Abstract
The subpolar North Atlantic (SPNA) is a region of high anthropogenic CO2 (Cant) storage per unit area. Although the average Cant distribution is well documented in this region, the Cant pathways towards the ocean interior remain largely unresolved. We used observations from three Argo-O2 floats spanning 2013-2018 within the SPNA, combined with existing neural networks and back-calculations, to determine the Cant evolution along the float pathways from a quasi-lagrangian perspective. Our results show that Cant follows a stepwise deepening along its way through the SPNA. The upper subtropical waters have a stratified Cant distribution that homogenizes within the winter mixed layer by Subpolar Mode Water formation in the Iceland Basin. In the Irminger and Labrador Basins, the high-Cant footprint (> 55 μmol kg−1) is mixed down to 1400 and 1800 dbar, respectively, by deep winter convection. As a result, the maximum Cant concentration is diluted (<45 μmol kg−1). Our study highlights the role of water mass transformation as a first-order mechanism for Cant penetration into the ocean. It also demonstrates the potential of Argo-O2 observations, combined with existing methods, to obtain reliable Cant estimates, opening ways to study the oceanic Cant content at high spatio-temporal resolution.
Similar content being viewed by others
Introduction
Since the beginning of the industrial revolution, human activities such as fossil fuel burning and changes in land-use have led to the emission of large amounts of carbon dioxide (CO2) into the atmosphere. This excess of carbon, referred to as anthropogenic carbon (Cant), represented an addition of 9.5 ± 0.5 GtC yr−1 to the atmosphere between 2011 and 20201. The ocean has absorbed 25 ± 2% of Cant emissions (170 ± 35 GtC out of 660 ± 65 GtC of total Cant emissions) since the beginning of the industrial era1. The ocean acts, therefore, as a net Cant sink and a moderator of climate change. This net Cant uptake occurs via air-sea exchange, driven by air–sea CO2 disequilibria resulting from the difference between the steady increase in atmospheric pCO2 over the years since the pre-industrial era and oceanic pCO22. The latter is primarily regulated by seasonal variations of temperature in the subtropics. In contrast, at higher latitudes, its oscillation is typically dominated by biological processes3 affecting dissolved inorganic carbon (DIC) and total alkalinity (AT). A low DIC/AT ratio will lead to a large CO2 uptake capability by the ocean in response to an increase in atmospheric pCO24. As a result of the accumulation of Cant, the Revelle factor, measuring the buffer capacity of the ocean, decreases5. Ultimately, changes in oceanic circulation also have the capacity to modify the saturation of Cant at the ocean surface, enhancing or dampening the ocean carbon uptake capability6.
The global Cant distribution in the ocean is not homogeneous, neither vertically nor horizontally7. As expected from direct air-sea exchange, Cant concentration is maximum at the surface and decreases with depth. Of the global ocean, the North Atlantic Ocean (Fig. 1) is the region with the largest Cant inventory per surface area8,9, storing up to 23–38% of the total oceanic Cant2,10. Within this region, Cant uptake mostly occurs in the subtropics6,7 due to the low Revelle factor of subtropical surface waters2. These subtropical Cant-loaded waters are then transported northward toward subpolar latitudes6,11 by the upper branch of the Atlantic Meridional Overturning Circulation (AMOC), where additional Cant uptake takes place12. The subpolar North Atlantic (SPNA) gyre has been identified as one of the most important regions for the injection of Cant towards the deep ocean7. This deep Cant penetration is mainly due to the formation of Labrador Sea Water (LSW) through deep winter convection in the Labrador and Irminger Seas7,13,14,15,16. Considering the role of the ocean in moderating the ongoing climate change through its capacity to uptake and store Cant, any change in its uptake and storage rates might have drastic consequences on the climate system. Changes in the Cant storage rates in the North Atlantic Ocean are fundamentally related to the increased atmospheric pCO2, the changes in the AMOC strength, the changes in the North Atlantic Oscillation (NAO) index, and deep water formation6,16,17. The current Cant increasing trends and related Cant storage rates might thus be modified in the future due to the increasing atmospheric Cant, and the projected weakening of both AMOC18 and deep-water formation19,20.
To date, Cant estimates are mainly based on methods that rely on scarce but valuable ship-based measurements of carbonate system parameters (carbon-based methods2,6,21,22,23) or transient tracers such as CFCs (transient tracer-based methods9,16). However, studying the spatio-temporal evolution of oceanic Cant storage and understanding the processes involved is a crucial challenge that requires a more detailed view of the upper and deep Cant distribution and of its main pathways into the ocean interior. Additionally, the distribution of Cant on timescales shorter than GO-SHIP cruises is necessary to understand the effect of CO2 emissions reduction and carbon removal strategies on the ocean. Considering the unrivaled spatio-temporal sampling provided by the Argo-O2 network24, the purpose of this study is to demonstrate that Argo-O2 data combined with existing neural networks (i.e., ESPER_NN25, CANYON-B and its routine CONTENT26) and a back-calculation method (φCTO method17,27) can be used to obtain reliable Cant estimate at the finest spatio-temporal scale to date. As a case study, we selected three Argo-O2 floats to describe Cant deepening as it enters and propagates along the SPNA region. We highlight the long journey of Cant from the surface towards the interior of the SPNA gyre. Our results open up perspectives to investigate Cant storage changes and transport at seasonal to decadal timescales with the growing global and seasonally unbiased Argo-O2 data network compared to ship-based measurements.
Results
Distribution of Cant associated with the North Atlantic Current
Four main regions with the different vertical distributions of Cant were identified along the floats’ pathways: region 1, located East of the Grand Banks of Newfoundland, and regions 2, 3, and 4, comprising the Iceland, Irminger, and Labrador basins, respectively (Fig. 2a). Region 1 is located inside a dynamic transition zone separating the subpolar and subtropical domains of the North Atlantic Ocean28. This transition zone is typically characterized by a sharp thermohaline and density front, which, in this case (Fig. 2c), indicates that float 5904988 entered the North Atlantic Current (NAC)29. The thermohaline front is also accompanied by a sharp Cant and oxygen gradient (Fig. 2b and Supplementary Fig. 5). In the first 400–600 dbar of the water column, region 1 is characterized by saline, warm, and poorly oxygenated North Atlantic Central Water (NACW) of subtropical origin. At this location, the NACW comprises the northern extension of the upper and thermocline subtropical waters, defined as water masses located between the surface and σθ = 27.0 kg m−3 12. These waters are typically characterized by a low Revelle factor5, which favors atmospheric Cant uptake along their journey from the subtropics. As a consequence, the NACW presents the highest Cant values (maximum Cant > 65.0 μmol kg−1) of all water masses identified along the pathway of float 5904988 (Fig. 2a). The values of Cant,def close to zero (Supplementary Fig. 3) indicate that the NACW is saturated in Cant due to the air-sea Cant uptake occurring during the journey of the water masses that form the NACW. Furthermore, just eastward of the outcrop of the isopycnal 27.2 kg m−3, the high Cant concentrations (Cant > 55.0 μmol kg−1) reach the base of the winter mixed layer depth (MLD) at ~400 dbar (Fig. 2b). In this region, the high Cant signal is vertically transferred as the MLD deepens, highlighting the vertical Cant homogenization due to mixing during winter convection. Below the mixed layer, between August 2017 and March 2018, Cant concentration in region 1 decreased sharply to a relative minimum of 26.2 ± 6.6 μmol kg−1 near 800 dbar (Fig. 2b). These concentrations increase again up to 36.1 ± 7.5 μmol kg−1 at 1050 dbar, associated with the Labrador Sea Water (LSW). As inferred by the progression of the float, the Cant-loaded NACW is transported northwards by the NAC (Fig. 1), toward the Iceland Basin (region 2), following the general circulation pattern of the SPNA gyre30.
Distribution of Cant in the Iceland Basin and over the Reykjanes Ridge
Region 2 encompasses the Iceland Basin, including the eastern flank of the Reykjanes Ridge, and is characterized by the presence of Subpolar Mode Water (SPMW) in the first 600 dbar of the water column. In region 2, the SPMW contains the largest Cant values over the water column, whose average is 57.2 ± 5.8 μmol kg−1. The SPMW is formed during winter convection31,32 by intense air–sea buoyancy loss, leading to the densification of the Cant-loaded NACW on its way toward subpolar latitudes33. This process is well illustrated by the good agreement between the winter MLD and the depth of the SPMW base (Figs. 3b and 4b). In the Iceland Basin, the depth of the isoCant 50 μmol kg−1, which mostly follows the winter MLD, reaches up to 600 dbar (Figs. 3b and 4b), that is, 200 dbar deeper than in region 1. Below the SPMW, Cant concentration decreases rapidly, highlighting trapping of Cant within the SPMW and revealing that the formation of this water mass is a precursor in the deepening of Cant in the subpolar gyre. To further assess this deepening, Fig. 5 illustrates the mean Cant concentration along four density ranges describing the main water masses found in the four regions12,34. In the second-density layer (27.2 < σθ < 27.52 kg m−3) of region 2, the averaged Cant concentration along the float 6901023 pathway increases from 40.1 ± 6.2 μmol kg−1 in the south-easternmost part of the region to 55.5 ± 6.5 μmol kg−1 on the east side of the Reykjanes Ridge (Fig. 4b). This east-west Cant enrichment is likely due to the trajectory of the float, going from an area containing a thin SPMW layer, to an area with a thick SPMW layer (Fig. 3b). In this region, Cant,def fluctuates between −20 μmol kg−1 and 0 μmol kg−1 (Supplementary Fig. 3), suggesting that air–sea Cant enrichment takes place, explaining part of the east–west Cant enrichment in the Iceland basin. In the western part of Region 2, the Reykjanes Ridge entails a natural topographic barrier separating the Iceland and Irminger basins, two basins with different Cant distribution and water masses. The isoCant 35 μmol kg−1, which fluctuated between 600 and 800 dbar east of the Reykjanes Ridge, deepens until ~1200 dbar on the western side of the ridge (Figs. 3b and 4b). The second density layer of region 2 (27.2 < σθ < 27.52 kg m−3) disappears west of the ridge during the winter months (Fig. 5, Supplementary Figs. 6c and 7c) due to outcropping and subsequent surface water densification in the vicinity of the ridge12. As a consequence, the third oceanic layer (27.52 < σθ < 27.71 kg m−3) shallows on the western side of the Reykjanes Ridge, illustrated by the shallowing of the isopycnal 27.71 kg m−3 from ~1000 dbar east of the ridge to less than 500 dbar west of the ridge (Fig. 5b, c). Along with this densification of the surface water, Cant increases by 34% and 31% in the third layer for float 6901023 and float 6901026, respectively, marking a rapid transfer of Cant into the third oceanic layer. However, if it cannot be excluded that the Cant increase is only due to vertical entrainment, this increase could partly be due to lateral entrainment of water masses from another origin. These results highlight the role of the Reykjanes Ridge, separating the Cant distribution and dynamics between region 2 and region 334. In the fourth oceanic layer (27.71 < σθ < 27.80 kg m−3), Cant concentrations are the lowest and slightly increase along the float trajectories (Fig. 5b, c). This increase stays within the averaged uncertainty range (±5.9 μmol kg−1) and is therefore not significant.
Deep transport of Cant in the Irminger and Labrador basins
The Irminger Sea (region 3) is characterized by high concentrations of O2 (Supplementary Figs. 5b, 6b, and 7b), delimiting the core of the LSW34. These large O2 values are indicative of the intense air-sea exchange that characterizes this region13,35,36. In contrast to regions 1 and 2, here, the vertical Cant distribution has been subject to a large water-column homogenization. This leads to a weaker Cant gradient, with a gradual decrease of mean Cant concentrations from 52.0 ± 5.8 μmol kg−1 at the surface to 41.2 ± 7.9 μmol kg−1 at the lower bound of the LSW (~1400 dbar) (Figs. 2b and 4b). The high Cant values are constrained within the LSW, meaning that the formation of this water mass traps and transports Cant toward the deeper ocean. Yet, the maximum Cant concentrations are still found at the surface, with values reaching 52.0 ± 5.8 μmol kg−1 (Figs. 2b and 4b). As suggested by the negative Cant,def in the surface waters of the Irminger Sea (Supplementary Fig. 3b and 3c), these high surface concentrations are likely due to the intense air–sea CO2 exchange, following the same behavior of O2 exchanges35. Following the Western Boundary Current (WBC) (Fig. 1), Cant is transported out of the Irminger Sea. In the WBC, we observe the occurrence of occasional high-Cant pulses throughout the water column (black arrows on Figs. 2b and 3b) that we relate to anticyclonic mesoscale eddies (Supplementary Fig. 8). These anticyclonic features lead to a punctual downward isopycnal displacement of surface waters containing high Cant concentration, explaining the Cant pulses identified in our sections. Such eddies could carry Cant away from their formation regions and when they collapse, they might contribute to the isopycnal mixing of this tracer. Once in the Labrador Sea, float 6901023 illustrates how the isoCant 35 μmol kg−1 reaches a maximum pressure level of 1800 dbar in the inner part of the basin (Fig. 3b). That is, implying a ~ 200 dbar deepening of this isoCant compared to what is observed in the Irminger Sea. As in the Irminger Sea, in the Labrador Sea, the high Cant concentrations are also contained within the LSW, confirming that the formation of this water mass is a first-order deep-trapping mechanism for Cant.
Discussion
This study describes the Cant evolution along its pathways from the upper to the intermediate layers (up to 2000 dbar) of the SPNA gyre. It is based on a combination of existing neural networks and the back-calculation φCTO method, using Argo-O2 observations as input data. Our results are based on three floats following the main branches of the subpolar gyre circulation. From this quasi-lagrangian perspective, we identified four main Cant deepening stages and related them to their corresponding physical mechanisms. At the southernmost boundary of our study region, in the subtropical/subpolar transition zone near Flemish Cap, Cant is mainly trapped within the NACW layer (σθ < 27.2 kg m−3). The NACW has the largest mean Cant concentration (62.8 ± 5.9 μmol kg−1) of all water masses intercepted by the studied floats. At this location, winter mixing favors vertical Cant homogenization, by which high Cant concentrations (>45.0 μmol kg−1) reach down to ~400 dbar (1st deepening stage). The air–sea heat fluxes and mixing processes along the northward progression of the Cant-loaded NACW lead to a densification of NACW and its transformation into SPMW as it reaches the Iceland Basin. This transformation causes a deepening of Cant down to ~600 dbar in this region (2nd deepening stage). The average Cant concentration within this newly formed SPMW (57.2 ± 5.8 μmol kg−1 for σθ < 27.52 kg m−3) is lower than upstream due to the Cant content being spread out in the vertical. Further west, the Reykjanes Ridge is identified as a key topographic feature that marks a steep transition between the Iceland Basin and the Irminger Basin with markedly different vertical Cant distributions. In the Irminger Sea, the Cant signal (>35.0 μmol kg−1) deepens until 1400 dbar (3rd deepening stage) and is constrained within the LSW layer, highlighting the role of the LSW formation on the Cant deepening. The deepening is even more pronounced in the Labrador Sea, where the Cant signal, still confined within the LSW, reaches 1800 dbar (4th deepening stage). The newly formed LSW joins the Deep Western Boundary Current as part of the North Atlantic Deep Water (NADW), spreading the Cant footprint downstream towards the ocean interior. Further evidence of this Cant trapping within the LSW and its propagation downstream is provided by our data east of the Flemish Cap transition zone, where a relative Cant maximum can be identified within the LSW layer underneath the NACW (Fig. 2b). As surface waters spread through the different basins of the SPNA, we also inferred water masses that have favorable conditions to uptake additional atmospheric Cant, by means of the Cant,def proxy. Our results indicate that air-sea Cant fluxes are likely to occur all over the region, and as far west as in the Labrador Basin.
Deep water formation is a broadly known and acknowledged mechanism for Cant penetration in the interior of the North Atlantic Ocean2,6,7. Relying on Argo-O2 data and existing methods, we presented an original approach that allowed us to describe this Cant deepening at the finest spatiotemporal scale to date. Our results revealed the role of the SPMW formation as an early stage of subpolar Cant deepening. This result reinforces the idea that Cant deepening occurs throughout the whole SPNA gyre and is not only linked to the main convective sites (i.e., Irminger and Labrador Basins), as traditionally attributed7,16. Mainly uptaken from the atmosphere in the subtropics2, Cant is transported northwards towards the SPNA region by the NAC6, which at these latitudes is the main component of the upper AMOC limb. The quasi-lagrangian Argo-float trajectories are a suitable and timely means for describing this horizontal Cant journey. Despite lateral advection by the upper AMOC limb, being the main source of Cant in the region, the contribution from ocean Cant uptake is also relevant6,12. By means of the Cant,def proxy, the Argo-O2 data also support the hypothesis of surface air-sea Cant enrichment all along the subpolar gyre domain by an outcropping of the low-carbon nutrient stream isopycnal range12. Given the dominant role of lateral advection in Cant distribution, it is not surprising that AMOC variability affects the northward Cant transport towards the SPNA and its storage rates6,11. Under current climate change, the AMOC is projected to weaken37, and we could thus expect a concomitant decrease in Cant storage rates and content in the SPNA. Nevertheless, the northward oceanic Cant transport might still be subject to continuous increase in response to the rise in anthropogenic CO2 emissions11,38. Estimates on a finer spatio-temporal scale, such as those based on Argo-O2 observations, should be sustained in the long term to determine which effect will be dominant in the future.
Our case study has shown that neural networks combined with high-quality in situ Argo-O2 measurements and a back-calculation method can effectively be used to retrieve Cant concentration through the three-dimensional estimates of oceanic variables (nutrients, DIC, total alkalinity). These oceanic variables are needed to compute Cant at a higher spatio-temporal resolution than that available from the current biogeochemical (BGC) observation networks. Hence, the methodology presented here, enchaining Argo-O2 observations and existing methods, opens up a promising step forward to study Cant distribution, its variability, and the processes driving these variations, which are crucial challenges regarding the ocean carbon sink39. The global nature of the Argo-O2 array gives the opportunity to significantly increase the spatio-temporal coverage of current Cant estimates. However, we can only guarantee the validity of our approach in the case-study region and for the period (1970–2020) over which the ship-based data used for training and validation of the neural networks were collected (GLODAPv2.202040). Neural networks can only reproduce what they have learned. Therefore they must be trained with new data as the ocean changes, which is particularly true for the SPNA gyre where interannual to decadal changes are considerable (e.g., cold blob41 or great salinity anomaly42). The importance of high-accuracy in situ data measurements for training the neural networks cannot be disregarded, as such, the continuation of ship-based observations should be assured. Another important consideration of our results and their interpretation lies within the averaged total uncertainty of our method, which is ±5.9 µmol kg−1. Any spatio-temporal change in Cant concentration below two times this value should be interpreted with caution. This current methodological uncertainty could possibly be reduced by using data from the growing BGC-Argo database43,44 to train the neural networks. Moreover, BGC-Argo floats, equipped with pH and nitrate sensors, could be directly used to derive Cant, hence providing independent validation Cant estimates when both BGC-Argo and Argo-O2 floats are concomitant in space–time. However, BGC-Argo floats were not included in our study because available floats did not match our selection criteria. In view of the above, there is, therefore, a complementarity between the scarce high-quality ship-based BGC measurements and the high spatiotemporal BGC-Argo and Argo-O2 observations. This complementarity is even more marked because Argo data need to be adjusted (QCed) against ship-based data. The Argo-O2 float data presented in this case study reached a maximum pressure level of 2000 dbar, limiting our result to the top 2000 dbar of the water column. The OneArgo mission calls for the deployment of 1250 floats capable of reaching 4000 or 6000 dbar45. Using Deep-Argo floats equipped with dissolved oxygen sensors could thus strengthen our findings and improve our knowledge about the mechanisms controlling the deep/abyssal Cant transport. For instance, the role of the Iceland-Scotland Overflow Water (ISOW), flowing below 2000 dbar, could not be identified here even though this water mass is known to contain Cant11.
Finally, it is worth mentioning that Earth system models tend to misrepresent the vertical penetration of Cant due to the shallow and weak formation of NADW46, even if this issue is better addressed in the latest generation of CMIP6 models18,47,48. The results of this study provide insights subject to improving the representation of the key pathways for the deep Cant propagation in Earth system models. Implementing a better representation of these pathways would definitely ameliorate the simulated vertical distribution of Cant and, ultimately, improve climate predictions.
Methods
Argo float observations
To highlight the Cant penetration into the deep ocean, we use Argo-O2 floats, giving the opportunity to obtain one Cant profile every 10 days. We selected the Argo-O2 floats that (1) followed the cyclonic pathway of the SPNA gyre, (2) had a lifetime longer than 3 years, and (3) crossed the A25 OVIDE hydrographic section49, which will serve as a reference dataset for validation. Three floats matched these criteria, comprising a total of 651 profiles of pressure (P), temperature (T), salinity (S), and oxygen (O2)50 within the region 47–65°N; 15–65°W (Supplementary Fig. 1). Data span from July 2012 to April 2018. Only data adjusted in delayed mode with a quality flag of 1 or 2 (good or probably good data)51 were used in our analysis (Supplementary Table 1). The accuracy of the data is assumed to be 0.002 °C, 0.01, and 2.4 dbar for temperature, salinity, and pressure, respectively52, and better than 3 μmol kg−1 for oxygen53.
Estimating biogeochemical variables with neural networks
To derive the biogeochemical variables (nutrients, DIC, and AT) needed to compute Cant, we rely on predictive neural networks. Neural networks (NN) are machine learning algorithms based on a multi-layer perceptron54,55, that are trained and validated against observations. The input data of the neural networks are, in this study, Argo P/T/S/O2 measurements. We used two neural networks, namely, ESPER_NN25 and CANYON-B, associated with its routine CONTENT26. ESPER_NN was used to obtain the macronutrients (phosphate, nitrate, and silicate) using P, T, S, O2, location, and time as predictors. ESPER_NN was adopted over other ESPER methods (i.e., ESPER_LIR and ESPER_Mixed) because it gives the lowest biases and root mean square errors over the global ocean for the predicted macronutrients25. ESPER_NN reproduces the validation dataset with average biases and errors of ~2% for macronutrients (Supplementary Table 1)25. However, in the North Atlantic Ocean, ESPER_NN gives uncertainties of ~1.3% for the predicted carbonate variables (AT and DIC), which is higher than the previous NN. Consequently, AT and DIC were computed with CANYON-B26, and the outputs of CANYON-B were passed through the CONTENT routine. This routine ensures consistency between carbonate variables and thus reduces the uncertainties of the carbonate system variables to ~0.5% for AT and DIC (Supplementary Table 1). We used neural networks rather than climatological products (such as, e.g., GLODAPv2 climatology56) because gridded climatologies are unable to capture changes in carbon variables in regions where water masses move laterally due to mesoscale processes or rapid circulation changes. This is because climatologies represent the mean field in which the variability has been smoothed out. In contrast, neural networks, such as ESPER_NN and CANYON-B, derive biogeochemical variables based on water mass characteristics, hence they can cope with such changes. In particular, the use of Argo-O2 float data as input to the neural networks reinforces this water-mass change tracking capability due to the quasi-lagrangian behavior of the Argo-O2 floats and their temporal resolution, three times greater than monthly climatological products.
Anthropogenic carbon estimates
To determine the anthropogenic carbon fraction (Cant) from DIC, we used the carbon-based back-calculation φCTO method17,27. This method has been widely applied to study the inventory of Cant, its storage rates, its variability14,17,57, and the influence of Cant on ocean acidification58. In any back-calculation method, Cant is computed as a two-step approach: (1) the changes in DIC due to biological activity occurring since a water parcel has left the ocean surface are removed, and (2) the pre-industrial preformed DIC is also removed, with the residual being interpreted as Cant. The φCTO method (Supplementary Equation 1) presents two main advantages compared to other back-calculation methods. First, this method accounts for the spatiotemporal variability of the preformed AT. Second, the parameterization of the preformed AT and disequilibrium terms are determined using the subsurface layer as a reference27. These two modifications improve the Cant estimates in cold and deep water formation regions subject to strong mixing processes, such as the SPNA gyre. A study comparing observational methods to estimate Cant in the Atlantic Ocean, including the TTD9, the TrOCA59, the CoIPSL60, the ΔC*61, and the φCTO method17, demonstrated that the latter provided the closest value to the average of all methods for the whole latitudinal range27. Based on that, we selected the φCTO method to perform our analysis. The input variables for this method are date, geographical location, T, S, O2 (in this study, the Argo variables), the macronutrients (obtained from ESPER_NN), plus AT and DIC (computed from CANYON-B and CONTENT). The φCTO method requires date as input because it adjusts the oceanic Cant concentration against the corresponding atmospheric pCO2. It also requires the geographical location (latitude and longitude) to define the water masses involved in the computation of the preformed AT and disequilibrium terms. All the input variables have associated uncertainties that propagate through Cant calculations. The uncertainties on Argo pressure, temperature, and salinity data are 2.4 dbar, 0.002 °C, and 0.01, respectively52. These uncertainties relate to the instrument’s performance. The uncertainties on O2 data are about 3.5 μmol kg−1 (Supplementary Table 1). The uncertainties of the predicted variables AT, DIC, and macronutrients are, on average, 10.85, 10.48, and 0.73 μmol kg−1, respectively (Supplementary Table 1). These NNs uncertainties are provided as part of the NNs predictive outputs and they result from the combination of training measurement uncertainties, the Bayesian method uncertainties, and the NNs input sensitivities26. Taking all of them into account, we quantified the uncertainty associated with our Cant estimate by randomly generating 100 Cant fields (Supplementary Table 1) using a Monte Carlo method62. The standard deviation from these Cant fields fluctuates between ±5.4 μmol kg−1 and ±10.2 μmol kg−1 (Supplementary Fig. 2) and the overall average value (for the three floats) is ±5.9 μmol kg−1. It is worth highlighting that the uncertainties of the modeled tracer fields (Supplementary Table 1) are evidently greater than the corresponding in situ measured variables (i.e., by 1.5 μmol kg−1 for oxygen). However, the uncertainties of the predicted variables are yet reasonably small to serve as inputs to the back-calculation method. Consequently, they allow an estimation of Cant that is statistically significant (and whose final uncertainties lie within 9% to 19% of the signal). To further validate our approach, we compared these ArgoTSO2-NN-based Cant estimates with ship-based Cant estimates obtained during OVIDE cruises49,63. We identified Argo-O2 profiles within the 35 km radius and less than 7 month-timespan of an OVIDE CTD cast (Supplementary Fig. 2). For each CTD cruise cast, Cant is estimated from in situ bottle-sample measurements of nutrients, oxygen, AT and pH, via the φCTO method (standard procedure11). Emulating the approach proposed in this study for the Argo-O2 floats (TSO2-NN procedure), we re-estimated Cant by using the ship-based P/T/S/O2 data as input to the neural networks to obtain nutrients, DIC, and AT, and then applied the φCTO method. These Cant profiles (ship-based standard and ship-based TSO2-NN) were then compared to our Argo-based Cant estimates (Argo-based TSO2-NN procedure), and all of them agreed well within the averaged method uncertainty of ±5.9 μmol kg−1 (Supplementary Fig. 2). This comparison indicates that the Cant values obtained from ship-based measurements and Cant estimates calculated via Argo-O2 data and neural networks give consistent results.
We also computed a proxy that allows the indirect assessment of the air–sea Cant uptake capacity, the Cant deficit (Cant,def, Supplementary Fig. 3). This proxy was defined by Ridge & McKinley12 as the difference between the Cant concentration of a water parcel and a reference (\({\langle {{{{{\rm{C}}}}}}_{{{{{\rm{ant}}}}}}\rangle }_{{\sigma }_{\theta } < 26.5}\)) computed as the average Cant concentration between the surface and the potential density (σθ) of 26.5 kg m−3. This isopycnal delimits the subtropical surface waters assumed to be in equilibrium with the current atmosphere. A water parcel with a negative Cant,def value indicates that this parcel has a deficit of Cant. If transported to the surface, this water parcel will have a lower pCO2 compared to a water parcel without this deficit, hence favoring the air–sea Cant uptake. Usually, Cant,def becomes more and more negative with depth (Supplementary Fig. 3), meaning that the deep waters would uptake a higher amount of Cant compared to the subsurface waters if they were transported to the ocean surface. The reference value is computed for 2002 with the GLODAPv2 dataset56, considering the averaged Cant concentration in surface waters (σθ < 26.5 kg m−3) offshore the Venezuelan coasts (10–20°N; 70–64°W), and is \({\langle {{{{{\rm{C}}}}}}_{{{{{\rm{ant}}}}}}\rangle }_{{\sigma }_{\theta } < 26.5}^{2002}\,\) = 51.5 μmol kg−1. We assume that our reference water mass offshore the Venezuelan coasts takes 5 years to arrive in the Iceland and Irminger basins64. For a given float, we use that reference value rescaled in time to 5 years before the time of its first profile, using the exponential equation of Carter et al.25. Assuming that due to its lagrangian behavior, the float follows the transformation of the same water mass, we kept the same reference value along the float trajectory within the subpolar gyre. However, this assumption is not valid for float 5904988 when it drifted from Region 4 to Region 1 and entered the North Atlantic Current, where we re-initialized the reference value.
Mixed layer depth and water mass definition
To determine the mixed layer depth (MLD) along the Argo float trajectories, we use the threshold method, which is based on a density difference of 0.01 kg m−3 65 between the surface and the base of the mixed layer66. Since this method tends to overestimate the MLD during deep convection events in winter67, we conducted a visual inspection68 to verify the MLD values during these events. When the MLD values disagreed between the threshold method and the visual inspection, we kept the value determined by visual inspection.
The North Atlantic is composed of several water masses (Fig. 1), most of which can be identified by their potential temperature (θ), S, σθ, O2, and potential vorticity. Here, the potential vorticity is calculated via the Brunt–Väisälä frequency, the gravity, and the Coriolis parameter. In the upper layers, the North Atlantic Current (NAC) flowing from the subtropics to the subpolar regions (Fig. 1), carries North Atlantic Central Water (NACW)30. The lower limit of this water mass is defined by σθ = 27.2 kg m−3 69 (Supplementary Fig. 4). Continuous air-sea interaction and winter convection along the NAC path cools and freshens the NACW, leading to its transformation into Subpolar Mode Waters (SPMWs)31,32. SPMWs are characterized by nearly homogeneous properties, with S > 34.98, σθ < 27.71 kg/m3, and a potential vorticity lower than 6 × 10−11 m−1 s−1 31,70 (Figs. 3 and 4). At intermediate depths (400–2000 dbar), in the subpolar gyre, the Labrador Sea Water (LSW) is located between 27.71 < σθ < 27.80 kg/m3. The LSW is formed by deep winter convection in the Labrador Sea71 and Irminger Sea36,65, trapping large amounts of O2 and Cant during its formation72. In those basins, we defined the LSW by O2 ≥ 290 μmol kg−1 34. However, along its path, the LSW loses O2 due to mixing and biological activity. Therefore the O2 criteria is not fully reliable to identify the LSW core away from its source regions. As a consequence, outside the Irminger and Labrador Seas we identify the LSW by S < 34.9453. At deeper levels (up to 2000 dbar), the Iceland-Scotland Overflow Water (ISOW) flows along the Reykjanes Ridge and is identified by σθ > 27.80 kg m−3 and S > 34.9434 (Supplementary Fig. 4).
Eddy identification
To detect the eddies along the Argo-float pathways, we run the autonomous eddy identification scheme of Faghmous et al.73, which monitors mesoscale ocean eddy activity in global sea level anomaly (SLA) dataset (http://www.aviso.altimetry.fr/duacs/). The scheme starts from the simple notion that every eddy has a single extremum, defined as a grid cell whose SLA is higher or lower than its 24 neighbors in a 5 × 5 grid. To determine the contour of the eddy, the algorithms search for the largest possible contour that would allow the feature not to violate the assumption that an eddy can have only a single extremum within its interior. For each eddy feature identified at time t, the eddies at time t + 1 are searched to find the closest feature within a pre-defined search space. When a feature at time t is associated with another feature at time t + 1, their sizes are compared to ensure that they are reasonably similar from a physical point of view. There are two major uncertainties associated with this eddy identification method. First, the use of a geometric eddy definition, as opposed to a physical one, introduces uncertainties in the eddy boundaries because they are not necessarily associated with the physical properties of the eddy. Second, as the method constrains features to have a single extremum, small features that are in close proximity might cancel each other out and not be detected. Despite these limitations, the success discovery rate of the autonomous eddy identification algorithm is 96.4%73.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
The original Argo data can be freely downloaded on the Euro Argo Data Selection platform (https://dataselection.euro-argo.eu/). Our Cant estimates, needed to evaluate the conclusion of the paper, can be downloaded on Zenodo (https://doi.org/10.5281/zenodo.7071614).
Code availability
The Matlab codes used for the simulation are available from the corresponding author upon request.
References
Friedlingstein, P. et al. Global carbon budget 2021. Earth Syst. Sci. Data 14, 1917–2005 (2022).
Sabine, C. L. et al. The oceanic sink for anthropogenic CO2. Science 305, 367–371 (2004).
Takahashi, T. et al. Climatological mean and decadal change in surface ocean pCO2, and net sea–air CO2 flux over the global oceans. Deep Sea Res. Part II Top. Stud. Oceanogr. 56, 554–577 (2009).
Egleston, E. S., Sabine, C. L. & Morel, F. M. Revelle revisited: buffer factors that quantify the response of ocean chemistry to changes in DIC and alkalinity. Glob. Biogeochem. Cycles 24, GB1002 (2010).
Revelle, R. & Suess, H. E. Carbon dioxide exchange between atmosphere and ocean and the question of an increase of atmospheric CO2 during the past decades. Tellus 9, 18–27 (1957).
Pérez, F. F. et al. Atlantic Ocean CO2 uptake reduced by weakening of the meridional overturning circulation. Nat. Geosci. 6, 146–152 (2013).
Davila, X. et al. How is the ocean anthropogenic carbon reservoir filled? Glob. Biogeochem. Cycles 36, e2021GB007055 (2022).
DeVries, T. The oceanic anthropogenic CO2 sink: Storage, air-sea fluxes, and transports over the industrial era. Glob. Biogeochem. Cycles 28, 631–647 (2014).
Waugh, D., Hall, T., McNeil, B., Key, R. & Matear, R. Anthropogenic CO2 in the oceans estimated using transit time distributions. Tellus B Chem. Phys. Meteorol. 58, 376–389 (2006).
Steinfeldt, R., Rhein, M., Bullister, J. L. & Tanhua, T. Inventory changes in anthropogenic carbon from 1997–2003 in the Atlantic Ocean between 20S and 65N. Glob. Biogeochem. Cycles 23, GB3010 (2009).
Zunino, P. et al. Variability of the transport of anthropogenic CO2 at the Greenland–Portugal OVIDE section: controlling mechanisms. Biogeosciences 11, 2375–2389 (2014).
Ridge, S. & McKinley, G. Advective controls on the North Atlantic anthropogenic carbon sink. Glob. Biogeochem. Cycles 34, e2019GB006457 (2020).
Fröb, F. et al. Irminger Sea deep convection injects oxygen and anthropogenic carbon to the ocean interior. Nat. Commun. 7, 13244 (2016).
Fröb, F. et al. Inorganic carbon and water masses in the Irminger Sea since 1991. Biogeosciences 15, 51–72 (2018).
Pérez, F. F. et al. Meridional overturning circulation conveys fast acidification to the deep Atlantic Ocean. Nature 554, 515–518 (2018).
Raimondi, L., Tanhua, T., Azetsu-Scott, K., Yashayaev, I. & Wallace, D. W. A 30-year time series of transient tracer-based estimates of anthropogenic carbon in the Central Labrador Sea. J. Geophys. Res. Oceans 126, e2020JC017092 (2021).
Pérez, F. F. et al. Temporal variability of the anthropogenic CO2 storage in the Irminger Sea. Biogeosciences 5, 1669–1679 (2008).
Jackson, L. C. et al. The evolution of the North Atlantic meridional overturning circulation since 1980. Nat. Rev. Earth Environ. 3, 241–254 (2022).
Schweckendiek, U. & Willebrand, J. Mechanisms affecting the overturning response in global warming simulations. J. Clim. 18, 4925–4936 (2005).
Stouffer, R. J. et al. Investigating the causes of the response of the thermohaline circulation to past and future climate changes. J. Clim. 19, 1365–1387 (2006).
Gruber, N. et al. The oceanic sink for anthropogenic CO2 from 1994 to 2007. Science 363, 1193–1199 (2019).
Müller, J. D. et al. Decadal trends in the oceanic storage of anthropogenic carbon from 1994 to 2014. AGU Adv. 4, e2023AV000875 (2023).
Woosley, R. J., Millero, F. J. & Wanninkhof, R. Rapid anthropogenic changes in CO 2 and pH in the Atlantic Ocean: 2003-2014: anthropogenic CO 2 increase in Atlantic. Glob. Biogeochem. Cycles 30, 70–90 (2016).
Roemmich, D. et al. On the future of argo: a global, full-depth, multi-disciplinary array. Front. Mar. Sci. 6, 439 (2019).
Carter, B. R. et al. New and updated global empirical seawater property estimation routines. Limnol. Oceanogr. Methods 19, 785–809 (2021).
Bittig, H. C. et al. An alternative to static climatologies: robust estimation of open ocean CO2 variables and nutrient concentrations from T, S, and O2 data using Bayesian neural networks. Front. Mar. Sci. 5, 328 (2018).
Vazquez-Rodriguez, M. et al. Anthropogenic carbon distributions in the Atlantic Ocean: data-based estimates from the Arctic to the Antarctic. Biogeosciences 6, 439–451 (2009).
Schneider, L. et al. Variability of Labrador Sea water transported through flemish pass during 1993-2013. J. Geophys. Res. Oceans 120, 5514–5533 (2015).
Dutkiewicz, S., Follows, M., Marshall, J. & Gregg, W. W. Interannual variability of phytoplankton abundances in the North Atlantic. Deep Sea Res. Part II Top. Stud. Oceanogr. 48, 2323–2344 (2001).
Daniault, N. et al. The northern North Atlantic Ocean mean circulation in the early 21st century. Prog. Oceanogr. 146, 142–158 (2016).
de Boisseson, E., Thierry, V., Mercier, H., Caniaux, G. & Desbruyères, D. Origin, formation and variability of the subpolar mode water located over the Reykjanes Ridge. J. Geophys. Res. Oceans 117, C12005 (2012).
McCartney, M. S. & Talley, L. D. The subpolar mode water of the North Atlantic Ocean. J. Phys. Oceanogr. 12, 1169–1188 (1982).
Brambilla, E., Talley, L. D. & Robbins, P. E. Subpolar Mode Water in the northeastern Atlantic: 2. Origin and transformation. J. Geophys. Res 113, C04026 (2008).
Petit, T., Mercier, H. & Thierry, V. New insight into the formation and evolution of the East Reykjanes Ridge current and Irminger current. J. Geophys. Res. Oceans 124, 9171–9189 (2019).
Maze, G. et al. Mass, nutrient and oxygen budgets for the northeastern Atlantic Ocean. Biogeosciences 9, 4099–4113 (2012).
Pickart, R. S., Straneo, F. & Moore, G. K. Is Labrador Sea water formed in the Irminger basin? Deep Sea Res. Part Oceanogr. Res. Pap. 50, 23–52 (2003).
Caesar, L., Rahmstorf, S., Robinson, A., Feulner, G. & Saba, V. Observed fingerprint of a weakening Atlantic Ocean overturning circulation. Nature 556, 191–196 (2018).
Brown, P. J. et al. Circulation-driven variability of Atlantic anthropogenic carbon transports and uptake. Nat. Geosci. 14, 571–577 (2021).
Gruber, N. et al. Trends and variability in the ocean carbon sink. Nat. Rev. Earth Environ. https://doi.org/10.1038/s43017-022-00381-x (2023).
Olsen, A. et al. An updated version of the global interior ocean biogeochemical data product, GLODAPv2. 2020. Earth Syst. Sci. Data 12, 3653–3678 (2020).
Shi, J., Wang, J., Ren, Z., Tang, C. & Huang, F. Cold blobs in the subpolar North Atlantic: seasonality, spatial pattern, and driving mechanisms. Ocean Dyn. 73, 267–278 (2023).
Biló, T. C., Straneo, F., Holte, J. & Le Bras, I. A. ‐A. Arrival of new great salinity anomaly weakens convection in the Irminger Sea. Geophys. Res. Lett. 49, e2022GL098857 (2022).
Addey, C. I. Using biogeochemical Argo floats to understand ocean carbon and oxygen dynamics. Nat. Rev. Earth Environ. 3, 739–739 (2022).
Claustre, H., Johnson, K. S. & Takeshita, Y. Observing the global ocean with biogeochemical-Argo. Annu. Rev. Mar. Sci. 12, 23–48 (2020).
Roemmich, D. et al. The Argo Program. in Partnerships in Marine Research 53–69 (Elsevier, 2022).
Doney, S. C. et al. Evaluating global ocean carbon models: the importance of realistic physics. Glob. Biogeochem. Cycles 18, GB3017 (2004).
Goris, N. et al. Constraining projection-based estimates of the future North Atlantic carbon uptake. J. Clim. 31, 3959–3978 (2018).
Terhaar, J., Frölicher, T. L. & Joos, F. Southern Ocean anthropogenic carbon sink constrained by sea surface salinity. Sci. Adv. 7, eabd5964 (2021).
Sarthou, G. et al. Introduction to the French GEOTRACES North Atlantic transect (GA01): GEOVIDE cruise. Biogeosciences 15, 7097–7109 (2018).
Argo. Argo float data and metadata from global data assembly centre (Argo GDAC). Seanoe (2000).
Thierry, V., Bittig, H. & The Argo-BGC team. Argo Quality Control Manual for Dissolved Oxygen Concentration. (2021).
Wong, A. P. S. et al. Argo Data 1999–2019: two million temperature-salinity profiles and subsurface velocity observations from a global array of profiling floats. Front. Mar. Sci. 7, 700 (2020).
Racapé, V., Thierry, V., Mercier, H. & Cabanes, C. ISOW spreading and mixing as revealed by Deep-Argo floats launched in the Charlie-Gibbs fracture zone. J. Geophys. Res. Oceans 124, 6787–6808 (2019).
Bishop, C. M. Neural Networks for Pattern Recognition. (Oxford University Press, 1995).
Rumelhart, D. E., Hinton, G. E. & Williams, R. J. Learning representations by back-propagating errors. Nature 323, 533–536 (1986).
Lauvset, S. K. et al. A new global interior ocean mapped climatology: the 1x1 GLODAP version 2. Earth Syst. Sci. Data 8, 325–340 (2016).
Ríos, A. F., Velo, A., Pardo, P. C., Hoppema, M. & Pérez, F. F. An update of anthropogenic CO2 storage rates in the western South Atlantic basin and the role of Antarctic Bottom Water. J. Mar. Syst. 94, 197–203 (2012).
García-Ibáñez, M. I. et al. Ocean acidification in the subpolar North Atlantic: rates and mechanisms controlling pH changes. Biogeosciences 13, 3701–3715 (2016).
Touratier, F., Azouzi, L. & Goyet, C. CFC-11, Δ14C and 3H tracers as a means to assess anthropogenic CO2 concentrations in the ocean. Tellus B 59, 318–325 (2007).
Lo Monaco, C., Metzl, N., Poisson, A., Brunet, C. & Schauer, B. Anthropogenic CO2 in the Southern Ocean: Distribution and inventory at the Indian‐Atlantic boundary (World Ocean Circulation Experiment line I6). J. Geophys. Res. Oceans 110, 2004JC002643 (2005).
Gruber, N., Sarmiento, J. L. & Stocker, T. F. An improved method for detecting anthropogenic CO2 in the oceans. Glob. Biogeochem. Cycles 10, 809–837 (1996).
Metropolis, N. & Ulam, S. The Monte Carlo Method. J. Am. Stat. Assoc. 44, 335–341 (1949).
Mercier, H. et al. Variability of the meridional overturning circulation at the Greenland–Portugal OVIDE section from 1993 to 2010. Prog. Oceanogr. 132, 250–261 (2015).
Messias, M.-J. & Mercier, H. The redistribution of anthropogenic excess heat is a key driver of warming in the North Atlantic. Commun. Earth Environ. 3, 118 (2022).
Piron, A., Thierry, V., Mercier, H. & Caniaux, G. Argo float observations of basin-scale deep convection in the Irminger sea during winter 2011–2012. Deep Sea Res. Part Oceanogr. Res. Pap. 109, 76–90 (2016).
de Boyer Montégut, C., Madec, G., Fischer, A. S., Lazar, A. & Iudicone, D. Mixed layer depth over the global ocean: An examination of profile data and a profile-based climatology. J. Geophys. Res. Oceans 109, C12003 (2004).
Pauthenet, E. et al. Four–dimensional temperature, salinity and mixed-layer depth in the Gulf Stream, reconstructed from remote–sensing and in situ observations with neural networks. Ocean Sci. 18, 1221–1244 (2022).
Pickart, R. S., Torres, D. J. & Clarke, R. A. Hydrography of the Labrador Sea during active convection. J. Phys. Oceanogr. 32, 428–457 (2002).
Descriptive Physical Oceanography: An Introduction. (Academic Press, Amsterdam; Boston, 2011).
Thierry, V., de Boisséson, E. & Mercier, H. Interannual variability of the Subpolar Mode Water properties over the Reykjanes Ridge during 1990–2006. J. Geophys. Res. Oceans 113, C04016 (2008).
Talley, L. D. & McCartney, M. S. Distribution and circulation of Labrador Sea water. J. Phys. Oceanogr. 12, 1189–1205 (1982).
Rhein, M., Steinfeldt, R., Kieke, D., Stendardo, I. & Yashayaev, I. Ventilation variability of Labrador Sea Water and its impact on oxygen and anthropogenic carbon: a review. Philos. Trans. R. Soc. Math. Phys. Eng. Sci. 375, 20160321 (2017).
Faghmous, J. H. et al. A daily global mesoscale ocean eddy dataset from satellite altimetry. Sci. Data 2, 1–16 (2015).
Acknowledgements
R.A. has received funding, as part of the EuroSea project, from the European Union’s Horizon 2020 research and innovation program under grant agreement No. 862626. L.I.C., V.T., and R.B. acknowledge support from Ifremer. H.M. was supported by CNRS. F.F.P. was supported by the BOCATS2 (PID2019-104279GB-C21) project funded by MCIN/AEI/10.13039/501100011033. This work is a contribution to CSIC’s Thematic Interdisciplinary Platform PTI WATER:iOS. The authors gratefully acknowledge financial support by the Brittany Region for the CPER Bretagne ObsOcean 2021-2027 and from the French government within the framework of the “Investissements d’avenir” program integrated in France 2030 and managed by the Agence Nationale de la Recherche (ANR) under grant agreement no ANR-21-ESRE-0019 for the Equipex+ Argo-2030 project. The data were collected and made freely available by the International Argo Program and the national programs that contribute to it (https://argo.ucsd.edu, https://www.ocean-ops.org). The Argo Program is part of the Global Ocean Observing System. The GLODAPv2 database is freely available as a numeric data package at the Carbon Dioxide Information Analysis Center (CDIAC) http://cdiac.ornl.gov/oceans/GLODAPv2. The Ssalto/Duacs altimeter products were produced and distributed by the Copernicus Marine and Environment Monitoring Service (http://www.marine.copernicus.eu).
Author information
Authors and Affiliations
Contributions
R.A., L.I.C., and V.T. designed and developed the concept of the study. R.A. conducted the data analysis with inputs from L.I.C., V.T., H.M., R.B., and F.F.P. R.A. drafted the first version of the paper. All co-authors read and reviewed the paper, and all co-authors agreed on the final version of the paper.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Peer review
Peer review information
Nature Communications thanks Jens Müller, Lydia Keppler, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Asselot, R., Carracedo, L.I., Thierry, V. et al. Anthropogenic carbon pathways towards the North Atlantic interior revealed by Argo-O2, neural networks and back-calculations. Nat Commun 15, 1630 (2024). https://doi.org/10.1038/s41467-024-46074-5
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41467-024-46074-5
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.