This work presents a deep Learning framework to predict surface chlorophyll concentrations and anomalies across large marine ecosystems, with implications for fish catch forecasts. 

The abstract and the introduction well describe the core idea and its fundamentals: the interactions between Earth’s physical and biogeochemical fields is important for predicting future climate and the marine biogeochemical variability is critical to advance climate predictions based on bio-climate interactions. 

Nonetheless, the methods and results section can be deeply improved in order to clarify the purposes of the research, its development and its scientific novelty. 

The Method Section offers a detailed overview of the deep learning architecture, together with datasets information, sensitivity analysis and prediction performances information. Despite that, the description of the architecture lacks some important details, and the dataset description, though comprehensive, is presented in a confusionary way, without properly describing the variables collected and their role in training-validation-test phases, a fact that reduces readability and reproducibility. 

The architecture developed for this project is a Convolutional Neural Network. Despite the authors having dedicated a paragraph of Methods Section and a paragraph of Results section to the description of the architecture, several fundamental aspects remain unclear—particularly the dimensions of the input and output data—reducing the clarity of the project’s objectives and implementation.

The research question behind this project and consequently research purposes (i.e. the relevance of modeling mean chlorophyll within LMEs and the rationale behind using entire 2D maps to derive a single pointwise mean value for each LME) appears confusionary, a lack of clarity that is also reflected in the results. It is not particularly clear the task the manuscript intends to solve, and in particular the objective of some experiments (i.e. mechanisms underlying chlorophyll prediction skills, described in Figure 4, and the capacity to model interannual fish catch variations with chlorophyll anomalies as environmental drivers)  and which is the scientific novelty they bring. 

Moreover, the description of the experiments and of the results appears not always clear, and more explanations (i.e. a more detailed description of the content of Figure 4, and of the relationship between the anomaly correlation skill behavior described in figs 4a and 4c with the maps of figs 4b and 4d) would improve readability and strengthen the paper as they would support the research question posed by the authors. Finally, the descriptions of certain figures, such as Figures 4 and 5, lack sufficient detail, limiting the comprehension of both the analyses conducted and the importance and relevance of the results obtained. 

In consideration of the previous points, the paper is acceptable for publication after major revisions.

A list of punctual issues is listed below. 

ABSTRACT:

• (L13-14): Enhance clarity and focus on this sentence to be more consistent with the problem presented.
• (L20): The sentence emphasizes the relevance of physical–biogeochemical coupling processes; however, it remains unclear whether the network explicitly learns this coupling or merely reproduces its effects, as well as the mechanisms by which such learning or reproduction is achieved.
• (L22): The term “chlorophyll anomalies” is introduced, but it is not defined, together with the baseline used for its computation. The entire article strengthens on this aspect, but there is no formal definition of anomaly. 

1. INTRODUCTION: 

• (L35): The inclusion of references to the definition of ESMs would facilitate a deeper understanding of the purposes of the project. 
• (L50): Deep learning models are highly sensible on data coverage. In particular, observational gasps and data-sparse components represent a huge limitation for the majority of deep learning approaches. Even if their usage grows with the increasing availability of data, sparse coverage still represents a limit for these models. A clearer explanation of the statement asserting that deep learning methods are well suited to data-sparse components would strengthen the justification for adopting a deep learning approach for this application.
• (L62): The manuscript does not clearly describe the outputs of the deep learning model. Both chlorophyll concentrations and chlorophyll anomalies are presented as model products; however, the definition and interpretation of the anomaly are not provided. Clarification of this aspect would improve the reader’s understanding of the overall study. Furthermore, it is unclear whether each LME is modeled independently or whether the model produces a global output from which individual LMEs are subsequently extracted and analyzed.
• (L65-68): I think a re-organization of the last sentences of the introduction would enhance the comprehension of the project. The current description of the dataset appears overly detailed for an introductory section, while some key elements, such as a clear definition of the model outputs, are not sufficiently addressed. It is therefore recommended to revise these passages by emphasizing the general characteristics of the proposed algorithms and providing only high-level information about the dataset, while relocating the detailed dataset description to the dedicated method section.

1. METHODS: 

• Section 2.1: the architectural description lacks key details required for reproducibility, such as a comprehensive table of all hyperparameters and a clear rationale for the choice of the proposed architecture and its components, such as including the use of GELU activations and the selected loss function. To further improve the clarity of the manuscript, it is recommended to present the network architecture, dataset, and validation strategy in separate subsections.
• (L91): the concept of anomaly correlation coefficient is introduced, but not defined. Including its definition, along with a brief description, would enhance the reader’s understanding of the results.
• Section 2.2 describes the dataset used, including the input, validation, and test sets, and provides details on input data preprocessing. I recommend reorganizing this section to clarify the distinctions between datasets used for different purposes. Additionally, more detail on the input data preprocessing would improve clarity, as the structure of the input data is not fully specified.  Specify source, variables, spatial resolution, temporal frequency of data used; in particular, clarify which dataset collects the input variables used for training, validation and test. Use a table if it can help. Moreover, it is unclear whether the inputs consist of concatenated global 2D maps of SST and chlorophyll anomalies or of 2D maps defined separately for each LME. Likewise, the description of the network output lacks clarity: it is not evident whether the output represents a mean chlorophyll value across all LMEs or a spatial map over each LME, nor whether the model predicts chlorophyll concentrations, chlorophyll anomalies, or both. 
• (L103): The input mask fills missing values with zeros. It would be helpful if the authors could provide additional insight into the rationale behind this choice. In particular, further clarification on whether missing values and land points are treated differently, and on the network’s ability to distinguish between these cases, would enhance the reader’s understanding.
• (L124): Paragraph 2.3 introduces SHAP as a method for interpreting model predictions and identifying dominant spatial drivers (L118). However, the role of SHAP in this context is not entirely clear. Given that the network inputs consist of SST and chlorophyll anomalies, one would expect the analysis to highlight the relative importance of these input variables. Instead, at (L124) it is stated that feature (i) corresponds to a specific grid point in the input map, which introduces some ambiguity regarding what information the SHAP analysis is intended to convey. A clearer explanation of how features are defined and how SHAP results should be interpreted would improve the clarity and understanding of the results.
• Section 2.4: the scope of this section appears scientifically obscure. Chlorophyll (or its anomaly) timeseries from satellites or from ESM models can be directly used to predict catch timeseries. Which are the added values of using NN derived chlorophyll? One would expect, at least, a comparison between catch timeseries predicted using chlorophyll from satellites, or to see the advantage of using the NN derived chlorophyll.   

RESULTS: 

• Section (3.1) presents a very interesting and informative analysis; however, some of the architectural details discussed here would be more appropriately included in the Methods section, within the description of the model architecture. In addition, to improve the comprehensibility of the architecture described in the Methods section and to maintain focus on the model’s results, it is suggested to move this sensitivity analysis to a Supplementary Materials section.
• (L150): In the caption of Figure 2, the baseline model is described as sharing the architecture of the reference model, while differing in certain training settings, such as the loss function. This suggests that the reference model represents an optimized version of the baseline. However, at line 148 it is stated that the sensitivity analysis presented in this paragraph originates from the reference model, with a single component modified in each experiment. Could the authors clarify the contributions of these sensitivity experiments to the reference model, and how the reference model was optimized relative to the baseline? Providing this explanation would help improve the reader’s understanding of the experimental design and the relationship between the baseline, reference, and sensitivity models.
• (L155): The concept of prediction skill is not defined, and its meaning remains somewhat unclear. In particular, in Figure 2, it is not evident what exactly the prediction skill measures. Including a brief description would enhance both clarity and will facilitate the comprehension of the proposed results.
• (L175): Could the authors clarify the statement, “The inclusion of additional input datasets generally improved the model’s prediction skill”? It should be noted that adding input variables does not necessarily guarantee improved model performance; if the additional inputs have weak correlation with the target, their inclusion could potentially lead to overfitting. Providing a reference and a more detailed explanation would help clarify this point and strengthen the interpretation of the results. Moreover, the choice to include chlorophyll as input variable when predicting chlorophyll itself should be clarified.  Moreover, it would be helpful to provide a table which contains for each test the input variables used for it. It is somehow difficult to reconnect the text to names listed in figure 2. 
• From Figure 3a, it appears that the CNN output is represented as a single mean value for the entire LME, resulting in a uniform color. Could the authors clarify whether this interpretation is correct, or if the correlation is instead computed at the level of individual grid points? Providing this clarification would help improve the reader’s understanding of the figure and the network’s output. Based on Figure 1, the inputs appear to consist of timeseries of two-dimensional spatial fields, whereas the outputs correspond to timeseries of zero-dimensional quantities (i.e., single surface values). If this interpretation is correct, the rationale for adopting a two-dimensional–to–zero-dimensional mapping should be explicitly discussed.  In particular, it would be helpful to clarify the intended purpose and advantages of this approach compared to the use of a simple spatial average, as well as to articulate the scientific novelty that this methodology is expected to provide. 
• Improve the quality and clarity of the figure 3: y axis is missing the label and unit, and the text should be enlarged. 
• (L195): The exact number of CNN input variables is not entirely clear. While SST and chlorophyll anomalies are listed as inputs in the introduction (L62), a different description appears later, stating that the model was “tested with a combination of physical and biogeochemical inputs, that is, SST only, chlorophyll only, and both SST and chlorophyll.” Could the authors kindly clarify the reason for this apparent discrepancy? If a sensitivity analysis was conducted to determine the optimal set of input variables, it would be helpful to briefly describe the procedure. Otherwise, specifying the exact input variables used in the current model would improve clarity for the reader.
• (L213): The manuscript states that “including surface chlorophyll anomalies, either alone or as an additional predictor, substantially increased the number of LMEs where the model achieved high prediction skill.” In the introduction, chlorophyll anomalies are already presented as an input to the model, whereas here it appears that they are added subsequently. Could the authors kindly provide a more detailed explanation of how the input data are structured and used? Clarifying this point would improve the reader’s understanding of the model setup and the role of different predictors.
• (L215-220): The caption of Figure 4 lacks clarity, and the prediction task described in lines 215–218 would benefit from a more detailed explanation. In particular, the inputs and outputs of the task should be explicitly specified, and the procedure used to compare predictions with observations should be described more clearly. For example, it is unclear which quantities are being compared at each grid point in Figures 4a and 4c. Additionally, the captions for Figures 4b and 4d are ambiguous; as currently presented, it appears that two sequences of three maps are shown. Consideration could be given to splitting this content into two separate figures in order to improve readability and facilitate the reader’s understanding of the task.
• The analysis presented in the latter part of paragraph 3.2 is interesting, and the results shown in Figure 4 are valuable. Nevertheless, the paragraph would benefit from a more detailed explanation of what is the content and the relevance of figures 4a and 4c, together with the implications between Figures 4a and 4b, as well as between Figures 4c and 4d. Clarifying these connections would greatly enhance the reader’s understanding of the results and their interpretation.
• (L239): The manuscript states that “the recurrence of this pattern in the model’s predictions indicates that it captures subsurface ocean memory in addition to surface signals.” Could the authors clarify why the recurrence of this pattern is interpreted as evidence of subsurface ocean memory, given that subsurface variables do not appear to have been used or introduced as input to the model? Providing additional explanation would help improve the reader’s understanding of this conclusion.
• (L248): For the sake of comparison, it would be helpful to include the ENSO dynamics in a Supplementary Material section, providing a baseline for reference alongside Figures 4b and 4d.
• L265-270: move in material and method the description of models. 
• Explain better how correlation between satellite chlorophyll and predictions by DL and dynamics are computed.
• In figure 5, use labels (DL and dynamics) that are consistent throughout the paper and are clear; if figure 5a and 5b provide the same information, consider simplification and use only one, otherwise, clarify the distinction. 
•  In Fig. 5a, the numbers of significant correlations are 15 for DL and 16 for dynamics. It appears to me to have quite poor performance results. Please reformulate L281-282.
• (L284): Some regions are listed as examples of comparable performance habits, but the Fig 5a does not show explicitly these regions. Indicating at which bars of the plot they correspond would increase the clearness of the results. Moreover, a map with the 66 LME is missing in the paper. 
• To evaluate the validity of using NN chlorophyll predictions instead of observed chlorophyll data for fish catch prediction, it would be informative to include a comparison, for example with results obtained from a linear regression model using satellite chlorophyll observations. Alternatively, please clarify the reason for this methodological choice.  
• Figure 6 shows only 2 LME. Providing additional information about the correlation between chlorophyll and fish catch in the other LMEs could strengthen the results.  
• Improve the quality and clarity of the figure 6: y axis is missing the label and unit and the text needs to be enlarged. Does y-axis represent the correlation coefficient between fish catch and chlorophyll anomalies, or the comparison between predicted and observed fish catch?

DISCUSSION & CONCLUSION: 

The conclusion and discussion section clearly summarizes strengths and limitations of the approach and the value of the sensitivity analysis. However, a few aspects could be better presented.

• (L340):  The phrase “while capturing physically interpretable signals underlying chlorophyll variability” could benefit from clarification. Since the CNN inputs are SST and chlorophyll anomalies, it would be helpful to specify whether this comment refers specifically to SST or to other physical signals. Providing this clarification would improve the reader’s understanding of the model’s interpretation.
• (L340): The statement that “the model successfully reproduces the known ocean–climate process” could benefit from further elaboration. Providing a brief explanation of which specific ocean–climate processes this sentence refers to would help strengthen the interpretation of the results and improve clarity for the reader.
• (L362): The statement that “sensitivity tests show that surface chlorophyll anomalies captured subsurface variability” would benefit from further clarification. From the manuscript, it appears that the sensitivity analysis was primarily performed to optimize the network architecture and input data. It is therefore not immediately clear how this analysis supports the conclusion regarding subsurface variability. Providing a more detailed explanation of the connection, or the underlying correlations, would help the reader better understand the interpretation of the proposed results.

I recommend major revision: the paper is promising and well positioned, but the current significance and evaluation framework does not convincingly rule out chance findings across many regions/species/lead times, and several methodological choices need clarification or strengthening.

The manuscript develops a CNN-based system to forecast surface chlorophyll anomalies for Large Marine Ecosystems (LMEs) using three consecutive months of SST and chlorophyll anomaly maps as inputs, trained on CMIP6 simulations plus a coupled reanalysis, and evaluated against SeaWiFS/MODIS satellite products. It additionally benchmarks against an ESM-based dynamical biogeochemical forecast system and explores fisheries relevance via regressions between predicted chlorophyll anomalies and reported fish catches.

Strength points:

- Clear problem framing and a relevant niche: you directly target known limitations in ESM biogeochemical predictability (observation sparsity, structural uncertainty, and computational cost), motivating a data-driven complement.

- Global scope with an application-relevant unit: the LME-scale framing is practical for coastal management and fisheries, and the system is trained/validated/tested on long records spanning CMIP6, reanalysis (1965–1997), and satellites (1998–2021).

- Interpretability attempt linked to dynamics: SHAP attribution is used to relate skill to recognizable mechanisms (ENSO-related patterns for the Pacific Central-American Coastal LME and Rossby-wave–like propagation in the Agulhas region).

Major concerns (must address):

- Multiple testing / field significance: annual skill is presented as “significant” using p < 0.10 markers across LMEs, and you then show time series for eight LMEs with significant skill when using both SST and chlorophyll inputs. With many LMEs tested, p < 0.10 without a multiple-comparison correction (e.g., FDR control) is not sufficient to claim that the set of “significant LMEs” exceeds what would occur by chance; this is especially important because the paper’s central headline is “skillful predictions in many LMEs.”

- Monthly forecast evaluation appears cherry-pickable: monthly forecasts (up to 24-month lead) are shown for only two LMEs (Pacific Central-American Coastal and Agulhas Current), chosen because they “exhibit significant annual mean chlorophyll prediction skill.” This selection criterion is not adequate to avoid post-selection bias for the monthly lead-time maps; readers will reasonably ask how typical these two are across all LMEs and whether the same lead-time structure occurs elsewhere. Relatedly, statements like “significant correlations extending up to 12-month lead times for forecasts initialized during boreal winter” (for LME 11) require stronger controls for the large number of initialization-month × lead-time tests shown in Fig. 4.

- Benchmark comparison needs uncertainty quantification: the dynamical benchmark is a strong part of the paper (it is well described as a 12-member, 2-year forecast system initialized monthly over 1991–2017). But the “outperformance” map that uses a correlation-difference threshold (≥ 0.2) at a nominal significance level raises questions: correlation differences should be accompanied by uncertainty estimates and a paired test (e.g., block bootstrap or Fisher-z with effective sample size) to show where differences are robust, not just large.

- Fish-catch results: selection and multiplicity: fisheries analysis uses Sea Around Us catches, selects top-10 species per LME, and applies linear regression using NDJ-initialized chlorophyll forecasts with different lags, reporting significant correlations for a subset of LME–species pairs. As written, it is unclear whether LMEs, species, and lags were pre-specified or chosen after looking at results, and there is no correction for the very large hypothesis space (LMEs × species × lags). Also, this section currently feels only loosely connected to the core ML/forecasting contribution (and it is linear regression, not a neural network), so it either needs a more rigorous, pre-registered-style evaluation or should be reframed as exploratory/supplementary.

Methodological issues and clarifications:

- Zero-filling of missing ocean color: satellite chlorophyll has missing values (clouds/polar night), and you apply “zero-filling,” masking missing pixels and filling them with zeros. This can inject artificial anomalies and create learnable artifacts (especially near persistently cloudy regions and high latitudes); at minimum you should quantify sensitivity (e.g., compare with masked-loss training, add a missingness channel, or use a learned imputation/gap-filled product).

- “Five ensemble members per initialization” is unclear: monthly forecasts are said to use five ensemble members per start date. For a deterministic CNN, this needs explanation (different random seeds? Monte Carlo dropout? perturbations of inputs?); also report how ensemble mean/spread are used in ACC computation and whether ensemble spread relates to skill.

Requested revisions:

- Add a formal multiple-testing treatment for annual LME skill (e.g., Benjamini–Hochberg FDR on p-values).

- For monthly forecasts, provide summary skill maps/statistics across all LMEs (or a clearly pre-specified subset) for initialization month × lead time, and then discuss the two highlighted LMEs as case studies.

- For DL vs dynamical comparison, replace the ad hoc “0.2 correlation difference” rule with a statistically grounded paired comparison and uncertainty intervals on skill differences.

- For fish catch, explicitly predefine the hypothesis set (LMEs, species, lags), and show aggregated results. If this cannot be done within scope, label the section clearly as exploratory and move it to Supplement.

- Clarify what constitutes “ensemble members” for the CNN monthly forecasts and how they affect reported significance

Suggested revisions:

- Revisit missing-data handling: include a missingness mask as an input channel and/or use masked losses, and report sensitivity relative to current zero-filled preprocessing.

With these changes, the manuscript could make a strong, credible contribution: the global LME framing, dynamical benchmarking, and mechanistic interpretability angle are all compelling, but the statistical evidence needs to be made robust before the main claims can be supported.

Overall assessment (summary for editor and authors):

This manuscript tackles an important direction: using machine-learning emulation to support analysis of chlorophyll variability and uncertainty in Earth System Model (ESM) settings. The topic is timely and relevant to ESD readership, and the paper has potential value if it can clearly justify (i) why chlorophyll is the right motivating problem in an ESM context, (ii) how the proposed training/validation strategy avoids circularity and inherited model biases, and (iii) how interpretability claims (e.g., SHAP) are supported with visible results and biogeochemical discussion.

At present, however, the paper has several conceptual and technical gaps that make the narrative and methodology feel under-justified or internally inconsistent. In particular, the motivation around “chlorophyll problems” and the role of satellite ocean-colour data in ESMs needs to be sharpened; the reliance on CMIP6 simulations for training needs stronger justification given the manuscript’s stated concerns about ESM limitations; the evaluation framework and skill metrics need to be stated explicitly; and the interpretability section is currently incomplete (SHAP plots are referenced but not visible).

I therefore recommend major revision. With rigorous clarification, strengthened motivation, clearer methods, and improved validation/interpretability presentation, the study could become a solid ESD contribution.

Major comments

Line 35–48 (motivation / problem statement on chlorophyll):

I am not finding enough rationale behind the “problem” with chlorophyll here. The manuscript should be more precise: what specific deficiency in simulated chlorophyll is being targeted (e.g., seasonal timing, bloom onset/termination, amplitude biases, regional pattern errors, long-term trends, nutrient limitation regimes, or vertical structure)? At the same time, ocean-colour observations are among the richest global observation networks and are widely used for short-term evaluation and reanalysis-type applications. Since the manuscript frames the work in the context of ESMs and climate projections, satellite ocean colour will not “solve” the forward-projection problem (it mainly constrains the historical period and near-real-time monitoring). This makes it difficult to pin down the motivation: why is chlorophyll observation discussed here in the ESM context, and what is the precise gap the proposed ML method fills? The authors should sharpen the problem formulation and explicitly connect it to ESM-relevant uncertainty and projection needs.

Line 42–43 (“Contributions to the substantial uncertainties …”):

This sentence is unclear. If the authors mean parameterisations contribute substantially to uncertainty, please specify which parameterisations (e.g., phytoplankton physiology, photoacclimation, grazing closure, remineralisation, mixing-light coupling) and in what way they contribute. As written, “parameterisations” is too broad and reads vague.

Line 58–59 (context / recent related work):

It may be worth acknowledging that there are very recent works on physics-based AI integration within biogeochemical models (including preprints). For example, a preprint Banerjee et al., 2026 (https://doi.org/10.31223/X5C74R) could be cited as a related direction to position the study in the rapidly evolving landscape.

Line 60 (“three consecutive months …”):

Please clarify what “three consecutive months” refers to (training window? evaluation period? event definition?) and why this temporal criterion is chosen. If it is a key design choice, it needs justification.

Line 64 (training with CMIP6 coupled models / conceptual consistency):

Training with CMIP6 model output will inevitably bring CMIP6’s own uncertainties and biases into the learned emulator. Earlier, the manuscript discusses limitations of ESMs, yet the ML model is trained entirely on CMIP6 simulations, which feels self-contradictory unless carefully justified. The authors should explicitly explain: (i) what the emulator is learning (model space vs observational truth), (ii) why this is still useful for the stated aims, and (iii) how inherited CMIP6 biases are managed or acknowledged (e.g., domain limitation, bias-aware training, uncertainty propagation, or careful interpretation that results are “CMIP6-consistent” rather than “truth”).

Figure 1 (infographic colour palettes):

In the infographic, the colour plots for CHL and SST (even if symbolic) would be clearer if the colour palette is differentiated between SST and chlorophyll. SST appears as a feature (not anomaly), yet it is shown using a coolwarm-type palette, which can be misleading. A more conventional SST palette (or a clearly distinct palette) would improve clarity.

Line 86–87 (“Seasonal or annual …”):

Why seasonal or annual mean? At this stage the reader should be specifically informed what temporal aggregation is used and why it is appropriate for the scientific question. Please be explicit.

Line 87 (“piControl …”):

Not sure a broad readership will understand “piControl” without explanation. Please add a short definition (e.g., long pre-industrial control simulation under fixed forcing, used as a baseline for internal variability).

Line 88 (“Satellite-derived chlorophyll …”):

Why not train using satellite chlorophyll/ocean-colour data, given the abundance and long record, and instead use it only for testing? If the goal is ESM-relevant emulation, the authors should explain the choice clearly (e.g., differences in variable definition, sampling, error structure, coverage, or mismatch between satellite CHL and model CHL). As written, this choice needs stronger rationale.

Line 103–104 (handling SST = 0 / masking strategy):

A key methodological concern: if SST values are set to zero (or if missing values are replaced with zero), then the emulator could incorrectly learn that 0°C (or 0 K depending on conventions) is a meaningful physical signal in regions where it is not, and this could contaminate training. The authors should justify the zero-handling strategy and ideally use a robust missing-data approach (masking, NaN-aware methods, explicit land/ice masks, or physically meaningful fill values plus a missingness indicator feature). Please clarify how this is handled and why it is safe.

Line 106–108 (“Reanalysis data …” validation choice):

Why use reanalysis data as validation, given known potential biases, especially for biogeochemical fields where assimilation and availability of ocean-colour observations vary and can affect the reanalysis itself? If satellite observations exist for chlorophyll, why are they not used more directly for validation? If reanalysis is used, please justify the choice and discuss limitations.

Line 118 (“To interpret …” SHAP figures not visible):

The manuscript mentions SHAP-based interpretation, but I am not able to see the SHAP feature plots in the main text. If they are in supplementary material, please ensure they are included and clearly referenced. If not provided, they should be added, as interpretability is part of the paper’s core claims.

Line 141–142 (“statistical significance …” skill metric definition missing):

I am not able to find the formulae/definitions for the prediction skill metrics used. The authors should define the prediction skill metric(s) clearly in the methods section (and provide the statistical test used for “significance,” if applicable). Without explicit definitions, it is difficult to interpret the results.

Line 181–183 (log-transform claim vs Fig. 2):

Figure 2 appears to show prediction skill is better with log-transformed chlorophyll, which seems contradictory to the claim made in lines 181–183. Please reconcile this: either revise the claim or clarify what specific aspect improves/worsens with the transform (e.g., relative vs absolute error, low-CHL regimes, extremes).

Line 210–211 (feature set choice is too narrow):

The feature set appears limited (e.g., mainly SST), but there are well-established publicly available datasets for other drivers that are biogeochemically relevant, such as mixed layer depth (MLD), PAR, winds (proxy for mixing), and potentially nutrients or stratification indices. Why were these not included? Even if the authors aim for minimal features, this choice needs to be justified given known controls on chlorophyll variability.

Line 242–244 (SHAP plots + biogeochemical interpretation vs black-box):

Again, I cannot see the SHAP plots referenced. In addition, for each SHAP result, the authors should explain whether the learned feature importance makes biogeochemical sense (not just statistical attribution), otherwise it risks being presented as black-box interpretability. The discussion should connect SHAP patterns to plausible mechanisms (e.g., SST as stratification proxy, seasonal light limitation, mixing control, etc.) and acknowledge where interpretation is uncertain.

Major revision.

The study is promising and timely, but it currently does not strike a clear balance between Earth-system modelling theme and the AI-method framing. A major revision that sharpens the ESM-relevant motivation, resolves the CMIP6-training self-consistency issue, clarifies the evaluation framework (metrics + significance), strengthens validation choices, and fully provides/justifies interpretability outputs (SHAP) would substantially improve the manuscript and make it suitable for ESD.