This manuscript introduces a new multi-frequency ice particle size distribution retrieval and evaluates the algorithm against airborne in situ observations during the Characterising CiRrus and icE cloud acrosS the specTrum-Microwave (CCREST-M) that took place in 2024 at the Chilbolton Observatory in the UK.  The paper is well written, and the analyses support the findings.  The subject is appropriate for Atmospheric Measurement Techniques.  My primary concern is the length of the manuscript.  This arises because the paper includes a detailed description of the CCREST-M campaign/measurements as well as the algorithm (which has several elements) and its evaluation using multiple diagnostics from three separate case studies.  While this has the advantage of presenting all this information in one place, it results in a very dense manuscript that was somewhat challenging to read.  Though this is not a critical flaw (I believe there are no strict word/page limits), I encourage the authors to explore opportunities to make the manuscript more concise.  For example, there is some repetition in the results and explanations of algorithm behavior that could probably be shortened.  Since this is just a suggestion, my recommendation is to accept with minor revisions.

1. The abstract could also be more concise. Some of the background in the first paragraph could be left for the Introduction and some of the details in the 3^rd paragraph could be consolidated into a more concise description of the results.
2. The authors have decided to combine the description of CCREST-M and the algorithm evaluation into a single manuscript. While I think I understand the motivation for this choice, these topics could probably be covered in two distinct, concise, and equally interesting, papers.
3. The results (Section 5) are very thorough, but somewhat repetitive. For example, many figures are repeated for different variables, temperature ranges, and different cases and several points are reiterated multiple times.  I recognize that each case illustrates a distinct aspect of algorithm performance, but I wonder if the results would be more impactful if they were conveyed more succinctly, perhaps consolidating the complementary findings from individual cases/comparisons where possible.  Also, could some of the figures that represent variants of the same plot be moved to an appendix or supplementary material?
4. The labels on several figures are very small (e.g. Figures 4, 5, 8, 9, and 10), some to the point of being illegible (11, 12, 14, 17, 20 - 23).

Haoran Li lihr@cma.gov.cn Chinese Academy of Meteorological Sciences This work presents a retrieval algorithm that integrates machine-learning-derived prior constraints with dual-frequency radar inversion. The 3 GHz radar observations provide prior information for the machine learning model, whereas the 35 GHz and 94 GHz radars are adopted for subsequent microphysical parameter retrieval. To the best of my knowledge, CCREST-M data is very valuable from the perspective of aircraft validation to triple-frequency retrievals. The results were compared to aircraft observations and G-band radar observations. I do not have critiques on the presented results, while I have a feeling that this algorithm could benefit from a direct triple-frequency retrieval. Since triple-frequency radars were practically used in this work, it remains unclear why the synergistic retrieval of 3 GHz, 35 GHz and 94 GHz radars was not adopted directly. If the value of single-frequency plus dual-frequency retrieval is well supported, I would be very happy to see its publication. Major comments, 1. As discussed above, I am very confused that you did not use a triple-frequency approach. Alternatively, is it possible to use the Ka-band or W-band radar in the ML model? Then, you do not need the third frequency. 2. The adequacy of using a single ice type. It seems to be a bold assumption to me that a single ice type was used in the radar retrieval. I understand that the CPI imagers suggest the presence of ice rosettes, but the sampling area is very limited compared to radar observations. I would encourage a thorough discussion on this limitation. 3. Redundant figures for PSD comparisons. It is recommended that in-situ observations, ML predictions and dual-frequency radar retrievals be integrated into a single panel. In this way, direct and intuitive comparisons among PSD results from different methods can be easily conducted. 4. Similarly, Tables 2,3,4 should be integrated into one table for a direct comparison. The same to all the PDF plots of different moments (e.g., fig9&10). 5. ML predictions for PSD moments were validated to in situ observations, but the dual-frequency validations are missing. 6. I like the G-band validation part. Since the G-band radar was collocated with other radars, I would recommend a long-term validation. I believe it is very handy to implement. Some technical comments, 1. Figure 2. Mark the periods where the aircraft validations were made. 2. L204. Liquid clouds are not uniformly distributed. How did you do the liquid attenuation correction with LWP? 3. L212. L267. Looks conflicting. You definitely need gaseous attenuation for Ka- and W-band radars. 4. L210. It should be 1 dB. In addition, I would not say ‘much less’ than 1 dB. We recently compared different parameterizations, and it is sub-dB difference. Li, Q., Li, H., Sun, X., et al. (2026). A survey of snow growth signatures from tropics to Antarctica using triple-frequency radar observations. Atmospheric Chemistry and Physics, 26(2), 1249-1264.