See attachment.

The paper presents a machine learning framework to model individual glacier thicknesses based on global glacier inventory data, supplemented with other data. The paper is motivated by differences in estimated ice volume between current physics-based models. The results obtained in the paper are promising and appears to improve the ice thickness estimates compared to the existing models. Estimating the World’s total glacier volume is relevant to projections of ice mass loss and sea level rise, and making use of machine learning to address this task is interesting and timely.

However, while the results are promising, the presentation needs to be improved. The method is not sufficiently explained but it is clear that it builds on knowledge from previous work and physics-based knowledge. The results needs to be systematically assessed and limitations discussed in more detail. Including these comments and arguments will increase confidence in the results and improve the impact. These points are explained below.

I have several concerns with the paper:

• The selection of features used to train the IceBoost are not substantiated and explained. The datasets used to produce the features are from well know data repositories or data products, and clearly referenced, but it is not explained how the features in Table 1 were chosen. For example, how are the three aspect features defined, and what is the basis for the selected smoothed slopes? Were other smoothings and other variables investigated, but not included in the analysis in the end?
• Furthermore, the main text does not define or describe the features, but the reader is referred to Appendix A and B. This makes the presentation of the model to appear rather superficial and makes it impossible to read Table 1, as well as understanding the discussions later. I suggest that some of Appendix A and B is included in the main text, so the presentation of the model appears clear and self-contained.
• This brings me to the next point, which is that the physical basis for the selected features is not described or commented on. I acknowledge that the machine learning method is not based on physical relationships, but the selection of the features are building on knowledge of physics-based relationships with ice thickness and volume. I miss some comments about the relevance of these features. Some may be obvious, like slope, distance to margin, etc., but others are not obvious, and other obvious parameters are not included, like e.g. precipitation. In this context, I also miss some references to previous works of what are the key parameters, in addition to the references to the textbook and the review of scaling relationships by Bahr et al. For example work by Oerlemans (e.g. summarized very briefly in Oerlemans (2005): Extracting a climate signal from 169 glacier records. Science 308, 675-677, https://www.science.org/doi/10.1126/science.1107046 or more elaborated in Oerlamans’s online booklet: https://www.cambridge.org/core/journals/journal-of-glaciology/article/j-oerlemans-2011-minimal-glacier-models-second-edition-utrecht-utrecht-university-igitur-utrecht-publishing-archiving-services-90pp-isbn-9879067010221-paperback-available-free-from-joerlemansuunl-and-as-free-download-from-httpigiturarchivelibraryuuni/A49EB2866D3E6E38040AECFD7B3E24AA .
• It is later discussed in section 3 that some features are not important for the results, while other features are. In this discussion, it would be helpful to include an assessment of how this analysis compares to the previous work, i.e. whether the same features come out as important, or other parameters are more important in the work here. It is for example expected that slope is important. It is not a surprise that ice velocity is less important, but it is indeed very interesting to have the importance of including velocity quantified. This should be clarified and assessed in relation to previous knowledge. I also miss some discussion of why velocity is only important in high-latitude regions, not elsewhere.
• The assessment of the model results, including comparison with other results and discussion of any limitations/difficulties appears sporadic. Perhaps because it is distributed over several sections (2.4, table 2, section 4). The quantitative assessment is summarized in Table 2, but only referred to in the methods section, and the application section is very qualitatively focused on the examples. The discussion in section 4 is very interesting, but could perhaps be summarized into a more general discussion of limitations/advantages of the method. It would be very helpful/interesting to see examples where the model do not perform well. Overall, I recommend that the assessment of the results and discussion of limitations should be more systematically organized to become more convincing.
• Another problem in this context is that the figure captions are generally deficient, in particular figures 3, 4, 5, and it took me very long time to understand what they actually show (see below).

Specific points:

• Section 2.4, lines 146-149: Explain why some regions are similar, so evaluations can be indicative in these regions.
• Appendix B2: I don’t think it is clear why missing glacier aspect data are substituted with zero. Please explain.
• Figure captions: all figure captions need to be reviewed and checked for completeness and clarity. Add letters a, b, etc
• Figure 3 and 4: it was not clear that the two models in lower row show the same results as in the upper row, but now with the observations added. Please clarify that observations are added in all three figure in lower row. This will help underline how much better the IceBoost is in these examples.
• Figure 5: It is not clear what big and small circles mean. Only by zooming in, it is possible to distinguish. It is not explained what the background image is.
• I suggest to merge section 4 and 4.1, as section 4.1 adds to the impression that the assessment in the main paper is lacking.