This manuscript presents a deep learning-based statistical downscaling model for surface winds over complex terrain. The model consists of two parts, correction of winds from a regional atmospheric model, and conversion from a coarser grid to a finer grid, based on information of high-resolution topography data and atmospheric conditions. Results indicate that the proposed model better represent winds over Western Alps, for which the model is trained. The manuscript is generally well written, and conclusions are clear. In particular, analyses on the explainability increases its potential for real applications, where reliability of the model matters. There are, however, still room for improvement in the presentation quality. Therefore, my recommendation is publication after minor revisions are made.

Minor comments:

1. I understand that the model names, such as AROME and ARPS, are well known in the community, but it would be better to present their full names somewhere in the manuscript, such as the lines at which they appear for the first time or a list of names at the end of manuscript.
2. What is the source of high-resolution topography data?
3. In this manuscript, the data are divided into training and test datasets. However, in many practices, people divide data into three sets: training, validation, and test, where the validation dataset is used to tune hyperparameters. How did you tune the hyperparameters?
4. Figure 4 and 5: Labels are too small. Please enlarge them.
5. Figure 6: In the main text, the bottom panel (Fig. 6b) appears first, and the top panel (Fig. 6a) later. Because this is a bit confusing, I would suggest swapping the panels. In addition, I would suggest citing the figures in the main text as Fig. 6a and Fig. 6b by explicitly indicating the panel ID.
6. Figure 7: Please enlarge the labels. The lines are difficult to distinguish. You can use different line types, such as dashed or dotted.
7. Why does this study start downscaling at the forecast lead time of 6 h? I am curious how these models perform at shorter lead times. In other words, do you assume the same model error in AROME_forecast for the lead times from 6 to 29 hours?
8. Related to the question above, have you applied “Neural Network+DEVINE” to AROME_analysis? “AROME_analysis+Neural Network+DEVINE” may serve as a good analysis product.
9. Figure 9: The color bars for (d) and (e) have no labels for negative values.
10. Page 25, line 3: A link to a reference is broken.
11. Section 5.2: In this section, each paragraph looks very long. I would suggest splitting the paragraphs for readability.
12. Figure 10 and 11: The yellow shadings and lines are difficult to read.
13. Figure 10: ALE increases as the wind speed increases at 515, 126, and 10 m. In contrast, ALE is neutral for wind speed at 50 m, and ALE decreases as wind speed at 5 m increases. Do you have any interpretation on this behavior?
14. Page 26, line 558-: Although this paragraph states that input variables perhaps have large interactions, in the following sentence, it is also stated that input variables are not correlated, which is a bit confusing. Could you clarify this point?
15. There are many typos and grammatical errors. Please doublecheck.

The paper presents a deep learning-based strategy for downscaling wind fields over mountainous terrain. The innovation of this work lies in the versatility of the approach, achieved by separating the downscaling in two parts: i) Correcting large scale data provided by a NWP model that serves as input for ii) a statistical downscaling model that assumes perfect large scale input data. This decoupling ensures the downscaling model remains independent of the NWP models providing the input data. The paper is well structured, the method is mostly well presented and the results are discussed in a satisfactory manner. I only have minor comments of which many are mere suggestions.

• In general figures should be made a lot cleared with significantly larger font sizes (in particular for ticks) and thicker lines. Please carefully revise all Figures for readability.
• The wording is sometimes slightly peculiar. I would advice having a native English speaker look over the manuscript.
• P2 L28: NWP are --> NWP models are
• P2 L34: the literature --> literature
• P2 L35-37: It is not clear to me how Dupuy et al. (2021) and Goutham et al. (2021) relate to Zamo et al. (2016) and Hoehlein et al. (2020). What is the different way? I suggest to either mention what the difference is, or don’t distinguish them so clearly in the text.
• P2 L38: Not only these methods --> These methods not only
• P2 L50: the literature --> literature
• P3 L58: incorrect --> imperfect
• P3 L59: imprecise initial condition and errors due to the assimilation procedure may be redundant?
• P3 L59: Perhaps etc. is better than …
• P3 L59-63: These sentences arouse the expectation that the manuscript addresses uncertainty quantification of errors. However, I feel that the quantification of errors and uncertainty of errors is not clearly distinguished throughout the manuscript.
• P3 L65: lower --> smaller
• P3 L68: sufficiently the input wind --> the input wind sufficiently
• P3 L72: of DEVINE --> of the DEVINE
• P3 L72: I don’t understand how the effect of downscaling is accounted for.
• P3 L76: Again, I would rather use etc than …
• P4 L103: Why discard forecast lead times 1-6? Aren’t these very relevant for various applications? Is it because short term forecasts are good enough and would not benefit as much from a corrective step? Please shortly motivate your choice.
• P4 L108: I believe “dispose of” is misused throughout the manuscript. It means “getting rid of”.
• P4 L114: introduce acronym AWS
• P4 L120: I suggest to delete the sentence where you mention you use some stations for training and others for evaluations here. It is well explained later and a bit confusing to state it here.
• Caption Figure 1: observation --> observations
• Caption Figure 1: Add (not shown) after “Note that an additions data set”
• P7 L168: time --> times
• P7 L171: How did you choose the hyperparameters?
• P7 L185: The term “label” might be confusing to some, as it has not been introduced before. I suggest reducing the number of synonymous terms throughout the manuscript. For example, choose either labels, targets or ground truth and use it consistently.
• P7 L191-195: I don’t think it adds value to add technical information here. This information is relevant for those who are interested in the code and should therefore be provided with the code publication.
• P10 L226: I don’t understand this sentence. I believe you are trying to explain how you deal with the mismatch between the location of the observation and that of the model grid points, but I don’t understand it. Please try to clarify.
• P10 L229: This is related to my confusion on how the effect of downscaling is accounted for (see comment on P3 L72). How exactly is DEVINE involved in the neural network for the corrective step? Also, is the “observed ground truth” observations here?
• P1- L235: Perhaps emphasize again that the motivation for not wanting to create a new downscaling model is versatility.
• P10 L245: to produce squeezed --> to produce a squeezed
• Figure 2: In my opinion contains too much details on DEVINE, both in the caption and the figure. Instead, the neural networks could be highlighted more. For example, I was a little confused on how exactly the skip layers are incorporated. For example, I was confused whether the network outputs a correction of the full wind speed/direction fields.
• P13 L308: steps --> step
• P14 first paragraph: If TPI and elevation are strongly correlated, it makes no sense to compare their respective influences. Please compute the correlation, and, if it is high, I suggest to state that it is not possible to identify any distinction between the two in terms of influence on the result. You do note this to some extent, but this should be emphasized and could reduce the discussion of this paragraph significantly and rendering the corresponding figure moot.
• Figure 5: I am unfamiliar with wind roses (though I expect most readers won’t be). Does the size of the spoke indicate the frequency with which the direction in the correspond bin occurs for all panels? So only the color coding is different than the rest for panel d (observations)?  
• P15 L350: Please define RMSE and MAE (not just in figure/table captions)
• P16 L366: reinforce previous --> reinforce a previous
• P16 L369: elevations --> elevation
• P16 L375: (Fig 5) However --> (Fig 5). However
• P20 L426: observation --> observations
• I really like paragraph 4.5. It is indeed emphasizing the importance of separating the corrective step from the downscaling step!
• P22 L457: modification --> modifications
• P22 L458: Related to comments on P3 L72 and P10 L229, I don’t understand how the network is aware of anything related to DEVINE.
• P22 L467: application --> applications
• P23 L481: I don’t understand this sentence. What challenges? Do you mean the evaluation is more critical? If yes, please rephrase. Otherwise, please explain.
• Figure 10: Do I understand correctly that ALE are approximated local gradients around the value of interest, averaged over time and space? And the shaded region represents to corresponding standard deviation? Are the solid lines (the means) also accumulated?
• P25 L524: I don’t see any indication of skip connections in Fig2. Perhaps it is because the lines are too thin, or the figure too small (also see previous comment about Fig 2).
• P25 L528: What exactly is the difference between real elevation and model elevation?
• P25 L535: Again, I am confused by this. How had neural network seen downscaled simulations?
• P27 L589: I would delete the i) and ii). At least for me it was only confusing.
• P28 L595: “best results” should be replaced with something like “most improvements”.