Observational studies report a rapid decline of ocean CO2 uptake in the temperate North Atlantic during the last decade. We analyze these findings using ocean physical-biological numerical simulations forced with interannually varying atmospheric conditions for the period 1979–2004. In the simulations, surface ocean water mass properties and CO2 system variables exhibit substantial multiannual variability on sub-basin scales in response to wind-driven reorganization in ocean circulation and surface warming/cooling. The simulated temporal evolution of the ocean CO2 system is broadly consistent with reported observational trends and is influenced substantially by the phase of the North Atlantic Oscillation (NAO). Many of the observational estimates cover a period after 1995 of mostly negative or weakly positive NAO conditions, which are characterized in the simulations by reduced North Atlantic Current transport of subtropical waters into the eastern basin and by a decline in CO2 uptake. We suggest therefore that air-sea CO2 uptake may rebound in the eastern temperate North Atlantic during future periods of more positive NAO, similar to the patterns found in our model for the sustained positive NAO period in the early 1990s. Thus, our analysis indicates that the recent rapid shifts in CO2 flux reflect decadal perturbations superimposed on more gradual secular trends. The simulations highlight the need for long-term ocean carbon observations and modeling to fully resolve multiannual variability, which can obscure detection of the long-term changes associated with anthropogenic CO2 uptake and climate change.
 The temperature, freshwater balance, and circulation of the North Atlantic Ocean have undergone substantial alterations during recent decades due to anthropogenic climate change and natural climate variability, resulting in overall warmer, more saline surface waters in lower latitudes and fresher surface waters in higher latitudes [e.g., Fung et al., 2005; IPCC, 2001; Curry et al., 2003; Curry and Mauritzen, 2005]. Many of the changes in ocean climate and circulation are in response to variations in the North Atlantic Oscillation (NAO), the dominant atmospheric climate mode in the basin on interannual to decadal timescales. During positive phases of the NAO, surface pressure decreases in the Icelandic low and increases in the Azores high, leading to stronger surface westerly winds, a northward shift of storm tracks, and altered cooling, evaporation and precipitation patterns.
 Here we analyze the mechanisms governing surface water pCO2 variability in the context of NAO driven variability using an ocean model hindcast simulation [Lovenduski et al., 2007; Doney et al., 2007, 2008a], which broadly reproduces the observed North Atlantic trends of sea-air ΔpCO2 during the last decade. Our results suggest that a component of the positive sea-air ΔpCO2 trend over the last decade occurred as the result of a shift from largely positive NAO index prior to 1995 to a period of mostly weakly positive or negative NAO. The rapid temporal shift in ΔpCO2 appears, in fact, to be anomalous with respect to the longer-term secular trends, which are significantly less pronounced. Moreover, our simulations exhibit intermittent periods of opposing trends (more negative ΔpCO2) in certain regions and time periods. After introducing our methods, we discuss relevant hydrodynamic responses of the North Atlantic Ocean to changes in the NAO index. Then we relate these findings to the variability of the CO2 system before closing with a synthesis of our results in the light of recent observations.
 Historical simulations using a global ocean carbon model are used to estimate the long-term trends of CO2 air-sea fluxes and related parameters over the North Atlantic on a monthly resolution. The ocean model includes the biogeochemical cycling of C, O, N, P, Fe, Si, and alkalinity and four phytoplankton functional groups (diazotrophs, diatoms, pico/nano-plankton, and coccolithophores). The ecosystem module [Moore et al., 2004] coupled to a modified version of the OCMIP-2 biogeochemistry code [Doney et al., 2006; Najjar et al., 2007] is embedded in the coarse-resolution Parallel Ocean Program (POP) ocean component of the Community Climate System Model (CCSM) [Yeager et al., 2006]. The hindcast (1979–2004) is forced with interannually varying atmospheric physical and CO2 data [Doney et al., 2007, 1998a]. Model validations for selected parameters have been given by Doney et al. [2007, 2008a, 2008b]. We compare our simulations with CO2 system observations [Lüger et al., 2006; Corbière et al., 2007; Schuster and Watson, 2007] for selected regions. For our mechanistic analysis we also refer to a model run with constant atmospheric CO2 conditions (preindustrial run) in order to unravel physical climate driven variability from trends due to rising atmospheric CO2 conditions.
 We employ salinity normalized alkalinity (AT,norm) and dissolved inorganic carbon (DICnorm) concentrations as water mass tracers using a mean salinity of 35; that is DICnorm = DIC*35/Salin situ and AT,norm = AT*35/Salin situ. The current model setup includes riverine freshwater inputs with zero salinity and with zero concentrations of DIC and AT; river inputs, therefore, are equivalent to net precipitation minus evaporation. Thus, the application of the simple salinity normalization procedure, as compared to more sophisticated analyses such as the one proposed by Friis et al. , introduces only minor error into our analysis.
 A regression analysis is used to unravel the mechanisms driving the simulated long-term trends of surface water pCO2 and sea-air ΔpCO2. Spatial trend maps are computed for each model property X from the local slope of the linear regression against time (δX). Regressions are calculated for the full hindcast (1979–2004) and for selected subperiods, including the 1995–2004 period, which covers the time when observational estimates show rapidly evolving ΔpCO2. The δpCO2 trends are decomposed using a linear Taylor expansion [Lovenduski et al., 2007; Doney et al., 2008a] into its governing processes: changes in salinity (S), temperature (T), salinity normalized DIC (DICnorm) and salinity normalized alkalinity (AT,norm):
Precipitation and evaporation of freshwater and river inputs drive correlated variations in surface water DIC and AT, which have opposing effects on pCO2. To remove this effect, we use salinity normalized DICnorm and AT,norm, and the partial derivative with respect to salinity ∂/∂SFW includes the effects of freshwater dilution on DIC and AT.
3.1. Responses of the North Atlantic Circulation to NAO Variability
 In this section, we briefly discuss the NAO driven variability of the circulation of the North Atlantic, particularly variations in surface velocity patterns. Under positive NAO conditions, such as during the 1989–1995 period, the North Atlantic Current (NAC), accelerates in the (north) eastward direction in response to the higher wind stresses (Figure 2). As a result more warm, saline subtropical water is transported toward the eastern subpolar gyre and polar seas. There is a corresponding reduction in transport into the return flow of the eastern subtropical gyre, as indicated by the temporal variations in integrated eastward transport across 28°W between 30–45°N. The transport in the eastern subtropical gyre is anticorrelated with the NAO with essentially zero lag. During positive NAO periods, the subtropical gyre also expands northward, evident from the buildup of a positive salinity anomaly at the boundary between the subpolar and subtropical gyre off the North American continent at approximately 45°N during the 1989–1995 period (Figure 3). The opposite circulation patterns occur during neutral or negative NAO conditions.
 The southward velocity of the Labrador Current and the eastward velocity of the southern limb of the subpolar gyre both also increase during positive NAO periods, bringing more fresh and cold waters of Arctic origin into the subpolar gyre (Figures 2b and 2c). Eastward transport is positively correlated with NAO, again with zero lag, across the northern (45–60°N) end of the 28°W section. Under neutral or negative NAO conditions, the NAC and the Labrador Current velocities weaken (Figure 2c), permitting Labrador Current Waters to spread along the eastern coast of North America (Figure 3, e.g., years 1996–2000). Our simulations are consistent in these respects with observational studies [e.g., Flatau et al., 2003].
 Positive NAO phases enhance the deep convection in the Labrador Sea and thus lead to an enhancement of the Meridional Overturning Circulation (MOC) with a 5 to 10 year time lag after the onset of a positive NAO [Eden and Jung, 2001; Eden and Greatbatch, 2003] (Figure 2a). Moreover, it has been reported that the strong westerly winds during positive NAO phases cause an anticyclonic anomaly (i.e., a slowing down) of the subpolar gyre [Greatbatch, 2000; Eden and Jung, 2001; Eden and Willebrand, 2001; Eden and Greatbatch, 2003]. Both processes tend to partially counteract the NAO related surface circulation patterns reported above, but their effect is significantly weaker than the more dominant and immediate wind-driven reorganizations.
3.2. Simulated ΔpCO2 Variability and Temporal Trends in the North Atlantic
 The CCSM simulations exhibit substantial anomalies in annual mean, deseasonalized sea-air ΔpCO2 (+/−20 ppm) in response to varying atmospheric forcing and ocean circulation (Figure 4). The ΔpCO2 anomalies have spatial scales of a few hundred to a few thousand km and often persist over multiannual timescales. The anomalies are typically subgyre scale; that is, multiple positive and negative anomalies occur simultaneously within either the subtropical or subpolar gyre.
 Over the full simulation (1979–2004), model surface ocean pCO2 rises gradually throughout the entire basin (Figure S1), with areas that both slightly exceed or lag behind the atmospheric rate of approximately 1.6 ppm a−1. This is illustrated in maps of the trend over time in sea-air ΔpCO2 (Figure 5a). Positive trends for the sea-air ΔpCO2 imply declining CO2 uptake in undersaturated areas or increasing CO2 release to the atmosphere in supersaturated areas. The broad similarity of the sea-air ΔpCO2 trends between the anthropogenic CO2 case (time-varying atmospheric CO2) (Figure 5a) and the preindustrial CO2 case (constant atmospheric CO2) (Figure 5b) indicate that these patterns arise primarily because of ocean dynamics acting on surface water pCO2 rather than the atmospheric CO2 perturbations.
 Trend estimates for sea-air ΔpCO2 are complicated on shorter, decadal timescales by regional interannual variability in surface water pCO2. For example for the recent positive NAO period (1991–1996), the absolute magnitude of the simulated sea-air ΔpCO2 trends (Figure 5c) are noticeably larger than the low-frequency (1979–2004) secular trends (Figure 5a). The 1991–1996 sea-air ΔpCO2 trends are strongly positive in the subtropics and western subpolar gyre (surface water pCO2 values growing faster than the atmospheric rate of 1.6 ppm a−1) and strongly negative in the eastern subpolar gyre and polar seas. The subpolar sea-air ΔpCO2 trends switch signs from the positive NAO period 1991–1996 to the neutral/negative NAO period 1997–2004 (Figure 5d).
3.3. Partitioning the Factors Driving ΔpCO2 and pCO2 Trends
 The NAO influences sea-air ΔpCO2 by altering the surface water thermodynamic properties (S, T, DICnorm, AT,norm) governing seawater pCO2 (equation (1)). These interactions occur via changes in air-sea heat and freshwater fluxes, biological fluxes, lateral transport, and vertical mixing. Visual inspection shows broad similarities in the annual sea-air ΔpCO2 anomaly patterns (Figure 4) with comparable anomaly maps for salinity (Figure 3), DICnorm (Figure S4), and temperature (Figure S5). For example, salinity and ΔpCO2 anomalies are typically anticorrelated, particularly evident during years of Great Salinity Anomaly events such as during 1984–1987 or 1991–1993.
 A linear decomposition of pCO2 variations following equation (1) leads to a more quantitative analysis of the factors governing pCO2 temporal trends (see Figure S1). In the anthropogenic CO2 case (1979–2004), surface water pCO2 trends are uniformly positive as expected because of anthropogenic CO2 uptake and the resulting positive trends in surface water DICnorm (Figure S1d). We remove this anthropogenic CO2 trend and isolate the signals caused by ocean circulation variability using a companion preindustrial CO2 simulation with constant atmospheric CO2 conditions (but the same time-varying atmospheric physical forcing). The preindustrial case pCO2 trends and the components due to S, T, DICnorm, AT,norm trends are shown in Figure 6 for 1991–1996 and in Figure 7 for 1997–2004.
 The dominant factors driving simulated pCO2 trends differ by region and time period, and the overall pCO2 trend often reflects the net balance of larger, opposing forcing trends. For 1991–1996 (positive NAO), warming of the NAC and subpolar gyre (positive pCO2 trend) is countered by a trend to more saline, lower-DICnorm conditions (negative pCO2 trend). The relative strengths of warming versus saline/lower DICnorm differ from the western to eastern basin, resulting in a dipole pattern in overall pCO2 trend. The effects of DICnorm and AT,norm are largely anticorrelated, dominated in their sum pCO2 (DICnorm & Alk,norm) by the DICnorm pattern. The thermodynamic forcing terms are more in phase in the subtropics, with warming, freshening, and elevated DICnorm all contributing to positive pCO2. As discussed above for ΔpCO2 trends, the 1997–2004 (neutral/negative NAO) pCO2 trends flip sign in the NAC and subpolar gyre because of a reversal of the regional trends in warming/cooling, salinity and DICnorm. The longer-term secular trends (1979–2004) in pCO2 due to T, S and AT,norm are substantially weaker or even vanish (Figure S1), which is the case for the ΔpCO2 as well (Figure 4).
 To further illustrate the impact of climate variability on ΔpCO2 trends, we examine time series of surface water at two locations in the eastern (56.1°N/18.5°W, Figure 8) and western (45.8°N/43.6°W, Figure 9) NAC/subpolar gyre. Both locations exhibit substantial multiannual variability in DICnorm, salinity, temperature and sea-air ΔpCO2. Consistent with the regional maps (Figures 6 and 7), periods of sharply increasing DICnorm at each location tend to occur when surface salinity is declining and temperature is rising. The western and eastern sides of the subpolar gyre are out of phase, with large surface freshening and increasing DICnorm in the west prior to 1995 and in the east after 1997.
 The dipole-like spatial behavior in model pCO2 trends at these two locations can be attributed to variations of the NAO index, which was anomalously positive from 1989 to 1995 and then reverted mostly to more neutral or negative values. NAO-driven variations in wind-driven surface water transport (section 3.1) modify the relative supply of warm, saline, low-DICnorm subtropical water masses from the NAC and cold, fresh, high-DICnorm polar water masses from the Labrador Current (Figure S3). The anticorrelation of salinity and DICnorm at both subpolar gyre locations is evident from the DICnorm versus salinity plot (Figure 10). Note that the relationship is between NAO and the multiannual trends in surface water properties, not necessarily the absolute magnitude of the anomalies themselves. The NAO response at the western location is relatively immediate, while the response at the eastern location is lagged by approximately 2 years reflecting the transport time of the surface waters across the North Atlantic [see, e.g., Belkin, 2004].
3.4. Mechanisms Driving Variability of the North Atlantic CO2 System
 We suggest the following mechanism to explain some of the major CO2 system variability features in the northern North Atlantic. During the early 1990s, a period of positive NAO, increased northeastward NAC transport (Figure 2) enhanced the supply of subtropical high-salinity/low-DICnorm waters, i.e., an enhanced supply of a CO2 deficiency, to the eastern subpolar gyre. The sea-air ΔpCO2 thus became more negative with time in the eastern subpolar gyre (Figures 6 and 8). The gradual buildup in surface salinity and DICnorm anomalies in the eastern basin due to changes in NAC flow lagged by a couple of years relative to the NAO [see, e.g., Belkin, 2004]. The effect of the enhanced intrusion of the Labrador Current Water into the eastern subpolar gyre was minimal relative to the enhanced delivery of subtropical North Atlantic Current Water, evident by the increasing salinity at the eastern subpolar gyre station (Figure 8).
 During the early 1990s positive NAO period, the corresponding decreased transport of Gulf Stream/NAC waters to the eastern subtropical gyre resulted in a reduced supply of low-DICnorm water from the west. This lead to positive trends in DICnorm and ΔpCO2 during positive NAO periods and to opposite trends after 1996 during the subsequent neutral/low-NAO period. The eastern subtropical gyre was thus out of phase with the eastern subpolar gyre in the simulations over the recent observation period.
 In the western North Atlantic, positive NAO conditions in the early 1990s caused a northward shift of the intergyre boundary. This resulted in a zonal band east of Nova Scotia of high salinity, negative DICnorm, and negative ΔpCO2 anomalies along the subtropical/subpolar boundary. Further north off Newfoundland and the Grand Banks, the acceleration of the southward flowing, low-salinity Labrador Current resulted in declining salinity and increasing DICnorm (Figure 9). The ΔpCO2 trends in this region were somewhat more damped because of a warming trend, which was particularly pronounced in the western subpolar gyre during the 1990s. Some of the warming trends can be attributed to NAO while some may be secular warming due to anthropogenic climate change.
 Perturbations to these multiannual, NAO-related carbon cycle patterns arise from transient events such as the propagation of surface Great Salinity Anomalies (GSA) around the subpolar gyre. Positive NAO conditions have been related to the occurrence of GSAs, which have been observed as episodic events in the polar and subpolar North Atlantic [e.g., Belkin, 2004]. Pulses of low-salinity/high-DICnorm waters (Figure 3) intrude into the subpolar gyre via the Labrador Sea and Labrador Current, and GSAs spread across the entire North Atlantic basin on relatively short timescales (1–3 years) via the NAC (Figure 3). These anomalies temporarily reduce CO2 uptake because of their high-DICnorm characteristics (Figure S4). Our simulation clearly reproduces the GSAs during 1982/1983, 1991 and 1993 [Belkin, 2004] (Figure 3).
 NAO-driven changes in wind speed constitute another factor that could contribute to CO2 flux variability [Doney et al., 2008a]. Neutral or negative NAO conditions are characterized by a reduction in wind speed over the subpolar gyre, in particular over the northeastern Atlantic Ocean. These wind conditions would thus act synergistically with the fundamental changes in the surface circulation pattern and support reduced CO2 uptake during neutral or negative NAO conditions.
 The simulated air-sea CO2 fluxes show a weak, but still significant, correlation with the NAO in two regions, the western subpolar gyre and the eastern North Atlantic north of 60°N (Figure 11). The CO2 flux–NAO correlation in the western subpolar gyre is positive and strongest at zero phase lag and is caused primarily by surface water pCO2 responses to varying SST. Substantial cooling under positive NAO phases lowers the pCO2 and permits higher CO2 uptake (Figures S4 and S5). The CO2 flux–NAO correlation in the eastern North Atlantic north of 60°N is negative and strongest with an approximately 2–4 year phase lag. This negative correlation can be attributed to warming (Figure S5) and enhanced delivery of more saline (Figure 3) waters into the polar region, as triggered by the enhanced surface circulation during positive NAO stages. In this region the temperature effect apparently dominates, leading to positive ΔpCO2 anomalies (Figure 5).
3.5. Comparison of Model CO2 System Trends With Observations
 The positive model 1997–2004 sea-air ΔpCO2 trends in the eastern subpolar gyre and polar seas (Figure 5d) are consistent with the trends evident from North Atlantic observational studies [Lefèvre et al., 2004; Omar and Olsen, 2006; Olsen et al., 2006; Lüger et al., 2006; Corbière et al., 2007; Schuster and Watson, 2007], which are based primarily on data from 1995 to 2002/4 (Figure 1). For the same time period, the model eastern subtropical gyre also exhibits positive, but weaker ΔpCO2 trends. In contrast, the ΔpCO2 trends for an earlier period 1991–1996 (Figure 5c) are characterized by a strong negative trend in the eastern Atlantic north of approximately 50°N, and by a strong positive ΔpCO2 trend in the eastern subtropical gyre. Over the full simulation (1979–2004), model surface ocean pCO2 rises gradually throughout the entire basin, with areas both only slightly exceeding or lagging behind the atmospheric rate of approximately 1.6 ppm a−1 as illustrated in the sea-air ΔpCO2 trends (Figure 5a). To the degree then that the model simulations capture the behavior of the real ocean, we suggest that the observed recent ΔpCO2 trends may be partially aliasing multiannual variability and do not fully reflect a secular change in the North Atlantic CO2 sink.
 To assess the skill of the model simulations, we compare our model results against field data for four selected regions, with figures provided in the supplement (Figures S6–S8). Observations at Bermuda Atlantic Times Series station (BATS) show the surface water pCO2 and atmospheric pCO2 rising at comparable rates [Bates, 2001, 2007]. This implies a constant sea-air ΔpCO2, which is reproduced in our model trends for the full simulation (1979–2004; Figure 5a). The simulations for the BATS region are discussed in detail by Thomas et al. . Further north in the western Atlantic Ocean, Lüger et al.  suggest increasing CO2 uptake, which is in accord with our simulations for some parts of the northwest Atlantic over the 1997–2004 period (Figure 5d). Despite a slight rise of the annual average of the sea surface temperature, the primary cause for this increase appears to be a cooling of the winter surface waters, which lowers the pCO2 and facilitates CO2 uptake (Figure S6).
 Several studies [Lefèvre et al., 2004; Omar and Olsen, 2006; Olsen et al., 2006; Lüger et al., 2006; Schuster and Watson, 2007] report declining sea-air ΔpCO2 and CO2 uptake in the NAC and eastern subpolar gyre. We show in detail in Figure 12a the simulated and observed pCO2 time series for one location in the eastern subpolar gyre, where the study areas of Lüger et al.  and Schuster and Watson  overlap. The observed and simulated surface ocean pCO2 rise more rapidly than the atmospheric pCO2, though the observed rise exceeds the simulated one. At this particular location both simulations and observations exhibit an increase in seasonality; however, our simulations reproduce a decrease in the seasonality for regions further northeast as reported by Schuster and Watson . Drivers for these patterns are the increase of winter minimum temperatures as well as a decrease of the summer maximum temperatures. The increase in winter temperature can be attributed to the reduced wintertime cooling of the North Atlantic during neutral or negative NAO phases because of the weakened westerly winds. Furthermore, the NAC slows down during neutral or negative NAO phases leading to a reduced delivery of warmer subtropical water contributing to the cooler summer temperatures (Figure S7). The observed decrease of seasonality [Schuster and Watson, 2007] thus might be caused by the decreased seasonality in surface temperature during neutral or negative NAO phases.
 Declining sea-air ΔpCO2 and CO2 uptake values observed south of Greenland [Corbière et al., 2007] have been assigned to rising temperatures. The simulations for this location (Figures 12b and S8) reproduce the observations, in particular the vanishing of the ΔpCO2 minima during the 1995–2004 period. For this period, our simulations give a surface water pCO2 trend of +3.0 ppm a−1 while Corbière et al.  report +2.5 μatm a−1. Subtracting the rise of atmospheric CO2 of approx. +1.6 ppm a−1 yields a remainder of approximately 1.4 ppm a−1 to be explained. Our simulations suggest warming as the major driver for this region (Figure S8), raising surface water pCO2 by +1.2 ppm a−1. Our simulations further report trends for DIC and AT of approximately +0.6 μmol DIC (kg a)−1 and −0.3 μmol AT (kg a)−1. Corbière et al.  did not identify (nor statistically reject) these findings, however it has to be noted here that these trends are relatively small and close to the detection limit for observational studies. Perhaps more importantly, the DIC and AT trends approximately cancel out with respect to pCO2.
 Using historical hindcast simulations (1979–2004), we show that the ocean CO2 system in the temperate North Atlantic exhibits substantial multiannual variability on sub-basin scales in response to NAO-driven reorganizations in ocean circulation and surface warming/cooling. The simulated temporal evolution in sea-air ΔpCO2 is consistent broadly with reported observational trends. We argue that the large CO2 system trends estimated recently from field data may be heavily influenced by multiannual variability. For example, many of the observational estimates [Lefèvre et al., 2004; Omar and Olsen, 2006; Olsen et al., 2006; Lüger et al., 2006; Corbière et al., 2007; Schuster and Watson, 2007] cover a period after 1995 of mostly negative or neutral NAO. In our simulations negative or neutral NAO conditions result in reduced NAC transport of warm, saline, low-DICnorm subtropical water into the eastern subpolar gyre, resulting in a substantial decline in CO2 uptake along the NAC and in the eastern subpolar gyre. We suggest therefore that air-sea CO2 uptake may rebound in the eastern temperate North Atlantic during future periods of more positive NAO, similar to the patterns found in our model for the sustained positive NAO period in the early 1990s. Our analysis provides support that long-term, coherent ocean carbon observing systems are of critical relevance for the understanding and interpretation of secular changes, associated with climate change and uptake of anthropogenic CO2 [e.g., Doney et al., 1998b; Fung et al., 2005; IPCC, 2001; Curry et al., 2003; Curry and Mauritzen, 2005] on the one hand and interannual to interdecadal variability on the other [Levine et al., 2008].
 S. C. Doney and I. D. Lima were supported by NASA grant NNG05GG30G. H. Thomas holds a Canada Research Chair. We are grateful to an anonymous referee and associate editor C. LeQueré, whose constructive comments greatly helped improve the manuscript. We are grateful to N. Metzl (LOCEAN/IPSL, Paris, France) and H. Lüger for making observational data available. This work contributes to CARBOOCEAN, an EU-FP6 project.