|The authors followed the suggestion of including multiple parameter estimation experiments (started with different parameter values). I welcome this addition, yet I would like to see more results included from these new experiments.|
# general comments
It is great to see that the authors repeated the parameter estimation experiment with different initial values. However, the results could be presented better and should probably be given more space. First of, the repeat experiments could already be mentioned and motivated in Section 2.4. Currently they are only presented in a relatively short passage in Section 4.1. Here, some statements create more questions than answers: "this was achieved by similar subsets of the optimized parameters" (line 397) Which subset of parameters remained the same, which only in a few experiments?
Apparently, the different experiments had similar results according to this passage "[...] and there was up to 76% of the reduction in the model-observational misfit (vs. 58% of the reduction in the reference case; Table B1) These results suggest that no matter where in parameter space the optimization started from, the adjoint/optimization scheme took the model cost function to similar local minima. (line 395)" Yet, a similar reduction in the cost function value does not imply similar parameter values.
In the current version, Table 1 paints a nice picture with uncertainty intervals for some parameters that are locally derived from the Hessian matrix. Having multiple cost function minima with different uncertainty intervals (and different parameters that are optimized/constrained) may distort that picture a bit and the presented uncertainty intervals may just be representative for 1 out of 16 experiments. Because this information is not available to the reader, it is unclear how generalizable the results in Table 1 are with respect to the other 15 experiments. A few more details/results to clarify, and maybe some discussion would be very helpful here.
In the updated manuscript, it is now much clearer how parameters are "removed" from the optimization. I am a bit surprised that removed parameters are abruptly reset to their initial values. Given the strong correlations that may be present between the parameters, a sudden reset of parameter values after several iterations could lead to "shocks" in other parameters. Or is the estimation restarted with all parameters starting from their initial values and fewer included in the optimization?
Section 4.3 is currently divided into two paragraphs. One is listing the changes brought to ecosystem indices by the optimization, without indicating if these changes are an improvement. The second paragraph then compares the model output to data and results from other studies and only briefly touches on the changes brought by assimilation. Here, it would be really useful to mix the two paragraphs and report the changes with reference to the data that is available.
# specific comments
l 29 "we discuss fully potential underlying reasons": Sounds a bit convoluted, maybe delete the "fully", also because it is difficult to claim an exhaustive discussion.
l 53: "its strength": I would suggest that there are multiple, changing it to "its strengths".
l 75: I think the first "dominated" in this sentence refers to large phytoplankton only, and the second one to smaller ones but I still think that two "dominated" are a bit confusing here.
l 142: "In principle, optimization should be able to capture the elevated diatom Chl by adjusting free parameters unless: 1) the right parameters are not adjusted and/or the baseline (non-optimized) parameters need significant adjusting, and/or 2) the model equations are not adequate even with the optimized parameters." What if the nutrient initial values are too low, would errors in the state estimates be a third option?
l 145: I am not sure if there has been much evidence for it in the WAP region but could their thick shells be a reason for less preferential grazing on diatoms?
l 250: "or estimated using a subset of the observations, without examining the effects of the initial parameter values on the model results prior to optimization": It's not clear how this should work. A subset of observations is used but the effect of the parameter values is never examined? What is the subset of observations used for then?
l 258: "with one parameter per each state variable, the change of which yields the largest decrease in the total cost function": How was this determined? Was the one parameter per state variable variable put in place first or did it turn out that the parameters yielding the largest effect, were one for each state variable?
l 271: "If parameters are optimized to ecologically unrealistic values, they are kept back to the initial parameter value": Even if they have undergone some changes in the previous steps, they are reset to the initial values? If so, could this have an effect on the the other parameters which may be correlated?
Eq 6: It would be good to clearly state the difference between the mean that is used in the computation of CV and the climatological mean that multiplied with it.
l 300: "J equivalent to J/M hereafter": Why not introduce it immediately?
l 317: What about increased wind-driven turbulence as the ice disappears, is this a concern?
l 319: "Also, because our model simulates only the spring-summer growth season, winter sea-ice growth is less of a concern.": Use a different term for sea-ice growth or change first instance of growth to something like "phytoplankton growth", so that phytoplankton growth won't be confused with sea-ice growth in this sentence.
l 331: I know that I had a question about this in my last review and I still think it should be explained better or just made more explicit. "Initial conditions are prepared by first optimizing the full growth seasonal cycle forced by climatological physics and assimilated with climatological observations and with the same bottom boundary conditions used in the optimization of the 2002-2003 growth season"
I think it should be pointed out here what kind of optimization is performed. Talking about the initial conditions, one could assume that optimization implies state estimation here, i.e. adjusting the initial conditions directly. However, based on the comments to my question, it appears that parameters were estimated for a climatological simulation which was then used to create the initial conditions. But where do the initial conditions for that climatological simulation come from? I think my problem is that I don't still understand what exactly "first optimizing the full growth seasonal cycle" really means.
l 390: "presented in the manuscript": Change to "presented above".
l 471: Is the decreased or increased (for NCP and POC) correlation realistic?
Fig. 4: Use the same coordinate system in both plots. Preferably, combine both plots into one, with different symbols for prior and posterior solution and different colors for the different observation types.
Fig. 6A/B: Join into the same figure, just like Fig 5B.