|The manuscript by Belikov et al. has been substantially revised. As requested in the initial review, the authors provided clarity in several areas, and expanded their evaluation of the forward model using additional observations. Overall, I find the revision to be a significant improvement. That being said, there are still a few rough spots that could use some clarification of the writing, where things are either not clear or simply vague. Also, the paper requires proof-reading and editing for grammar throughout. I provided a few comments for the abstract and introduction, but stopped after that point. |
16: The authors did not comply with my previous request to be quantitative in their abstract. The state “However, we do not consider it is necessary to include any numbers in the introduction.” This isn’t the introduction, it is the abstract. As such, quantitative assessments of “improvements” are not only hight recommended but essentially a requirement. I strongly suggest they reconsider, if they want their work to be relevant and useful.
4.25: It would be more correct to say high dimensional inverse modeling.
Eq 1: This is difficult to read. Try putting the summation indices above and below the summation sign, rather than to the right. Use a horizontal bar for fractions rather than forward slash.
Section 2.1. Both reviewers requested clarification on the time-coupling described in this section. Despite the revisions, unfortunately it still isn’t very clear. Could the authors consider some sort of schematic or equations to explain, given that the writing is a bit hard to follow? For example, statements such as “add contribution from surface sources” are vague.
Figs 2-5: You might refer to Table 1 in the figure captions for definition of the cases.
Fig 2-4 and associated text: The use of additional measurements to evaluate the forward model is appreciated. However, the discussion and analysis of these results is a bit lacking. Why are is correlation degraded at some locations (e.g., PRS)? Why is the correlation better for some sites than others? Statements like “Most significant improvement” on 11.13 are quite vague. Please be more quantitative. Other statements aren’t really correct, such as 11.21: The improvements….are higher than those obtained by increasing the resolution… It’s easy to find cases that contradict this statement, for example there is a 50% reduction in STD between Case 1 and 2 for the ALT site (Fig 4a), but no apparent improvement between the Eulerian-only and coupled mode (Fig 4b). Recognizing, vaguely, that there are issues at some sites in the final paragraph of this section (11.28 - 12.3) doesn’t really count; try merging this content into a singe cohesive, precise, discussion. I’m not arguing that the coupled mode isn’t better overall. Definitely it is! It’s just that the authors need to be much more precise and careful with the writing and discussion of their results for a scientific paper.
13.8: Another potential drawback is that discrete adjoints of nonlinear advection routines have been shown to have poorer performance for 4D-Var optimization than the continuous adjoints (Liu and Sandu, 2008, doi 10.1002/Fld.1547).
15.29: In my understanding, checkpointing or recalculation of variables for an adjoint model are needed when nonlinearities arise. Yet your model is linear, no? So what is it that needs to be checkpointed?
2.16: reproducing —> reproduction
3.15: an —> of an
3.17: using for the first time of —> for the first time using
3.18: use —> using
3.19: develop —> , and development
4.4: effective —> effective in (or effective at)
4.4: remove “of”
4.5: The Lagrangian —> Lagrangian
4.6: tracks —> track
4.13: with set —> with a set
4.16: gradient —> gradients
4.20: data assimilation —> assimilation
5.4: adjoint, that —> adjoint that