AMOC fingerprints influence seasonal SST predictability in the North Atlantic

We investigate the impact of the strength of the Atlantic Meridional Overturning Circulation (AMOC) at 26◦N on the prediction of North Atlantic sea surface temperature anomalies (SSTA) a season ahead. We consider the physical mechanism proposed by Duchez et al. (2016a) and test the dependence of SST predictive skill in initialised hindcasts on the phase of AMOC at 26◦N. We use initialised simulations with the MPI-ESM-MR seasonal prediction system. First, we use the assimilation experiment between 1979-2014 to confirm that the AMOC leads a SSTA dipole pattern in the tropical and 5 subtropical North Atlantic, with strongest AMOC fingerprints after 2-4 months. Going beyond previous studies, we find that the AMOC fingerprint has a seasonal dependence, and is sensitive to the length of the observational window used, i.e. stronger over the last decade than for the entire time series back to 1979. We then use a set of ensemble hindcast simulations with 30 members, starting each February, May, August and November between 1982 and 2014. We compare the changes in skill between composites based on the AMOC phase a month prior to each start date to simulations without considering the AMOC 10 phase. We find higher SST hindcast skill at 2-4 months lead time for SSTA composites based on the AMOC phase for February, May and November start dates. Our method shows major benefit for May start dates, where mean summer SST hindcast skill over the subtropics increase by a factor of 2, reaching up to 80% agreement with ERA-Interim SST.

. Transports mean values, standard deviations and seasonal ranges (in parentheses) for the model  and observed AMOC (2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014). All values in Sv. field averaged at each latitude, and the AMOC is defined as the vertical maximum of the stream function. We verify the modelled AMOC using observations from the RAPID array at 26 • N. The RAPID AMOC is defined as the sum of three components: the Florida Strait transport, the surface Ekman transport (EKM), and the geostrophic upper-mid-ocean transport.
A detailed description of the calculation of the individual components is provided in Smeed et al. (2018).
We evaluate the atmospheric contribution to the SST variability using EKM and ASF. We evaluate both the EKM relationship 90 to SST, as well as the AMOC without its EKM component, i.e. AMOC-EKM (Mielke et al., 2013). EKM is calculated from the zonal wind stress component τ x integrated over the Atlantic, i.e. EKM = − τx ρf dx, where ρ is the reference density (1025 kg m −3 ) in MPIOM and f is the Coriolis parameter. For ASF we use the total surface heat fluxes over sea, which include shortwave, longwave, latent and sensible heat fluxes. ASF is parameterized as described in Marsland et al. (2003), with fluxes defined positive downward. 95 We calculate monthly means of AMOC, EKM, SSTA and air-sea heat fluxes. Our main data set consists of model output, in addition to AMOC observations from RAPID and SST from the ERA-Interim reanalysis (Dee et al., 2011). This data set is deseasoned by removing the 12-month climatology obtained from the monthly data and the linear trend is removed. We refer to these detrended, deseasoned quantities as anomalies. Time series are smoothed using a 3-month running average to filter out high frequency variability. Seasonal means are defined as December-January-February (DJF) for winter, March-April-May 100 (MAM) for spring, June-July-August (JJA) for summer and September-October-November (SON) for autumn.
To assess the variability of the AMOC fingerprint and to evaluate its role on seasonal SST predictability, we perform lagged correlations from 0 up to 12 months, with the AMOC leading SSTA. Additionally, we compute lagged correlations for ASF, EKM and AMOC-EKM leading SSTA to explore the atmospheric contribution. For our hindcast skill analysis, we assess predictive skill of the hindcast simulations against the ERA-Interim data with the point-wise Anomaly Correlation Coefficient 105 (ACC, Collins (2002)). Significance is assessed via a bootstraping method at the 95% confidence level using 1000 samples.

Model verification for AMOC
We verify the modelled AMOC and the components EKM and AMOC-EKM to observations as a first step to our analysis.
We evaluate the AMOC seasonal cycle using both anomalies and full values. The anomalies are calculated by removing the annual mean of each year (grey lines in Fig.1.a-f) of the full time series , after which the data were smoothed with a 3-month running average. The observed AMOC shows minimum transport in March and maximum in August (c.f. Fig.1.a). Minimum transport for the modelled AMOC is achieved slightly later, in April-May, while it peaks twice in August and December. The seasonal cycle for both the observed and the modelled AMOC agree with the ones discussed by Mielke  (Cunningham et al., 2007) and a high resolution MPI ocean model spanning the same period. For EKM (c.f. Fig.1.c), the seasonal cycle for observations and model are slightly out of phase, but both show a clear maximum in summer (July-August) and minimum in spring (March-April). The seasonal range for the modelled EKM is 3.36 Sv, compared to lower 2.26 Sv for the observations. The opposite is found for the AMOC seasonal range, which is smaller for model (2.79 Sv against 3.90 Sv). These differences in range and phase for AMOC and EKM can explain the The correlation pattern for the subpolar region is also pronounced, however the strongest negative correlations of -0.4 are only present up to 2 months lag (c.f. Fig.2.a, b). These negative correlations have been previously associated with the NAO imprint in the Atlantic (Fan and Schneider, 2012), and are not explained by D16's physical mechanism which we investigate in this study. D16's physical mechanism attributes an active role of ocean heat advection on the SST variability at the timescale of 140 a few months, due to anomalous convergence or divergence of OHT. Therefore, we restrict our analysis to the AMOC influence on SST over tropical and subtropical North Atlantic, and exclude the subpolar pattern.
To further investigate the variability of the SST dipole pattern, we here analyse the role of SST seasonality. Using the assimilation experiment, we perform correlations of the AMOC anomalies at a given month with the mean seasonal SSTA 2-4 months ahead (Fig.5). By doing so, we provide a detailed view of the temporal variability of the SST dipole pattern, making it easier to link the observed pattern to other drivers that could potentially affect the SST variability, such as ASFs or EKM.
We find a strong fingerprint in spring (MAM), with average (maximum) correlation of the order of 0.4 (0.52) (c.f. Fig.5).

170
During summer (JJA), the fingerprint is less pronounced than in spring, with lower average (maximum) correlation magnitudes of around 0.3 (0.44). In contrast, we find that autumn and winter seasons lack a characteristic dipole pattern, showing instead only a narrow region of negative correlations over the subtropics of the order of -0.2 (-0.1) for winter (autumn). In particular, for autumn SSTs (Fig. 5.c) we find significant positive correlations of the order of 0.6 over the subpolar region, suggesting that atmospheric drivers potentially supersede the AMOC fingerprints during this season.

The atmospheric contribution
At the seasonal timescale, much of the SST variability in the North Atlantic is response to atmospheric forcing (Deser et al., 2010). The two main processes responsible for the atmospheric imprint in the large-scale SST variability are anomalous ASFs  D16's physical mechanism suggests that via convergence (divergence) of OHT in the subtropics (tropics), a strong AMOC at 26 • N drives the sub-and tropical SST variability at a maximum of 2-5 months lead time. Therefore, as a consequence of the ocean's thermal memory, we expect that hindcasts initialised after strong AMOC phases may result in more (less) predictable SSTs in the subtropics (tropics), given weaker contribution from the atmosphere to the SST variability over the areas of strong 240 AMOC fingerprints.
For strong AMOC phases, we find higher hindcast skill for DJF, JJA and MAM SSTAs over the subtropics in comparison to ACCs considering the full period (cf. Fig.8a-b, g-h, d-e, respectively), in agreement with the physical mechanism. ACCs above 0.6 are found in the Gulf Stream region and along the North Atlantic Current for JJA and DJF SSTAs (cf. Fig.8h, b, respectively). In particular, we find higher skill for JJA SSTAs in a zonal band between 30 • -40 • N extending up to 40 • W, 245 where mean ACCs increase by a factor of 2 and reach maximum above 0.8 in comparison to ACCs considering the full period (cf. Fig.8g, h). ACCs for MAM SSTA increase over the subtropics particularly in the Sargasso Sea, where ACCs above 0.8 cover most of its western side. While the mechanism does not solely explain the hindcast skill behaviour in the tropics, we find an improvement over the hurricane main development region (MDR, e.g. Hallam et al. (2019)) with ACCs above 0.8 over large parts (cf. Fig.8e).

250
We now examine ACC changes in the SST composite based on weak AMOC phases. These changes in hindcast skill, albeit less pronounced than for strong AMOC phases, agree to some extent with D16's physical mechanism for DJF and JJA SSTAs.
For example, compared to the full time series (Fig.8a, g), we find better hindcast skill for DJF and JJA SSTA composites in the tropics, with ACCs above 0.8 mainly occurring over the central MDR (cf. Fig.8c, i, respectively). As opposed to these findings, the physical mechanism fails to explain the ACC changes with respect to SON SSTA composites (cf. Fig.6k, l). This seems 255 to confirm that significant atmospheric contribution to the SST variability (e.g. over the subtropics, cf. Fig.6e, f)  scales (e.g. Zhang (2008); Muir and Fedorov (2015); Borchert et al. (2018)), only recently the extent to which AMOC influences SST at seasonal time scales has been addressed (Duchez et al., 2016a). In this study we explore the influence of AMOC strength at 26 • N on North Atlantic SST seasonal variability and predictability in the MPI-ESM-MR model. We specifically test whether our model AMOC fingerprints agree with the physical mechanism proposed in D16, and could therefore be considered in our predictions to condition seasonal SST hindcast skill over North Atlantic tropics and subtropics. Our findings suggest that the 265 strength of AMOC is a potential regional source of SST predictability by controlling the variability of heat advection north or south of 26 • N.
In line with D16, we find pronounced AMOC fingerprints at 2-5 months lag when considering the RAPID period (2004 -2014). Going beyond this study, however, we find that AMOC fingerprints are sensitive to the length of the observational window used. Although our findings are in good agreement with D16 when we restrict the analysis to the most recent decade 270 (cf. Fig.2), we find less pronounced AMOC fingerprints with respect to the full time series back to 1979, at a maximum of 2-4 months lag. A possible reason for these differences could be the decadal AMOC variability, i.e. AMV (e.g. Ba et al. (2014); Knight et al. (2005)). The RAPID period corresponds to a positive AMV phase (Zhang, 2007), i.e. warmer SSTs over the North Atlantic due to changes in the AMOC dynamics at the multidecadal time scale, resulting in stronger OHT that could potentially enhance the AMOC influence at the seasonal timescale. HadGEM3-GC2. They tested the robustness of the AMOC fingerprints on the SST through time, finding a good agreement with D16 at the 5-month lag, when taking the mean of 11-year segments of the full time series. However, when considering the full 120-year time series, this agreement was overall lower than when analysing the 11-year segments. Likewise, we find weaker AMOC fingerprints when analysing 30-year segments selected from the MPI-ESM-MR historical simulation (not shown).

280
In addition to the sensitivity of the observational window length, a key aspect that distinguishes our analysis from previous studies is that we find a significant seasonal dependence on the AMOC fingerprints. This dependence is coherent in both initialised and free-running model (not shown), with the strongest AMOC fingerprints occurring during spring and early summer.
In line with Alexander-Turner et al. (2018), we argue that a main driver for this seasonal dependence is the contribution of stochastic atmospheric variability, and in a lesser extent the Ekman transport. This has a direct implication on the consideration 285 of the physical mechanism in our seasonal prediction system, which is thus dependent on the initialisation month.
The impact of this seasonal dependence can be illustrated as the distinguished effects of the physical mechanism on the hindcast skill for each start month (cf. Fig.8). Over parts of the subtropics, we achieve high SST hindcast skill for NOV (DJF SSTAs), FEB (MAM), and particularly for MAY (JJA SSTAs), for strong AMOC phases. Such windows of opportunity for skilful summer SST predictions (Mariotti et al., 2020) in turn may benefit winter NAO predictions, with consequent influences on the storm track activity starting from October (e.g. Cassou et al. (2004)), as well as on the development of Blocking regimes (e.g. Guemas et al. (2010)), and extreme events (e.g. Arora and Dash (2016)). Moreover, after weak AMOC phases, we find better hindcast skill for DJF, MAM and JJA SSTAs over the tropics, in particular over the central MDR. These improved SST predictions over the MDR could be extremely beneficial for assessing seasonal hurricane formation probabilities (Saunders and Lea, 2008). 295 We highlight that for FEB hindcasts, skill of MAM SSTAs increases over most of the North Atlantic for strong AMOC SST composites. Therefore, besides D16's mechanism only explaining the improvement over the subtropics, our results suggest that a more active ocean anticipating the initialisation in February could potentially overwrite the higher frequency variability of the SST dominated by the atmosphere (e.g. Yeager et al. (2012), Robson et al. (2012) and Borchert et al. (2018)). This interpretation cannot, however, be extended to AUG hindcasts, which showed no significant improvement in the subtropical SON SST skill for 300 strong AMOC SST composites (Fig.8j). This supports the evidence of a strong influence of stochastic atmospheric variability for this region at 2-4 months lag (cf. Fig.6.e-f) and calls for other physical mechanisms, that if incorporated in the prediction could result in a more prominent effect on the hindcast skill. Recently, a similar approach invoked a chain of physical processes in the prediction and achieved improved skill for European summer climate predictions (Neddermann et al., 2018).
As a first step to investigate whether the predictive skill is conditional to the phase of individual AMOC components, 305 we assess the role of the non-Ekman ocean heat transport in driving the seasonal variability of SST. We thus create SST composites based on the AMOC-EKM strength at 26 • N for MAY and NOV hindcasts (not shown). Over the subtropics, the Ekman transport is largely responsible for anomalies in the SST, due to its coupling with both strong winds and pronounced temperature gradients (Frankignoul, 1985). Roberts et al. (2017) carried out a detailed observation-based heat budget analysis for short-term variations of the ocean heat content, supporting the idea of potential predictability of SSTs in the mid-latitudes 310 from an active ocean dynamics. Moreover, Ossó et al. (2018) use the ERA-Interim reanalysis and suggest the Ekman advection associated with weaker westerly winds in late winter and early spring to be a major driver of spring SSTA east of Newfoundland.
We find a relatively weak improvement on the SST skill over the subtropics for AMOC-EKM SSTA composites, suggesting an important role for the Ekman transport on the predictability of summer and winter SSTs.
Although we achieve significant improvement in SST hindcast skill for important regions of the North Atlantic with our 315 method, one major limitation of our analysis is that we only consider a limited set of predictors over the North Atlantic (e.g. AMOC strength at 26 • N, air-sea heat fluxes and Ekman transport), and further test a single physical mechanism to explain SST variability at the seasonal time scale. Several studies have shown that one of the most robust remote impacts of ENSO is the teleconnection to tropical North Atlantic SSTs in boreal spring (e.g. García-Serrano et al. (2017)). The incorporation of another physical link into the prediction, such as ENSO, could show additional refined information on the North Atlantic SST 320 prediction skill.
Our analyses support further investigation of the AMOC strength and its associated heat transport as complementary information for the seasonal prediction of SSTAs. Both high-resolution coupled models and the two ongoing AMOC monitoring programs RAPID-MOCHA (Cunningham et al., 2007) and OSNAP (Lozier et al., 2017(Lozier et al., , 2019 are essential for a thorough understanding of the mechanisms analysed here. In fact, the seasonal relationship between AMOC and SSTA could contribute 325 to the knowledge of the potential applications of a real-time data delivery system, when finally implemented in the RAPID array (Rayner et al., 2016).

Conclusions
We assess the impact of AMOC fingerprints on North Atlantic seasonal SST variability and predictability. We consider the physical mechanism proposed by D16 in the hindcast skill analysis of a 30-member ensemble hindcast initialised every Febru-330 ary, May, August and November, and evaluate the effect of this mechanism by exploring the changes in SST hindcast skill for tropical and subtropical North Atlantic, when compared to the hindcast analysis without considering this mechanism. Our findings using MPI-ESM-MR suggest that: 2. In extension to D16, this AMOC fingerprint has a seasonal dependence, and is sensitive to the length of the time window used. This sensitivity affects both the intensity and structure of the fingerprints, which are stronger in spring and summer, and over the last decade, than for the entire time series back to 1979. 3. The AMOC fingerprints seasonality can be attributed to i) the influence of stochastic atmospheric variability on SST via cumulative ASFs, which is most pronounced for autumn SST variability over the subtropics and seems to weaken the effects of AMOC fingerprints; ii) Ekman transport, which explains part of the variability over the subtropical lobe of the AMOC fingerprint; except during spring, when the AMOC fingerprint is instead mainly determined by the Ekman component.