Since the industrial revolution, atmospheric CO2 levels have been increasing almost exponentially as a result of human activities such as fossil fuel burning, cement production and land use changes. The global ocean has acted as a strong sink for this anthropogenic CO2 and is currently taking up approximately 25 % of the annual emissions . While the ocean has the capacity to store almost all of the anthropogenic CO2 released to the atmosphere, the emissions currently outpace the oceanic absorption rates . This is because the transport of anthropogenic CO2 from the atmosphere into the ocean interior is limited by the rate of vertical exchange between the surface and the deep ocean . Warming of the ocean will decrease this rate as a consequence of the increased stratification, and Earth system models predict a decline in oceanic anthropogenic CO2 uptake efficiency over the 21st century . Warming of the ocean will also affect CO2 solubility, primary production and other factors governing the distribution and inventory of natural carbon in the ocean . It is important to constrain the magnitude of these feedbacks for policy planning, but current estimates vary significantly among models. Observational-based quantitative and qualitative insight into carbon cycle climate interactions are important for the further improvement of projections of the future ocean carbon cycle.

Over the past 3 decades, ocean CO2 chemistry data have been collected on a regular basis in the world's oceans. As the observational record grows, direct evidence of the climate sensitivity of the marine carbon cycle emerges. For example, the Southern Ocean carbon sink exhibits clear variations in response to atmospheric circulation patterns; the sink was weakening from the early 1980s to the early 2000s , but has strengthened in the more recent decades . In the subarctic western North Pacific, measurements from 1992 to 2008 at the two time series stations KNOT and K2 reveal decadal trends in total dissolved inorganic carbon (DIC) related to alkalinity-driven reductions in CO2 outgassing . In the Mediterranean Sea, changes in the large-scale circulation result in variability in the anthropogenic CO2 concentration . Within the subpolar North Atlantic, high-quality carbon data have been collected almost every second year since the early 1990s , enabling the determination of subdecadal variability. This shows relationships between the anthropogenic CO2 storage rate and the extent and intensity of ventilation processes primarily driven by the North Atlantic Oscillation (NAO) .

The subpolar North Atlantic is a key region for the storage and transport of CO2 in the global ocean . While a response of anthropogenic CO2 storage to atmospheric forcing has been determined as mentioned above, less is known about variations in natural DIC, any relations to atmospheric forcing and relevance for total DIC inventories . Here, we analyse changes in total DIC and its natural and anthropogenic components in the central subpolar North Atlantic, in the Irminger Sea, in relation to the distribution and evolution of water masses over a 24-year period from 1991 to 2015, covering three periods of variable convective activity .

The Irminger Sea, a central sea in the subpolar North Atlantic (Fig. 1), is a climatically sensitive area with strong hydrographic contrasts. The subpolar North Atlantic circulation pattern has been extensively presented in the literature; here the description follows and . In the upper ocean, the East Greenland Current carries cold and fresh water of Arctic origin southwards in the west close to the shelf of Greenland. In the east, the Irminger Current carries warm and salty water northwards along the Reykjanes Ridge. The salinity and temperature signature of these warm water masses is affected by the strength and shape of the subpolar gyre . South of the Denmark Strait, they largely recirculate to the south. In the centre of the cyclonic circulation of the Irminger Gyre, preconditioning for convection is fulfilled and depending on heat loss, deep convection can occur . The extent and strength of convective processes are mainly driven by the state of the NAO, which is the leading mode of atmospheric variability over the mid-North Atlantic .

At depth, the circulation in the Irminger Sea is mainly characterized by the Denmark Strait Overflow Water (DSOW) and the Iceland–Scotland Overflow Water (ISOW). DSOW to the west is a relatively recently ventilated water mass enriched in oxygen (O2) and other dissolved atmospheric gases. It is composed of several water masses originating from the Arctic Ocean and the Nordic Seas . ISOW originates from the intermediate waters of the Nordic Seas, which are modified as they flow through the Iceland Basin to the Irminger Sea . In combination with ISOW and DSOW, Labrador Sea Water (LSW) forms North Atlantic Deep Water (NADW) , the key component of the lower limb of the Atlantic Meridional Overturning Circulation. Two main LSW classes are identified in the subpolar North Atlantic: the LSW1987-1994 and the LSW2000. LSW1987-1994, from now on called classical LSW (cLSW), is a dense, cold and relatively fresh water mass with high concentrations of dissolved atmospheric gases formed by the recurring winter convection in the mid-1980s and mid-1990s in the Labrador and Irminger seas . After 2000, the lighter LSW2000 or upper Labrador Sea Water (uLSW) has largely replaced cLSW .

Data from 15 cruises in the Irminger Sea covering 1991–2015 are used in this study (see Table 1). Data from the first 13 cruises were extracted from the GLODAPv2 data product, which provides bias-corrected, cruise-based, interior ocean data . The more recent data are from the 2012 OVIDE cruise (expocode: 29AH20120622) and the 2015 SNACS cruise (expocode: 58GS20150410). In order to minimize seasonal bias due to primary production, the upper 100 m of the water column are excluded from the inventory analysis. The region between 40.5 and 31.5∘ W was covered by all 15 cruises.

All cruises intersect the ocean currents of the Irminger Sea described in the previous section. The cruises occupied either WOCE section A01/AR07E, the FOUREX or the OVIDE section, and locations are presented in Fig. 1. In order for the sections to be fully comparable, a coordinate transformation was performed for the 1997 FOUREX data. The latitude and longitude coordinates were rotated to the AR07E section using Cape Farewell as a pivot point. Adjusting the distance between the stations ensured that the adjusted coordinates of the station over the Reykjanes Ridge on the FOUREX line matched the station over the Reykjanes Ridge on the AR07E line. The inventory estimates are sensitive to depth, and therefore the pressure coordinates of all cruises were normalized. The location of every station for all cruises was mapped to the 1 arcmin global relief model of the Earth's surface using a nearest-neighbour interpolation. The ratio between the bottom depth of this bathymetry and the reported cruise station bottom depth was multiplied with the pressure coordinates of each station. This normalization step mainly affected the adjusted FOUREX data, while for the other cruises the normalization changed sampling depths by less than 20 m.

The accuracy of the GLODAPv2 data product is better than 0.005 in salinity, 1 % in O2, 2 % in nitrate (NO3), 2 % in silicate (SiO2), 2 % in phosphate (PO4), 4 µmol kg-1 in DIC and 6 µmol kg-1 in total alkalinity (AT) . For 29AH20120623, the overall accuracy of NO3, PO4 and SiO2 was 1 %; the accuracy of DIC was 2 µmol kg-1 and for AT it was 4 µmol kg-1 . For the SNACS cruise in 2015, pressure, conductivity, temperature and dissolved O2 were directly measured with a Sea-Bird 911plus CTD profiler. At every station, water samples were obtained at 12 depths using Niskin bottles and used to calibrate the CTD measurements following the Global Ocean Ship-based Hydrographic Investigations Program (GO-SHIP) calibration procedure . The accuracy of bottle salinities analysed with a salinometer was ±0.003. The accuracy of O2 concentrations measured with Winkler titration using a potassium iodate solution as a standard was 0.2 µmol kg-1. The precision was better than 2 % in PO4, 1 % in SiO2 and 1 % in NO3 as evaluated using samples drawn from sets of Niskin bottles tripped at the same depth. DIC and AT were measured according to with an accuracy of 2 µmol kg-1 for both .

The seawater CO2 chemistry can be fully described if at least two of the four variables DIC, AT, CO2 partial pressure or pH are known. The measured variables at each of the 15 cruises are listed in Table 1. For six cruises, AT and pH were measured, and therefore DIC was calculated for these using the dissociation constants of . For three cruises, only DIC was measured. For these, AT was approximated using the salinity–alkalinity relationship for the North Atlantic from . This relationship is defined for the surface ocean only, and therefore its validity for the deep Irminger Sea was tested (Appendix A). The mean difference between the approximated and the measured AT data available was less than 5 µmol kg-1; this is better than the target accuracy of AT in the GLODAPv2 data product. No bias with depth or position was evident.

The total DIC concentration is partitioned into its natural and anthropogenic components (DIC = DICnat + Cant). The Cant concentration was estimated with the φCT∘ method (see Sect. 4.1). The DICnat concentration is the difference between DIC and Cant. For all cruises, the column inventories were estimated for DIC, DICnat and Cant. The inventories are sensitive to depth, and therefore column inventories were only estimated for the part of the transect covered by all 15 cruises between 40.5 and 31.5∘ W. The column inventory is the concentration profile integrated over the entire water column : Invs=∫0zcs⋅ϱdz. Here, Invs is the column inventory of any species s, cs its concentration, ϱ the density at in situ temperature and pressure and z the depth of the water column. The storage rate is the slope of a linear least-squares regression over the mean column inventories with time. The standard error of the slope is the error of the storage rate. Changes in inventories can be caused by changes in the distribution of water masses with different species concentrations or by changes in species concentration within the water masses. The distribution of water masses was determined using an extended optimum multiparameter analysis (eOMP; see Sect. 4.2). While the hydrographic parameters that describe a set of source water types (SWTs) used for the eOMP analysis are assumed to be time independent, the concentrations within each water mass of species such as DIC or Cant vary over time and can therefore only be resolved by evoking the concept of water mass mixing-weighted average concentration, i.e. archetypal concentration (see Sect. 4.2). Finally, the inventory changes can be decomposed into contributions from changes in the archetypal concentration of the source water types and from changes in layer thickness of each water mass assuming linearity: dInvtotdt=∑WM∂Inv∂cdcdt+∂Inv∂zdzdt. Here, ∂Inv∂cdcdt is the mean layer thicknesses with variable archetypal SWT concentrations, while ∂Inv∂zdzdt can be calculated as the mean archetypal SWT concentrations multiplied by the layer thickness changes over a specific time period. Hence, the two drivers of the observed inventory variability in total DIC and its natural and anthropogenic components can be identified.

Over the 24-year period considered here, the frequency and the amplitude of the mean winter NAO changed significantly . The convection in the subpolar gyre during winter, which is driven by the large-scale atmospheric circulation , has varied in strength and extent accordingly. Three distinctive time periods can be characterized from 1991 to 2015 by different levels of convective activity in the Irminger Sea. In the first period from 1991 to 1997, several consecutive positive NAO winters led to extensive deep convection in the entire subpolar gyre . In the second period from 2000 to 2007, the NAO was in a more neutral state and only shallow convection occurred in the Irminger Sea. In the third period from 2008 to 2015, three deep convective events took place in the Irminger Sea in 2008, 2012 and 2015 . The DIC, DICnat and Cant inventory changes are presented with respect to these three periods and for the entire 24 years of observations in the following sections.

The temporal change in DIC, DICnat and Cant concentration in the Irminger Sea is clearly visible in Fig. 5, which shows interpolated cruise station data for 1991, 1997, 2007 and 2015 as the start and end years of the three periods considered here. The increase in DIC is evident throughout the basin. From 1991–1997, cLSW with a low DIC signature dominates the basin, but a tongue of older IW over the Reykjanes Ridge transports relatively high DIC concentrations into the Irminger Sea. By 2015 the DIC concentration had increased by at least 10 µmol kg-1 compared to 1991, as visualized by the disappearance of the 2150 µmol kg-1 contour line (Fig. 5a). Temporal changes in DICnat concentration (Fig. 5b) are small compared to those in DIC and less systematic. The Cant concentration increases over time, not only at the surface, but also over the entire water column, which is indicated by the disappearance of the 20 µmol kg-1 contour line below 1500 m from 1991 to 2015 (Fig. 5c).

The column inventory time series of DIC, DICnat and Cant shown in Fig. 6 quantifies this large temporal change in the Irminger Sea sections. Typically, the DIC inventory increased by 1.43 ± 0.17 mol m-2 yr-1 from 1991 to 2015, from approximately 5645 to 5685 mol m-2. The Cant storage rate was 1.84 ± 0.16 mol m-2 yr-1 in the same time period, which is larger than the rate of DIC inventory change. At the same time, the DICnat inventory decreased at a rate of -0.57 ± 0.22 mol m-2 yr-1. Therefore, the annual change in the DIC inventory is mainly driven by the large Cant storage rate but partially offset by the loss in DICnat inventory. The variability in the DIC and Cant inventories over the 24-year period is of similar magnitude, as indicated by the error of the slope in Fig. 6, whereas the DICnat inventory varies slightly more. DICnat increases as water masses age and DICnat from the remineralization of organic matter accumulates, while it decreases during water mass renewal and/or ventilation, which brings water with preformed, relatively low DICnat concentrations into the ocean interior. It is notable that the Cant inventory increased sharply from 2012 to 2015, while there was a comparably large decline in the DICnat inventory. This is not an artefact of the method, but can be explained by the fact that the 2015 data were obtained during active convection in the Irminger Sea . During the strong convection, older water masses enriched in DICnat were replaced by water masses high in Cant due to their most recent contact to the atmosphere and relatively low in DICnat as DICnat from remineralization has not yet accumulated. In contrast to that, the peak in 2005 in the Cant inventory could be related to the advection of Cant-enriched water masses formed in the Labrador Sea or in the region south of Cape Farewell into the Irminger Sea , but this strong signal may also reflect the true error or reveal measurement bias.

The data that have been collected in the Irminger Sea over the past decades provide unequivocal evidence for climate forcing of the carbon cycle in this oceanic region. Over long timescales, the steady trend due to the uptake of anthropogenic CO2 clearly dominates, but at shorter timescales it varies and can also be significantly masked by variability in DICnat. In particular, this was the case in the period from 2000 to 2007 when the negative DICnat storage rate partially offset the increasing Cant storage, resulting in a DIC storage rate that was only about one-third of the storage rates in the preceding and later time periods considered here. In that period, when convection was shallow, the replacement of relatively DICnat-rich cLSW with relatively DICnat-poor UW and uLSW led to the loss of DICnat. Ageing might have increased the DICnat concentration, but the water masses in which these processes occur were flushed out of the study area. Climate feedback mechanisms that involve natural carbon cycling in the ocean are relevant to elucidate. While found steady-state conditions for the natural carbon cycle in the entire eastern subpolar North Atlantic between 2002 and 2010, this was not the case for the Irminger Sea from 1991–2015. A possible explanation for this discrepancy could be the longer time period considered here or that if both the Irminger Sea and the Iceland Basin are regarded, inventory changes might cancel each other out. This was shown to be the case for Cant by as a consequence of opposite changes in cLSW volume east and west of the Reykjanes Ridge.

The results presented here do not indicate a consistent response in the storage rates for anthropogenic CO2 to the NAO. The storage rates were similar for the first two periods, with predominantly high and low values of the NAO index, while it was clearly larger for the last period, with a predominantly high NAO index. In that time period, the Cant storage rate increased by a factor of 1.6 compared to the period from 2000 to 2007 (Fig. 9). The lack of change in Cant storage rates between the first two periods contrasts with the results of , who found that the storage rates were low from 1997–2006. This is a result of the differences in data: when using the same cruises and the same periods of time (i.e. 1997–2006 for the middle period) as in to estimate Cant storage rates, we estimated a significant decline in Cant storage rates for the middle period compared to the first (Cant storage rates 2.78 ± 0.28 mol m-2 yr-1 for 1991–1997; 1.06 ± 0.47 mol m-2 yr-1 for 1997–2006).

The DIC storage, on the other hand, does show a consistent response to the NAO index; it is larger in the periods with predominantly high-NAO-index winters (the first and the last) than in the middle period, with a predominance of low-NAO-index winters. This response was also found when using the same cruises and time periods as to calculate the storage rates. Altogether, this shows that while the calculation of Cant storage rates over small time periods is sensitive to data selection, the calculation of DIC storage rates is not. This is not unreasonable, as estimates of Cant inventories involve more variables (for example AOU and alkalinity), which increases the risk of introducing sampling or measurement biases. Regardless, convective events increase the storage of Cant; this is clearly demonstrated for the 2015 event also presented in detail by , who also used an independent approach for estimating Cant. Most likely because the thick layer of Cant-rich cLSW was mostly established in 1991 when the first data were collected, no large storage rate is observed for the early NAO positive period included here .

calculated North Atlantic column inventory changes in DIC using data extracted from GLODAPv1 and CARINA . To obtain further insight on the governing processes, they also included O2, AOU and DICabio, which is DIC corrected for the fraction of remineralized carbon and is closely related to anthropogenic carbon. However, their inventories were only estimated over the upper 2000 m of the water column. For DIC, the storage rate in the Labrador and Irminger seas combined was estimated to be 0.57 mol m-2 yr-1 . This is only one-third of our estimates of 1.43 ± 0.17 mol m-2 yr-1. This difference is likely the result of the different depth ranges, region and time periods evaluated. It is, however, noteworthy that the DICabio storage rate calculated by is not significantly different from zero. The entire increase in DIC inventory is explained by an increase in the remineralized fraction, as determined from AOU and their assumed C : O ratio. This implies either that there is no storage of Cant in the region or that any increase in Cant is completely offset by reduced CO2 solubility.

In order to compare to the results here, the DICabio, O2 and AOU storage rates are shown in Fig. 10 for the same periods as in Fig. 9. For the Irminger Sea cruise data from 1991–2015, the DICabio storage rate is 1.40 ± 0.17 mol m-2 yr-1 and explains the DIC storage rate almost entirely. The negligible trend in AOU explains the lack of difference between the DIC and DICabio storage rates. The estimated DICabio storage rate is lower than the Cant storage rate (1.92 ± 0.17 mol m-2 yr-1, Fig. 6), probably due to a decline in preformed DIC values, i.e. a loss of solubility as a consequence of the positive surface temperature trends in the Irminger Sea from 1991–2015 . The long-term warming trends in the surface ocean and thus the decreasing O2 solubility leads to a deoxygenation of -0.8 ± 0.3 mol m-2 yr-1 over the 24-year period, which is consistent with the suggested loss in preformed DIC.

On shorter timescales (Fig. 10b and c) all variables show large variations in the storage rates consistent with the already discussed changes in hydrography. In contrast to the findings of , DIC and DICabio storage rates are positive over all time periods. This might be partly due to the increased data coverage in our study. Over the entire time period, DICabio underestimates Cant by about 25 %. However, if the time period is too short, DICabio becomes much more variable because it includes solubility effects. For example, from 2000–2007 the DICabio underestimates the Cant storage rate by 40 % (compare Figs. 9c and 10c), while from 2008–2015 the DICabio overestimates the Cant storage rate by 20 % (compare Figs. 9d and 10d).

Similar to the Cant storage rates, both periods before and after 1997 show a loss in O2 despite stronger convection in the early 1990s. At the same time, the small increase in AOU does not reflect the constant inventory rate in DICnat before 1997, nor does the constant AOU storage rate reflect the loss in DICnat after 1997. Again, this is attributed to the fact that cLSW was mostly formed before 1991 . The mean O2 saturation degree over the entire water column was 89 % in the period from 1991 to 1997, which resulted in a high O2 inventory of 750 ± 4 mol m-2 indicative of the recent ventilation. After 1997, no re-ventilation took place and the mean O2 saturation dropped 2 %, while the inventory decreased to 732 ± 4 mol m-2. From 2008 to 2015, the O2 saturation values increased to 89 % again due to the strong inventory increase of 4.3 ± 2.6 mol m-2 yr-1, which was mainly driven by the deep convection in 2015 . Here, the strong loss in AOU matches the loss in DICnat as the water column is re-ventilated so that remineralized organic matter is replaced by a larger fraction of Cant. In fact, the convection in 2015 was strong enough to restore the O2 inventory levels so that the mean inventory from 2008 to 2015 was 749 ± 4 mol m-2, the same level as the well-ventilated early 1990s.

The repeat observations within the Irminger Sea show significant changes in total, natural and anthropogenic CO2 inventories from 1991 to 2015 with large interannual variability in the natural component. The eOMP method results and the decomposition of the inventory changes give valuable insight into the driving mechanisms to interpret the observed variability. Overall, changes in layer thickness of the main water masses appear most important for the DICnat inventory, while concentration change within these water masses is the key factor for Cant. Cant is typically more important for changes in total DIC inventories than DICnat. While the DIC inventory changes show a clear signal associated with the NAO, for Cant the signal is less robust, especially before and after 1997, likely because the data used here do not cover the period before 1991 when the thick layer of cLSW was formed. From 1991 to 2015 the mean Cant saturation over the entire water column increased from 52 to 67 %, increasing the Cant inventory from 53 ± 3 to 117 ± 3 mol m-2, mainly driven by the most recent convection in 2015. Despite the negative trend in O2 inventory from 1991 to 2015, the convection in 2015 was strong enough to replenish O2 levels at depth, leading to a mean saturation of 89 % and an O2 inventory of 749 ± 4 mol m-2, which was as high as in 1991. Cant is sensitive to the time period considered, while DIC appears more robust to sampling and measurement bias. Therefore, for a comprehensive view on carbon cycle feedback mechanisms, not only Cant but also natural and total DIC should be taken into account.