have been obtained following the Monte Carlo methodology. For each month, the OHC relative to the 1981–2010 mean is calculated using the EN4 temperature and temperature uncertainty fields. This process is repeated 1000 times with randomly sampled uncertainty values and used to derive the ∆OHC error estimates. The ∫\({Q}_{n}^{{\prime} }\) error estimates are determined using the same 1000 repeat Monte Carlo approach with monthly random errors specified for each of the four component heat flux fields that sum to form the net heat flux. These are the latent (specified error 30 Wm−2), sensible (20 Wm−2), longwave (15 Wm−2) and shortwave (20 Wm−2) fluxes. The error values in brackets are conservative one standard deviation estimates based on evaluations of NCEP/NCAR and other reanalyses using surface flux buoy observations50,51. The sensible and latent heat flux errors are potentially strongly interdependent as they are driven by errors in the wind speed and near-surface temperature and humidity gradients; to allow for this possibility the sensible heat flux error is taken to be 2/3 of the latent heat flux error on each repeat. The uncertainties in ∆OHC and ∫\({Q}_{n}^{{\prime} }\) are summed in quadrature to form the error on ∆OHT. All error estimates are specified at the one standard deviation level. Note, as a further consistency check, the surface flux dependent results reported in the paper have been re-determined using an alternative reanalysis, ERA552. Similar results are obtained with either reanalysis. For example, for March 1995 to February 1996 the box integrated net heat flux values with the two datasets both show net warming in the NE box: ∫\({Q}_{n}^{{\prime} }\) = 0.58 (0.59) ZJ for ERA5 (NCEP) and both are close to zero in the SE box: ∫\({Q}_{n}^{{\prime} }\) = 0.11 (−0.08) ZJ for ERA5 (NCEP).

The ∆SST pattern in Fig. 4a represents surface heat flux dominance over OHT and increased surface heat loss conditions. It shows the composite mean of years with |∫\({Q}_{n}^{{\prime} }\)| > |∆OHT| and |∫\({Q}_{n}^{{\prime} }\)| > 0.5 ZJ (12 members within period 1991–2019). Both negative (∫\({Q}_{n}^{{\prime} }\) < −0.5 ZJ and ∫\({Q}_{n}^{{\prime} }\) < ∆OHT) and positive (∫\({Q}_{n}^{{\prime} }\) > 0.5 ZJ and ∫\({Q}_{n}^{{\prime} }\) > ∆OHT) events are used to form the composite in order to increase the sample size. To prevent cancellation of the ∆SST patterns associated with negative and positive events (which are similar but of opposite sign) a scaling factor of −1 is applied to the patterns from the positive events. The ∆SST pattern in Fig. 4b represents OHT dominance over surface flux and reduced heat transport conditions. It shows the composite mean for years with |∆OHT| > |∫\({Q}_{n}^{{\prime}}\)| and |∆OHT| > 0.5 ZJ (11 members). As above, both negative (∆OHT < −0.5 ZJ and ∆OHT < ∫\({Q}_{n}^{{\prime} }\)) and positive (∆OHT > 0.5 ZJ and ∆OHT > ∫\({Q}_{n}^{{\prime} }\)) events are selected, and ∆SST for those members with ∆OHT > 0.5 ZJ is scaled by a factor −1 for sign consistency with reduced heat transport. Confidence limits for stippling of the composite fields have been determined using the Normal-Z statistic53. Values for ∫\({Q}_{n}^{{\prime} }\)and ∆OHT used to select the years that form the composites are calculated for the eastern SPNA (45–65 °N, 40–15 °W). The effective sampling size for testing of correlation significance has been determined according to54.

Note the boxes used to determine IOHT and IQn in Fig. 4 are process-based (i.e., specified according to the pattern resulting from OHT or surface heat dominance) and were defined after the analysis used to produce Fig. 4a, b patterns. They are different to the NE and SE boxes used for Fig. 2 that are regional sub-divisions chosen for a sensitivity analysis. We have explored also whether the IOHT and IQn time series shown in Fig. 4c, d are sensitive to the box definitions employed and find this is not the case. For this sensitivity analysis, each of the box boundaries is varied by 1° steps within a 10° range and the time series are recalculated in each case. The standard deviation of the resulting ensemble of time series is determined and shown by shading about the original results (solid lines) in Fig. 4c, d. It can be seen from the figure that the uncertainty due to possible alternative box definitions is small and our main result that there is a change in the relative behaviour of the time series after 1960 is robust to the choice of boxes used to define the IOHT and IQn indices.