No Emergence of Deep Convection in the Arctic Ocean Across CMIP6 Models

As sea ice disappears, the emergence of open ocean deep convection in the Arctic, which would enhance ice loss, has been suggested. Here, using 36 state‐of‐the‐art climate models and up to 50 ensemble members per model, we show that Arctic deep convection is rare under the strongest warming scenario. Only five models have convection by 2100, while 11 have had convection by the middle of the run. For all, the deepest mixed layers are in the eastern Eurasian basin. When that region undergoes a salinification and increasing wind speeds, the models convect; yet most models are freshening. The models that do not convect have the strongest halocline and most stable sea ice, but those that lose their ice earliest ‐because of their strongly warming Atlantic Water‐ do not have a persistent deep convection: it shuts down mid‐century. Halocline and Atlantic Water changes urgently need to be better constrained in models.


Introduction
The Arctic Ocean is changing.The resulting reduction in Arctic sea ice extent and thickness (Mallett et al., 2021;Meier & Stroeve, 2022) could enhance vertical mixing: More brine may be rejected year-round as younger, saltier ice desalinates (Peterson, 2018) or as sea ice reforms in winter over the now seasonal-ice areas (Onarheim et al., 2018), while the ice-freed regions may become more susceptible to wind stirring (Timmermans & Marshall, 2020).In the Eurasian Arctic, the process known as Atlantification (Årthun et al., 2012), whereby waters of Atlantic origin are becoming warmer (Smedsrud et al., 2022) and penetrate further into the Arctic, may be further weakening the stratification and enhancing sea ice melt.These led Polyakov et al. (2017) to hypothesize that the Arctic may start exhibiting deep convection in winter, which, along with an unabated heat loss to the atmosphere (Shu et al., 2021), would enhance the Atlantification, leading to even more sea ice loss and polar amplification.Besides, the net effect on the ecosystem and on carbon uptake of deeper mixed layers, and resulting upwelling of warm waters and enhanced ice melt, are unclear (e.g., Mortenson et al., 2020).
Arctic deep convection under high emission scenarios has indeed been found in individual climate models by Lique et al. (2018) using HiGEM, and Brodeau and Koenigk (2016) and Bretones et al. (2022) using EC-Earth, and mixed layers depths approaching 500 m have been observed north of Svalbard by Pérez-Hernández et al. (2019).However, these most likely are the exception.The latest observations in the deep Eurasian Arctic (Schulz et al., under review) yielded mixed layers no deeper than 130 m, even in winter, in agreement with past observations of such a strongly stratified ocean as reviewed by Timmermans and Marshall (2020).Peralta-Ferriz and Woodgate (2015) also found that a deepening of mixed layers in the emerge and persist in the Arctic in the majority of Climate Model Intercomparison Project Phase 6 models, despite a drop in the Nordic Seas • Arctic deep convection occurs only when both surface salinity and winds are increasing, year round, yet most models are freshening • A deep halocline and stable ice cover hinder deep convection, while early ice loss due to warmer Atlantific water can also stop it rapidly

Supporting Information:
Supporting Information may be found in the online version of this article.
Arctic is unlikely, since stratification greatly dominates over the wind effect.Besides, in a recent Arctic study using models that participated in the Climate Model Intercomparison Project Phase 6 or CMIP6 (Eyring et al., 2016), Muilwijk et al. (2023) showed that there was no agreement among models regarding future stratification and the effect of Atlantification under the strongest warming scenario (SSP5-8.5, O'Neill et al. (2016)).
We here determine whether deep convection emerges in the Arctic in the future scenario SSP5-8.5 using all CMIP6 models and all their ensemble members for which the mixed layer depth (MLD) output was available, as described in Section 2. In Section 3, we detail the spatial and temporal patterns of future Arctic mixed layers and discuss possible reasons for these, focusing on variability and trends in surface properties, halocline and Atlantic layer depth, sea ice and heat fluxes, and winds.We conclude in Section 4.

CMIP6 Data
To investigate deep convection and its potential drivers in the future Arctic, we use all CMIP6 models and all their ensemble members that had monthly mixed layer depth ("mlotst"), ocean surface and full depth salinity ("sos" and "so") and temperature ("tos" and "thetao"), sea ice concentration ("siconc"), net heat flux into the ocean ("hfds"), and surface wind speed ("sfcWind") available on any of the Earth System Grid Federation nodes for the future scenario SSP5-8.5, for January 2015-December 2100.The models and their ensemble members are listed in Tables S1 and S2 in Supporting Information S1.
For AWI-CM-1-1-MR and GISS-E2-1-G, we used the sea ice thickness ("sivol") because sea ice concentration (SIC) was not available.We used the models' bathymetry ("deptho") when it was available, but had to generate it from the full-depth salinity for 11 models, as the last level with salinity data.Similarly, we used the models' grid cell area ("areacello") when it was available, and generated it from the grid boundary coordinates otherwise.Finally, for CESM2-WACCM, GFDL-CM4 and MRI-ESM2-0, several grid types were available and we chose for simplicity the regularized grid ("gr").For the other models, we took the one grid type available; see Tables S1  and S2 in Supporting Information S1.

Methods
The thresholds and choices of this subsection are discussed in Text S1 in Supporting Information S1.In agreement with Lique et al. (2018), we consider that there is deep convection in the Arctic if the mixed layer depth (MLD) exceeds 500 m.We here only perform a binary detection of deep convection, so we select the overall maximum MLD, in space or time depending on the analysis.The ensemble members that exhibit deep convection in the Arctic over the observed part of SSP5-8.5 (2015SSP5-8.5 ( -2023) ) are shown on the first figure and subsequently removed from the study, as they are already inaccurate at the beginning of the run.Consequently, of the originally 36 models, 27 remain for most of the analysis.We do not take biases in Nordic Seas MLD into account for model selection, as all CMIP6 models have spurious deep convection there (Heuzé, 2021).
The sea ice edge (Figure S1 in Supporting Information S1) is detected as the contour of 15% concentration or 10 cm thickness, averaged over 2040-2060 and 2080-2100.We define the Eurasian Basin (EB) as the region north of 80°N, longitudes 20°W-140°E, deeper than 1,000 m, and the Nordic Seas as the region of latitudes 66-80°N, longitudes 30°W-20°, deeper than 1,000 m.For the last analysis, we further split the EB into western and eastern parts along the 60°E meridian.See contours on Figure 1.Sea ice area over each region is the sum of the SIC multiplied by the grid cell area.For the models with no SIC, in order for all models to be on the same axes, we show instead the min-max normalized average sea ice thickness over the region, multiplied by the other models' maximum sea ice area there (70 and 80 million km 2 in the western and eastern Eurasian basins, respectively).
Following Shu et al. (2022), we computed the halocline depth as the first depth level where: where α and β are the thermal expansion and haline contraction coefficients, respectively, θ the potential temperature, S the salinity, and z the depth.As in for example, Heuzé et al. (2023), we computed the Atlantic Water

Geophysical Research Letters
10.1029/2023GL106499 core depth (AWCD) and temperature as the depth and temperature of the subsurface (deeper than 100 m) temperature maximum, respectively.
Besides, we compute trends for each grid cell, after interpolating all the parameters onto each model's mlotst grid.The trends are computed for each ensemble member separately, and averaged afterward.We consider that there is ensemble agreement if more than 50% of the ensemble members have a significant correlation/trend (determined using a t-test at 95% significance) of the same sign; we then present the median value and its ensemble-spread.Finally, we present the composite trends after grouping the models based on their Arctic deep convection  S1 and S2 in Supporting Information S1); when this number is larger than one, the figure shows the ensemble median.Black contours are the 1,000 m isobath.Red squares indicate the models for which all ensemble members have deep convection in the Arctic already over 2015-2023, and which are therefore not considered for further analysis.Locations discussed in the manuscript are indicated on the top-right panel: Cyan contour "EB" is the Eurasian Basin, further subdivided into western "W" and eastern "E" at 60°E; indigo dashed contour "GIN" is the Nordic Seas; and magenta arrow "SAT" is the St Anna Trough.
behavior, as described in the result section.We show potential drivers for the month prior to the deep mixed layers (see also Text S1 in Supporting Information S1) to try and determine causality.

Arctic Deep Convection Is Rare, Restricted Both in Space and Time
The maximum MLD reached in the Arctic over 2015-2100 varies strongly across models (Figure 1, note the logarithmic scale).The value does not exceed 100 m for some, such as CAMS-CSM1-0, while others such as CMCC-CM2-SR5 exceed 2,000 m in the majority of the EB.All models with deep MLD agree that the deepest values are in the EB, most commonly by St Anna Trough.Nine models have MLD exceeding the 500 m threshold for deep convection already over 2015 to present for all their ensemble members (red squares on Figure 1).All models presented on Figure 1 have spuriously deep MLD in the Nordic Seas, reaching all the way to the sea floor in some models, compared to the 500-1,000 m observed over the last decade (Abot et al., 2023).
After removing the models and ensemble members that are unrealistic in the present-day regarding their Arctic MLD (asterisks in Tables S1 and S2 in Supporting Information S1), 27 models remain.The temporal evolution of their MLD reveals four groups of models (Figure 2 and Table S3 in Supporting Information S1): 1.The first 11 models have no Arctic deep convection during the entire run.The Arctic MLD time series (blue lines, Figure 2) are mostly flat, with no year where the MLD exceeds 500 m, regardless of the ensemble member.The maximum MLD across these models and their ensemble members is often of the order of 100 m, that is, like currently observed (Schulz et al., under review).2. Six models have deep convection in the Arctic on rare occasions, in the middle of the run.The maximum number of years with deep convection for this group is 17 out of 86 (Table S3 in Supporting Information S1), but is most often four or fewer.3. Five models also have deep convection in the Arctic in the middle of the run, more often.It starts by 2030, peaks in the first half of the run, and then declines slowly, with the ensemble average back under the 500 m threshold by the end of the run.4. The last five models start convecting in the second half of the run, by 2070, and appear "stably" convecting at the end of the run.Unfortunately, only one ensemble member of one of the models (ACCESS-CM2) is available beyond 2100, so we cannot tell whether deep convection in these models would also decline later.
It is worth noting that ensemble members usually have a consistent behavior (the shading on Figure 2 usually agrees with the thick line), and models of the same family tend to belong to the same group.The only exception are the two ACCESS models (Figure 2), but there are large differences in their designs and implemented schemes of relevance for polar regions and deep convection in particular, as discussed in Mohrmann et al. (2021).
The Nordic Seas MLD (red lines, Figure 2) falls below 1,000 m for all models, and even below 500 m for two thirds of them.Five models do have maximum mixed layers east of Greenland deeper than 500 m (Figure 1).In observations, such comparatively shallow MLD nonetheless ventilate the East Greenland Current (Våge et al., 2018).Investigating whether this is the case here is beyond the scope of this paper, especially so since, to the best of our knowledge, the representation of that current in CMIP models has not been evaluated yet.Nordic Seas deep convection is not related to the models' behavior in the Arctic.Therefore, unlike suggested by for example,  (2015) argued that mixed layers would not deepen because of surface freshening and warming, regardless of changes in wind.We therefore now investigate the relationship between MLD and these four variables (surface salinity, temperature, SIC, and wind speed) in CMIP6 models.
Salinity variations are correlated to MLD variations (Table 1) and the correlation differs depending on the Arctic deep convection behavior of the models.Most models that have no deep convection in the Arctic or rarely (first two blocks) have a positive correlation between their March MLD and the surface salinity 1 month before, and a negative correlation with the ocean surface temperature: These models have shallower MLD when they are fresher and warmer (as expected in a changing Arctic, e.g.Peralta-Ferriz and Woodgate (2015)).In the two groups of convecting models, especially so in the convective region (lines "DC," Table 1), the correlation is positive with salinity and temperature: deeper MLD are associated with saltier and warmer surface waters the month prior.Of the four drivers investigated, the MLD is most strongly correlated with salinity or temperature for the vast   1).Correlations are similar when considering possible drivers the summer before (not shown, see also Text S1 in Supporting Information S1).Unsurprisingly, the possible drivers are not independent (Table S4 in Supporting Information S1).The correlations between temperature and salinity, and between salinity and either SIC or wind speed are of different signs depending on the convecting behavior, suggesting that different processes and/or water masses are involved (see Section 3.3).Correlations are similar when considering the decadal means rather than yearly values (Table S5 in Supporting Information S1), which now leads us to investigating trends.
The trend analysis reveals that: 1.The models with no deep convection in the Arctic become fresher throughout the Arctic, throughout the run (Figure 3, first two lines).Their trends are rather weak compared to the other model groups.Their winter sea ice does not retreat far into the Arctic, even by the end of the run (Figure S1 in Supporting Information S1), therefore their heat loss is small.2. The models with deep convection at the middle of the run, be it rarely (Figure 3, lines 3 and 4) or peaking and declining (lines 5 and 6) exhibit similar trends.Their ocean surface becomes saltier and winds stronger at the location where mixed layers deepen in the first half of the run; they freshen in the rest of the Arctic.In the second half, when mixed layers are shallow again, the ocean surface freshens everywhere.The main difference between these two groups is in their sea ice trends, with the models whose convection peaks and slowly declines having no winter sea ice by the end of the run (Figure S1 in Supporting Information S1).Their heat loss trend is strongest of all the groups, in the first half of the run, but then insignificant in the second half, suggesting that convection shuts down because the heat reservoir is depleted.We investigate this further in the next subsection.3. The models that convect at the end of the run (Figure 3, last two lines) have the opposite salinity trends: first a freshening, then a salinification in the region where mixed layers deepen, along with stronger winds.
Note that the trend patterns are similar in summer (Figure S2 in Supporting Information S1), indicating that the changes occur year round.The conclusion is that Arctic deep convection is associated with both a saltier ocean surface and stronger winds.Stronger winds alone are not enough (see e.g. the no deep convection group, second half of the run).

Long Term Drivers: Sea Ice, Stratification, and Atlantification
Deep convection rarely emerges as a stable feature in the Arctic.Both a sea ice loss too low or in contrast too intense, too early (and associated intense heat loss) suppress deep convection.That intense heat loss, along with the salinification of convecting models, alludes to a possible feedback mechanism with the Atlantic water layer (as suggested in the simple model of Davis et al. (2016)): Deep convection leads to the upwelling of warm and salty Atlantic Water, which induces more ice loss and more deep convection.We now investigate this possible feedback, focusing on the eastern EB where deep convection emerges (western shown on Figure S3 in Supporting Information S1).
As found by Heuzé et al. (2023), the AWCD varies widely across models (red lines on Figure 4) and is deeper than in observations for all but the E3SM models.Likewise, in agreement with Khosravi et al. (2022), we find that changes in AWCD are model-dependent, but in general the AWCD becomes shallower and passes the 500 deep convection threshold by the end of the run.Nevertheless, we find marked differences between the four groups of models: Note.Models are ordered based on their Arctic deep convection behavior, as per Figure 2.Only correlations significant at 95% are shown; for models with more than one ensemble member, median of the correlations of the dominating sign.For models with more than one ensemble member, standard deviation is the across-ensemble spread; spatial spread otherwise.Bold fonts highlight the maximum correlation for each model.Correlation between parameters is shown in Table S4 in Supporting Information S1. 1.The models that do not convect (rows 1, 2, and NorESM2-LM) have the deepest halocline (across-model median of 115 ± 94 m at the beginning of the run, blue lines on Figure 4), and their sea ice area (gold lines) hardly changes even by the end of the run, probably because the deep halocline isolates it from the subsurface oceanic heat (Shu et al., 2022).Their AWCD and its changes are inconsistent across models, but they are the models with the lowest increase in AW core temperature (AWCT, 2.2 ± 1.0°C).2. The models that convect sporadically, early in the run (ACCESS-ESM1-5 and row 3) also have a comparatively deep halocline (100 ± 67 m), but less than the no-convection models.Similarly, their increase in AWCT is modest (2.4 ± 1.3°C).Therefore, they also have a somewhat stable sea ice area, but it decreases after 2075.3. The models that convect in the first half of the run and then cease (row 5) have an area-median halocline that is very shallow (10 ± 41 m) and their AWCT warms the most of all models (6.0 ± 1.0°C).This is probably why they lose their sea ice early in the run.The average MLD (black lines) and average AWCD (red) envelopes overlap; the MLD reaches into the Atlantic Water layer, which explains these models' large heat loss trends (Figure 3) and early ice loss.The AW changes seem unaffected by deep convection, showing once again that Atlantification is caused by advection rather than local changes (e.g., Polyakov et al. (2017); Shu et al. (2022)).
That is, were it a local process, deep convection would steadily cool the AW. 4. The models that convect at the end of the run (bottom row on Figure 4) do not have the shallowest halocline, and have a significant sea ice decline starting around 2070.Their timeseries are rather similar to the those of the models that exhibit convection only sporadically, suggesting that deep convection would not persist, but we cannot tell without longer runs.
The E3SM models really stand out: they have somewhat realistic halocline and AW depths, and a strongly reduced sea ice cover by 2070, yet no deep convection.It could be because their design is very different from that of the other CMIP models, notably using Spherical Centroidal Voronoi Tessellations instead of standard grid for the ocean, sea ice and land ice (Golaz et al., 2019), but investigating these individual models further is beyond the scope of this paper.
In summary, as hypothesized from observations by Peralta-Ferriz and Woodgate (2015), it is the interplay between halocline depth, that is, stratification, and sea ice retreat (coupled with increased wind stirring) that seems to determine whether a model will exhibit deep convection.The role of Atlantification is complex, as was found by for example, Muilwijk et al. (2023), as AW temperature and depth are inconsistent across models.In particular, early convection into the AW layer does not lead to a positive feedback enhancing deep convection as expected by Polyakov et al. (2017), but rather shuts down convection rapidly.It is therefore crucial to better constrain projections of changes in the halocline depth and properties (Muilwijk et al., 2023) and its drivers (Cai et al., 2023;Khosravi et al., 2022) as well as the sea ice (Årthun et al., 2021;Keen et al., 2021).Most likely, the exact chain of events controlling changes in the mixed layer is model specific, based on their choices of parameterizations or even their definition of the mixed layer (Griffies et al., 2016), but individual model studies require access to each model's code, and are way beyond the scope of this paper.

Conclusions
We used all CMIP6 models and all their ensemble members for which the MLD and its potential drivers the surface and subsurface properties, SIC, and wind speed, were available for the strongest warming scenario SSP5-8.5.After removing the ensemble members that had spurious Arctic deep convection (defined, as in Lique et al. (2018), as MLD deeper than 500 m) over 2015-2023, we were left with 27 models, of which 11 had no deep convection in the Arctic over 2015-2100; six that had it extremely rarely (usually 4 years or fewer), by the middle of the run; five for which deep convection peaked in the first half of the run and then declined and disappeared; and five in which deep convection emerged in the Arctic in the second half of the run and still convected in 2100.Deep convection ceases in the Nordic Seas for most models, with MLD often shallower than 500 m, showing that

Geophysical Research Letters
10.1029/2023GL106499 deep convection in the Arctic is not simply a northward migration of the Nordic Seas ventilation.The Arctic MLD was most strongly correlated with the surface salinity, and the sign of this correlation depended on whether the model convected or not.Similarly, when and where the models are not convecting, their surface salinity freshens; at the location where they do, when they do, it becomes saltier and surface winds are increasing.A strong halocline and associated absence of sea ice loss resulted in models not convecting, whereas models with a weak halocline lose their ice and convect early in the run, but then cease after a period of intense heat loss to the atmosphere.The AWCD was unaffected by the deep convection behavior of the models, even in the rare cases where the models' mixed layer reached it, further showing that Atlantification is an advected signal (Shu et al., 2022) and plays a complex role in setting future stratification (Muilwijk et al., 2023).The exact mechanism triggering deep convection is most likely model-specific though, and its determination would requires in-depth sensitivity studies for each model.Such in-depth investigation could also lead to model improvement.CMIP6 models consistently exaggerate deep convection both in the North Atlantic and in the Southern Ocean (Heuzé, 2021).Understanding why they have no such consensus in the Arctic could hold the key to a more realistic representation of mixing, globally.

Figure 1 .
Figure 1.Maximum mixed layer depth over January 2015-December 2100 (shading, logarithmic scale) for SSP5-8.5 at each grid cell.For each model, parentheses indicate the number of ensemble members available (see TablesS1 and S2in Supporting Information S1); when this number is larger than one, the figure shows the ensemble median.Black contours are the 1,000 m isobath.Red squares indicate the models for which all ensemble members have deep convection in the Arctic already over 2015-2023, and which are therefore not considered for further analysis.Locations discussed in the manuscript are indicated on the top-right panel: Cyan contour "EB" is the Eurasian Basin, further subdivided into western "W" and eastern "E" at 60°E; indigo dashed contour "GIN" is the Nordic Seas; and magenta arrow "SAT" is the St Anna Trough.

Figure 2 .
Figure 2.For the 27 models that do not have deep convection in the Arctic over 2015-2023, time series of their yearly maximum mixed layer depth (MLD) in the Nordic Seas (red) and in the Arctic Ocean (blue).For each model, parentheses indicate the number of ensemble members remaining; when this number is larger than one, the figure shows the range across these ensemble members (shading) and the ensemble median (thick line).Horizontal black lines indicate the 500 and 1,000 m MLD thresholds, indicative of deep convection.Vertical black line is the year 2023.Models are ordered based on their Arctic behavior: First two rows and NorESM2-LM, no Arctic deep convection; ACCESS-ESM1-5 and fourth row, rare convection, by the middle of the run; fifth row, convection peaks by the middle of the run and then declines; bottom row, convection starts late in the run.

Figure 3 .
Figure 3. Composite trends based on the models' Arctic deep convection behaviors of Figure 2, for the first half of the 21st century run (top rows) and the second half (bottom rows), in March mixed layer depth (first column), and February ocean surface salinity (S, second), heat flux into the ocean (third), sea ice concentration (fourth) and surface wind speed (last).Behaviors are described to the left of the figure, and number of models for each behavior is given in parentheses.For each panel, stippling indicates non-significant trends and/or model disagreement regarding the trend's sign.Straight stippled lines across the North Pole are an artifact of the necessary interpolation.

Figure 4 .
Figure 4.For the 27 models that do not have deep convection in the Arctic over 2015-2023, time series of their yearly median mixed layer depth (black), sea ice area (gold), halocline depth (blue) and Atlantic Water core depth (red) over the Eastern Eurasian basin (see Figure1).For GISS-E2-1-G, gold line is the normalized median sea ice thickness (see text).When more than one ensemble member remains, the figure shows the range across these ensemble members (shading) and the ensemble median (thick line).Models are ordered based on their Arctic behavior: First two rows and NorESM2-LM, no Arctic deep convection; ACCESS-ESM1-5 and fourth row, rare convection, by the middle of the run; fifth row, convection peaks by the middle of the run and then declines; bottom row, convection starts late in the run.

Potential Triggers: Salinification and Stronger Winds Lique
Lique et al. (2018), deep convection does not migrate to the Arctic in response to its cessation further south.In the next sections, we investigate what else could be causing the different Arctic deep convection behaviors.et al. (2018) found that deep convection emerges in response to a local increase in surface salinity caused by enhance surface circulation; Bretones et al. (2022) and Pérez-Hernández et al. (2019), a reduction in sea ice.In contrast, Peralta-Ferriz and Woodgate

Table 1
Correlation Coefficient and Its Standard Deviation For Each Model Between the March Mixed Layer Depth (MLD) and the Previous Month (February) Ocean Surface Salinity (S), Ocean Surface Temperature (T), Sea Ice Concentration and Surface Wind Speed (Wind), For the Eurasian Basin (EB) and, For the Models With Arctic Deep Convection, Where the Maximum MLD of Figure 1 Exceeds 500 m (DC)