This manuscript is a nice addition to the long-standing scientific debate about what drives decadal to multi-decadal variability in the ocean carbon sink. The manuscript is comprehensive and mostly well written and clearly structured. The methods and results are well explained and appear robust.

My one truly major comment is regarding the lack of proper statistical analysis. This will have to be added before the manuscript is published. In addition I want to raise several other issues that should be considered and which I think would help improve the manuscript and increase its impact.

General:

Occasionally the reading is marred by awkward sentences. The second sentence in the abstract exemplifies this: Observing the ocean carbon sink is not challenging because there are high uncertainties, but the uncertainties are high because observing the ocean is challenging.

Throughout the manuscript the author states that observation-based pCO2 products extrapolate observations. I do not like the use of the word “extrapolate” here. When we extrapolate we extend into an unknown situation, but the observation-based products primarily attempt to interpolate between known situations. It is a bit nit-picky, and probably boils down to semantics, but I’d prefer the term “gap-filling”. In my mind that is more comprehensive.

Introduction:

The introduction begins with a statement that the ocean has removed about 25% of all anthropogenic emissions since the onset of the industrial revolution. This is not factually incorrect, but I still find the statement somewhat misleading. Based on Table 8 in Friedlingstein et al. (2023) the cumulative ocean uptake (both since 1750 and since 1850) amounts to approximately 25% of the total emissions (FF+LUC). However, given this statement at the beginning of section 3.9 in the same paper “The cumulative land sink is almost equal to the cumulative land-use emissions (220±70 Gt C), making the global land nearly neutral over the whole 1850–2022 period.” we understand that ocean has been more important in storing human-made emissions than the “has taken up around one quarter of all anthropogenic emissions” would indicate. In a paper highlighting the ocean sink I think this is a nuance worth noting in the introduction. But I stress again that the statement is not actually incorrect.

In the introduction (lines 92 onwards) the point is made that differences between observation-based products and GOBM products could be due to the data sparsity. Here it should be noted that it is not the sparsity that is the major problem, but rather the uneven sampling in time and space. Hauck et al. (2023) showed that if observations were regularly spaced the differences largely disappear. This is worth mentioning because it is a problem much more difficult to remedy than just having too few observational data.

Throughout the manuscript it can be difficult to understand what “drivers of the decadal trends” actually mean. My understanding is that this study looks at defining the underlying causes of multi-decadal variability, that is, variability in the decadal trends. I could be mistaken, but this should regardless be stated more clearly in the introduction.

In the introduction several limitations to studies using the observation-based and GOBM products are presented. This is correct and fair, but there are also limitations to using ESMs so a few sentences describing these should be added at the end of the introduction.

Methods:

In section 2.2 the author describes how the magnitude of the ocean carbon sink in the 12 different ESMs was corrected/adjusted. First, I would think that the magnitude of the ocean carbon sink is related to the model’s climate state since it is closely linked to the ocean circulation. So what are the consequences of doing such a correction to the models? Second, it is stated as fact that a “negative bias in the magnitude of the carbon sink also introduces a negative bias in the decadal trends”. I do not understand why this will always be the case. This part of the method requires a bit more description, and a bit more discussion.

In the introduction the author argues that one of the benefits of using ESMs over other types of data products is the ability to perform robust statistics. I completely agree, but it seems as if no, or very little, statistical analysis has been performed. At least there is no description of such analyses in the methods. This I think is a major weakness of the manuscript as it stands now. At the very least a table with statistics for Figures 3, 4, 5 should be added, but a more thorough statistical approach to the analysis in section 5 and 7 would also be beneficial. I also note here that the author presents a p-value for the regression analysis (Figures 3, 5, 8). The p-value has been a topic of discussion for several years now, and is often misused. At the very least you must state what null hypothesis you are testing. However, given that you have a lot of data points the p-value is perhaps not the most useful metric. Given the amount of data available when using several ESMs I would suggest you try a Bayesian hypothesis testing instead for a more useful metric for significance.

Results:

In Figures 1 and 2 it is hard to tell which line represents which scenario. Please choose colors with more contrast.

In Figures 3-5 it is difficult to tell the difference between the observation-based points and the GOBM-based points. They are both too small (given the black outline), and the colors are too similar to easily differentiate.

I find Figure 4 very interesting and would have liked more discussion about it. It looks like there may be some regional variation in at what global sink strength the regional sink deviates from the expected trend. Intuitively this makes sense to me, but it would be interesting to see whether it really is the case or just me seeing things. Either way it would add interesting discussion about why the relationship breaks down. Also, considering that the regional analysis, for which no correction of the low bias in sink magnitude was performed (section 2.2), produces results so comparable to the global analysis, why is the bias-correction in section 2.2 necessary? It would be good if it could be tested what the results would be if no correction was done.

Section 4 warrants a more robust statistical analysis and more discussion. Right now no reasons are given for the presented differences. Also, given the short timeframe for most of the observation-based products how robust are the presented results?

Minor comments:

Line 26: add “and” before “the ocean”

Line 71: in the abstract you state “3 to 7 decades”, here it is “four to seven”

Line 75: the observations are not just relatively sparse, they are very sparse

Line 88-89: this sentence is unclear and needs rewriting for clarity

Line 104: move “from phase 6 of the Coupled Model Intercomparison Project (CMIP6)” to directly after “12 ESMs” on line 103

Line 158: change “effect” to “affect”

Line 190: It is not easy to see these jumps in the figure. Consider highlighting them somehow (different colors?)

Line 240: replace the first “and” by “with”, and the second “and” by “or”

Line 261-262: This sentence is incomplete

Line 305: It is unclear whether this is the Pearson correlation coefficient (r) or the coefficient of determination (r^2). Based on the rest of the section I would guess the latter, but please specify and use the correct terminology.

Line 322-323: This sentence (beginning with “As the atmospheric …”) is incomplete

Line 351: earlier in the manuscript “we” is used – be consistent

Line 378: add “final” before GOBM

Line 390: “causes changes”

Figure 6: Add details in caption about the vertical line and shading in subplot c)

Figure 9: It is next to impossible to tell the lines on this figure apart. Please choose different colors. I would also recommend making the lines for individual models thinner.

Line 563: Are these the same five in every decade? Please specify and if not this warrants more discussion

Dr. Terhaar studies CMIP6 historical and future projections to assess the drivers of trends in the global ocean carbon sink. The author compares decadal trends in atmospheric pCO2 to decadal trends in the ocean sink, and proposes a 1 decade lag of the ocean behind the atmosphere. A mechanistic explanation for this lag is not offered. The author also proposes that this analysis of ESMs demonstrates that the decadal trends in pCO2 products are too large.

Major comments

• The a 1 decade delayed response of the ocean sink to trends in the atmospheric growth rate in CMIP6 is intriguing. But why? A proposed mechanism for this effect is missing in the manuscript. In McKinley et al. 2020, we show that change in the atmospheric growth rate impacts the ocean sink with no lag, due to the impact on the delta pCO2. Lovenduski et al. (2021) demonstrate this mechanism with CANESM5 for changes in the atmospheric growth rate consistent with COVID19. What is different mechanistically, or about the analysis performed here, that leads to a very different conclusion here? The author needs to address this directly.
• What is the impact on the findings of the adjustments to model output following Terhaar et al. (2022)? Dr. Terhaar and colleagues’ previous findings are interesting, but not conclusive. Others, such as Goris et al. (2018) propose alternative metrics for such a constraint. It is important to understand the impact of this adjustment on these results.
• The author focuses the introduction on the weaknesses of pCO2 products and GOBMs, but does not adequately acknowledge that ESMs also have weaknesses. The fact that the author will adjust and detrend the ESMs before doing his analysis needs to be acknowledged here, as just one example of a weakness. Please adjust this discussion to be more balanced.
• The author needs to be more precise about ESMs vs. GOBMs. Papers such as Gruber et al 2023 and the associated literature, as well as RECCAP2, focus on comparing pCO2 products to GOBMs, not to ESMs as indicated on Line 458 and below. Please check throughout the paper and make sure the discussion does not confuse.
• There is a lot of discussion of detailed features from the figures that are very difficult for the reader to see due to a lack of annotation. For example, on figures 1 and 2 where atmospheric CO2 concentration or growth rate are plotted against the ocean sink, the author discusses features at specific dates. It is not possible to see these dates on such a figure. The author needs to make sure the reader can identify the features he discusses.

Minor

Line 33 - 12 members is not a “large ensemble” in the common understanding of this terms. This would be at least many 10s of members.  Please revise.

Line 59 – Ridge and McKinley 2021 Biogeosciences should also be cited

Line 78 – Please add Gloege et al. 2022 JAMES, Bennington et al. 2022 GRL, Bennington et al. 2022 JAMES

Line 80 – Please add LaCroix et al. 2020

Line 87 – Says “pCO2 products” here, should be GOBMs

Line 93 – unclear to what “both products” refers

Line 94 – Gloege et al. 2021 did not use GOBMs; large ensembles of ESMs were used.  

Line 110 – Please reference McKinley et al. 2023 ERL as a study that considers the full CMIP6 suite.  

Line 193 – Clarify here briefly that 2014 is the end of the historical period of forcing - i.e. “After 2014, when SSP scenario forcing begins,… “ or similar

Line 218 – A square root relationship should be dependent on the units. Please include units. Please also mark this square root relationship on the figure

Line 225 – “ocean carbon sink does not go back close to zero but remains almost stable (Fig. 2b).” This cannot be easily seen on the plot

Section 2.1 Some of these ESMs provide multiple ensemble members to CMIP6. How are the ensembles used here? Just the first one taken? An average made? If the latter, then it could impact results by averaging out some of the internal variability of individual members, and this would need to be discussed. Please make this clear, and discuss any impacts on results.

Section 3.2, and Line 251-262. It is difficult to follow these discussions. Adding annotation on figures (as suggested below), increasing the size of the figures so that these features can be seen, and/or revising the text to more clearly describe the features being discussed.

Line 245-246. The first phrase of this sentence is incomplete, and also and above it was said there is a square root relationship. Please revise.

Line 260. Replace “done” with “down”

Line 264. It is not true that it is “not possible” to compare to a linear trend. For example, it would be possible to calculate linear trends for 30 years, and use this as comparison. The author needs to find a better way to justify the approach taken here.

Line 270. What is the mechanism of this delayed response?

Line 318. SSP1-2.6 scenario is not clearly labeled in figure 3a. Please ensure the reader can follow this discussion.

Line 390. Strike “to”

Line 455-482. ESMs may get the climate modes, but not necessarily the ocean carbon sink response to these modes.  The following discussion of biogeochemical vs physical driven variability in GOBMs and of higher resolution models does try to address this, but still it is not conclusive because these are still models being used as the point of comparison. There is circularity in this discussion that needs to be acknowledged – i.e. though sampling of pCO2 may lead the pCO2 products to overestimate decadal variability, it also remains possible that the pCO2 products are capturing real signals that we are not modeling.  

Line 486. “ESMs used here”

Line 523. Again, what is the mechanism of the decadal lag in the ocean response to the atmospheric growth rate?

Figure 1.

• Please add a legend on the figure. Please make the blue more clearly distinguishable as not black/gray.
• Consider marking 1920, 1960, 1990, 2000 on panel c, d. This is needed to more easily follow the discussion about “jumps” at line 189-191.

Figure 2

• It is not clear where 1920 and 1940 are in panels c and d; please these clearly mark on the plots

Figure 3

• The text suggests that here it is ocean sink in decade 2 compared to atm CO2 trend in decade 1, but it isn’t stated in this way in the caption. The caption suggests they are concurrent trends. This needs clarification.