This study looks at the change of circulation regimes under climate change for an Arctic and Atlantic-Eurasian domain. It finds that the NAO+/AO- regimes will occur more often in the future, while the NAO-/AO+ regimes will occur less.

Overall, the study is a valuable addition to the scientific literature on weather regimes, presenting important results for the Arctic. However, the description and discussion of the methods and results needs to be substantially improved before I would find this ready for publication.

I have put comments in the pdf and summarise my main points here.

My major concerns are:

1. The authors use pseudo-PC’s to identify and assign regimes for the CMIP6 model data. While I think this is the most suitable approach for studying changes in regime occurrence, I question whether it is also suitable to study changes in the regime patterns. By projecting the CMIP6 data onto the ERA5 PC’s you introduce a bias of sorts and I wonder how different the results would be when using the model’s own PC’s?
2. Also the other choices made in the method section are not in all case thoroughly argued for, for example the use of 10 PC’s and not e.g. 15. I believe the paper would benefit significantly from rewriting and clarifying the arguments behind the choices made.
3. The authors study both a projected and simulated approach for k-means, but not for SANDRA. Could you clarify why? A similar question for arguing toward the best number of regimes.
4. In the results the authors argue for just using the projected approach because it is “more accurate”, where to me it appears the smaller spread between models is by design. By that I mean that the using each models own regimes by definition will create more variability in the frequency, as now also the regime patterns differ between the models. Can the authors elaborate on this and argue for their choice?
5. The authors repeatedly mention the patterns being similar between the two methods considered (KME and SAN), where looking at the figures I do not think the differences are non-significant for all regimes. This is also indicated by the different response to climate change of some regimes. I do not have an intuition about how to interpret the SANDRA regimes with respect to those of k-means. It would be helpful if the authors could elaborate on the difference between the methods and its impact on the regime interpretation.
6. Figure 4. I do not know what is shown in this figure. Which regimes are compared to the reference? I would say each model has their own, whereas the authors refer to the common EOFs of CMIP6 models. I could not find anything in the methods section discussing this. I think it would be valuable to show the spread between models in a figure like this, also finding which models have a regime representation that is closer to reality.
7. The authors only look at changes in the regime occurrence. It would be valuable to also study changes in the persistence of the regimes, to get a more dynamical understanding of the effects of climate change, e.g. is a regime becoming more frequent because it becomes more persistent? I suggest adding results on persistence.
8. Many of the studies references for comparison study different domains, e.g. Europe in Boé (2019). It would be good to acknowledge that in the text, as it can impact the results. The same for studies used in comparison that study the winter season. Furthermore, I suggest adding some references on the different number of regimes: https://www.nature.com/articles/nclimate3338 and https://rmets.onlinelibrary.wiley.com/doi/pdfdirect/10.1002/qj.3818.
9. I do not see the added value of discussing storylines for this paper. Since each storyline only contains two models, it is near-impossible to draw any robust conclusions. Can the authors clarify why they chose to also include the storylines in this work?
10. The conclusion and outlook section could benefit a lot from placing the work in the context of the wider literature. Are the finding in line with other studies, with what we expect linked to jet stream shifts, … As is, the outlook part is very brief and lacks context.

If I were convinced that indeed the projected approach is valid in this case, the article is very well presented and presents a significant contribution to the field. The authors provide an explanaition to why the cluster frequency changes under one storyline and the other, which is a significant contrubution to understanding the impacts of large scale ciruclation changes and the model uncertainty.  However, the article would highly benefit from carefully confirming that the regimes for which the frequencies are studied are indeed representative of how the circulation is organized in models and in the future. I recommend major revisions but I highly encourage the authors to address and resubmit.  

Major comments:

1. The projected approach to weather regimes is an established methodology. However, the paper would be a much more valuable contribution if the authors were able to provide a more robust validation to justify the approach. 

- The pseudo-PCs are already a projection of the data onto the EOFs from ERA5, which is already an great assumption. I would suggest that the projected approach was compared with the complete clustering approach performed with the models.  To my understanding, only if the circulation in the model can be considered as representative of the reference, the two can be comparable. Formally, this means that the model sub-space is contained in the reference space. One way of quantifying this is the "quantization error" proposed by Quagraine et al. (2020). 

- Currently the evaluation is performed using Taylor diagrams and the statistical significance of the changes in the frequency of occurrence under the influence of GHG is performed with a Welch's test. However, these metrics are valid once the above has been tested (i.e. that the model subspace is contained in the reference subspace). 

2. Storylines:

- Why didn't you consider the strong ArcAmp and strong Barents-Kara-Sea warming storyline and the opposite?  I can appreciate from Lavine et al. that depending on the target variable, the change under each storyline is different, why use these two only?

- Can you show that the circulation in the models undergoes the circulation changes that you consider relevant to your regime changes?

- I think that the assumptions behing using the storylines should be justified. Why can the authors expect to find different regimes of regime frequency under these two storylines? 

3. Different clustering algorithms. I would move the evaluation of regimes using SAN to the Supporting Information or further exploring  the results that are obtained when  using SAN for the complete analysis. The authors claim that these is not much different between one method or the other, however, I find this questionable: First, the correlations in Figure 3 are not impressive and second, the SCAN and DIPOLE regimes are very different from a dynamical perspective. Low slp and high slp over the British Isles respectively can lead to very different weather conditions. What is relevant related to the weather regimes are the regional impacts. At least this is what is claimed in the introduction. Can you show that the precipitaiton and temperature impacts associated with these regimes are comparable?

4. I am not convinced by the claim "Thus, the spatial structure of the summer circulation regimes does not change significantly under the influence of rising GHG emission in the future time period compared to the historical time period". Referring to my comment above, I would consider that evaluating the regimes independently for the future and the historical period would be a first step to ensure that projecting the forced simulations is a valid assumption. If this is the case, then one can claim that the spatial structure of the regimes does not change in the models. I also consider that this is not a "weak external forcing". Given that SSP5-8.5 is used and the authors refer to this as a strong forcing in a previous pragraph.

Line 280: This first sentence is not clear. Do you mean that "In the models representing the BK+ and PA- storylines the NAO+ occurs significantly more?

Line 295: The immediate above paragraphs were describing the results in Figure 6, while the paragraph starting in l288 refers to the KAN and SAN comparison. I don't follow the logic. 

Caption Figure 6: The markers associated with the storylines are not escribed.

Overall the silhouette and elbow results are not impressive, 6 or 7 could have been a perfectly reasonable number of clusters. Line 440 says "The KElbow Visualizer suggests that five clusters represent an optimal number of clusters for both regions." No description of what is seen in the figure is provided. Why is this suggested by the figure?