The study presents a novel methodology to retrieve snowflake mass and geometrical properties from the simultaneous observation of the snow particle by multiple viewing angles. The authors developed a GAN algorithm to address this problem. It is worth noting that this is no trivial challenge since the shape of snowflakes is highly irregular.

Finding a methodology for the automatic estimation of snowflake mass, on a single particle level is still an area of active research since it poses significant challenges. The problem is also very relevant for many applications ranging from microphysical studies of snow to remote sensing simulations. For this reason, I consider the study of great interest for an AMT publication.

The paper is logically structured and easy to follow. The graphics are clear and the text is well written.

The methodology is well described and made easier to apply by other researchers thanks to the open code and data sharing.

The limitations of the proposed methodology are appropriately discussed and the application example helps to illustrate the relevance of the study.

I did not find any major flaw in the paper and I certainly recommend it for publication after minor revisions. I will take this opportunity to list a few minor comments that I hope will help the authors refining their work.

The thing that puzzles the most is the fact that the GAN is trained with model data. This point is appropriately discussed at line 155 and following, but I am still thinking about how this affects the results presented in table 4. It appears to me that the exponent of the power-law fit for the mass-size relation is always approximately 2 and only at very high degrees of riming (75-100%) it significantly deviates from this value. This is coherent to what was already shown in Leinonen and Szyrmer (2015) where the model predicts that riming mostly affects the prefactor of the mass-size relationship, while the exponent is not affected. In that study, a completely different growth model (i.e. rime growth, model C) is needed in order to affect the exponent of the mass-size relation. Because of that, I am suspicious about the fact that this result might be a consequence of the model employed.

On the other hand, it seems that there is some evidence to support the idea that the exponent of the mass-size power law should vary more continuously between the unrimed aggregates b=2 and the 3D scaling b=3 (see e.g. Mason et al 2019 Retrievals of Riming and Snow Density from Vertically Pointing Doppler Radars, figure 1 and equation 8).

In my view, when snowflakes grow by riming they increase mostly in mass (fill-in theory and Seifert et al. 2019) and will start increasing in size more and more while their rime mass fraction increases. In my view, the smallest particles should reach this limit first, hence their size would start increasing earlier than one of larger aggregates, thus the exponent b of the m-D power low starts increasing. This conceptual view comes from the idea that the riming degree of single particles is not constant inside a snowflake population.

If I understand correctly (please correct me if I am wrong), snowflakes coming from various particle populations (different times/ weather events) are stratified according to their riming degree in Table 4. If the logic of my previous paragraph is valid, the m-D fits derived in Table 4 are not well representing natural snowflake populations because, in general, one should assume a size-dependent riming degree.

My additional questions related to this point are: How do the results of Table 4 compare with those of Mason et al. 2019? How confident are you in the quality of the used riming model to represent real physics? Is it possible that the results are affected by a biased model? Can you elaborate on which improvements to the snowflake model might be needed in order to make it producing snowflakes whose mass scale with exponents in between 2 and 3?

A few more specific comments:

Line 141: I guess also the orientation of the particle matters for the simulated silhouette. In other parts of the paper, the orientation of the particle is discussed as a source of uncertainty. I wonder if it would be possible to constrain orientation by means of hydrodynamic models and exploit it to constrain also the GANs retrieval

Line 185. I guess that also the surface properties of ice (roughness mainly) are needed in order to simulate the interaction of light with the snowflake and that is also something that the used model does not provide.

Line 191. I guess there is a typo PhotoTOnic

Section 4.1 The used apparatus seems very expensive and has significant limitations nonetheless (max size, fragile material). Maybe the authors can give some indications on which are the technical specifications for a 3D printer suited to replicate the experiment.

Line 202 Is the fall speed impacting the measurement capabilities of MASC? I guess the material used for 3D printing has a different density with respect to ice and thus a different fall speed.

Line 255. In Equation 2, if I got it right, m^i is the same quantity as m_i in Eq.1 just with the voxel identifier i shifted from subscript to superscript. If that is the case I would suggest using the same notation in both cases (I have a personal preference for subscripts)

The article presents a proposed method for hydrometeor mass retrieval and geometric 3D modelling by the application of a GAN trained for these purposes. The GAN is trained on simulated data and the reasoning for this is well explained, however this might warrant further work with captured data. 14 printed flakes for 198 images seems like the minimum (although I may be wrong).

The paper is well written and does an admirable job of explaining a difficult topic. The authors make an effort to outline the limitations of their work and discuss their means to address them. The figures and tables are clean and support their work.

I have no concern publishing this document, although I do have some general questions for the authors.

General Questions:

One thought I have is if the snowflakes used (3D printed) for evaluation were part of the training? Or were they generated specifically for evaluation. Is there any thought on how much error is introduced by using faux snowflakes in validation testing on a network trained on simulated snowflakes? I expect it to be minimal but wonder what your intuition is. This is a difficult problem to solve, and I commend your approach.

Did the printing allow for the introduction of air pockets? How solid were these printed flakes? Did any flakes have cavities? How did the GAN perform with irregular shapes? You site Kleinkort et al with a “volume reconstruction (using a standard 3-camera MASC) is quantified to be 27% in terms of absolute error…”. Kleinkort found improvement by introducing additional camera angles. Have you given thought to including additional angles for the GAN?

Line Specific:

170- What were the 3 PCA components kept? Or what were they related to. Might help in reproducibility if we had that information.

192- Were there any experiments with different printer material? A larger validation set would be beneficial and perhaps a more durable material could assist in that.

289- Mean terms of mean NSE. Want to make sure this is not a typo.

319 – Table 5 and 4… this causes me fits. I don’t know if there is a rule of numbers being listed in order, but it certainly stands out.

              Repeated in Fig. 6 so at least the authors are consistent.