The Southern Ocean as a constraint to reduce uncertainty in future ocean carbon sinks

. Earth system model (ESM) simulations exhibit large biases compares to observation-based estimates of the present ocean CO 2 sink. The inter-model spread in projections increases nearly 2-fold by the end of the 21st century and therefore contributes signiﬁcantly to the uncertainty of future climate projections. In this study, the Southern Ocean (SO) is shown to be one of the hot-spot regions for future uptake of anthropogenic CO 2 , characterized by both the solubility pump and biologically mediated carbon drawdown in the spring and summer. We show, by analyzing a suite of fully interactive ESMs simulations from the Coupled Model Intercomparison Project phase 5 (CMIP5) over the 21st century under the high-CO 2 Representative Concentration Pathway (RCP) 8.5 scenario, that the SO is the only region where the atmospheric CO 2 uptake rate continues to increase toward the end of the 21st century. Furthermore, our study discovers a strong inter-model link between the contemporary CO 2 uptake in the Southern Ocean and the projected global cumulated uptake over the 21st century. This strong correlation suggests that models with low (high) carbon uptake rate in the contemporary SO tend to simulate low (high) uptake rate in the future. Nevertheless, our analysis also shows that none of the models fully capture the observed biophysical mechanisms governing the CO 2 ﬂuxes in the SO. The inter-model spread for the contemporary CO 2 uptake in the Southern Ocean is attributed to the variations in the simulated seasonal cycle of surface p CO 2 . Two groups of model behavior have been identiﬁed. The ﬁrst one simulates anomalously strong SO carbon uptake, generally due to both too strong a net primary production and too low a surface p CO 2 in December–January. The second group simulates an opposite CO 2 ﬂux seasonal phase, which is driven mainly by the bias in the sea surface temperature variability. We show that these biases are persistent throughout the 21st century, which highlights the urgent need for a sustained and comprehensive biogeochemical monitoring system in the Southern Ocean to better constrain key processes represented in current model systems.


Introduction
Since the industrial revolution, a steady increase in anthropogenic CO 2 emissions from fossil fuel burning, cement production, and land-use change have led to an increase in atmospheric CO 2 concentration of about 43 % in 2014 relative to its preindustrial value according to the latest measurements from the Earth System Research Laboratory in Mauna Loa (www.esrl.noaa.gov).This represents the highest CO 2 concentration for at least the last 800 000 years.Increasing atmospheric CO 2 is one of the most important drivers for ongoing, and likely future, climate change, and it affects the ocean carbon reservoir.By taking up about approximately 26 % of the anthropogenic CO 2 emissions annually (Le Quéré et al., 2015), the ocean slows down the growth of the atmospheric CO 2 concentration and therefore the rate of climate change.However, the ocean carbon uptake rate will decrease in the future owing to the lowered buffer capacity of the surface waters and the potential weakening of carbon transport from the surface to the deep ocean, leading to a positive climate feedback (Arora et al., 2013;Heinze et al., 2015).
The oceanic carbon sink is mainly controlled by the physical and the biological pumps, which are both affected by the changing climate (Volk and Hoffert, 1985).The physical pump depends mainly on two processes: dissolution of CO 2 gas in seawater and transportation of dissolved inorganic carbon into the deep ocean by mixing and circulation processes.The biological pump is predominantly governed by the population of marine phytoplankton, which consumes the dissolved inorganic carbon (DIC) in the seawater to produce organic matter or soft tissue via photosynthesis.Through gravitational forcing, this organic matter sinks into the ocean interior where it is remineralized back into DIC.The biological pump is also affected by the global increase in temperature and by changes in circulation.The solubility of atmospheric CO 2 in the ocean's surface is expected to be negatively impacted by global warming since the solubility of CO 2 gas in seawater decreases with warmer temperatures (Sarmiento et al., 1998).Additionally, the oceanic circulation that links the low DIC in the ocean surface to the CO 2 -rich deep ocean could be altered in the next few decades through weaker upwelling and a slowdown in the Atlantic meridional overturning circulation (AMOC) (Rahmstorf et al., 2015).Both of these carbon pump processes are represented in the latest earth system model (ESM) simulations from CMIP5, which include for the first time a coupling between the atmosphere-ocean global climate models (AOGCMs, as in previous CMIPs) and the biogeochemical fluxes between the ocean, atmosphere, and terrestrial biosphere reservoirs (Taylor et al., 2012).
It has been shown that the trend in current anthropogenic CO 2 emissions closely follows the Representative Concentration Pathway (RCP) 8.5 scenario (Peters et al., 2013;Fuss et al., 2014), which is the most pessimistic future scenario, with high atmospheric CO 2 concentrations leading to 8.5 Wm −2 additional radiative forcing by 2100.The estimated emissions reached 37.0 ± 1.3 Gt CO 2 yr −1 in 2014 (Friedlingstein et al., 2014), matching the RCP8.5 scenario leading to the highest increase in global mean temperature from 3.2 to 5.4 • C at the end of this century relative to 1850-1900.This study focuses on analyzing the fully interactive ESM simulations from CMIP5 for the 2001 to 2099 period from the experiments "esmHistorical" and "esmrcp8.5".We compare the 2001-2010 period from the simulations with observational data.
Figure 1 presents the time series of the global annual CO 2 uptake by the ocean computed from nine different CMIP5 models (Sect.2.2) from the years 2001 to 2099, including the observation-based estimate of carbon flux for the period 1998-2011 (Landschützer et al., 2014).In addition to the large present-day inter-model spread, the figure also highlights the increase in the inter-model spread projected into the future.The magnitude of the standard deviation (i.e., of the inter-model variation) increases by a factor of 2 from 2001 (±0.3 Pg C yr −1 ) to 2099 (±0.6 Pg C yr −1 ).During this period, the projected cumulative oceanic carbon sink ranges from 340.4 to 488.5 Pg C. The 148 Pg C difference in the size of the ocean carbon sink translates into roughly a 70 ppm dif- ference in atmospheric CO 2 concentration by the end of the 21st century.In order to improve the fidelity of future projections provided by the climate modeling community, it is necessary to identify and attribute the mechanisms responsible for the growth in the inter-model spread of ocean CO 2 uptake and to determine methods to constrain this.Motivated by these growing uncertainties and the need to constrain them, our study focuses on analyzing relationships between regional and global uncertainties in ocean carbon uptakes as simulated in CMIP5 models.This is necessary because the strength and variability of the ocean carbon sinks vary considerably from one region to another and are attributed to region-specific mechanisms.For instance, in the equatorial Pacific, the long-term trend in CO 2 uptake is strongly influenced by the El Niño variability (Feely et al., 2006).In other regions, such as the Southern Ocean, the variability is related to the Southern Annular Mode (see, e.g., Le Quéré et al., 2007;Landschützer et al., 2015).
The need to reduce this inter-model spread is imperative to reduce uncertainty in future climate projections and enable policy makers to make the most informed decisions.The simulated uncertainty in ocean carbon uptake could arise from different factors.Feedbacks from ocean and terrestrial biospheres on the CO 2 concentration are expected but highly uncertain and thus difficult to predict (Denman et al., 2007).Differences in (i) basin-scale ocean evolution of ocean carbon uptake rates, (ii) timing and amplitude of physical and biogeochemical processes driving the regional pCO 2 seasonal cycle, as well as (iii) responses to transient future climate change are among potential contributors to future uncertainty in ocean carbon uptake, and addressing these points will be the main focus of this study.
The paper is organized as follows: in the next section we describe the observations and models used in this study as well as the terms and metrics used to investigate the relationships between present-day and future carbon uptake and the regional boundaries.Section 3 discusses the results of the analyses.Additional discussions and comparison with previous studies are presented in Sect. 4. Finally, the study is summarized in Sect. 5.

Observation-based estimates
We used the monthly data set documented by Landschützer et al. (2014) for the 2001-2010 period.It includes the surface ocean partial pressure of CO 2 (pCO 2 ) and the sea-air CO 2 flux (fgCO 2 ) gridded with a 1 • × 1 • horizontal resolution corresponding to 360 by 180 points in longitude and latitude, respectively.The pCO 2 data set is originally extrapolated in space and time from SOCAT (Surface Ocean CO 2 Atlas) version 2 (Bakker et al., 2014) by a two-step neuralnetwork approach as described in Landschützer et al. (2013).The sea-air CO 2 flux is computed based on this pCO 2 field, applying a standard bulk formulation and high-resolution Cross-Calibrated Multi-Platform (CCMP) wind speeds (Atlas et al., 2011).The monthly averaged sea surface salinity (SSS) was downloaded from the Simple Ocean Data Assimilation (SODA; Carton and Giese, 2008) and has been regridded to the pCO 2 data set.The sea surface temperature (SST) is from the National Oceanic and Atmospheric Administration (NOAA) Optimum Interpolation (OI) sea surface v.2 (Reynolds et al., 2002).
The climatology of net primary production (NPP) used for the seasonal model-data assessment in the Southern Ocean (Sect.3.2) is documented by Nevison et al. (2015) and is computed over the 1997-2010 period using data derived from the Sea-viewing Wide Field-of-view Sensor (SeaW-iFS).It uses an empirical chlorophyll (Chl) algorithm for the Southern Hemisphere that was tuned to in situ Chl in the Southern Ocean and spatially blended with the standard Sea-WiFS Ocean Color (OC) 4 algorithm (Kahru and Mitchell, 2010).

pCO 2 decomposition
In order to allow decomposition of pCO 2 variability into its physical and biogeochemical components, we estimated the alkalinity (ALK) and DIC at the same resolution as the pCO 2 data.The alkalinity was computed from the SST and SSS estimates depending on the region using the Lee et al. (2006) formulation.When temperatures are out of range in the selected region, the computation returns a missing value (NaN).The DIC was computed using the CO 2 inorganic carbon chemistry program CO2SYS developed in Matlab (van Heuven et al., 2011) using the gridded SST, SSS, pCO 2 , and alkalinity as input parameters.The global average surface silicate and the phosphate concentrations were used: 10 and 0.75 µmol kg S W −1 , respectively.We expect that this choice has a relatively small influence on our results since a shift in the concentration of silicate and phosphate by 4 and more than 6 times to their original mean values would generate only a 0.05 and 0.2 % change in the DIC computation.To complete the CO2SYS input, we applied the dissociation constants K1 and K2 introduced by Mehrbach et al. (1973) and refitted by Dickson and Millero (1987).
Decomposition of the total pCO 2 seasonal variability was based on the following approximation (Tjiputra et al., 2014): where the pCO tot 2 variation in time (dpCO tot 2 / dt) is approximately equal to the sum of the four decomposed pCO x 2 variations in time, where x is DIC, ALK, SST, or SSS.The pCO x 2 terms represent a set of thermodynamic equations that relates to the inorganic carbon species, taking into account variation in x, while the other components are kept at their long-term local average values.In this way, dpCO DIC 2 / dt is an estimate of the temporal variability of the local pCO 2 field as a result of changing DIC only.The same estimates were applied for the other three parameters (i.e., ALK, SST, and SSS).

Model descriptions and post processing
The nine participating CMIP5 ESMs in alphabetical order, are the (1) Beijing Climate Center Climate System Model (BCC-CSM1.1(m)),(2) Canadian Centre for Climate Modelling and Analysis ESM (CanESM2), (3) Community Earth System Model (CESM1-BGC), (4) Geophysical Fluid Dynamics Laboratory ESM (GFDL-ESM2G), (5) Hadley Global Environment Model 2 (HadGEM2-ES), (6) Japan Agency for Marine-Earth Science and Technology (MIROC-ESM), (7) Max Planck Institute for Meteorology ESM (MPI-ESM-LR), (8) Japanese Meteorology Research Institute ESM (MRI-ESM1), and (9) Norwegian Climate Centre ESM (NorESM1-ME).These models have also contributed to the last Intergovernmental Panel on Climate Change Assessment Report (IPCC-AR5).All outputs were downloaded directly from the Earth System Grid Federation (ESGF; http: //esgf.llnl.gov),and we analyzed the fully interactive CO 2 emissions-based "esm" simulations.These esm simulations take into account carbon fluxes between the land-atmosphere and ocean-atmosphere interfaces to prognostically simulate the atmospheric CO 2 concentration; thus, they include more realistic spatially varying atmospheric CO 2 concentration.The selection of these models is based on the availability of www.earth-syst-dynam.net/7/295/2016/all variables necessary to discuss the impact of the pCO 2 seasonal cycle on the carbon uptake: fgCO 2 , pCO 2 , SST, SSS, DIC, ALK, and NPP.However, the BCC-CSM1.1 and the MIROC-ESM models do not provide some of these variables (ALK and DIC) and therefore were only analyzed for the contemporary and future uptake relationship (Sect.3.1).In order to compare the global and regional fgCO 2 between the models presented in Sect.3.1, fgCO 2 outputs were also interpolated to the observational grid of 360 × 180 points.Model outputs from the historical (esmHistorical) experiments were added to the RCP8.5 (esmrcp8.5)to complete the 2001-2099 period of study.We used the same "r1i1p1" realization from each model.Tables 1 and 2 summarize the physical and marine biogeochemical components and features of each model.
The marine primary productivity is one of the key components that governs the carbon cycle in the ocean, impacting the oceanic pump through alteration of the buffering capacity and the CO 2 remaining in the atmosphere.The ocean NPP is controlled by nutrient availability and other physical factors such as temperature and light.As presented in Table 2, the CMIP5 models use different representations of multiple nutrient limitations, varying from one to five explicit nutrients in CMOC-NPZD and BEC-TOPAZ2 and from one to three phytoplankton species, NPZD-HAMOCC-CMOC and BEC-TOPAZ2.This highlights the wide range of biogeochemistry complexity, which can also contribute to the inter-model spread in their respective outputs.We note that the MOM4-L40 uses the OCMIP2 biogeochemistry module, which does not include an explicit marine ecosystem; therefore, in this case the primary production is simulated only as a function of surface phosphate concentration.

Uptake efficiency
In Sect.3.2 we compute for each model (at the original model resolution) the "uptake efficiency" (uptake y eff ) of carbon in the ocean, where y represents the different basin regions as defined in Sect.2.6 as well as the global (glb) and other ocean region excluding the SO (eSO).The uptake eff measures the efficiency of a specific water mass in taking up carbon for a given change in atmospheric pCO 2 .A high uptake eff value represents a good capacity of the ocean to contain DIC for a certain change in atmospheric pCO 2 and vice versa.This term is computed as follows: uptake (2) The DIC y and pCO y 2 in Eq. (2) represent the respective area-weighted mean surface concentration of DIC and pCO y 2 within the domain y.RF y is the regional mean Revelle Factor (RF) (Revelle and Suess, 1957), also known as the inverse of the ocean's buffering capacity for atmospheric CO 2 uptake, i.e., to convert CO 2,aq into different carbon species (carbonate and bicarbonate) within domain y.Water masses with a lower Revelle Factor are more efficient at taking up anthropogenic carbon (Sabine et al., 2004).The increase in atmospheric CO 2 has pushed the surface CO 2,aq concentration to a higher level, resulting in an increase in the Revelle Factor (Zeebe and Wolf-Gladrow, 2001) and thus a decrease in ocean's buffering capacity.This mechanism represents a positive climate feedback, which reduces the uptake rate of atmospheric CO 2 in the future (Wallace, 2001).The DIC and pCO 2 fields were taken directly from the model outputs, whereas the RF were computed with CO2SYS.

Inter-model correlation
In order to assess qualitatively the existence of any patterns or consistencies between the simulated CO 2 uptakes among the different models, two metrics of inter-model correlation coefficients have been computed: R y mean and R y cum .A high correlation coefficient indicates a strong relationship between the contemporary and the projected inter-model spread.Moreover, a statistically significant positive correlation coefficient denotes that models that project weak (high) uptake in the contemporary period tend to project weak (high) future uptake.The two correlation coefficients were computed as follows: (4) R y mean represents the inter-model correlation coefficient between contemporary annual mean CO 2 uptake in the different y regions and global uptake rate in the last decade of the 21st century (i.e., 2090-2099).Then, R y cum represents the inter-model correlation coefficient between the contemporary annual mean uptake rate and the cumulative global carbon uptake over the 2001-2099 period.

Regional boundaries
Regional characteristics of anthropogenic-carbon uptake can be assessed through division of the global ocean into eight basin-scale regions.The regional distribution is defined according to the low, mid-and high latitudes, motivated by the large-scale difference in carbon uptake mechanisms occurring in these regions (see, e.g., Mikaloff Fletcher et al., 2007).The regions are as follows: Southern Ocean (SO; 45-70 • S), midlatitude Southern Ocean (mSO; 15-45 • S), tropical Pacific (TPa; 15 • N-15 • S), tropical Atlantic (TAt; 15 • N-15 • S), Indian Ocean (Ind; north of 15 • S), North Pacific (NPa; 15-60 • N), North Atlantic (NAt; 15-60 • N), and Arctic Ocean (Arc; > 60 • N).Computation for the analysis pertaining to the relationships between contemporary and future CO 2 uptake (Sect.3.1) is made at the resolution of the observational data.We note that the selection of 45 • S as a boundary between the mid-and high-latitude SO could pose problems since the SO region has sophisticated dynamics, and, dependent on the models, the 45 • S latitude could cut into regions of dominant carbon sources or sinks.To address this issue, we also perform additional analyses where we use a dynamic boundary separating the mid-and high-latitude Southern Ocean applying a surface density of 26.5 kg m −3 .For instance, Séférian et al. (2012) apply this density line to separate the subtropical mode water (TMW; region of weak increase in future CO 2 uptake) and the subantarctic model water (MW; region of strong increase in future CO 2 uptake).

Inter-model contemporary and future CO 2 uptake relationships
The relationship between contemporary and future global CO 2 uptake (R glb mean ) as simulated by the CMIP5 models is shown in the Fig. 2a, and the relationships relating to R glb cum are shown in Fig. 2b. Figure 2a shows that the models have positive correlation but weak linear relationships between the present and future CO 2 uptake rate.However, the linear relationships become more pronounced for cumulative carbon sinks (Fig. 2b), as shown by a R glb cum value of 0.77.This is also consistent with Fig. 1, which shows that the inter-model spread in CO 2 uptake evolves in a relatively similar manner into the future.
Next, for each region defined in Sect.2.6 (as depicted in Fig. 3a), we computed the inter-model correlation coefficient metrics following Eqs.(3) and (4).The SO region yields the highest correlation coefficient as depicted in Fig. 3 with R SO cum = 0.65 and R SO mean = 0.76, in blue and red bars, respectively.Other regions remain weakly correlated with a correlation coefficient close to zero or under 0.40 for both correlation fields (TPa, TAt, Ind, NPa, NAt, and Arc regions).Only the mSO region reveals quite a strong negative correlation coefficient, with R mSO mean = −0.55.Nevertheless, in all regions except the SO, the correlation coefficients are statistically insignificant at the 90 % confidence level, while SO shows a statistical significance of over 99 %.For the remain-  der of the analysis, we therefore combined the seven regions with an insignificant correlation coefficient into one new region: global ocean excluding SO (eSO).Globally, for the annual oceanic CO 2 uptake during the 2001-2010 period, all models except the NorESM1-ME are well within the range of the observation-derived estimate from Landschützer et al. (2014) of about 1.99 ± 0.50 Pg C yr −1 .The two highest uptake estimates are simulated by the BCC-CSM1.1(m)and NorESM1-ME models (2.47±0.15 and 2.63±0.07Pg C yr −1 , respectively), and the lowest uptake estimates come from the MPI-ESM-LR and MRI-ESM1 models (1.88 ± 0.10 and 1.78 ± 0.13 Pg C yr −1 , respectively).
Figure 4 depicts the inter-model relationships in the SO and eSO domains for R SO mean , R eSO mean , R SO cum , and R eSO cum .In SO, four models (BCC-CSM1.1,NorESM1-ME, MPI-ESM-LR, and CESM1-BGC) overestimate the carbon uptake flux from the atmosphere to the ocean when compared with two independent observationally based estimates of about 0.15 ± 0.12 Pg C yr −1 (Landschützer et al., 2014) and 0.27 ± 0.13 Pg C yr −1 (Lenton et al., 2013); the latter is derived from Takahashi et al. (2009) data sets.The highest estimates are simulated by the BCC-CSM1 and NorESM1-ME models at 1.03 ± 0.09 and 0.64 ± 0.11 Pg C yr −1 , respectively.Two models (CanESM2 and MRI-ESM1) underestimate the flux, with the lowest estimate simulated by the CanESM2 model, which is also the only model to simulate the Southern Ocean as a source of carbon to the atmosphere at about −0.64±0.09Pg C yr −1 (Fig. 4a).Three other models (GFDL-ESM2G, MIROC-ESM, and HadGEM2-ES) project CO 2 uptake within the two observationally based estimates.
Figure 4a and c illustrate the strong inter-model linear relationships between the contemporary CO 2 uptake rate in the SO and the projected future uptake rate (Fig. 4a) and cumulated carbon uptake over the 21st century (Fig. 4c).In the last decade of the 21st century, the CMIP5 models project ocean carbon uptake rates ranging from 4.30 to 5.92 Pg C yr −1 .The cumulative oceanic CO 2 uptake during the 21st century is projected to be between 340.4 and 488.5 Pg C.
Figure 4 also demonstrates the peculiarity of the Southern Ocean as compared to the rest of the world's ocean (eSO), where the contemporary CO 2 uptake rate in the latter has a relatively strong negative correlation with R eSO mean = −0.52 and R eSO cum = −0.24.To investigate the robustness of this cor- relation coefficient with time, we computed R y cum for all of the original eight regional basins for all 10-year windows between 2001 and 2099 (shown in Fig. 4e).The SO region is shown to have a consistently strong positive correlation coefficient across time with R SO cum values of 0.69 ± 0.04.Other regions have more pronounced temporal variations, particularly the North Atlantic (NAt), which goes from a negative correlation in the early 21st century (−0.39) to a positive correlation after 2020; the latter increases to 0.59.However, the correlation in the NAt remains statistically insignificant at 90 % of confidence level.
As stated in Sect.2.6, we also computed the correlation coefficient metrics for the SO region using a dynamic boundary (instead of a fixed 45 • S latitude) along the surface water with a density of 26.5 kg m −3 .Figure 5c illustrates the model-dependent dynamic boundaries as simulated for August 2005.Figure 5a and b show that the linear inter-model relationships remain strong (correlation coefficient of at least 0.76) when the dynamic boundary is used, suggesting that the inter-model relationships in the SO are relatively robust.

Carbon uptake evolution in the Southern Ocean
In this section, we examine why the SO has the highest R SO mean and R SO cum relative to the other regions.Only seven out of the nine models previously used to establish the correlations are used; the BCC-CSM1 and MIROC-ESM models are excluded because they do not provide the monthly ALK and DIC fields needed for the uptake eff analysis.The remaining models are CanESM2, CESM1-BGC, GFDL-ESM2G, HadGEM2-ES, MPI-ESM-LR, MRI-ESM1, and NorESM1- ME. Figure 6 shows the time series anomalies (relative to the year 2001) of CO 2 uptake, net primary production (NPP), and uptake efficiency (uptake eff ) in SO and eSO as simulated by the CMIP5 models.
There is a general increase in CO 2 uptake for both SO and eSO, as would be expected from the increasing atmospheric CO 2 concentrations under the RCP8.5 scenario.However, in SO (except for the CESM1-BGC model) the simulated uptake rates steadily increase towards the end of the 21st century, and the multi-model mean increases to 1.2 ± 0.3 Pg C yr −1 higher than in the present day (Fig. 6a).The CESM1-BGC model simulates stabilization of CO 2 uptake during the last 2 decades of the 21st century.In the other regions (Fig. 6b), the multi-model mean reaches a saturation point of 1.9 ± 0.4 Pg C yr −1 in the 2070s before the uptake strengths go down to 1.5 ± 0.4 Pg C yr −1 in 2100.This "peak and decline" pattern is consistently shown in all models analyzed here.
The unique SO region benefits from the strong link between deep and surface ocean through the southern upwelling (Sallée et al., 2013a).Earlier studies analyzing the previous generation of ESMs also demonstrated that this region will be an important sink of future atmospheric CO 2 although the efficiency of the sink may decrease (Roy et al., 2011).The increasing CO 2 sink in the SO was shown to be associated with a reduction in the fractional ice coverage which alleviates the light limitation on photosynthesis and increases in surface ocean temperature, both of which would increase the phytoplankton growing season.
A global mean decrease in NPP of about −3.12 ± 3.54 Pg C yr −1 is projected by the CMIP5 models, predominantly attributed to the increase in surface temperature leading to stronger stratification and hence reducing the nutrient supply to the surface ocean through vertical mixing (Bopp et al., 2013).The large differences between the structure of the ecosystem models of the CMIP5 models no doubt contribute to the large inter-model uncertainty.For example, the GFDL-ESM2G and MRI-ESM1 models simulate global annual NPP estimates which differ by more than a factor of 2 at 66.7 and 25.9 Pg C yr −1 , respectively (not shown), and are outside of the multi-model standard deviation for the NPP estimates.The MRI-ESM1 model considers only one nutrient limitation and simulates only one type of phytoplankton, while the GFDL-ESM2G model uses a more sophisticated ecosystem module with five types of nutrients and three classes of phytoplankton (Table 2).However, this alone is insufficient to determine the reason why the GFDL-ESM2G NPP is so strong.is at the low end of the inter-model range.Inter-model variations in the physical and biogeochemical interaction important for the surface primary productivity, such as irradiance and upwelling, should be analyzed further to seek to address this question.This, however, is beyond the scope of the present study.
Figure 6c-d show the NPP anomalies relative to 2001 in the SO and eSO regions through the 21st century, highlighting that surface primary production in SO is either stable or weakly increasing by roughly 0.5±0.3Pg C yr −1 at the end of this century, while it is clearly decreasing in the other regions (by −2.6 ± 0.1 Pg C yr −1 ).This is consistent with findings by Laufkötter et al. (2015), who show that the NPP increase in the SO is predominantly attributed to the weakening temperature limitation for phytoplankton growth projected in the future.They also indicate that, despite the inter-model agree-ment in long-term trend, the regional inter-model variation is substantial.
In SO, steady biological production may also be responsible for maintaining low pCO 2 in the summer and keeping a higher buffer capacity than in the other regions (Hauck and Völker, 2015).The two last panels of Fig. 6 show the anomaly of uptake eff (Sect 2.4 for definition), which is expected to decrease as the Revelle Factor increases under future high ambient atmospheric CO 2 concentrations (Heinze et al., 2015).Nevertheless, the decreasing trend is weaker in the SO than the eSO region at −0.28 ± 0.01 vs. −0.37 ± 0.02 µmol kg −1 ppm −1 respectively by the year 2099.Indeed, for the same change in pCO 2 and roughly the same Revelle Factor change, the SO experiences a smaller change in DIC (not shown), indicating a unique process occurring in the high-latitude SO.Deep winter mixing at polar regions is very efficient in transporting the anthropogenic carbon from  2014) for CO 2 uptake and SST anomaly, and the NPP was taken from SeaWiFS data as described in Nevison et al. (2015).
surface to depth, resulting in an increase in uptake efficiency (Marinov et al., 2007).Moreover, the increase in the meridional temperature gradient from the tropic, to the high latitudes projected in the future could lead to an enhancement of the Antarctic Circumpolar Current via a stronger wind stress at the surface (Gillett and Fyfe, 2013).As a result, this could translate into enhancement in intermediate water formation and more efficient transport of anthropogenic carbon from the surface into depth.
The steady increase in carbon uptake in the SO region could be the reason for the strong correlation of this region with future global sinks (Sect.3.1 and Fig. 4a).In the next subsections, we therefore focus on analyzing the mechanism for ocean uptake in the SO region as simulated by the different models.

Inter-model division in the Southern Ocean carbon uptake
Figure 4a-b show that the CMIP5 models exhibit diverse contemporary carbon uptake sinks in SO, from an outgassing of −0.64±0.09Pg C yr −1 (CanESM2) to an uptake of 1.03± 0.09 Pg C yr −1 (BCC-CSM1.1).To investigate the mechanisms driving the inter-model heterogeneity, we compute the carbon uptake on a seasonal timescale and compare it to estimates derived from observations (Landschützer et al., 2014).We divided the seven CMIP5 models into two groups.The first group (hereafter referred to as "G1") represents those that simulate an anomalously stronger annual CO 2 uptake rate in the SO as compared to the observationally based estimates and consists of the CESM1-BGC, HadGEM2-ES, MPI-ESM-LR, and NorESM1-ME models.The second group (hereafter referred as "G2") comprises models that simulate anomalously weaker CO 2 uptake, consisting of the CanESM2, GFDL-ESM2G, and MRI-ESM1 models.Figure 7 shows the inter-model mean and spread of these two Earth Syst.Dynam., 7, 295-312, 2016 www.earth-syst-dynam.net/7/295/2016/ A. Kessler and J. Tjiputra: The Southern Ocean as a constraint to reduce uncertainty in future ocean carbon sinks 305 groups in their projections of the seasonal cycle of carbon fluxes, NPP, and anomalies of SST for both the contemporary (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) and future (2090-2099) periods.Figure 7 illustrates that the G1 models have nearly the opposite seasonal cycle to the G2 models.The G1 models (Fig. 7a, red lines) simulate a strong mean ocean CO 2 uptake in December-January of about 0.30 mol C m −2 month −1 , which has the same direction as, but is more than 4 times stronger in magnitude than, the observation-based estimate of 0.07 mol C m −2 month −1 .This overestimation corresponds to the period of the highest NPP, where the G1 model means simulate an NPP maximum more than 3 times stronger than the observations: 35 compared to 11 g C m −2 month −1 , respectively (Fig. 7c, black and red lines).Moreover, the SST anomaly during this 2-month period has a negligible effect on the CO 2 flux; there is no significant change in carbon flux occurring when the SST anomaly increases from +0.3 to +1.5 • C (Fig. 7e, red line).This highlights that the biological activity in G1 models is the primary driver for the CO 2 flux seasonal cycle in SO during the highproductivity season, while the impact of the seasonal temperature on the surface pCO 2 appears to play only a secondary role (Takahashi et al., 2002).
In contrast to the G1 models, G2 models (Fig. 7, blue lines) simulate strong outgassing during the summertime with a negative CO 2 flux of nearly −0.10 mol C m −2 month −1 .This is in disagreement with the observation-based estimates and is predominantly driven by the SST changes.The magnitude of the SST anomaly from the G2 models is 2 times stronger than the observations, whereas the NPP cycle is similar in amplitude.The rapid warming and cooling of SST simulated in the G2 models during the spring and fall seasons lead to a higher and lower surface pCO 2 , respectively.As a result, the G2 models simulate strong CO 2 uptake during the fall season, which also implies that the solubility pump is a primary driver.
During the austral winter, in August, observationbased estimates show a maximum outgassing, roughly −0.03 mol C m −2 month −1 (Fig. 7a, black line).G1 models simulate the same mechanism at this period but the maximum outgassing is reached earlier in May-June instead of August.Concurrently, G1 models simulate a minimum of NPP at this time, pushing up the pCO 2 at the surface.Thus, the SO turns into a source of CO 2 for the atmosphere despite the SST anomaly of about −0.6 • C, which would tend to push the pCO 2 in the opposite direction.The same shift appears for G2 models but for an opposite CO 2 flux seasonal cycle, with a maximum uptake in May of nearly 0.10 mol C m −2 month −1 .The NPP is twice as strong in May-June for the G2 models than for the G1 models at 1.5 vs. 0.6 g C m −2 month −1 , respectively.This, in addition to a stronger magnitude of SST anomaly, leads to a CO 2 uptake in April-June being projected by the G2 models as depicted in Fig. 7a (blue line).
Figure 8 illustrates the mean spatial variation in CO 2 uptake for each season from Landschützer et al. (2014) and as simulated by the G1 and G2 models during the contemporary period.In the G1 models, the seasonal cycle that is too strong as compared to the observation-based estimates occurs throughout most parts of the SO, especially during October-December and January-March, with considerably stronger carbon sinks found in the south of the circumpolar current.Between 50 and 60 • S, the outgassing in G1 models is noticeably stronger during April-June and July-September.In the G2 models, the largest source of bias during October-December and January-March when compared to the Landschützer et al. (2014) estimates is found in the Atlantic and Indian sectors of the SO where the models simulate a relatively uniformly strong outgassing.
Future period simulations (Fig. 7, right panels) show that the seasonal phase in the carbon flux will be relatively similar, but the amplitude will grow considerably as compared, to the current seasonality.The distinctions in NPP and SST seasonal cycle between G1 and G2 models are also maintained.Therefore, the bias in the present-day seasonal phase of CO 2 fluxes is projected to persist toward the end of the 21st century.We note that there is a 1-month shift in simulated SST anomaly seasonal cycle where a maximum SST anomaly appears in March instead of February.

Drivers for the Southern Ocean carbon uptake
Following Eq. (1) in the methods section, we decomposed the pCO 2 seasonal cycle anomalies into four drivers: DIC, ALK, SST, and SSS for both the contemporary and future periods (shown in Fig. 9).The SSS-induced variations are not shown because the magnitude is negligible relative to the other variables.As with the previous subsection, we focused our analysis on the two contrasting model groups (G1 and G2).The amplitude of pCO 2 from the G1 (Fig. 9a, red line) overestimates the Landschützer et al. (2014) estimations.Nevertheless, it fits closer with later measurements from Merlivat et al. (2015), though they estimated the sea-air carbon flux only from a section of the Southern Ocean where pCO 2 anomaly amplitudes can reach roughly 120 µatm.In both groups as well as the observation-based estimates, the SST-and ALKinduced pCO 2 anomalies are generally in phase with each other and they are the opposite of the DIC-induced variability.
The DIC-induced pCO 2 anomaly from the G1 models simulates a minimum in January instead of March, a 2-month shift as compared to the values derived from Landschützer et al. (2014).The G1 models, which simulate anomalously strong SO carbon uptake, generally simulate too low a surface pCO 2 during December-January (an anomaly of −40 µatm; Fig. 9a, red line) due to the NPP that is too strong (Fig. 7c, red line).The driving parameter seems to be the DIC at the water surface (−38 to −64 µatm of anomalies in December-January).Thus, DIC consumption by the phyto- plankton via photosynthesis confirms the importance of biological activity for carbon uptake in this group of models during this high-productivity period.The G2 models project nearly the same amplitude as that derived from observations but depict an opposite phase for carbon uptake.Figure 9 shows that G2 models simulate surface pCO 2 during December-January that is, anomalously, too strong, of roughly 7 to 12 µatm (Fig. 9c, blue line).There, the driving parameters are the SST and alkalinity with anomalies of about 38 and 19 µatm in January, respectively.Indeed, these two components tend to push up the pCO 2 at the surface in the summer season (December-March) and present also a 1-month shift as compared to the values derived from observations (Fig. 9c, gray squares and circles).
The future simulations accentuate even more the pCO 2 seasonal cycles for the G1 models (Fig. 9b, red line).The amplitude of this seasonal cycle approximately doubles in 2090-2099 relative to 2001-2010 (i.e., the standard deviation increases from 24.5 to 50.3 µatm), mostly due to the DICinduced variability (Fig. 9b, red triangles).The amplitudes of the SST-induced and ALK-induced variability are projected to double as well; however, their combined magnitude is still weaker than DIC-induced variability.The G2 models maintain roughly the same amplitude of total pCO 2 through the 21st century (Fig. 9d, blue line), though the pCO 2 -induced components also increase by a factor of 2. Nevertheless, the DIC-induced seasonal cycle of the G2 models, conversely to the G1, is of about the same order of magnitude as the SST-induced seasonal cycle, thus balancing the change in the pCO 2 .

Discussion
The ocean plays an instrumental role in buffering the increasing atmospheric CO 2 concentration and the ongoing climate change.In this study, for the first time, we evaluate the relationships between present-day regional ocean carbon sinks with future cumulative carbon sinks over the 21st century under the high CO 2 RCP8.5 scenario as simulated by a suite of fully interactive CMIP5 ESMs.The SO is found to be a good predictor for future global carbon uptake.We therefore examined the representation of oceanic carbon uptake and its future evolution in the SO.Specifically, we assess the model capability to simulate the observed seasonal pCO 2 cycle for the present-day period.
With respect to the annual mean CO 2 uptake in the Southern Ocean, Jiang et al. (2014) evaluate a set of CMIP5 models but from different simulations, i.e., with prescribed atmospheric CO 2 concentrations, and over a slightly smaller domain (56-62 • S).Despite these differences, our present findings are very much comparable to the prior study, with the CanESM2 and GFDL-ESM2G models simulating net outgassing and close to neutral relatively stronger carbon sinks, especially during austral summer.For the 2001-2010 period, only one model (MRI-ESM1, ∼ 362 µatm) simulates a mean surface pCO 2 in the SO that is lower than the observed ∼ 364 µatm; the majority of models simulate a stronger mean carbon sink than the observational estimate.This indicates the need to consider the seasonal cycle when evaluating carbon uptake projections in the SO.Jiang et al. (2014) also examine the CMIP5 models' simulated pCO 2 seasonal cycle in the Drake Passage as compared to shipboard measurements.They show that the pCO 2 seasonal cycle in this confined domain is representative of the broader circumpolar region.Our analysis using Landschützer et al. (2014) data further indicates that the pCO 2 seasonal cycle in the SO region is reasonably homogeneous in phase, with the maximum pCO 2 in austral winter (August) and the minimum in summer (January), as illustrated in Figs.7 and 9.We show that most models simulate a larger seasonal pCO 2 amplitude than the observation-based estimates.Despite overestimating the mean carbon sinks, both HadGEM2-ES and MPI-ESM-LR reproduce the seasonal phase of the data.The MPI-ESM-LR simulates carbon uptake in the early spring period that is too strongly biologically mediated, consistent with the anomalously strong late winter mixing (August-September) (Jiang et al., 2014;Sallée et al., 2013b), which causes the required nutrients to upwell to fuel biological production.The GFDL-ESM2G, one of the two models that simulate carbon sinks closest to the data-based estimate, simulates the opposite pCO 2 seasonal phase, which is largely attributed to the bias in absolute values and amplitude of SST seasonal cycle.None of the CMIP5 models analyzed here are able to reproduce the observed seasonal cycle and annual mean carbon sinks within the uncertainty range in the SO.This highlights the difficulty in simulating the correct variability in hydrography and biogeochemistry in this region.
Based on the linear inter-model relationship presented in this study, the GFDL-ESM2G, MIROC-ESM, and HadGEM2-ES models simulate contemporary CO 2 fluxes in the SO closest to the observationally based estimate (see, for example, Fig. 5a) and therefore are likely to have more credibility with regard to their future projections.Nevertheless, from our seasonal cycle analysis, it is not clear if these models simulate the observed mechanisms governing the CO 2 fluxes.According to Landschützer et al. (2015), the nonthermal component of the pCO 2 variation is an important driver for the long-term CO 2 fluxes in the SO. Figure 10 shows the seasonal anomaly of a nonthermal CO 2 seasonal cycle in the SO from models and observation-based estimate.The CanESM2 and GFDL-ESM2G simulate comparable amplitude and seasonal phase with the observation-based estimate, but the former model has anomalously high surface pCO 2 (i.e., it simulates a net source of CO 2 to the atmosphere in the SO).Taking this as an additional constraint, our analysis suggests that the GFDL-ESM2G performs best in capturing the observed CO 2 fluxes in the Southern Ocean.
In their model study applying an ad hoc parameterization of wind stirring, Rodgers et al. (2014) demonstrate that changes in wind stirring have a large impact on the mean carbon uptake and seasonal cycle phasing in the SO (south of 45 • S).They show that resultant changes in the seasonal onset of stratification influence both entrainment and the biological pump.Furthermore, Sallée et al. (2013b) identify the annual mean freshwater flux as the primary source of error for the SO mixed layer depth in CMIP5 models.This uncertainty arises from the lack of accurate estimates of buoyancy fluxes from observations in the region.
In the SO, the CO 2 flux and its evolution in response to climate change also depend critically on the spatial and temporal variation of convection processes (see, e.g., Sallée et al., 2012).Due to the coarse spatial resolution in CMIP5 models, convection processes along the continental margin that form the AABW (Antarctic Bottom Water) are not well reproduced (Heuzé et al., 2013).Similarly, Bernardello et al. (2014) suggest that the anthropogenic CO 2 uptake in the Weddell Sea is closely linked to the size and timing of deepwater convection.It remains to be investigated how these un-certainties contribute to the inter-model spread of the projected CO 2 uptake in the SO shown here, especially with the next round of CMIP6, which includes models with a higher resolution According to the analysis preformed in this study, improving the representation of amplitude and seasonal phase of contemporary surface pCO 2 in SO has the potential to reduce the uncertainty of the future ocean carbon uptake in CMIP5 models.Bias in amplitude is identified to be associated with the magnitude of primary production in the spring-summer seasons, whereas bias in the seasonal phase is attributed by poor representation of SST seasonal cycle.Seasonally varying surface primary production data along with relevant biogeochemical and ecosystem state variables (e.g., nutrients and oxygen) would help constrain process parameterization in the model.In order to improve the SST simulation, improvements in the representation of physical processes across the air-sea interface and between mixed layer and ocean interior supported by high-quality observation would be needed.
Despite a steady increase in surface pCO 2 observations in the SO region over recent decades, it remains markedly undersampled, both spatially and temporally.Presently, there are only very few locations where the full annual cycle of observations is available (Bakker et al., 2014).The SO region also has the largest differences in the net CO 2 fluxes as estimated from different methods involving observations and models (Landschützer et al., 2014).Recently, Resplandy et al. (2014) found that DIC-induced small spatial scale pCO 2 structures existing in the SO are non-negligible.Such small-scale processes are generally missing in the coarseresolution CMIP5 models and sparse observations.In addition, the strong interannual variations in the air-sea CO 2 fluxes identified in this region (see, e.g., Landschützer et al., 2015;Lovenduski et al., 2015) could also contribute to the discrepancies in the observed and model-simulated pCO 2 seasonality presented in this study (see also Hauck et al., 2013).
Beyond surface processes, in the subsurface circulation patterns could also to the surface biases simulated in the SO.Here, regions critical for biological production and carbon uptake are associated with mode and intermediate water formation locations (Sarmiento et al., 2004;Sallée et al., 2012).Despite the fact that the simulated net uptake rates of atmospheric CO 2 in the SO are mostly overestimated compared to the values derived from observations (as shown in this study), Frölicher et al. (2015) show that the CMIP5 models underestimate the anthropogenic-carbon storage in the Southern Ocean.This indicates either a shortcoming in the simulated large-scale overturning circulation or too strong a sink of non-anthropogenic carbon simulated here.Therefore, a better constraint of the former mechanism should be prioritized in order to improve the projection of the long-term evolution of air-sea CO 2 uptake.
New US-led initiatives that aim to enhance our understanding of the Southern Ocean processes are emerging, for instance, the Southern Ocean Carbon and Climate Observation and Modeling project (SOCCOM; http://soccom.princeton.edu).The biogeochemical Argo floats planned to be deployed will provide novel measurements that will help identify how the changing physical processes influence the biogeochemistry dynamics, and vice versa.The EXport Processes in the Ocean from RemoTe Sensing (EXPORT; http: //cce.nasa.gov/cce/ocean.htm)campaign, which studies the export and fate of ocean NPP using remote sensing observations, will also provide a better constraint for ecosystem process parameterization in the model.In addition, a multimodel intercomparison involving observational data such as this study is useful to elucidate the complex interplay among physical and biogeochemical processes, which ultimately would reduce uncertainties in climate projections.

Conclusions
The latest-generation ESMs project ocean carbon uptake with considerable uncertainty, and this uncertainty is projected to grow 2-fold by the end of the 21st century under a high future CO 2 emissions scenario.In this study, the evaluation of the CMIP5 ESMs was focused on assessing the ability of the models to project the future CO 2 fluxes between ocean and atmosphere by looking at the correlation coefficient of each region with the global future ocean CO 2 uptake.
We found that the highest inter-model correlation is in the Southern Ocean (SO) region (R SO mean = 0.76, R SO cum = 0.65), meaning that most models agree regarding the evolution of their CO 2 uptake behavior through the 21st century.The majority of models simulate a steady increase in CO 2 sink rate due to a weaker decrease in buffer capacity and due to a relatively stable NPP throughout the 21st century.We show that models that take up anomalously low CO 2 in the SO today would project low cumulative CO 2 uptake throughout the 21st century and vice versa.This suggests that the carbon uptake in the SO can be used to constrain future global uptake uncertainty.We highlighted that in other regions, the models simulate a decrease in CO 2 uptake during the second half of this century but with a large inter-model spread in the timing of the decreasing trend, thus affecting the multi-model correlation in these areas.
We have identified a strong bias in the amplitude of carbon uptake simulated by the CMIP5 models for the period 2001-2010 in the Southern Ocean, ranging from a source to the atmosphere to a sink almost 3 times more powerful than has been observed (1.03 ± 0.09 vs. 0.27 ± 0.13 Pg C yr −1 ; Lenton et al., 2013).Inter-model spread in the SO carbon sink arises from variations in the surface pCO 2 seasonality, which is attributed by the bias in the simulated timing and amplitude of primary production and SST.By analyzing the differences in the simulated pCO 2 seasonalities, we classified two groups of models according to two different behaviors.Models that simulate anomalously strong 2001-2010 CO 2 uptake in SO (Fig. 7a, red color) reproduce the observed pCO 2 seasonal cycle but its amplitude is 2.5 times stronger than the observation-based estimates.This is because of the strong surface DIC variations, which push down the pCO 2 by simulating too strong a biological production.The effect on the projected simulations is an increase in CO 2 uptake due to the weaker decrease in seawater CO 2 buffering capacity.Other models that simulate anomalously low 2001-2010 CO 2 uptake in SO (Fig. 7a, blue color) simulate a comparable seasonal pCO 2 amplitude to that estimated from the observations but a wrong seasonal phase.This is due to the bias in the SST seasonal cycle with stronger amplitude, which tends to push up the pCO 2 in January-March, therefore simulating the SO as evolving towards a source of carbon for the atmosphere.
These biases in time and magnitude show the difficulty in simulating the observed marine ecosystem and the biogeochemical processes that contribute and govern the ocean surface pCO 2 .Consequently, simulating the right contemporary seasonal cycles of biological processes NPP and SST in the Southern Ocean would allow us to constrain the bias in terms of future oceanic carbon flux.
Seasonal timing and amplitude in pCO 2 are shown to be critical in order to accurately simulate the present and future CO 2 uptake, and therefore accurate monitoring of these biogeochemical processes in the SO is critical in order to constrain the assessment of the contemporary and future ocean carbon sink and subsequently the uncertainty in future climate change.However, the Southern Ocean remains one of the most poorly sampled ocean regions with respect to biogeochemistry.The observational data analyzed here were generated through extrapolations from the limited direct measurements; this could add an extra uncertainty to the analysis.This emphasizes the urgent need for a sustained and comprehensive observational campaign in this region, which is emerging as the key region to better constrain the evolution of future ocean carbon sinks.

Figure 1 .
Figure 1.Time series of annual ocean carbon uptake as simulated by nine ESMs for the 2001-2100 period.Thick gray line indicates the multi-model mean, shaded areas represent ± one standard deviation of the inter-model variations.The observation-based estimate (black circle) is from Landschützer et al. (2014).

Figure 2 .
Figure 2. Global annual contemporary carbon uptake rate vs.(a) global future (2090-2099) carbon uptake and (b) cumulative global carbon uptake over 2001-2099.The straight gray lines show the best-fit linear regression across all models.

Figure 3 .
Figure 3. Panel (a): the eight regional ocean basins adopted in this study, defined according to different ocean basins and latitudinal lines of 70, 45, 15 • S, 15, and 60 • N, and (b) correlation coefficient between regional contemporary CO 2 uptake rate with (red bars) future uptake rate in the 2090s and (blue bars) cumulative carbon uptake in the 21st century.The numbers over the x axis on panel (b) represent the area of each region (in 10 8 km 2 ).

Figure 4 .
Figure 4. Annual contemporary carbon uptake rate vs. global uptake rate projected for the last decade of the 21st century by CMIP5 models in the (a) high-latitude Southern Ocean (SO) and (b) other ocean regions excluding the Southern Ocean (eSO).Annual contemporary carbon uptake rate vs. cumulative 21st-century carbon uptake projected by CMIP5 models in the (c) SO and (d) eSO.Panel (e): time series of correlation coefficient computed in a similar way as that in Fig. 3b (blue bars) for every decade between 2001 and 2099.The straight gray lines in panels(a-d) show the best-fit linear regression across all models.Vertical shades on panels (a, c) depict observation-based estimates(Landschützer et al., 2014).

Figure 5 .
Figure 5. Annual contemporary carbon uptake vs. global uptake rate projected for the last decade of the 21st century by CMIP5 models.Here the SO is defined using dynamic boundaries separated by a surface water density of 26.5 kg m −3 .Panels (a) and (b) show the contemporary SO carbon uptake on the x axes in Pg C yr −1 and mol C m −2 yr −1 , respectively.Panel (c) illustrates the 26.5 kg m −3 density lines that separate the MW from the TMW for the month of August 2005 as simulated by the different models (same color convention as in a and b).

Figure 6 .
Figure 6.Time series of anomalies (relative to the year 2001) of annual mean (a) carbon uptake, (c) net primary production, and (e) carbon uptake efficiency in the Southern Ocean.Panels (b), (d), and (f) also show the same fields as the panels on the left but for the ocean region outside of SO.Thick gray lines represent the inter-model mean, with gray shadings representing ±1 standard deviation of inter-model variations.

Figure 7 .
Figure 7. Seasonal cycle of (a-b) carbon flux, (c-d) NPP, and (e-f) SST anomalies for the Southern Ocean region simulated by CMIP5 models for the 2001-2010 (left panels) and 2090-2099 (right panels) decades.The red (blue) shadings represent the range of G1 (G2) models that simulate anomalously strong (weak) carbon uptake (see also text in Sect.3.3).The colored lines represent the mean of the respective model groups, black lines represent the observational estimate adopted fromLandschützer et al. (2014) for CO 2 uptake and SST anomaly, and the NPP was taken from SeaWiFS data as described inNevison et al. (2015).

Figure 9 .
Figure 9. Anomalies of decomposed pCO 2 components in the Southern Ocean for the 2001-2010 (left panels) and future 2090-2099 (right panels) periods: (a, b) overestimating models; (c, d) underestimating models.Values shown are from multi-model mean.The gray lines and markers are estimates derived from observations according to Landschützer et al. (2014).

Figure 10 .
Figure 10.Anomalies of nonthermal pCO 2 seasonal cycle, (pCO DIC+ALK 2 ) as simulated by seven ESMs for the 2001-2010 period.The gray dashed line indicates the observation-based estimate of the pCO DIC+ALK 2 seasonal cycle.The numbers within the parentheses represent the amplitude for each model as well as for the observation-based estimate.

Table 1 .
Description of the models analyzed in this study, indicating ocean model resolutions (levels in the vertical and horizontal resolution in degrees), the thickness of the surface layer, and the marine biogeochemical components included in the ESM.

Table 2 .
Main characteristics of the marine biogeochemical components of the nine ESMs used in this study: list of nutrients limiting the phytoplankton growth and the number of explicitly represented phytoplankton and zooplankton groups.