This study is based on a global configuration of the ocean model system NEMO (Nucleus For European Modelling of the Ocean) version 3.2 (Madec, 2008). The quasi-isotropic tripolar grid ORCA (Madec and Imbard, 1996) has a resolution of 0.5∘ in longitude and 0.5∘ × cos(φ) in latitude (ORCA05) and 46 vertical levels whereof 10 levels lie in the upper 100 m. It is coupled online to the Louvain-la-Neuve sea ice model version 2 (LIM2) and the biogeochemical model PISCES-v1 (Pelagic Interaction Scheme for Carbon and Ecosystem Studies; Aumont and Bopp, 2006). Parameter values and numerical options for the physical model follow Barnier et al. (2006) and Timmermann et al. (2005). Two atmospheric reanalysis products, DFS4.2 and DFS4.4, were used for this study. DFS4.2 is based on ERA-40 (Brodeau et al., 2010) and covers the period 1958–2007, while DFS4.4 is based on ERA-Interim (Dee et al., 2011) and covers the years 2002–2012. The simulation was spun up over a full DFS4.2 forcing cycle (50 years) starting from rest and holding atmospheric CO2 constant to levels of the year 1870 (287 ppm). Temperature and salinity were initialized as in Barnier et al. (2006). Biogeochemical tracers were either initialized from climatologies (nitrate, phosphate, oxygen, dissolved silica from the 2001 World Ocean Atlas, Conkright et al., 2002, and preindustrial dissolved inorganic carbon (CT) and total alkalinity (AT) from GLODAP, Key et al., 2004) or from a 3000-year-long global NEMO–PISCES simulation at 2∘ horizontal resolution (iron and dissolved organic carbon). The remaining biogeochemical tracers were initialized with constant values.

At the end of the spin-up cycle, two 143-year long simulations were started in 1870 and run in parallel. The first one, the historical simulation, was forced with spatially uniform and temporally increasing atmospheric CO2 concentrations (Le Quéré et al., 2014). In the second simulation, the natural simulation, the mole fraction of atmospheric CO2 was kept constant in time at 287 ppm. Both runs were forced by repeating 1.75 cycles of DFS4.2, interannually varying forcing over 1870 to 1957. Next DFS4.2 was used from 1958 to 2007. Simulations were extended up to 2012 by switching to DFS4.4 in 2002. No significant differences were found in tracer distributions and Cant related quantities between both atmospheric forcing products during the years of overlap (2002–2007). Carbonate chemistry and air–sea CO2 fluxes were computed by PISCES following the Ocean Carbon Cycle Model Intercomparison Project protocols (http://ocmip5.ipsl.jussieu.fr/OCMIP/, last access: 25 June 2018) and the gas transfer velocity relation provided by Wanninkhof (1992). Climate change trends and natural modes of variability are part of the forcing set used to force both simulations. Hence, any alteration of the natural carbon cycle in response to climate change (e.g., rising sea surface temperature) will be part of the natural simulation. The concentration of Cant, as well as anthropogenic CO2 fluxes, is calculated as the difference between the historical (total C = natural + anthropogenic contribution) and natural simulations following Orr et al. (2017).

The model simulates a global ocean inventory of Cant in 2010 of 126 PgC. It is at the lower end of the uncertainty range of the estimate by Khatiwala et al. (2013) of 155 ± 31 PgC (Fig. 1). At the global scale, the error of the model is close to 6 % (values excluding arctic region and marginal seas). The underestimation of the simulated Cant inventory compared to Khatiwala et al. (2013) is largely explained by the difference in the starting year of integration (Bronselaer et al., 2017): 1870 for this study as opposed to 1765 in Khatiwala et al. (2013). The coupled model configuration is referred to as ORCA05–PISCES hereafter. The reader is referred to Bourgeois et al. (2016) for a detailed description of the model and the simulation strategy.

Observations used to evaluate the transport of Cant in ORCA05–PISCES were collected along the Greenland–Portugal OVIDE section and at 24.5∘ N following the tracks presented in Fig. 2. Simulated air–sea fluxes of CO2 were compared to the observation-based gridded sea surface product of air–sea CO2 fluxes from Landschützer et al. (2015a). Programs and/or data sets are briefly summarized below.

Heat transport across a section was computed from horizontal velocity orthogonal to the section times the heat term estimated from temperature and salinity using the international thermodynamic equations of seawater (TEOS 2010). Heat transport is used in Sect. 4.1 to evaluate model performance to reproduce the well-known mechanism controlling its interannual variability correctly and to compare to results for Cant transport.

In the model, over one-third of Cant entering in the southern box at 25∘ N (0.092 ± 0.016 PgC yr-1) is transported across the OVIDE section before leaving the domain through the Greenland–Iceland–Scotland sills (Fig. 3). The comparison between online and offline estimates of Cant transport across the OVIDE section confirms the dominant contribution of advection (Fig. S3) in line with Tréguier et al. (2006). Simulated transport of Cant (Fig. 3) is clearly underestimated: it is 3 times smaller than observations at 25∘ N (Zunino et al., 2015b) and at the OVIDE section (Pérez et al., 2013; Zunino et al., 2014, 2015a), whereas it is 1.5 to 2 times smaller than observations at the sills (Jeansson et al., 2011; Pérez et al., 2013). In order to identify the reasons for the underestimation of TCant in the model, simulated volume transport and concentration of Cant are compared to in situ estimates (Eq. 2) in the following paragraphs.

Simulated air–sea fluxes of total and anthropogenic CO2 (Fig. 3) are higher than those derived from in situ data by Pérez et al. (2013) their Fig. 3. For total CO2, the model–data difference is 0.013 PgC yr-1 (northern box) and 0.103 PgC yr-1 (southern box). For the anthropogenic component, it is 0.028 PgC yr-1 (northern box) and 0.036 PgC yr-1 (southern box). While the model overestimates CO2 uptake, the ratio of anthropogenic to natural flux is comparable to previous estimates (Gruber et al., 2009; Schuster et al., 2013), which implies a similar overestimation of both components. To understand the origin of the overestimation of fluxes, simulated air–sea fluxes of total CO2 were averaged over 2003–2011 and compared to observation-based estimates from Landschützer et al. (2015a), taken as representative of the SOCCOM exercise (Rödenbeck et al., 2015). The model overestimates CO2 uptake mainly between the OVIDE section and the Greenland–Iceland–Scotland sills (Fig. 8a, b). The month of occurrence of the seasonal maximum or minimum air–sea CO2 flux was diagnosed. It is presented in Fig. 8c, d. A seasonal phase shift between simulated fluxes and data-based estimates is observed north of 50∘ N, where the model strongly overestimates gas exchange. Fluxes peak in winter in observations, while they reach their maximum in summer in the model. The seasonal change in surface water pCO2 is dominated by biological activity north of 40∘ N and by temperature (or thermodynamics) between 20 and 40∘ N (Takahashi et al., 2002). The model reproduces the main driving process of seasonal variability of air–sea CO2 fluxes in the subtropical region. However, the dominant effect of temperature extends too far north in the model because the latter failed to reproduce the air–sea gradient of winter pCO2. As a result, the seasonal change in CO2 fluxes is dominated by the thermodynamical effect in the subpolar gyre, which enhances the ocean sink for atmospheric CO2. Despite the seasonal phase shift noted in the subpolar gyre, the amplitude of the interannual variability of total air–sea CO2 fluxes (defined as the standard deviation of air–sea fluxes computed over 1982–2011 after removing the seasonal cycle; Fig. 9) is well reproduced by the model over the total domain, including north of 40∘ N where the variability is the largest.

Figure 10 presents annual time series (1958–2012) of the MOC intensity and the transports of heat and Cant across 25∘ N, 36∘ N and OVIDE. The analysis of annual time series (Table 5a) reveals a strong correlation between the intensity of the MOC and the heat transport across all three sections. Conversely, the transport of Cant only correlates with the MOC intensity at 36∘ N. As expected, circulation is the main driver of the interannual to decadal variability of heat transferred across the three sections (Johns et al., 2011; Mercier et al., 2015). Its impact on the Cant transport variability is, however, masked by additional processes.

The transport of Cant across all sections increased continuously over the period of study (Fig. 10, Table 5c). Neither heat transport, nor MOC intensity, nor the net volume of water transported across the sections display a similar increase (Table 5c). Zunino et al. (2014) attributed essentially the increase in the northward transport of Cant since 1958 to its accumulation in the northward flow of the MOC upper limb. In order to isolate the circulation effect, we removed the positive trend from the time series of Cant transport. The correlation (r) between the detrended Cant transport and the intensity of the MOC increased from 0.30 to 0.74 at 25∘ N and from 0.67 to 0.70 at 36∘ N (Table 5a and b). It did, however, not change at the OVIDE section (Table 5a and b). Circulation emerges as the dominant control of interannual to decadal variability of Cant transport at 25 and 36∘ N, but not across the OVIDE section.

Figure 11 shows the budget of Cant from 1959 to 2011 for the three boxes. Each budget is composed of the storage rate of Cant, the air–sea flux of anthropogenic CO2 and the divergence of Cant transport. The storage rate of Cant increased continuously over the North Atlantic. The increase was largest in Box 2 (36∘ N–OVIDE), where the storage rate is multiplied by 3, followed by Box 1 (25–36∘ N), where it is doubled. Figure 11 also shows that the air–sea flux of anthropogenic CO2 and the divergence of Cant transport contributed equally to changes in Cant inventory in the southern box between 1959 and 2011. From 36∘ N to the OVIDE section, the contribution of air–sea flux dominated prior to 1985. From 1985 onward, the transport divergence gained in importance, albeit with a pronounced interannual variability. In the northern box, changes in Cant inventory followed air–sea fluxes, with a weak contribution of transport divergence limited to interannual timescales. The significant positive correlation (Table 6a, no trend removed) between storage rate and air–sea flux in all three boxes suggests that during the past 53 years the latter controlled the Cant storage rate on multi-decadal scales. The transport divergence of Cant increased continuously from 1985 onward in boxes 1 and 2 and is positively correlated with changes in Cant storage rate over 1959–2011 (Table 6a, no trend removed). The Cant transport divergence did, however, not contribute to the long-term change in Cant inventory between the OVIDE section and the Greenland–Iceland–Scotland sills (Table 6a), where it is close to zero (incoming TCant = outgoing TCant).

The trend in response to increasing atmospheric CO2 levels dominates the signal and the correlation at the expense of interannual variability. In order to identify the controls of interannual variability, the analysis was repeated with detrended time series. It reveals a strong correlation between the storage rate of Cant and its transport divergence for all three boxes (Table 6b). The correlation with air–sea fluxes is either not significant or weak (Table 6b). The analysis of model outputs suggests that while long-term changes in Cant storage rate are controlled by air–sea flux of anthropogenic CO2, its interannual variability is, on the other hand, driven by the divergence of Cant transport. Additional analyses were made to identify which role is played by the circulation in the annual evolution of the storage rate of Cant. For each box, correlations between detrended time series of Cant transport divergence and incoming or outgoing transport of Cant were assessed. These estimates, summarized in Table 7, show that the divergence of Cant transport is always correlated with the incoming transport of Cant and not with the outgoing transport of Cant. The interannual variability of the North Atlantic Cant storage rate is thus driven by the transport of Cant coming from the south. The intensity of the MOC controls the interannual variability of both terms at 25 and 36∘ N (Sect. 4.1). The analysis of the full 53-year period corroborates conclusions drawn for the period 2003–2011 (Sect. 3.3) and is in line with previous studies (Pérez et al., 2013; Zunino et al., 2014).

In this section, we identify major water masses making up the upper and lower limb of the MOC to evaluate their contributions to the regional Cant storage rate over the period 1959–2011. The North Atlantic circulation is well documented. Based on previous studies (e.g., Arhan, 1990; McCartney, 1992; Hernández-Guerra et al., 2015; Daniault et al., 2016) and on the vertical distributions of volume transports integrated zonally at 25∘ N, 36∘ N, OVIDE and the Greenland–Iceland–Scotland sills (Fig. 12), we defined three water classes: North Atlantic Central Water (NACW, Class 1), Intermediate Water (IW; Class 2) and North Atlantic Deep Water (NADW, Class 3).

NACW (Class 1) is transported by upper ocean circulation, either northward (Class 1N) by the Gulf Stream and the NAC, or southward (Class 1S) by the subtropical gyre recirculation in the western European basin. The southward recirculation is composed of colder and denser waters (Talley et al., 2008), allowing the distinction of Class 1S from Class 1N in our study (Fig. 12). NACW loses heat during its northward journey, which increases its density. As a result, the density limits between Classes 1N and 1S and 2 change with latitude. Based on Fig. 12, we defined Class 1N from the surface to σ1=29.1 kg m-3 at 25∘ N, 30 kg m-3 at 36∘ N and 31 kg m-3 at the OVIDE section. This class is not found at the Greenland–Iceland–Scotland sills. Class 1S, proper to the subtropical region, is found from 29.1 to 31 kg m-3 at 25∘ N and from 30 to 31 kg m-3 at 36∘ N.

IW (Class 2) encompasses the densest water masses of the MOC upper limb, such as Antarctic Intermediate Water (AAIW), Subantarctic Intermediate Water (SAIW) or Mediterranean Water (MW). Class 2 circulates northward between σ1=31 and 31.8 kg m-3 from 25∘ N to OVIDE and between σ1 = 31 and 31.9 kg m-3 through the Greenland–Iceland–Scotland sills (Fig. 12).

NADW (Class 3) supplies the lower limb of the MOC. It flows southward from the subpolar gyre to the subtropical region. In the model, it is found below σ1 = 31.7 kg m-3 at 25∘ N, 36∘ N and OVIDE and below σ1 = 31.9 kg m-3 at the Greenland–Iceland–Scotland sills (Fig. 12).

Analysis of the long-term changes in the simulated transport of volume and Cant across the four sections and for the three specified classes led to identify two periods, before and after 1995 (Fig. S4). The distinction between these two periods is based on Class 1N (northward NACW) at the OVIDE section and Class 2 (IW) at 36∘ N for which Cant and volume transports were nearly constant before 1995 but strongly increased after 1995 (Fig. S4). Based on these two periods, the discussion focuses first on 1959–1994 to understand how each water mass contributed to the North Atlantic Cant storage rate (Fig. 13a). The period 1996–2011 is analyzed next to evaluate the impact of the strong increase in Cant transport after 1995 on the storage rate of Cant.

Before 1995, more than 50 % of Cant transported by NACW flowing northward (Class 1N) at 25∘ N crossed 36∘ N, whereas 30 % recirculated southward within Class 1S. At the OVIDE section, the transport of Cant was equal to 12 % of the 25∘ N Cant transport, whereas it was close to zero at the sills (Fig. 13a). Figure 13a also reveals positive anthropogenic CO2 air–sea fluxes for the three boxes as well as a non-negligible Cant storage rate between 25 and 36∘ N. The net transport of Cant within Class 1 was nevertheless positive in all three boxes and higher than the associated Cant storage rate, which suggests a vertical transport of Cant from Class 1 to Class 2. The preceding comment suggests that NACW plays a key role in the Cant storage rate between 25∘ N and the OVIDE section, as well as in the transfer of Cant to a lower (denser) layer during its northward transport.

This cross-isopycnal transport between Class 1 and Class 2 (Fig. 13a) causes a decrease in the volume of Class 1 waters and an increase in the volume of Class 2 waters transported northward from 25∘ N to the OVIDE section (Fig. S4). This is in line with results from De Boisséson et al. (2012) who highlighted the densification of subtropical central water by winter air–sea cooling and mixing with intermediate waters along the NAC path. Moreover, results from Cant transport (Fig. 13a) also suggest that IW was enriched in Cant between 25∘ N and the OVIDE section over the study period. The large Cant uptake north of 36∘ N is explained by regional winter deep convection occurring along the NAC that mixes NACW, rich in Cant, with IW, poor in Cant. It should be noted that the Cant budget of Class 2 in Box 2 has a deficit of 0.01 PgC yr-1. This result suggests that an additional source (e.g., MW, Álvarez et al. 2005) supplies Cant to IW between 36∘ N and the OVIDE section.

Figure 13a also shows that 62 % of Cant entering in Box 3 by advection of Class 1 and Class 2 waters and by air–sea flux was converted into Class 3 inside the box and exported southward. The remainder was stored in Box 3 (18 %) or transported northward through the Greenland–Iceland–Scotland sills as Class 2 waters (19 %). NADW was thus strongly enriched in Cant between the OVIDE section and the Greenland–Iceland–Scotland sills by entrainment of NACW/IW and deep convection, which is in agreement with results from Sarafanov et al. (2012). Finally, a small fraction of Cant entering in Box 2 within Class 3 left the area across 25∘ N (24 %, Fig. 13a). The remainder was stored within Class 3 between 36∘ N and OVIDE.

After 1995, 27 % of Cant entering within Class 1 at 25∘ N flowed northward across the OVIDE section, which is 2 times higher than for the previous period (Fig. 13b). As discussed above, this relative increase in Cant transport at OVIDE was associated with a significant increase in volume transport across the section (Fig. S4a). The latter was multiplied by 1.9 after 1995 at the expense of the diapycnal transport between Class 1 and Class 2 waters, which decreased by 60 % compared to the previous period. As a result, less Cant is transferred from NACW to IW. Figure 13b shows that changes in Class 1 waters in Box 2 went along with a relative but small decrease in air–sea flux and in the net Cant transport across 36∘ N. In Class 1 of Box 3, the relative increase in Cant transport at OVIDE was concomitant with a similar increase in the contribution of the vertical transport of Cant to Class 2 waters as well as with a small decrease in the contribution of air–sea flux. Moreover, the relative increase in Cant transferred into Class 2 (Box 3) is associated with a relative increase in Cant transported within Class 2 waters throughout the Nordic sills, in Cant transported vertically into Class 3 waters and in the regional Cant stored inside the box (Class 2 Box 3), but also to a relative decrease in the Cant transport of Class 2 at OVIDE. The excess of NACW rich in Cant entering the northernmost box (OVIDE–sills) was transferred into IW before being exported to the Nordic regions or stored in the subpolar gyre.