In this manuscript, deep learning methodology based on CNN architectures is evaluated for downscaling CMIP6 simulations and projection with standard resolutions to a higher resolution, i.e. 1/10th deg. The analysis focuses on the Iberian Peninsula, and the ability of four CNN architectures to reproduce the surface T, Tmin, Tmax and Pr climates is evaluated. The DL algorithm is trained using ERA5 and compared to the high-resolution regular gridded dataset over the historical period, and then used to downscale future projections (relative to historical climate) in agreement with four future scenarios and multi-models. 

Overall, the manuscript is well organized in format and the writing is clear. Although the physical process is rarely touched (due to the limitation of DL), the downscaled results based on DL are sound. I have some comments regarding clarifications in the main text. 

Line 23: “Notably, …. climate scenario”. Please specify the region (Iberia) exhibiting the temperature increase. 

 Line 129: I am not familiar with any paper using “democratic” to describe a simple average. It would be better to just use “simple average”, which is straightforward and easy to follow. I noticed that “democratic” has been mentioned several times in this manuscript. Please also revise those instances accordingly. 

Line 145 (Fig. 1) Can you mention why this domain of predictors is chosen (dashed line)? Have you tested the sensitivity of the downscaled fields when the domain of predicators is changed?

Lines 170-173: These sentences would be better placed in the introduction. 

Table 1: It would be better to provide the full name of CMIP6 ESGCM. For example, CM6A-LR -> IPSL-CM6A-LR.

Lines 220-222: I am not able to follow this sentence. Could you rephrase it? 

Lines 222-223: How are these terms (alphas, beta, and distribution) used in this study? I couldn’t find more discussions about these terms from CMIP6 runs. 

Line 238: How were the missing data filled? Were they filled with zeros?

Line 276: How about changing “members” to “models”, as the members are also used to describe 4 DL architectures (Line 384)? Similarly, in Line 48, 7 members -> 7 models. 

Line 283: Why was the base period chosen as 1981-2010 and not 1979-2014? The projected temperature increase depends on the reference period, which should be mentioned clearly in the corresponding section. 

Fig. 2 and subsequent figures: I am not sure how the error bar of each boxplot is calculated. Is it related to the uncertainty of parameters in the DL model?

The error bar in Fig. 10 and the following figures seem to have different meanings compared with the previous figures, I guess. Are these related to the spread from 4 CNN methods and 7 models?

Soares et al. build on previous work on CNN downscaling to compare the performance of 4 different deep learning architectures on the Iberian Peninsula’s temperature and precipitation. The study is unique in using the high-resolution observational product for Iberia and in its findings that the predictive skill was not sensitive to architecture complexity. Overall the paper was clear. Some aspects of the methodology were hard to keep track of. My comments are mostly minor. 

Comments

A schematic would be helpful to show the main aspects of the methodology in terms of datasets used to train the model and datasets which are downscaled with the model.

Do you use the same DL models that were trained on ERA5 and Iberia01 on ESGCMs or do you generate 28 unique DL models that represent the relationships between each ESGCM and Iberia01? If the former, can you discuss the assumption that the relationship between coarse and fine scales is the same between ERA5 and Iberia01 as it is between ESGCM and Iberia01 (maybe in the discussion section)? I think your figures show that this assumption is valid enough. If the latter, I couldn’t find that detail in the text.

I see from the comments that the DL data is to be added; are the parameters provided as well as the output?

L240 “The ESGCMs were bias corrected in relation to ERA5 through a simple mean-variance scaling method.” How much does this affect the work on extremes, i.e., how does the variance differ?

L270 Is it reasonable to assume that these DL models trained on the full period would have similar attributes/predictive skill as the ones trained on shorter period?

L556 I’m not sure why you call this the climate change signal since this appears to be just the anomaly relative to some historical baseline. Are you making the case that the average of 7 members is enough to remove internal variability and give only climate signal? If so, that should be stated.

L565 It's difficult for me to make out the purple dots in contrast with the background. These figures need to be provided at higher dpi.

L565 Can you address whether the DL architecture is producing excessively noisy results? I don’t know what to make of this figure.

Minor comments

L20 Can you explain what you mean by "overall agreement"?

L30 “The extensive results presented are” awkward phrasing

L34 “projects worrying changes in what concerts” awkward phrasing

L36 “view of those changes being especially intense” awkward phrasing

L39 “Such disadvantage fosters” awkward phrasing

L61 unclear what “it” refers to

L64 “promising ones” >> “promising”

L64 delete “ which, in turn, is a subdomain of AI” this feels like unnecessary context

L65 “In ML…automatically” This doesn't really make sense to me. There is a training algorithm, and training data, so the models aren't training by themselves.

L73 “an artificial analogous to” >> “an artificial analog to”

L79 “there have been early” >> “Early”

L80 Delete “but the results” and “enough”

L88 “reduces for precipitation” reduces what?

L89 “are perceived with precaution” awkward phrasing

L91 “There have been attempts…” false parallelism here

L95 “DL also presents…” Break this into several sentences

L103 “to strong impacts on the” >> “to high”

L104 “stronger” relative to other seasons or relative to other regions?

L106 “Also, it is projected a significant reduction” make active voice. “Along” >> “throughout”

L107 “aligned” >> “Concomittant”

L121 Delete “and tested”, which is unclear at this point in the paper

L131 “liability” >> “shortcomings” or similar. Unclear what “liability” means in this context

L161 “understandably”?

L162 “what concerns to predictors data” awkward phrasing

L230 We need an additional column for the activation function type to make clear the difference between BMlinear and BM1

L236 “was” >> “were”

L271 “downscale the CMIP6 ESGCMs” here it is ambiguous whether you downscale them together or individually

L306 “slightly improvements” >> “slight improvements” Compared with Iberia01?

L308 “(3) the four…” you mean they don't quite capture extremes, right? There’s probably a more straightforward way to say this

L440 I find the labeling of these plots confusing because only the colored bars are called by the ESGCM's name but it seems like the grey bars also are representing those individual ESGCM outputs

L448 “7 members” here it is unclear that these are referring to different CMIP6 models.

L591 “Therefore, we opted to consider…” this feels more like a methods statement

L647 In what sense are they consistent? Consistent with one another in terms of 2/3 agreement?