Deep water formation in the North Atlantic Ocean in high resolution global coupled climate models

Abstract. Simulations from seven global coupled climate models performed at high and standard resolution as part of the High Resolution Model Intercomparison Project (HighResMIP) have been analyzed to study the impact of horizontal resolution in both ocean and atmosphere on deep ocean convection in the North Atlantic and to evaluate the robustness of the signal across models. The representation of convection varies strongly among models. Compared to observations from ARGO-floats, most models substantially overestimate deep water formation in the Labrador Sea. In the Greenland Sea, some models overestimate convection while others show too weak convection. In most models, higher ocean resolution leads to increased deep convection in the Labrador Sea and reduced convection in the Greenland Sea. Increasing the atmospheric resolution has only little effect on the deep convection, except in two models, which share the same atmospheric component and show reduced convection. Simulated convection in the Labrador Sea is largely governed by the release of heat from the ocean to the atmosphere. Higher resolution models show stronger surface heat fluxes than the standard resolution models in the convection areas, which promotes the stronger convection in the Labrador Sea. In the Greenland Sea, the connection between high resolution and ocean heat release to the atmosphere is less robust and there is more variation across models in the relation between surface heat fluxes and convection. Simulated freshwater fluxes have less impact than surface heat fluxes on convection in both the Greenland and Labrador Sea and this result is insensitive to model resolution. is not robust across models. The mean strength of the Labrador Sea convection is important for the mean Atlantic Meridional Overturning Circulation (AMOC) and in around half of the models the variability of Labrador Sea convection is a significant contributor to the variability of the AMOC.



Abstract.
Simulations from seven global coupled climate models performed at high and standard resolution as part of the High Resolution Model Intercomparison Project (HighResMIP) have been analyzed to study the impact of horizontal resolution in both ocean and atmosphere on deep ocean convection in the North Atlantic and to evaluate the robustness of the signal across models. The representation of convection varies strongly among models. Compared to observations from 40 ARGO-floats, most models substantially overestimate deep water formation in the Labrador Sea. In the Greenland Sea, some models overestimate convection while others show too weak convection.
In most models, higher ocean resolution leads to increased deep convection in the Labrador Sea and reduced convection in the Greenland Sea. Increasing the atmospheric resolution has only little effect on the deep convection, except in two models, which share the same atmospheric component and show reduced convection. Simulated convection in the Labrador Sea is 45 largely governed by the release of heat from the ocean to the atmosphere. Higher resolution models show stronger surface heat fluxes than the standard resolution models in the convection areas, which promotes the stronger convection in the Labrador Sea. In the Greenland Sea, the connection between high resolution and ocean heat release to the atmosphere is less robust and there is more variation across models in the relation between surface heat fluxes and convection. Simulated freshwater fluxes have less impact than surface heat fluxes on convection in both the Greenland and Labrador Sea and this 50 result is insensitive to model resolution. is not robust across models. The mean strength of the Labrador Sea convection is important for the mean Atlantic Meridional Overturning Circulation (AMOC) and in around half of the models the variability of Labrador Sea convection is a significant contributor to the variability of the AMOC. widespread convective activity down to 2200m (Yashayaev and Loder, 2017). Irminger Sea convection was also strong in recent years and reached record levels in the winter 2014/ 2015 (de Jong and de Steur, 2016;de Jong et al., 2018). The transport of freshwater from the Fram Strait along the Greenland coast into the Labrador Sea is another contributing factor, with increased freshwater fluxes leading to reduced salinities that suppress deep convection in the Labrador Sea (Holland et 105 al., 2001;Jungclaus et al., 2005;Koenigk et al., 2006).
Deep convection in the Nordic Seas may also play an important role (Lozier et al., 2019). Langehaug et al. (2012) linked the variability of the AMOC to the variability of the overflows across the Greenland-Scotland Ridge.
Because of the coupled nature of the involved processes, coupled climate models are in principle well suited to study the interactions between deep convection and other climate-related processes. However, Heuze (2017) stated that "the majority 110 of CMIP5 models convect too deeply, over too large an area, too often and too far south". Further, Heuze (2017) found that deep convection is best simulated in those models with realistic ice edges in the North Atlantic.
In this study, we analyze the impact of increasing the horizontal resolution on the deep convection in the North Atlantic. We use simulations from seven models participating in the High-Resolution Model Intercomparison Project (HighResMIP, Haarsma et al., 2016), which have been performed in the EU-H2020-project PRIMAVERA. High resolution has been shown 115 to improve many aspects of the ocean circulation. Gutjahr et al. (2019) showed a reduction of temperature and salinity biases in the MPI-ESM1-2 model with eddy resolving ocean resolution. Grist et al. (2018) showed a more realistic northward ocean heat transport in high-resolution models that results consequently in a more realistic representation of the sea ice in the Atlantic sector of the Arctic Ocean (Docquier et al., 2019). High resolution ocean models also substantially improve the position of the North Atlantic Current (Chassignet and Marshall, 2008;Sein et al., 2018). Furthermore, higher horizontal 120 resolution might lead to a more realistic simulation of freshwater exports out of the Arctic (Fuentes Franco and Koenigk, 2019) and a better representation of the properties and position of the dense overflows.
After this introduction, we proceed with describing the models, the data and the methods in section 2. Sections 3 to 5 will show the results from this study and we will conclude in section 6.
2 Models, data and method 125

Models and simulations
In this study we analyze seven global coupled climate models (see e.g. Vannière et al., 2019), which participated in the HighResMIP experiment within the H2020-EU-project PRIMAVERA. These models are ECMWF-IFS (Roberts et al., 2018), HadGEM3-GC31 , MPI-ESM1.2 (Gutjahr et al,. 2019), CMCC-CM2 (Cherchi et al., 2019), CNRM-CM6.1 (Voldoire et al., 2019), AWI-CM-1.0 (Sidorenko et al., 2014, HR andLR setups: Sein et al., 2016) and  Earth3P (Haarsma et al., 2020). We use the historical coupled simulations from 1950-2014 and the 100-year control simulations (using constant 1950-forcing) from these seven models for our analysis. All models performed the simulations in at least two different resolutions following the HighResMIP-protocol. Changes in oceanic and atmospheric parameters are https://doi.org/10.5194/os-2020-41 Preprint. Discussion started: 13 May 2020 c Author(s) 2020. CC BY 4.0 License. kept to a minimum between low and high resolution simulations, so that all changes can be directly attributed to the change in resolution. 135 The resolution varies among models. A few of the models vary both ocean and atmosphere resolution at the same time while others separately changed ocean or atmosphere resolution. This allows us to analyze also the effect of increasing the resolution in only one component of the system. Five of the seven models use the NEMO-ocean model as ocean component.
While this might limit the robustness of our conclusions across models, it has to be noted that the NEMO-model configurations differ quite substantially from each other with different sea ice models (LIM2, LIM3, GELATO, CICE) and 140 differences in parameters (e.g. Gent McWilliams versus Smagorinsky). AWI-CM-1-0 and MPI-ESM1.2 use the same atmosphere component but different ocean components.
More details on the models and the simulations used in this study are provided in table 1.

Observational data
To compare the mixed layer depth (MLD) from the models to observations, the typical variable used to represent ocean 145 convection depth, we use data from ARGO-profiles (Holte et al. 2017) provided on a 1° grid. In the ARGO-data, the de We use turbulent latent and sensible heat fluxes from the global ocean-surface heat flux products  developed by the Objectively Analyzed air-sea Heat Fluxes project at the Woods Hole Oceanographic Institution (WHOI-OAFlux) to 155 evaluate the surface heat fluxes in our models. We use monthly means of the WHOI-OAFlux data on a 1° grid from 1958 onwards.

Method
Several different indices have been defined for the deep convection in the ocean (e.g. Schott et al., 2009;Yashayaev and Loder, 2009;Lavergne et al., 2014;Koenigk et al., 2007;L'Heveder et al., 2012). These indices take into account either the 160 deepest reaching convection and/ or the horizontal extent of the MLD. However, none of them excludes convective events that are too shallow to contribute to deep water formation. To overcome this problem, Brodeau and Koenigk (2016) defined the so-called "Deep Mixed Volume" (DMV), which only considers the convective mixing below a specific depth (critical depth z crit ) and integrates the volume of these deep mixed water masses in different convection regions of the North Atlantic.
In our study, we use the DMV index for monitoring the deep convection. Following Brodeau and Koenigk (2016), we use a critical depth of 1000m for the Labrador Sea and 700m for the Greenland Sea. In the Labrador Sea, convection needs to reach a depth of around 1000m to be able to sustain the renewal of Labrador Sea water and eventually become North Atlantic Deep Water (Yashayaev, 2007). In the Nordic Seas, convection needs to at least reach down to the depth of the Denmark Strait and Faroe Bank Channel, which is around 600-700m in the models. We define the Labrador Sea region as and Greenland Seas convection in all models fall into these regions. Although intermittent deep convection can occur in the Irminger Sea as well, we focus here only on the two regions with deepest convection.
We use monthly mean values of the March MLD of the model simulations to calculate the DMV. Note, that short convection episodes that exceed z crit might thus be missed.
We also calculate the DMV from the ARGO data as comparison to the model results. We infilled grid-points with missing 175 data in the ARGO-data by interpolating the nearest neighbours. This and the short time series of the ARGO-data lead to uncertainties in the calculations of the DMV from ARGO and therefore it only provides a rough estimate for the real world.
As an additional comparison, we also calculate the DMV based on critical depths of 0m, thus considering the full mixed layer.
For correlations, we calculate the Pearson correlation coefficient (r). We call a correlation significantly different from 0, if 180 the p-value of the Pearson correlation coefficient is 0.05 or smaller based on a two-sided student-t distribution. Assuming 98 (N-2) degrees of freedom (assuming independence of each year of data in the 100-year (N=100) 1950-control simulations), the correlation is significant if |r| exceeds 0.2. When taking the autocorrelation of the variables into account, the degrees of freedom are reduced and differ depending on model and variable.

Deep Convection in the North Atlantic 185
This section analyzes first the MLD in March, the month with the strongest convection in both observations and models, in the North Atlantic in the different models and in ARGO. Then, we focus on the DMV in the models, its variability and potential trends in the historical simulations. Although, this is a considerable amount, given the relatively long averaging period, the MLD differences due to increased resolution from 1° to a 0.25° in the NEMO-models are larger. Figure 2f shows the DMV in the Labrador Sea for 220 the ensemble mean and the four ECMWF-IFS-HR members. There is substantial spread across members but no generally different behavior in amplitude and time-scales of variability and trends across model members can be seen.

Mixed layer depth
Even though four members are not sufficient to capture the total natural variability, these results suggest that natural variability cannot explain the differences in MLD due to a change in spatial resolution.

Deep Mixed Volume 225
In the following, in order to consider the horizontal extension of convection patterns and discard shallow convection events that have limited impact on the oceanic circulation such as the AMOC, we will concentrate on the DMV index to investigate the deep convection in the Labrador and Greenland Seas in more detail.
https://doi.org/10.5194/os-2020-41 Preprint. Discussion started: 13 May 2020 c Author(s) 2020. CC BY 4.0 License. Figure 3 shows the DMV in the Labrador Sea in March in the historical model simulations. In agreement with Figure 1, 230 increasing the resolution from around 1° to 0.25° in the ocean generally leads to an increased DMV in all models using NEMO, while the opposite is true for AWI-CM-1-0. A further increase in ocean resolution to 1/12 ° in HadGEM3-GC31-HH does not further increase the DMV. The DMV varies strongly among models: ECMWF-IFS-LR does not show any deep convection events in the entire historical period, CNRM-CM6.1 and EC-Earth3P simulate only a few events with deep convection and AWI-CM-1-0-LR and CMCC-CM2 simulate strong deep convection every winter. 235 Table 2  Sea. If deep convection occurs, the ocean is often mixed down to the bottom, while in-situ observations indicate that deep convection rarely exceeds 2000m (Yashayaev and Loder, 2016;Yashayaev and Loder, 2017).

Labrador Sea
If we use a critical depth of z crit =0 m instead of 1000 m in the Labrador Sea and thus consider the total mixed layer depth, the relative deviation of the DMV in the models from ARGO is reduced as expected (not shown). However, AWI-CM-1-0-LR and CMCC-CM2 still overestimate the DMV based on ARGO by a factor of three and two, respectively. On the other hand, 245 ECMWF-IFS-LR simulates only 20% of the mixed volume compared to ARGO. The comparison between z crit0 and z crit1000 reveals also some non-linearites in the deep convection. While CNRM-CM6.1-HR has a nine times higher DMV (z crit1000 ) compared to ARGO, it is only 16% higher for z crit0 , whereas the DMV (z crit1000 ) for MPI-ESM1-2-XR is 4.6 times higher compared to ARGO but 14% smaller for z crit0 . The strength of the deep convection in March is reflected in the vertical density distribution in the Labrador Sea. Naturally, the models with more frequent and deeper convection show a much weaker vertical stratification than the models that do not exhibit deep convection. More interesting as the density distribution during the convection period itself is the vertical stratification at the beginning of the winter, which indicates the preconditioning of the ocean for convection events later in 260 winter. Figure 4 shows the vertical density stratification of the upper 600m in the Labrador Sea. All models show a near https://doi.org/10.5194/os-2020-41 Preprint. Discussion started: 13 May 2020 c Author(s) 2020. CC BY 4.0 License.
surface low density layer, mainly due to a combination of low surface salinity and relatively (compared to late winter) warm water near the surface in November. Generally, the models with lower ocean resolution show a stronger stratification in the upper ocean than models with higher resolution (except for AWI-CM-1-0). The two model simulations, which do not simulate any deep convection, ECMWF-IFS-LR and EC-Earth3P, show particularly strong upper ocean density gradients. 265 Consequently, a large buoyancy flux would be needed during winter until deep convection could set in in these two models.
MPI-ESM1-2 and AWI-CM-1-0 show a more stratified upper ocean in November with increased atmospheric resolution.
This agrees with a weaker convection in their higher resolution versions. The density profiles of the high ocean resolution models agree relatively well with the observed one from ARGO although the near surface low density layer is too shallow in most of these models. This might contribute to the overestimation of the deep convection in late winter in these models 270 (compare Figures 1 and 3) but is probably not the only reason as will be further discussed in section 4.
Twelve of 19 simulations indicate a significantly negative trend of the DMV in the historical period ( Figure 3, Table 3). To investigate whether this trend is really due to external forcing and not to model drift due to the rather short spinup period, we compared the DMV in the historical simulations with that from the 100-year 1950-control simulations ( Figure 5 and Table  275 3). Most of the control simulations do not show any large trends and in 9 out of 17 historical simulations, the DMV trends in the historical simulations are significantly more negative compared to the first 65 years of the control simulations indicating that external forcing is the major cause for the DMV reduction. As found with the mean DMV, the negative trends are larger with higher ocean resolution in the historic simulations.
A reduction of DMV in the historical period would be in line with some recent studies by Caesar et al. (2018), Thornalley et 280 al. (2018) and Brodeau and Koenigk (2016).
We calculated the power spectrum of the DMV in the Labrador Sea in order to investigate the predominant variability of the DMV in each simulation in more detail ( Figure 6). For better comparison, we detrended and normalized (using the standard

Greenland Sea
The DMV in the Greenland Sea shows also a large spread across models (Figure 7, Table 2 Table. 3). Similarly, the strong negative trends in CMCC-CM2 can partly be explained by similar drifts in the control simulations, although the reduction in the historical runs is significantly larger than in the control runs.  Increasing the atmosphere resolution has a minor impact on the DMV in the Labrador Sea, except for MPI-ESM1.2 and AWI-CM-1-0, where DMV is reduced with increased resolution.
The resolution dependency of the DMV in the Greenland Sea in single models is smaller than in the Labrador Sea. However, 335 all the models, except for CNRM-CM6-1, show a decreased DMV when increasing the resolution to around 0.25°. The response to increased atmosphere resolution is not robust across models.

The impact of heat and freshwater fluxes on the deep convection in the North Atlantic
Deep convection depends strongly on the buoyancy of the ocean surface layer in the convection regions -the heat loss to the atmosphere and the influx of fresh water into the convection regions. 340 Brodeau and Koenigk (2016) showed that the turbulent surface heat flux (SHF) is the main driver for interannual variability in the DMV. Thus, in the following, we will mainly focus on the SHF. As in the Labrador Sea, northerly winds are the main cause for large oceanic surface heat loss to the atmosphere in the Greenland Sea. The northerly winds are connected to low pressure anomalies over northern Scandinavia and the Barents Sea. 375

Surface heat fluxes
The DMV in the Greenland Sea is correlated to the SHF as well. However, here we find a stronger model dependency of this relation. The correlation is weak to moderate in HadGEM3-GC31 (r=0.22 for LL; r=0.5 for HH) and in CNRM-CM6.1 (r=0.35; r=0.5 for HH) but high correlation is found for ECWMF-IFS (r=0.64 for HR; r=0.85 in LR) and EC-Earth3P (r=0.61; r=0.69 for HR). As for the Labrador Sea, the relation between SHF and DMV shows no clear resolutiondependency. 380 Labrador Sea, which tends to reduce the convection. Figure 12 shows for the two model simulations with the highest correlation between freshwater transport through Denmark Strait and DMV in the Labrador Sea (HadGEM-GC31-LL, EC-395

Freshwater and sea ice exports
Earth3P-HR) that increased freshwater transport leads to a substantial reduction of the MLD in the Labrador region and thus contributes to the variability of the DMV. For most other model simulations, the effect is rather small compared to the impact of SHF-variability on the DMV.
In some models, the southward transport of liquid freshwater through Baffin Bay is positively correlated with the deep convection in the Labrador Sea (up to r=0.35 in HadGEM3-GC31-LL). This may seem counterintuitive, but northerly winds 400 in the Baffin Bay cause strong SHF in the Labrador Sea and dominate the convective conditions and simultaneously lead to increased fresh water transports to the south.
We do not find any resolution dependency of the correlation between freshwater exports and convection in the Labrador Sea.
This result is in contrast to a recent study from Fuentes Franco and Koenigk (2019) where they analyzed a set of HadGEM3-GC2 simulations at different resolutions and found larger correlations with increased resolution. 405 Overall, there is only a weak relationship between freshwater export through the Fram Strait and convection in the Greenland Sea, although it shows some dependency on the respective model (not shown). In some of the simulations, more freshwater export out of the Arctic is associated with reduced deep convection in the Greenland Sea, but in the majority of the simulations larger exports occur at the same time as increased convection. In the latter case, the increased convection is driven by northerly winds, which at the same time increase the freshwater exports through Fram Strait. 410

The linkage of the DMV to the AMOC
The effect of high resolution on the AMOC in the HighResMIP model simulations has been studied in more detail in a parallel study to ours (Roberts et al. submitted). They found that "the AMOC tends to become stronger as model resolution is enhanced, particularly when the ocean resolution is increased from non-eddying to eddy-present and eddy-rich". Roberts et al., (submitted) also analysed the relation between temporal mean values of the DMV and the average AMOC. As shown in 415 section 3.2, only few models simulate a DMV that is consistent to observed estimates. However, these models underestimate https://doi.org/10.5194/os-2020-41 Preprint. Discussion started: 13 May 2020 c Author(s) 2020. CC BY 4.0 License.
the AMOC (except for CNRM-CM6-1) compared to the RAPID-observations whereas some of the models (HadGEM3-GC31-MM and -HM, MPI-ESM1.2-HR, AWI-CM-1-0-LR) markedly overestimate the DMV in the Labrador Sea but simulate a realistic AMOC. Thus, the observations show a stronger AMOC with a lower DMV compared to the models, indicating other shortcomings in the representation of processes that govern the AMOC in the models. 420 There is in general a strong relationship between DMV in the Labrador Sea and the AMOC strength across models; models with more deep water production in the Labrador Sea have a stronger AMOC. Also for all single models (apart from AWI-CM-1-0), simulations with larger DMV are linked to a stronger AMOC. This relationship is less robust between DMV in the Greenland Sea and the AMOC as expected from the reduced DMV with increased resolution in most of the models (see sections 3.2.2). 425 To investigate the impact of variability in the deep water formation on the variability of the AMOC, we performed crosscorrelation analyses between the DMV in Labrador and Greenland Seas and the AMOC (at 26°N) for lags between -/+ 10 years. In agreement with results by Brodeau and Koenigk (2016)

Conclusions
We analyzed historical and 1950-control simulations in different resolution from seven global climate models following the HighResMIP protocol and investigated the impact of increasing the resolutions in ocean and atmosphere on deep convection in the North Atlantic Ocean. 445 The main results are summarized as follows: https://doi.org/10.5194/os-2020-41 Preprint. Discussion started: 13 May 2020 c Author(s) 2020. CC BY 4.0 License. realistic simulation of deep convection is important for the large scale ocean circulation, in particular the AMOC, the northward heat transport in the ocean and related impacts on the atmosphere. It also raises serious questions of the future behaviour of the AMOC in climate models and its consequences for local and global climate. 450 -The ocean resolution clearly affects the deep water formation in the Labrador Sea. Convection activity enhances with increasing ocean resolution in four out of five models in this study. However, all these models use NEMO3.6 (although in somewhat different configurations) as their ocean component. It remains therefore unclear whether global models with other ocean models respond differently to an increased resolution since the reduced convection in the fifth model (AWI-CM-1-0) 455 results very likely from the simultaneously increased atmosphere resolution.
-Increasing the ocean resolution from 1° to 1/4° in the models with NEMO as the ocean component has a larger impact on the convection than increasing the atmosphere resolution in these models. In contrast, MPI-ESM1-2, in which only the atmosphere resolution has been increased, and AWI-CM-1-0 (increased resolution in both atmosphere and ocean) show 460 substantially reduced convection in the Labrador Sea at high resolution. Both models (AWI-CM-1-0, MPI-ESM1-2) use the same atmospheric component (ECHAM6.3) and the reduction of DMV with increased atmospheric resolution can likely be linked to reduced atmospheric winds in ECHAM6.3 in the high resolution version (Gutjahr et al., 2019;Putrasahan et al., 2019). The models with higher ocean resolution show more dominate variability at the decadal time scale in the Labrador Sea compared to their lower resolution counterparts. 465 -In the Greenland Sea, increasing the ocean model resolution to around 1/4° reduces the convection in most models.
Increasing the atmosphere resolution tends to reduce the convection but the result is not robust across models. Many models show dominant variability between 10 and 25 years but no clear dependence on the resolution could be found.

470
-The turbulent surface heat fluxes are strongly related to the deep convection in both the Labrador and Greenland Seas and seem to be more important for the variability of the DMV than freshwater exports out of the Arctic. In the Labrador Sea, we find that higher resolution leads to increased ocean heat release to the atmosphere in all the NEMO models but to reduced heat release in MPI-ESM1-2. This is in close agreement with the resolution dependency of the deep convection. Thus, increased turbulent surface heat flux with high resolution is the main explanation for increased DMV in the Labrador Sea. 475 The correlation between surface heat fluxes and DMV in the Labrador and Greenland Seas does not show any robust resolution-dependency.
-The 10-year low pass filtered DMV in the Labrador Sea is highly positively correlated (r=0.6-0.8) with the AMOC at 26 °N in around half of the model simulations. In these simulations, the DMV leads the AMOC by a few years. In the other simulations, the correlations are also positive but much lower (0.3-0.4) and time lags of the highest correlations are not robust across these simulations. The DMV in the Greenland Sea and the AMOC are only significantly correlated in few simulations and no clear lead/ lag relationship can be established. The correlations between DMV and AMOC are not dependent on the resolution.

485
-Increasing the resolution improves the vertical stratification of the upper ocean in late autumn but it does not generally improve the representation of the deep convection. In a few of the low-resolution models, the convection is overestimated compared to ARGO and this positive bias becomes even larger with higher resolution. However, the high resolution models have not been tuned and the main purpose of HighResMIP is to investigate the impact of increasing the resolution rather than to improve existing biases. 490

Model
Trend /