Comments on manuscript:

Combining low and high frequency microwave radiometer measurements from the MOSAiC expedition for enhanced water vapour products, by Walbröl et al.

Overview

The work presents retrievals of vertically resolved temperature, humidity, and integrated water vapor (IWV) from combined measurements of 2 microwave radiometers, including frequencies at or higher than 183 GHz. The retrievals were trained with ERA5 data from the European Centre for Medium-Range Weather Forecasts (ECMWF) and synthetic MWR observations simulated with a radiative transfer code. Retrievals were evaluated against ERA5 and radiosondes launched during MOSAiC. The combined retrievals noticeably improved the IWV and the profiles compared to retrievals where only lower frequencies were used.

General comment

The paper is generally well written and complete in its explanation, my main question is of a general nature and goes back to the broader issue of the advantage of using radiometers vs. reanalysis for profiling. As the authors state in the introduction current reanalysis have difficulties in accurately representing the Arctic winter atmosphere but, in this context the paper does not shed any light on whether this combination of frequencies provides an improvement over the reanalysis. For example, in figure 5, rather than seeing errors related to MWR-MOSAIC and MWR-ERA5, it would be better to show MWR-MOSAIC and ERA5-MOSAIC. This way we can see if the retrievals offer a better representation of the low troposphere in the Arctic compared to ERA5.

Similarly, looking at Fig. 10 b, it appears that ERA5 still does a better job than the synergy to represent the humidity vertical structure. Therefore, my question is, why not just use ERA5? It appears that even after all the improvements shown in this work, radiometric profiles are still not good enough to provide useful vertically resolved humidity. Are they just a smoother version of the reanalysis dataset used to train them? Or they actually improve on the reanalysis biases?

Thank you

This is a well-written paper that is easy to follow and clearly articulates its goals and its approach to investigating them.  Overall, this paper is appropriate and suitable for publication in AMT after some relatively simple issues have been addressed. Unfortunately, I feel that this includes some small data analysis which raises the threshold from minor to major revisions.  Still, these revisions should not be difficult to implement.

Line 191:  The authors state that they use NNs "because they can deal better with the nonlinear relationship between IWV and TB measurements in the G-band." I feel that this statement needs more supporting justification.  What is this comparison being made to (i.e. NNs are better than what other technique?) Why and how are they better? One of the chief advantages of NNs is that, after they have been trained, they are computationally very quick to perform retrievals as compared to physical/OE methods. However, physical methods have a significant advantage in that they are better suited to retrieving observations of conditions that are not represented in the training dataset, which could be important in a changing arctic. I am not suggesting that the authors redo their entire analysis with a brand new retrieval method; however, I feel that they need to more strongly justify why they chose this method.

Line 210: Why was this grid chosen? There is growing awareness that the selection of a retrieval grid has a notable influence on DOF and vertical resolution (see Loveless et al. 2023 doi:10.1109/IGARSS52108.2023.10283250). Did you assess any of the influences that the grid choice may have had on the results?

Lines 265 and onward: I am approaching this problem from the perspective of an observationalist, not a modeler or a strong user of ERA5. With that in mind, it seems intuitive that an NN that is trained on ERA5 will produce profiles that show stronger agreement with ERA5 than with an external data source not just because of the uncertainties of the radiosondes or the spatiotemporal matching issues (though those concerns are important) but because the NN is implicitly including all of the known biases and uncertainties of ERA5. To me, the comparison to radiosondes is more interesting and important than the comparison to ERA5, but it is that reanalysis comparison that gets the stronger focus of attention. One step forward may be stronger justification of why the ERA5 comparisons are important. 

Line 305: It sounds like the radiosonde analysis was performed with respect to sondes that had been interpolated to the retrieval grid, and not smoothed to account for the influence of weighting functions.  Given that you have averaging kernals for a random subset of the cases, it would be possible to vertically smooth the sondes. This would permit an analysis of the bias and RMSE relative to sondes that would remove the influence of the large vertical gradients and provide a more direct comparison between the statistics calculated relative to sondes versus those from ERA5.

Lines 411 and onward: the DOF is an important quantity for evaluating the information content of a retrieved profile.  Also important is the vertical resolution, which is easily obtained by scaling the vertical grid spacing by the inverse of the averaging kernal. As you already have these pieces of information, a small discussion of the changes in the vertical resolution brought on by the synergy would be achievable and relevant to this paper's aims. 

Line 432: Why did you opt for radiosonde pressure instead of a hypsometrically-derived pressure profile from the T and specific humidity profiles? This would allow for a self-contained measurement of RH without relying on the relatively-infrequent sonde observations.

Line 448: The paper may be stronger if this particular section is removed. As it stands, it introduces a somewhat counterintuitive example (the synergy makes things worse) then promises to investigate further in a future paper.  It may be best if the entire discussion is held to that future paper so that it can be investigated in detail and the nuances can be explained.

Line 486: Here is a good example of why the inclusion of a vertical resolution discussion is important. Improvements are seen in the lowest 1500 m, but I'm guessing that even with the synergy and the non-zenith stares, the true vertical resolution is on the order of 1500 m or worse given the broad weighting functions of the microwave band. Noting the changes in the vertical resolution would help better attribute the causes of these improvments.