Review of Technical note: Emulation of a large-eddy simulator for stratocumulus clouds in a general circulation

model by K. Nordling et al.

General comments

The article demonstrates how a computationally light Gaussian Process Emulator is used as a parameterization of shallow clouds in the ECHAM climate model. The emulator was trained on LES simulations.  This is important, because the representation of shallow clouds in GCMs is a major source of uncertainty, as they occur on spatial scales smaller than the model grid. The emulator approach is an interesting way of using high-resolution models to inform the parameterization of these clouds in a GCM, and this work demonstrates it in practice. The fact that the emulator can replace a parameterization in a GCM, and run a long time without crashing the model is by itself an achievement.

I like the disciplined approach of deciding for which conditions to apply the emulator (low clouds) and using the emulator for only a small number of well-chosen quantities (here vertical velocity and precipitation rate).

I have concerns with how the simulations used for emulator training were implemented, detailed below. These should be addressed in the text. I recommend the article is published with a minor revision (and with the simulations as they are).

Specific comments

The emulator appears designed especially for stratocumulus clouds (title and L58). However the criteria for using the emulator in the GCM are not very strict, and will allow the emulator to be used also for other cloud types, in particular shallow cumulus. I worry that the LESs used for training are small (10x10 km) and run for a short time, 3.5 hours. This means that any mesoscale cloud organization effects (see e.g. Bony et al 2020, https://doi.org/10.1029/2019GL085988) will be missed. For stratocumulus such effects may not be decisive, but for shallow cumulus they are, and then strongly influence the precipitation. Such organization effects are probably not well captured by the original GCM either, so in my view it would be very valuable if the emulator could include them in the parameterization.

A second issue related to cumulus is whether the set of input variables to the emulator is adequate also for cumulus. For example, the surface fluxes of heat and moisture might be important parameters for cumulus. 

The LES simulations are mentioned as the main computational bottleneck in this work. Could you state how expensive they are?  My feeling is that longer and larger simulations would be possible with current computational resources, even if 1000s of them are required.

Some more details of how the emulator is constructed would be useful. Did you have to or choose to implement it yourself, or was an existing library used? In the source code I saw the GPF library. It's not easy to find the author of it, is there something you could cite?

Does the reference Rasmussen and Williams fully define the algorithms used, or did you make additional design decisions? If so, it would be good to document them, the description is now quite brief.

Does the emulator give measures of uncertainty?  Can such a measure be used to see where additional LES runs would be beneficial, if one wants to extend the training data set?

Was a new LES dataset constructed for this paper, or were the runs from Ahola 2022 used again? This was mentioned in several places but gave contradictory impressions.

End of section 2.3 - the vertical distribution of precipitation: This seems a good solution to a non-obvious issue that appears in this approach. I appreciate that it is documented here.

Minor remarks:

L8: "Although especially..." incomplete sentence.

L56: the superparameterization of Jansson et al 2019 (also 2021 https://doi.org/10.1029/2021MS002892) uses an LES specifically aimed at improving the parameterization of shallow clouds (at a high cost and not globally).

L56: For the discussion of previous approaches and emulators for stratocumulus: Glassmeier et al., 2020 (https://doi.org/10.5194/acp-19-10191-2019) construct a Gaussian Process Emulator for stratocumulus clouds based on LES results not for GCM use but to understand the different states of stratocumulus, and should be mentioned.

L135: the definition of the jumps is hard to understand.

L157: The removal rate procedure is hard to understand.

Fig 5: The x axis ticks on the two top rows could be left out, to be consistent with the look of the y-axis and to avoid clutter

Code and data availability:

I appreciate the code and data being openly available.

Line 420: data reference is broken "ECHAM simulation results are available from ? (Nordling et al., ?; last access ?)"

The emulator implementation (https://github.com/kallenordling/eclair_emulator): a minimal README and a license statement would be helpful. Additionally, you could consider archiving a specific version in a permanent repository such as Zenodo, which generates a citable DOI. GitHub and Zenodo work well together.

The repository of UCLALES-SALSA could also be mentioned here.

I am genuinely hopeful that new methodologies of the type described in this manuscript will lead to important insights into the physics of the atmosphere and our ability to more effectively model and predict weather and climate. But unfortunately, this particular manuscript does not yet achieve these goals. It can be said that this work often appears as a ‘solution looking for a problem’. It may well be that as a proof of concept this technical note may be publishable, but my suggestion is for the authors to consider the following points and improve the manuscript accordingly.  

1) If the authors wish to reach the parameterization/modeling community (which I hope they do), they should make an effort to improve the description of their approach. These new methods are so different from what is traditionally done by the parameterization and modeling communities that they do require much clearer explanations (including better schematics).

2) A key issue is that there appears to be an inconsistency between the problem/regime that is being addressed and the ‘solution’ that is being proposed. This link between the ‘problem’ and the ‘solution’ needs to be established in a much more effective way - see (3) and (4) below.

3) If the focus of the paper is on shallow marine clouds, why would the authors select cloud base updraft velocity (in the context that it is being used) and rain water formation rate as variables for their study? Please clarify.

4) The marine boundary layer clouds problem in climate and weather models is, to first order, a turbulence-convection-macrophysics problem. So, why are the authors more focused on microphysics?

5) A demonstration of how far this manuscript is from the more traditional parameterization and modeling research is the sparsity of discussion (and corresponding references) of the work that has been done in this field in the last few decades. This aspect needs to be clearly improved. This work (which is novel and interesting) needs to be grounded in what has been attempted over the last decades and its failures and successes.

6) LES are extremely powerful tools to help develop and improve parameterizations of turbulence, convection, and cloud macrophysics. But this is not necessarily the case for cloud and aerosol microphysics: LES and other atmospheric models suffer from many similar issues in this context. The LES problems with microphysics have been clearly reported in the literature. So, why are the authors focusing on microphysics? Why should LES be trusted? Please clarify.

7) Independently of their accuracy, parameterizations and the models in which they are implemented often display internal consistency. The method advocated in this manuscript appears to potentially break this consistency in a variety of ways. Although the authors briefly discuss this issue in the context of the variables that they are focused on, they should openly address this critical issue in broader terms.

8) The manuscript needs a more explicit discussion of cloud cover and cloud fraction profiles, including adding figures with global maps of cloud cover (such as is done for other variables) and with profiles of cloud fraction (as in figure 5).