Using a biogeochemical ocean model that includes the iron cycle, we carry out preindustrial (control, CTL) and glacial (Last Glacial Maximum, LGM) climate simulations focusing on changes in export production (EP). The model successfully reproduces general trends of a paleoclimate reconstruction of EP at the LGM except over the Atlantic Ocean. By conducting a series of sensitivity simulations, we investigate the mechanism controlling EP at the LGM in each basin. In the Southern Ocean, the model successfully reproduces the dipole pattern of the paleoreconstruction: the higher-latitude decrease and lower-latitude increase of EP. It is found that the lower-latitude increase of EP comes from iron fertilization effects by enhanced dust deposition, while the higher-latitude decrease of EP is caused by the reduction of surface shortwave due to spreading of sea ice there. We also find that increased dust input in other basins remotely affects EP in the Southern Ocean. In the Pacific Ocean, the model suggests that iron fertilization effects are dominant in open ocean regions. In the Atlantic Ocean, the model simulates overall reduction of EP, whereas the paleoreconstruction suggests the increase in some regions. We propose that the Atlantic response is strongly affected by distribution of iron limitation in a control climate. In our CTL simulation, the biological production is limited not by iron but by phosphate in the Atlantic Ocean, which leads to the decrease of EP in spite of the significant increase of dust deposition there. It is implied that the accurate evaluation of iron limitation in the present ocean is critical for evaluating changes in EP and associated reduction of atmospheric CO2 concentration at the LGM.
 The atmospheric CO2 concentration at the Last Glacial Maximum (LGM) is lower than the interglacial periods by roughly 80 ppm [e.g., Petit et al., 1999]. Reasons for changes in the atmospheric CO2 concentration during glacial and interglacial cycles are widely discussed but still not clarified yet [e.g., Sigman and Boyle, 2000; Archer et al., 2000]. One of plausible mechanisms is the so-called “iron hypothesis” proposed by Martin . The Southern Ocean is the most significant high-nutrient, low-chlorophyll (HNLC) region in the present climate. This is presumably due to lack of iron, which is confirmed by iron fertilization experiments where addition of iron indeed causes the increase of primary production [e.g., Martin et al., 1994] and export production [e.g., Buesseler et al., 2004]. Antarctic ice core data suggest that dust deposition at the LGM is significantly higher than interglacial periods [e.g., Petit et al., 1999]. Martin  proposed that increased dust flux rises iron concentration and enhances biological production in the Southern Ocean at the LGM, which acts as a sink of atmospheric CO2.
 Many modeling studies suggest that strengthening of biological production in the Southern Ocean can contribute to the drawdown of atmospheric CO2 concentration [e.g., Sarmiento and Orr, 1991; Popova et al., 2000], but there are a few studies which evaluate the iron hypothesis by using a model with explicit iron cycle. By using a multibox model, Watson et al.  showed atmospheric CO2 concentration drops by 40 ppm due to iron fertilization in the Southern Ocean. Archer et al. , who used an ocean general circulation model (OGCM), obtained the reduction of 8 ppm. Bopp et al.  conducted OGCM simulations under glacial and interglacial conditions by changing not only atmospheric dust deposition but also physical fields such as SST and ocean circulation. They obtained the results that changes in dust deposition induce 15 ppm drawdown and combined changes in dust input and physical fields lead to 32 ppm drawdown. Complex processes related to iron cycle in the ocean are not fully understood yet, and modeling of iron cycle itself is still an important topic. In the study by Bopp et al. , the iron in the deep ocean is rapidly scavenged when its concentration exceeds 0.6 nmol kg−1 because it had been believed that iron concentration in the deep ocean is nearly constant (0.6 nmol kg−1). However, recent observational data suggest that there is significant difference in deep ocean iron concentration among the Atlantic, Pacific, and Southern Oceans. By improving representation of scavenging process, Parekh et al.  developed an iron model which reproduces the overall gradient of iron concentration among the basins. Using their iron model, Parekh et al.  obtained 8 ppm drawdown of atmospheric CO2 concentration by dust input changes. Recently, Tagliabue et al.  performed LGM simulation with more complex biogeochemical model and obtained 11 ppm drawdown by LGM dust deposition. Although all the studies suggest that dust changes have impact on atmospheric CO2 concentration, there is discrepancy in its quantitative evaluation among the previous studies.
 If the iron hypothesis is true, the biological production must be strengthened in the Southern Ocean at the LGM. Kohfeld et al.  compiled various paleoproxies and provided an assessment of changes in export production from the Holocene to the LGM. In their method, the change in each site is categorized into a five-point scale (lower, slightly lower, no change, slightly higher, and higher) and its reliability is also evaluated based on the quality of age model, the measurement methods, the number of proxies, and the degree of consensus among them. Their data suggests that export production indeed increases in lower latitudes of the Southern Ocean, but it decreases in higher-latitude regions in the Southern Ocean. In addition, increase of export production also takes place over the Atlantic Ocean. Therefore, changes in export production at the LGM do not seem simple as the iron hypothesis suggests. Although data are still sparse, comparison of simulated export production with their data is very useful for validating model simulations and model estimate of atmospheric CO2 drawdown. Detail comparison is also expected to improve our understanding of role of iron and other factors in atmospheric CO2 drawdown at LGM more quantitatively.
 In this study, by using the ocean biogeochemical model, we simulate export production at glacial (LGM) and interglacial (control, CTL) climate. As in the same way as Tagliabue et al.  and Bopp et al. , simulations under glacial and interglacial conditions are conducted by changing not only atmospheric dust deposition but also physical fields. The simulated results are compared with paleoreconstruction of Kohfeld et al. . Because Tagliabue et al.  and Bopp et al.  already compared their results with Kohfeld et al.'s  data set (Bopp et al.  used the earlier version of Kohfeld et al.'s  data), our attempt is not the first one. However, our study focuses directly on changes in the export production at the LGM. We discuss detail mechanisms controlling the changes in export production at the LGM. The responses in the Atlantic, Pacific, and Southern Oceans are individually investigated and we clarify that dominant process is different in each basin. For this purpose, we conduct various sensitivity simulations where physical and dust input effects are separately investigated. As for the physical effects, we further decompose them into those from velocity field, vertical mixing, and surface shortwave radiation. Dust input effects are also separately evaluated in the Atlantic, Pacific, and Southern Oceans, by which we demonstrate that increase of dust input in each basin not only enhances in situ export production but also affects export production in other basins remotely. Owing to the above mentioned various sensitivity simulations, we can systematically investigate changes in export production at the LGM and propose their controlling mechanisms in a clear manner.
 The paper is organized as follows. The model and experimental design of numerical simulations are described in section 2. The numerical simulations are shown in section 3 where we explain our main results. In section 4, we provide discussion, implication, and conclusion elicited by the results obtained in section 3.
 The preindustrial (CTL) and glacial (LGM) simulations are conducted by MIROC. The design of simulations follows PMIP2 protocol and the performance of MIROC has been investigated in several papers [e.g., Braconnot et al., 2007; Yanase and Abe-Ouchi, 2007: Otto-Bliesner et al., 2009]. Weber et al.  and Otto-Bliesner et al.  reported the simulated Atlantic ocean deep circulation at the LGM. Although MIROC simulates the stronger Atlantic meridional overturning circulation (AMOC) at the LGM than CTL therein, updated simulations of MIROC reproduce the weaker AMOC at the LGM (Figure 1). In the updated simulations, two modifications are applied to the original simulations. First, the layer thickness diffusivity [Gent et al., 1995] is increased from 3 × 10−7 to 7 × 10−7 cm2 s−1. Second, downward sea surface heat flux is declined between 80°S and 45°S (by 15 W m−2 at 80°S, 0 W m−2 at 45°S, and linearly in between) for removing a warm bias around the Antarctic Sea. The former modification affects the response of deep water formation and the latter leads to realistic sea ice distribution in the Southern Ocean. These modifications are considered to result in more realistic response of the AMOC at the LGM. These updated simulations described above are used in this study.
where C is concentration of tracer, v is velocity, KH and KV are horizontal and vertical diffusivity, and SC is a source/sink term of the tracer associated with biogeochemical processes. Note that the isopycnal and layer thickness diffusions are not explicitly written in (1) but actually included in the model calculation. The velocity, temperature, salinity, and other physical variables required for calculation of (1) are simulated in advance by MIROC. In addition, the horizontal resolution of these physical variables is reduced to ∼2.8° resolution for the sake of computational efficiency. The method for reducing the resolution is basically the same as that described by Aumont et al. . The calculation of (1) is conducted under the prescribed v, KH, and KV by using the same tracer transport algorithm as COCO.
2.3. The Biogeochemical Model
 The biological model is based on the work by Parekh et al. , where iron cycle is explicitly calculated in the model. Values of model parameters displayed in Table 1 of Parekh et al.  are used in this study. Following Parekh et al.  rather than original expression of Parekh et al. , uptake rate of phosphate by phytoplankton (Γ) is determined from availability of light (I), phosphate (P), and iron (Fe);
where ze is depth of euphotic zone, α is maximum export rate, and KI, KP, and KFe are half saturation constants for I, P, and Fe, respectively. We note that the export production (EP) is determined from vertical integral of (2) within the euphotic zone;
where μ is fraction of nutrients transferring to dissolved organic phosphate. In the model, dust deposition is considered as iron source into the ocean. Iron in the ocean is uptaken in proportion to EP at the fixed Fe/P ratio and is removed from the ocean by scavenging onto sinking particles. The scavenging takes place only against free form of iron and the scavenging efficiency depends on particle concentration which is diagnosed from sinking particle flux by assuming constant sinking velocity. Note that total ligand concentration (LT) controls amount of free form of iron and the value of LT is set to be 1 nM, larger than other iron models (e.g., 0.6 nM of Archer and Johnson ). Parekh et al. [2004, 2005] indicate that larger LT is appropriate for simulating recently observed iron concentration in the deep ocean.
 We also incorporate ocean carbon cycle following the procedure proposed in OCMIP2 [Najjar and Orr, 1999] and atmospheric pCO2 is calculated in a well-mixed 1-box atmosphere.
2.4. The Iron Input
 Global distribution of dust deposition flux is obtained from results of SPRINTERS in preindustrial and LGM simulations [Takemura et al., 2005, 2009]. The iron flux (FFe) is assumed to be 3.5 wt % of total dust deposition flux [Fung et al., 2000]. Figure 2a displays distribution of FFe in CTL. The iron input is high in downwind region of large deserts such as Sahara, Australian, and Gobi deserts. The globally integrated value of bio-available iron input (i.e., βFFe) is 1.9 × 109 mol yr−1. This value is comparable to previous studies (e.g., 1.7 × 109 mol yr−1 of Parekh et al.  and 3.65 × 109 mol yr−1 of Aumont et al. ).
Figure 2b indicates changes in iron input from CTL to LGM. In most of areas, the iron input increases in LGM, which is consistent with paleoproxy data (DIRTMAP [Kohfeld and Harrison, 2001]). However, in Figure 2b, the iron input decreases over the Southern Ocean. Because Antarctic ice core data suggest that dust deposition at glacial periods is up to 20 times as large as that in interglacial periods [e.g., Petit et al., 1999], the dust deposition over the Southern Ocean appears underestimated in the LGM simulation by SPRINTERS. Mahowald et al.  point out the importance of dust source from ice sheet area which is not considered by Takemura et al. . We may need more careful treatment of dust source for improvement of dust simulation at the LGM. In this study, in order to evaluate potential effects of increased dust deposition over the Southern Ocean, we additionally conduct a sensitivity simulation where the iron input over the Southern Ocean is increased at the LGM.
2.5. Experimental Design
 In CTL, physical fields (i.e., v, KH, and KV in equation (1)) are obtained from preindustrial simulation of MIROC, and dust input field is obtained from SPRINTERS. Simulated dust input is basically the same as the simulation PRE of Takemura et al. , although we simulate it under SST taken from the MIROC result, whereas Takemura et al.  used the observed SST. As for LGM, in the same way as CTL, we use LGM simulations from MIROC and SPRINTERS for physical and dust fields, respectively.
 In addition to CTL and LGM, we carry out series of sensitivity experiments in order to investigate physical and dust input effects on EP in detail. In PHYS, the physical field is taken from LGM while the dust input is from CTL (vice versa for DUST). In order to further investigate results of PHYS, we carry out PHYS-V, PHYS-KV, and PHYS-I which are the same as CTL except that circulation (v in the equation (1)), vertical mixing (KV in the equation (1)), and surface shortwave radiation (I in the equation (2)) are replaced by those of LGM, respectively. The simulation DUST-ATL (DUST-PAC) is the same as DUST expect that the dust input is replaced by that of LGM only in the Atlantic (Pacific) Ocean. In DUST-SO20, the dust input in the Southern Ocean (south of 60°S) is increased from CTL by a factor of 20 while the other conditions are the same as CTL. The simulation LGM2 is the same as LGM except that we modify the dust input in the Southern Ocean to be 20 times larger than CTL. We also conduct CTL-B0.1 and LGM2-B0.1 where iron solubility parameter is significantly reduced from CTL and LGM2, respectively. The simulations conducted in this study are summarized in Table 1.
Table 1. Simulated Export Production (EP) and Atmospheric pCO2 in Each Simulationa
Name of Simulation
EP (Gt C yr−1)
Parenthetical values represent difference from CTL (for LGM2-B0.1, difference from CTL-B0.1).
LGM physical field
LGM dust-input field
LGM vertical mixing
LGM dust in Atlantic
LGM dust in Pacific
20 times dust in SO
LGM + 20 times dust in SO
CTL (small iron solubility)
LGM2 (small iron solubility)
 Observed database of the World Ocean Atlas 2001 [Conkright et al., 2002] is used for initial condition of phosphate, and GLODAP [Key et al., 2004] is for dissolved inorganic carbon and alkalinity. We use the constant initial value (0.6 nM) for the iron concentration. In all the simulations, the model is integrated for 2000 years from initial conditions described above and the average over the last 100 years is used for analysis.
3.1. Control Simulation
 The observed and simulated surface phosphate distributions are shown in Figures 3a and 3b, respectively. The basic structure of observed field, high concentration in subpolar and equatorial regions and low concentration in subtropical regions, is well reproduced in the model. This structure appears controlled mainly by wind-driven circulation where upwelling and downwelling regions reveal high and low concentrations, respectively. The simulated equatorial upwelling in the eastern Pacific Ocean is too strong and locates too southward due to the coarse resolution of MIROC, which causes the difference between simulated and observed phosphate concentrations there. Besides such disagreement, the model well reproduces the observations as confirmed by scatterplot of Figure 3c.
Figures 3d and 3e display the observed and simulated surface iron concentrations, respectively. Although the observed data are limited, the model reproduces the basin-scale observed features such as high concentration in the Atlantic Ocean and low concentration in the Pacific Ocean. This tendency is also confirmed by scatterplot of Figure 3f, where the concentration in the Atlantic Ocean (circle in Figure 3f) is higher than that in the other basins (cross for the Pacific Ocean and triangle for the Southern Ocean) in both the simulation and the observed data. Higher concentration in the Atlantic Ocean than other basins is explained by the larger dust input in the Atlantic Ocean (Figure 2a). Ocean upwelling also controls the iron concentration by providing iron-rich deep water to the surface as seen in the Southern Ocean. Figure 3f indicates that the model somewhat overestimates the high value in the Atlantic Ocean. On the other hand, the model underestimates the concentration in some regions of the Pacific and Southern Oceans. This may imply that another iron source is important there as suggested by recent studies [Nishioka et al., 2007; Moore and Braucher, 2008].
Figure 3h shows the distribution of the simulated EP. As in phosphate distribution, Figure 3h reflects the wind-driven upwelling/downwelling pattern and captures the overall feature of the satellite-based estimation (Figure 3g). The data-model comparison of EP (Figure 3i) is not as good as that of phosphate (Figure 3c) since there is uncertainty in both the model and the satellite-based estimation. The globally integrated EP is 9.0 Gt C yr−1 in CTL (Table 1), and this value is within the range of observed estimate (e.g., 5.3 Gt C yr−1 of Louanchi and Najjar  and 9.6 Gt C yr−1 of Schlitzer ). Therefore, we believe the model reasonably simulates the present EP. We also note that the simulated atmospheric CO2 concentration is 305.4 ppm (Table 1), and this is almost a steady state value.
3.2. LGM Simulation
Figure 4b displays EP simulated in LGM in the form of the difference from CTL. As shown in Figure 2b, dust input significantly increases in LGM except over the Southern Ocean. It is anticipated that EP increases (decreases) where the dust input increases (decreases). It is the case in the Pacific Ocean: EP increases where the dust input increases. However, in the Atlantic Ocean, despite the most significant dust increasing takes place there, EP decreases in most of the Atlantic Ocean. In the Southern Ocean, even though the dust input slightly decreases there, EP increases in the north of 60°S. Because dust deposition in the Southern Ocean seems underestimated in LGM as mentioned before, result from LGM2 is also shown in Figure 5. The figures suggest that the spatial pattern of changes in EP is common between LGM and LGM2, although the increased dust input over the Southern Ocean modify the absolute value of EP changes. It is suggested that the pattern of EP changes is not necessarily directly linked to that of dust input changes.
Figure 4a represents paleoreconstruction of EP compiled by Kohfeld et al. . The model simulations capture some characteristic features of the paleoreconstruction. In the Southern Ocean, the dipole pattern, EP increases in the north of 60°S and decreases in the south of 60°S, is well accorded with the reconstruction. Tagliabue et al.  and Bopp et al.  also referred to this dipole pattern although their simulated transition from increased to decreased EP occurs more poleward than the reconstruction. Bopp et al.  interpreted that this dipole pattern is caused by changes in both dust input and ocean circulation. We will investigate controlling mechanism of this dipole pattern in detail latter in this study. In the western North Pacific Ocean, EP decreases near the Kamchatka island and increases in south of it in the model. The limited reconstructed data appears to support this pattern, although the low resolution of the model fails to capture the accurate structure (e.g., in the Okhotsk Sea, EP decreases in the data but increases in the model). In the eastern equatorial Pacific Ocean, the model simulates the increased EP. Some data support this response but the others do not. On the other hand, the significant decrease takes place in the North Atlantic and Indian Oceans in the model, which is not accorded with the data.
3.3. LGM Sensitivity Simulations
Figure 6 displays EP in the additional sensitivity simulations (also see Table 1). Based on these results, we try to discuss the mechanisms which control the response of EP at the LGM focusing on individual response in the Atlantic, Pacific, and Southern Oceans in this section.
3.3.1. The Southern Ocean
 As already mentioned, the dipole pattern in the Southern Ocean indicated by the paleoreconstruction (Figure 4a) is well reproduced in the model (Figures 4b and 5). Figure 6g and the other figures clearly demonstrate that the decrease in the south of 60°S comes exclusively from changes in shortwave radiation. Because sea ice near Antarctica increases and spreads toward low latitudes in LGM, high albedo of sea ice makes sea surface shortwave flux smaller there in LGM than CTL. As for the increase of EP in the north of 60°S, Figure 6g suggests that spreading of sea ice also slightly increases EP there because of northward transport of unused nutrients but this effect is too small to explain the response of Figures 4b and 5. Figure 6h clearly indicates that the increase of EP in the low-latitude Southern Ocean is caused by increased dust deposition there in LGM2. In LGM where the dust deposition decreases over the Southern Ocean, the dust input increase in the Pacific Ocean contributes to the low-latitude increase of EP in the Southern Ocean remotely as seen in Figure 6d.
 This study gives the clear explanation on the controlling mechanism of the dipole pattern reported by Kohfeld et al. . In spite of using relatively simple biological model compared with those of Bopp et al.  and Tagliabue et al. , we succeed in simulating the dipole pattern. This implies that response of EP in the Southern Ocean is mainly controlled by changes in physical fields (more specifically, changes in sea ice distribution) and details of biological response may be of secondary importance.
3.3.2. The Pacific Ocean
 In the open ocean region of the Pacific Ocean, Figure 6b captures the overall pattern of Figure 4b, which indicates that the response of EP is basically controlled by the increase of dust input. We also find that Figure 6d is almost the same as Figure 6b there. This means that the Pacific EP is enhanced by local dust deposition there in LGM. In DUST-SO20 where dust deposition is increased over the Southern Ocean, Figure 6h suggests the possibility that the Pacific EP is also stimulated by dust input over the Southern Ocean as a result of interbasin transport of iron.
 As for the decrease of EP near the Kamchatka island in the western North Pacific Ocean, this is caused by changes in physical fields (Figure 6a). Sea ice increases and vertical mixing reduces there at the LGM, both of which contribute to this decrease of EP (Figures 6g and 6e).
3.3.3. The Atlantic Ocean
 The Atlantic deep circulation becomes weaker at the LGM (Figure 1) and the associated upwelling of deep circulation is reduced, which causes the decreased response of EP especially at low latitudes in the Atlantic Ocean (Figure 6c). In addition, EP also decreases in the northern North Atlantic Ocean since the vertical mixing decreases in the downwelling branch of the deep circulation (Figure 6e). These results demonstrate that changes in physical fields at the LGM robustly contribute to the decrease of EP in the Atlantic Ocean (Figure 6a). Therefore, it is expected that changes in biological processes at the LGM seem responsible for the increase of EP in the Atlantic Ocean reported in the paleoreconstruction. In fact, Figure 6b suggests that the changes in dust input explain the large increase of EP around 30°S in the South Atlantic Ocean. This increase is caused by enhanced dust input over the Atlantic Ocean (Figure 6f). However, except for this large increase around 30°S, EP slightly decreases over the Atlantic Ocean in spite of significant increase of dust input there.
Figure 7 plays a key role in understanding the above mentioned decrease of EP in the Atlantic Ocean in our model. As seen in the equation (2), the biological production depends on both phosphate and iron concentrations in the nonlinear way. Here, we call area where P/(P + KP) is larger (smaller) than Fe/(Fe + KFe) as the iron-limited (phosphate-limited) area. Figure 7a displays the difference between P/(P + KP) and Fe/(Fe + KFe) and indicates that the Pacific Ocean is the iron-limited area but the Atlantic Ocean is the phosphate-limited area in CTL. This situation is also the same in LGM, although parts of the North Pacific Ocean shift to the phosphate-limited area (Figure 7b). The equation (2) means that increased iron concentration does not affect the biological production at all in the phosphate-limited area. This explains the fact that increased iron input raises EP in the Pacific Ocean but does not in the Atlantic Ocean. In addition, Figure 6f suggests that the increased dust input not only “does not raise” but also “decreases” EP in the Atlantic Ocean. For explaining this response, we have to consider changes in phosphate concentration. We discuss this point in the case of DUST-ATL. Figure 7a indicates that the boundary between the iron-limited and phosphate-limited areas locates near 30°S in the Atlantic Ocean. Some parts of increased iron in north of 30°S are advected southward, which raises iron concentration over the whole Atlantic Ocean including the iron-limited area of the South Atlantic Ocean (Figure 7c). Owing to the increasing of iron concentration, EP raises in the iron-limited area of the South Atlantic Ocean. At the same time, the increased EP makes phosphate concentration decline there (Figure 7d). This decreasing anomaly of phosphate concentration is advected northward in turn, which leads to slight decrease of phosphate concentration and EP in the phosphate-limited area of the Atlantic Ocean. In this way, in the Atlantic Ocean, EP is decreased by the increasing dust input due to decreasing phosphate concentration there. The response in the Indian Ocean is also explained by the same mechanism.
4. Discussion and Concluding Remarks
4.1. Importance of Evaluation of Iron-Limited Area in the Present Climate
 Comparing simulated EP (Figure 4b or 5) with the paleoreconstruction (Figure 4a), we find that the model succeeds in reproducing overall structure in the Pacific and Southern Oceans. On the other hand, the model fails to reproduce the increase of EP over the low-latitude Atlantic Ocean reported in the paleoreconstruction. The model result is explained by the distribution of the iron-limited area in CTL (Figure 7) as discussed in section 3. In phosphate-limited areas, EP does not increase but decreases by increasing dust input. Therefore, it seems difficult for our model to reproduce the Atlantic increase of EP found in the paleoreconstruction, as long as the Atlantic Ocean is the iron-limited area in our control simulation.
 The above mentioned finding suggests that the distribution of iron-limited area in the present climate has a key role in understanding the response of EP at the LGM especially over the Atlantic Ocean. However, the observed information on this distribution is very limited although some numbers of iron-fertilized experiment prove that the iron-limited area actually locates over the Pacific and Southern Oceans [e.g., Takeda and Obata, 1995; Tsuda et al., 2003; Boyd et al., 2004]. In addition, there is no consensus about the iron-limited area among current models. Schneider et al.  compare four different coupled climate carbon cycle models and report the simulated iron-limited area in each model. Schneider et al.'s Figure 5 indicates that the pattern differs significantly each other; one model indicates that almost all of the ocean is iron-limited, and another model suggests the iron-limited area is restricted to the equatorial Pacific Ocean and the Southern Ocean. Therefore, with our current knowledge, it is difficult to evaluate how realistic Figure 7a is.
 The distribution of the iron-limited area depends on model parameters in addition to the used physical and dust fields. Especially, there is large uncertainty in parameters related to iron cycle. For example, we assume that 3.5% of dust consists of iron and 1% of this iron becomes bio-available at the surface ocean. These numbers are very uncertain and could have horizontal variation although we assume they are globally constant. In addition, the iron flux of Figure 2a depends on used dust transport simulation which may also have difficulty in reproducing the actual dust distribution in quantitatively realistic manners, although we believe the overall pattern is well simulated by the present dust transport model.
 In order to evaluate impact of uncertainty of iron input to reproducibility of EP at the LGM, we conduct the additional simulation where the iron solubility parameter (β) is reduced from 1% to 0.1% (CTL-B0.1). We also conduct the same LGM simulation as LGM2 except that β is set to be 0.1% (LGM2-B0.1). Note that 0.1% of β is probably too small, and here we intend to demonstrate possible impact of uncertainty coming from β, the weight concentration of iron, and amount of dust flux. In CTL-B0.1, almost all of the ocean becomes iron-limited area. In LGM2-B0.1, the changes in EP become very different from Figure 5 especially in the low-latitude Atlantic Ocean (Figure 8). In Figure 8, we successfully reproduce the increase of EP in the low-latitude Atlantic Ocean found in the paleoreconstruction. Whereas the low-latitude Atlantic Ocean is the phosphate-limited area in CTL, it is the iron-limited area in CTL-B0.1. This causes totally different response of EP in the low-latitude Atlantic between Figures 5 and 8. We also find that very large atmospheric pCO2 drawdown takes place in LGM2-B0.1. We should note that we do not think the value of β made above (0.1%) is realistic, because CTL-B0.1 seems unrealistic in terms of too large surface phosphate concentration (not shown), too high atmospheric pCO2, and too small globally averaged EP. In CTL-B0.1 and LGM2-B0.1, we merely demonstrate that the difference in the iron-limited area in the present climate actually controls the response of EP and atmospheric pCO2 at the LGM.
 Although we have demonstrated the importance of control simulation by changing the value of β here, we also point out that the distribution of iron-limited area can be affected by other reasons. Bopp et al.  reproduce the increase of EP in the low-latitude Atlantic Ocean, whereas our study, and Tagliabue et al.  do not. One possible explanation for the difference is that the distribution of the iron-limited area of Bopp et al.  is significantly different from those of our results and Tagliabue et al. . For more quantitatively accurate representation of the iron-limited area, the biogeochemical model used in this study may be required to be improved by such as more realistic input of iron flux and modification of our simple representation of colimitation between phosphate and iron. In addition, coastal regions are recently recognized as important iron source areas in addition to surface dust input [e.g., Nishioka et al., 2006; Moore and Braucher, 2008], which should also be considered in the model since the iron-limited area is controlled by not only dust input but also such iron sources. In the future studies, we expect that more realistic representation of the iron-limited area in control could lead to more accurate simulation of EP at the LGM.
4.2. Implication for Changes in the Atmospheric pCO2 at the LGM
 The globally averaged EP is displayed in Table 1. To clearly demonstrate the effects of each mechanism on EP, we provide Table 2. Table 2 suggests that the globally averaged EP of LGM2 is almost the same as that of CTL. In PHYS, the globally averaged EP decreases by 1.0 Gt C yr−1 from CTL. On the other hand, dust input changes contribute to increase of EP (0.3 Gt C yr−1 of increase in DUST-ATL, 0.7 Gt C yr−1 of increase in DUST-PAC, and 1.3 Gt C yr−1 of increase in DUST-SO20). Even though the globally averaged value of EP in LGM2 does not differ from CTL so much, this is a result of large cancelation between effects of changes in physical and dust input fields.
Table 2. Simulated Changes in Export Production (EP) and Atmospheric pCO2 at the LGM Summarized by Focusing on Each Mechanism
EP (Gt C yr−1)
LGM2 − CTL
PHYS − CTL
PHYS-V − CTL
PHYS-KV − CTL
PHYS-I − CTL
LGM2 − PHYS
DUST-ATL − CTL
DUST-PAC − CTL
Southern Ocean dust
DUST-SO − CTL
 The simulated atmospheric pCO2 is summarized in Table 2 (also displayed in Table 1). The atmospheric pCO2 drawdown from CTL is 50.2 ppm in LGM2. About half of this decrease comes from changes in physical field (−28.5 of PHYS), and the other half originates from changes in dust input. This implies that role of dust input is comparable to that of physical fields. As for physical effects, a large portion of atmospheric pCO2 drawdown results from changes in oceanic solubility of CO2 due to decreasing SST (not displayed) as in previous studies [e.g., Brovkin et al., 2007; Tagliabue et al., 2009; Kurahashi-Nakamura et al., 2010]. PHYS-I suggests that the reduction of EP due to spreading sea ice leads to the increase of atmospheric pCO2 as reported by Kurahashi-Nakamura et al. . We note that changes in EP are not necessarily in antiphase with those in atmospheric pCO2 [e.g., Marinov et al., 2008] as the difference in PHYS-V from CTL, because supply of deep water with higher carbon concentration by ocean upwelling or mixing also affects atmospheric pCO2. As for dust input effects, comparison among DUST-ATL, DUST-PAC, and DUST-SO20 indicates that increase of dust input in the Southern Ocean have the largest contribution to atmospheric pCO2 drawdown as previous studies suggest. In addition, we note that increase of dust input in the Pacific Ocean has a nonnegligible contribution. Rough estimate from Table 1 indicates that about two thirds of dust effects come from changes in dust input over Southern Ocean, and the other one third results from those over the Pacific Ocean. At the same time, dust input changes in the Atlantic Ocean have only a minor effect.
 Although recent OGCM studies seem to indicate that dust input effects on atmospheric pCO2 is small (e.g., 8 ppm of Archer et al. , 15 ppm of Bopp et al. , and 11 ppm of Tagliabue et al. ) and suggest that dust input effects may play only a second-order role in atmospheric pCO2 variation, we obtain relatively larger pCO2 drawdown (24.9 ppm in DUST-SO20) than these studies. One of its reason may come from the fact that the dust flux used in this study is different from those in the studies cited above: Takemura et al.  in our study, Mahowald et al.  in the study by Tagliabue et al. , and Mahowald et al.  in the other studies. The globally integrated values of these dust flux in the present climate are significantly different from each other (in calculation using our land-sea mask, the globally integrated value of bio-available iron input is 1.9 × 109 mol yr−1 for Takemura et al. , 5.6 × 109 mol yr−1 for Mahowald et al. , and 10.9 × 109 mol yr−1 for Mahowald et al. ), and that of Takemura et al.  is the smallest among them, which may lead to high sensitivity of pCO2 drawdown at the LGM. The used biogeochemical model are also different among the studies. Our model is relatively simple compared to the complex biological model of Tagliabue et al.  and Bopp et al. . Our model may be too simple to represent the actual EP changes at the LGM. For example, it is suggested that N fixation is sensitive to changes in dust deposition [e.g., Falkowski, 1997; Tagliabue et al., 2008]. It is also indicated that an Fe/C ratio of phytoplankton changes in response to increased dust deposition [Sunda and Hunstman, 1997; Tagliabue et al., 2009]. Such biological response is not considered in our model, and different treatment of biological processes may improve the Atlantic response of EP where our model does not reproduce the paleoreconstruction. In addition, the used physical fields are also important as we demonstrate that the dipole pattern in the Southern Ocean is explained by spreading sea ice; our model successfully reproduces this pattern in spite of relatively simple biogeochemical model. We emphasize here that the paleoreconstruction of EP is very powerful information for evaluating how realistically the model estimates changes in EP and atmospheric pCO2 at the LGM. More detail paleoreconstruction of EP at the LGM and careful comparison with the model will help to improve our quantitative evaluation of “iron hypothesis” and other factors in atmospheric CO2 changes at the LGM.
 Comments from A. Tagliabue and anonymous reviewers were helpful for improving the manuscript. R. Ohgaito helped us prepare the data set of coupled model simulations. H. Ichijo conducted the dust transport simulations by using SPRINTERS developed by T. Takemura. N. Mahowald kindly provided us with the dust simulation results of Mahowald et al. [1999, 2006]. Thanks are extended to T. Kurahashi-Nakamura, M. Yoshimori, T. Sueyoshi, and R. O'ishi for their helpful comments and discussions. The numerical simulations were performed by HITACHI SR11000 at Information Technology Center, University of Tokyo, and NEC SX-8 at the National Institute for Environmental Studies. The figures were prepared with the Dennou Library (developed by the GFD-Dennou Club) and GMT.