Overall comments:

The study presents a k-means cluster analysis of the recent ACCESS model in the spirit of (Williams and Webb 2009) and (Haynes et al. 2011). The paper could do more to concretely link biases in cloud properties to biases in radiation- radiative biases are suggested to be related to biases in various cloud properties, but this appears to be by eye rather than quantitative. In several places the writing is difficult to follow and especially in the analysis section it is hard to follow whether cloud RFO or cloud properties are being referred to (in several cases clouds are referred to as being simulated correctly or incorrectly, but it is unclear what that means in the regime framework) and often the ability of the model to replicate these quantities is described in vague relative terms.  It is also not clear if the authors are comparing in-cloud and area-averaged water paths.

Abstract:

In several places the authors refer to incorrectly or correctly simulating clouds. It is ambiguous what they mean by this. It  seems to be only referring to phase and frequency (which I think it equivalent to cloud fraction). If this is the case, it would be good to clarify that  we only  care  about phase and  frequency in the abstract and not other things like optical depth and condensed water path (for instance).

L44: what ensemble is being referred to?     

L60: What aspect of Bodas-Salcedo 2014 demonstrates a need for consistent evaluation techniques?

L62: it is unclear what the first two sentences of this paragraph are referring to. What are climate-scale runs? Why wouldn’t the synoptic meteorology be the same? I think what the authors are getting at is the difference between coupled and AMIP runs.  However, there is not any guarantee that the synoptic state will be the same across AMIP runs and the authors just discussed Field and Wood 2007 above, which composites on synoptic state- making it immaterial whether low pressure centers and other synoptic features are occurring in the same place.

L75: It is somewhat vague what the authors mean by ‘incorrectly’ or ‘correctly’ simulated… Is this just in terms of phase and frequency, or in terms of all characteristics in a more abstract way?

L175- It’s a little ambiguous here whether the authors are referring to IWP and LWP averaged over the grid box, which is what the model outputs (aka clivi and clwvi-clivi), or if they are talking about in-cloud liquid and ice water path, which is what MODIS would see. It is also somewhat mysterious how propagating errors would affect LWP and IWP and not other cloud properties in COSP. Some additional discussion of this is needed to instill confidence in their evaluation.

L193: Consistent with which previous studies?

L235 and 245: Is CFL/CFI random overlap, or just what is seen from space? Could biases be driven mostly by this cirrus in the model if it is just what is seen from space, with minimal relevance for the PBL cloud that drives SWCRE?

L249: Consider citing:

Mülmenstädt, J., Salzmann, M., Kay, J. E., Zelinka, M. D., Ma, P.-L., Nam, C., et al. (2021). An underestimated negative cloud feedback from cloud lifetime changes. Nature Climate Change, 11(6), 508–513.

Field, P. R., & Heymsfield, A. J. (2015). Importance of snow to global precipitation. Geophysical Research Letters, 42(21), 9512–9520. https://doi.org/10.1002/2015GL065497

L253: again, it is unclear if the authors are comparing in-cloud LWP and IWP to area-mean LWP and IWP.

L273: This discussion is fairly qualitative in terms of relating various cloud properties to radiative bias. Quantitative estimates of how (for instance) cloud fraction relates to radiation exist:

Bender, F. A. M., Engström, A., Wood, R., & Charlson, R. J. (2017). Evaluation of Hemispheric Asymmetries in Marine Cloud Radiative Properties. Journal of Climate, 30(11), 4131–4147. https://doi.org/10.1175/JCLI-D-16-0263.1

Can the authors show whether the CF bias in their simulations can explain the actual radiative bias?

Section 5.2.1: this section and the associated figure 7 are quite hard to follow. The authors may benefit from more clearly distinguishing errors in RFO and in cloud properties for a given cluster. The writing is somewhat ambiguous- clouds are referred to as being ‘correctly’ simulated- is this in terms of getting enough of them, or in terms of them looking right when they show up? In particular, the second paragraph of this section could be improved by using fewer vague qualifiers (‘comparatively well captured’, ‘relatively well captured’, ‘somewhat correct’,…) what is the baseline for these statements? These statements are then used to make causal statements about what biases in clouds are leading to biases in radiation, but without any support – wouldn’t it be possible to do a more quantitative assessment of where biases in SWCRE are coming from?

Figure 7 is pretty hard to follow. The authors may need a cartoon with annotations or something to illustrate this. There are dots, colors, outlines, months, clusters, and 5 different quantities. A single cartoon for one of the subplots would be helpful.

Overall, I would suggest moving the summary of the findings before section 5.2.1 to give the reader an overview of what is going to be discussed.

L555: the authors bring up a good point- is some of their RFO bias simply due to nudging? Can the k-means clustering be replicated on a free-running simulation to see what the biases look like?

This study successfully extends an increasingly popular tool for model analysis (cloud cluster analysis) in several innovative ways. The most important development is the introduction of an approach for comparing the performance of a model when cloud type is correctly simulated to when it is not. This is made possible through the usage of nudged model input and provides previously undiscovered information on the quality of model simulation of these cloud types. While limited to one model, this study provides a pathway to determining the answer to a fundamental research question in this field (“If models can accurately simulate cloud types, what will the impact on the biases in other cloud variables be?”). There are some places within the manuscript where the text could be updated to be more clear or concise and a few methodological queries that I feel need to be addressed. Otherwise, I believe this is an outstanding paper.

Major Comments:

In the current manuscript the normalisation of the cloud top pressure - cloud optical thickness joint histograms is ambiguous. For the majority of prior papers using these histograms, they are normalised by the cloud fraction value so that the sum of the cells of the histograms add to the cloud fraction value. If I had to guess based on figure 2, each of the histograms in this study are normalised to a value of 1 (I could be wrong about this). If this is the case it will have some implications on the interpretation of the results with respect to prior studies. This alternative normalisation could be justified by arguing the paper's focus is on phase and vertical structure which may be better captured with this approach, however currently I cannot find any discussion of this in the paper and it should definitely be discussed. While not expected in this paper, it could be interesting to compare results of these different normalisations.

I am surprised by the lack of discussion of supercooled liquid water throughout the manuscript. There are several places, particularly in your results, where I think some discussion is warranted. Many of your results show too much ice fraction and not enough liquid in key cloud types which is indicative to me of issues with the model representation of supercooled liquid water. Some good places to add this would be the paragraph starting on line 273, the discussion of figure 7-9 and the conclusions.

There are some figures that are passingly discussed in the text but not currently in the paper which would make great additions to an appendix. In particular, I am thinking about the phase property versions of Figure 5 and the individual sub-region versions of Figure 6. I know I would be interested in seeing those figures and the most appropriate place via ACP guidelines appears to be in an appendix.

Minor Comments/Typos/Suggested Text Changes:

Line 7: Sentence starting on this line should be simplified due to its complex clausal structure.

Line 20: Consider changing “simulation by models of cloud properties” to "simulation of cloud properties within models"

Line 46: Consider changing “compensate the” to “compensate for the”

Line 84: I think something has gone wrong with the citation formatting here

Line 153: Can you please be more specific about the identification of clear sky cases and their removal from the dataset

Line 190: I think you have cited the wrong paper by mistake here. From what I can see Pendregosa 2011 does not have that information. 

Line 201: Sentence starting on this line is a little clunky. Perhaps change to something along the lines of them being defined globally and only examined in detail over the Southern Ocean

Section 3 could possibly do with some references to past paper which have identified similar seasonal (Bodas-Salcedo et al. 2012) or spatial (eg. Kuma et al. 2020) biases

Line 225: Consider changing “less zonal” to “less zonally coherent”

Figure 1 appears to have an issue where some rows are more magnified than others leading to some straight edges

Line 261: I think “indicating that this is a complex system to understand” should be changed to something that stresses the behaviour in the system is complicated inplace of how understandable it is.

Line 282: Simplify the wording of “using 12 clusters for five years of daily-mean joint histograms”

Line 288: Consider deleting “while this is important to note,”

Figure 3 and 4: The last sentence in the caption is a little unclear and awkwardly worded. Consider simplifying.

Line 327: Consider deleting “region of interest”

Line 328: Consider rewording “are spatially consistent in sign and for some, magnitude” as it is a little confusing.

Line 333: The sentence stating on this line may need some rewording as it could imply you do not look at subregions instead of the intended meaning that non-SO data is excluded.

Line 371: I found the wording of the first three sentences of this paragraph quite hard to follow. I would suggest rewriting them with an emphasis on clarity.

Section 5.2.1 even though it is incredibly rare, a sentence discussing the extreme phase based biases shown in the TC class could be a valuable addition here.

Line 401: The transition to a new sentence discussing mid-level clouds feels off, because I assume the previous sentences were already discussing them. I think this could probably just be resolve with some more clear wording

For figures 7-9 I think the exact definition of the total column is unclear. I can’t determine if it shows the overall biases associated with accurately simulated clouds or with all clouds. If it is showing for both “incorrectly” and “correctly” represented cloud types combined, it could be interesting if there was a way to decompose this information and show each independently. Regardless of if this is possible or not, some text needs to be added clarifying what this column represents.

Line 446: The sentence starting on this line is confusing due to its clausal structure. Can you simplify the sentence or split it into two.

Line 449: Consider deleting the “For” at the start of the sentence.

Line 454: Please simplify “Considering now what may be contributing to”

Line 531: I think the sentence starting on this line is a little confusing and would benefit from being rewritten.

Line 533: Consider deleting “biases, again whether the cloud regime is correctly simulated or assigned as something else,”