Antarctic Bottom Water and North Atlantic Deep Water in CMIP6 models

Deep water formation is the driver of the global ocean circulation, yet it was poorly represented in the previous generation of climate models. We here quantify biases in Antarctic Bottom Water (AABW) and North Atlantic Deep Water (NADW) formation, properties, transport and global extent in 35 climate models that participated in the latest Climate Model Intercomparison Project (CMIP6). Several CMIP6 models are correctly forming AABW via shelf processes, but in both hemispheres, the large majority of climate models form deep water via open ocean deep convection, too deep, too often, over too 5 large an area. Models that convect the least form the most accurate AABW, but the least accurate NADW. The four CESM2 models with their “pipe” / overflow parameterisation are among the most accurate models. In the Atlantic, the colder AABW, the stronger the abyssal overturning at 30◦S, and the further north the AABW layer extends. The saltier NADW, the stronger the Atlantic Meridional Overturning Circulation (AMOC), and the further south the NADW layer extends. In the Indian and Pacific oceans in contrast, the fresher models are the ones who extend the furthest regardless of the strength of their abyssal 10 overturning, most likely because they also are the models with the weakest fronts in the Antarctic Circumpolar Currents. There are clear improvements since CMIP5: several CMIP6 models correctly represent or parameterise Antarctic shelf processes, fewer models exhibit Southern Ocean deep convection, more models convect at the right location in the Labrador Sea, bottom density biases are reduced, and abyssal overturning is more realistic. But more improvements are required, e.g. by generalising the use of overflow parameterisations or by coupling to interactive ice sheet models, before deep water formation, and hence 15 heat and carbon storage, are represented accurately.

. The 35 CMIP6 models used in this study; their ocean component; horizontal resolution in • latitude x • longitude; vertical grid type (ρ means isopycnic, σ terrain-following, several symbols a hybrid grid) and number of vertical levels; and official reference. N/A indicates that no paper has been published yet for the CMIP6 configuration.

Model name
Ocean component Horizontal Vertical Reference 1 ACCESS-CM2 MOM5 1 x 1 z* 50 N/A 2 ACCESS-ESM1-5 MOM5 1 x 1 Although we used the full historical run for robustness verifications, we present only the results for the period January 1985 to December 2014, for consistency with the observational products. Note that we neither detrended the CMIP6 historical run nor substracted the pre-industrial control run, again for consistency with observations (which feature the climate change 70 trend). These observations are the full-depth ocean temperature and salinity climatologies from the World Ocean Atlas 2018 (Locarnini et al., 2018;Zweng et al., 2018, respectively), the annual mixed layer depth climatology of de Boyer Montégut et al. (2004), and the global bathymetry GEBCO (GEBCO Compilation Group, 2019).

Computations: deep water properties, transports and extents
To start with, when necessary, we computed the monthly mixed layer depth (MLD) of the CMIP6 models as per the CMIP6 75 procedures by first computing the monthly mean density σ θ from their monthly practical salinity and potential temperature.
The MLD is then detected as the depth where σ θ differs from that at 10 m depth by more than 0.125 kg m −3 . Note that a different threshold of 0.03 kg m −3 is used in the observational reference (de Boyer Montégut et al., 2004). We could then  Brodeau and Koenigk (2016). That is, for each month and each region, we keep only those grid cells where the MLD exceeds a critical value and sum the product MLD x cell area. We work with the maximum value of each year. As in Brodeau and Koenigk (2016) and Koenigk et al. (2020), we use a critical value of 700 m in the Nordic seas as it is the depth of the sill that connects them to the rest of the world ocean, and 1000 m in the Labrador Sea. As in e.g. Heuzé et al. (2013); De Lavergne We quantify biases in the models by computing the root mean square error (model minus reference) in temperature, salinity and density σ θ at the sea floor grid cell. To do so, all models had to be interpolated onto the reference's grid. Note that we purposely keep σ θ instead of σ 4 as σ θ is the density used in the models' code to notably compute the MLD. For later calculations, we also compute the temperature and salinity of the water masses AABW and NADW by taking their average 90 properties over a specific region. As we will show in section 3.1, the AABW formation region really differs from model to model; as such, instead of using a limited region as in Johnson (2008), we detect AABW as having the temperature minimum anywhere deeper than 2000 m and south of 50 • S. For NADW, we produce two flavours: NADW SPG as having the salinity maximum anywhere deeper than 1000 m in the small area of SPG defined by Johnson (2008, 55 • W to 54 • W, 53 • N to 63 • N, yellow box on Fig. 2); and NADW GIN , the salinity maximum anywhere deeper than 1000 m in the GIN sector defined above.
floor to surface of the velocity vo, the AMOC is defined as the southward subsurface maximum. We could not directly use the meridional overturning circulation output "msftmz" as it is provided by hardly a third of the models; it is in kg s −1 instead of m 3 s −1 , requiring division by the density, which we only have the monthly mean of;

105
for most models, the Indian and Pacific oceans are provided as one joint region, so we could not have obtained the SMOC in each basin.
Having to interpolate the irregular model grids onto the sections instead of directly using the model output may have introduced some errors. But as the AMOC results of this manuscript and that of Menary et al. (2020) for the models and experiment we have in common are similar, we are confident in our MOC values. Note that two models, GFDL-ESM4 and NorCPM1, did not 110 provide vo, limiting our transport analysis to 33 CMIP6 models.
Finally, to investigate in CMIP6 the link found in Heuzé et al. (2015) between the SMOCs and the northward extent of AABW layer, we chose to re-create for CMIP6 models the Johnson (2008) maps of AABW and NADW volumes in the global ocean. However, using the same approach as Johnson (2008) whereby we would have to determine the characteristics of every water mass in each basin for each model is beyond the scope of this paper. Instead, as below the core of NADW the global 115 ocean (excluding the Arctic) is a mixture of NADW and AABW only, we determine at each depth the NADW and AABW contents from a conservative property χ using the mixture equation of Jenkins (1999): and Here, as in Johnson (2008), we consider the practical salinity and potential temperature as conservative enough to be used for these calculations. We then take the 50% content depth as the border between the NADW and AABW layers, i.e. anything with more than 50% AABW or less than 50% NADW is in the AABW layer, and the AABW thickness is the difference between the depth of that border and sea floor. We finally take the median of all the combinations: temperature or salinity, NADW properties from SPG or GIN, and AABW or NADW contents. For the NADW layer, we detect the NADW core as the maximum NADW 125 content with an extra criterion that the maximum must be larger than 80% NADW. Tests with values ranging from 60 to 100% yield similar values (not shown). Then the so-called NADW thickness is the thickness from the depth of the core to the NADW-AABW boundary (or to sea floor if there is no AABW). By working with a mixture of two water masses only, we could not try and detect the top of the NADW layer. Note that traditional methods of using a fixed temperature and/or salinity for water mass determination cannot be applied to potentially biased climate models. The northward extent of AABW in each basin is 130 defined as the northernmost latitude of the uninterrupted contour of thickness = 2000 m that starts in the Southern Ocean. We do the same for the southward extent of NADW in the Atlantic Ocean. For this part of the analysis, we show ony the results for NADW that originated in SPG; NADW that originated in GIN seems to leave the Nordic Seas in no model (not shown).

Results
In this section, we first look at deep water formation and properties in the Southern Ocean, then deep water formation and 135 properties in the North Atlantic. It is only in the last section that we analyse both water masses together, by determining their global transports and volumes. In this section, we talk only about CMIP6. The comparison with CMIP5 will come in section 4.

150
-GFDL-ESM4, HadGEM3-GC31-LL, IPSL-CM6A-LR, SAM0-UNICON and UKESM1-0-LL may overflow in the Ross sector, but we would need higher temporal resolution data to be certain; the other 11 models occasionally have a plume of dense water leaving the shelf, but it is nowhere near as dense as the shelf water it originates from (see video example of ACCESS-CM2).
In summary, in no model is there any ( ice cover, and the two MIROC who are ice-free (Mohrmann et al., subm.;Roach et al., 2020). In the Amery and Ross sectors, we need to distinguish between the models that have non zero DMV because of open ocean deep convection, and those with coastal polynyas. In the Amery sector, aside from MCM-UA-1-0 whose deep convection area is but a continuation of the (1850-2014, not shown), with GISS-E2-1-G convecting once in the Weddell sector and once in the Ross sector, and GISS-E2-1-G-CC thrice in the Weddell sector. The four CESM2 models do not, but they have an overflow parameterisation that artificially moves dense water from the shelf to the deep basins (Briegleb and Large, 2010). A pipe sucks the dense water on the shelves and releases it in the deep basin without having to cascade. This is why we cannot detect it on the overflow movies.

AABW properties
Does the way CMIP6 models form their AABW impact its characteristics, as it did in CMIP5 ( Fig. A1). The other 6 models have relatively accurate bottom salinity, but are biased warm (supp. Fig. A2). No model has regional biases, which means that 205 CMIP6 models are overall biased light or biased dense in the deep Southern Ocean (excluding the shelves).
The models with low biases in bottom density also tend to have zero to low DMVs in the Weddell Sea, but the relationship does not hold for maximum DMVs larger than 200 10 13 m 3 (Table 2). NorESM2-LM and -MM notably have low biases but very high DMVs, but they also do shelf overflows. Open ocean deep convection leads to a warming and salinification of bottom waters (Zanowski et al., 2015); one hypothesis is then that models that hardly convect stay closer to the bottom density 210 value they were initialised with. In the case of the CESM2 suite, the overflow parameterisation may help form accurate bottom water. Biases as RMSE are not the whole story though. As expected, we do find significant relationship (95% level) between the actual temperature and salinity of AABW and the DMV: in the Amery and Ross sectors, more deep convection leads to warmer AABW (correlation of +0.33 and +0.29 respectively) as in Wang et al. (2017). In the Weddell sector however, more deep convection leads to fresher AABW (correlation of -0.35), which in fact is consistent with the short-term response of the 215 Southern Ocean to deep convection in Zanowski et al. (2015). The multimodel mean AABW salinity is 34.606 ± 0.154; the reference value from Johnson (2008), 34.641. The multimodel mean AABW temperature is -0.45 ± 0.73 • C; the reference value from Johnson (2008), -0.88 • C. That is, the multimodel mean AABW is warmer and fresher than the reference, and more DMV worsens these biases. Note that the values of the individual models are given in supp. Table. A2.
To summarise, in the Southern Ocean, most models form their AABW by open ocean deep convection. In the Weddell 220 Sea, this convection seems tied to the Weddell Polynya activity, and impacts the AABW salinity most: more deep convection, fresher bottom salinity. In the Amery and Ross sectors, it is linked more to the bottom temperature: more deep convection, warmer bottom salinity. Models which seem to form dense water via shelf processes also exhibit deep convection, so we cannot determine whether overflows alone would make the Southern Ocean more accurate. Models that convect the least or not at all tend to be the most accurate; four or these, the CESM2 suite, may be aided by their overflow parameterisation (Briegleb 225 and Large, 2010; Snow et al., 2015); another one, NorCPM1, assimilates observations (Counillon et al., 2016).
We will study the impact of these biases on the global transport of AABW in section 3.3. But as we cannot do so without investigating the AABW -NADW tug of war in the Atlantic basin, let us first evaluate the representation of NADW in CMIP6 models.  11 https://doi.org/10.5194/os-2020-66 Preprint. Discussion started: 10 July 2020 c Author(s) 2020. CC BY 4.0 License.

North Atlantic deep water in CMIP6 models
In the North Atlantic, all 35 CMIP6 models of our study exhibit deep convection in the subpolar gyre (black contours on Fig.   2 and Table 3). As in CMIP5 (Heuzé, 2017), a large proportion of them convects not only in the Labrador Sea as the reference, but also intensely south of Iceland (Irminger Sea): is not surprising. Consequently, there is also a strong correlation between the DMV in the Weddell Sea and in SPG (+0.57): 250 models that convect a lot in the Weddell Sea convect a lot in SPG as well.
All models except INM-CM5-0 and NorCPM have deep convection in the GIN seas as well. Moreover, in GIN, models convect most years, with a minimum as high as 24/30 years (Table 3). There is more variability in the SPG, but likewise the majority of models convect all years. Besides, they convect too deep. While in the Southern Ocean, deep convection to the sea floor can happen (Killworth, 1983), in the North Atlantic it should not go much beyond 1000 m (e.g. Våge et al., 2009

North Atlantic bottom properties
The picture is less grim regarding bottom properties biases (shading on Fig. 2). Three models have bottom density biases -ACCESS-ESM1-5 is accurate in the SPG sector but biased dense (salty) in GIN. which was in fact the original motivation for that parameterisation (Briegleb and Large, 2010). NorCPM1 is somewhat disappointing; it is built on NorESM2-LM and is supposed to have improved performances thanks to data assimilation (Counillon et al., 2016). Its bottom density is indeed better than NorESM2-LM's in SPG, but is biased even denser (saltier) in GIN. There variability. Investigating the exact cause of the biases in GIN is beyond the scope of this paper, not least because in the next section, we will show that NADW GIN does not contribute to the global NADW in CMIP6 models. For now, we can conclude that the bottom property biases in GIN are not related to deep water formation in the region.

Global transport of NADW and AABW in CMIP6 models
In this last section, we shall determine the global fate of NADW and AABW once they leave their source regions. For NADW,   (Huussen et al., 2012), so unsurprisingly, from the weak MCM-UA-1-0 (1.5 ± 1.6 Sv) to the strong GFDL-CM4 (11 ± 18 Sv), all models are in that range and the 325 multimodel median is 3.0 ± 2.5 Sv. This is a remarkable improvement since CMIP5, where a majority of models had an Indian SMOC close to 0 (Heuzé et al., 2015). In the Pacific finally, Lumpkin and Speer (2007) estimated the MOC to be 11 ± 5 Sv.
MCM-UA-1-0 is again the weakest (3.9 ± 1.9 Sv), and the only model that falls out of the observational range, resulting in a multimodel median of 5.9 ± 3.0 Sv. In summary, the AMOC and southern MOCs are rather accurately represented in CMIP6 models!

330
The across-model correlations between the transports are strong and significant (95% level): the stronger the SMOC in the Indian Ocean, the stronger as well in the Pacific Ocean (correlation of +0.37). In contrast, a strong SMOC in either of these basins corresponds to a weak SMOC in the Atlantic (Atlantic-Indian, correlation of -0.45; Atlantic-Pacific, -0.34). And a weak SMOC in the Atlantic corresponds to a strong AMOC (correlation of -0.30), as previously found by Patara and Böning (2014) in the NEMO model. We are obviously not implying causation from the correlations, but it is interesting to find relationships 335 between the biases quantified in sections 3.1 and 3.2 and the transports. In agreement with Patara and Böning (2014), a stronger Atlantic SMOC is associated with lower temperature biases (correlation of 0.29), that is, colder AABW (-0.35), whereas a stronger Pacific SMOC is associated with stronger density biases (+0.36). A stronger AMOC is associated with larger biases in temperature and salinity in SPG (correlations of +0.33 and +0.37 respectively), and in particular a saltier NADW SPG (+0.34, as in the paleoclimate simulations of Menviel et al., 2020). The Atlantic SMOC is the only transport that is linked to the . Unlike e.g. Menary et al. (2015) or Koenigk et al. (2020), we find no link between the MOCs and the models' horizontal resolution.

345
In line with Heuzé et al. (2015), we expect the transports to impact the interbasin spread of NADW and AABW, that is, that the stronger the transport, the further from its source the water mass will travel. To investigate this, we recreated the AABW and NADW thickness maps of Johnson (2008)  As explained in the methods, the NADW layer in the Indian and Pacific oceans is most likely biased by our calculation method 355 that takes into account only two water masses, and thus shall not be discussed further.
After extracting the southernmost extent of NADW and northernmost extents of AABW for each model (see supp. tables B1 and B2), we do find, as expected, that the stronger the AMOC, the further south NADW extends in the Atlantic (correlation of 0.32). And the stronger the Atlantic SMOC, the further north AABW extends in the Atlantic (correlation of 0.40). As we previously found an anticorrelation between the AMOC and the Atlantic SMOC across CMIP6 models, the Atlantic balance 360 is complete: models with strong AMOC and weak SMOC have their Atlantic dominated by NADW (e.g. CESM2-WACCM), whereas those with a weak AMOC and strong SMOC are filled with AABW (e.g. IPSL-CM6A-LR). And although there was no significant correlation between the DMV and the transports, we do find that the larger the DMV, the further the extent of NADW (DMV SPG, correlation of +0.34) or AABW (DMV Weddell, +0.51). We found no significant correlation between the northward extent in the Indian or Pacific oceans and either the SMOCs or DMVs, or with the strength of the ACC. There 365 are however relationships with their bottom properties: the northward extent of salty models is less than that of fresh models (correlations of -0.31 in the Indian and -0.44 in the Pacific). As we also find a strong positive relationship (correlation of +0.72) between the salinity of AABW and the salinity gradient across the ACC computed by Beadling et al. (2020), i.e., we find that the fresh models have a weak gradient to overcome, this result is not surprising. We can even speculate that in the absence of NADW, AABW would expand further north in the fresher models regardless of their SMOC. 18 https://doi.org/10.5194/os-2020-66 Preprint. Discussion started: 10 July 2020 c Author(s) 2020. CC BY 4.0 License.
In conclusion, in CMIP6 models as in the real ocean, deep convection impacts bottom water characteristics and biases: in the Southern Ocean, deep convection seems associated with more biased deep waters; in the North Atlantic, the more the models convect, the least biased they are. Either way, these biases then impact the deep water transport: a saltier NADW is associated with a stronger AMOC; colder AABW, stronger Atlantic SMOC. These transports then impact the location of the "NADW - This relationship does not hold anymore for CMIP6, and it is the models that convect the least that tend to be the most accurate ( Fig. 1 and table 2). It may be because many models are now artificially prevented from opening polynyas and convecting in the 385 Weddell Sea (Mohrmann et al., subm.). However, as the Weddell Polynya has now reopened in the real ocean (Campbell et al., 2019), future models may remove their "polynya-prevention" schemes again. Another reason for CMIP6 models seemingly not needing Southern Ocean deep convection to have accurate bottom properties may be that, as we showed in this paper, several CMIP6 models successfully represent shelf processes. This was an unexpected result considering that horizontal resolutions have not increased much since CMIP5, suggesting that models have improved their parameterisations instead (Danek et al.,390 2019). Regardless of the formation process, bottom density biases are smaller in CMIP6 than they were in CMIP5 (RMSEs on  (Zanowski et al., 2015).
In the North Atlantic, to the best of our knowledge, most CMIP5 studies focussed on the relationship between deep water formation and the AMOC or the warming hole (e.g. Menary and Wood, 2018) but did not investigate bottom property biases.
The one exception is Ba et al. (2014) who found a recurrent cold bias; with the World Ocean Atlas 2018 as reference, we find 400 in contrast that most CMIP6 models have a warm bias at the bottom of the North Atlantic. Deep water formation in the North Atlantic in the majority of CMIP5 models occurred too often, too deep, over too large an area (Heuzé, 2017). This sentence is still valid for CMIP6 ( Fig. 2 and table 3). One noticeable improvement (?) is that the models whose CMIP5 predecessor convected only in the Irminger Sea now convect in the entire subpolar gyre, including the Labrador Sea. Unfortunately, some of the models that performed well in CMIP5 when considering the location of deep convection in the SPG, i.e. had a relatively 405 small area in the Labrador Sea, have also expanded to the entire SPG (e.g. the CNRM family). That is, the inaccurate models may be on the way to improvement, most likely because the Arctic sea ice is better represented in CMIP6 than in CMIP5 (Shu et al., 2020), but the ones that were relatively accurate have degraded. The same holds for the Nordic seas: CMIP6 models are convecting even more than CMIP5 models did, and they already were convecting too much. In an increasingly warmer and icefree climate, Lique and Thomas (2018) predict that deep water formation would migrate from the North Atlantic subpolar gyre 410 to its subtropical gyre, and from the Nordic seas to the Arctic. Liu et al. (2019) adds that this will depend on whether meltwaters will most strongly impact the stratification, shutting down deep convection, or the horizontal gradients and hence the winds, pushing meltwater away from convection areas. For now, we observe that from the very icy CMIP5 to the more accurately de-iced CMIP6 models, deep water formation regions just expanded to occupy most of the space available in SPG and GIN. It is unclear whether increasing the resolution of future models would solve this issue: Danek et al. (2019) dramatically reduced 415 mixed layer depths in SPG by using an adaptative mesh with 5-15 km resolution, while Koenigk et al. (2020) finds that DMVs in the SPG become even larger in the high resolution versions of the models that participated to HighResMIP. Without changing the horizontal resolution, a more systematic inclusion and better representation of the stratosphere may be enough to reduce deep convection in the North Atlantic (Haase et al., 2018).
Regarding the transports, as noted by Menary et al. (2020) the AMOC is stronger in CMIP6 than in CMIP5, which they 420 blame on the aerosol forcing. Except for INM-CM5 that is now way too strong, or which uploaded incorrect velocity fields, this increase is not that strong and most models are in the obervational range. In the case of the CNRM family, a stronger AMOC is in fact a much more accurate AMOC (from 12 Sv in CMIP5 to 19 Sv in CMIP6). The NorESM models have a weaker AMOC in CMIP6, which is more accurate than their CMIP5 version (from 32 Sv in CMIP5 to 21 Sv in CMIP6). The two highest resolution models have weakened so much that their AMOC is too low (GFDL-CM4 and MPI-ESM1-2-HR). This 425 seems in contradiction to Koenigk et al. (2020) who found that increased resolution in HighResMIP leads to a stronger AMOC, but their result is mostly true when the models reach an eddy-resolving resolution. Which they do not, here, in CMIP6. It is harder to determine whether the Southern MOCs at 30 • S have improved since the values from inverse modelling (Lumpkin and Speer, 2007) and observations (Huussen et al., 2012) have very large uncertainties. All that we can say is that the Atlantic SMOC is stronger in CMIP6, so that only the GISS family continues having an Atlantic SMOC around 0 Sv. In the Indian 430 Ocean, no model has a transport of 0 anymore, which resulted in a doubling of the multimodel mean from 1.6 Sv in CMIP5 to 3 Sv in CMIP6, giving it the same importance as the Atlantic SMOC. The Pacific SMOC remains the strongest of the three and sees no significant difference between CMIP5 and CMIP6 except for the two models that used to be around 0, INMCM4 (INM-CM5-0 is now at 10 Sv) and GISS-E2-H (GISS-E2-1-H now at 7 Sv). As in CMIP6 the Southern Ocean representation from the bottom (this manuscript) to the top (Beadling et al., 2020) has improved, as well as the ACC (also Beadling et al.,435 2020), it is no surprise that more models are now capable of exporting AABW to the rest of the world ocean. To the best of our knowledge, the global extent of AABW and NADW, presented here for CMIP6 on Figs 3 and 4 respectively, was not assessed in CMIP5, so we cannot determine whether improved Southern Ocean characteristics lead to an improved global water mass distribution.
What can we expect from a hypothetical CMIP7? Higher resolution, most likely, although that was already expected from 440 CMIP6 and did not happen. As explained above and by Koenigk et al. (2020) or Danek et al. (2019), a higher resolution would not necessarily improve deep water formation. Holt et al. (2017) goes as far as stating that shelf processes will not be correctly represented until the horizontal resolution remains lower than 1/72 • , which they expect might be reachable by the most advanced computers within 10 years. Unfortunately, we do not all have access to these computers, so that even now, computing the global monthly mixed layer depth of the highest resolution model (GFDL-CM4, 1/4 • ) required over 600 core 445 hours for the 165 years of the historical run. Higher resolution output will be impossible to manage, unless cloud-computing solutions such as PANGEO become the norm (Odaka et al., 2020). Instead of increasing the resolution, a seemingly easier solution would be to improve parameterisations (Holt et al., 2017), especially overflow parameterisations (Snow et al., 2015). Briegleb and Large (2010) first showed that an overflow parameterisation to transport water from the Nordic Seas to the rest of the North Atlantic resulted in an improved representation of the ocean there. In CMIP6, the CESM2 models with their 450 "pipes" in the North Atlantic and Antarctic shelves were among the most accurate models, especially for AABW. It would be interesting to see whether such a parameterisation on a different model would yield the same results, or whether the CESM2 models are just very accurate. Efforts could also concentrate on improving other components of the climate model, for example the atmosphere, as an improved representation of the stratosphere would supposedly decrease unrealistic deep water formation (Haase et al., 2018). But where most progress can probably be made is in the cryosphere. As deep water formation is tied to convection, especially the CESM2 family that has an overflow parameterisation. In the North Atlantic (section 3.2), models convect too often, too deep, over too large an area, but in the subpolar gyre that area has migrated from the Irminger Sea (in CMIP5 models) to the more accurate Labrador Sea. The models that convect the most in the North Atlantic subpolar gyre also have the least biased NADW. NADW that forms in the subpolar gyre is the only one that occupies the world ocean; NADW from the Nordic seas appears to stay in the Nordic seas. The saltier NADW, the stronger the AMOC, and the further south 470 the extent of NADW (section 3.3). That extent is limited by the strength of the abyssal overturning in the southern Atlantic or SMOC, with stronger Atlantic SMOC (caused by colder AABW) resulting in a further northward extent of AABW. In the Indian and Pacific oceans, the extent is directly related to the AABW properties, not the SMOCs: models with a comparatively fresh AABW are also the ones with weak fronts across the Antarctic Circumpolar Current, and hence can travel the furthest north. In summary, for both deep water masses in CMIP6, their formation impacts their properties, which impact their transport 475 and global extent, which in turns will have large impacts on global predictions of thermal expansion and sea level rise (Zickfeld et al., 2017), carbon storage (Tatebe et al., 2019), ecosystem changes (Sweetman et al., 2017) etc. Although CMIP6 models represent AABW and NADW more accurately than CMIP5 models did, a lot still need to be improved, especially deep water formation (section 4).
How to improve deep water formation in climate models then? A higher horizontal resolution may not be the answer as, 480 depending on the model, it either reduces (Danek et al., 2019) or increases even further deep convection (Koenigk et al., 2020).