This work tried to set up an Arctic SIE prediction system based on the deep learning method, which is important for not only the SIC short-term prediction but also the understanding of the SIE changes. However, there are gaps between the manuscript and publication, and several key issues should be clarified before publication.

Below I include a number of comments, both minor and more significant in no particular order, that I’d prefer to be considered before publication.

1. Lines 3-4: In fact, sea-ice melting does not raise the sea level.
2. The authors used two words, forecasting/forecast and prediction/predict, in the manuscript. As the timescales are different between forecast and prediction, I recommend that authors use predict/prediction in the manuscript.
3. Line 13: The authors only gave the prediction results in advance 9 months for 2021, and they should clarify more accurately in the abstract in case of misleading readers. The authors can evaluate more cases for lead times as 9 months.
4. Line 15: has some physical interpretability -> has a physical interpretability
5. Line 37: Wei et al. (2021) -> (Wei et al., 2021)
6. The introduction is too long and redundant. For example, in Lines 49-54, it seems that the data assimilation is not close to the manuscript's key point and can brief these to one sentence. Lines 37-38, what’s the purpose of “For example, the average SIE…”. The authors should reorganize the introduction structure.
7. Lines 94-97: It seems there is a high overlap between contributions #1 and #2. It’d be better to merge them into one.
8. There is a missing part about the description of the data used in the study. It could be added after section 2.
9. Figure 2, it’s better if authors list the variables used in the IceTFT framework and give the output clearly (similar to illustration input SIE).
10. Line 158: 39.23°-90°N?
11. Figure 3: the variables in this figure do not match the variables in the IceTFT framework, such as SW, LW. Meanwhile, according to the authors discussed in section 5.7, I wonder if the results become better using the SW and LW instead of DSWRF, CSDSF, USWRF and DLWRF, CSDLF.
12. The NCEP-NCAR Reanalysis 1 was used in this work. I wonder if the results will change while changing the data to ERA5 or JRA-55. In other words, does the framework depend on the dataset?
13. Line 223: Evaluation -> Evaluation method
14. Tables 3 and 4, what does the percentage mean? The authors should describe the new statistics variable clearly in the manuscript.
15. Line 246: By using the short input length (6 months) leads to worse results. So, what if the input length increases to 18 or 24 months?
16. From figure 6, it can be seen that the biases are much larger in Sep than in winter or other seasons. So, it’s better that the authors can evaluate the IceTFT model's ability in different seasons, which may be more helpful for using the IceTFT model and understanding the sea-ice prediction ability. In fact, the prediction ability in summer (JJAS) is also more important than in other seasons.
17. Figure 8: as we know, SIO is the prediction results from June, July, and August. The authors should clarify what kind of prediction data of SIO used in this figure.

The study developed a deep-learning-based SIE prediction model and investigated the skills in SIE prediction. The SIE prediction model called IceTFT model shows higher prediction skill than Sea Ice Outlook. However, the overall structure of the ms is not very reasonable and rigorous. For example, the introduction should not overly emphasize the present study and main point, but focus on reviewing the methods and shortcomings of SIE predictions in the previous studies, then gradually introduce one's own work and emphasize the advantages of the current work. Overall, the present version of the ms is too coarse and there are a lot of grammar errors. The readability of the ms needs to be greatly improved. In term of these shortcomings, I do not recommend acceptance at the present version. Specific comments and suggestions are as follows.

1. Line 95, the contributions of this paper are not concise enough. For example, (1) is repeated with (2), (3) is a part of (4).
2. Add the reference about the definition of SIE in Line 150
3. The contemporaneous correlation in Table 1 is calculated between the global mean monthly one of eleven variables (e.g., SST) and SIE? And what’s the lag-correlation since we are more concerned about the prediction not simulation?
4. The metrics (MAE, RMSE, RMSD) have units. What’s the meaning of the percentages and those in parentheses in Table 3 and Table 4? Please describe the calculation.
5. Please give the detailed meaning of the x-axis in Figures 5,6. Time along the x-axis means the started month or target month? How to obtain Error and SIE along the y-axis?
6. Many figures and tables are not cited in ms, such as table 4, figure 5 and figure 6.
7. What’s the INITIAL experiment in Table 5? The INITIAL experiment seems to be described near Line 388 but the table 5 is first cited in Line 368. Similar issues exist in other figures.
8. What’s PRATE?
9. The authors mentioned that VSN in IceTFT can filter the spurious correlations (Line 137). In Section 5.7, the authors only analyze the variable sensitivity. Results in Tables 5 and 6 reflect some common key variables (such as SST and DSWRF), and also demonstrate some individual high-correlated variables (such as PRATE in 2019). From the perspective of statistics, can I consider that the individual high-correlated variables are belong to spurious correlations? Please add some explanations on such phenomenon under different ice cases from the perspective of dynamics.
10. From figure 8, the authors compare IceTFT with other models in SIPN. I focus that the others mainly belong to dynamical numerical models and there are no deep learning models. This also means the authors have not compared IceTFT with ordinary deep (machine) learning models. I don't have a clear quantification on how much improvement of TFT and others (such as MLP, LSTM). It is quite a consensus that a deeper (more complex) network is better, but I still want to know if this improvement matches the time and computation comsumption.
11. The authors construct the model based on the original TFT (Lim et al. 2019), please give a detailed parameter configuration, including the hidden layer number and hidden dimensions. Is the current configuration the best? Please give some discussions on the upper limit of IceTFT’s forecasting skill and why it surpasses other models.
12. Why a higher sensitivity value indicates that the variable makes significant contributions to predictions? I don’t understand why the sensitivity can be estimated by adding random noises. To investigate the contribution of different variables to SIE prediction in the model, one can artificially weaken the signal of the concerned variable and investigate the changes in prediction skills.

Other suggestion:

Line 35: “showed” => “shown, “… the errors were …” => “… that the errors were …”

Line 40: changed as “there is still a certain gap between these forecasts and observations”

Line 71: “recursive” => “a recursive”

Line 80: “lack” => “the lack”

Line 88: “that increases” => “which increases”

Line 141: “encorder” => “encoder”, this misspelled is also seen elsewhere

Line 141: delete “from that”

Line 149: “input” => “inputs”

Line 150: “in original TFT” => “in the original TFT”

Line 160: “in order to” => “to”

Line 177: “ SHUM” => ”and SHUM”

Line 285: “the SIE continued decline with steep slope” => “the SIE continued to decline with a steep slope”

Line 295: “As a results” => “As a result”

Line 332: “lead time” => “lead times”

Line 395: delete “has”

Line 396: “the more errors” => “more errors”

Line 403: “on 2021 than 2019” => “on 2021 than on 2019”

Line 450: “predict the next year SIE” => “predict the SIE in the next year”

Line 451: “the previous year data” => “the previous year’s data”

Line 456: “IceTFT model” => “the IceTFT model”. Missing “the” in many places of the ms.