Please find the review in the attachment. 

The manuscript documents relationships between above- and below-cloud CCN retrievals from a HSRL (and a neural network) with cloud microphysics estimated from the Research Scanning Polarimeter during ORACLES field campaign, over the Southeast Atlantic Ocean.

This is an interesting manuscript that makes a clever use of airborne remote sensors for investigating the aerosol indirect effect.  The integration of vertically resolved lidar retrievals is particularly appealing, with a methodology carefully crafted to minimize sampling biases through a spatiotemporal collocation that accounts for the typical spatial variability of aerosols. This manuscript will be an important reference for advancing the way we use remote sensors for investigating aerosol-cloud interactions and cloud adjustments. My recommendation is acceptance after minor revisions.

Comments:

1. It is somewhat intriguing that the below-cloud CCN shows a weaker slope with cloud droplet number concentration (Nd) than the above-cloud CCN. Of course, chances are that some unknown artifacts could be conspiring to reduce the correlation between below-cloud CCN and Nd. For instance, there is a big contrast in relative humidity between the boundary layer (BL) and the free troposphere (FT), which might be affecting the CCN estimate. Or, this could reflect issues with ERA-5 and its inability of properly represent the BL depth (e.g. an underestimation of the BL height). A way to partially test the “artifact” hypothesis would be to derive ACI and correlations between in-situ CCN and Nd for a number of specific (in-situ) profiles and see whether this in-situ ACI is comparable to the remotely-sensed ACI. This in-situ estimates for ORACLES might be already available in the literature.
2. The machine learning CCN follows the methodology of Redemann and Gao (2024), however, the input parameters are not identical, which is reduced in Lenhardt et al. This is a reasonable approach, however, the whether the performance of this ML is similar to Redemann&Gao is unknown. Is this information reported in another article? Is the performance similar?
3. While I understand the scientific motivation of estimating CCN with machine learning techniques, especially given the CCN importance in the context of the aerosol indirect effect, submicron aerosol concentration is a variable that makes more sense to derive from a HSRL (e.g. Sawamura et al., 2017, https://acp.copernicus.org/articles/17/7229/2017/). If it is possible to apply a physically based algorithm to compute aerosol concentration, then I wonder whether deriving submicron aerosol concentration using a neural network would yield better results than deriving CCN. Moreover, the choice of supersaturation is not guided by physical arguments but by data availability. For example, I doubt that the real supersaturation values in these stratocumulus clouds exceed 0.2%, whereas the value used in the article is 0.4%. I am mainly interested to know the authors’ opinion regarding whether aerosol concentration is a variable that we should focus on rather than CCN (I am not expecting a revised manuscript that address this point, but I am very interested in learning from the authors’ perspective).
4. In 2013-2014 we wrote a pre-ORACLES article (Painemal et al., 2014) aiming at documenting how cloud microphysical properties changed over the SE Atlantic and the role of cloud top height and aerosol layer altitude. Similar to Lenhardt et al. (2026), we found a strong control of stability, manifested in a regional distribution that gave rise to two distinct patterns north and south of 5˚S. The area north of 5˚S was particularly interesting because cloud droplet effective radius (Re) and LWP were anticorrelated, which appeared to be mediated by changes in the boundary layer height. The regional analysis also described key differences in cloud top height and aerosol base height from CALIPSO between the northern and southern part of the ORACLES domain. Of course, ORACLES data are richer than CALIPSO and MODIS, but I am mentioning this paper because Lenhardt et al. captured in greater detail those processes we observed with satellite data. While I am hoping that our paper might help provide a regional context to this ORACLES analysis, please don’t feel obligated to cite it.

Painemal, D., S.Kato, and P.Minnis (2014), Boundary layer regulation in the southeast Atlantic cloud microphysics during the biomass burning season as seen by the A-train satellite constellation, J. Geophys. Res. Atmos., 119, 11,288–11,302, doi:10.1002/2014JD022182.

Other comments:

Line 31: For a technical paper, this information is a bit redundant.

Line 131-132: “cloud edge humidification effects are considered in the prediction of NCCN under dry conditions”. What type of cloud edge humidification effects can be captured if reanalysis cannot replicate this? Moreover, gradients in relative humidities for cloud edges occur at scales much smaller than the reanalysis resolution.  It seems that the only humidification effect that can be captured is that dictated by the environmental relative humidty.

Line 144-145, you mean overlying smoke layer tends to underestimate satellite Reff and cloud optical depth if not accounted for in the algorithm? (Meyer et al. 2013, JGR).

• When it is stated that lower tropospheric stability (LTS) provides a meteorological constraint, it is pertinent to clarify the spatiotemporal scale. LTS is somewhat a good predictor of cloud coverage for describing annual cycle and large scale processes. However, LTS and cloud coverage (and cloud microphysics) poorly correlate at synoptic scales. Because LTS in this study is really describing climatological changes between 2 regions over the SE Atlantic, it is more accurate to say that LTS provides a way to separate climatological regimes for the investigation of ACI (the concept of meteorology is just too broad).
• Do the generals findings of the article differ if instead of using the ML-based CCN, the aerosol extinction coefficient at 532nm from the HSRL were directly used for the analysis?
• Figure 4 and similar bivariate histograms. The green color bar is a bit difficult to see. I would suggest to use a color palette with a discrete number of colors (no more than 12).

David Painemal, NASA LaRC