This paper investigates in some detail the vertical structure of the lower atmosphere in the Arctic using MOASiC data. While the attempt as such is commendable and the methods are interesting, the paper does not harvest. Nothing new or unexpected comes to light; instead most of the conclusions either follows logically at a textbook level from cited criteria, or worse, in some cases are side effects of the sampling. The authors seem unaware of a lot of material already published on this topic, but are hung up on a few widely spread misconceptions that they belive they have proven but not, and also seem to lack insight into turbulent flows and boundary-layer dynamics.

These concerns are grave enough that I must recommend rejection at this point, while hoping the authors come back and do a better job of this topic because it is important.

Major concerns:

Off the bat, the authors erroneously claim that the Arctic ABL is almost always stably stratified and then does whatever they can to make that the truth, although their own results actually shows this to (still!) not be the case. The most common stability in the Arctic ABL is in fact near neutral; after all, what is the stability difference near the surface between “near-neutral with stability aloft” and “shallow and well-mixed”? It is not the stability but the depth!

These authors need to read up more literature on the vertical structure of the Arctic ABL, perhaps beginning with the Tjernstrom and Graversen paper (2009 in QJRMS), performing a similar analysis on the SHEBA data, however, without the benefit of SOM. Yes, SHEBA was a while ago, but all things old are not useless.

In the end, the most common of the 12 stability classes used in this paper comes out to be NN in all seasons except summer (Figure 6) and adding in the well-mixed shallow cases, that all has a median Rib close to zero (the very definition of near neutral), it is clear that the initial statement and the conclusions by these authors in the text is just plain wrong (cf. e.g. lines 39, 363 and 693 in the manuscript)!

The SOM approach is interesting but poorly both explained and executed. Why so very many nodes? Is there no way to objectively determine the optimal number of nodes, e.g. by minimizing the inter-node and maximizing the intra-node variances? In fact, by reducing everything to 12 classes in Table 2, the authors themselves pretty much abandons the SOM, and does something else that does not require it; the SOM part becomes underutilized and, in the end, doesn’t add much to the final results, mostly based on the 12 classes and not on the 30 nodes. I would also like to know much more about how the SOM analysis handles structure versus geometry. Will, for example, profiles profiles with similar structure but very different geometry end up in the same node? Consider, for example, two inversion capped but well-mixed ABLs, one with an inversion base at 200 m and the other at 1 km, end up in the same node? A lot is confusing in the way the SOMs are described, for example there seems to be way more than a 100% of soundings, adding the numbers in each node for a total. BTW, many of the numbers printed in the large node boxes in the figures are so small I neede a magnifying glass to read them. If you insist that the SOM analysis is fruitful, and I'm not saying it isn't, then do it properly and explain it well; that could very well be a useful paper all by itself!

Then make a separate paper out of the 12 stability classes, but do a proper analysis and try and find something new. Work more with the crieria; ask yourself, for example, what would be necessary to characterize a decoupled stratocumulus, with a high inversion and most of the turbulence is generated in the cloud layer, or an inversion-and-stratocumulus-capped relatively deep but coupled ABL where the buoyancy from the cloud top generates a much deeper total ABL than motivated by the surface fluxes (e.g. Brooks et al. 2018 in JGR). There is a lot to be gained here but it does need an insightful analysis. A very stable boundary layer is, as we all know, shallow with large stability, large wind shear and Rib but low u*; everything in Figure 7 is just a confirmation of textbook ABL meteorology. If you select your criteria this way, cases will have all these other characteristics automatically; no use even looking at the result and the question has become the answer, while no one learned anything new.

In Figure 8, nothing except panel d, is significantly different from anything else and most of the conclusions are hand waiving. WS & NN are the deepest and hence expected to have the largest jet wind speeds simply by their distance from the surface; only surface friction can reduce wind speed; ABL turbulence just mixes things around. Temperature inversions (Figure 9) are common – everywhere! You find them also in the mid-latitude convective ABL. What I think is special for the Arctic is that: i) the ABL is so often quite shallow; ii) when not shallow, it is often capped  by stratocumulus providing the extra energy explaining the depth, and iii) when it is stable (which is mostly in winter), it can be so very stronly stable for a long time, since there is no diurnal cycle that resets the stability once per day. If you base a category on having weak stability aloft, is it surprising that this class stands out when you analyze that inversion, and what does a median inversion-base height at 2 km has to do with the ABL? And how can the WS alone class have a larger “inversion intensity” than the WS-MSA and WS-SSA classes; ins’t that counterintuitive? Going on, the WS ABL is formed by surface cooling, so it stands to reason that it has no clouds or at least a high cloud base; clouds higher than a few km in winter has very little effect on the surface energy budget. The lowest clouds seem to be found in well mixed or near neutral cases; one may wonder if that is because these low clouds force that particular stability? Or is it the other way around?

Finally, drop the AUV data! For two reasons: i) I can’t see that it adds anything useful, and ii) the risk that it contaminates the statistics; radiosounding where done every day regardless of weather but UAVs flew only occasionally during a part of the year.

I could go on for many more pages, and list many detailed complaints and things I don't understand, and I would have if I could find it in me to recommend - at least - major revision. It makes me somewhat sad to have to be so negative; obviously this looks like a failure in supervision. But I wouldn’t be doing my part in upholding the quality of the journal if I let this through – I'm so sorry.

This manuscript describes a statistical analysis for a year-round period of Arctic boundary layer observations based largely on radiosondes during the MOSAiC project accompanied by data observed by the DataHawk2 unmanned vehicle. As a central tool, a "self-organizing map" approach was applied to the data. There is no doubt that such a statistical analysis is extremely helpful in addition to all the case studies that have been and are being evaluated. Therefore, the enormous amount of work is greatly appreciated, and after major improvements to the manuscript, I also support the publication of this analysis. However, the manuscript needs a thorough revision that goes far beyond classical "major revisions". I will try to justify this in detail in the attached review.