Spatio-temporal variability of marine primary and export production in three global coupled climate carbon cycle models

This study compares spatial and temporal variability in net primary productivity (PP) and particulate organic carbon (POC) export production (EP) from three different coupled climate carbon cycle models (IPSL, MPIM, NCAR) with observation-based estimates derived from satellite measurements of ocean colour and inverse modelling. Satellite observations of ocean colour have shown that temporal variability of PP on the global scale is largely dominated by the permanently stratified, low-latitude ocean (Behrenfeld et al., 2006)\nocite{Behrenfeld06} with stronger stratification (higher SSTs) leading to negative PP anomalies and vice versa. Results from all three coupled models confirm the role of the low-latitude, permanently stratified ocean for global PP anomalies. Two of the models also reproduce the inverse relationship between stratification (SST) and PP, especially in the equatorial Pacific. With the help of the model results we are able to explain the chain of cause and effect leading from stratification (SST) through nutrient concentrations to PP and finally to EP. There are significant uncertainties in observational PP and especially EP. Our finding of a good agreement between independent estimates from coupled models and satellite observations provides increased confidence that such models can be used as a first basis to estimate the impact of future climate change on marine productivity and carbon export.


Introduction
Marine net primary productivity (PP) is a key process in the global carbon cycle, con-20 trolling the uptake of dissolved inorganic carbon (DIC) in the sunlit surface waters of the ocean and its transfer into organic carbon (OC). Subsequent gravitational sinking of detrital particulate organic carbon (POC) through the water column results in the export of POC (EP) from the surface into the ocean's interior, where it becomes partly or entirely remineralised and eventually transported back to the surface. The export of A quantitative understanding of the processes that control PP and EP and their implementation in coupled climate biogeochemical models is essential to project the effect of future climate change on marine productivity, carbon export fluxes (Bopp et al., 2005;Plattner et al., 2001;Joos et al.,1999;Maier-Reimer et al., 1996) and their possible feedbacks on the climate system (Friedlingstein et al., 2006). Unfortunately, produc-15 tivity and export are not well constrained by observations, making it difficult to validate corresponding results from climate models. Productivity estimates from coupled models and satellite observations are to a large degree independent in construction, and cross-comparison of the two approaches provides a promising way to assess jointly their overall skill and identify the main underlying mechanisms that control PP and EP 20 variability. To do so, this study investigates results from three different fully coupled climate carbon cycle models (IPSL, MPIM, NCAR) that include interactions between the atmosphere, ocean circulation and sea-ice, marine biogeochemical cycles and the terrestrial biosphere. As all three models differ in their major components (atmosphere, ocean, terrestrial and marine biospheres), the aim of this study is to give a description 25 of the present day situation of marine PP and EP as simulated by coupled models. Note that it would not be sufficient to investigate the primary and export production in a (far cheaper) set of forced ocean-only model experiments, since climate in the fully coupled models (e.g. ocean currents and resulting nutrient distributions, cloud cover data are available, the latter is rather a comparison between different models.

Models
All models used in this study consist of fully coupled 3-D atmosphere-ocean climate models that have also contributed to the IPCC Fourth Assessment Report (AR4; Solomon et al., 2007;Meehl et al., 2007). The models were run with additional carbon 10 cycle modules for the terrestrial and oceanic biospheres (Friedlingstein et al., 2006).

IPSL
The IPSL-CM4-LOOP (IPSL) model consists of the Laboratoire de Météorologie Dynamique atmospheric model (LMDZ-4) with a horizontal resolution of about 3 • ×3 • and 19 vertical levels (Hourdin et al., 2006), coupled to the OPA-8 ocean model with a 15 horizontal resolution of 2 • ×2 • ·cosφ and 31 vertical levels and the LIM sea ice model (Madec et al., 1998). The terrestrial biosphere is represented by the global vegetation model ORCHIDEE (Krinner et al., 2005), the marine carbon cycle by the PISCES model (Aumont et al., 2003). PISCES simulates the cycling of carbon, oxygen and the major nutrients determining phytoplankton growth (PO 4 , NO 3 , NH 4 , Si, Fe). Phy-20 toplankton growth is limited by the availability of nutrients, temperature and light. The model has two phytoplankton size classes (small and large), representing nanophytoplankton and diatoms, as well as two zooplankton size classes (small and large), representing microzooplankton and mesozooplankton. For all species the C:N:P ratios are assumed constant (122:16:1), while the internal ratios of Fe:C, Chl:C and Si:C Introduction EGU of organic carbon in the model: semi-labile dissolved organic carbon (DOC), with a lifetime of several weeks to years, as well as large and small detrital particles, which are fueled by mortality, aggregation, fecal pellet production and grazing. Small detrital particles sink through the water column with a constant sinking speed of 3 m/day, while for large particles the sinking speed increases with depth from a value of 50 m/day 5 at the depth of the mixed layer, increasing to a maximum sinking speed of 425 m/day at 5000 m depth. For a more detailed description of the PISCES model see Amount and Bopp (2006) and Gehlen et al. (2006). Further details and results from the fully coupled model simulation of the IPSL-CM4-LOOP model are given in Friedlingstein et al. (2006).

MPIM
The Earth System Model employed at the Max-Planck-Institut für Meteorologie (MPIM) consists of the ECHAM5 (Roeckner et al., 2006) atmospheric model in T63L31 resolution with embedded JSBACH terrestrial biosphere model and the MPIOM physical ocean model, which further includes a sea-ice model (Marsland et al., 2003) and the 15 HAMOCC5.1 marine biogeochemistry model (Maier-Reimer et al., 2005). The coupling of the marine and atmospheric model components, and in particular the carbon cycles is achieved by using the OASIS coupler. HAMOCC5.1 is implemented into the MPIOM physical ocean model (Marsland et al., 2003) using a curvilinear coordinate system with a 1.5 • nominal resolution where 20 the North Pole is placed over Greenland, thus providing relatively high horizontal resolution in the Nordic Seas. The vertical resolution is 40 layers, with higher resolution in the upper part of the water column (10 m at the surface to 13 m at 90 m). The marine biogeochemical model HAMOCC5.1 is designed to address large scale/long term features of the marine carbon cycle, rather than to give a complete description of the 25 marine ecosystem. Consequently, HAMOCC5.1 is a NPZD model with two phytoplankton types (opal and calcite producers) and one zooplankton species. The carbonate chemistry is identical to the one described in Maier-Reimer (1993 EGU description of HAMOCC5.1 can be found in Maier-Reimer et al. (2005), while here only the main features relevant for the described experiments will be outlined. Marine biological production is limited by the availability of phosphorous, nitrate and iron. Silicate concentrations are used to distinguish the growth of diatoms and coccolithophorides: if silicate is abundant, diatoms grow first, thereby reducing the amount of 5 nutrients available for coccolithophoride growth. The production of calcium carbonate shells occurs in a fixed ratio of the phytoplankton growth (0.2). The model also includes cyanobacteria that take up nitrogen from the atmosphere and transform it immediately into nitrate. Please note that biological production is temperature-independent, assuming that phytoplankton acclimate to local conditions. Global dust deposition fields are used to define the source function of bioavailable iron. Removal of dissolved iron occurs through biological uptake and export and by scavenging, which is described as a relaxation to the deep ocean iron concentration of 0.6 nmol/m 3 . In the experiments used here, export of particulate matter is simulated using prescribed settling velocities for opal (30 m/day), calcite shells (30 m/day) and organic carbon (10 m/day). Reminer-15 alisation of organic matter depends on the availability of oxygen. In anoxic regions, remineralisation using oxygen from denitrification takes place. HAMOCC5.1 also includes an interactive module to describe the sediment flux at the sea floor. This component simulates pore water chemistry, the solid sediment fraction and interactions between the sediment and the oceanic bottom layer as well 20 as between solid sediment and pore water constituents.

NCAR
The physical core of the NCAR CSM1.4 carbon climate model (Doney et al., 2006;Fung et al., 2005) is a modified version of the NCAR CSM1.4 coupled physical model, consisting of ocean, atmosphere, land and sea-ice components integrated via a flux 25 coupler without flux adjustments (Bolville et al., 2001;Bolville and Gent, 1998). The atmospheric model CCM3 is run with a spectral truncation resolution of 3.75 • (T31) and 18 levels in the vertical (Kiehl et al., 1998)

EGU
Model (NCOM) with 25 levels in the vertical and a resolution of 3.6 • in longitude and 0.8 • to 1.8 • in latitude (T31×3) . The water cycle is closed through a river runoff scheme, and modifications have been made to the ocean horizontal and vertical diffusivity's and viscosity's from the original version (CSM1.0) to improve the equatorial ocean circulation and interannual variability. The sea ice component model runs at the same resolution as the ocean model, and the land surface model runs at the same resolution as the atmospheric model. The CSM1.4-carbon model includes a modified version of the terrestrial biogeochemistry model CASA (Carnegie-Ames-Stanford Approach) (Randerson et al., 1997), and a derivate of the OCMIP-2 (Ocean Carbon-Cycle Model Intercomparison Project Phase 2) oceanic biogeochemistry model (see Najjar, R. G. and Orr, J. C., 1999, unpublished manuscript, OCMIP-2 Biotic-HOWTO; available online at http://www.ipsl. jussieu.fr/OCMIP/). In the ocean model the biological source-sink term has been changed from a nutrient restoring formulation to a prognostic formulation, thus biological productivity is modulated by temperature, surface solar irradiance, mixed layer 15 depth, and macro-and micronutrients. Following the OCMIP-2 protocols (Najjar et al., 2007), total biological productivity is partitioned 1/3 into sinking POC flux, equivalent to EP, and 2/3 into the formation of dissolved or suspended organic matter, much of which is remineralised within the model euphotic zone. Total productivity thus contains both new and regenerated production, though the regenerated contribution is probably 20 lower than in the real ocean. While not strictly equivalent to primary production as measured by 14 C methods, rather net nutrient uptake, NCAR PP is a reasonable proxy for the time and space variability of PP if somewhat underestimating the absolute magnitude. Please note that in this study for reasons of simplicity net nutrient uptake for NCAR is labelled and treated as PP, even though it is not essentially the same. The 25 ocean biogeochemical model includes in simplified form the main processes of the organic and inorganic carbon cycle within the ocean, and air-sea CO 2 flux. A parameterisation of the marine iron cycle has also been introduced (Doney et al., 2006). It includes atmospheric dust deposition/iron dissolution, biological uptake, vertical parti-Introduction

Experiments
All results of the current study are obtained from simulations with the coupled climate carbon cycle models explained above (IPSL, MPIM, NCAR), that simulate fully coupled interactions between the atmosphere, ocean circulation, sea-ice, marine biogeochemical cycles and terrestrial biosphere. The models are a subset of the models that 10 contributed to the C4MIP project (Friedlingstein et al.,2006) and follow the C4MIP protocols. Two of the models (IPSL, MPIM) are only forced by the historical development of anthropogenic CO 2 emissions due to fossil fuel burning and land-use changes from preindustrial to present and the A2 scenario from the year 2000 on. In addition to that, NCAR is also forced by CH 4 , N 2 O, CFCs, volcanic emissions and changes in so-   4,2007 Marine productivity in coupled climate models

Ocean circulation and biogeochemical cycling
In a Taylor diagram, all three models show good agreement for 3-D global annual mean distributions of temperature ( Fig. 1). Global SST distributions including the seasonal cycle are even better reproduced, reaching correlation coefficients above R=0.98 and 5 normalised standard deviations very close to 1. Salinity distributions are known to be less well reproduced by coupled climate models, which is largely attributed to deficiencies in the hydrological cycle of the atmosphere models (Schneider et al., 2007). This is also the case for the models of the current study. However, correlation coefficients for both the global annual mean 3-D salinity distributions and the SSS values including 10 the seasonal cycle are still on the order of R=0.78-0.90. In contrast to temperature, the agreement between modelled and observed salinity distributions is poorer when going from the global annual mean 3-D field to seasonal surface water values, which is due to a stronger influence of possible misfits from the atmospheric hydrological cycle, as mentioned above. While correlation coefficients remain similar, the normalised 15 standard deviations are considerably higher than observed. Generally, in terms of T and S all three models of the present study perform at least as well as the majority of models that contributed to the IPCC AR4 (Meehl et al., 2007;Schneider et al., 2007). The mixed layer depth (MLD) is a dynamically important variable for upper ocean water mass transport (Gnanadesikan et al., 2002), and especially the maximum (MLD max ) 20 during winter time strongly affects the formation of mode, intermediate, deep and bottom waters. Furthermore, ocean mixing plays an important role in surface ocean nutrient availability and thus biological productivity (Najjar et al., 2007). Observation-based estimates of MLD, however, are uncertain, and different methods to determine MLD yield different results for both models and observations (Boyer-Montégut et al., 2004). 25 MPIM strongly overestimates MLD max in the Southern Ocean, while the other models show too shallow mixing in this area as compared to climatological MLD observations (Boyer-Montégut et al., 2004;Fig. 2   • N the models overestimate zonal average MLD max up to a factor of two. Spatial correlations of modelled and observed MLD max are only moderate and on the order of R=0.3-0.4. The tendency towards too deep mixing is expressed by high ratios of normalised standard deviations (std mod /std obs ) in a Taylor-diagram ( Fig. 1). For MPIM 5 the overestimation of MLD max in the Southern Ocean is known to be caused by too little ice cover in the Weddell Sea (Marsland et al., 2003). This misfit leads to a normalised standard deviation above 6.5 in a Taylor diagram (not shown). However, when applying the definition of MLD as used in Schneider et al. (2007) to temperature and salinity data from the models and the climatological observations (Conkright et al., 2002) the result-10 ing correspondence of MLD max distributions is not considerably better (not shown).
A further approach to assess the modelled ocean circulation fields is a comparison of modelled and observed water mass transports (Schneider et al., 2007;Schmittner et al., 2006;Large et al., 1997). The Atlantic Meridional Overturning Circulation (AMOC) is an important measure for ocean circulation, quantifying the amount of deep waters 15 formed in the North Atlantic, which is of high climatic relevance for the global redistribution of heat and energy by the ocean. In the current study AMOC is defined as the maximum meridional water mass transport below 300 m depth in the North Atlantic, given in Sverdrup (1 Sv=10 6 m 3 s −1 ). In MPIM the simulated water mass transport lies well within the range of equally defined observation-based estimates for the 1990ies, 20 whereas in IPSL the AMOC is low and in NCAR high, but within the error-bars of one standard deviation both also match the observed range (IPSL: 11.4±1.3 Sv; MPIM: 18.1±0.9 Sv; NCAR: 23.6±1.0 Sv, OBS: 11.5-22.9 Sv; Smethie and Fine, 2001;Talley et al., 2003).
The fully coupled carbon-cycle climate models used in the current study generate 25 their own internal climate variability including coupled ocean-atmosphere modes such as El Nino Southern Oscillation (ENSO). To assess the models' capability to reproduce ENSO variability, the power spectra of SSTs averaged over the Nino3 Box (150 • W-90 • W, 5 • S-5 • N) are calculated (Fig. 3). For IPSL and NCAR there are maxima of the BGD 4, 2007 Marine productivity in coupled climate models EGU power spectra between two and seven years, which is the typical range of El Nino frequency (Randall et al., 2007), while MPIM does not show a maximum, but also strongly increasing power within this range. This may still be reasonable, as this analysis was performed over a 20 year time-period, which may be too short to exhibit representative ENSO dynamics.

5
Phosphate concentrations (PO 4 ) are a good indicator for biogeochemical cycling in the models, as PO 4 is not affected by air-sea gas exchange, which is treated differently in the models. Furthermore, in the current study PO 4 is a fully prognostic tracer in contrast to the models contributing to the OCMIP-2 study, where surface PO 4 was restored (Najjar et al., 2007). The comparison with climatological PO 4 values (Conkright et al., 2002) reveals a good reproduction of annual mean PO 4 fields by all models with correlation coefficients around R∼0.85. IPSL and MPIM perform especially very well with normalised standard deviations close to 1 (Fig. 1). For all models the agreement gets weaker when going from the annual mean 3-D distribution of PO 4 to surface water values including the seasonal cycle. Maps of surface water PO 4 concentrations, av-15 eraged over the top 0-100 m, show that all models roughly match the general pattern of observed PO 4 distributions with lower values in the oligotrophic subtropical gyres and higher values in the high latitudes, particularly in the Southern Ocean, and also in the equatorial and continental upwelling regions (Fig. 4). MPIM underestimates the surface water PO 4 gradient from high towards low latitudes. The NCAR model over- 20 estimates surface PO 4 in the tropical Pacific because excessive iron limitation overly reduces biological phosphate uptake (Fig. 5).
In all models vertically integrated PO 4 concentrations (not shown) are low in the North Atlantic and increase along the conveyor-belt circulation pathway, ending up with a high PO 4 inventory in the North Pacific (Sarmiento et al., 2003). However, all models Interactive Discussion EGU formation. The factors that limit phytoplankton growth in the models are computed from Michaelis-Menten equations for nutrient concentrations, according to MM=N/(N+K), whereby N is the nutrient concentration (PO 4 , NO 3 , iron and silicate) and K is the respective half-saturation constant as used in each model. The nutrient yielding the 5 lowest MM-value is considered to be the most limiting. Other factors (temperature and light) are assumed to be limiting when MM of the limiting nutrient is still above 0.7. In Fig. 5 the limiting factors are shown when the respective variable is limiting phytoplankton growth during at least one month of the year, showing that in some regions different variables may limit productivity at different times of the year. In the MPIM 10 model iron limitation is even stronger than in NCAR and occurs almost everywhere. In IPSL iron is mostly limiting in the equatorial east Pacific, the Southern Ocean and the North Pacific, which are known to be high nitrate low chlorophyll areas (HNLC) that are strongly iron limited. Furthermore, in IPSL NO 3 is the most limiting factor in the intermediate to low latitudes. In the high latitudes also temperature and light are important 15 for nanophytoplankton growth and diatom production is also limited by the availability of silicate.
In summary, all models show a reasonable agreement with observations of temperature and salinity fields, MLD max , PO 4 concentrations and also with estimates of water mass transports, giving confidence in the models that the simulated circulation fields 20 and biogeochemical cycles are reasonable and comparable to those of other state-ofthe-art coupled climate models (Meehl et al., 2007;Randall et al., 2007).
3.2 Annual mean and seasonal cycle of primary productivity and export production Vertical integrals of PP have been computed over the entire depth of the water column in order to compare with estimates from satellite observations of ocean colour and to 25 derive global integrals for PP (

EGU
(MPIM) to 31 GtC/yr (IPSL), and is considerably lower than satellite-based estimates of around 48 GtC/yr (Behrenfeld et al., 2006). Carr et al. (2006) reported a factor of two between global integrals for PP (35-70 GtC/yr) obtained from different algorithms that derive PP from ocean colour, indicating that there is still some uncertainty among those methods. PP from the coupled models is still lower than the low-end satellite 5 estimates. Despite their large spread the different PP algorithms for the observations show patterns of spatial and temporal variability that are similar between different approaches (Carr et al., 2006) and can therefore be used for comparison with results from climate models. The spatial distributions of modelled PP agree moderately well with the pattern of 10 observation-based PP (Fig. 6). In general, satellite-based and model PP are high in the equatorial upwelling regime, especially the equatorial east Pacific and the equatorial east Atlantic. PP is also elevated in the North Atlantic and the Southern Ocean north of the polar front, while it is low in the subtropical oligotrophic gyres. High PP values as derived from observations for the coastal zones are strongly underestimated 15 by all three models, which is a known fact from coarse resolution models. This underestimation of modelled PP in the coastal regions is mainly responsible for the lower than observed global annual PP. For MPIM and NCAR the fairly low PP suggests that iron limitation in the models is too strong (Fig. 5). In the MPIM model this causes too high PP in the low-latitude Atlantic and too little elsewhere. Furthermore, dust input in 20 MPIM is at the lower end of observed values (Timmreck and Schulz, 2004), leading to low iron deposition from the atmosphere, which further increases the iron stress for marine productivity. The lower PP values in the NCAR model also reflect the fact that the calculated productivity is intermediate between new and primary production, which also partly explains the lower than observed PP values (Fig. 6). 25 The seasonal variability of PP is shown in Hovmöller diagrams with the zonally averaged PP plotted versus time from observations and models (Fig. 6). Areas of highest productivity are observed in the North Atlantic, where also seasonal variability is highest. In the oligotrophic gyres PP is low throughout the year and variability is also tension of the oligotrophic gyres. Also in the Southern Ocean PP is overestimated and extending farther southwards than in the observations. In MPIM the periods of high productivity in the higher latitudes are very short, indicating that here PP occurs in very short (strong) pulses, while in the equatorial region PP is higher than the observations with some moderate seasonal variability. The pattern of spatio-temporal variability in 10 NCAR is very similar to the one from the observations, but like in IPSL, PP in the intermediate southern latitudes is biased high and the area of oligotrophic gyres is reduced. The latter is a signal which is arises in the Atlantic, while in the Pacific the expansion of the gyres is too large (Fig. 6). As mentioned above, the derivation of PP from satellite data entails large uncer-15 tainties, as it is a stepwise approach from measurements of ocean colour over the determination of chlorophyll concentrations (and sometimes organic carbon biomass) to PP. While the first steps consider concentrations, PP is a time-dependent rate, requiring even more complicated assumptions about the underlying mechanisms. With each further step from ocean colour to PP the errors propagate, leading to consider-20 able uncertainty in the value of global annual PP, reflected in the factor of 2 spread in the study of Carr et al. (2006). For IPSL, where chlorophyll concentrations are available, a comparison of modelled and satellite-based chlorophyll distributions shows a better agreement than PP values (Fig. 7). Although the model slightly underestimates the surface Chl concentrations, especially in high (northern) latitudes and around the 25 equator, the pattern and the order of magnitude of spatial variability for Chl are slightly closer to observations than those obtained for PP estimates (Fig. 8). Export Production (EP) describes the amount of particulate organic carbon (POC) that is transported from the surface ocean to depth across a certain depth level (IPSL: 4,2007 Marine productivity in coupled climate models EGU 100 m; MPIM: 90 m; NCAR: 75 m). In this study only the material settling through gravitational sinking is regarded, while the total export also encompasses subduction and mixing of suspended particles and dissolved organic matter (DOM) due to water mass transports. The fields of annual mean EP have similar patterns to those of PP for all three models (not shown), i.e. as a first approximation those areas with higher PP 5 also have higher EP. In NCAR this relation is prescribed by using a fixed ratio (1/3) of PP that is exported as POC, while for the other models the amount of EP and its ratio to PP is less straightforward, depending on particle flux and food-web dynamics. The global annual rates of modelled EP are given in Table 1, ranging from 5 GtC/yr (MPIM) to 9 GtC/yr (NCAR), being thus below observation-based estimates reaching from 11

Highest EP in IPSL and NCAR occurs in the intermediate to high latitudes between 40-60
• North and South, respectively (Fig. 9). A secondary maximum is situated around the equator, where upwelling of nutrient rich deep waters permits high PP (Fig. 6) and thus high EP. Between those areas, in the latitude band of the oligotrophic 15 subtropical gyres (∼10-40 • North and South) EP is low. In MPIM, zonal average EP increases from low values in the high latitudes to its maximum around the equator, however, in general EP seems to be underestimated everywhere except for the equatorial region if compared to the other models and the observation-based estimates. This distribution will also be caused by the strong iron stress due to too little atmospheric iron 20 deposition (Fig. 5).
In a Taylor diagram the correlation coefficients for the spatial distributions of annual mean EP from the models are rather poor with highest values for IPSL (R=0.35). These correlations are lower than those determined for PP (Fig. 8). One problem for the comparison of different EP estimates are the different reference levels, depending on the 25 definition of EP, as mentioned above. Consequently, the shallowest estimate (NCAR, 75 m) should give highest values and the deepest estimate (Schlitzer, 133 m) lowest export fluxes, which, however, is not the case (Table 1). Another probable reason for the mismatch between modelled and observation-based EP is the large uncertainty of BGD 4,2007 Marine productivity in coupled climate models EGU observation-based estimates for EP. Similar to the PP estimates, they also vary by a factor of two, from global annual mean values of 11 to 22 GtC/yr. Furthermore, the two estimates of Schlitzer (2000) and Laws et al. (2000), which have about the same amount of EP (11 GtC/yr), show very different spatial distributions (Fig. 9) and also spatial correlations between the two are not better than those from comparison with 5 modelled EP (Fig. 8).

Interannual variability
Primary production is known to be sensitive to climate impacts like for example El Nino Southern Oscillation (ENSO). Consequently, during an El Nino period PP in the tropical Pacific is reduced whereas at a La Nina situation it is enhanced (Behrenfeld et 10 al., 2001;Chavez et al., 1999). From satellite observations it has also been shown that the global signal of interannual PP variability can largely be attributed to the permanently stratified, low-latitude oceans (Behrenfeld et al., 2006). What is more, such anomalies are highly correlated with shifts in the climate system in a way that stronger stratification and the resulting surface ocean warming, which correspond to more El

15
Nino-like conditions, lead to negative PP and chlorophyll anomalies globally over much of the tropics and subtropics. At first sight this relation seems to be counter-intuitive, as phytoplankton growth is positively related to rising temperatures (Eppley, 1972). On the other hand, stronger stratification results in less nutrient supply from deep waters, which in turn limits phytoplankton growth. Algorithms that derive PP from ocean 20 colour make only indirect assumptions about nutrient concentrations via the temperature effect on PP. More detailed answers regardinig the links between stratification, temperature, nutrients and biological productivity can be provided by ocean biogeochemical models. All three models examined in the current study show exactly the same behaviour which was described by Behrenfeld et al. (2006) with the global signal 25 of PP anomalies largely controlled by the permanently stratified low-latitude oceans that have annual average SSTs above 15 • C (Fig. 10 EGU permanently stratified ocean (PP strat ) to global PP (PP glob ) approximates the fraction of the stratified areas to the global ocean surface area, while the anomalies of PP glob and PP strat are positively correlated (Table 1), which highlights the dominant role of the low-latitude ocean in setting the global signal of PP variability. Observation-based PP strat anomalies are negatively correlated with changes in strat-5 ification and SST over the same area ( Fig. 10; Behrenfeld et al., 2006). Note that in interpreting Fig. 10, one should focus on the magnitude and frequency of the PP variability, not the phasing of specific PP events. Because the models are fully coupled, they each generate their own unique internal climate variability that can only statistically be compared with other models and observations. For the model results, a stratification 10 index (SI) was calculated as by Behrenfeld et al. (2006) as the density difference between the sea surface and 200 m water depth. In the IPSL model, PP integrated over the entire domain of the low-latitude, stratified water is strongly anti-correlated with SST and SI. In the NCAR simulation, the biological-physical anti-correlations are more apparent when the analysis is restricted to the area of the Nino3 Box (150-90 • W, 5 • S- Fig. 10). For MPIM no correlation can be found between the respective anomalies of PP and stratification (SI) or SST, neither for the whole low-latitude, permanently stratified domain nor for any other individual sub region. This may be due to the fact that in MPIM the equatorial Atlantic ocean is dominated by strongly oscillating predator-prey cycles, which lead to phase shifts in phytoplankton and zooplankton 20 abundance. This can also explain the high amplitude and frequency of interannual variability of PP simulated by MPIM (Fig. 10). The slopes that can be derived from the anomalies of PP strat versus stratification (SSLOPE) and SST (TSLOPE) are very similar in IPSL and the observation-based estimates (Fig. 11). In the satellite observations for the area of the permanently strat- 25 ified oceans there is a PP anomaly of -876 TgC/month per unit stratification increase (kg m −3 ) and a decrease of -151 TgC/month per degree SST increase (Table 1). For IPSL SSLOPE is slightly weaker (-787 TgC/month) and TSLOPE somewhat stronger  EGU the entire low-latitude, stratified domain, but considerably higher correlations are found for the region of the Nino3 Box (Table 2). This area is, however, too small to explain the bulk of the PP variability for the permanently stratified ocean, which governs the global signal in the NCAR simulations. These results indicate that PP variability in NCAR and MPIM, where no good relationship between PP and climate could be detected, 5 acts differently in different areas, summing up to the common signal of the low-latitude, stratified ocean. Mechanisms for PP variability in these two models may be different or different sub regions of the low-latitude stratified ocean may dominate the signal at different times. A further step is to examine the mechanisms behind the climate-productivity cor-10 relations as found in IPSL for the entire low-latitude ocean domain and in NCAR for the tropical East Pacific. As mentioned before, one might expected that PP increases under higher SSTs due to the temperature sensitivity of phytoplankton growh rates (Eppley, 1972). To find out which other mechanisms contribute to PP variability, maps of cross correlations of different anomalies averaged over the whole region (Fig. 12) versus the local PP variability are drawn, highlighting those regions in the models where PP reacts most sensitive to changes in other variables like stratification, SST and surface water nutrient concentrations. Here, the correlation coefficient R 2 is multiplied by the sign of the regression slope to show next to its strength also the direction of the correlation (positive or negative). To isolate the large scale climate impact on productiv-20 ity "anomalies" of temperature, stratification and nutrient concentrations are averaged over the whole low-latitude, permanently stratified ocean, rather than showing the local signal on each grid-point. The pattern found for anomalies of PP and stratification (SI) in the IPSL model is very similar to the one seen in the observation-based estimates with strongest anti-correlations (negative R 2 -values) occurring at the borders of the EGU one would assume to find also a strong impact of iron variability on PP towards the center of the high productivity tongue. The absence of such a signal in IPSL, however, may be caused by the formulation of the iron cycle in IPSL, where iron concentrations are restored to a minimum value of 0.01 nM l −1 . By doing so, the natural variability of iron is suppressed, strongly dampening the signal that otherwise would be transferred 5 into PP variability. Furthermore, as iron is strongly scavenged by sinking particles in the surface ocean, the nutricline for iron is deeper than the one for NO 3 , so that for iron stronger upwelling is needed to efficiently resupply the surface ocean with iron from below. In NCAR, the highest anti-correlations for PP and SI anomalies are also found in the equatorial Pacific (Fig. 12). Especially for the area of the Nino3 Box higher cor-10 relation coefficients between PP and SI/SST than for the entire low-latitude, stratified domain have already been mentioned above. The respective correlation coefficients are given in Table 2, showing also that in NCAR PP in the area of the Nino3 Box is mainly controlled by iron availability, while there is no correlation to PO 4 concentrations. In MPIM there are no considerable correlations between anomalies of PP and 15 PO 4 concentrations, which can be attributed to the effect of too strong iron stress in the model.
From the correlations and the signs of the regression slopes of the variability of PP and SI (stratification index) versus different nutrient concentration anomalies, the chain of cause and effect from climate impacts to nutrient concentrations and further to PP 20 and finally to EP, may be tentatively reconstructed (Table 2). Accordingly, in IPSL the global signal of PP variability is well explained by the behaviour of the entire permanently stratified, low-latitude ocean. Even though in some cases there are moderate correlations between PP and iron variability the slope is negative, which means in the opposite direction than expected (Table 2). This shows the effect of restoring the iron 25 concentrations on hiding the potential impact of iron limitation on PP variability. Instead, PP variability is dominated by NO 3 and PO 4 (Fig. 12, Table 2). Both nutrients show high correlations with climate indicators like SI and SST as well as with PP for the entire stratified, low-latitude domain and also for the Nino3 Box only ( EGU PP variability is largely independent of PO 4 and other nutrients (Fig. 12), even though the general feature that global PP variability is controlled by the low-latitude oceans is also very well reproduced (Fig. 10). A similar behaviour is found for NCAR, where the North Atlantic also contributes considerably to PP variability (Fig. 6), but this region alone is also not able to explain large parts of the global PP variability. Iron is the main 5 limiting nutrient for productivity in the low latitudes of the NCAR model (Fig. 5), and especially in the area of the Nino3 Box there are high correlation coefficients between stratification, SST and iron variability and further on for iron and PP variability ( Table 2). One reason for the absence of a higher correlation between the anomalies of PP and nutrients of the whole low-latitude area maybe the fact that in the NCAR model PP 10 stands for net nutrient uptake, which includes processes like grazing and heterotrophic respiration.
The pattern of the sensitivity of PP to climate is transferred into EP variability. EP reacts most sensitive to changes in PP in those areas, where PP reacts most to climate variability (Fig. 13). In IPSL, variability in EP is strongly correlated with the average PP 15 variability at the borders of the high productivity tongue in the equatorial Pacific, a pattern which is also reproduced by the EP correlations with NO 3 anomalies (Fig. 13). In MPIM there are no considerable correlations between EP and PP or other climate and nutrient anomalies when regarding the entire low-latitude ocean. In the NCAR model, EP shows the strongest correlation with PP variability in the North Atlantic, where PP 20 (and thus EP) exhibits the strongest interannual variability. Correlation coefficients different from 1 can be found, even though EP is fixed to be 1/3 of PP, as local EP variability is correlated with PP variability averaged over the entire low-latitude, permanently stratified domain. For the global integrals of PP and EP, however, the correlation is 1 (Table 2). Surprisingly, neither IPSL nor NCAR show considerable correlations of Introduction

Conclusions
The current study has shown that while it is still difficult to reproduce absolute values of observation-based annual mean PP with ocean biogeochemical models, both the mechanisms and the amount of interannual variability are well captured by the models. In good agreement with recent satellite observations, the global signal of modelled 5 PP variability is largely dominated by the low-latitude, permanently stratified ocean (Behrenfeld et al., 2006). This result is very robust across the models, even though they perform differently in representing biogeochemical cycles (Fig. 5). The finding of the dominating role of PP in the low-latitude ocean, however, has also be regarded with some care, as both 3-D models and satellite observations may be biased towards lower 10 latitudes. High-latitude satellite-derived PP values especially for the Southern Ocean are less well constrained than those for the low-latitudes due to cloud coverage and the poorer availability of reference measurements. The 3-D coupled models are sometimes tuned to reproduce climate and biogeochemical variability in the low latitudes, while in the high latitudes difficulties with sea-ice interactions may occur.

15
In particular, in this study we are able to demonstrate that under more El Nino-like conditions with surface ocean warming and stronger stratification in the low-latitude ocean, nutrient supply to the surface ocean is reduced, and results in lower PP and EP. In a future warming climate with permanent El Nino like conditions both PP and EP probably will be reduced, which has already been shown by other climate model 20 simulations (Sarmiento et al., 2004;Boyd and Doney, 2002;Bopp et al., 2001). TS-LOPE was determined in the present study from observation-based estimates and the IPSL model to quantify PP sensitivity to SST ( Table 2). As the model simulations were run until the year 2100 using the A2 scenario, results are available to check whether TSLOPE also holds for the future impact of climate change on PP in the low-latitude 25 ocean. Accordingly, one would expect PP (EP) in the low-latitude ocean to decrease by 3 GtC/yr (0.6 GtC/yr) per 1 • C temperature increase. For this particular area the IPSL model predicts an increase of the average SST of 1.6 • C from the years 1985-2005 until BGD 4, 2007 Marine productivity in coupled climate models EGU 2090-2099. This should lead to a decrease in PP (EP) by 4.8 GtC/yr (1 GtC/yr), which amounts to a decline of 27%. However, the model predicts PP over the same time period and area to reduce only by 1.6 GtC/yr (-9%) and EP by 0.7 GtC/yr (-19%). This shows that TSLOPE, which was derived from modelled interannual variability during the two decades between 1985 and 2005, can not be used for extrapolation into future 5 climate conditions. Consequently, to estimate the impact of future climate change on marine productivity further mechanisms and, of course, also the high latitude ocean will have to be considered, which will be done in a complementary study that analyses the continuation of the here presented model simulations in a scenario of future climate change (SRES A2) until the year 2100. 10 In summary, the current study has illustrated a strong link between marine productivity and climate variability, both derived from satellite observations, with coupled climate carbon cycle modelling. On the one hand observation-based estimates have been used to constrain model results for ocean circulation and biogeochemical cycling, while on the other hand results from the climate models have been used to interpret and re-15 construct the mechanisms that underly climate-driven shifts in marine productivity that have been observed. The strong agreement between the areas and mechanisms that dominate PP variability on the global scale for both types of investigation increases confidence in results obtained by both models and observation-based estimates, and it demonstrates the power and importance of model data intercomparisons on the way 20 to better understand the links between climate and marine biogeochemical cycles.  4,2007 Marine productivity in coupled climate models BGD 4,2007 Marine productivity in coupled climate models Depth integrated net primary production (PP) and Export of POC (EP) for the different models and observation-based estimates. PP glob : global annual PP; PP strat : global annual PP in the area of the permanently stratified, low-latitude ocean; Area strat : percentage of the permanently stratified, low-latitude ocean to the global ocean domain; R 2 PP ano correlation coefficient for anomalies of PP glob versus anomalies of PP strat ; SSLOPE P P : slope of the regression line for PP ano (=difference of monthly PP value to the climatological mean of PP) and SI ano (=difference of monthly stratification value to climatological mean stratification); R 2 SSLOPE P P : correlation coefficient for PP ano and SI ano ; TSLOPE P P : slope of the regression line for PP ano and SST ano ; R 2 TSLOPE P P : correlation coefficient for PP ano and SST ano . Notation for EP correlations is equivalent. Values in brackets correspond to one standard deviation.  Table 2. Correlation coefficients (R 2 ) and signs of regression slopes (+/-) of cross correlations of local anomalies of PP and stratification (SI) versus the anomalies of different variables averaged over the area of the low-latitude, stratified ocean (LL-STRAT) and the Nino3 Box (NINO3; 150 • W-90 • W, 5 • S-5 • N).