Review Selivanova, Iovino, Cocetta: Past and future of the Arctic sea ice in HighResMIP climate models

General: 

The authors use a sub-set of HighResMIP model (those that have been used as part of the EU-project PRIMAVERA) to investigate the effect on increasing ocean and atmosphere resolution on the representation of ice area and ice thickness/ volume in the Arctic and within Arctic sub-regions and their variability and past and future trends. The main finding is that increasing the resolution does not lead to a robust signal in the analysed sea ice parameters across the investigated models. 

The article is generally clearly structured and easily understandable, and also the figures are easy understandable. Not all the presented results are entirely new, mainly because high resolution often does not show substantially changes in the results and thus studies based on historical and future simulations with “standard-resolution” CMIP6 models have shown comparable results. However, there are still sufficient new and interesting results, which would justify a publication after some additional analysis and clarifications.

Major points:

1. The authors are not discussing the role of internal variability, and how it might affect differences between low and high-resolution versions of the models. Since only one member of each model configuration is analysed, internal variability itself can lead to large differences in both mean values and trends (compare for example results from Swart et al. 2015, Jahn et al. 2016 using single-model large ensembles or Karami et al. 2023 looking at long term internal variability), and internal variability can probably explain quite a part of the differences between the low and high resolution versions - at least for some parameters and regions. Internal variability might also be a contributing factor to the opposite response to increasing resolution that has been found in the different models, where the authors so far state that the response to increasing resolution depends on the model. 

To solve this problem, authors could either include more HighResMIP models into their analysis to investigate if more models show the same response to increasing resolution, or they could include all existing model members of the models they used in their analysis to find a more robust response in each single model. At the least, the authors should use proper significant tests to find out if the differences that have been found between LR and HR versions of the models are statistically significant.

2. Linked to the internal variability: At certain places the manuscript gives the impression (but I hope this was only somewhat misleading formulated) that the authors expect that global coupled models should be able to represent observed extremes. For example, they state that none of the models is able to reproduce the observed sea ice minima in 2007 and 2012 (section 3.4). Due to internal variability, we can never expect that a coupled model will represent a certain extreme in a certain year in long-term historical and future simulations as performed in CMIP6 / HighResMIP. Also, in section 3.4, the authors correlate annual sea ice values from the models with the satellite data to investigate how well the interannual variability is represented. If this is really what has been done (if I misunderstood please clarify what really has been correlated), this analysis must be removed because for the same reason as historical coupled model simulations cannot reproduce observed extremes at the observed time, they cannot represent observed interannual anomalies. That the correlations for SIV nevertheless provide relatively high values might be due to similar trends in satellite data and models – in case no detrending has been done before the correlation (even this should be clarified). 

3. I am a bit puzzled by the regional results in section 3.3. Figure 5 shows large differences (up to 10-15%) in sea ice area between models in e.g. CA, LV, ESS even in winter when we would expect that ice concentration is rather close to 100%. To me it seems that the area of the same sub-region is of different size in the different models. If you look at Figure 5 you will see that the sea ice area in models often is not changing at all between late autumn and spring, which indicates that the maximum ice area for this region has been reached. And if not the max sea ice concentration in some of the models have been hard-coded to 90 or less percent, then the area of the region just has to be smaller in those models which show a much smaller sea ice area. This would be a major issue since then the ice areas/ volumes from the different models and between models and satellite data and maybe even between the same model in different resolution cannot be compared. For proper comparison all model and observational data sets should be interpolated to the same grid and it has to be ensured that land-sea masks are comparable between the models. 

Please explain more in detail how you calculated the area of the regions and carefully check all results in section 3.3, and if necessary update them.

Minor points:

1. The HighResMIP-protocol foresees that differences between high and low resolution versions should be kept to a minimum, however, some models include automatized adaptions of parameter values to the resolution, other model groups had to make some small modifications to make the model running properly at high resolution. Further, the LR-versions normally were tuned and the HR-versions not at all. While this is in line with the HighResMPI-protocol, an untuned model might lead to substantial net surface and top of the atmosphere net energy imbalances, which leads to drifts and could affect the response of the climate.  

I suggest to shortly mention that these factors might also contribute to the response to increasing the resolution.

2. Line 31-33: The sentence is formulated such as it sounds as the strongest trend is in September and the 2nd strongest in March. But I guess what is meant is that the strongest trend is in September and the lowest trend is in March. Please check and clarify.

3. Line 50: ‘…no significant changes in the area of the Arctic MIZ… ” But the position of the MIZ changed? If yes, maybe worth to mention.

4. Line 98/99: ‘For the past, sea ice properties …’ This sentence is badly placed here. I suggest to remove it here and mention the period else where (e.g. around line 112).

5. Table 1: Since the focus is on sea ice, it might make sense to compare some basics of the sea ice model and sea ice paramterizations as well.  

6. Line 152: ‘…does not change significantly …’. Please clarify how you calculated the significance.

7. Line 159: ‘Configurations with finer ocean resolution have a better fit to CDR …’ How do you estimated this? Bye eye? Except for the ECMWF-model, I have difficulties to clearly see this. 

8. Line 172: It seems that a number of models simulate despite too thick ice, too small ice concentrations compared with satellite data. I think this could be an important finding for guiding sea ice model improvements and their tuning, and should be mentioned somewhere, e.g. in the discussion or conclusion section. 

In the same context, see also Fig. 3, some models, particularly, both CMCC-versions, EC-Earth3P-HR and ECMWF simulate the ice edge pretty well (and the ice area) but SIT (and ice volume) is much too large. Normally, I would expect that too thick ice should be to a certain degree linked to too large ice areas as well but this does not really seem to be the case. Do you have any explanation for this piling up of sea ice in the Arctic? Sea ice model short comings? Too strong lateral melting? Or too strong Beaufort Gyre keeping all the ice in the Central Arctic?

9. Line 184: Models tend to have the SIA minimum too early, several already in August. Maybe worth to mention.

10. Line 219-221: The two sentences are saying more or less the same. I suggest merging them to one, e.g.:  It is worth noting that the evaluation of the simulated MIZ area is highly dependent on the reference product used, particularly in summer.

11. Line 292-295: It is not surprising that models agree in winter sea ice area better than in summer ice area in Arctic Ocean areas since in winter the Arctic Ocean is entirely ice covered. In summer, winter/ spring thickness is one important factor for how quickly the area is getting ice free and certainly differences between models in summer atmospheric circulation and temperature play a major role as well. I am sure heat input (and volume input) of rivers also matters in reality, however, I am not aware, that any of the models explicitly simulates temperature of the inflowing river water. Do not most models use the ocean temperature at the river mouth/ ocean points where the runoff enters the ocean? If you would like to state that the heat input from rivers is the main reason for the summer spread across models, please provide more details on how the different models differ in terms of their river representation and how large the difference in terms of heat-input into the Arctic from rivers is between models.  

12. Section 3.3: The rather lengthy description in this section is partly a bit difficult to follow. A table - for at least SIA - showing winter (March) and summer (September) ice area for each region in each model and in satellite data would help.

13. Line 349: EC-Earth3P-HR simulates much thicker ice than the LR-version. This is likely explaining why the trend in ice area is smaller. 

 Technical corrections:

Line 98: Add ‘8.5’ to SSP5-scenario: SSP5-8.5.

Line 130 ‘north’ instead of ‘North’

Line 170: ‘addition’ instead of ‘Addition’

Line 408: I suggest to change ‘14 models’ with ’14 simulations’ , since only 6 different models are used.

Selivanova et al., examine past and future Arctic sea-ice variability and changes in a subset of models participating in HighResMIP. The authors investigate how high-resolution (HR) and low-resolution (LR) versions of a model affect a range of Arctic sea ice variables, including sea ice thickness, volume, area, and concentration both over the historical record and under a future emissions scenario. The authors find that increasing the horizontal model resolution does not lead to a any significant difference in Arctic sea ice.



This manuscript is clear and the figures are quite clear. Despite the main result showing that HR models does not substantially change Arctic sea ice trends (when compared to LR models), it is important to document and has important implications for future modeling efforts with refined grids. However, I think there are some overlooked aspects of this result that might change the key message. Thus, I think this manuscript should be published after some additional analysis and clarifications. Below I describe these concerns and suggestions.



Major 



I am concerned that the authors overlooked the role of internal variability on Arctic sea ice trends and variability. Internal variability is known to be highly model dependent (Bonan et al., 2021) and strongly influence sea ice trends (Swart et al., 2015). I think it would be helpful for the authors conduct additional analyses that examine other members of each model. A quick glance at the HighResMIP archive (https://esgf-node.llnl.gov/search/cmip6/) suggests this is possible for at least some models. For instance, CNRM-CM6-1 has 10 members. If not all models have more ensemble members, it could be worthwhile to focus on comparing HR and LR results in a model with 10 ensemble members (e.g., CNRM-CM6.1). My belief is that HR and LR models will have different "forced" responses and this results itself could broaden the study. I also think the HR and LR models will likely have different internal variabilities based on Fig. 3 which shows that HR and LR models have different SIV mean states.



In summary, I strongly suggest the authors conduct additional analyses that essentially repeat this analysis but with a more robust quantification of the "forced" response and internal variability.



Minor



L33 and L38: Cite Fetterer et al., 2016 and Stroeve & Notz, 2018 instead of the https links.



Line 170: Remove capital "A" from addition.

L184-185: It is probably worth mentioning that some models have biases as their summer minimum is in August rather than September like in observations.



L407-409: I would suggest changing models to simulations. Only six models were used.