Air-sea CO2 fluxes and the controls on ocean surface pCO2 seasonal variability in the coastal and open-ocean southwestern Atlantic Ocean: A modeling study

We use an eddy-resolving, regional ocean biogeochemical model to investigate the main variables and processes responsible for the climatological spatio-temporal variability of pCO2 and the air-sea CO2 fluxes in the southwestern Atlantic Ocean. Overall, the region acts as a sink of atmospheric CO2 south of 30◦S, and is close to equilibrium with the atmospheric CO2 to the north. On the shelves, the ocean acts as a weak source of CO2, except for the mid/outer shelves of Patagonia, which 5 act as sinks. In contrast, the inner shelves and the low latitude open ocean of the southwestern Atlantic represent source regions. Observed nearshore-to-offshore and meridional pCO2 gradients are well represented by our simulation. A sensitivity analysis shows the importance of the counteracting effects of temperature and dissolved inorganic carbon (DIC) in controlling the seasonal variability of pCO2. Biological production and solubility are the main processes regulating pCO2, with bio10 logical production being particularly important on the shelves. The role of mixing/stratification in modulating DIC, and therefore surface pCO2, is shown in a vertical profile at the location of the Ocean Observatories Initiative (OOI) site in the Argentine Basin (42◦S, 42◦W).


Introduction
Shelf regions are amongst the most biogeochemically dynamical zones of the marine biosphere 15 (Walsh, 1991;Bauer et al., 2013). Even though they comprise only 7 − 10% of the global ocean area (Laruelle et al., 2013), continental shelves could contribute to approximately 10 − 15% of the ocean 1 primary production and 40% of the ocean's carbon sequestration through particulate organic carbon (Muller-Karger et al., 2005). Global discussions about the role of continental margins as a sink of atmospheric CO 2 gained momentum after Tsunogai et al. (1999) suggested that these shelf regions 20 take up as much as 1 PgC/year of atmospheric CO 2 . Recent estimates range from 0.2 PgC/year (Laruelle et al., 2013) to roughly 0.6 PgC/year (Yool and Fasham, 2001), somewhat more modest than initially thought (Gruber, 2015), but still relevant to the global ocean sink estimated around 2.3 PgC/year (Ciais et al., 2014). 25 Continental shelves tend to act as a sink of carbon at high and medium latitudes (30 • − 90 • ), and as a weak source at low latitudes (0 • − 30 • ) (Chen et al., 2013;Hofmann et al., 2011;Bauer et al., 2013;Laruelle et al., 2014), i.e., they tend to follow similar meridional trends as the open ocean CO 2 fluxes (Landschützer et al., 2014;Takahashi et al., 2009). 30 However, continental shelves present a higher spatio-temporal variability of air-sea CO 2 fluxes than the adjacent open ocean, with the inner shelf and near coastal regions generally acting as a source of CO 2 to the atmosphere, while the mid/outer shelf and the continental slope generally act as sinks (Cai, 2003). This pattern can be explained by the increased primary production and decreased terrestrial supply towards the outer shelf (Walsh, 1991). Seasonality of the upper ocean (e.g. 35 mixing and stratification) may also be important to the air-sea exchange of carbon. For example, the United States southeast continental shelf acts as a sink of CO 2 in the winter and as a source in the summer (Wang et al., 2005).
In the southwestern Atlantic Ocean, the shelf region presents distinct features. To the south, the 40 Patagonian shelf is one of the world's largest shelves with an area close to 10 6 km 2 , broadening to more than 800 km from the coastline (Bianchi et al., 2009). To the north, the Brazilian shelf narrows to around 100-200 km from the coastline. This region is one of the most energetic regions of the world's ocean with the confluence of the warm southward-flowing Brazil Current (BC) and the cold Malvinas Current (MC) flowing northward (Piola and Matano, 2001). The extension of the conflu-45 ence roughly divides the subtropical and subantarctic oceanic gyres in the South Atlantic and might be a hotspot for shelf-open ocean exchange (Guerrero et al., 2014).
In the open-ocean, the South Atlantic is thought to absorb between 0.3-0.6 PgC/year south of 30 • S, while acting as a source to the atmosphere north of 30 • S (Takahashi et al., 2002). Aside from 50 global open-ocean estimates, only a few local studies were conducted on the continental shelves in this region. The Patagonia shelf was characterized as a source of CO 2 to the atmosphere on the inner shelf, and as a sink in the mid-outer shelf (Bianchi et al., 2009). The southeast Brazilian shelf and continental slope were characterized as sources of CO 2 to the atmosphere during all seasons (Ito 2 et al., 2005). Such regions are often neglected, or poorly resolved, on relatively coarse global mod-55 elling assessments, although they may contribute up to 0.2 PgC/year of global ocean CO 2 uptake (Laruelle et al., 2014).
Regional marine biogeochemical models have been used to assess the ocean carbonate system and CO 2 fluxes, including the continental margins. For example, along the US east coast, the seasonal-60 ity of pCO 2 was found to be controlled mainly by changes in the solubility of CO 2 and biological processes (Fennel and Wilkin, 2009). Along the California coast, biological production, solubility and physical transport (e.g. circulation) were found to be the most influential processes on pCO 2 variability, both spatially and temporally (Turi et al., 2014).

65
In this study we use a regional marine biogeochemical model coupled to a hydrodynamic model to investigate the parameters and processes regulating the variability of ocean surface pCO 2 in the southwestern Atlantic Ocean. Our model domain includes the location of the global node mooring that is soon to be deployed as part of the Ocean Observatories Initiative (OOI) at 42 • S, 42 • W (oceanobservatories.org). 70 We compare modeled surface pCO 2 distribution with observations and use the results to investigate the relative importance of the parameters (DIC, temperature, alkalinity and salinity) and processes (biological production, air-sea CO 2 flux, CO 2 solubility and physical transport) in controlling surface pCO 2 distribution and variability on the continental shelf and open ocean in the southwestern 75 Atlantic Ocean.

Model
The physical model used in this study is the Regional Ocean Modeling System (ROMS) (Shchep-  The biogeochemical model is an NPZD type, including the following state variables: phytoplankton, zooplankton, nitrate, ammonium, small and large detritus, and a dynamic chlorophyll to carbon ratio for the phytoplankton . A carbon component is also coupled to the model, with the addition of calcium carbonate, DIC and alkalinity to the system of state variables (Gruber 3 et al., 2011;Hauri et al., 2013;Turi et al., 2014). Parameters utilised in the biogeochemical model 90 are listed in Table 1 of Gruber et al. (2006), and the CaCO 3 parameters as in Hauri et al. (2013).
These parameters represent phytoplankton types with large nutrient requirements and relatively fast growth rates, usually large organisms . Since our domain encompasses several ecological provinces (Gonzalez-Silvera et al., 2004), we may not represent all regions equally well with only one phytoplankton functional type.

95
The initial and boundary conditions used for the physical variables were obtained from a climatology of the Simple Ocean Data Assimilation (SODA) (Carton and Giese, 2008), and for the biogechemical variables from a Community Earth System Model (CESM) climatological model product (Moore et al., 2013). The model is forced at the surface with climatological winds from 100 QuikSCAT (Risien and Chelton, 2008) and heat and freshwater surface fluxes from the Comprehensive Ocean-Atmosphere Data Set (COADS) (Da Silva et al., 1994). We used a fixed atmospheric pCO 2 of 370 µatm without CO 2 incrementation throughout the years and without seasonal variations. Starting from rest we ran the model for 8 years and used a climatology from years 5 through 8 for our analyses.

105
Even though some processes as river runoff and tides are locally relevant (i.e., la Plata River, and Patagonia shelf), we are not considering them in the present study (see conclusions section). The low salinity waters from the La Plata river are included in the climatological forcing from COADS which are "nudged" into the model. These shortcomings may effect the results in some regions, but 110 it is unlikely that they will affect the overall pCO 2 results in the wider domain.

Analysis
Ocean surface pCO 2 is the most important variable determining the air-sea CO 2 flux. This is because the variability of ocean pCO 2 is much greater than that of atmospheric pCO 2 , and variations in the 115 gas transfer coefficient are usually several times smaller than those of ocean surface pCO 2 (Takahashi et al., 2002). Seawater pCO 2 is regulated by the concentration of dissolved inorganic carbon (DIC), alkalinity (ALK), temperature (T ) and salinity (S). While T and S are controlled solely by physical factors, DIC and ALK are affected both by biological production and physical transport.
DIC concentration is also affected by air-sea CO 2 fluxes (Sarmiento and Gruber, 2006).

120
In our model, ocean surface pCO 2 is calculated through a full model implementation of the seawater inorganic carbon system, i.e., as a function of the state variables T , S, DIC, and ALK, with the dissociation constants k 1 and k 2 from Millero (1995). In order to assess the impact of different parameters on pCO 2 variability, we decompose pCO 2 with respect to T , S, DIC and ALK, follow- , 130 where the ∆'s are anomalies, either spatial or temporal, relative to a domain or an annual mean, The perturbation applied here was 0.1% of the domain mean.
In order to investigate the parameters and processes controlling pCO 2 on the continental margin, we limited our temporal analysis to three regions with depths shallower than 1000 m: the Southeast tially averaging the parameters within each region, and using the temporal anomalies (subtracting the annual mean) on Eq. 1. For the spatial analysis, we used the whole study area and then calculated in each grid cell the spatial anomalies (subtracting the domain mean of that grid cell), finally applying it to Eq. 1.

155
In order to identify the main processes responsible for the variability of surface pCO 2 , we used a progressive series of sensitivity experiments as in Turi et al. (2014), focusing on the processes of biological production, CO 2 solubility, air-sea CO 2 fluxes, and physical transport. To quantify these processes, we made three additional model runs, progressively excluding each process. In the first experiment (E1), we set the CO 2 gas exchange flux coefficient between the atmosphere and the ocean 160 to zero, inhibiting gas exchange in the surface layer. In the second experiment (E2), we started from E1 and also turned off the photosynthetically available radiation (PAR), preventing phytoplankton growth. Finally, in experiment E3, the CO 2 solubility was set to a constant value, calculated with the domain-averaged surface salinity and temperature of 34.66 and 12.33 • C, respectively, while maintaining the changes of E1 and E2. The control run minus E1 represents the impact of gas ex-165 change between ocean and atmosphere, E1 minus E2 represents the impact of biology, E2 minus E3 represents the impact of variable solubility. The last experiment (E3), in which there is no air-sea flux, no biology and constant solubility represents the impact of physical transport (Turi et al., 2014).
Given the short model integration times, the vertical gradients in the E3 simulation have not come 170 in to steady-state with the processes. So our physical transport is working on the vertical DIC gradients established by the biological pump. Since the lateral boundary conditions are the same for all experiments, these simulations are therefore only approximations of the impact of each process on pCO 2 . Further, this separation assumes a linear additionality of each process, which is clearly a strong simplification given the non-linear nature of the inorganic carbonate system (Sarmiento and 175 Gruber, 2006). The same spatial and temporal analysis described for the variables (ALK, DIC, T and FW) was also applied for the processes experiments (air-sea CO 2 flux, biology, CO 2 solubility, physical transport).

180
Model results were evaluated against data from the Surface Ocean CO 2 Atlas (SOCAT) version 2 (Bakker et al., 2013). SOCAT f CO 2 observations were converted into pCO 2 using the set of equations from Körtzinger (1999) and then compared with modeled pCO 2 to assess the overall skill of the model. Due to the paucity of in-situ observations, particularly on the continental shelves, we used monthly climatologies for the comparison. The seasonal model evaluation was made over the whole b,c,f,g). In the mid-outer shelf the ocean generally acts as a sink, while to the north the ocean is in equilibrium with the atmosphere particularly during winter.
The monthly analysis was restricted to three offshore areas (A1, A2 and A3 in Fig.4a). We compared the spatial monthly mean modeled surface pCO 2 with the monthly average of the SOCAT Modeled pCO 2 results for A1 agree very well with the observations, representing the pCO 2 evolution throughout the year with maximum values in summer (Fig.4b). All statistical indicators char-230 acterized the model with a good/reasonable skill in A1 (Table 1).

7
A2 is the region with the largest pCO 2 standard deviation from both model and observations (Fig.4c). This region is near the confluence between the warm Brazil Current and the cold Malvinas Current, generating one of the most energetic regions of the world's oceans. Moreover, this 235 region comprises the shelfbreak front, with differences in stratification, local dynamics and salinity between shelf waters and Malvinas current waters (Fig.2a). Consequently, ME was estimated as poor/bad in this region, probably due to the high pCO 2 data variability. But CF and PB were both rated as "good/reasonable" (Table 1).

240
In A3 the model consistently underestimated pCO 2 (Fig.4d). This bias is seen in the seasonal comparison and in the monthly analysis, where summer is the only season for which modeled pCO 2 is within the standard deviation of the observations. ME was estimated as poor/bad in A3, but PB and CF rated our model as reasonable and good, respectively. (Table 1). Both A2 and A3 regions are close to an area of elevated eddy kinetic energy (Fig.4a), which could explain the large standard 245 deviation and biases in these regions.
The Taylor diagram is consistent with the model efficiency (ME) estimate, showing good/reasonable results in A1, with a correlation of 0.8, and poor results in A2 and A3, with negative correlations (Fig.5). Only in A1, the correlation was found to be statistically significant. Aside from greater pCO 2 250 variability in these regions, the poor results found in A2 and A3 could also be due to the paucity of the observational data both in space and time.
Furthermore, in order to validate the baseline of our model, seasonal climatologies of modeled sea surface temperature and chlorophyll-a were compared with climatologies from AVHRR and 255 MODIS-aqua, respectively. Results and a detailed discussion of this validation are shown in the appendix.
In conclusion, our model reproduces the most important north-south and inner-outer shelf gradients seen in the pCO 2 observations. While there is clearly room for improvement, we deem this level outer Patagonia Shelf. DIC s has the highest impact on the spatial variations, being counteracted by ALK s and T (Fig.6). In contrast, the fresh water flux has a minor influence on the spatial anomalies 270 of pCO 2 , agreeing with Turi et al. (2014) and Doney et al. (2009). Even though with a smaller role, ALK s influence on pCO 2 anomalies presented absolute values higher (−100 to 100 µatm) than previous studies in other regions (Lovenduski et al., 2007;Turi et al., 2014). The higher contribution of both DIC s and ALK s to the spatial variations in pCO 2 could be explained by the more heterogeneous domain that encompasses several distinct surface water masses and frontal zones. Also, the 275 elevated contribution of ALK s could be due to our relatively high CaCO 3 to biological production ratio of 0.07.
The changes in the state variables affecting pCO 2 are ultimately being driven by physical and biogeochemical processes. We investigate the role of each of these processes in controlling the changes 280 in surface pCO 2 from our sensitivity experiments (E1, E2, E3). The most important processes affecting the spatial distribution of pCO 2 are biological production (E1 -E2) and physical transport (E3) (Fig.7). When physical transport (vertical and horizontal) is the only process altering pCO 2 , we observe an increase in pCO 2 of up to 800 µatm on the continental shelves, due to the upwelling and

pCO 2 drivers -temporal analysis
In order to identify the seasonal variability of the contribution of each parameter, we used local grid temporal anomalies over the seasonal cycle (Fig.8). DIC s and T are still the most influential parameters, with increasing importance on the continental shelves. The contribution by ALK s is relevant 310 only on continental shelves south of 32 • S, and FW have a minor influence (not shown). It is important to highlight that the magnitude of the signals seen in this analysis is one order of magnitude smaller than the previous spatial analysis. This is likely due to our large and heterogeneous domain, which results in much spatial gradients than what is modeled over the seasonal cycle. This pattern of opposing contributions of T and DIC was also found along the North American 340 east coast by Signorini et al. (2013), who attributed winter mixing and the spring-summer biological drawdown as the processes responsible for pCO 2 and DIC variability. In the offshore subtropical region (ST) the pCO 2 anomalies have higher amplitudes than in the adjacent continental shelf (SEBS), and are driven mainly by Temperature, with the other variables having minor contributions (Fig.11). In the offshore southern region (SA), DIC s controls pCO 2 variability, with T and ALK s 345 dampening pCO 2 anomalies (Fig.10), similar to the adjacent shelf (PS).
The analysis of the processes underlying this seasonal variability using our progressive sensitivity simulations shows that on all shelf regions, biological production and CO 2 solubility mostly control pCO 2 variability (Fig.10). Physical transport, although weaker than biological production, acts to 350 diminish the pCO 2 variability by counteracting the effects of biology and increasing DIC concentrations. In our case, physical transport controls pCO 2 spatially, but the temporal effects of physical transport are much weaker than in Turi et al. (2014) along the California coast. This is probably because the much stronger upwelling in that region acts to dampen the effects of biology by bringing DIC rich waters to the surface. Along western boundaries, upwelling is weaker and more localized.

355
Physical transport is therefore more related to processes that modulate vertical mixing and stratification, thereby controlling the seasonal enrichment of surface waters, and horizontal advection due to the presence of two major western boundary currents. Finally, air-sea CO 2 fluxes are only a minor contribution to the pCO 2 anomalies.

360
In conclusion, on the Patagonia Shelf (PS), the biological production is the most important contributor to pCO 2 variability, with a peak summer contribution of −80 µatm and a maximum in the winter of 70 µatm. On the South Brazilian Shelf (SBS), solubility is the most influential process (up to 90 µatm), followed by biological production and physical transport, during all seasons. On the Southeast Brazilian Shelf (SEBS), the pattern is the same as in the SBS, although with a smaller 365 magnitude and variability. Physical transport, although large in absolute contributions in the spatial analysis, has a lower contribution to pCO 2 variability in the temporal analysis.
In the subtropical region, processes that control the temporal variability of pCO 2 on the shelf and offshore are different. In the open ocean (ST) (Fig.11) pCO 2 is mainly controlled by solubility, 370 with the biological production having the least effect on pCO 2 . This contrasts with the importance of biology on mid/low latitude continental shelves (SEBS). In the subantarctic region, the processes controlling pCO 2 are similar for both the offshore region (SA) and the adjacent continental shelf (PS) (Fig.9). In this case biological production is the most important process being countered mainly by solubility, although with a smaller magnitude in the offshore region. the Brazilian shelves act as weak sources of CO 2 (0 to 1 molCm −2 yr −1 ). Nevertheless, we found variability on each continental shelf, with regions on the inner Patagonia Shelf acting as a source 400 or in equilibrium with the atmosphere (0 to 2.0 molCm −2 yr −1 ), and regions on the outer Brazilian shelves acting as sinks of CO 2 .

Vertical Structure -Case Study at Argentine OOI Site
Seasonal variations in mixing and stratification control the evolution of the mixed layer depth and 405 consequently the vertical structure of the state variables of the carbonate system. Diapycnal fluxes of DIC and DIC sinks from primary production are important processes regulating ocean surface pCO 2 (Rippeth et al., 2014). Therefore, the mixed layer depth is linked with the surface pCO 2 variability.

410
In order to understand the seasonal evolution of the upper ocean vertical distribution of the state variables in the region and how it affects surface pCO 2 , we chose the location of the Ocean Observatory Initiative (OOI) site in the Argentine Basin at 42 • S, 42 • W (Fig.1 a), as it will soon become a test-bed for the validation of biogeochemical models globally and regionally. We extracted modeled climatological vertical profiles of DIC concentration, temperature and chlorophyll-a, and compared 415 with the modeled surface pCO 2 and mixed layer depth (Fig.13).
During the entire year, this location acts in our model as a sink for atmospheric CO 2 , with modeled surface pCO 2 ranging from 280 µatm to 320 µatm. The contribution of DIC s and T are again driving surface pCO 2 anomalies. In this case DIC s is controlling the anomalies signal, being damp- waters. This affects pCO 2 much more than the decrease in temperature, resulting in positive pCO 2 anomalies. After winter, this excess of DIC is consumed by biological fixation during spring and summer, thus reducing surface pCO 2 .

430
In this study, we used a regional hydrodynamic model coupled to a biogeochemical model to investigate, in a climatological sense, the main parameters and processes that control ocean surface pCO 2 and air-sea CO 2 fluxes in the southwestern Atlantic Ocean. Modeled ocean surface pCO 2 compared well with the available in-situ data, reproducing the expected meridional and cross-shelf gradients of pCO 2 , with elevated pCO 2 in the inner shelves and at lower latitudes. Our results highlight that 435 the most important variables controlling the spatio-temporal variability of pCO 2 are T and DIC s .
These two variables have opposing effects on pCO 2 and have been shown to be the main drivers of pCO 2 both in global (Sarmiento and Gruber, 2006;Doney et al., 2009) and in other regional studies (Turi et al., 2014;Signorini et al., 2013;Lovenduski et al., 2007). Following DIC s and T , we found that ALK s is a secondarily important spatial regulator of pCO 2 , with increasing importance on the Our model does not include river inputs of carbon, which are known to be an important factor regulating pCO 2 (Bauer et al., 2013). The lack of tides may adversely affect our model results in the inner shelf of Patagonia, where tidal amplitudes reach up to 12 meters at some points (Kantha, 1995;460 Saraceno et al., 2010) and tidal fronts are known to impact oceanic pCO 2 (Bianchi et al., 2005). In future regional studies focused on the Patagonia shelf, tides and river run-off should be included.
Modeling studies such as this one depend heavily on in-situ observations, the lack of which hampers our ability to properly refine our model. This will certainly be improved by future efforts of data