Two superimposed cold and fresh anomalies enhanced Irminger Sea deep convection in 2016 – 2018

Sea deep convection in 2016 – 2018 Patricia ZUNINO , Herlé MERCIER , Virginie THIERRY [3] [1] Altran Technologies, Technopôle Brest Iroise, Site du Vernis , 300 rue Pierre Rivoalon, 29200 Brest, France, [2] CNRS, University of Brest, IRD, Ifremer, Laboratoire d'Océanographie Physique et Spatiale (LOPS), IUEM, ZI de la pointe du diable, CS 10070 29280 Plouzané, France [3] Ifremer, University of Brest, CNRS, IRD, Laboratoire d'Océanographie Physique et Spatiale (LOPS), IUEM, ZI de la pointe du diable, CS 10070 29280 Plouzané, France

and 7, because the northern limit of the pink box used for computing the preconditioning SECF has been modified. -) Fig. 8, because the previous point 59°N -40°W is not centered in the SECF box, we now present the results at 58°N -40°W, instead. Moreover, we decided to present all the anomalies in place of only anomalies larger than one-standard deviation.
-) Fig. 10 in the previous manuscript was removed.
There is code readily available to derive MLD from Argo profiles using all variables (Holte and Talley, 2009; http://mixedlayer.ucsd.edu/). I suggest the authors use this, or some adjustment of their own code, to rederive the MLD for all profiles and adjust the results of the paper accordingly.
My final main comment is that the title could be rephrased to represent the content/conclusions better. The fresh anomaly that seems to be referred to a deep one, the lowering of the halocline. The surface freshwater anomaly, which is discussed in detail elsewhere but is only touched upon here, is was not enhancing convection. It is only the cold surface anomaly that worked to enhance somewhat, but even that is only touched upon. Still, those who have not yet read the abstract may think this paper is about the big surface Sanom currently going around. While in fact, the paper focuses in detail on favorable preconditioning which is not mentioned in the title. So, it is not clear why this title was chosen.
Thank you very much for your constructive comments. In the following we answer point by point to your comments and indicate how the manuscript is going to be revised.
Following your suggestion, we revised the manuscript to define the MLDs based on density, temperature and salinity criteria (and not density criteria only). We adapted our method to include temperature and salinity criteria in addition to density criteria and we compared our results to two alternative methods of determination of the MLD previously used by de Jong et al. (2012) and Pickart et al. (2002). In our revised method, we determined the MLD as the shallowest of the three MLD estimates obtained separately from temperature, salinity and density profiles using the threshold method (de Boyer Montégut et al., 2004). The threshold criteria were the differences in property between the surface (30 m) and the MLD set to 0.01 kg m -3 in density (Piron et al. 2017), 0.1°C in temperature and 0.012 in salinity. The temperature threshold of 0.1°C and the salinity threshold of 0.012 were selected because they correspond to a threshold of 0.01 kg m -3 in density that was previously shown to perform well in the subpolar gyre (Piron et al., 2016). Indeed, MLD based on this density threshold favorably compared to those estimated by the method of Thomson and Fine (2003) as demonstrated in Piron et al. (2016;2017) and visual inspection. We used de Jong's methodology as follows. First we interpolated the Argo data into 10 m depth steps. Then, we estimated the standard deviations of density, temperature and salinity from the surface to each depth level. Following de Jong et al. method's, three MLD were defined as the depths were the standard deviations were smaller than 0.05 kg m -3 , 0.05°C and 0.005 for density, temperature and salinity, respectively. The final MLD was the shallowest of the three estimates. The Pickart's methodology was applied as follows. We used the estimates of our threshold method as a first guess for the MLD. Then, the mean and standard deviation of the density, temperature and salinity were estimated from the surface to the initially defined MLD. Finally, we plotted the twostandard deviation envelope overlaid on the original profile. The mixed layer depth was determined as the location where the profile permanently crossed outside of the two-standard deviation envelope. The MLDs resulting from our method are shallower than the MLD resulting from the method of de Jong et al. (see examples in figures R1 -R3). Moreover, sometimes, the MLD defined by de Jong's method in terms of temperature or salinity is not placed at the base of the mixed layer (as visually defined), e.g. profiles 6900446 -213 (Fig. R1) or 5904772 -33 (Fig. R3). Otherwise, the MLDs estimated by our method are coherent with the MLDs resulting from the method of Pickart et al. (2002): see the envelopes (discontinuous vertical lines in figures R1 -R3) of mean ± two -times the standard deviation of density, salinity and potential temperature, from the surface to the MLD estimated with our method. Finally, we also compared our results with the MLDs determined using Holte & Talley (2009)'s method and available in the web. However, MLDs were not available for all our floats, e.g. float 6900446, or the method provides too shallow MLD, e.g. profile 6901171 -101 (89 m, see Fig. R2). Figure R1. Vertical profiles of potential density, salinity and potential temperature of profile 6900446 -213. The black points are the MLD estimated by our threshold method. The blue points indicate the MLDs resulting from the method of de Jong et al. (2012): in the density plot the MLD derived from density profile, in the salinity plot the MLD derived from salinity profile and in the temperature plot the MLD derived from temperature profile; the final MLD is the shallowest of the three defined MLDs. Following Pickart et al. (2002), the envelopes of mean ± two -times the standard deviation of the density, salinity and potential temperature from the surface to the MLD estimated using as a first guess for the MLD our threshold method were estimated and represented as discontinuous vertical lines. We have also recalculated all the properties showed in the table 1 of the previous version of the paper. Note that these properties are now estimated considering only the profiles inside the SECF box (pink box in Fig. R4.) The new results (table R1 in this document) are in line with the results of the submitted paper. 3.263± 0.031 2 *W2018 line corresponds to the properties of the mixed layer in W2018 in SEFC when the data of Float 5903102 were considered in the analysis. Finally, following the suggestion of referee 3, we decide to exclude the data of float 5903102 of our analysis because their MLDs matched with the maximal depth dived by the float. We want also to clarify that in the previous version of the paper and in the new results, the deepest MLD observed in the SECF in winter 2017 was recorded by profile 6900446 -213 and not by profile 4901809 -35 (bright blue profile in the Fig. 2 of the previous version of the paper) as indicated by the referee. Note that for 6900446 -213, the new MLD is the same than in the previous version of the manuscript.
Concluding, when recalculating the MLDs as suggested by the referees, the maximal MLD observed in the SECF was deeper than 1300 m in winters 2016, 2017 and 2018 (see fig. R4 and table R1). It indicates that deep convection occurred during the studied winters. This is the first important result of our paper, which does not change when recalculating the MLD.
Concerning the title, in order to avoid preconceived ideas to the reader, in the revised manuscript we change it to: "Why did deep convection persist over four consecutive winters (2015-2018) Southeast of Cape Farewell?" Below are some more minor comments Introduction Line 94. "In the Labrador Sea, deep convection occurs almost every year, yet with different intensity. In the Irminger Sea…". In the Irminger Sea some convection (_400 m) always occurs as well, and the intensity varies not unlike the Labrador Sea. Please rephrase or add a definition of "deep". We agree. Following Piron et al. (2015), we focus on convection deeper than 700 m, which is the minimum MLD for LSW renewal. We clarified the sentence that now reads : "In the Irminger Sea, Argo and mooring data showed that deep convection deeper than 700 m happened in the Irminger Sea during winters 2008, 2009, 2012, 2015 and 2016 (…)." Data Why is the TEOS-10 toolbox used, but profiles of theta and practical salinity are still shown instead of CT and SA? TEOS-10 allows the computation of theta and practical salinity.
Please explain briefly why 35 is chosen as a reference. This sentence is going to be deleted because we do not use FW in the paper. Sorry for the confusion it may have caused.
The ERA Interim reanalysis is replaced by ERA5. Best to do a check whether the results are robust to the choice of reanalysis. It could be interesting to check the results obtained using the new ERA5 dataset. However, the first author of this paper, who processed the data, is now working in a private company and she has not the time of redoing calculations with this new database.

Method
De Boyer and Montégut criterion is not suitable for these profiles as discussed above.
See above our answer to the major comment.
The definition of the Irminger Sea, with 48_W as the limit is rather unusual. The area in Figure 1 southeast of Cape Farewell is not typically referred to as the Irminger Sea as it fall outside of the central Irminger Gyre and profiles here are very likely to have been recently advected from the Labrador Sea. To be more consistent with previous literature it would be better to split this region in three areas: the Labrador Sea, the Irminger Sea and in between the area south of Cape Farewell. We agree that 48° W is not the limit between Labrador and Irminger Sea. When splitting the region in three areas as in Piron et al. (2017) we did not observe deep convection in the northernmost Irminger Sea (note that with the previous method of MLD computation we had a few deep MLD in the northernmost Irminger Sea in winter 2016 (those MLD corresponding to profiles not homogenous in temperature and salinity were not diagnosed with the new MLD method). In the new version of the paper we define a new pink box that we refer to as Southeast Cape Farewell (SECF) region (see Figure R4). The only change in the pink box is its northern limit: 61°N/59.3°N in the previous/revised version of the manuscript. The new box encloses all the profiles showing deep MLD during winter 2016, 2017 and 2018 Southeast of Cape Farewell. Note that the pink box is also used to estimate the atmospheric forcing and the preconditioning of the region. We recalculated it: the new results are very similar to the results shown in the previous version of the paper and do not change the conclusions of the paper.
Equation 1 and others. There are periods (.) instead of multiplication symbols. Thank you for noting it. We change all of them.

Results
What is Q3? "Q 3 is the MLD value that is exceeded by 25% of the profiles showing MLD deeper than 700 m and is equivalent to the aggregate maximum depth of convection defined by Yashayaev and Loder (2016).", as it was indicated in lines 152 -153 of the submitted manuscript.
Part of the results paragraph will have to be rewritten when MLD are rederived. Right, we are going to rewrite this section with the new results.
Line 268: Mean over which period? 1993 -2016, as indicated in the figure caption of Figure 4. We add 1993 -2016 in the text.
Line 296: This is true only when the upper 600 m already has a density close to that of the layer below (which for example could not be the case when a lot of freshwater is added). Otherwise additional buoyancy fluxes will still be required. We are describing the buoyancy profiles from the mean (2008 -2014) and we see that the thermal component of the buoyancy dominates the total buoyancy. We agree that if a large amount of freshwater is added to the upper ocean, we would find an important contribution of the haline component of the buoyancy, but it is not what we see in the mean (2008 -2014) buoyancy profiles. We added Fig. 6 at the end of this sentence to make clear that we are describing the results of this figure and that the statement is not a general statement. This comment has also motivated us to reduce the SECF or pink box. The new estimates of atmospheric forcing correspond to a reduced region closer to the position where deep convection took place.
The method used to predict the MLD does not take advection into account. This is counterintuitive because we see advection play a big role throughout winter in the field. The fact that the reanalysis do not quite match with the actual fluxes observed at OOI (Josey et al, 2018) may also be needed to take into account here. It will be interesting to see how much of a match between prediction and observation remains once new MLD are derived, likely the prediction will overestimate more. Your comment makes sense, but note that the new estimates of MLD continue matching adequately with the predicted MLD. In the new version of the manuscript we will mention that the differences between the predicted and observed convection depth could be due to errors in the atmospheric forcing (Josey et al., 2018), lateral advection and/or spatial variation in the convection intensity within the box that was not captured by the Argo sampling.

Discussion
Line 366: This was seen throughout the 1990s and is not quite as surprising as the authors state. We deleted "surprisingly".
Line 397: The Labrador Sea is always more favorably preconditioned, it is quite visible in the hydrographic sections and has been noted before. The Labrador Sea is usually more favorably preconditioned than the Irminger Sea. However, we see that the water column from the surface to 1,300 m in winter 2017 is more favorably preconditioned in the SECF than in the Labrador Sea (see Fig. 7 in the previous version). For example, in order to homogenize the water column down to 1,300 m, 1.80 x 10 9 J m -2 is required in the SECF whereas 2.13 x 10 9 J m -2 is needed in the Labrador Sea.
Line 406: Bit of a chicken and egg problem. The halocline is also deeper in the Labrador Sea because convection is deeper there. Would rephrase. Not really a chicken and egg problem, if you are thinking in terms of preconditioning. To clarify our point we modified the sentence as : "The deep halocline acts as a physical barrier for deep convection in both the Irminger Sea and the Labrador Sea, but because the deep halocline is deeper in the Labrador Sea than in the Irminger Sea, the preconditioning is more favorable to deeper convection in the Labrador Sea than in the Irminger Sea." Line 416 / Fig 10. The depth is chosen such that it is always in the convective regime in the Labrador Sea, hence the nice steps. It is mostly too deep for this in the Irminger Sea, so a lot of the variability is caused by advection except in exceptionally deep convection years. You are right and the figure is confusing even when the discussion is limited to deep convection events in the SECF region. Because of your comment and the comments of reviewer 2 we decide to delete Figure 10 and paragraph 415 -433 in the revised manuscript.
Line 430: There is a multitude of evidence that there was very deep convection in the Irminger Sea in the 1990s (but no Argo program). The LSW was advected to the Irminger Sea in the subsequent years and hence properties converged. Please rephrase.
We decide to remove Figure 10  , so, it does not concern the period we study in our paper. Otherwise, we think that Holliday et al (2019) has not been published yet (V. Thierry is co-author of the paper).

Conclusions
Line 450 "in or near the Irminger Sea" In the revised manuscript this sentence is changed to: "During 2015 -208 winter deep convection happened in SECF reaching deeper than 1,300 m".
Line 473: was this only caused by advection of LSW or was the layer eroded by the 1600 m deep convection in 2016? Our sentence was confusing. We will mention that deep convection of W2016 also favored the preconditioning for winter 2017 -2018.

Received and published: 25 June 2019
This paper reports on a very interesting analysis of recent Argo data in the subpolar North Atlantic. They claim that deep convection in the Irminger Sea, which began in 2015, persisted through 2018 because of favorable preconditioning. They perform some novel analyses and the findings will be of great interest to the community. However, I agree with the review posted by Femke de Jong, which raises issues with the way that mixed layers are defined in the study. This is central to the interpretation and conclusions of the study, and I think that at the very least some major re-framing of the work is necessary. I recommend this work for publication after major revisions. Thank you very much for your constructive review. In the following we answer to each of your comments and describe how we are going to take into account your suggestions in the revised manuscript.
Major comments I would like to echo de Jong's comments regarding the mixed layer depth derivation. In order to show that deep convection occurred in 2016-2018, they should show that all properties (including temperature and salinity) were homogeneous throughout, not just that a density threshold was exceeded at a very deep depth. The three referees agreed on this point. Consequently, we have adapted our methodology to estimate the MLD considering density, temperature and salinity profiles. Please, refer to the beginning of the answer to de Jong (referee 1) in order to see how we have modified our methodology to estimate MLD. The new results (MLD and properties) do not change the main conclusions of our paper.
Regardless of the method selected by the authors in their revision they should include much more detail on it in the text as it is a central calculation. They should also be clear about how sensitive their results are to the method used and to the thresholds that are selected. They should also detail how their methods relate to the methods used by previous studies in the region. Right, we will explain our revised methodology to estimate the MLD indicating the threshold of density, temperature and salinity used. Moreover, we will add a figure in supplementary material showing that our estimates of MLD favorable compare with the estimates resulted when using the methods of Pickart et al. (2002) or de Jong et al. (2012) as discussed at the beginning of response to referee 1.
The authors should address how sensitive their results are to the Argo float coverage, and what portion of the Argo floats present have deep mixed layers. They report how many floats have mixed layers deeper than 700m, but not how many were present. Does the percentage of floats with deep mixed layers decrease over time? The authors should comment on why they think so few Argo floats have deep mixed layers. Is it consistent with their buoyancy forcing analysis? Does the sampling in time account for some of this: i.e. Are deep mixed layers seen more commonly late in winter? Thanks to your comment we realized that our discussion was misleading because it was based on the percentage of profiles showing deep convection during the entire winter, which is small by construction because only profiles at the end of winter show deep convection. We rather should have count the number of floats showing deep convection during a given year. Accordingly, we now identify the period when deep convection occurs as the period when at least one profile shows MLD > 700 m (the period begins when a profile with MLD > 700 m is detected for the first time for the given winter and it ends when there is no more profiles with MLD > 700 m). Then, we quantified the percentage of floats with deep MLD present in that period and region (pink box in figure R4 of answer to referee 1). This information is summarized in table R2 of this document and will be included in section 4.1 of the revised manuscript. The percentage varies between 33% and 73%. In 2017, the three profiles with deep mixed layer were recorded by three different floats, all located in the southwest corner of our region. This shows that the convection area was confined to a small area of the SECF region and explains that the lowest percentage is observed in 2017. Table R2. Sensitivity study about the Argo float coverage in the SECF region (pink box in Figure R4 of the answer to referee 1). Period is the period during which floats with deep mixed layers were observed. We indicate the total number of floats found in the SECF region during the indicated period, and the number of floats showing deep convection. Finally, the percentage of floats showing deep convection is indicated. The mixed layers reported in winter 2018 are almost all to the south of Cape Farewell, and not in the Irminger Sea. Further, the TS properties in 2018 are much more similar to Labrador Sea properties than Irminger Sea properties ( Figure 3). This is consistent with the SCF box properties reported in Piron et al. 2017. Some of the properties in 2016 and 2017 may also fall in that category, I don't think the author's should be calling this "Irminger Sea convection". Right. In the revised manuscript we changed the northern limit of the pink box to 59.3°N instead of 61°N previously and refer to the pink box as Southeast Cape Farewell (SECF).
The author's show a very interesting analysis of Labrador Sea properties which are advected into Irminger Sea and contribute to the deepening of the Irminger Sea halocline (Figures 7 and 9). Is this advection limited to the 1200-1400 range they consider? How does advection from the Labrador Sea in other depth ranges fit in? Advection from the Labrador Sea certainly contributed to vary the properties from the surface to 1000 m. However, the buoyancy budget showed that this is minor contribution compared to the buoyancy loss due to the local air-sea flux. We add a comment about it in the revised manuscript.
The title is confusing, and I wonder if, in general, the author's should shift their focus from the Irminger Sea in particular and instead focus on the important connections between the intermediate waters in the subpolar North Atlantic. I think the authors have an opportunity here to clarify that intermediate waters are formed in many places and how the connections between these basins affect intermediate water mass properties. Their focus on salinity in addition to temperature would make this angle particularly interesting.
In the revised manuscript we change the title to: "Why did deep convection persist over four consecutive winters (2015-2018) Southeast of Cape Farewell?" Moreover, we now mention several times the role of advection from the Labrador Sea. We also added to the discussion the following paragraph: The

Minor comments
The link between anthropogenic forcing and the recent convection that is drawn in the first few sentences in the abstract and throughout the introduction is a bit of a stretch. The motivation could be made more direct and convincing, and this type of speculation could remain in the discussion where it is more relevant. Ok, in the revised manuscript we exclude the references about the anthropogenic forcing by deleting the first sentence in the Abstract and the first paragraph in the Introduction.
L109: "for the first time to our knowledge in this region" -this is a broad claim and not necessary. Ok. Deleted.
Section 3.1: Please add significant detail on the mixed layer estimation method. Right, it has been added in the revised manuscript as indicated at the beginning of this document.
L222: The 2018 profiles with deep mixed layers are not in the Irminger Sea. Right, as explained above, we changed the limit of the pink box and refer to our pink box as SECF.
L237: "Water masses formed are very similar" It should at least be acknowledged that they are formed much closer to the Labrador Sea than in previous years.
Right, when excluding the floats south of Cape Farewell as requested by referee 3, the properties of the water mass formed in the SECF region in W2018 is not similar to the formed in the Labrador Sea in W2018. The sentence "Water masses formed are very similar" is excluded in the new version of the manuscript. L331: "despite they were also fresher" ! "despite the fact that they were also fresher" Ok, we will change it.
L340: Refer to figure 6. Ok, we will write, "The predicted convection depths are determined as the depth at which B(zi) (Fig. 6a), equals the atmospheric forcing." L348: Clarify what happened here. These floats only profiled down to 1,100m?
Exactly. We would rewrite the sentence as: "This result is in line with the fact that among the 10 profiles that we used to compute Q3 in W2018, 6 showed deep convection down to 1,100 m and were recorded by floats with a maximum profiling depth of 1,100 m, most likely leading to an underestimation of the MLD." However, this sentence is going to be deleted in the revised manuscript. In the revised paper, and following the suggestion of referee 3, we exclude from the analysis the profiles that do not extend beneath the base of the mixed layer, because it results in bias in the properties related to the mixed layer.
L351: I was also confused by the fact that the author's claim to neglect advection, but cite advection of properties from the Labrador Sea as a reason for favorable preconditioning. Perhaps remove that claim. Additionally, the fact that the T and S properties are not homogeneous goes against the idea that deep convection is occurring locally. We agree that this paragraph was confusing. We now identify lateral advection as a possible cause for the buoyancy budget residuals. The profiles with non-homogenous TS in the mixed layers are now excluded from the analysis.
L370: hydrological ! hydrographic. Yes, hydrographic, we will change it. This figure and paragraph were not essential for the conclusions of the paper. We decide to remove them in the revised manuscript.
Figure 4: Note the differences between the axis ranges in the caption. This figure could be featured earlier as it provides important context. Ok, we will write in the figure caption: "Note the differences between the axis ranges". We refer to this figure earlier in the section 4.2 of the revised manuscript. , that corresponds to sig0 equals to 27.754 Kg m 3 which is the density of the mixed layer in the SECF for these winters (Fig. 3). These results support local formation. Accordingly, we add at the end of the first paragraph of section 4.3: "The denser density of the core of the thick layers in 2017 -2018 compared with 2015 -2016 agrees with the densification of the mixed layer SECF shown in Table 1

Anonymous Referee #3
In this manuscript persistence of deep convection in the Irminger Sea is investigated. One winter of particularly severe atmospheric forcing and deep convection was followed by three winters of climatological strength which also had deep mixed layers. The authors quantified the buoyancy loss required for deep convection to commence each winter and concluded that the preconditioning arising from the previous winter's homogenization of the water column was a main reason for the persistence of deep convection. I think this manuscript has the potential to be an important and valuable contribution to better understand deep water formation in the Irminger Sea /subpolar North Atlantic. However, as made clear also by the other reviewers, I have concerns about the determination of mixed-layer depths. As such, I recommend that the paper be revised before publication. Thank you for your valuable comments; they help us improve our work. In the following we answer point by point to all of your comments and explain how we will modify the manuscript accordingly.

Major comments:
I am not convinced that automated routines, such as the threshold or split and merge methods, are particularly suitable for determining the vertical extent of the mixed layer. These routines generally perform well when applied to summer and fall profiles, when the upper ocean is stratified and there is a pronounced density difference between the mixed layer and the lower part of the profile. However, they are less accurate during periods of active convection when stratification is eroded. Furthermore, such routines cannot identify mixed layers that are isolated from the surface, either in the form of vertically stacked mixed layers or by early stages of surface restratification. Such isolated mixed layers are prevalent in the Labrador and Irminger Seas during winter (e.g. Pickart et al., 2002). As pointed out by the other referees, if the density profile is considered in isolation, changes in temperature and salinity may be density-compensated such that the water column can appear to be homogenized while in reality it is not. Examples of that can be seen in If widespread deep convection occurred also during these winters, there should be many more profiles with deep mixed layers. Is it possible that the mixed-layer depths determined by the automated routines are remnants of deep convection from a previous winter or from the Labrador Sea where mixed layers are generally deeper? The percentage of profiles with deep MLD depends on the period during when we compute the statistics. Our previous method was misleading because we considered the entire winter for computing the statistics and not only the convection period (see also answer on this point to referee 2). We now identify the period during which deep MLDs > 700 m were observed for each winter in the Southeast Cape Farewell (SECF) region (pink box in Fig. R4 in referee 1 answer) (see answer to reviewer 2 for more details). Then, we quantified the percentage of floats that measured deep MLD in the region and during the period of deep convection. The results are shown in table R2. The lower % is found for winter 2017, but it is still substantial and reflects the fact by the fact that the floats showing deep MLD were found southwest of the SECF box suggesting that convection did not occur over the full box. The results of this sensitive study will be added to the section 4.1 of the revised manuscript. Table R2. Sensitivity study about the Argo float coverage in the SECF region (pink box in Figure R4 in the answer to referee 1). Period is the period during which floats with deep mixed layers were observed. We indicate the total number of floats found in the SECF region during the indicated period, and the number We do not think that the observed MLD are remnants of deep convection from a previous winter or from the Labrador Sea because the new estimates of MLD are from profiles homogenous in terms of density, temperature and salinity. Most importantly, the fact that the 1D-buoyancy budget is nearly closed (section 4.3) is also an indication that deep convection occurred locally in the SECF box during winters 2016, 2017 and 2018.
To get a more robust estimate of convection in the subpolar North Atlantic these winters, I suggest dispensing with the 700 m "deep convection" criterion and showing if not every mixed layer at least the 50-80% deepest mixed layers encountered by each float every winter. That would remove shallow mixed layers arising from early phases of the seasonal evolution of the mixed layer and profiles obtained within stratified eddies, while the remaining mixed layers would allow for more robust quantification of the general depth of convection.
OK, this seems to be a nice idea, but it would bias low the estimate of convection depth if the statistics of MLD were made using the profiles for the entire winter. The criteria should be applied to the convection period that we select here by considering profiles deeper than 700 m because it is the minimum depth that should be reached for LSW renewal. If apply to those profiles your criteria would not be much different from our Q3. Note that our estimate of convection depth based on the statistical criteria Q3 is equivalent to the aggregate maximal convection depth used by Yashayaev and Loder (2017) and allows direct comparison with this author's results.
Profiles that do not extend beneath the base of the mixed layer (there may be some examples in Figure 2d-f) would result in a shallow bias of the mixed-layer depth estimate and should be excluded from the analysis. We agree. These profiles located between 48°W and 45°W are not consider in our new results.
Specific comments: Line 95: It should be: "...Argo and mooring data..." Corrected Lines 106 and 361: Mixed layers exceeding 1400 m depth were determined also from shipboard measurements in the Irminger Sea in April 2015 (Fröb et al., 2016). We add this reference to the revised manuscript.
Line 122: If the TEOS-10 convention is used, conservative temperature and absolute salinity should be used instead of potential temperature and salinity. TEOS-10 allows the computation of theta and practical salinity.
Line 123: Please explain why a salinity of 35 was chosen as a reference value. This sentence is deleted in the manuscript because we do not use FW in the paper. Sorry for the confusion it may have caused.
Line 124: Please provide more information about the gridded products. Are different time periods and resolutions the only difference between the products? What are the errors, in particular for the EN4 product which extends back to 1900 and covers some very data-sparse periods? ISAS and EN4 are optimal interpolation of in situ data, but the optimal interpolation method is not exactly the same in both products due to different choices for the spatial and temporal correlation functions used for the optimal interpolation. Details about both databases are described in the references given in the manuscript ( Note that we used EN4 data from 1993 afterwards and that the monthly temperature and salinity fields at a given time only depends on the data found in a short time window around the date of the analysis. The data sparse-period at the beginning of the 1900 did not influence our results.
Line 130: Does the net air-sea heat flux include radiative fluxes or only turbulent fluxes? It includes both radiative and turbulent fluxes. We indicate it in the revised manuscript.
Line 149: I do not think that 48•W is commonly used as a border between the Labrador and the Irminger Seas. Many of the deep mixed layers were recorded directly south of Greenland, in a region that is not really part of either the Labrador or the Irminger Seas. Ok, the limit at 48°W was used just to include in the analysis of the MLD properties the profiles found between 48°W and 45°W in 2018. In the revised computation we used only profiles inside the pink box which limit is at 45°W and we now refer to the pink box as Southeast Cape Farewell (SECF) instead of Irminger Sea. Note that the northern limit of the box is changed from 61°N to 59.3°N. We calculated the atmospheric forcing and the preconditioning considering this new box limit and it does not change the main results and conclusions of our work.
Line 156 and elsewhere: Please insure that all papers cited in the text are included in the References section. For example is Gill (1982) missing. Ok, thank you for noting it.
Line 174: How was the depth of the Ekman layer estimated?
We used the Ekman transport and we considered that the SST is representative of the temperature in the Ekman layer. We will clarify this point in the revision.
Line 179: For consistency, it might be better to use SST also from the EN4 product. Ok, we have estimated the Ekman Buoyancy Flux (BFek) using EN4 SST. The horizontal Ekman Buoyancy flux in the SECF region (pink box in Fig. R4 in response to referee 1), accumulated from 1 September to 31 August the year after was estimated with: i) with EN4 SST and EN4 SSS and ii) with ERA SST and EN4 SSS; they are represented in Figure R5. Both time series show the same behavior but the results obtained with EN4 SSS and EN4 SST are smoother than the results obtained with ERA SST and EN4 SSS. Thank you for your comment, we switched to EN4 SST. Line 430: There were no wintertime measurements in the Irminger Sea in the early 1990s, but there is strong indirect evidence that deep convection occurred in the Irminger Sea at that time (see for example publications from the group of R. Pickart). Right, there are evidences that deep convection occurred in the Irminger Sea in early 1990s (Pickart et al., 2003). In any case, the three referees find something wrong in this paragraph and Figure 10. Because this paragraph and figure is not important for the conclusions of our paper we decide to remove them in the revised version of the paper.
Line 481: Acknowledgement is misspelled. Right, thank you.
Lines 519 and 522: The name de Jong is inconsistently capitalized. Right, corrected.   methods (see section S1, Fig. S1 and Fig. S2 in supplementary material). 168 In this paper, deep convection is characterized by profiles with MLD deeper than 700m (colored big 169 points in Fig. 1) because it is the minimum depth that should be reached for Labrador Sea Water 170 (LSW) renewal (Yashayaev et al., 2007;Piron et al. 2016). 171 The winter MLD and the associated θ, S and  properties were examined for the Labrador Sea and 172 the SECF region by considering the profiles inside the cyan and pink boxes in Fig. 1 (Table 1). We additionally showed that for both 262 winters Q3 was about 1300 m (Table 1). Now, we describe the convection of W2017 and W2018. In 263 W2017, deep convection was defined from four three Argo profiles in the Irminger Sea (see Fig. 1c  264 and Fig. 2a-c). The maximum MLD of 1,400 m was observed on 16 th March 2017 at 56.65°N -265 42.30°W. The aggregate maximum depth of convection Q3 coincided with the maximum MLD 266 because the estimates are based on only four profiles. In W2018, ten profiles showed MLD deeper 267 than 700 m in the Irminger Sea (Fig. 1d, 2d-f). Tthe maximum MLD of 1,300 m was observed on 24 268 February at 58.12°N, 41.84°W. (Fig. 1d, 2d-f). Float 5903102 measured MLD of 1,100 m South of Cape 269 Farewell (Fig. 1d), but the estimated MLDs coincided with the deepest levels of measurement of the 270 float so that these estimates, possibly biased low (see Fig. 2d-f), were discarded from our analysis. 271 The aggregate maximum depth of convection Q3 was 1,100 m. Float 5903102, which was localized 272 South of Cape Farewell, did not profile deeper than 1,100 m in any of its six cycles (see Fig. 2d The properties (σ 0 , S and θ) of the end of winter mixed layer were estimated for the four winters 287 (Table 1 and Fig. 3). We observed that, between W2015 and W2018, the water mass formed by deep 288 convection significantly densified and cooled by 0.019 kg m -3 and 0.3060.215°C, respectively (see 289   Table 1). 290 In the Labrador Sea, Q3 increased from 2015 to 2018 (see Table 1). Deep convection observed in the 291 Labrador Sea in W2018 was the most intense since the beginning of the Argo era (see Fig. 2c in 292 Yashayaev & Loder, 2016). From W2015 to W2018, newly formed LSW cooled, salted and densified 293 by 0.134°C, 0.013 and 0.023 kg m -3 , respectively (Table 1). 294 The water mass formed in the Irminger SeaSECF is warmer and saltier than that formed in the 295 Labrador Sea (Fig. 3); the exception is in W2018 when the characteristics of the water masses formed 296 in each of the basins are very similar. The deep convection in the Irminger SeaSECF is always 297 shallower than in the Labrador Sea. Both results are discussed later in Sect. 5. The seasonal cycles of B surf * and Q are in phase and of the same order of magnitude, while the FWF* 301 , which is positive and one order of magnitude lower than Q and , does not present a seasonal cycle 302 (Fig. S1S3). The means (1993 -2018) of the cumulative sums from 1 September to 31 March of Q, 303 FWF* and B surf * estimated over the Irminger SECF box (Fig. 1) are -2.46 ± 0.43 x 10 9 J m -2 , 0.28 ± 0.10 304 x 10 9 J m -2 and -2.22 ± 0.49 x 10 9 J m -2 -2.52 ± 0.43 x 10 9 J m -2 , 0.31 ± 0.11 x 10 9 J m -2 and -2.26 ± 305 0.51 x 10 9 J m -2 , respectively. Despite B surf * is mainly explained by Q, the accumulated FWF* amounts 306 to ~10 % of the accumulated Q with opposite sign. The air-sea buoyancy flux atmospheric forcing 307 estimated in terms of buoyancy is therefore 10% lower on average than when estimated in terms 308 ofthe air-sea heat flux. Considering the Ekman transports, the 1993 -2018 means of the 309 accumulated BF ek , HF ek and SF ek from 1 September to 31 March amount to 0.37 ± 1.15 x 10 8 J m -2 , -310 0.35 ± 1.36 x 10 8 J m -2 , and 0.02 ± 2.04 x 10 8 x 10 9 J m -2 , respectively. The horizontal Ekman heat flux 311 is negative, while the Ekman buoyancy flux is positive. This buoyancy gain indicates a southeastward 312 transport of surface freshwater caused by dominant winds from the southwest. Noteworthy, BF ek is 313 one order of magnitude smaller than the B surf *. 314 Piron et al. (2016) found that "the wind stress led to an Ekman-induced heat loss that reinforced by 315 about 10% the heat loss induced by the net air-sea heat fluxes". Here, we considered the buoyancy 316 flux induced by the Ekman response to the wind stress and we estimated the buoyancy, heat and salt 317 Ekman fluxes (BF ek , HF ek and SF ek ). The means (1993 -2018) of the accumulated BF ek , HF ek and SF ek 318 from 1 September to 31 March amount to -0.0004 ± 0.04 x 10 9 J m -2 , 0.0446 ± 0.04 x 10 9 J m -2 , and 319 0.0626 ± 0.04 x 10 9 J m -2 , respectively. So, on average, HF ek and SF ek compensate each other resulting 320 in an almost zero J m -2 BF ek . However, for particular years with strong wind stress as it was the case in 321 2015, there is no such compensation and BF ek is different from 0 (see Fig. 4). 322 We now compare the accumulated B surf * from 1 September to 31 March the year after for the last 323 four deep convection years. It amounted to -3.21 x 10 9 ± 0.05 J m -2 , -2.29 ± 0.04 x 10 9 J m -2 , -2.23 ± 324 0.05 x 10 9 J m -2 and -2.58 ± 0.05 x 10 9 J m -2 for W2015, W2016, W2017 and W2018, respectively. The 325 cumulative sum of BF ek from 1 September 2014 to 31 March 2015 was -0.27 ± 0.04 x 10 9 J m -2 ; the 326 estimates for the following winters were near 0 J m -2 . When the BF ek is added to the B surf *, the 327 resulting atmospheric forcing is -3.48 x 10 9 ± 0.05 J m -2 , -2.19 ± 0.04 x 10 9 J m -2 , -2.20 ± 0.05 x 10 9 J 328 m -2 and -2.57 ± 0.05 x 10 9 J m -2 for W2015, W2016, W2017 and W2018, respectively. The estimate 329 for W2015 is ~30% larger than the estimates for 2016 -2018. Time series of atmospheric forcing 330 anomalies in Fig. 4 show that this strongly negative W2015 anomaly of accumulated B surf * was 331 caused by very negative Q and FWF anomalies and a negative BF ek as well. During W2016, W2017 332 and W2018 however, all atmospheric forcing terms were close to zero. 333 The total atmospheric forcing SECF was quantified as the sum of B surf * and BF ek . The anomalies of 334 accumulated fluxes from 1 September to 31 August the year after, with respect to the mean 1993 -335 2016, are displayed in Fig. 4 for the SECF box. The grey line in Fig. 4a is the total atmospheric forcing 336 anomaly (B surf * plus BF ek ). We identify years with very negative buoyancy loss in the SECF region, e.g. 2015 were caused by the very negative anomalies in both B surf * (Fig. 4a) and BF ek (Fig.4d). This 339 correlation was not observed for all the years presenting a negative anomaly of atmospheric forcing. 340 Noteworthy, during W2016, W2017 and W2018, the anomaly of atmospheric forcing was close to 341 zero. 342 From these results we conclude that, cContrary to the very negative anomaly in atmospheric fluxes 343 over the Irminger SeaSECF region observed for W2015, the atmospheric fluxes were close to the 344 mean during W2016, W2017 and W2018. 1,000 m) became colder than the years before and, despite a slight decrease in salinity, the cooling 353 caused the density to increase (Fig. 5c). Fig. 5d shows  1 =0.01 kg m -3 layer thicknesses larger than 354  Table 1 and Fig. 3. 360 B(zi) is our estimate of the preconditioning of the water column before winter (see Method). Fig. 6a  361 shows that, deeper than 100 m, B was smaller for W2016, W2017 and W2018 was smaller than for 362 W2015 or B for the mean W2008 -W2014. Furthermore, for W2016, W2017 and 2018, B remained 363 nearly constant with depth between 600 and 1,300 m, which means that once the water column has 364 been homogenized down to 600 m, little additional buoyancy loss results in homogenization of the 365 water column down to 1,300 m. Both conditions (i) less buoyancy to be removed and (ii) absence of 366 gradient in the B profile down to 1,300 m indicate a more favorable preconditioning of the water 367 column for W2016, W2017 and W2018 than during W2008 -W2015. profiles (discontinuous blue lines in Fig. 6). B θ accounts for most of the increase in B from the surface 375 to 800 m and below 1,400 m (see Fig. 6a and Fig. 6b). The negative slope in the B s profile between 376 800 -1,000 m (Fig. 6b) slightly reduces B (Fig. 6a) and is due to the decrease in S associated with the 377 core of LSW (see Fig. 3 in Piron et al. 2016). In the layer 1,000 -1,400 m, the increase in B (Fig. 6a) is 378 mainly explained by the increase in B s (Fig. 6b), which follows the increase in S in the transition from 379 LSW to Iceland Scotland Overflow Water (ISOW). This transition layer that will be referred to  Fig. 7a). Finally, we note that the profiles of B(z i ), B θ (z i ) and B s (z i ) for 395 W2016 and W2018 are more similar to the profiles of W2017 than to those of W2015 or to the mean 396 2008 -2014 (see Fig. 6), which indicates that the water column was also favorably preconditioned 397 for deep convection in W2016 and W2018 for the same reasons than in W2017. 398 The origin of the changes in B is now discussed from the time evolutions of the monthly anomalies of 399 θ, S and σ 0 at 58°N -40°W that is at the center of the SECF box (Fig. 8) and can be summarized as follows. On the one hand, the properties of the surface waters (down to 413 500 m) were colder than previous years and, despite they were also fresher, they were denser. The 414 density increase in the surface water reduced the density difference with the deeper-lying waters. 415 The intermediate layer (500 -1000 m) was also favorably preconditioned due to the observed 416 cooling. Additionally, in the layer 1,100 -1,300 m, the large negative contribution of B s in 417 relationwith restpect to its mean is explained by the decrease in S in this layer, which caused a 418 decrease in σ 0 and, consequently, reduced the σ 0 difference with the shallower-lying water. The 419 decrease in S also resulted in a deepening of the deep halocline. 420 4.4. Atmospheric forcing versus preconditioning of the water column 421 We now use the estimates of the accumulated atmospheric forcing (B surf * + BF ek ) from 1 September 422 to 31 March the year after (see Fig. S4) to predict the maximum convection depth for a given winter 423 based on September profiles of B. The predicted convection depth is determined as the depth at 424 which B(zi) (Fig. 6a) equals the accumulated atmospheric forcing. The associated error was estimated 425 by propagating the error in the atmospheric forcing (0.05 x 10 9 J m -2 ). The accumulated atmospheric 426 forcing amounted to -3.21 x 10 9 ± 0.05 J m -2 , -2.21 ± 0.04 x 10 9 J m -2 , -2.01 ± 0.05 x 10 9 J m -2 and -2.47 427 ± 0.05 x 10 9 J m -2 for W2015, W2016, W2017 and W2018, respectively. We found predicted 428 convection depths of 1,085 ± 20 m, 1,285 ± 20 m, 1,415 ± 20 m and 1,345 ± 20 m for W2015, W2016, 429 W2017 and W2018, respectively. We consider Q3 as the observed estimate of the MLD (Table 1) favorable preconditioning of the water column. Note that, the cooling affected the layer from surface 472 -to 1,400 m and the freshening affected the layer from near surface to 1,000 -1,5001,600 m (Fig. 8), 473 but the cooling and the freshening were intensified at different depth ranges (Fig. 8) intensified over the whole SPNA in February 2017. By this mechanism, the advection from the 492 Labrador Sea contributed to create property anomalies in the water column. However, the buoyancy 493 budget showed that this was a minor contribution compared to the buoyancy loss due to the local 494 air-sea flux, even if it was essential to preconditioning the water column for deep convection. 495 We now compare the atmospheric forcing and the preconditioning of the water column in the 496 Irminger Sea SECF region with those of the nearby Labrador Sea where deep convection happens 497 each almost every year. As noted by Pickart et al. (2003), tThe atmospheric forcing over the Labrador 498 Sea is ~15 % larger than that over the Irminger SeaSECF region: the means (1993 -2018) of the 499 atmospheric forcing, defined as the time -accumulated B surf * + BF ek from 1 September to 31 March 500 the year after, are -2.61 ± 0.55 x 10 9 J m -2 in the Labrador Sea and -2.18 ± 0.54 -2.26 ± 0.58 x 10 9 J m -2 501 in the Irminger SeaSECF region. The difference was larger during the period 2016 -2018 when the 502 atmospheric forcing equaled -3.10 ± 0.19 x 10 9 J m -2 in the Labrador Sea and -2.23 ± 0.23 -2.31 ± 0.21 503 x 10 9 J m -2 in the Irminger SeaSECF region. In terms of preconditioning, the 2008 -2014 mean B 504 profile (blue continuous lines in Fig. 7) was lower by ~0.5 x 10 9 J m -2 in the Labrador Sea than in the 505 Irminger SeaSECF for the surface to 1,000 m layer and by more than 1 x 10 9 J m -2 below 1,0001,200 506 m. It indicates that the water column was more favorably preconditioned in the Labrador Sea than in 507 the Irminger SeaSECF region during 2008 -2014. Differently, B for W2017 shows slightly lower values 508 from the surface to 1,300 m in the Irminger SeaSECF region than in the Labrador Sea (see orange 509 lines in Fig. 7). However, B in the Labrador Sea remains constant down to the depth of the deep 510 halocline between LSW and North Atlantic Deep Water (NADW) at 1,700 m. In the Irminger SeaSECF 511 region, the deep halocline remained at ~1,300 m between 2016 and 2018 (see B s lines in Fig. 7a). In the following we consider the time -evolution of , S and σ 0 at the layer 700 -900 m (Fig. 10), 533 considered here as the core of the LSW in both the Irminger Sea and the Labrador Sea.  Table 1). In spite of the cooling and densification that occurred 539 during the last winters, LSW is warmer and lighter than that formed during W1994 and W1995 540 (2.85°C, 27.78 kg m -3 , Pickart et al. 2003). We note a long-term (1994 -2018) warming of LSW 541 observed in the Irminger Sea. The comparison with the LSW properties in the Labrador Sea over 2002 542 -2018 (Fig. 9) shows that the LSW observed in both basin has the same density while that of the 543 Irminger Sea is warmer and saltier than that of the Labrador Sea. Interestingly, this behaviour was 544 also observed along the 90s (Pickart et al, 2003). It is also worth noting that  and S observed in 545 Labrador Sea and Irminger Sea converged at the end of the 90s (Fig. 6 in Pickart et al., 2003) and 546 along our period 2015 -2018 ( Fig. 3 and Fig. 10) Considering the expected increase in freshwater inputs, the atmospheric forcing and preconditioning 572 of The atmospheric forcing and preconditioning of the water column was evaluated in terms of 573 buoyancy. We showed that the atmospheric forcing is 10% weaker when evaluated in terms of 574 buoyancy than in terms of heat because of the non-negligible effect of the freshwater flux. The 575 analysis of the preconditioning of the water column in terms of buoyancy to be removed (B) and its 576 thermal and salinity terms (B θ and B s ) revealed that B θ dominated the B profile from the surface to 577 800 m and B s reduced the B in the 800 -1000 m layer because of low salinity of LSW. Deeper, B s 578 increased B due to the deep halocline (LSW-ISOW) that acted as a physical barrier limiting the depth 579 of the convection. 580 During 2016 -2018, the air-sea buoyancy losses were close to the climatological values and the very 581 deep convection was possible thanks to the favorable preconditioning of the water column. It was 582 surprising that these events reached convection depths similar to those observed in W2012 and 583 W2015, when the latter were provoked by high air-sea buoyancy loss intensified by the effect of 584 strong wind stress. It was also surprising that the water column remained favorably preconditioned 585 during three consecutive winters without strong atmospheric forcing. In this paper, we studied the 586 reasons why this happened.