The authors present a method for model validation in terms of extreme values based on classical extreme value theory, they further discuss a clustering algorithm to detect spacial dependencies and tendencies for concurrent extremes.  CTU TS and AWI-ESM model output are used to tested the statistical approaches. The manuscript was writen well, and the results are interesting. I suggest to accept it after minor reviseion.

1) The caption of Fig 1 does not agree with the figs.

2) Since only one model are used in this manuscript, I would suggest the authors choose a model that agree most to the observations. What you need to do is to compare the performance among all available models first, and then apply the methods to that model output.  This would more interesting. 

The authors present a method for climate model validation which focuses on extremes rather than on classical measures such as means and covariances. Second, they apply a clustering algorithm to detect regions in which extremes occur simultaneously. The paper is well written, the authors clearly expose the different steps and methods used in their approach. However, I have some methodological remarks, mainly concerning the use of extreme-value theory. 

General comments:

• The introduction does not seem to cite other papers that study model validation focused on extremes, yet I think it is already an active research topic. For an example, see for instance 

Timmermans, B., Wehner, M., Cooley, D., O’Brien, T., & Krishnan, H. (2019). An evaluation of the consistency of extremes in gridded precipitation data sets. Climate dynamics, 52(11), 6651-6670.

• Page 2, lines 35-36: please mention that there are two popular approaches, block maxima and peaks-over-thresholds. 
• Page 4, lines 110-111: I would remove ``especially for the yearly maximum of daily average precipitation’’ since the GEV has been used extensively for many types of data.
• Page 5, lines 131-132: this sentence is rather unclear, could you maybe reformulate?
• Page 5, lines 145-146: please note that the lower bound of the extremal index corresponds to perfect dependence (comonotonicity), which is more general than a Pearson correlation of 1 (a Pearson correlation of 1 implies comonotonicity but not vice versa). 
• Page 6, equation (8): does it make sense to « weight » the three marginal parameters equally? Maybe the shape parameter could play a bigger role than the mean and scale parameters?
• Page 9, line 210: it is not necessary to use bootstrap-based confidence intervals for the PWM estimators, since their asymptotic covariance is known and has a simple expression; see 

Hosking, J. R. and J. R. Wallis (1987). Parameter and quantile estimation for the generalized Pareto distribution. Technometrics 29(3), 339–349.

Ribereau, P., P. Naveau, and A. Guillou (2011). A note of caution when interpreting parameters of the distribution of excesses. Advances in Water Resources 34(10), 1215– 1221.

• Page 13, Fig 7: the graphs are difficult to compare because of the many clusters and colors. Could you please resume the main differences in the text?
• Page 13, line 250: ``While parametric copula families are applicable only to a very limited extent in high dimensions…’’ I disagree, there are many possibilities to model high-dimensional data using parametric copulas, for example, through vine copulas.
• Page 14, line 259: aren’t the spatially stationary method well suited to model the clusters as identified previously?

Typographical errors:

• Page 4, line 114: resultung -> resulting
• Page 9, line 224: Fig 4 should be Fig 7?