My initial impressions is that this is a good article and that it should be accepted to this journal with some revisions.

The authors have looked at a multi-fidelity algorithm that uses additive surrogates based correction models. Furthermore, the approach does not require gradients from the high-fidelity model. The author made these choices so that the method is general purpose. The authors then applied this method to a blade design problem, controller tuning problem and finally a wind farm layout optimization problem. In all cases the authors were able to show that the multi-fidelity method was able to obtain a better solution (according to the high-fidelity analysis) than single discipline low fidelity optimization. While at the same time solving these problems faster than the single fidelity high fidelity optimization. The article is interesting and relevant work to the community and has been well written. However, there are some areas for improvement and I feel that some revisions can help make a stronger manuscript. Those are discussed as follows:

First, I have concerns about the multi-fidelity algorithm that I think the authors should comment on in greater detail. The first point of concern is that none of the multi-fidelity optimization solutions matched the high-fidelity solution. This suggests that at the limit the corrected low fidelity problem is not equivalent to the high-fidelity solution. Some corners of the multi-fidelity literature try to show that the corrected multi-fidelity algorithm is equivalent or not. It would be good for the authors to comment on whether their algorithms tries to achieve this or not. This lack of equivalence is obvious in the design solution for the blade optimization. For readers and engineers that are more concerned with the final design configuration than the objective, it would be helpful if the author could comment on that and explain to what extent the surrogate correction played a role in this discrepancy.

Also I think that this shortcoming is related to convergence of the surrogate to the high-fidelity model. The authors show some plots that qualitatively show convergence. However, it is small differences in the gradients at the limit that ultimately define an optimal solution. I think that it would be good for the authors to show quantitatively the convergence rate of the surrogate model at the optimal high fidelity optimization point for both the solution and the gradients.

Furthermore, surrogate based methods suffer greatly from the curse of dimensionality. Despite this, I was quite pleased that the authors were able to obtain good results despite this typical weakness. To overcome this weakness the authors explain how they used a specific type of surrogate that internally performs model reduction to overcome this dimensionality. I personally would like to know more details about this. For example, do many dimensions cause the surrogate longer to start converging? How many high-fidelity evaulations were needed before good convergence was achieved? Maybe more details on the internal model reduction methods used in their surrogate would be nice.

At the end of the article in figure 10, the authors summarize the different cases purely in terms of computational savings. However, I think different readers are more interesteed in a range of metrics. Since none of the multi-fidelity solution achieved the same performance of the pure high-fidelity, I think that you also need to include this in the summary. Furthermore, sometimes it's not the objective value that matters, but the final design configuration. Thus, I think that it's also important that you discuss limitation. Finally many readers would consider the distance in the design space from the high fidelity solution an important metric. So I would also show this in the summary. Most would see the objective and final design configuration more important...

Concerning the cited literature, you have many citations of the work within wind energy. However, I think you could give a better summary of the different contributions in surrogate based optimization. This is an old topic in the aerospace community. There are a mix of different methods that have a range of speed improvements and accuracy in terms of reproducing the high fidelity solution. It would be good if you could present these different methods and explain how your work fits with these.

Another comment is the choice of problems. The IEA Task 37 has created various optimization test cases for this purpose. It would be better for the larger community if your multi-fidelty was applied to these test cases so that the community can see how your results compare with the larger community results. This is of course a big task ... but maybe this is something to consider for future articles.

In your future work discussion, I would elaborate a bit more on how this method could be improved. The multi-fidelity literature in aerospace shows faster optimization with better correspondence with the pure high fidelity solution. You made choices to improve generality and applicability, so this not a critism, but I think in light of other methods it would be good to comment on what would be needed to get closer the high fidelity solution faster.

This manuscript talks about multifidelity optimization in wind energy. Now "multifidelity optimization" is a large umbrella under which many different approaches fall - things that people have been doing for years in order to speed up their optimizations, but also more recent ideas how to do this in a more structured way. The manuscripts starts with an informative review of some of the literature and recent domain-specific works that explicitly considers the multifidelity optimization idea, and this contains useful references and is valuable. The authors then quickly describe their method of choice, which is the use of trust-region optimization with a "Kriging partial least squares" surrogate model. Unfortunately no details at all are given about their particular algorithm used and what parameters or other choices have been made. This is a major weakness of this paper, especially so since trust-region optimization is not something that most people will be familiar with (even though it is extensively used with highly nonlinear problems). The used surrogate model is a modern variant of the Gaussian process model, which has received much attention recently (especially in wind energy applications), due to its intrinsic ability to estimate the uncertainty in its predictions. I would assume that this feature will be used when performing trust-region optimization with it, so this should be explained and discussed. There are many interesting options and ideas here that would interest readers, and generally these methodological details are more interesting than the shown example studies and their results. In fact, as long as the method is not fully clear, the results are not that interesting - to a scientifically-minded reader. One might be inclined to forgive the authors this lack of details here given the typical space requirements of journals (and the cost of publication), but without this information the article reads more like an advertisement for the authors' method than what was probably intended. At the very least a reference to supplementary material (e.g. in the form of a non-peer reviewed technical report - this is a prime example of why we still need and should value this type of publication!) should be given where readers can find all implementation details.

The manuscript continues to discuss three example studies where this approach - correcting a low fidelity simulation with a meta-model based on few evaluations of a high-fidelity simulation - seems to perform somewhat better than either optimization based on the low-fidelity model alone (being more accurate than this) or when using the high-fidelity model alone (being faster than this). Again, it is unfortunately completely unclear how the optimization with these models (only the low- or high-fidelity ones) is performed!

Somewhat surprisingly, the high fidelity models used in the case studies have only slightly higher complexity than the low fidelity models. For example, in the blade design study, both models are based on blade element momentum (BEM) theory. The low fidelity model is using the original steady state BEM approach, whereas the high fidelity model is using its unsteady generalization. For this pedagogic demonstration this is sufficient, but in real applications one would expect to use some kind of CFD simulations as the high fidelity model, I assume?

All in all, the topic is relevant, and the authors seem to have some success with their approach. If now they could only describe the approach with enough details that readers could reproduce it (or learn from this paper how to do something similar with their own wind energy optimization problems), then this would be a very nice contribution to the literature.

Major comments

1. The authors need to explain the details of the trust-region optimization and KPLS surrogate model used, in sufficient detail that readers can (potentially) reproduce the results in the paper.

2. page 4: "This approximation is devised such that it is equal to the high-fidelity model at the points where we have high-fidelity data." - So it is what we would call an interpolating approximation. Note that this property is not strictly necessary (although it usually makes sense).

3. Fig. 2: It is unclear how the trust region is changing between iterations.

4. Fig. 3: There seem to be samples missing in the figure (e.g. in Fig. 3a, the function depicted is not piecewise-linear between the sample on the far left and the cluster of samples in the right third). Also, are all three figures based on the same number and location of samples from the high-fidelity model?

5. In the blade design study, the low-fidelity optimized blade must be cheaper (less materials used) than the high-fidelity or multi-fidelity optimized blade? By how much? The comparison (Table 2) would be more informative and convincing if the optimization was using the same route (e.g. by enforcing only increases in twist and chord?) instead of converging on completely different paths (toward completely different local optima).

6. How the method works is intuitively clear in 1D (second case study), but what happens in higher-dimensional cases? The first case study is 7 dimensional, the third one is 14 dimensional. How does the approach work with such high-dimensional data, what are its limits or challenges with it? For example, how are the surrogate models initialized, and does this effort not become too computationally expensive in higher dimensions?

7. Fig. 6: Abbreviations of what simulation outputs are shown should be explained. Not every reader is so familiar with OpenFAST software that they will immediately understand what they see here.

8. page 6: "By not using simple additive or multiplicative factors ..." - This seems to suggest that only the last example uses a surrogate model as correction function, and the first two case studies use something simpler? Or is it just this statement that is somewhat misleading?

9. Where is the data availability statement (required by the journal)? Where can readers find the code and the data used for the results in this paper?