Modeling how surface nitrogen fixation influences subsurface nutrient patterns in the North Atlantic
Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, Yokohama, Japan
Corresponding author: C. Yoshikawa, Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, G1-25, 4259 Nagatsuta-cho, Midori-ku, Yokohama 226–8503, Japan. (firstname.lastname@example.org)
 We represented mechanistically the process of nitrogen (N) fixation and associated N* anomalies in the Atlantic Ocean using a three-dimensional coupled physical/biogeochemical model. Available direct measurements of N fixation rates in the Atlantic Ocean are compiled, and these, along with observed N* anomalies, constrain the model. The model N fixation rate for the whole Atlantic domain is 2.1 × 1012 molN yr–1. The model-generated N* anomaly shows the observed feature of a subsurface maximum. When plotted on isopycnal surfaces, the model-generated N* anomaly bears little relation to the pattern of N fixation at the surface. However, the highest N fixation rates should be spatially related to N* distribution if particulate export is remineralized at depths in the same region where the N fixation occurred. We performed case studies varying remineralization and advection to clarify the genesis of the N* anomaly and to determine the reasons underlying differences between N* anomalies and N fixation rate patterns. These studies indicated that the difference between these two patterns was created by both horizontal advection of excess N compared to phosphorus (P) and preferential remineralization of P compared to N. N fixation and preferential P remineralization create high N* anomalies both at the surface and in subsurface waters in the tropical Atlantic, which are transported into the northwestern North Atlantic by western boundary currents and subsequently subducted. As a result, the highest N* anomalies are located not in the tropics but in the northwestern North Atlantic.
 Nitrogen (N) fixation is important because of its role as a N source in the oceanic fixed N budget [Gruber and Sarmiento, 1997; Codispoti et al., 2001; Gruber and Sarmiento, 2002; Codispoti, 2007; Gruber, 2008] and also because of its potential to drive net carbon (C) export [Hood et al., 2000]. High rates of N fixation have been observed in the tropical and subtropical Atlantic Ocean, where nitrogenous inorganic nutrients are deficient and iron (Fe)-rich dust deposition is very high [e.g., Capone et al., 1997, 2005]. Basin-wide N fixation rates for the North Atlantic have been estimated by various methods: prognostic model runs, geochemical estimates based on observed departures from Redfield stoichiometry of N and phosphorus (P), C drawdown and isotopic calculations, and observational extrapolations based on direct rate measurements [Gruber and Sarmiento, 1997; Hood et al., 2001; Lee, 2001; Coles et al., 2004; Hansell et al., 2004; Capone et al., 2005; Coles and Hood, 2007; Deutsch et al., 2007; Moore et al., 2009; Monteiro et al., 2010]. However, these estimates vary between 0.6 × 1012 molN yr–1 and 3.2 × 1012 molN yr–1 in the tropical North Atlantic domain (10°N–30°N) and 2.0 × 1012 molN yr–1 and 8.7 × 1012 molN yr–1 in the whole Atlantic domain. Here, we update a prior estimate [Coles and Hood, 2007] using a higher-resolution prognostic model constrained by both measured N to P ratio anomalies and N fixation rates.
Gruber and Sarmiento  introduced a tracer,
which is a deviation from the Redfield proportion [Redfield et al., 1963], to investigate atmospheric dinitrogen fixation, denitrification, and inputs from outside of the ocean. Here we use the Gruber and Sarmiento  definition of N* for comparison with previous historical work, but equally applicable is the Hansell et al.  definition of the DINxs parameter, which is equivalent to the N* equation, but unnormalized and unbiased:
 Although biogeochemical cycling in Redfield proportions does not generate N* anomalies, atmospheric dinitrogen fixation, atmospheric N deposition, and denitrification with non-Redfield elemental cycling can generate N* anomalies. N fixation often occurs in oligotrophic regions, when diazotrophs take up dinitrogen gas instead of nitrogenous nutrients (i.e., nitrate and ammonium). When this happens, N* increases because of the N addition in the absence of a commensurate (Redfield) addition of P. Similarly, atmospheric deposition of N and P that is out of Redfield proportion can also alter N*. In contrast, in low-oxygen waters and poorly oxygenated sediments, denitrification, which involves the utilization of nitrate instead of oxygen in bacterial respiration, occurs, which decreases N* because nitrate is consumed (converted to N2 gas) in the absence of a commensurate loss of P. Similarly, chemosynthetic bacteria can also lower N* in low oxygen environments by consuming nitrite/nitrate in the anammox reaction. N* then has utility in the tracing of water masses affected by N fixation/deposition and denitrification/anammox.
 However, the N* field is also affected by the advection of nutrients and suspended and sinking particles. Palter et al. , Moore et al. , and Palter et al.  suggested that the horizontal advection of nutrients is important in determining the availability of nutrients in the North Atlantic subtropical gyre. Moreover, Siegel and Deuser  showed that the export trajectories of individual sinking particles have a very strong horizontal component. In fact, Coles and Hood  found that the horizontal advection of low N* water from the Arctic was an essential boundary condition in their model (following observational work of Yamamoto-Kawai et al. ). Within the basin itself, they found that the N* anomaly on an isopycnal surface bears little relation to the pattern of N fixation at the surface and suggested the possibility that the advection of both fixed nitrogenous nutrients and N-rich particulate matter have a strong influence on N* distribution patterns. However, few studies have investigated directly the route by which surface, tropically generated excess N is delivered into the thermocline of the North Atlantic.
 Moreover, the N* field is also possibly affected by the non-Redfield remineralization of particulate and dissolved organic matter (DOM). Several studies [Shaffer, 1996; Clark et al., 1998; Vidal et al., 1999; Abell et al., 2000; Aminot and Kerovel, 2004; Castro et al., 2006; Landolfi et al., 2008] have reported C:P and N:P enrichment of DOM. Although sinking particles have C:N:P ratios much closer to the Redfield ratio compared to DOM, several studies [Paytan et al., 2003; Faul et al., 2005; Benitez-Nelson et al., 2007; Paytan and McLaughlin, 2007] have reported typically higher C or N relative to P, especially organic P compounds, in sinking particles compared to the Redfield ratio. This provides potential evidence for the preferential remineralization of P compared to N, which would also give rise to positive N* anomalies. Indeed, the recent model results of Zamora et al. , who investigated a similar topic related to subsurface positive N* anomalies in the North Atlantic, showed that preferential P remineralization leads to elevated subsurface nitrate relative to phosphate (PO43–), although they concluded that preferential P remineralization alone could not account for the high N* subsurface signal. Their model did not include N fixation but did include atmospheric N deposition and the non-Redfieldian uptake of nutrients. Here, we use a model that includes N fixation processes and investigate the contribution of N fixation to subsurface N* anomalies. However, to maintain N fixation at the surface, mechanisms of differential remineralization of N and P [Coles and Hood, 2007], dissolved organic P (DOP) uptake [Mather et al., 2008], P mining [Karl et al., 1992], and coupling between denitrification and N fixation [Deutsch et al., 2007] have been invoked to provide a low N/high P environment. The Atlantic is not a region of open ocean denitrification, and thus the mechanism of coupling between denitrification and N fixation explains the observed N fixation via low N* input from other regions where denitrification occurs. Here, we investigate differential remineralization as one means of generating an N* anomaly; however, we believe that P mining and DOP uptake would show a very similar pattern in N* results, as P would become more available at the surface where N fixation is active.
 In this study, we present results from a 3-D coupled physical/biogeochemical model with both dynamic and externally imposed representations of N fixation and the resultant N* anomalies. The model horizontal resolution is 0.5°, which is finer than that in the previous 2° resolution study by Coles and Hood , to improve the realism of the ocean circulation. We compile and summarize available direct measurements of N fixation rates in the Atlantic Ocean and evaluate our model results using both measured N fixation rates and N* anomalies calculated from observed nutrient data sets. Lastly, we perform case studies, varying remineralization and advection to clarify the processes that create the difference between surface N fixation patterns and the subsurface N* anomaly in the Atlantic.
2 Model Description
2.1 Biogeochemical Model
 The biogeochemical model used in this study is a modified version of the Hood et al.  model [see also Coles et al., 2004; Hood et al., 2004; Coles and Hood, 2007], which includes N, P, and Fe cycling and limitations (Figure 1). Here, we provide a general description of the model. A detailed discussion of the model equations and all of the parameter settings can be found in Coles and Hood , so only a general overview is presented. The model parameters are slightly revised and are presented in Table 1.
Table 1. Model Parameters
Decreased slightly from 0.01 in Coles and Hood  to maintain observed nitrogen fixation rates in the western tropical and subtropical Atlantic shown in Table 2.
Decreased from 0.35 d-1 in Coles and Hood  to reproduce observed near-surface DIP and DIFe concentrations from Conkright et al.  and Anderson and Henderson  and reasonable P and Fe limitations from Coles and Hood .
 The biogeochemical model includes three functional compartments with variable elemental ratios: Inorganic nutrients [dissolved inorganic N (DIN), dissolved inorganic P (DIP), dissolved inorganic Fe (DIFe)], organic nutrients [dissolved organic N (DON), DOP, dissolved organic Fe (DOFe)], and detritus (DetN, DetP, DetFe), as well as three biological compartments with fixed but not necessarily Redfield nutrient ratios: phytoplankton, a diazotroph patterned after Trichodesmium, and a heterotroph. In this study, the phytoplankton and heterotroph compartments have Redfield N:P stoichiometry, whereas the diazotrophs are substantially N-rich [e.g., Krauk et al., 2006; White et al., 2006; see Table 1]. The Fe:N ratios of the phytoplankton and heterotrophs are converted from published Fe:C ratios for these groups (Tortell et al.  and Fung et al. , respectively) using appropriate conversion factors (Vrede et al.  and Redfield et al. , respectively). Diazotrophs are parameterized to have a substantially higher Fe:N ratio than phytoplankton or heterotrophs using conversion factors from Sañudo-Wilhelmy et al. . In contrast, the inorganic nutrient, organic nutrient, and detritus compartments have emergent stoichiometric ratios, which are determined by differential uptakes and remineralizations of functional compartments that are out of the Redfield ratio, differential remineralizations of compartments that are in the Redfield ratio, and differential sources and sinks at model lateral boundaries. Detritus is produced by mortality from all plankton groups and also via heterotroph egestion. Organic nutrients are produced by exudation from phytoplankton and diazotrophs and excretion by heterotrophs; thus, they represent only rapidly remineralized labile organic matter pools. Eventually, detritus and organic nutrients are recycled to inorganic nutrients. Detrital P, N, and Fe can be recycled at different rates in the model [Shaffer, 1996; Clark et al., 1998; Castro et al., 2006]. The specific remineralization rates were determined by Coles and Hood  to reproduce the observed levels and patterns of P, N, and Fe limitation observed by Mills et al.  and also to give reasonable DIP, DIN, and DIFe concentrations at the surface and approximately correct N* patterns. The model-generated N* is calculated by setting NO3– = DIN and PO43– = DIP with equation (1); thus, the potential for organic nutrient contribution to N* is only represented when the organic nutrients have been mineralized.
 Fe deposition in the surface mixed layer is specified using model-generated climatological dust fluxes [Luo et al., 2003; Mahowald et al., 2003]. A fraction of this dust is assumed to be Fe, some fraction of which is bioavailable [Christian et al., 2002; Moore et al., 2002b; Table 1]. This bioavailable fraction of the Fe is added directly to the detrital Fe compartment to allow for further biological or chemical processing to make the dust-associated Fe available for uptake by phytoplankton. Following Christian et al. , Fe is “scavenged” from the DIFe pool at a rate that is proportional to the total detritus concentration using a scavenging rate and a nondimensional multiplier. River input is incorporated in the simulations as surface freshwater mass flux for the four largest Atlantic Basin rivers (the Amazon, Congo, Orinoco, and Mississippi), with modest P and Fe concentrations of 0.4 mmol m–3 and 0.007 mmol m–3, respectively. However, because observations show that Amazon nitrate is removed on the shelf prior to the plume influencing the open ocean [DeMaster and Aller, 2001], and the Amazon is the largest river source to the region of N fixation, no N inputs from the rivers are assumed in this model.
 The model was tuned to reproduce surface chlorophyll based on SeaWiFS climatology, the observed N* anomaly distribution, and magnitude and both surface and subsurface DIN, DIP, and DIFe concentrations based on comparison with both historical ship transects and databases (WOCE A16N and A20 sections) (see section 3). As a result, the parameters for the Trichodesmium natural mortality rate, the enhanced DP and the DFe recycling rate are modified from Coles and Hood  (Table 1).
2.2 Physical Model
 The 3-D general circulation model is the Hybrid Coordinate Ocean Model [Bleck, 2002; Chassignet et al., 2003; Halliwell, 2004]. This model is formulated with an arbitrary vertical coordinate that transitions from fixed depth coordinates at the surface to variable isopycnal following coordinates in the ocean interior. The implementation extends from 35°S to 65°N, with 24 vertical layers, and 0.5° spatial resolution (0.5° in latitude, and 0.5° cos (lat) in longitude), which makes it eddy-permitting in tropical regions, but does not resolve the Rossby radius sufficiently at midlatitudes to ensure that the Gulf Stream separation point is well simulated. Atmospheric forcing comes from the ERA40 Reanalysis [Kallberg et al., 2005], and surface salinity is relaxed back to World Ocean Atlas 1994 monthly values to minimize model drift [Levitus and Boyer, 1994; Levitus et al., 1994].
 The model is initialized to the World Ocean Atlas 1994 levels of temperature, salinity, nitrate, and PO43– [Conkright et al., 1994; Levitus and Boyer, 1994; Levitus et al., 1994]. Initial DIFe is set proportionally to DIN using an Fe:DIN ration of 25:1 (µmol:mol) [Fung et al., 2000]. The phytoplankton, diazotroph, heterotroph, DOM, and detritus pools are initialized to low and spatially uniform values using Redfield ratios for the DOM and detritus. At the southern, northern, and Mediterranean Sea boundaries, the temperature, salinity, and inorganic nutrient concentrations are relaxed below 25 m to seasonal averages to ensure that the inflowing water masses include high N* values from the Mediterranean and low values from the Arctic, unless otherwise specified.
 The physical model is spun for 60 years before the biogeochemical model is initialized and run. The modeled nutrient distributions have a rapid initial adjustment, stabilizing after 20 years, although they continue to adjust gradually as the model's physical and biogeochemical cycles equilibrate. The simulations presented here are shown after 60 years of simulation.
3 Model Validation
3.1 Surface Phytoplankton Chlorophyll
 Figures 2c and 2d show the monthly climatologies of surface chlorophyll concentrations in April and August for the model run to be compared with Figures 2a and 2b, which are monthly climatologies for April and August constructed from 5 years of SeaWiFs satellite ocean color data. Note that the color bar is nonlinear to highlight differences in oligotrophic regions. Figures 2e and 2f show the root mean square error (RMSE), which is an estimator of the absolute magnitude of the differences between the observed and modeled values. The model generally reproduces the seasonal cycles of upwelling and the spring blooms. The phytoplankton growth is limited by N in the Atlantic in both April and August, as shown by previous coarse-resolution results [Coles and Hood, 2007]. Some model biases exist, e.g., the modeled oligotrophic gyres systematically overestimate chlorophyll a. Although the model resolution is four times finer, the Gulf Stream still follows the North American shelf north of Cape Hatteras as expected in a noneddy resolving model. As a result, the RMSE shows a tongue-shaped high value around the northern subtropical gyre boundary in April created by the advection of low-nutrient Sargasso Sea water northward rather than eastward. The previous model also tended to overestimate equatorial upwelling and underestimate coastal upwelling, as is generally the case for coarse-resolution ocean models, leading to excess phytoplankton biomass estimates along the equator and low phytoplankton biomass estimates along the coastlines [see also Hood et al., 2004; Coles and Hood, 2007]. In this model, these trends are improved to some degree, but the RMSE is still high along the coastlines in some regions (e.g., in the northern Gulf of Mexico, off the northwestern coast of Brazil, and off the west coast of Africa in spring). These are all areas where riverine sediment and CDOM inputs from the Mississippi, Amazon, Orinoco, and Congo rivers may cause overestimation bias in SeaWiFS chlorophyll estimates; thus, the model may not be as low as the comparison would suggest [Carder et al., 1989; Del Vecchio and Subramaniam, 2004; Hu et al., 2004]. Although the four times finer-resolution model improves the ocean circulation and phytoplankton distribution, a much finer-resolution model is needed to accurately reproduce the high phytoplankton biomass along the coastlines in upwelling zones such as, e.g., the Gulf of Guinea.
3.2 N2-fixation Rate
 Here we compile and summarize the available direct measurements of N fixation rates from the Atlantic between 40°N and 20°S, and compare them to the model-generated rates. These measurements are reported in Table 2. We also generate seasonal maps of N fixation rates for direct comparison of observed and modeled rates (Figure 3). Hood et al.  compiled measurements of Trichodesmium biomass. Using these measurements, they confirmed that the spatial pattern and seasonality of Trichodesmium abundance predicted by our previous model was essentially correct, and thus indirectly validated the N fixation rates. Here, we directly compare the modeled rates with measurements. This is now possible because many N fixation rate measurements have been published since the Hood et al.  study.
Table 2. Observed N2 Fixation Rates in the Atlantic Ocean; Rates Were Recalculated Using Molar Ratios of 3:1 Moles C2H2:N2 Reduced Where Noted
Observed rates (µmol Nm–2 d–1)
Original paper used a molar ratio of 4:1 moles C2H2:N2 reduced.
Original paper used a molar ratio of 6.3:1 moles C2H2:N2 reduced. Converted from hourly rate assuming N2 fixation persisted for 10 h per day.
Converted from hourly rate assuming N2 fixation persisted for 24 h per day. Data base on direct 15N2 uptake. Average rates are assumed over the top 20 m and have been increased by 50% to account for activity below 20 m.
Data base on direct 15N2 uptake.
Nitrogen fixation rates of Trichodesmium.
Nitrogen fixation rates of unicellular diazotrophs.
Nitrogen fixation rates of the entire nitrogen fixing community.
 The meridional extent of N fixation predicted in the model is consistent with observations (Figure 3); i.e., the model has finite N fixation rates in a region that is largely confined to tropical and subtropical waters [e.g., Carpenter, 1983; Capone et al., 1997]. Variations in N fixation rates in the model are determined primarily by the depth and duration of winter mixing [Hood et al., 2004]. In tropical and subtropical regions, persistent net surface heating results in thinner mixed layers, higher mean light levels, and DIN depletion, which favor Trichodesmium growth. In the model, Trichodesmium growth is co-limited by both P and Fe over most of the tropics (20°N and 10°S) and is limited by P in the southern Sargasso Sea and Gulf of Mexico, as shown by previous coarse-resolution results [Coles and Hood, 2007].
Trichodesmium is commonly observed and highly variable in the Gulf of Mexico. While Trichodesmium concentrations during November-April in the eastern Gulf are about 0.75 col L–1, summer concentrations are generally much higher, averaging 20 col L–1 [Lenes et al., 2001] and bloom concentrations in excess of 1000 colonies L–1 have been observed [Vargo et al., 2004]. Mulholland et al.  reported N fixation rates in the eastern Gulf of Mexico from July 2001, June 2003, and November 2003. The observed rates in summer and fall were high, reaching values of 199 and 165 μmolN m–2 d–1, respectively (Table 2). The model-generated N fixation rates show similarly high values in summer and fall, with a summer average rate of 218 μmolN m–2 d–1 and a fall average rate of 212 μmolN m–2 d–1, which is much higher than the winter average model rate of 122 μmolN m–2 d–1 (Figure 3). The differences in the mean values between modeled and observed rates in the Gulf of Mexico are +19 μmolN m–2 day–1 in summer and +47 μmolN m–2 d–1 in fall. Thus, the model appears to reproduce the main features of the observed seasonality, but the rates are slightly overestimated. This discrepancy may be due to exceptionally low Trichodesmium concentrations in July 2001 (between 0.2 and 7.8 col L–1), which are closer to winter values, and may have provided a low bias to the observational averages [Mulholland et al., 2006].
 Many measurements of N fixation rates have been made in the western North Atlantic Ocean. Capone et al.  summarized the rates during six cruises in various seasons. They showed that the highest rates were localized on the western side of the Atlantic Ocean above 10°N, and that some seasonality existed, with the lowest rates in January, increasing in the spring to reach maximal rates during the late summer through early fall. The model-generated N fixation rates show the same high values on the western side of the basin above 10°N and the same seasonal features as those observed by Capone et al.  and Subramaniam et al. , with summer and fall average rates of 160 and 177 μmolN m–2 d–1, respectively, which are much higher than the winter model average rate of 117 μmolN m–2 d–1 (Figure 3). Moreover, Carpenter and Price  measured N fixation rates on three cruises to the Caribbean and western Sargasso Sea during September–October 1973, February–March 1974, and August 1974 and found that the average N fixation rates in the Caribbean Sea were 162 μmolN m–2 d–1, while those in the Sargasso Sea were only 6.9 μmolN m–2 d–1. The model-generated N fixation rates showed the same spatial features, with an annual average rate of 220 μmolN m–2 d–1 in the Caribbean Sea, which is much higher than the annual average rate of 86 μmolN m–2 d–1 in the Sargasso Sea. Furthermore, Orcutt et al.  conducted a seasonal study measuring rates of N fixation on monthly cruises over a 2.5 year period at the Bermuda Atlantic Time-series Station (BATS) site. The observed seasonal averages are high in summer (13 μmolN m–2 d–1) and fall (22 μmolN m–2 d–1) compared with the winter (1 μmolN m–2 d–1) and spring (1 μmolN m–2 d–1) rates. The model-generated rates similarly show high values of 8 μmolN m–2 in summer and 69 μmolN m–2 in fall around the Bermuda Atlantic Time-series Station site (31°N, 64°W) compared with low values in winter (1 μmolN m–2 d–1) and spring (1 μmolN m–2 d–1). The differences in the mean values between modeled and observed rates in the western North Atlantic Ocean are +59 μmolN m–2 d–1 in winter, –1 μmolN m–2 d–1 in spring, –53 μmolN m–2 d–1 in summer, and +30 μmolN m–2 d–1 in fall. Thus the model appears to reproduce the main features of the observed rates and seasonality, but the winter rate is overestimated and the summer rate underestimated. The model cannot reproduce observed locally heterogeneous events like blooms [see Hood et al., 2004] because such events are not reconstructed by a model at this scale without data assimilation. This provides at least a partial explanation of why the model tends to overestimate the observed N fixation rate in prebloom and postbloom seasons, and underestimate the rate observed during bloom events, i.e., events in the model have lower amplitudes and frequencies compared to those in nature. Similarly, the apparent underestimation of rates during summer may be due to the fact that the observed summer rates include some stations where N fixation rates were in excess of 1000 μmolN m–2 d–1, observed during transient surface bloom events that cannot be reproduced by the model. Given the limited observations, estimates of the standard deviations of observed and modeled N fixation rates, which would provide better model-data comparison, cannot be made.
 To date, only a few measurements of N fixation rates have been made on the African side of Atlantic. Staal et al.  measured N fixation rates along a north-south transect in the eastern Atlantic Ocean from January to February 2000. Nitrogenase activity was detected between latitudes of 14°N and 13°S. The region off the coast of DR Congo had the highest measured rates on the transect of up to 11.3 μmolN m–2 d–1. Voss et al.  and Montoya et al.  measured much higher rates of 140 μmolN m–2 d–1 in the eastern North Atlantic along a 10°N transect from October to November 2002, which declined westward to 24 μmolN m–2 d–1. Moreover, Tyrrell et al.  reported abundance only, but showed that Trichodesmium was commonly observed at high abundances in the region off the coast of Guinea, based on eight Atlantic Meridional Transect cruises between 50°S and 50°N during both spring and fall during 1995–1999. The model-generated N fixation rates also show high values all year-round in a band centered at about 10°N, which extends westward from the coast of Guinea (annual average: 205 μmolN m–2 d–1; Figure 3). The differences in mean values between the modeled and observed values in this area are +189 μmolN m–2 d–1 in winter and +84 μmolN m–2 d–1 in fall. Thus, the model reproduces the observed feature of high rates in this area, but it appears that the winter rate is substantially overestimated. This discrepancy may be due to the fact that the winter rates measured by Staal et al.  were within the lower range published for the Atlantic Ocean because of their lack of encounters with blooms or typical colony forms.
 We cannot directly evaluate the modeled rates for the western South Atlantic because no published rates from this region exist. However, Tyrrell et al.  reported Trichodesmium abundance in the western South Atlantic and showed that Trichodesmium concentrations were high in the area along the 10°S in spring where high N fixation rates are generated in spring by the model (Figure 3).
3.3 N* Distribution
 Although the preceding comparison between the model and observed rates of N fixation provides some insight into the accuracy of the model, and confirms high rates of N fixation in the Atlantic, the resolution of the observations are not sufficient to allow for a comprehensive assessment of the performance of the model. Therefore, we also use the observed anomaly in the N and P ratio (N*) generated by the balance of import, denitrification, N fixation, and non-Redfield remineralization to evaluate the correctness of the model-generated N fixation rate.
 We calculate N* by equation (1) on the 26.8σθ density surface using the WOCE Global Hydrographic Climatology [Gouretski and Koltermann, 2004; Figure 4a] and in our model run (Figure 4b). For comparison, the annually averaged N fixation rate generated by the model is also shown in Figure 4c. The model shows a similar pattern to the observations on this density surface, with the maximum N* anomaly occurring in the northwestern subtropical gyre and extending northward into the eastern subpolar gyre. However, the model-generated N* maximum is slightly lower than that observed on this density surface due to the somewhat deep N* maximum as compared with observations (see Figure 5). The model also fails to represent the observed low N* pattern in the Gulf of Mexico. This is an intriguing feature of the climatology. Since N* inputs to the Gulf of Mexico are high, this minimum requires a shelf denitrification signal that reaches the depth of the 26.8σθ density surface, which is much shallower in the Gulf of Mexico than in the western Atlantic Ocean. A pronounced N* minimum occurs in the Caribbean and Gulf of Mexico at a depth of 500 m, which is likely related to shelf denitrification (data not shown). This process is not represented in our model, but note that no quality reference data were available for the Gulf of Mexico, and the data quality may not be as high as for the Atlantic basins. The model-generated high N* patterns are in good agreement with the model-generated high N fixation rate patterns in the western North Atlantic, the area off the coast of Guinea, and the area along 10°S (Figures 4b and 4c).
 The N* patterns are dissimilar to modeled and observed surface N fixation rate patterns north of 30°N. High N fixation rates are not generated by the model in the Atlantic Ocean north of 30°N, yet high N* values above 2 μmolN L–1 occur on the 26.8σθ density surface in this area in both the model and in observations. The highest N fixation rates should be spatially related to N* distribution if particulate export is remineralized at depth in the same region where the N fixation occurred. Coles and Hood  suggested that the difference between N* and surface N fixation rate patterns is due to the advection of both fixed N at the surface and exported particulate N and P at depth, which transports the N* anomaly to higher latitudes. The validity of this hypothesis is discussed in detail in the next section.
 Figure 5 shows vertical N* distributions along 25°W in July and 52°W in August, compared with observations from WOCE lines A16N and A20. The model represents the basin-wide N* pattern, although more spatial variability exists in the WOCE data. The magnitude of the N* anomaly in the model is approximately correct (3–5 μmolN L–1) but the maximum is too deep (around 700 m) compared to observations (around 400 m). The previous coarse resolution model showed the same deeper N* maximum, and Coles and Hood  suggested that the sinking rate of detritus was too fast, or the remineralization of detritus too slow. However, the subsurface N* maximum might be created not only by local export of the anomaly from above and subsequent remineralization, but also by the advection and diffusion of the anomaly. The model has a much stronger N* minimum at the sea surface (above 500 m) than that observed in the tropical region. This is because we used fixed remineralization and sinking rates for every region. Data from the Atlantic Meridional Transect cruises, however, do show a frequent negative N* anomaly between 100 and 250 m, which is consistent with the model results [Mahaffey, 2003, Figure 3.40]. The discrepancy between the modeled and observed N* anomaly may suggest that the difference between N and P remineralization rates in the model is too large in these locations, or may be related to more localized preferential P remineralization. The model also fails to represent the low N* signal at the southern end of both transects below 600 m in the observations. This feature may be due to the advection of a low N* shelf/slope denitrification signal from the east, and denitrification processes are not represented in the model [Coles and Hood, 2007]. The largest N* values in the model are centered between 10°N and 20°N along 25°W with both surface and subsurface maxima, collocated with high N fixation rates at the surface. In contrast, along 52°W, the N* maxima in the model are centered between 15°N and 30°N at the surface and between 30°N and 40°N just below 1000 m. Thus, the surface N fixation rates are not collocated with subsurface maxima. In fact, the intense N* signal spreads well beyond 30°N at depth in both sections, which is the northern edge of the model simulated N fixation. The mechanisms creating subsurface N* maxima in the model are discussed in detail in the next section.
4 Case Studies
 Four case studies which vary detrital N and P remineralization rates (Case 1, 2), horizontal advection of N and P (Case 3), and remineralization rates of DON and DOP (Case 4) are explored to determine how N fixation creates a positive N* anomaly in the Atlantic Ocean and the reason for the spatial discrepancy between surface N fixation rate and subsurface N* patterns (Figure 6). Because nitrogen fixation rate is strongly dependent on the ratio of nitrogen and phosphorus in the euphotic zone, we cannot separate nitrogen fixation variability from factors such as remineralization rate and advection in a mechanistic nitrogen fixation simulation such as the control run. In the case studies, we hold nitrogen fixation rate constant, and instead approximately the same amount of N that is fixed by Trichodesmium in the control run is added as detritus N (DetN) to the model when and where the temperature exceeds 25°C. This allows us to compare the relative influences of remineralization and advection to setting subsurface nutrient ratios. If we did not hold the nitrogen fixation rate constant it would vary with each test case and confound comparison and interpretation of the results. The spatial and vertical distributions of the N additions are shown in Figure 7, and importantly, the DetN addition is localized to the south of the subsurface N* maxima. All of the experiments were initialized with subsurface N* set to 0, and the boundary conditions were also fixed to ensure that inflowing N* was 0, which allowed us to isolate changes in N* that occur within the Atlantic Basin exclusively. Thus, the emergent N* patterns do not reflect reality, but rather demonstrate the contrasting effects of local processes acting on a fixed nitrogen supply. The results presented here are shown after 60 years of simulation.
 Case 1, which uses the same parameter settings as in the control run, except with the amount of N fixation specified as described above (Table 1), is shown in Figures 8a, 8b, 9a, and 9b. As the initialization and boundary conditions are different, the control and Case 1 runs cannot be directly compared. However, although the N* values in Case 1 are much lower than those in the control run, the 26.8σθ surface and vertical N* distributions in Case 1 are qualitatively similar to the control run (Figures 4b, 5c, and 5d), displaying a crescent-shaped high N* pattern in the western Atlantic and a subsurface high N* anomaly extending north of 40°N.
 In Case 2, the P remineralization rate of detritus (eDP) is lowered from 0.24 day–1 to 0.14 day–1 so that it equals the N remineralization rate of detritus (eDN) (see Table 1). This case was intended to separate out the contribution to the N* signal generated by differential remineralization. The surface N* minimum and subsurface N* maximum disappear in this case (Figures 8c and 8d). This occurs because no mechanism exists for creating the negative surface N* anomaly and positive subsurface N* anomaly, which are generated by differential remineralization, although two mechanisms exist for exporting excess N to depth through settling and via physical subduction processes. The Case 2 simulation illustrates how high positive N* values advect to outside the zone where N is added and subsequently subduct there (compare Figures 8c, 8d to 9c, and 9d). The depth at and extent to which this surface inorganic N-rich pool is subducted is seen in Figures 8c and 8d near 40°N, where in the former, the positive N* anomaly extends down to 2000 m. The difference between the N* pattern on the 26.8σθ density surface and the area where N is added is also significant in Case 2, and most of the excess N* is lost through the northern edge of the domain.
 In Case 3, the horizontal advection of inorganic nutrients (DIN and DIP) is stopped [the advection of other tracers, for example, DOM (DON and DOP), detritus, and others, is not shut off] while setting the remineralization rates of N and P to be equal to those used in Case 2. Thus, this case isolates the role of the advection of inorganic nutrients in moving the N* anomaly out of the tropics. Here, the high N* anomaly is located in the tropical region, which corresponds with the area where N is added (compare Figure 7 to 8e, 8f, 9e, and 9f).
 Comparison between Case 2 and Case 3 illustrates how the horizontal advection of the N* anomaly creates the difference between N* and surface N fixation rate patterns. The high N* water created by N addition in the tropical gyre spreads toward the north on the sea surface via horizontal advection (compare Figures 9c, 9d to 9e, 9f) and is then subducted in the northern North Atlantic (compare Figures 8c, 8d to 8e, 8f). This is consistent with a suggestion by Moore et al. . In our model, 28% of the added N advects north of 30°N. However, the advection of high N* water created by N addition alone cannot generate the typical observed subsurface maximum unless a mechanism for preferential P remineralization is also invoked. This is illustrated by the differences between the results with preferential P remineralization (Case 1) and without preferential P remineralization (Case 2). Whereas the N* maximum is located subsurface in Case 1 (Figures 8a and 8b), it is located at the surface in Case 2 (Figures 8c and 8d). We also conducted a case study in which horizontal advection of DIN and DIP was turned off in the presence of differential remineralization (data not shown). In this case, the surface N* minimum was intensified relative to that in Case 2, and the subsurface maximum of N* was strongest in tropical and equatorial regions. Thus, differential remineralization cannot generate a subtropical N* maximum in the absence of lateral advection and subduction processes. In our model, 72% of the added N sinks or subducts below 200 m. In addition to these processes, preferential P remineralization also creates excess N below 200 m, which is 1.2 times more N than that added. Thus, N fixation through sinking and subduction processes and preferential P remineralization contribute 37% and 63%, respectively, to the development of excess N* below 200 m.
 Recent work has suggested the possibility that preferential P remineralization in DOM also affects the N* field [Shaffer, 1996; Clark et al., 1998; Vidal et al., 1999; Abell et al., 2000; Aminot and Kërouel, 2004; Castro et al., 2006] and that this process can also generate high N* anomalies at the subsurface. We therefore conducted Case 4, in which the remineralization rate of detrital P (eDP in Table 1) is 0.14 day–1 and the P remineralization rate of DOM (eDOP in Table 1) is 0.24 day–1, so that the preferential P remineralization of DOM is turned on and that of detritus is turned off. This case generates a weak subsurface N* maximum (Figures 8g and 8h). However, the impact of turning on differential remineralization of DOM compared to detritus is very different. Note that the positive N* anomaly is transported to deeper layers in the tropical region, but not at higher latitudes compared to Case 1. Overall, this result shows that the effect of the preferential remineralization of P in DOM on the N* field is much smaller than that of the preferential remineralization of P in detritus. However, because the model does not represent a semi-labile DON pool, the effect of the preferential P remineralization in DOM on the N* field is likely underestimated.
 Recent work has also pointed to the importance of the advection of high N:P particulate and DOM [Duce et al., 1991; Hansell et al., 2007; Knapp et al., 2008; Landolfi et al., 2008]. A possibility exists that the particulate and dissolved organic pools contribute to the sequestration of fixed N at depth through advection, with subsequent remineralization to inorganic nutrients, as explored by Landolfi et al. . We conducted an additional case study in which we stopped the horizontal advection of detritus and DOM. Stopping the horizontal advection of detritus and DOM had little effect on the N* pattern (data not shown). However, the model only represents a very labile DON pool (consistent with the results of Meador et al. , which suggested that N fixation does not contribute significantly to the total DON pool) with a much shorter remineralization time scale than that considered by Landolfi et al. , and thus, the results are not comparable. However, generating a subsurface N* anomaly is clearly possible with a model that reproduces the broad pattern in observations without invoking the multiyear mineralization of organic nutrients.
 To clarify the process of N fixation and associated N* anomalies in the Atlantic Ocean, we developed a 3-D coupled physical/biogeochemical model. The horizontal resolution is four times finer than in our previous model [Coles and Hood, 2007] to improve the representation of physical processes. We compiled the available direct measurements of N fixation rates in the Atlantic Ocean to constrain the model, along with observed N* anomalies. The meridional extent of N fixation predicted by our model is consistent with the limited observations. The model generates N fixation rates that are largely confined to tropical and subtropical waters [e.g., Carpenter, 1983; Capone et al., 1997]. The model N fixation rate for the whole Atlantic domain is 2.1 × 1012 molN yr–1. This estimate is in the range of the 2.0 × 1012 to 8.7 × 1012 molN yr–1 reported by prognostic modeling, geochemical estimates, and observational direct measurements [Lee, 2001; Capone et al., 2005; Coles and Hood, 2007], but is probably higher than the recent estimates of Hansell et al.  and Landolfi et al. , who both used a different domain for integration. The model generates a similar N* pattern to that observed, and a typical subsurface maximum. The model-generated N* anomaly on isopycnal surfaces bears little relation to the pattern of observed or modeled N fixation at the surface.
 We performed case studies, varying remineralization and advection to determine the fraction of the N* anomaly due to N fixation, and the reason for the difference between N* at depth and surface N fixation rate patterns. These studies indicated that the difference between these two patterns is created both by the horizontal advection of excess DIN and preferential remineralization of P relative to N. Preferential P remineralization is required to deliver high N* particles to depth to form the subsurface maximum. Advection is also important, as N fixation and preferential P remineralization create pools of N-rich water at the surface and subsurface, respectively. This water is transported northward by the western boundary currents and subducts around the subtropical subduction zones of the North Atlantic. As a result, the highest N* values are located subsurface in subtropical regions, downstream from the tropical source.
 We conclude that the N* pattern in the ocean is strongly affected by N fixation, water circulation, and preferential P remineralization. P mining, the idea that diazotrophs like Trichodesmium use vertical migration to collect P at depth and then return with it to the surface [Karl et al., 1992], would also tend to create subsurface excess N, and therefore might generate a similar pattern to the preferential P remineralization results shown here. However, this mechanism needs to be explored quantitatively in a model. Since N fixation occurs in the tropics, but subduction and advection processes occur in the subtropics, these processes are likely more sensitive to interannual or climate variations such as the North Atlantic Oscillation that affect circulation and subduction patterns. Efforts to improve the model's representation of the physical processes associated with particle export and differential remineralization should therefore be emphasized. Moreover, more direct measurements of N fixation rates would help with the validation and development of the dynamic representation of N fixation.
 We thank M. Voss, J. Montoya, and M. Mulholland for their nitrogen fixation rate data. The contributions of three anonymous reviewers also significantly improved the manuscript. C. Yoshikawa was supported by Grant-in-Aid for Japan Society for the Promotion of Science (JSPS) Fellows, Global Environment Research Fund by the Ministry of the Environment Japan (A-0904) and Global COE program “From the Earth to Earths.” The contributions of V. Coles and R. Hood to this study were supported by NOAA grant Z780701 NA17EC1483 01-5-27807 and NSF grants OCE-0622276 and OCE-0933975. The simulations were carried out using the computer resources of the Tokyo Institute of Technology. This paper represents UMCES contribution 4451.