Net Arctic Ocean primary production (PP) is expected to increase over this century, due to less perennial sea ice and more available light, but could decrease depending on changes in nitrate (NO3) supply. Here Coupled Model Intercomparison Project Phase 5 simulations performed with 11 Earth System Models are analyzed in terms of PP, surface NO3, and sea ice coverage over 1900–2100. Whereas the mean model simulates reasonably well Arctic-integrated PP (511 TgC/yr, 1998–2005) and projects a mild 58 TgC/yr increase by 2080–2099 for the strongest climate change scenario, models do not agree on the sign of future PP change. However, similar mechanisms operate in all models. The perennial ice loss-driven increase in PP is in most models NO3-limited. The Arctic surface NO3 is decreasing over the 21st century (−2.3 ± 1 mmol/m3), associated with shoaling mixed layer and with decreasing NO3 in the nearby North Atlantic and Pacific waters. However, the intermodel spread in the degree of NO3 limitation is initially high, resulting from >1000 year spin-up simulations. This initial NO3 spread, combined with the trend, causes a large variation in the timing of oligotrophy onset—which directly controls the sign of future PP change. Virtually all models agree in the open ocean zones on more spatially integrated PP and less PP per unit area. The source of model uncertainty is located in the sea ice zone, where a subtle balance between light and nutrient limitations determines the PP change. Hence, it is argued that reducing uncertainty on present Arctic NO3 in the sea ice zone would render Arctic PP projections much more consistent.
 Rapid Arctic sea ice retreat has occurred over the last 10 years [e.g., Stroeve et al., 2012], culminating in 2012 with a September annual minimum nearly 50% below the 1979–2000 average [University of Colorado at Boulder, 2012]. This retreat is associated with younger and thinner ice [Maslanik et al.2007; Kwok and Rothrock, 2009] and an increasingly long melt season [Stammerjohn et al.2012]. Summer ice retreat has the potential to profoundly impact the production of organic carbon by marine phytoplankton. Indeed, a 20% increase in net primary production (PP) annually integrated over the whole Arctic Ocean (hereafter, integrated primary production, IPP, TgC/yr) from 1998 to 2009 has been inferred using a combination of satellite ice concentration and ocean color products [Arrigo and van Dijken, 2011]. This increase has been shown to be associated with sea ice retreat, not with increasing primary production per unit area (hereafter, specific primary production, sPP, mgC/m2/yr). Ocean color data also suggests earlier bloom onsets in the Arctic [Kahru et al., 2011]. Those changes in primary production have a number of potential important consequences on marine food webs [see Tremblay et al., 2012, and references therein], as well as on the carbon cycle [Bates and Mathis, 2009; Cai et al., 2010]. Along those changes, other ongoing physical and geochemical changes in the Arctic Ocean have also been suggested: warming, upper layer freshening, altered stratification, increased acidification, and, in the subarctic Pacific, increased hypoxia [e.g., Carmack and McLaughlin, 2011].
 Under increasing greenhouse forcing, sea ice retreat is projected to continue into the future [e.g., Massonnet et al., 2012]. Updated climate projections—performed in the framework of the World Climate Research Program Coupled Model Intercomparison Project Phase 5 (CMIP5) [Taylor et al., 2012]—suggest a summer ice-free Arctic between 2041 and 2060 in a high-emission climate scenario [e.g., Massonnet et al., 2012]. How primary production will evolve in the future Arctic Ocean depends on the present status and the future evolution of the limiting factors. Intuitively, sea ice retreat promotes phytoplankton growth via an increase of light availability. Arrigo and van Dijken  project a >60% increase in IPP for a summer ice-free Arctic using a simple linear extrapolation of the observed recent trends. This admittedly neglects potential nutrient limitation, as nitrate already limits primary production in some seasonally ice-covered Arctic locations [Tremblay and Gagnon, 2009]. However, the present and future role of nitrate limitation is difficult to assess. First, nutrient data are extremely scarce in the Arctic, and, second, the roles of factors directly controlling nitrate concentrations (ocean circulation, river runoff, atmospheric inputs, and biogeochemical recycling, including photo-oxidation) are not well understood [Wassmann and Reigstad, 2011].
 These uncertainties on phytoplankton growth limitation in the present Arctic are also obvious in ice-ocean model simulations of primary production forced by atmospheric reanalyses [Popova et al., 2012]. These models show consistent spatial patterns of Arctic primary production and reasonable agreement with the satellite ocean color-based inferences of Pabi et al. . Moreover, whereas those models manifest similar levels of light limitation owing to the general agreement on the ice distribution, they disagree on the amount of nutrients available for primary production [Popova et al., 2012]. Hence, the predictive capabilities of such models may be questionable as nutrient limitation could become increasingly important.
 Indeed, in an early model intercomparison using four Earth-System Models (ESMs) focusing on the response of marine primary production to climate change [Steinacher et al., 2010], the Arctic Ocean was identified as one of the regions where the models do not agree, with three ESMs showing increased primary production at the end of century, whereas one model shows decreased primary production. Here we find that 11 ESMs used for the new CMIP5 projections still disagree on the sign of future primary production change, primarily due to uncertainties in NO3 concentrations (see Figure 1). Model simulations are analyzed using simulated vertically integrated net primary production, ice concentration, and surface nitrate concentration. In section 2, methods are described, including models, simulations, variables, and diagnostics of interest, as well as observational data sets used to evaluate models. In section 3, we first evaluate the ESMs simulations using historical data sets (section 3.1); then, we analyze projections for the late 21st century. Projections are analyzed for the whole Arctic region (section 3.2) and then separately for the open ocean and sea ice zones (section 3.3). We finally briefly discuss the stratification changes and their potential impact on NO3. Results are discussed and conclusions are drawn in section 4.
2.1 Description of Models and Simulations
 Table 1 lists and briefly describes the 11 ESMs used for this study. ESMs were selected on the requirement that they archived sea ice, primary production, and nitrate fields up to 2100 by the time of preparing this paper. We used the historical simulations from 1900 to 2005, and one climate change scenario, the representative concentration pathway (RCP) 8.5, for which the radiative forcing increases nearly steadily over the 21st century to peak at +8.5 W/m2 in 2100 relative to preindustrial [Moss et al., 2010]. Among the several ensemble members that were ran for each model, only the first one was analyzed.
Table 1. A Brief Description of the ESMs Used in This Study
aReferences: Canadian Earth System Model (CanESM2) [Arora et al., 2011; Christian et al., 2010], Geophysical Fluid Dynamic Laboratory (GFDL) models [Dunne et al., 2012, 2013], Hadley Centre Global Environmental (HadGEM) models [HadGEM2 Team, 2011; Collins et al., 2011], Institut Pierre-Simon Laplace (IPSL) models [Dufresne et al., 2013; Aumont et al., 2003], Model for Interdisciplinary Research On Climate (MIROCs) [Watanabe et al., 2011; Kawamiya et al.2000]; Max-Planck-Institut (MPI) models [Ilyina et al., 2013; M. A. Giorgetta et al., Climate change from 1850 to 2100 in MPI-ESM simulations for the Coupled Model Intercomparison Project 5, submitted to Journal of Advances in Modeling Earth Systems, 2013].
bNumber of vertical levels, longitude/latitude resolution of model grid.
dStructure of the pelagic ecosystem model (NPZD = fixed quota Nutrient-Phytoplankton-Zooplankton-Detrital Matter Model, if more elaborated than NPZD, the number of transported tracers is indicated); number of phytoplankton groups (PG), limiting nutrients.
eCompared to MIROC-ESM, MIROC-ESM-CHEM includes an atmospheric chemical component.
 ESMs include representations of the general circulation and physics of the atmosphere, ocean, and sea ice, as well as several so-called Earth System processes, including biogeochemistry and carbon cycle for land and ocean. Among the models, different decisions were made with respect to the complexity of these components. Atmospheric (oceanic) horizontal grid resolution ranges over 1°–4° (0.5°–2°) and the number of vertical levels ranges from 24 to 95 (31 to 59).
 All models have reasonable sea ice dynamics. Regarding sea ice thermodynamics, most models still use the Semtner  zero-layer model which is known to misrepresent the phase of the sea ice seasonal cycle. Four models include a representation of the subgrid-scale ice thickness distribution, which significantly improves the seasonality of both ice extent and thickness [Massonnet et al., 2011] and sea-ice related feedbacks [Holland et al., 2006]. Typically, there is no effective transmission of light through the ice itself, only the open water in the pack contributes: solar radiation which penetrates through the ocean surface linearly increases with simulated open water fraction.The only exceptions are the Geophysical Fluid Dynamics Laboratory (GFDL) models, which also enable light penetration through ice, depending on surface albedo, snow depth, and ice thickness.
 As far as the representation of marine biogeochemistry and microbial marine ecosystems is concerned, models can be divided in three groups. CanESM2 and the two MIROCs are the most simple in that respect, using an NPZD component with a single phytoplankton group and one limiting nutrient (N). The MPI-ESMs and HadGEM2s use enhanced NPZD models, including more limiting nutrients and an additional phytoplankton group for the HadGEM2s. At the end of the range, the IPSL and GFDL models use >20 tracers, including several phytoplankton and zooplankton groups, as well as an exhaustive ensemble of limiting nutrients, with, for some of them, varying quotas in the different model groups. Besides these differences in ecosystem model structure, there are also differences on how the model treat riverine and atmospheric inputs, sediment remobilization, and nitrogen fixation.
 In order to eliminate any spurious model drift, the initialization of the models starts from climatologies or from CMIP3 output. It then typically includes several thousand years of the offline ocean, sea ice, and marine biogeochemical component and then, a few hundred years of the fully coupled model (see Table 1). Therefore, the initial fields reflect the model equilibrium state rather than the observed values.
2.2 Diagnostics Used for Analysis
 The Arctic Region is our domain of interest. It encompasses all the oceanic pixels located north of the Arctic Circle (66.56°N) as in Arrigo and van Dijken . We use model output for ice concentration, nitrate concentration, and PP, as described below. All fields were interpolated on a 1° geographic latitude-longitude grid for the analysis.
 To characterize sea ice, we use monthly fields of ice concentrationA(λ,φ,t) (sicin the CMIP5 terminology), which is the fractional surface of a grid cell covered by sea ice. λ and φ are latitude and longitude. Ice concentration is used to split the Arctic Region into time- and model-dependent zones: (i) the ice zone (IZ), i.e., the ensemble of cells with at least 15% of ice at least 1 month a year, (ii) the perennial ice zone (PIZ), i.e., the cells with at least 15% of ice in September, month of the annual minimum, (iii) the seasonal ice zone (SIZ), i.e., the cells that are in the IZ but not in the PIZ, and (iv) the open ocean zone (OZ), i.e., the cells that are not in the IZ. Note that 15% is a commonly used threshold to remove cells with small ice concentration values from the IZ. The extent of each zone is simply the integrated surface of area of its grid cells.
 Since nitrogen is the limiting nutrient in the Arctic Ocean, we characterize nutrients based on the annual mean fields of surface nitrate concentration NO3(λ,φ,t) (mmol/m3) (no3 in the CMIP5 terminology). Ideally, the annual maximum nitrate value would have been a better measure of the potential for primary production, but was not available for all models. However, the annual maximum time series could be computed for nine models and was found to be highly correlated (0.99) with the annual mean, over 1900–2100. Therefore, annual mean nitrate can be used as a valid replacement for trend analysis in the present context. In contrast, mean nitrate is not as good as a predictor for actual summer NO3 limitation, but analysis in section 3.2 shows it is reasonably efficient for interpretation of the integrated primary production time series.
 We characterize primary production based on the annual fields of vertically integrated net primary mole productivity of carbon by all types of phytoplankton (intpp in the CMIP5 terminology), and convert it into specific primary productionsPP(λ,φ,t) (gC/m2/yr). Integrating spatially, we obtain the integrated primary productionIPP(t) (TgC/yr). Dividing by the area of a given region, we obtain the specific regional primary production(gC/m2/yr) for that particular region.
 Analysis of time series is facilitated if normalized values are used. Normalized integrated primary production for a given year and a given model is defined as . Normalized average nitrate concentration is defined as , where mmol/m3 is the half-saturation concentration for nitrate uptake by diatoms [Sarthou et al., 2005]. The multiplication by 1.5 is performed because nitrate limitation as formulated in the models starts for . Using this formulation, is >0 for supposed eutrophic conditions and <0 for oligotrophic conditions.
 We had no access to the actual values of the factors limiting diatom production in the models, but we could to some extent reconstruct them, since the different limitation functions are generally expressed quite similarly among the models. For nitrate, we used the standard hyperbolic dependence:
For limitation by light, we first noticed that, assuming no light transmission through sea ice (true for most models), the annual average incoming photosynthetically available radiation (PAR) is well correlated with September ice extent, or equivalently, the perennial ice zone extent SPIZ (correlation = 0.89). Hence, for light, we use a limitation function based on the PIZ extent and the total surface area of the domain Sdom:
The combined light-nutrient limitation function is defined as: . We refer to the resulting expressions as empirical limitation functions, to emphasize that those are not the actual limitation functions, but rather an indirect measure of diatom growth limitation in the models. Empirical limitation functions have a value ∼1 for weak limitation of diatom growth and ∼0 for strong limitation.
2.3 Data Used for Validation
 Several observational products are used to evaluate the models historical simulations. Primary productiondata were calculated using the latest reprocessing (R2010.0) of surface Chl-a derived from 8 day Level 3 binned SeaWiFS ocean color measurements [Pabi et al., 2008; Arrigo and van Dijken, 2011]. This derivation is based on the following. Chl-a is assumed to be constant over a prescribed (20 m) mixed-layer and to decrease exponentially further down into the water column. Using formulations for temperature response and light limitation of phytoplankton growth, daily net primary production (mgC/m2/d) is derived for each satellite pixel. Using sea ice data and spatiotemporal extrapolation procedures, integrated primary production (TgC/yr) is derived for each year. The temporal (March–September) and spatial coverage are limited, as detailed in Pabi et al. . Here the 1998–2011 time series of integrated primary production (TgC/yr) (see Figure 2b) and the spatial distribution of annual specific primary production on a 9 km grid (1998–2005) (see Figure 3e) are used. Let us caution that satellite primary production retrievals underestimate vertically integrated production. This is because they are based on surface chlorophyll estimates and do not resolve the frequent subsurface chlorophyll maxima [Tremblay et al., 2008]. A brand new synthesis of primary production data in the Arctic Ocean [Hill et al., 2013] suggests that Arctic-integrated PP increases from 466 ± 94 TgC/yr to 993 ± 94 TgC/yr when subsurface maxima are included.
 We use surface nitrate data from the World Ocean Atlas (WOA) [Garcia et al., 2010] (see Figure 2c). The 2544 valid data points could be extracted above the Arctic circle, covering 1928–2005. Data coverage is sparse, with practically no data in the Arctic basin and an important summer bias [Popova et al.2012].
Sea ice concentration data (1979–2007) derive from scanning multichannel microwave radiometer (SMMR) and the Special Sensor Microwave/Imager (SSM/I), reprocessed by the EUMETSAT Ocean and Sea Ice Satellite Application Facility [Tonboe et al., 2011].
3.1 Model Evaluation
 To evaluate the models, we analyze the results of the historical simulations and compare them to observational products. We find that primary production—both in terms of domain integral and spatial distribution—is rather realistic in the mean model, but highly variable among individual models. This is due to (i) a correct localization of the PIZ for the mean model and (ii) large intermodel differences in the factors controlling PP (sea ice and, in particular, NO3).
Domain-integrated values. The first three columns of Table 2 display the values of IPP, NO3, and the surface of the perennial ice zone (SPIZ) over the last decade of the 20th century. Figure 2 shows the time series of integrated production, mean surface nitrate, and PIZ extent for all models over 1900–2100 and how they compare to observational estimates.
Table 2. Simulations Statistics for Each Model and the Mean Modela
aAnnual primary production (domain integral, IPP); surface NO3 concentration (domain average, ); surface area of the perennial ice zone (SPIZ); and seasonal maximum mixed layer depth (domain average, ); over 1980–1999 and 2080–2099. We also give ΔIPP, the 2080–2099 to 1980–1999 change in IPP(domain integral); , the linear regression coefficient from the fit IPP(y)=a·y+b(over 2080–2099); Tolig, the year where oligotrophy is attained (e.g., ) for the first time; and TSIZ, the year where SPIZ falls below 106 km2 for the first time.
bMixed layer depth directly diagnosed from the vertical mixing scheme (omlmax).
cT and S-derived mixed layer depth (mlotst) provided by individual modelling groups, as omlmaxwas not available from the CMIP5 archive.
 Mean model IPP is reasonably realistic, reaching 511 ± 12 TgC/yr over 1998–2005, compared with the 469 ± 16 TgC/yr derived from satellite ocean color data. The range from in situ observations [Hill et al.2013] is much larger (466–993 TgC/yr). However, the intermodel spread is large, reaching 40% of the mean, MIROC-ESM-CHEM (206 TgC/yr) and GFDL-ESM2M (947 TgC/yr) being the least and most productive models over 1998–2005, respectively.
 Mean model surface NO3 is 4.5 ± 2.5 mmol/m3 over 1928–2005, compared with an observed 3.0 mmol/m3. This comparison has to be interpreted carefully, as the uncertainty on observed NO3 is high above the Arctic circle, with almost no WOA observations in the Arctic Basin [see Figure 3f and Popova et al., 2012]. Besides, the intermodel spread in NO3 is large, reaching 55% of the mean, IPSL-CM5A-LR (1.7 mmol/m3) and MPI-ESM-LR (9.1 mmol/m3) showing the least and most surface NO3 concentrations, respectively.
 The extent of the perennial ice zone (PIZ) over 1979–2005 is relatively well represented in the mean model (7.1 ± 0.5 million km2), compared with an observed 7.4 ± 0.6 million square kilometers. The intermodel spread in PIZ extent is less than for IPP and NO3, reaching only 17% of the mean. CanESM2 (4.6 million km2) and GFDL-ESM2G (9.0 million km2) are the two extreme models in terms of PIZ extent over 1979–2005.
 The spread in IPP is driven by the differences in the factors controlling primary production, namely nitrate concentration and sea ice. The associated corresponding empirical limitation factors are highly variable among models (see Figure S2). For 1980–1999, phytoplankton growth is typically more limited by sea ice than by nutrients, as indicated by the smaller empirical limitation factor for sea ice (0.49) than for nitrate (0.73) (mean model, 1980–1999). The intermodel range in both limitation factors is significant, from 0.36 (GFDL-ESM2G) to 0.66 (CanESM2) for sea ice; and from 0.51 (IPSL-CM5A-LR) to 0.85 (MPI-ESM-MR) for NO3.
Spatial distributions. The spatial distribution of specific primary production (see Figures 3a, 3e, and S1) is similar to satellite ocean-color retrievals, with near-zero values in the PIZ and maxima on the Siberian shelves, in the Chukchi Sea, and in the North Atlantic sectors of the Arctic Ocean. Large intermodel variations exist, however. For instance, the simulated maxima found near river mouths on the West Siberian shelves are highly variable and not as large as observed. In addition, both GFDL models have a significant share of under-ice production. Under-ice production cannot be detected from satellites but is supported by in situ observations [e.g., Alexander and Niebauer, 1981; Mundy et al., 2009; Arrigo et al., 2012]. The two HadGEM2s have high productivity on the East Siberian shelf and in the Chukchi Sea, and the two MIROCs simulate small specific primary production overall (see Figure S1).
 The spatial distribution of surface NO3 is even more variable than that of sPP (see Figures 3b–3f and S1). The simulated NO3 maximum in the North Atlantic is also found in WOA observations (see Figures 3b–3f), while the Chukchi Sea maximum is not, which could be due to a sampling summer bias in the WOA observations or to model parameterizations affecting Pacific water inflow. Indeed, one would expect an Atlantic-Pacific NO3 gradient across the Arctic, as Pacific waters entering the Arctic tend to have higher NO3 concentrations than Atlantic waters [Codispoti et al., 2013]. Each model has specific regional characteristics. For instance, the mean model NO3 maximum in the Chukchi Sea is only due to the contribution of four models (MPIs and HadGEM2s). CanESM2, IPSL, and MIROC models generally show weak NO3 variations. The other models typically feature one large regional maximum, in the North Atlantic for the GFDL models, on the East Siberian Shelf for the HadGEM2 models, and a broad maximum centered in the Pacific Arctic for both MPI models (which have the largest NO3 values in the sample).
 The 1979–2005 spatial distribution of the PIZ is well simulated in the mean model and by most models (see Figures 3c, 3d, 3g, 3h, and S1). The IZ extent is typically slightly overestimated, especially on the East of Greenland. The September concentration field is more variable among models, but intermodel differences are consistent with the location of the PIZ. Notable models are CanESM2 and HadGEM2-ES (summer low bias), as well as IPSL-CM5A-LR (winter high bias).
3.2 Response of Primary Production to Climate Change: The Arctic Ocean
 Summer ice extent and nitrogen levels both sharply decrease over the 21st century for all models (Figure 2 and Table 2). In contrast, the IPP time series widely differ among models: for some models, IPP increases continuously (e.g., GFDL's); for some, it remains rather stable (e.g., CanESM2); and for the rest of the models, IPP initially increases, reaches a temporary maximum, then decreases (e.g., IPSL's). To quantify changes, we define ΔIPP as the difference in 2080–2099 mean IPP from the 1980 to 1999 mean (see Table 2and Figure 4a). Over 11 models, two models predict a large increase (GFDLs); six predict a mild increase (e.g., ΔIPP<100 TgC/yr, CanESM2, HadGEM2-CC, MIROCs, and MPIs); and three predict a decrease (IPSLs and HadGEM2-ES). In addition, among the six models with a mild positive ΔIPP, IPP is stable over 2080–2099 in two models and decreases in four models (see in Table 2). For those models, the cumulated decrease has not been sufficient yet to bring IPP below the 1980–1999 levels. Note that because of different choices for ecosystem formulations and parameters, the GFDL models have an efficiently recycling microbial loop, inducing that a significant part of production is regenerated (much higher fraction than in the other models). This partly explains why IPP increases in response to warming for those two models.
 These intermodel differences in the response of IPP to climate change are due to the differences in the absolute values of the limiting factors, namely summer ice extent and NO3. As their effect on primary production is opposite, the response of IPP to climate change results from a subtle balance between an increase due to sea ice retreat and a decrease due to decreasing NO3 stocks.
 To illustrate this, two models with extreme and opposite responses were selected (see Figure 5). GFDL-ESM2G has large initial NO3 stocks and maintains NO3 surface concentrations above the alleviated oligotrophic threshold throughout the 21st century (, orange line, Figure 5c). Hence, integrated annual production (green) monotonically increases, following the empirical limitation factor (lim, black), which is primarily driven by the sea ice retreat (limsie, blue).
 In contrast, IPSL-CM5A-MR has much lower initial NO3 levels, which are below the oligotrophic threshold already at the beginning of the 20th century (, orange line, Figure 5d). Hence, the release of light limitation due to sea ice retreat (blue) is counter-balanced by the increase in NO3 limitation (red). In the early 21st century, the effect of sea ice retreat initially dominates, and the limitation factor (lim, black) and primary production (green) both increase. By about 2030, the effect of decreasing nutrients takes over, production decreases, and reaches levels below its 20th century values by about 2050.
 This analysis was performed for all models (see Figure S2), and for each of them, the change in IPP is determined by the balance between the gains due to decreasing sea ice and the losses due to decreasing NO3 stocks. This information is illustrated in a more synthetic fashion in Figure 6, where annual values of normalized integrated production IPP* are plotted versus the limiting factors in several forms. Analysis indicates the following. (i) Above oligotrophy (, Figure 6a), IPP* decreases with , just because sea ice decreases and nutrient increases are simultaneous. As expected, below oligotrophy, IPP* is primarily controlled by and proportional to NO3. (ii) In all models, production positively responds to decreasing summer ice extent (increasing limsie, Figure 6b). When the summer ice extent is too small, it cannot affect IPP*, as shown by the large scatter of points at the right-end of Figure 6c. (iii) Only the combined effect of NO3 and sea ice completely explains the changes in IPP* (Figure 6c). Hence, we argue that changes in IPP can be explained using the combination of nitrate and sea ice limitation factors.
 In this context, it seems that the uncertainty in NO3 rather than the uncertainty in perennial ice extent explains the wide range in ΔIPP. Indeed, oligotrophy is attained (see Tolig in Table 2) occurs as early as 1921 in IPSL-CM5A-LR, in the course of the 21st century for most models, or does not occur at all in the GFDL and MPI models. This range is much larger than the range in the first year where the perennial ice zone extent falls below 106 km2(2021 to 2076, see TSIZ in Table 2).
 The mean model also shows an initial increase in primary production due to ice retreat, followed by a decrease driven by the NO3 progressive depletion. This latter decrease is not sufficient to compensate for the initial sea ice effect, and the net change in IPP is a +58 TgC/yr anomaly by the end of the 21st century (Table 2). Given the range of future IPP projections in the Arctic, the mean model is not as meaningful as the full range of model responses, which well illustrates the role of initial NO3 stocks on the uncertainty in IPP projections.
3.3 Response of Primary Production to Climate Change: Spatial Variations
 The spatial distribution of the end-of-21st century changes in sPP (see Figure 7 and S3) indicates that surface NO3 concentrations decrease in the whole domain and for all models, while sPP tends to consistently decrease in the OZ and to increase in the IZ, with large intermodel variations. As the OZ and IZ seem to be characterized by different responses, their contribution to IPP was split for 1980–1999 and for 2080–2099 (Figure 4).
 By the end of the 20th century, the mean model IZ dominates (78%) IPP. The IZ share is nevertheless smaller than the areal fraction of the IZ (89%), because sPP is more than twice as large (68.9 gC/m2/yr) in the OZ than in the IZ (29.7 gC/m2/yr). This sPP difference is attributable to specific physical characteristics of the two provinces (latitude, the presence of sea ice, and ocean vertical mixing). The share of the PIZ contribution to IPP varies among models and is notably the largest in the GFDL models (25–44%of production in the PIZ), which allow light penetration through sea ice itself.
 By the end of the 21st century, the perennial ice zone typically disappears and the seasonal ice zone slightly decreases (Figure 8b); hence, the share of the IZ to IPP decreases, while the specific primary production slightly increases to 33.4 gC/m2/yr in the IZ (longer open water season) and decreases to 51.3 gC/m2/yr in the OZ (less nutrients).
 To better understand how the OZ and IZ contribute to (2080–2099) to (1980–1999) changes in IPP, their different contributors (TgC/yr) are depicted in Figure 8c for the mean model and detailed in Figure 4 and Table 3 for all models. The numbers in Table 3 are obtained as follows. We first split the domain change in IPP into contributions from the sea ice and open ocean zones:
For any subregion of the domain, IPP can be decomposed into , where Sis the surface of the subregion and is the mean specific primary production over the subregion. Hence, one can further separate ΔIPP into contributions from changing S and :
Using (4) to expand both terms on the right-hand side of equation (3) and using the fact that ΔSIZ=−ΔSOZ, we obtain a second decomposition for ΔIPP:
where the first and second terms correspond to changing specific primary production in the sea ice and open zones, respectively, and the third term is the replacement of ice-covered regions into year-round open ocean.
Table 3. (2080–2099) to (1980–1999) Changes in Integrated Primary Production (TgC/yr) Over the Arctic Region, for Each Model and the Mean Modela
aThis change (ΔIPP) is decomposed into contributions from changes in area and specific primary production over the sea ice and open ocean zones. See section 3.3 for details.
 We first analyze the mean model 2080–2099 increase in IPP with respect to 1980–1999, ΔIPP=58 TgC/yr. Based on (3), we find that ΔIPP results from an OZ gain of 82 TgC/yr that is not fully canceled by an IZ loss of 24 TgC/yr. These 24 TgC/yr result from losses due to a reduced IZ extent (−68 TgC/yr), which dominate the gains due to increasing sPP in the IZ (43 TgC/yr). The 82 TgC/yr contribution of the OZ to ΔIPP result from the gain due to the areal increase of the OZ (129 TgC/yr), dominating the losses due to the decrease in sPP in the OZ (−46 TgC/yr). This mean model behavior is characteristic of most models.
 Using equation (5), these 58 TgC/yr can be summarized as the sum of three processes: (i) an increase in sPPin the IZ (+43 Gt C/yr), likely because sea ice is more seasonal; (ii) a decrease in sPP in the low latitude OZ, typically the North Atlantic (-46 Gt C/year), because the ocean is more oligotrophic; and (iii) the conversion of >2 million km2 of ice-covered regions into about twice as productive permanent open ocean (61 TgC/yr), which is the main contributor. The areas where sea ice disappear yearlong are typically the Greenland and Barents Seas (see Figure 8a).
 All models consistently feature an increase in open ocean IPP due to the increase in open ocean area, and a simultaneous reduction in sPP in the OZ. In the IZ, the models disagree on the sign of the IPP change and predict an increase for some (GFDLs, MPIs) and a decrease for others. Note that the GFDL models are peculiar, because they lose their perennial ice cover in the 21st century, recycle nitrogen more efficiently than the rest of the models, and have under-ice light transmission. In general, the models with increasing IPP in the IZ are the ones for which increases in specific primary production dominates the losses due to the conversion from IZ to OZ. Those models are also the ones that are not oligotrophic (GFDL, MPI). This underlines again that information on nutrient levels in the IZ are crucial.
 We finally come back to the actual spatial distributions of sPP and NO3 changes for 2080–2099 as compared to 1980–1999 (Figure 7). The Barents and Chukchi Seas see the most notable sPP increases. The Barents Sea moves from the IZ to the OZ in most models, while the Chukchi Sea typically remains in the IZ. More importantly, both regions are inflow shelves [Carmack and Wassmann, 2006], in the sense that they receive an external NO3 supply (from Atlantic and Pacific waters, respectively) which can support the sPP increase. NO3 decreases everywhere, particularly in the Arctic Basin and in the Chukchi Sea, and less so in other shelf regions, including Barents Sea. The factors explaining this contrast are: contrasting mixed layer trends in deep basins (decrease) and in the vicinity of the shelf break (increase); and long advective time scales between the Arctic Basin and Pacific, Atlantic (T= 15–20 years) and shelf (T= 5 years) waters sources [Popova et al., 2013]. The NO3 decrease is larger in the Chukchi Sea than in other shelf regions, because of the increasing NO3 uptake associated with the sPP increase.
3.4 Factors Contributing to the NO3 Uncertainties and Trend
 The large intermodel NO3 scatter is present from the beginning of the simulations; therefore, it must be linked to various initialization strategies. NO3 initial fields reflect the model equilibrium state, reached after long (>1000 year) spin-ups simulations.
 Despite the large intermodel scatter, surface NO3 consistently decreases in all models at a mean rate of 2.3 ± 1 mmol/m3/century over the 21st century. This can be due to increased stratification, decreased Pacific and Atlantic NO3 concentrations, weaker Atlantic and Pacific water inflows, and increased N export to depth by sinking particles. Quantifying those processes through a N budget is not feasible with the available CMIP5 data. A quick look at the evolution of the seasonality of NO3 indicates that winter NO3 concentrations are smaller and smaller every year, indicating that nutrient stocks decrease. This can be explained by the decreasing NO3 physical supply, being less and less able to compensate for the biological N uptake and export to depth. In the rest of this section, we show that throughout the 21st century, most models simulate (i) a stratification increase in the entire Arctic region and (ii) a surface NO3 decrease in the North Atlantic and North Pacific that is quite comparable to the Arctic one.
 To quantify stratification, consistently defined mixed-layer depths (hmld) for all models were not available. The best compromise was to use the monthly maximum mixed layer depth from vertical mixing scheme (omlmax in the CMIP5 terminology) for seven models and the mixed layer diagnosed from T and S vertical profiles (mlotst) for the HadGEMs and MIROCs (see Table 2). Since these two definitions are not consistent, the mean model is not really meaningful and was given for indication only.
 Over 1980–1999, simulated mixed layers show an ample seasonal cycle with a March maximum of m (mean model, Arctic Region) and a summer minimum of ∼ 10 m. The simulated is much larger in the OZ (∼600 m) than in the IZ (∼100 m).
 By 2080–2099, is typically 25 m smaller than the 1980–1999 average for the whole model ensemble. In Figure 9, time series of the following stratification index
are plotted for the entire Arctic, the sea ice, and open ocean zones. refers to a 5 year running mean and refers to a regular mean. SI cancels the differences arising from the different definitions of mixed layer. Figure 9 suggests that mixed layer depth decreases over the whole Arctic, and in both, the open ocean and sea ice zones over the 21st century. Notable exceptions are the HadGEM models which have increasing mixed layer depths from ∼ 2030 onward in the sea ice zone. Regionally, mixing is not always decreasing everywhere: it increases in several coastal and shelf areas (not shown), but the regional average is dominated by the mixed layer shallowing in the central basin.
 Increasing stratification contributes to reduce the pre-bloom nutrient stocks. A simple correlation analysis (Table 4) indicates that the variance in annual mean Arctic explains more than 60% of the variance in NO3 variance for seven models over 11. The two largest exceptions are the HadGEM models, which also feature some of the smallest mixed layer depth standard deviations. For the IPSL models, mixed layer variations explain only about 50% of the NO3 changes, but the latter are the smallest in the model sample.
Table 4. Correlation Statistics of Annual Mean Surface NO3 and Seasonal Maximum Mixed Layer Depth (Averaged Over the Whole Domain)a
aComputed from annual mean time series over 1900–2100: standard deviations (, ), raw and squared Pearson correlation coefficient rXY.
 In neighboring regions of the North Atlantic and Pacific Oceans (>50°N), surface NO3 also decreases. The mean model rate of decrease of about 2 mmol/m3/century for both regions is very similar to the 2.3 mmol/m3/century in the Arctic Region. Correlations between the surface Arctic NO3 and corresponding values for the North Atlantic (mean c.c. = 0.89) and the North Pacific (0.87) are >0.8 for every model. This suggests a substantial contribution of the Pacific and Atlantic waters flowing to the Arctic NO3 trend. Changes in the intensity of the ocean circulation could also affect nutrient supply to the Arctic, as the consistent decrease in the Atlantic Meridional Overturning Circulation intensity found throughout the 21st century in 10 CMIP5 models [Cheng et al., 2013].
4 Discussion and Conclusions
 The response of primary production to future climate change in the Arctic Ocean analyzed here is inconsistently simulated by 11 ESMs used in the framework of the CMIP5 exercise. Among these models, three models predict a steady increase in Arctic-integrated primary production, and eight models have an initial increase followed by a decrease in the course of the 21st century. For three models among those eight, the decrease is sufficient to reach, by 2080–2099, integrated primary production levels that are below those simulated for 1980–1999. This result was envisioned by Popova et al. , using five ice-ocean models (none is fully part of the CMIP5 models used here) forced by atmospheric reanalyses and ran over the recent past in the framework of the Arctic Ocean Model Intercomparison Project (AOMIP). This result was also apparent in the previous generation CMIP3 models [Steinacher et al., 2010]. In any case, the potential for Arctic primary production increase (mean model +58 TgC/yr) appears to be for most models much smaller than the ∼300 TgC/yr obtained from a simple linear extrapolation of observed trends [Arrigo and van Dijken, 2011], because limitation by nitrate stocks is playing a primary role.
 Indeed, uncertainties in Arctic future primary production change for the 21st century are primarily driven by the large range in initial NO3 levels. Whereas NO3 consistently decreases over 1900–2100, models reach oligotrophy very early in the 20th century, or not even before 2100, because of this initial NO3 scatter. This scatter directly stems from differences in the equilibrium state of ocean biogeochemical model components, reached after several thousand years of “spin-up” simulations, and used to initialize the CMIP5 simulations. In contrast, the scatter in the time of disappearance of perennial ice does is minor (2021–2076) and does not seem to affect uncertainties. However, the sea ice zone is the largest contributor to Arctic primary production and is where uncertainties in the sign of primary production changes are most visible. In the open ocean zones, virtually, all models agree on which changes to expect in the 21st century: more integrated PP due to wider open ocean zone, replacing less-productive portions of the ice zone; and less PP per unit area due to less nitrate.
 Despite the large uncertainties, the same mechanisms are at play in all models. Models agree on less perennial sea ice (hence, more light) and less nitrate in the course of the 21st century. Why nitrate decreases was not identified. But an associated increase in stratification, and simultaneous nitrate reductions in the North Atlantic and the North Pacific Oceans were identified as potential contributors. Either way, the simulated future physical ocean regime—through mixing, upwelling or eddy supplies—is overall not able to promote recurrent nitrate renewal, compensating for the losses due to increased primary production and export to depth.
 The agreement on the general mechanisms at play gives hope, since reducing the uncertainty on the present observed and simulated nutrient status in the Arctic Ocean, would render projections much more consistent. Nitrate data are scarce in the Arctic, but a few profiles exist and their recent analysis broadly indicates what are the spatial and seasonal patterns in the Arctic [Codispoti et al., 2013]. Future model calibrations should at least use these observations as a baseline evaluation. Collecting an exhaustive Arctic observational NO3 climatology is already beyond reach because of recent changes. However, assessing the levels of pre-bloom present Arctic nitrate stocks is important, and it might still be possible to collect pre-bloom NO3 observations, which do not deviate too far from the 20th century climatology. In this context, a minimal requirement for data collection would be to get pre- and post-bloom NO3 profiles at a few well-chosen locations (or over an Arctic transect) and over a few years. Optical nitrate sensors are promising tools to accomplish this goal [see, e.g., McLaughlin and Carmack, 2010].
 In addition, an evaluation and an improvement of the model mechanisms influencing NO3 and more generally PP in the Arctic are required. The model equilibrium state expresses a combined response to ocean circulation and biogeochemical parameterizations and their tuning. Hence, three natural directions of model improvement seem suitable: (i) a better model representation of ocean circulation and ice distribution; (ii) a better representation of Arctic biogeochemical processes, in particular in the sea ice zone; and (iii) improved calibration of the equilibrium state of the ocean biogeochemical model components. The Arctic Ocean circulation far from being fully covered by observations, and its simulation is challenging. Large intermodel differences in stratification, stratification type, and ocean circulation—which are the main physical NO3 drivers—were found among nine AOMIP models [Holloway et al., 2007].
 The large uncertainties in the sea ice zone suggest that the part of the problem could be linked to sea ice. Some of the expectedly major physical effects of sea ice on ocean biogeochemistry—namely, the input of brine and freshwater due to sea ice growth and melt, and the barrier role of sea ice to wind stirring of the water column—have at least a first-order model representation. However, several sea ice processes relevant to ocean biogeochemistry are neither properly understood nor represented in ESMs [Vancoppenolle et al.2013]. First, sea ice is a host for microbial communities [Thomas and Dieckmann, 2010], and stores nutrients, organic material, and iron [e.g., Lancelot et al., 2009; Vancoppenolle et al., 2010], which are released during sea ice melt. To address that issue, some groups are developing biogeochemical sea ice components for large-scale Earth system studies [e.g., Deal et al., 2011]. Second, the model representation of light transmission through sea ice has generally not been developed for biogeochemistry problems, leading to significant uncertainties in under-ice light fields [see, e.g., Lavoie et al., 2005]. Under-ice light fields are likely a primary factor in the trigerring of under-ice blooms, which importance has recently been re-evaluated [Mundy et al., 2009; Arrigo et al., 2012]. Other Arctic biogeochemical processes in the sea ice zone deserving more consideration from modelers include land-ocean interactions, in particular the large and fast changing riverine inputs [Peterson et al., 2006] spreading on the wide Arctic continental shelves [e.g., McClelland et al., 2007]. Rivers are likely not the primary actor, though, since the riverine contribution of nitrate to new production is very small at regional scales and negligible at the pan-Arctic scale [Le Fouest et al., 2012]. The representation of benthic nutrient cycling is also generally absent from models, which could matter on the wide Arctic shallow shelves.
 The evolution of primary production in the Arctic Ocean may have consequences on important ecosystem services, such as marine resources (i.e., fisheries) and the ocean anthropogenic carbon sink. Estimates of total marine (and estuarine) fish catches in the Arctic Ocean (FAO Area 18) amounts to 10,200 t/yr by the mid-2000s [Zeller et al., 2011]. This is very low when compared to other Large Marine Ecosystem areas of the world ocean, but these resources are an important part of food security in that region and have essential cultural significance. Current expectations are that increased accessibility of Arctic areas, as well as enhanced primary production due to reduced sea-ice cover, could result in increased fish catches in the coming decades [Tremblay et al., 2012]. Our analysis shows that climate change impacts on PP in the Arctic Ocean are quite uncertain and that increased oligotrophy could limit a potential light-driven increase in PP, and hence the existence of any large-scale commercial fisheries. In addition, recent observations suggest that the expected physical (stratification) and chemical (oligotrophy) changes could benefit small picophytoplankton cells at the expense of large diatoms [Li et al., 2009] which would support lower export of biogenic carbon for extraction (e.g., harvest).
 According to the compilation of Bates and Mathis , the Arctic Ocean presently takes up from 66 to 199 TgC/yr from the atmosphere, thus contributing to the ocean anthropogenic carbon uptake. The Arctic Ocean CO2 sink has likely been increasing due to sea ice retreat: from 24 to 66 TgC/yr over 1972–2002 according to observational estimates [Bates et al., 2006], and by ∼10 TgC/yr over 1997–2007 following a model study [McGuire et al., 2010]. How the Arctic Ocean CO2 sink will evolve in the future is not presently clear. On the one hand, potential increase in PP in the Arctic Ocean due to sea ice retreat could drive increases in new production or net community production, and hence a DIC drawdown that should increase the air-sea CO2 disequilibrium (i.e., ΔCO2) and increase the net oceanic uptake of CO2[Bates and Mathis, 2009]. On the other hand, shallow mixed layers, reduced sea ice cover, strong stratification, and relatively small biological production in low nutrient conditions can result in a quick increase of CO2 in Arctic waters, decreasing the air-sea CO2 difference which may limit the future efficiency of the Arctic CO2 sink. Such a setting has been observed in 2008 in the Canadian Basin by Cai et al. . In the end, the future of the Arctic CO2 sink will be determined by the modified seasonality of its driving factors [see, e.g., Steiner et al., 2013], in a complex combination that cannot be easily predicted [Bates and Mathis, 2009]. The inconsistency of the future projections of PP and oligotrophy in the Arctic Ocean shown in this paper casts further doubt upon the sign of the impact of changes in the biological pump on anthropogenic carbon uptake in the coming decades.
 Valuable help from François Massonnet, Pierre-Yves Barriat, Camille Lique, Roland Séférian, Ivan Grozny, Ioulia Nikolskaia, Hideki Okajima, and Michio Kawamiya; and financial support from project BISICLO FP7 CIG 321938 are gratefully acknowledged. We also acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for CMIP5, and we thank the climate modeling groups for producing and making their model output available. For CMIP5, the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. Here the French Portal Prodiguer from the Institut Pierre-Simon Laplace, Paris, France, was used to download CMIP5 output with great efficiency. Three groups are warmly thanked for providing data and associated support: EUMETSAT-OSISAF (sea ice concentration), the National Oceanographic Data Center (World Ocean Atlas nitrate), as well as Kevin Arrigo and Gert van Dijken (Arctic primary production), The two anonymous reviewers and the editor are warmly thanked for their significant contribution to this manuscript.