This is an interesting and well-presented analysis of aerosol-cloud adjustments in a convection-permitting model over Sc decks. The analysis technique (causal graphs) is novel and is used to compare to observations. The paper is well-written. I have a few concerns that are detailed below related to the appropriateness of comparing a convection-permitting model with fixed cloud condensation nuclei to observations using this particular framework. It is my opinion that these concerns can be handled by adding caveats.

If the authors find it appropriate and not excessively time consuming, they may find that the robustness of their analysis and their suggestion that there is a strong benefit to high resolution would be aided by including analysis of a lower resolution version of the same model with interactive aerosol and with fixed CCN so that they provide examples to argue (i) that their assumption of fixed CCN is not affecting their results and (ii) that the high resolution model actually improves model performance. In particular, constant CCN makes it a bit hard to interpret the causal graphs used in this analysis. To use such a high resolution model some parts of the simulation need to be sacrificed, but the ability of the aerosol, cloud, and precipitation to couple is key to reproducing observed present-day covariability and fairly compare observations and models (Stevens and Feingold 2009; McCoy et al. 2020; Wood et al. 2012; Gryspeerdt et al. 2019). Is there a way to modify the causal graph if one arrow linking precipitation, CCN, and cloud is disabled in the model, but exists in the observations?

Ln5: Observing = inferring if it is based on correlation. If there is a transient change in aerosol it may be possible to say there is an observe a causal response. On ln 39 the cited studies are correlative, rather than causal.

Ln 5: While high resolution models are probably generally good since fewer processes need to be resolved, we cannot provide a climate model that resolves the micro-scale processes that lead to precipitation suppression and entrainment thinning.

Ln 35: Also (Wall et al. 2022).

Ln 40: While I generally agree that models probably overpredict adjustments due to minimal representation of size-dependent entrainment (Karset et al. 2020). Please note that the cited papers here are correlative rather causal. As discussed later in the intro (Christensen et al. 2022) there are opportunities for dealing with causal ambiguity in observations of aerosol-cloud interactions. Also note that in (McCoy et al. 2020) both models and observations imply a negative adjustment from aerosol-cloud interactions in the present day based on similar techniques to (Gryspeerdt et al. 2019) even though the model response to anthropogenic aerosols is an increase in LWP.

Ln 44-59: I agree with all of this, but I think it would be good to qualitatively state what the range of length scales of interest are in Sc clouds. They are still definitely sub-km (Wood 2012).

Ln 65: While fixed CCN is useful especially in high resolution models where running full aerosol chemistry would be expensive, it is somewhat unphysical since scavenging can’t remove any CCN. This is an aspect of present day covariability between clouds, precipitation, and aerosol that seems like it would be needed when comparing between observations and the model output. Discussions of the source of covariability that goes cloud->precipitation->aerosol see discussion in (Gryspeerdt et al. 2019; McCoy et al. 2020; Wood et al. 2012). It would be interesting to contrast the analysis in this study with a lower resolution version of the same model with interactive aerosol and without interactive aerosol.

Ln70: One confounder is that aerosol sources tend to be on land, so there is spurious correlation between lower LWP and aerosol (Wood et al. 2012; Gryspeerdt et al. 2019). I am not sure nudging really impacts that all that much and since you have set CCN to a constant through the atmosphere this source of covariance is removed. Another strong source is that thicker clouds tend to rain more- reducing Nd and creating covariance that goes from clouds to precipitation to aerosol (McCoy et al. 2020; Wood et al. 2012; Gryspeerdt et al. 2019). This is also artificially turned off if CCN is set to a constant.

Ln 90: as the authors note- the simulations are not nudged, so it is hard to know how to compare these simulation by eye in the absence of comparisons of the ICON meteorology to observations/reanalysis.

Ln 100: Agree- ERA5 is just running reanalysis atmospheric structure through a parameterization- albeit one where aerosol-cloud interactions are not being represented. It is good that the authors explicitly point out that ERA5 clouds are not reanalysis in the same way that things like near surface temperature are.

Ln 108: Why isn’t it also explainable by low CCN? Nd is a function of CCN and updraft and the constant CCN could be unrealistically low.

Figure 4: Would be good to note that ERA5 RWC and LWC are the reanalysis thermodynamic fields run through a parameterization.

Line 171: Perhaps cite (Karset et al. 2020)

Figure 7 and Table 1: I am not familiar with the assumptions underlying causal graphs and it would help readers such as myself to understand and interpret these results if there was some discussion of how turning off arrows in the model like CCN being able to interact or size-dependent entrainment that we respectively know and suspect exist in the real world affects the causal graph. This seems like a broad discussion of comparing models, which will always be somewhat structurally incomplete to reality.

Section 4: This section does a good job summarizing the results of this analysis.

Christensen, M. W., and Coauthors, 2022: Opportunistic experiments to constrain  aerosol effective radiative forcing. Atmos. Chem. Phys., 22, 641–674, https://doi.org/10.5194/acp-22-641-2022.

Gryspeerdt, E., and Coauthors, 2019: Constraining the aerosol influence on cloud liquid water path. Atmos. Chem. Phys., 19, 5331–5347, https://doi.org/10.5194/acp-19-5331-2019.

Karset, I. H. H., A. Gettelman, T. Storelvmo, K. Alterskjær, and T. K. Berntsen, 2020: Exploring impacts of size‐dependent evaporation and entrainment in a global model. Journal of Geophysical Research: Atmospheres, 125, e2019JD031817.

McCoy, D. T., P. Field, H. Gordon, G. S. Elsaesser, and D. P. Grosvenor, 2020: Untangling causality in midlatitude aerosol–cloud adjustments. Atmos. Chem. Phys., 20, 4085–4103, https://doi.org/10.5194/acp-20-4085-2020.

Stevens, B., and G. Feingold, 2009: Untangling aerosol effects on clouds and precipitation in a buffered system. Nature, 461, 607–613.

Wall, C. J., J. R. Norris, A. Possner, D. T. McCoy, I. L. McCoy, and N. J. Lutsko, 2022: Assessing effective radiative forcing from aerosol–cloud interactions over the global ocean. Proceedings of the National Academy of Sciences, 119, e2210481119, https://doi.org/10.1073/pnas.2210481119.

Wood, R., 2012: Stratocumulus Clouds. Mon. Weather Rev., 140, 2373–2423, https://doi.org/10.1175/MWR-D-11-00121.1.

Wood, R., D. Leon, M. Lebsock, J. Snider, and A. D. Clarke, 2012: Precipitation driving of droplet concentration variability in marine low clouds. J Geophys Res-Atmos, 117, n/a-n/a, https://doi.org/10.1029/2012jd018305.

Review of "Investigating the sign of stratocumulus adjustments to aerosols in the global storm-resolving model ICON" by Fons, Naumann, Neubauer, Lang and Lohmann, ACP Manuscript egusphere-2024-195.

Summary

The authors present a statistical analysis of the drivers of aerosol-cloud interactions in marine boundary layer clouds in satellite observations and in ICON Sappire, a cutting-edge global storm-resolving model.  The statistical analysis aims to get at causal pathways driving changes in cloud properties and dynamics due to aerosol perturbations.  As the aerosol/microphysics/radiation treatment and coupling in the model is somewhat simplified, not all of the same pathways exist in the two frameworks.  However, the analysis yields some insight into the behavior of the model and its biases.

Recommendation: Minor revisions

This paper is advancing a new and seemingly promising statistical technique for understanding the (causal) mechanisms driving aerosol-cloud interactions.  I am supportive of this paper being published in ACP, though after minor revisions.  In the report below, I ask many questions about the causal approach and how the details or uncertainties in the satellite retrievals might affect comparisons with the model results.  As I am not an expert in either statistical analyses related to causal analysis or satellite retrievals, I would ask the authors to think of them as questions driven by curiosity rather than as criticism of the manuscript and study.  Perhaps, other readers new to these approaches might have the same kinds of questions.  Some of them may not be worth addressing in the manuscript.  Most all of the others might just require a reference or a sentence to address.

I would be willing to review the manuscript again if the editor thought it would be useful.

===========================

Major comments:

1. (p. 12/sec 3) I have some knowledge of cloud-aerosol interactions but little of causal approaches such as used in this study.  I have certainly learned a lot about them by reading this paper and some of its references.  My questions below may be simple minded, but I might not be the only reader to have such questions.  While addressing some of the questions might require a bit of extra exposition, I think only a sentence or two in the appropriate places would be sufficient.

 - My understanding from Fons et al (2023, specifically the caption to supplemental figure 12) is that the satellite causal graph reflects only daytime conditions.  Is that true?  If so, please state it clearly somewhere.  (Apologies if I've missed that.)  However, the availability of both day- and night-time data from ICON offers the possibility to see whether the causal networks change from day to night (at least in model world).  That seems interesting, because marine boundary layer clouds have significant diurnal cycles of radiative cooling, entrainment, liquid water path and precipitation, which could (?) change the causal connections between day and night conditions.

 - As GPM satellites with radar and/or microwave instruments are not continuously overhead, most precipitation data will rely on the same geostationary satellite that is also providing information about cloud optical depth, effective radius and cloud top temperature.  When combined with the adiabatic assumption, there seem to be many fewer degrees of freedom in the satellite data than arrows in the causal diagram.  How should the reader think about this?

 - Rainfall (measured at cloud base or the surface) matters to the dynamics of the boundary layer when it is strong and has little influence when it is weak (in comparison to the other drivers of turbulence and convection).  If the impact of rainfall on the boundary layer is nonlinear, how will this show up in the causal relationships, which seem to be linear and based on the standardized changes of rainfall across all times?

 - As susceptibility of low cloud quantities is often defined in terms of logarithmic changes (e.g., equation 4 in Terai et al, 2012, https://doi.org/10.5194/acp-12-4567-2012), would it be worth contrasting that briefly with the linear sensitivities represented by the alpha's and beta's in this study (although standardized as described as in Fons et al, 2003)?

 - Cloud fraction adjustments to increased aerosol occur in clouds with low background aerosol concentrations.  Are they worth including in the causal network, especially if the pixel size for the causal analysis is 0.5 degrees?

 - Are the causal relationships in the network based on the high CCN or the low CCN simulation, or a combination of the two?  My understanding of the time series analysis is that the causal relationships depend on time variations in predictor quantities in a single simulation.  That seems quite different from other approaches that difference quantities between the high and low CCN simulations, so it would be good to be clear on this point.

 - (inspired by p. 17/fig 8) If the causal effects are computed from time series with an increment in predictor (e.g., aerosols) at time zero, wouldn't the influence of that local in time-and-space aerosol increase move away from a fixed Eulerian location as the winds advect the airmass?  What is the meaning of these after-effects 24 hours later than an aerosol (or other) perturbation at a fixed Eulerian location?  Does the aggregation of data to larger scales (here 0.5 degrees and 10 degrees in Fons et al, 2023) avoid this issue?

 - Also, do satellite retrievals of effective radius correspond to cloud top values or some type of integral over the depth of the cloud?  Shang et al (2023, ACP, https://doi.org/10.5194/acp-23-2729-2023) suggest that different wavelengths may give information about different vertical levels.

If answers to these questions are in references like Fons et al (2023), Runge et al (2019) or elsewhere, the authors could emphasize that in the text.

===========================

Specific/minor comments (11/240 means p. 11, line 240):

 - Perhaps, the abstract could mention that the simulations are made for a period during boreal summertime.

5/109: Did ICON try to include an estimate for subgrid vertical velocity variance in the parameterization for activation?  While subject to uncertainty, that might have helped correct the low bias in Nd.  However, this might not be worth such an effort when using a Smagorinsky-type parameterization, because they do not perform well in predicting subgrid quantities in simulations with these kinds of grid spacings.  See e.g., Cheng et al (2010, https://doi.org/10.3894/JAMES.2010.2.3, sec. 3.1.2.1).

8/Fig 4: Do metrics related to decoupling change between the low and high CCN simulations?  The drizzling boundary layer might encourage such decoupling with rain evaporation below cloud base.  The differing rain mixing ratios (and presumably) rain rates suggest different levels of precipitation-related subcloud evaporation in the two simulations.

Also, vertical velocity variance might be useful for thinking about activation and cloud formation, alongside the mean vertical velocity.

9/Fig 5: The labels a/b in the caption refer to the wrong panels.  Also, why not also show the same plot for the low CCN simulation, or at least a contour showing how the joint Nd-LWP distribution shifts from low to high CCN?

Is the joint distribution based on grid point/satellite retrieval footprint data?  if one or both datasets coarsened to the 0.25 or 0.5 degree resolutions mentioned in the satellite data appendix, a short description of how the Nd or effective radius data are coarsened (e.g., average over all cloudy columns) would be useful.

With the high CCN simulation having uniform "aerosol" concentrations, I would expect that changes in cloud base Nd in ICON are induced mainly by changes in the updraft strength, while satellite Nd changes are most likely dominated by actual aerosol variability.  Is this consistent with the interpretation of the authors?  If so, how should this contrast make us think about the results?

9/163: The "invisible ship tracks" paper of Manshausen et al (2022, https://doi.org/10.1038/s41586-022-05122-0) does show positive LWP adjustments in trade cumulus regions and might be worth mentioning here alongside Jiang et al.

10/Fig 6: Are the Nd and r_d values shown here averaged over all grid points (cloudy and clear) or only over cloudy grid points?  The in-cloud values are a more useful reference when thinking about aerosol-cloud interactions.  I would be surprised if the in-cloud values of N_d and r_d would go to zero so smoothly at the top and bottom of the cloud layer, but perhaps this does happen in ICON.

11/182 and 14/266: Does ICON Sapphire really use a fixed effective radius for all cloud liquid even when the double moment microphysics is used?  I would understand this choice, but this should really be explicitly stated somewhere in this paper since I couldn't find it in the original Sapphire paper.

13/236: Could the magnitudes in the model graph be stronger because the model Nd is low and the clouds in the model simulation are precipitating more strongly than in the satellite data?  Is it clear that the alpha's should be invariant as the background aerosol/Nd changes?

top of p. 14: For the reader, it might be helpful to first talk about alpha's with clear signals in the satellite and model before later talking about those that are more complicated.

15/304: Stevens and Seifert (2008, https://doi.org/10.2151/jmsj.86A.143) is an earlier study along these lines.  Albrecht (1993, https://doi.org/10.1029/93JD00027) isn't a direct predecessor but makes clear the impact of precipitation on the depth of the marine boundary layer.

19/420: Possible reference for "increasing the vertical resolution": Bogenschutz et al (2023, https://doi.org/10.5194/gmd-16-335-2023).  More sophisticated subgrid turbulence closures also offer some promise of better representation of stratocumulus at kilometer-scale grid spacing: Shi et al (2018, https://doi.org/10.1175/JAS-D-17-0162.1), Bogenschutz et al (2023, https://doi.org/10.1029/2022MS003466).

20/449: Specify a particular altitude, not one relative to the inversion height (which is variable).

22: It would be worth mentioning the footprint of the satellite retrievals from SEVIRI, to give some context about how many pixels are being averaged into the 0.25 or 0.5 degree grids.

===========================

Typographical/rephrasing suggestions (OPTIONAL):

12/215: "along" --> "alongside"

16/318: "datam" --> "data"

19/421: "e.g.," should be before the reference.  Maybe add another <> before the {PossnerEtAl2014}?

20/444: "a an" --> "an"