General comments:



Manuscript by Paredes et al. titled "rsofun v5.0: A model-data integration framework for simulating ecosystem processes" describes an R-package that is mainly built around the P-model for efficient simulations and model-data synthesis workflows. Authors present a common use case where they calibrate the model with a single data-stream using a Bayesian approach and propagate the calibrated parameter uncertainty to model outputs.



The objectives and the rationale of the study are clearly stated, and the manuscript is written in a clear and concise manner. I appreciate the efforts of the authors and developers for this development towards reprudicle modeling results and a lower bar of entry to vegetation modelling. 



While I would like to highlight that the reported development is in the interest of the GMD community, I found the writing style of the paper a bit different than typical GMD papers where many scientific and technical details are omitted with referrals to the vignettes. Below I commented on the manuscript parts that could be enriched, and I ultimately defer to the editor's decision but I suspect that it currently is not aligning quite well with categorization as a GMD "model description paper" which are expected to be comprehensive, detailed, complete and rigorous. Please consider supplementing the manuscript with deeper details and discussions. 

Specific comments:



P2.L36-39: The strength of Bayesian approaches in combining information from multiple sources and scales could also be emphasized here. Also, van Oijen (2017) could be a nice addition to the citations.



P2.L59: Would the authors consider rephrasing this sentence? Currently, it reads as if a novel solution is about to be presented whereas Bayesian calibration of a process-based vegetation model is done more times than I can count. Perhaps try presenting it as an application or an implementation of an existing solution. 



P3.L67: It's a pity that the paper doesn't present this more sophisticated, and perhaps more realistic and valuable setup. 



Table 2: Could the parameters that were held fixed for the calibration be marked with a different character than asterisk (*) since soilm_thetastar and kc_jmax also have an asterisk in their symbols? Or the table caption could be modified to read "... marked with an asterisk in the last column"



P6.L34: Could you elaborate as to which functions are included in this set and why/how they were chosen?



P7.L44: As GPP is not a measured but a derived quantity, could you please mention the approach used to derive it? FLUXNET2015 dataset is fairly well-documented but for the sake of being more complete, please also add how the GPP values were gapfilled and aggregated to daily values (which I assume is the model time step).



P7.L61: Undermined how? Could you please elaborate on how the convergence was affected so that the readers can follow and apply the same logic in their applications when needed? Was there a strong correlation structure? Were the chains getting stuck? Was the result the same with different algorithms? Different chain lengths? The text also mentions before that kc_jmax and beta_unitcostratio (P3.L83) have previously been calibrated separately and fixed in this study, which was somewhat agreeable, but now saying that you decided to hold them constant because calibrating them with other model parameters undermined convergence is confusing. Please clarify and reconcile. 

P7.L62: Also, on Table 2 (because they're fixed) no range to vary them was given for those parameters that were held constant in calibration. But I believe for them to appear in the sensitivity analysis (Fig 2) you varied them in some range. In fact, I can see that in the vignette these are specified, but please include those ranges also somewhere in the manuscript for clarity and reproducibility. 



Figure 2: Would it be possible to add the symbols to the figure? Because the paragraph right before (P7.L56-62), refers to these parameters with symbols while the figure refers to them with parameter names which requires the reader to go back to Table2 to do the mapping. Admittedly it's a small list, but I suspect it wouldn't be difficult to add symbols to the figure. Same goes for Figure 3.



P8.L75: 24K iterations sounds like an interesting choice. A more typical number would be, for example, 50K or 100K. Could you please explain in the text how you decided on this number (since the goal of the package and paper is to lower the bar and provide guidance to the audience)?



P8.L77-78: I really appreciate the vignettes but I feel like some of the results reported there should go together with the paper (e.g. in the supplement). I think vignettes are a lot more practical and ultimately more useful to the end users, and I wouldn't object if this was an open source scientific software journal but I expect GMD papers to be more complete. For example there is a kphio_par_a and kphio_par_b correlation discussion in the vignette which is completely missing from the paper.



P9.L87: (from here onwards, including Figure 4) I believe there is a mix up in the way "model error" term is used. Perhaps the authors meant "error uncertainty" instead of "model error"? It is true that the credible interval is solely concerned about the uncertainty in the model parameters. Predictive interval here, however, is concerned with the overall residual error between the model and the data. In other words, the way the error term was jointly fitted in the calibration makes it intractable to decompose this term into data and model error. It is also known to dominate and cause overestimation of predictive uncertainty. Please refer to the relevant literature and revise (e.g. van Oijen, 2017). 



Results - I was a bit surprised to see no quantitative reporting of the model improvement after the calibration. The only visual comparison (Figure 4) reports the posterior performance with no reference to prior performance, and only for one year. While showing a single year is useful for practical purposes, please consider providing results for all years (e.g. in the supplement):

- It would be interesting for the readers to see how performance in different years compare, also with quantitative metrics.

- Furthermore, even though measurements from years 2013 and 2014 were deemed problematic, it would be good to show what the calibrated model predicts for these out-of-sample years that the calibration did not see.

- Last, but not least, there was no mention of further posterior predictive checks/diagnostics. One can for example plot residuals against predictors and employ formal tests that measure where the observed data falls on the distribution of simulated data. Is there a pattern in the residuals? Were the correct distributional assumptions made? Comparing credible and the predictive intervals is only a (very) limited part of the story.



Discussion - Please consider enriching the discussion section with the following:

- Were the design choices adequate? I mean all the decisions from coupling the model with BayesianTools as a package to the prior and likelihood forms you selected?

- What does it take to transfer this calibration to another site/variable, multiple sites/variables?

- How do you recommend iteratively updating these results as new data becomes available (either more of the same type of observations or with new data sources)? Does your implementation allow re-reading its own outputs as inputs?

- What are the other limitations of this study? What are your outlooks?

- While I understand that it is not the main goal of the paper, the inexperienced readers could benefit from pointing to the literature that provides guidance on the peculiarities of model calibration, e.g. see MacBean et al. 2016, van Oijen 2017, Oberpriller et al. 2021, Cameron et al. 2022. Therefore, in addition to points above, please also consider adding some discussion along those lines since the section is rather thin at the moment. 

P11.L13: "complementary observational constraints" such as?



P11.L16: Please elaborate some more. Is it a weakness of the study that FvCB parameters were kept constant? What was the reasoning? Is it recommended to do so? What are the expected challenges there? Are there future plans to address this?

MacBean et al. 2016 https://doi.org/10.5194/gmd-9-3569-2016

van Oijen 2017 https://doi.org/10.1007/s40725-017-0069-9

Oberpriller et al. 2021  https://doi.org/10.1111/ele.13728

Cameron et al. 2022 https://doi.org/10.1111/2041-210X.14002

The authors presented a Simulating Optimal FUNctioning {rsofun} R package, aiming to lower the bar of entry to ecosystem modelling and model-data integration for scientists. This R package providing the potential of efficiently model parameterization and estimation of uncertainty. The authors tested the {rsofun} R package at the site level by applying parameterization to the P-model and presenting a corresponding case study. The results demonstrate that this tool can be used to calibrate photosynthesis-related parameters of the P-model at site scale and to simulate GPP accordingly. The package shows potential to advance research in ecosystem modeling. However, further development is still required. The case study presentation lacks sufficient detail regarding key aspects of the tool, such as how model–data integration is implemented. Moreover, the core functionalities of the tool appear to rely heavily on other existing R packages. The authors are encouraged to clearly identify the unique contributions and core features developed specifically within this package—particularly whether it provides a generalized interface to invoke different ecosystem models. Overall, the {rsofun} R package is a promising tool to support ecosystem modeling studies, but it remains incomplete in its current form. Its potential for application at regional scales is still uncertain. Therefore, I recommend substantial revisions before the manuscript can be considered for acceptance.

Below are detailed comments:

[line number]: Please using continuous line numbers.

[Page 1, Line 11]: Whether site-scale simulations or parameterization need parallelizable computation？How about applying this package at a larger scale？

[Page 2, Line 58]: There is no table before this table, so please rename it as Tab. 1.

[Page 4, Line 08]: Was net radiation an output in P-model？and it is not suitable to list soil temperature  in ecosystem water balance.

[Page 4, Line 14]: There is no information about spin-up period or corresponding years in Table 1, please move this sentence to corresponding section.

[Page 7, Line 41]: Why the author selected this site, and as rsofun package so computationally efficient, why not test at more sites that can represent different conditions. What is the depth of the root zone soil at this site?

[Page 7, Line 60]: If the authors mentioned analysis, please give the result in manuscript or supplementary materials.

[Page 7, Line 62]: When all parameters are used for model calibration, the multidimensional parameter space increases exponentially, potentially leading to model non-convergence. This also indicates that different values of c*, beta, tau and b0 can influence the parameterization outcomes. Meanwhile, the fixed parameters should be explicitly justified in terms of how they were determined.

[Page 8, Line 71]: Given that c* is the third most sensitive parameter, why was it not included in the calibration, while β0, which ranks second to last in sensitivity, was? The authors are requested to further clarify the rationale behind the parameter selection.

[Page 10, Line 90]: It is hard to recognise the light green band, and please explain more about the model uncertainty and parameter uncertainty.

[Page 10, Line 99]: It's dark orange or red line, the authors described as red line before.