General comments:

In this paper, the author built a statistical downscaling model based on (Convolutional Neural Network) CNN. And used the reanalysis data as a reference for the training phase. Then, the author employed the trained model to downscale the ensemble of CMIP5 models over Europe. The results were compared with the CORDEX RCMs for the historical (1975-2005) and the projection (2006-2100) period. The results from the deepSD algorithm are contributing to CORDEX initiative, which is a breakthrough since the statistical downscaling based on artificial intelligence was not trusted for climate studies in the past few years.  In general terms, the manuscript is well structured, the methodology is well described, the figures and tables are well organized and the results are adequately discussed. This paper falls within the scope of this journal. In this sense, the article can be approved after minor revisions.

Recommendation: Accept after minor revisions.

Specific comments:

Abstract:

The sentence “To our knowledge, this is the first time that CNNs have been used to produce multi-model ensembles” is not that accurate since there are previous studies that employed CNN to downscale the model ensemble (e.g., Babaousmail et al. (2021)).

Introduction:

1. Introduction (third paragraph): “These methods are not computationally demanding…” This sentence needs a citation.
2. The author should justify why he selected the RCP8.5 scenario out of the other scenarios? Also, was there any method employed for the selection of the 8 GCMs?
3. The author didn’t justify why E-OBS v20 was selected as an observation in this study.

Data and methods:

1. Since the author is comparing the ensemble resulting from 8 GCMs with the RCM ensemble projection, shouldn’t be the number of GCMs equal to the number of RCMs?
2. Usually, when we train a neural net model, a validation phase is required after the training and it should be selected from the historical 25 years period, in this paper the author didn’t mention it.
3. Concerning the CNN algorithm, we noticed that the CNN used to downscale precipitation has one more layer than the one for temperature (one output layer). Can the author explain the reason?

Please find the comments in the attachment.

 general comments

The paper is an important contribution to the methodology of the downscaling of the climate change simulations.  The results show that neural networks can bring meaningful regionalized climate change fields that can be a good complement to those obtained from regional climate models. The method works well for precipitation and temperature, but question of its usefulness for other climate fields, specially 3d fields, remains open.

The paper is clearly written and well structured. However, there are a few thinks that can be improved in the paper. In particular, I thing that the answer to this comments could improve the paper.

1. with which criteria were chosen the predictand fields?
2. how expensive in computer resources is the method?
3. Why eobs was used? It is too smooth, what can be seen in the results, specially in places with high topography.
4. It seems that the use of more output layers for the precipitation than in temperature makes the biases in the downscaling of precipitation as small as for temperature, but reduces the standard deviation in the downscaling (Figure 3) I think that this fact is related to the methodology and should be commented by the authors.
5. Also, the fact that the simulation of R01 in DeepESD is closer to the RCMs that to the GCMs shows the importance of a good simulation of orographic precipitation, while SDII and Mean temperature in DeepESD and GCM are closer, probably reflecting the tuning of the GCMs (which usually is not made in RCMs) and the training with observations in DeepESD. The exception for temperature in ED looks strange for me and would be nice if you explain this behavior.
6. Results in figures 3,4 and could be also contributed by the use of stochastic (deterministic) approaches for the precipitation (temperature)    

 specific comments

A more detailed description of the methodology for not specialists (most of readers, I guess) should be interesting. Can be added as an appendix

How does the interpolation method influences the results?

Why did not used a regional, high resolution reanalysis as predictands?

In the Iberian Peninsula and the Scandinavian peninsula the climate change signal in DeepESD is similar to that of the global models, while the opposite is true in central Europe. Could you elaborate on this?

Review of GMD-2022-57

Downscaling Multi-Model Climate Projection Ensembles with Deep Learning (DeepESD): Contribution to CORDEX EUR-44

By Baño-Medina et al.

In this study, the authors develop a downscaling method for climate variables based on the prefect prognosis approach using convolution neural networks (CNN) and evaluate how it extrapolates to unseen climatic states as projected by multi-model Earth system models. They focus on temperature and precipitation as well as on the European domain. The authors compare the output of their method to regional high-resolved climate model output and show that the CNN-approach reduces biases for the historical period and extrapolates to future climate change conditions in a plausible way.

The downscaling approach is very interesting and useful, especially the evaluation of its extrapolating skill to a different system state. Overall, the manuscript is well written. I have a few general comments and a short list of specific comments. Thus, I recommend minor revisions before publication.

1. Could you make flowchart of your workflow and include as a figure in the methods section? It is a bit hard to follow your exact procedure.
2. Isn‘t the comparison to observations between unconstrained mechanistic models (i.e. GCMs) and CNNs trained on observations "unfair"? If you did some nudging procedure with GCMs you would also end up with model output better fitting observations. For the CNN training, did you split the observational data into train (validation) and test set (only train on 20 years  and show performance for 10 years)? Again, a flowchart would help to understand what you did. If you show the performance of DeepESD for the test set and compare that to GCM output, it’d be “more fair”, but still, just by design we would expect that the CNN reproduces observations better than GCMs.
3. You show that the CNN learns the "necessary" dynamics based on predictors of the historical period and extrapolates reasonable well using predictors from GCM output for projections. That is a very interesting point. I wonder if this simple bias correction for GCMs really does the trick, as the models considerably diverge over the climatic time-scales and very model specific regional biases emerge. Can you comment on whether other bias-correcting measures were tested? Overall, there is certainly a long list of potential further evaluation and testing steps that could be undertaken, but maybe it is enough for this model description paper.

Specific Comments:

L5: What is DeepESD standing for? Please introduce acronym before first usage.

LL33-34: The “perfect prognosis” approach is based on the assumption that GCMs don’t have systematic biases with respect to the observations that were used for training, right? Maybe you should include a short sentence here that addresses this aspect.

L55: I recommend to use another more static hosting platform for your code, e.g. Zenedo (https://zenodo.org/).

L60: Why did you use ERA-Interim reanalysis? It is outdated for quite some time now.

L62: I don’t understand your use of dashes (—) in the manuscript. Please check whether the make sense throughout the manuscript.

LL62-65: What about adding high-resolution orography description as static predictor?

L85: Why did you analyze both and can you provide the reason why you settled with the deterministic one?

L88: Please stick to the tenses (in this paragraph you mix present and past tense), i.e. do not switch between present and past tense when describing your results or methods. I recommend that you always use present tense when talking about your study, i.e. when describing your methods, your results etc., and use past tense when referring to already published studies.

L137: “contribute to increasing”

Figure 1: Add unit at lower right colorbar. Also, it’d be useful if you could include letter characters as pointers to subplots, e.g. a,b,c,d. This comment applies for all Figures.

Figure 4: Please be more specific about the numbers in the plots. Please provide more detailed information in the caption.

Figure 5: The mid-column misses a time axis. DeepESD is not “yellow” but “green”, no?