Summary:

The manuscript makes use of the subset of models from CMIP6 that provided high temporal resolution SST / chlorophyll output to investigate the sub-seasonal dynamics of simulated phytoplankton. The model output is processed first using a decomposition methodology that breaks down the variability into different temporal modes. Subsequently, the output is further processed using spatial decomposition to identify whether the simulated variability has the correct horizontal length scales (e.g. to distinguish where models exhibit seemingly comparable variability to observations but on much coarser spatial resolution). The manuscript finds that none of the models realistically represents the seasonal and sub-seasonal patterns of variability observed (unlike the situation with SST). However, the spatial decomposition teases out patterns between models that allows them to be separated into three groups with better or worse representation of real-world variability. The authors note that one group is hampered by its spatial resolution, while the other two exhibit excessive sub-seasonal variability apparently from tightly-coupled predator-prey cycles. The manuscript concludes by noting the value of high temporal resolution output for identifying unrealistic model behaviour, and with a call for modelling groups to endeavour to provide this in CMIP7.

Review:

Overall, I found the manuscript interesting and quite convincing about the realism or otherwise of current generation CMIP models. I have no major comments on the content of the manuscript, but have made a small number of minor suggestions about improvements. I recommend accepting the manuscript following these minor corrections.

Comments:

One overall comment I have is around the quality of the figures. There are some unhelpful choices here to my mind and I detail these below. However, I would accept that this is largely an aesthetic decision, and would not insist on my suggested changes being implemented. Another general comment I’d make is that it would be good to try to put the models examined into some sort of context within the wider CMIP6 ensemble – I’ve suggested an idea in what I say about Figure 2a, but there may be a more obvious or better solution.

Ln. 46: Inconsistent ordering of references; they’re neither in chronological nor alphabetical order (I prefer the former).

Ln. 64: Amend to “… produced by *a subset of* ESMs …”.

Table 1: The IPSL and CNRM models are lumped together (presumably because of a common ocean), but do they share a common atmosphere or atmospheric resolution?

Table 1: The MPI rows have a missing border between HAMOCC6 and 150 km cells.

Table 1: The MPI and NorESM2 rows mention piControl simulations, but I don’t believe that these are mentioned elsewhere in the manuscript.

Ln. 134: A period of 8 months is mentioned here for the so-called “multi-annual component”. Why 8 months and not 12 months? I’m sure I’m not understanding something.

Ln. 153: I’m not a fan of “Results and Discussions” sections, and would prefer the authors to properly separate results from discussion to improve the manuscript’s clarity. However, it can be difficult to separate them at this stage, so ignore this suggestion if it isn’t obvious to address.

Figure 1: Conventionally, darker colours are used to indicate lower values while brighter colours are used to indicate higher values. The choice here is confusingly the reverse.

Figure 2a: The models are distributed into two clear groups but the manuscript doesn’t reflect on this. Is there any straightforward distinction to be drawn between them? For instance, what would the mean fields of the two groups look like? Would there be any clear distinguishing patterns.

Figure 2a: Since the models examined fall into only 4 “families”, and given that they all perform fairly badly here, I wonder if it might be worthwhile somehow contextualising their performance against the wider CMIP6 ensemble? Possibly by adding other models that are outside of the analysis here? Either in this figure, or in a supplementary version of this figure. Even without those models being analysed in detail as here, it would provide context for the representativeness of the models used here.

Figure 3: This chart makes a sensible comparison between the variability modes of the obs and models. However, I wonder if there’s a way to put the information it presents onto a single axis where the models and observations can be seen together. For instance, “total” variability on the x-axis, and the fraction that’s sub-seasonal on the y-axis? You may have tried something like this already.

Figure 4: Add in the caption which models, and why, are missing here. Presumably data availability?

Figures 4, 5: A bit more consistency in style would be good for these bar chart figures. Figure 3 seems to make use of the space best, with Figures 4 and, especially, 5 using it less well (i.e. the bars are thinner).

Figure 7: I think this could be a much better figure if pie charts weren’t used. Each model (and possibly model region) could be given a simple x-y subplot in which the x-axis is period and the y-axis is geographical area or frequency. Each subplot could then also contain the same information for the observational data. This would add information currently hidden by the limited number of periods selected for the pie charts, and would make it easier to compare with the observational data. At present the reader has the unenviable task of squinting to try to work out how similar / different one pie chart is from another. Line plots would – I suggest – be much better here.

Figure 8: This is a horrible colour map. Not only is it a single colour, but the different shades of that colour are very difficult to discern, with an emphasis on darker shades that make any distinctions in the plots fairly invisible. Why not use one of the colour maps used elsewhere to make discerning the distinctions easier?

Ln. 502: The structure of biogeochemical models is alluded to here but no evidence is presented. Perhaps illustrating with time-series plots of representative differences between models might help clarify this here. Or even examine the low frequency output of the models involved to determine if they differ in their phytoplankton-zooplankton relationships. However, this is only a suggestion as it might be sending you on a wild goose chase.

Ln. 515: The authors advocate for CMIP modelling groups to submit daily outputs of biogeochemistry variables but don’t mention which ones specifically. Obviously chlorophyll but, per the preceding point, would they advocate for others like surface zooplankton too? This is a good opportunity to advocate for them.

General comments:

The writing is clear and the manuscript introduces new findings that contribute to our understanding of chlorophyll-a dynamics. I recommend accepting the manuscript with minor revisions. The following comments are merely provided as suggestions to further improve the manuscript's completeness and clarity.

Summary:

This study compares simulated surface chlorophyll-a (Schla) variability from a subset of CMIP6 Earth System Models (ESMs) with satellite observations and contrasts this performance with that of SST. The analyses highlight discrepancies in the ability of ESMs to simulate Schla across different timescales, with a specific focus on the understudied sub-seasonal variability. The ESM simulations are selected based on the availability of daily Schla and SST outputs. Temporal variability is decomposed into sub-seasonal, seasonal, and multi-annual scales, identifying three main groups: one showing an overestimation of sub-seasonal variability which is attributed to the coarse spatial resolution of the ESMs, a second group showing an underestimation of sub-seasonal variability, potentially linked to intrinsic predator-prey oscillations within the ESMs, and a third group displaying an overestimation of total variance but consistent temporal decomposition. The authors conclude that, unlike SST, ESMs do not adequately represent Schla variability, emphasizing the need for additional CMIP simulations with higher spatial and temporal resolutions to address these limitations.

Specific comments:

Perhaps the manuscript could include an explicit mention of the limitations of the approaches and how they affect the final findings, specifically concerning:

• The biases/uncertainties of satellite observations: While the manuscript uses satellite observations as the benchmark for comparison, it would strengthen the discussion to acknowledge the inherent biases and limitations of these datasets. For instance, biases introduced by gap-filling and uncertainties in the retrieval process could affect the representation of SChla variability. It would be beneficial if the authors discussed these biases and how they might influence the overall findings.
• Comparison of satellite and ESM timeseries: The analyses use satellite timeseries spanning 16 years and ESM simulations spanning 33 years. It would be helpful if the authors addressed whether this difference could impact the representation of multi-annual variability in the analyses and thereby affect their findings and conclusions.
• Thresholds for spatial coherence analysis: When mentioning the thresholds for the analyses of the spatial extent of coherence, the authors could clarify the rationale behind the choice for an upper threshold of 2400 km in diameter and the threshold value of 0.8 for correlations and how these influence the findings and interpretations.
• ESM future simulations: It could be mentioned why the future simulations analysis was limited from 2084 to 2100, rather than a longer time range.
• Use of a single ensemble member: The study currently uses one single ensemble member per ESM. It would be interesting to discuss the implications of this choice, as utilizing the ensemble mean could provide a more accurate representation of model performance and reduce variability introduced by individual simulations. Similarly, where possible, it would be valuable to discuss the mean across ESMs, as ensemble means often yield more accurate representations than individual models.

Technical corrections:

Ln 81: Umlaut on Müller

Ln 84: …more than ‘three’ times...

Ln 88: Keerthi et al. (2022) (comma is not necessary)

Table 1: A  border line is missing between HAMOCC6 and 150 km

Ln 126: Reference the CDO remapping tool remapdis (see reference on: https://code.mpimet.mpg.de/projects/cdo/wiki/Cite)

Ln 173:  ‘display’ in plural instead of displays

Ln 182: Perhaps mention the metric of correlation employed, I assume the Pearson Correlation coefficient?

Fig 2: Add degree symbol at 60°N and 60°S. For clarity, consider adding to the description that the dots represent models, while dashed lines represent observations. Additionally, complement the color scheme by using different symbols for each model to improve accessibility for color-blind readers.

Ln 216: There is no reference to Figure 3 in the text. It would improve clarity to reference Figure 3 here.

Fig 3: The description states ‘(Left Panel)’, however, I do not see a left and right panel nor a reference to a ‘(Right Panel)’. Consider adding for clarity: Normalized standard deviation ‘of globally averaged’ Schla…

Ln 266-267: Consider adding a reference to the figures in the sentence: The standard deviation across different timescales and the relative contribution of these timescales to the total SST variance ‘(Figure 4)’ show distinct patterns compared to SChl ‘(Figure 3)’.

Ln 273: The term 'ENSO' is used as an abbreviation without prior introduction. Additionally, 'El Niño' is mentioned in line 469. For coherence and clarity, consider introducing the term in full as 'El Niño–Southern Oscillation' upon its first use, then consistently using either 'El Niño' or 'ENSO' throughout the rest of the manuscript.

Description Fig 4: Is it standard practice to reference a previous figure or would it be helpful to include the full description again?

Fig 6 ln 353: 800 km in lowercase

Ln 473-474: Consider adding the following: The simulated change of the sub-seasonal variability of SChl ‘in response to X’,…