The ENS104 mean LGM surface air temperature (SAT) anomaly (ΔSAT) is -4.6±1.7, and the range is -2.5 to -10.4 ∘C. The mean is close to the observed ΔSAT of -4±1 ∘C (Annan and Hargreaves, 2013) and the range roughly equivalent to the range of previous model-based estimates (Kim et al., 2003; Masson-Delmotte et al., 2006; Schneider von Deimling et al., 2006; Braconnot et al., 2007; Holden et al., 2010a; Brady et al., 2013). The ENS104 mean LGM SST anomaly (ΔSST) is -1.8±0.8 ∘C, and the range is -4.5 and -0.7 ∘C. The mean is again close to an observational data-constrained model estimate (Schmittner et al., 2011) and within the range of estimates inferred from proxy data (MARGO Project Members 2009 in Masson-Delmotte et al., 2013). There is a positive correlation between ΔSAT and ΔCO2 (r=0.75, 0.05 significance level henceforth), most likely reflecting the radiative impact of atmospheric CO2 on SAT, as well as the effect of changing SAT on ΔCO2. As suggested above, decreasing SST may contribute to decreasing CO2 via the CO2 solubility temperature dependence. Changing SAT may also affect ΔCO2 via its effects on sea ice, ocean circulation, terrestrial and marine productivity (see below). The positive correlation is reproduced in ENS315 (r=0.74), and as shown in Fig. 2, ΔSAT and ΔSST tend to be less negative in ENS315 than in ENS104. In ENS16, ΔSAT and ΔSST are from the extreme or at least lower end of the ENS104 range.

The ENS104 mean ΔSAT and ΔSST spatial distributions are shown in Fig. 3. In line with observations (Annan and Hargreaves, 2013), the largest SAT decreases (>10 ∘C) are simulated over the Laurentide and Eurasian ice sheets. The Equator-to-pole temperature gradient is also broadly reproduced. The largest SST decreases (≥4 ∘C) are found in the North Atlantic and northeast Pacific, with more limited cooling (≤-2 ∘C) in the tropics and polar regions, again consistent with observations. However, the largest SST decreases ought to also be found in the Southern Hemisphere midlatitudes, whereas the simulated cooling is more moderate.

The ENS104 mean percentage increase in LGM salinity (and DIC, ALK, PO4, etc.) due to decreasing sea level is 2.84±0.62 %S, and the range is 2 %S to 4 %S. There is no significant relationship between %S and ΔCO2 in either ENS104 or ENS315, and the distribution of %S is similar in ENS104 and in both ENS315 and ENS16 (Fig. 4).

The ENS104 mean LGM global annual sea ice area anomaly (ΔSIA) is 18.6±7.4 million km2, and the range is 9.9 to 44 million km2. There is a negative correlation between ΔSIA and ΔSAT (r=-0.97) and between ΔSIA and ΔCO2 (r=-0.74). The negative correlation between ΔSIA and ΔCO2 likely reflects the impact of changing atmospheric CO2 on ΔSIA but may also include a smaller contribution from changing sea ice area to ΔCO2. Increasing LGM sea ice area could, for instance, have capped the outgassing of CO2 from the ocean, particularly in the Southern Ocean, and also reduced the net ocean–atmosphere CO2 flux by decreasing the AMOC strength (see below). The negative correlation between ΔSIA and ΔSAT, and between ΔSIA and ΔCO2, is reproduced in ENS315 (r=-0.96 and r=-0.74, respectively). ΔSIA in ENS104 also tends to be higher than in ENS315 and smaller than in ENS16 (Fig. 5).

As shown in Fig. 6, fractional sea ice cover increases in all regions where sea ice is present in pre-industrial simulations, although the largest increases take place in the North Atlantic.

The ENS104 mean spatial distribution of the LGM precipitation rate anomaly (ΔPP) is shown in Fig. 7. The LGM changes are mostly negative but regions of positive ΔPP do exist, notably over Siberia and Australia. The largest LGM precipitation decreases (>1.5 mm d-1, with a maximum of 2.25 mm d-1) are found over northern North America and from around the eastern North Atlantic to northwest Asia, coinciding with the location of the Laurentide and Eurasian ice sheets (and the largest increases in fractional sea ice cover), respectively. Relatively large precipitation decreases (>0.75 mm d-1) are also simulated in eastern Asia, equatorial Africa and other regions of enhanced fractional sea ice cover. Comparison against a pollen-based precipitation reconstruction (Bartlein et al., 2011; Alder and Hostetler, 2015) suggest that the simulated precipitation changes over Europe and equatorial Africa are of the right direction, while precipitation changes over western Siberia at least ought to be negative. The sign of the precipitation changes over North America is mostly consistent with observations, which record negative changes over most of the continent. However, positive changes, which are also observed, are not captured. Although not shown here, comparison of the ENS104 mean against the ENS315 mean suggests that the precipitation patterns in the two are very similar, but the decreases generally tend to be higher in the ENS104 mean. The precipitation decreases in ENS104 conversely tend to be smaller than in ENS16.

The ENS104 mean LGM AMOC strength anomaly (Δψmax) is -2.8±2.8 Sv, and the range is -8 to 4.7 Sv (Fig. 8). These estimates lie at the low end of the Δψmax predicted by nine Paleoclimate Modelling Intercomparison Project phase 2 (PMIP2) (Weber et al., 2007) and eight PMIP3 (Muglia and Schmittner, 2015) coupled model simulations. However, they do not include the more negative Δψmax predicted by Völker and Köhler (2013), for instance. The ENS104 mean LGM-PRE AABW cell strength in the Atlantic Ocean (Δψmin) is 0.1±1.2 Sv. A positive Δψmin represents an LGM decrease in cell strength as we keep the original (negative) sign for anticlockwise flow of Antarctic water. A negative Δψmin conversely represents an LGM increase in cell strength. However, the difference between the LGM and PRE Atlantic AABW here is not statistically significant. The range of Δψmin is -4.3 to 4.3 Sv, roughly comparable to the range of Δψmin predicted in Weber et al. (2007) (see also Muglia and Schmittner, 2015) but excluding the much larger ψmin⁡ increase predicted by Kim et al. (2003), for example. As shown in Fig. 9, the northern limit of the ENS104 mean LGM AABW cell is roughly at the same latitude as in the pre-industrial simulations. The maximum depth reached by the ensemble mean AMOC base is also similar to pre-industrial. Observations (Lynch-Stieglitz et al., 2007; Lippold et al., 2012; Gebbie, 2014; Böhm et al., 2015), conversely, suggest that the LGM AMOC shoaled to less than 2 km, raising its base depth by 2500 and 600 m at the north and south ends of the return flow, respectively. The LGM AABW, in turn, is thought to have filled the deep Atlantic below 2 km, reaching as far north as 65∘ N, which is approximately 25∘ north of its modern northern limit (Oppo et al., 2015).

Although not shown here, the ENS16 ensemble members tend to exhibit a shoaling of the AMOC and enhanced penetration of AABW. With regard to Δψmax and Δψmin, one can see from Fig. 8 that these tend to be more negative (i.e. weaker AMOC and stronger AABW) than in ENS104. The Δψmax and Δψmin in ENS315 tend to conversely be more positive. In ENS104, we also find a positive relationship between Δψmax and ΔCO2 (r=0.57) and a negative relationship between Δψmin and ΔCO2 (r=-0.42). The relationships are reproduced in ENS315 (r=0.59 and r=-0.36, respectively). We additionally find, in both ENS104 and ENS315, negative correlations between Δψmax and Δψmin (r=-0.62 and -0.63), Δψmin and ΔSAT (r=-0.4 and -0.4), and Δψmax and ΔSIA (r=-0.62 and -0.66), as well as positive correlations between Δψmax and ΔSAT (r=0.68 and 0.66), and Δψmin and ΔSIA (r=0.37 and 0.42). Based on these relationships, we hypothesise that increasing LGM AABW strength led to an expansion of the AABW cell. The latter in turn restricted the AMOC to lower depths and reduced its overturning rate (e.g. Shin et al., 2003). The increase in AABW strength was likely driven by increases in sea ice enhancing brine rejection. Sea ice increases in the North Atlantic may have additionally weakened the AMOC cell by locally reducing deep convection.

The relationships between Δψmax, Δψmin and ΔCO2 are also consistent with increasing ψmin⁡ (decreasing ψmax⁡) contributing to decreasing atmospheric CO2. The replacement of North Atlantic Deep Water (NADW) by AABW in the North Atlantic would, for instance, have led to a dissolution of deep-sea sediment CaCO3 due to AABW having a lower bottom water CO32- concentration than NADW (see, e.g. Yu et al., 2014). The increased CaCO3 dissolution flux would in turn have raised the whole ocean alkalinity, lowering the atmospheric CO2. Enhanced AABW production would also have caused the deep ocean to become more stratified, allowing more DIC to accumulate at depth and promoting further CaCO3 dissolution. A decrease in NADW formation could have additionally lowered atmospheric CO2 by reducing CO2 outgassing at the ocean surface and reducing the burial rate of deep-sea CaCO3 due to the concomitant increase in deep-sea DIC accumulation. Further investigation is, however, required to confirm these causal relationships.

As shown in Fig. 10, most of the ensemble members in ENS104 predict an LGM increase in terrestrial biosphere (ΔTerrC) and lithospheric

The ΔLithC stems from changes in the deep-sea CaCO3 burial flux and/or CaCO3 weathering/shallow water deposition flux and was initially calculated to ensure that carbon was being conserved over the LGM simulation.

The ENS104 mean ΔTerrC, ΔOceanC and ΔLithC, the signs of which are consistent with scenario 1, are reported in Table 4, alongside previous estimates from observational data- and model-based studies. From here, we can see that the mean ΔTerrC is only aligned with a handful of estimates and no studies so far report a negative ΔOceanC. Instead, ΔOceanC is estimated to be positive, primarily based on carbon isotope data. The loss of hundreds of petagrams of carbon from the ocean in response to terrestrial carbon growth has, however, been previously proposed (e.g. Zimov et al., 2006). Moreover, if we assume that 90 % of the atmospheric CO2 perturbation caused by the increase in terrestrial biosphere carbon reported in Table 4 gets removed by the ocean and sediments, the change in ocean carbon would be negative, even after adding the remaining carbon to be lost from the atmosphere to the ocean. We discuss what these results would likely mean for carbon isotope data in Sect. 3.2.6.

The positive ΔTerrC studies in Table 4 attribute the increase in terrestrial carbon to different factors: Zimov et al. (2009) and Zech et al. (2011) predict large increases in permafrost carbon, while Zeng (2003) ignores permafrost. Instead, the author attributes the glacial terrestrial carbon increase to the preservation rather than destruction of carbon in areas to be covered by the LGM ice sheets, lower soil respiration rates (in the active carbon pool) caused by a colder climate, as well as storage of carbon on exposed continental shelves. Here, neither the latter carbon accumulation mechanism, nor that of permafrost growth, are included. However, our model does attempt to capture the very slow rates of soil decomposition characteristic of permafrost (Williamson et al., 2006). We attribute the terrestrial carbon increase in our ensemble to a higher soil carbon inventory caused by a decreasing soil respiration rate. As can be seen from Fig. 10, vegetation carbon tends to conversely decrease at the LGM. As in Zeng (2003), we also do not assume that, as the ice sheets expand over terrestrial carbon, they destroy it. The lack of this potential loss term means that our ΔTerrC estimates may be higher than in many previous studies, irrespective of what the response of the terrestrial biosphere is to the LGM climate and CO2 forcings. O'ishi and Abe-Ouchi (2011), for instance, estimated that the LGM climate is responsible for the loss of 502 PgC of terrestrial carbon, while another 388 PgC is removed by the ice sheets. Zeng (2003) conversely proposed that 431 PgC is preserved under the ice sheets at the LGM. This number includes 315 PgC present during the interglacial and another 116 PgC accumulated in response to the glacial climate forcings, prior to insulation of the terrestrial carbon from the atmosphere by the ice sheet coverage. Here, analysis of ENS16 suggests that during the 1000 years of LGM ice sheet build-up, the terrestrial carbon inventory in the areas to be occupied by the ice sheets increases by between 6 and 444 PgC, yielding LGM “ice sheet or burial” carbon inventories between 318 and 1341 PgC (Table 5). This increase accounts for less than half of the total LGM change in terrestrial carbon (i.e. ΔTerrC) in the majority of simulations. However, if this burial carbon were to have been destroyed rather than preserved, ΔTerrC would be negative in all but three simulations, as opposed to positive in all but one simulation (Table 5).

Most of the terrestrial carbon increase in areas to be covered by the ice sheets in ENS16 is due to soil carbon, with vegetation carbon decreasing in all but one simulation. The range of terrestrial carbon increases and the associated burial carbon amounts include Zeng (2003)'s estimates. Our LGM burial carbon estimates would also accommodate an additional 250–550 PgC (Franzen, 1994) from increased glacial peat accumulation (Zeng, 2003). No observational data-based estimates of the LGM burial carbon inventory are available since only limited evidence exists for organic material being preserved by ice during glaciations (Franzen, 1994, and references in Weitemeyer and Buffett, 2006). Outside of the ice sheets, increases in the terrestrial carbon inventory in ENS16 are mostly due to soil carbon, which increases in all simulations. Vegetation carbon, conversely, decreases in the majority of simulations. Our range of carbon changes outside of the ice sheet areas include the 198 PgC increase predicted by Zeng (2003) as a result of reduced soil respiration.

Although not evaluated directly, it is likely that similar ice sheet/non-ice-sheet terrestrial carbon proportions than in ENS16 are found in ENS104 and ENS315 because of the similar climate change distributions in all three instances (see earlier sections). Although not shown here, the spatial distribution of ΔTerrC in ENS16 is also similar to that of the ENS104 (and ENS315) mean. The spatial distribution of ΔTerrC in ENS104 is shown in Fig. 11.

The largest increases in terrestrial carbon (≥20 kgC m-2) are found in North America and Europe/western Asia, both within and south of the Laurentide and Eurasian ice sheet margins (Fig. 11). Regions with smaller but still relatively large (≥10 kgC m-2) increases include the Andes and Patagonia regions, the southern tip of the African continent, eastern north Siberia and the grid cells just south of the Tibetan Plateau. The largest LGM decreases in terrestrial carbon (≥10 kgC m-2) conversely tend to be found in northwest North America, Beringia and the Tibetan Plateau region. Other regions with relatively large (≥5 kgC m-2) decreases include equatorial Africa and the deserts in central Asia. Everywhere else the LGM terrestrial carbon density increases by between 0 and 10 kgC m-2. Comparison against paleoecological reconstruction studies (Crowley, 1995) suggests that the simulated terrestrial carbon changes within the Laurentide and Eurasian ice sheet areas are of the wrong sign, except in northwest North America, since these studies assume the complete destruction of vegetation and soils in ice sheet areas. Discrepancies between the ENS104 and observations further arise from the rainforest regions, where the ensemble mean predicts terrestrial biosphere carbon density changes between -5 and 10 kgC m-2, well above observed changes of ∼-23 kgC m-2. It is important to note, however, that as suggested in Zeng (2007), the rate of decomposition of soil carbon at the LGM may have been slower than assumed in pollen data-based studies. The largest increases in terrestrial carbon density (∼40 kgC m-2) produced by the ensemble mean are comparable to those found in areas with permafrost growth (Zimov et al., 2006). However, the peaks are potentially misplaced, being located within and south of the Laurentide and Eurasian ice sheet covered areas, rather than in eastern Siberia and Alaska. Alternatively, terrestrial carbon increases in eastern Siberia and Alaska are simply underestimated in the ensemble mean and large increases in terrestrial carbon indeed took place within the ice sheet areas during glacial periods.

The large LGM decreases in terrestrial carbon in northwest North America and adjacent Beringia are likely caused by precipitation decreasing comparatively more than SAT and causing the decrease in photosynthesis to exceed the decrease in soil respiration. However, it is also noteworthy that, although not shown here, the regions with the largest decreases in terrestrial carbon density, namely northwest North America, Beringia and the Tibetan Plateau area, are also the regions with the largest terrestrial carbon densities in the pre-industrial ENS315 mean. We further note that the Tibetan soil carbon peak is overestimated in the latter, and the North American soil carbon peak misplaced, compared to observations. We attribute the first discrepancy to the lack of soil weathering in the model and the inclusion of land use effects in the observational data-based estimate (Holden et al., 2013b; Williamson et al., 2006). The second discrepancy is attributed to the lack of explicit representation of permafrost and the absence of moisture control on soil respiration (Williamson et al., 2006).

The ENS104 mean LGM total POC export flux anomaly (ΔPOCexp) is -0.19±1 PgC yr-1 and the range is -2.57 to 2.56 PgC yr-1, roughly consistent with previous model-based estimates (e.g. Brovkin et al., 2002, 2007; Bopp et al., 2003; Chikamoto et al., 2012; Palastanga et al., 2013; Schmittner and Somes, 2016; Buchanan et al., 2016). As shown in Fig. 12, compared to ENS104, the POC flux decreases in ENS315 and ENS16 tend to be smaller and larger, respectively. ΔPOCexp is positively correlated with Δψmax (r=0.72 and 0.79) and negatively correlated with Δψmin⁡ (r=-0.62 and -0.58) in both ENS104 and ENS315. The correlations potentially suggest that decreasing AMOC strength and increasing AABW production led to decreasing POC export. One possible mechanism is enhanced deep ocean stratification due to increasing AABW formation leading to not only more efficient trapping of DIC at depth (see above) but also nutrients and therefore reduced availability in the euphotic zone. All else held constant, a weaker and shallower AMOC cell would also inhibit the transfer of nutrients from the deep ocean to the surface. A negative correlation can additionally be found between ΔPOCexp and ΔSIA (r=-0.55 and -0.6), probably because no primary production occurs beneath the sea ice surface. Increasing sea ice area at the LGM therefore leads to decreasing POC export flux. This would also explain the largest ENS104 mean decreases in POC export flux, shown in Fig. 13, coinciding with increases in sea ice fraction.

The largest ENS104 LGM increases in POC export flux conversely occur at around 50∘ S, roughly in front of the Antarctic sea ice margins. Increases in POC export are also simulated close to the North Pacific and Atlantic sea ice margins, as well as in the eastern equatorial Pacific and the southwest Atlantic upwelling region. The increases in POC export flux at the sea ice margins are likely caused by the advection of unutilised nutrients from underneath the sea ice. However, they may additionally be due to the enhanced iron availability from the increased supply of aeolian dust, particularly in the Southern Ocean and North Pacific since these are strongly limited by iron (Ridgwell et al., 2007). Iron fertilisation may also explain the increases in POC export flux in the eastern equatorial Pacific and in the southwest Atlantic upwelling region. Comparison against observations suggests that the ensemble mean POC flux changes immediately north, and south of the Antarctic sea ice margins align with observations of increased and reduced marine productivity in the subantarctic (∼45 to 60∘ N) and Southern Ocean, respectively (Kohfeld, 2005; Kohfeld et al., 2013; Jaccard et al., 2013; Martínez-García et al., 2014). The simulated decreases in export flux in the Arctic and subarctic Atlantic (i.e. above approximately 50∘ N), and the increases in export flux immediately south of 50∘ N are also in agreement with previous reconstructions (Kohfeld, 2005; Radi and de Vernal, 2008). The mostly lower LGM export fluxes at the Equator and in the South Atlantic are conversely inconsistent with the observational data of Kohfeld (2005). The decreases may be caused by the increases in productivity in high-nutrient, low-chlorophyll (HNLC) regions reducing the phosphate (the other limiting nutrient in GENIE-1 besides iron) availability for photosynthesis in other regions. They may additionally be due to the model not simulating enhanced nutrient inventories in response to enhanced weathering or reduced shallower water deposition of organic matter. The model also does not vary wind speed, which may have resulted in stronger tropical upwelling in the Atlantic at the LGM. The evidence is more ambiguous (or missing) for the Pacific (Jaccard et al., 2010; Kohfeld and Chase, 2011; Kohfeld, 2005; Costa et al., 2016) and Indian oceans (Kohfeld, 2005; Singh et al., 2011) and is therefore not discussed in more detail here.

The ENS104 mean land-to-ocean bicarbonate flux scaling factor (GWS) is 1.16±0.24 (corresponding to a percentage change in the land-to-ocean bicarbonate flux, %LOC, of 38.67), and the range is 0.52 to 1.5 (corresponding to a %LOC between -49.33 and 50). As shown in Fig. 14, the GWS in ENS104 tends to be larger than in ENS315 and smaller than in ENS16. There is also a negative correlation between GWS and ΔCO2 (r=-0.52) in ENS315, suggesting that increasing the input of bicarbonate to the ocean leads to a decrease in CO2 by raising the inventories of ALK and DIC in a 2:1 ratio. In ENS104, however, r is below the 0.05 significance level, suggesting that it is less important.

The ENS104 mean global deep-sea CaCO3 burial flux anomaly (ΔCaCO3bur) is 0.036±0.045 PgC yr-1 and the range is -0.098 to 0.139 PgC yr-1. The mean value is approximately 3 times larger than the observed value (Catubig et al., 1998), although the latter still falls within the range of simulated values. As shown in Fig. 15, ΔCaCO3bur in ENS104 tends to be higher than in ENS315 and lower than in ENS16. The change in global deep-sea CaCO3 burial flux anomaly is strongly determined by GWS, as suggested by the positive correlation (r=0.88 and 0.9) between the two, in both ENS104 and ENS315. Increasing %LOC (the percentage change in the land-to-ocean bicarbonate flux) should indeed enhance the CaCO3 burial flux as increasing ALK means the deep ocean CO32- will eventually increase. The latter in turn would cause the saturation horizon to fall, allowing CaCO3 to accumulate over greater areas (which are now exposed to undersaturated waters) (Sigman and Boyle, 2000). The input of ALK to the surface ocean would also increase the rate of CaCO3 export production (enhancing the sediment deposition flux of CaCO3), since as discussed in Chikamoto et al. (2008), the latter is proportional to the production rate of POC (which is equal to the POC export flux), together with the sea surface saturation state with respect to CaCO3, in GENIE-1. There is indeed also a positive correlation between %LOC and the global change in CaCO3 export flux (r=0.27 and 0.4), and between the latter and ΔCaCO3bur (r=0.34 and 0.45) in both ENS104 and ENS315.

The ENS104 mean spatial distribution of ΔCaCO3bur is shown in Fig. 16. Relatively large increases in burial flux (≥0.5×10-5 mol cm-2 yr-1) can be found at around 50∘ S, in the North Pacific and to a lesser extent the North Atlantic. In other regions, the burial flux is significantly lower or negative, with the largest losses (≤-0.5×10-5 mol cm-2 yr-1) occurring in the North Atlantic and arctic regions. The only exception is the western North Atlantic, which exhibits a large increase in burial. A comparison of the results against the reconstructions of Catubig et al. (1998) is somewhat difficult, as the coverage is poor but overall CaCO3 burial was higher in the North Atlantic and the Pacific, and lower in the tropical and South Atlantic, and the Indian Ocean and Southern Ocean.

As shown in Table 4, a frequent argument for a lower glacial terrestrial carbon inventory is the reconstructed mean glacial ocean δ13C value of approximately 0.35 ‰ lower than present due to the fact that plants discriminate against 13C during photosynthesis. In our simulations, conversely, it follows that the increase in glacial terrestrial carbon inventory would have resulted in an increase in ocean δ13C. Decreasing SSTs and increasing CaCO3 weathering would have, moreover, likely raised it further (in the first instance, by enhancing fractionation at the air–sea interface, and in the second instance, through the input of isotopically heavy weathering products). However, as noted in Zeng (2007), the interpretation of the δ13C value can be complicated by factors such as the impact of enhanced glacial carbonate ion concentrations on δ13C in foramifera shells (Lea et al., 1999). In addition, there are other processes in our model which may have counteracted at least part of the increase in ocean δ13C. These include reduced marine productivity (e.g. Zimov et al., 2009), as phytoplankton discriminate against 13C during photosynthesis, giving the marine organic carbon reservoir a low δ13C. However, we note that the sign of this impact would additionally depend on the associated changes in organic matter remineralisation and burial. Another relevant process is greater sea ice area, which can lower the ocean δ13C by reducing the air–sea gas exchange and therefore the net transfer of 13C into the ocean (Stephen and Keeling, 2000). Moreover, we propose that adding missing processes could have decreased ocean δ13C even further. These include weaker surface winds (while in our model these are fixed), again through reduced air–sea gas exchange (Menviel et al., 2015), as well as enhanced weathering and reduced deposition of organic carbon at continental margins due to lower sea levels (Wallmann, 2014). Yet, further research is required here as one recent study suggests that taking into account these processes would most likely not (the possibility is not completely ruled out) allow reconciliation of a positive ΔTerrC with the observed mean glacial ocean δ13C value (Jeltsch-Thömmes et al., 2019).

Missing processes would likely also be needed to reconcile our lower POC export fluxes with deep ocean oxygen records. These records tend to indicate there was a decrease in LGM deep ocean oxygen concentration (Jaccard et al., 2016). Lower POC export fluxes would conversely have resulted in an increase in deep ocean oxygen due to reduced oxygen consumption at depth. When observed over sufficiently large areas, lower oxygen concentrations can support the presence of an enhanced ocean carbon inventory as deoxygenation can be explained by reduced ocean ventilation (the sole input of oxygen is from the ocean surface) (Wagner and Hendy, 2015). The reduced ventilation is, in turn, assumed to have led to the accumulation of a significant amount of DIC in the ocean interior. Thus, explaining the lower deep ocean oxygen concentrations without having to reduce ocean ventilation as extensively as suggested by previous studies would as a minimum likely require LGM export production to have increased rather than decreased. However, it may be possible to increase deep ocean oxygen consumption by increasing organic matter at depth but keeping the surface POC export flux constant. This would require adding missing processes such as increasing remineralisation depth with decreasing ocean temperature and increasing ballasting into our model (Kohfeld and Ridgwell, 2009; Menviel et al., 2012).