Parameterising cloud base updraft velocity of marine stratocumuli

Ahola et al

Overview

This article presents an LES assessment of 3 types of parametrisations used to derive cloud base vertical velocity in large scale models.

The methods tested include one “traditional” linear parametrisation where updraft velocity is derived from cloud top radiative cooling and two parametrisations derived from different machine learning techniques (Gaussian process emulation and random forest). The authors demonstrate that when compared to LES simulations, which are viewed as truth, the machine learning techniques produce a more accurate representation of the cloud base vertical velocity than the “traditional” method.

General Comments

The authors do a nice job of explaining machine learning methods and defining a workflow. In particular, the authors have structured the methods around the workflow design, which makes the method easy to read and reference. Using the workflow, the authors clearly show the application of the machine learning techniques can improve on more “traditional” parametrisation methods when all methods are compared to LES data. On face value this is a nice result, however, it is very difficult to understand if either parametrisation is performing well because the article fails to present enough information about simulations to understand the validity of the training data. In particular

1. the authors fail to present any cloud evolutions over time, e.g. cloud base and height plots, LWP and surface/cloud base precipitation. The authors do present statistics, but it is important to present a mean and a range of the time evolution of the simulated marine Sc, so that a reader can be confident the machine learning is training on sensible pseudo-data.
2. There are a lot of thresholds defined in the paper, which are used to include or ignore data, but many of these thresholds are not justified and it is not demonstrated how sensitive the techniques are to these thresholds.
3. It is unclear why the authors use a fixed solar zenith angle for daytime and separate daytime and night-time, rather than run diurnal simulations. This choice, without further justification, casts doubt on the simulations, i.e. a diurnal simulation was not possible.

Hence, while I think there is good explanation of the methods, the minimal presentation of the simulation results to demonstrate the simulations produce a good representation of marine Sc mean that I recommend major revisions.

Specific comments

• Line 70 – I was surprised to read that there is an autoconversion scheme in the bin model. It is important to note that such a configuration for a bin model is non-standard, possibly unique(?) since bin models tend to represent the (semi)-continuous growth of activated droplets to rain through condensation and collision-coalescence processes. From the description in this article, it is not clear whether SALSA includes collision-coalescence for cloud bins or do the cloud bins just experience condensation? This should be clarified in the text.
• Line 70 – The authors explicitly state that SB uses saturation adjustment but they do not state what SALSA uses. Does SALSA use a prognostic supersaturation scheme? This needs to be clarified and stated for completeness.
• Line 85 – The filtering applied to the ECHAM data isolates single layer clouds, which are below 3000 m. What do the authors mean by single layer cloud? Could they provide more information since it is not clear whether the authors are filtering on clouds that are one vertical grid in thickness or clouds that are multiple grids but continuous, i.e. not multi-layer clouds between the surface and 3000 m. Without more information, I have assumed that the filtering is on multi grid clouds and I am concerned that the filtering of this data will thus result in the inclusion of Marine Sc and Cu. If this is the case then the clouds could develop from different dynamical forcing, which could be a problem because the “traditional” parametrisation (Zheng 2016) is only applicable to marine Sc. Hence, could the authors confirm that Cu is filtered out of the data and if this can not be confirmed, could they discuss the impact of the inclusion of Cu in the data?
• Line 89 – How sensitive is the filtering to the cloud water threshold?
• Line 90-91 – Is there a justification for the IWP and LWP percentage thresholds, beyond the maintenance of the single-layer cloud.
• Line 92 – Why has 3000 m been selected as the cut-off for the clouds?
• Line 145 – the authors present examples of the initial profiles, which is good and they look very sensible for an idealised marine stratocumulus. However, the authors do not present any demonstration of the simulation beyond the statistics of wpos versus cloud top longwave cooling. This means that as a reader and a reviewer, I can not make any assessment of the validity of the training data, i.e. how does the daytime or nightime cloud evolve? Is the turbulence spun up? Is the simulated marine Sc coupled or decoupled from the surface, are they precipitating, etc? This is really important information, which the reader needs to understand the presented plots. Could the authors provide example plots of the evolution of cloud top and base, LWP and RWP and surface precipitation? Ideally, the lots should include some representation of the range of values.
• Line 150 – The article presents 2 sets of simulations – Daytime with fixed solar zenith and Nightime. I am puzzled why the authors did not do a diurnal cycle and sample the data in day and night? Could the authors justify the choice of fixed solar zenith angle over a diurnal cycle? Also, are the results sensitive to the choice of solar zenith angle?
• Line 175 – the simple, non-machine learning parametrisation is based on Zheng et al (2016), why was this the only non-machine learning parametrisation tested?
• Line 230 – 235 – Are the authors filtering for columns that include evaporating rain? For example, the cloud base is defined as the lowest grid where LWC > 0.01 g kg-1. If rain is evaporating then this threshold could be very low, since the evaporation will return drops to cloud sized bins in SALSA (this will not be an issue in SB), hence there is the risk the updraft velocity at that grid does not represent cloud base where activation occurs. Could the authors clarify that evaporating rains does not bias the cloud base vertical velocity?
• Line 245 – Data was thrown away if the cloud fraction was smaller than 0.61 or cloud top rose more than 10%, could the authors justify these thresholds and add some discussion about the sensitivity of the parametrisation comparison to these thresholds? (I note the discussion of Feingold et al but that does not justify the 0.61.
• Line 281 – Updraft velocity parametrsiations from SB are shown to be better than those from SALSA. I agree that this is mainly due to the difference in sample set number but, could the difference also be due to the use of a prognostic supersaturation in SALSA vs “all-or-nothing” scheme in SB.  

This is an interesting study that attempts to improve the parameterization of cloud-base updraft speed for marine stratocumulus using a large ensemble of large-eddy simulations (LES) driven by a GCM. Three methods are examined: (1) a linear physics-based scheme developed by a previous study, (2) a machine learning method that incorporates the (1), developed by the authors, and (3) a purely machine-learning-based method without incorporating physics, also developed by the authors. By comparing these three methods, it is found that (2) and (3) markedly outperform the (1), suggesting their potential usage for improving cloud-base updraft simulation in more coarse-resolution models. Some analyses about which predictor is more dominant were conducted. 

Overall, the manuscript is clearly written and straightforward to understand. The conclusion that the new parameterizations improve the previous one is robust and convincing. Also, the data of the large ensemble of LES, which are released publically, should be very useful datasets.  However, there are two major gaps that undermine the scientific quality, with one regarding the sampling issue and one regarding a lack of physical interpretation of the results. If these two issues can be well addressed, the manuscript should be a good fit for the ACP. See my detailed comments below.

Major comments:

(1) Marine stratocumulus only occurs under certain large-scale environments (e.g. see Wood (2012)'s review article). Your samples include many cases with delta_thetal of only several K degrees (Fig. 2b). Under this condition, an overcast stratocumulus deck can hardly sustain because of the strong cloud-top entrainment under a weak temperature inversion. Moreover, under weak inversions, the strong entrainment of warm air from above into the boundary layer, stably stratifying (or decoupling) the boundary layer so that your assumption of well-mixedness becomes invalid.  Although you removed some of these cases (in Lines 244-251), the standard seems arbitrary to me and it cannot guarantee the remaining samples are physically reasonable since they are sampled based on a purely statistical method. I would suggest plotting a map showing where those samples are, so that the readers can have a sense of what kinds of stratocumulus we are looking at: eastern subtropical stratocumulus deck? or stratocumulus in the postfrontal region of midlatitude cyclones? or polar stratocumulus under an unperturbed environment? The point is to build a physical context for understanding those samples. Actually, Zheng et al. (2016)'s study is limited to only two regions, which does not necessarily guarantee a universal relationship.

(2) I feel the discussion of the results is somewhat superficial. For example, entrainment is a crucial process governing the boundary layer dynamics. Stronger entrainment of warm air above can stabilize the boundary layer, reducing the turbulence. Such an important process, however, is not even mentioned in the discussion. Another important process is evaporative cooling that can also promote instability and enhance turbulence. Discussions should include these fundamental processes that contribute to the TKE budget of the boundary layer. 

Minor comments:

L335: "developed three xxx".  I think the first method is developed by previous work. The wording should be revised.