The study by Gao et al. performed a 3000-yr 128-ka simulation to highlight the importance of the meltwater release into the North Atlantic in reducing the model–data discrepancies over the Southern Ocean. The manuscript is clearly written, and the topic fits the CP. However, there are several concerns that should be addressed.

Major comments:

1. In section 2.2, the authors declared that they selected the most recent surface temperature data syntheses and SIC dataset. I am wondering the reason for ignoring the data from Turney et al. (2020), as they also provide annual SST data during the early LIG. If there is no special reason, comparisons between the SST data from Turney et al. (2020) and models should be added.

[Reference]

Turney, C. S. M., et al. (2020). A global mean sea-surface temperature dataset for the Last Interglacial (129–116 kyr) and contribution of thermal expansion to sea-level change. Earth System Science Data 12(4): 3341-3356.

2. In section 2.3, the experimental setups are not clear. The boundary conditions and external forcings (e.g. greenhouse gases, orbital parameters, ice sheet, vegetation, land-sea mask) at 128 ka need to be displayed or indicated directly.

3. In section 3.1, the authors evaluated the performance of PMIP4 models in simulating SST over the Southern Ocean, but they only provide the RMSE. The spatial correlation coefficients between PMIP4 models and HadISST1 dataset should also be provided. I suggest providing a Taylor diagram to show the performance of PMIP4 models more clearly.

4. Across the manuscript, the authors suggested that the orbital parameters, greenhouse gases and Antarctic ice sheet played a limited role in the warming of the Southern Ocean during the early LIG, and attributed the warming to the meltwater release into the North Atlantic. Zhang et al. (2023) indicated that the global sea level rising during the LIG (at 126 ka) can also reduce the model-data discrepancies over the Southern Ocean. I suggest that more discussions about this point need to be added.

[Reference]

Zhang, Z., et al. (2023). Atmospheric and oceanic circulation altered by global mean sea-level rise. Nature Geoscience 16(4): 321-327.

5. The authors indicated that long-time simulation (i.e. the 3000-yr simulation) of H11 is likely required to capture the full magnitude of the Southern Ocean and Antarctic warming, following the guideline of the modeled linear SST trend in a 1600-yr simulation by Holloway et al. (2018). In Figure A1, the linear trend of the SST from 0 to 1600 years is significant. However, from 1600 to 3000 years, the SST seems to fluctuate near a mean state, rather than shows a long-term increasing trend. So I am wondering the necessity of running a long-time simulation in reconciling the model-data mismatch based on the current results. I suggest providing the difference between the “short” simulation and “long” simulation at 128 ka to highlight the novelty of long-time simulation performed in this study.

Minor comments:

Line 87: Four most recent surface temperature data syntheses, may be three?

Last interglacial climate simulations provide an interesting insight into a warmer climate state, albeit under different orbital and greenhouse gas forcings than those of present-day. Here the authors address a notable issue in the PMIP4 last interglacial simulations, which show very little Southern Ocean SST warming in comparison to proxy records. The limited improvement between PMIP3 and PMIP4 in that respect highlights the work still to be done by both the proxy and modelling communities to resolve these differences. However, by adding additional freshwater into the North Atlantic in a classic hosing experiment, the authors find considerable improvement in the simulation of last interglacial Southern Ocean SST, at least in their model (HadCM3).

This is an interesting experiment fitting the scope of CP, and is of particular relevance given increasing interest in hosing experiments for both future and past climates. Nevertheless there are some important weaknesses in the design of the experiment that I believe should be addressed, as follows.

(1) Antarctica will also have contributed anomalous freshwater fluxes during the LIG. These may actually act against your northern influence, for example by cooling SST (Mackie et al. 2020 and several similar studies). Recent work has suggested Antarctica’s sea-level contribution was early in the LIG, e.g. before 126 ka (Barnett et al, 2023). In that case the Antarctic freshwater fluxes may have been just as important as North Atlantic freshwater fluxes in their influence of Southern Ocean SST. Even though I understand the logistical advantage of simply continuing an existing model run, and the computational constraints, this point needs some careful discussion in your manuscript as it is (in my view) a major weakness of your experiment. How do the magnitudes of Arctic freshwater-induced cooling compare with Antarctic freshwater-induced warming in the Southern Ocean? Previous modelling studies can help answer this. Bearing in mind that sea-level rise could also help explain the apparent cold bias in PMIP4 lig127k Southern Ocean SST (Zhang et al. 2023).

(2)  Comparison of simulated and reconstructed SST. See Section 2.5. Here the implication is that the simulated SST is being tested against the “truth”, which here comprises SST reconstructions. However, reconstructions themselves have considerable uncertainty as discussed in Chandler & Langebroek (2021a,b). Their recommendation was to focus comparison on regional averages, rather than site-by-site. I would suggest to follow that approach in this paper and (for example) use the proxies to get three regional SST anomalies (Atlantic, Indian, Pacific sectors) then evaluate your results on a regional basis rather than site-by-site basis.

(3) Use of HadISST1 1870-1899 as a PI reference dataset. Probably this has some precedent in other studies, but as acknowledged in Section 2.4 there are very sparse observations for the Southern Ocean during 1870-1899. Consequently this comparison is not very convincing or useful without a lot more information about the errors in this region, in HadISST1. I suspect they will be fairly substantial! Why not use a more recent period? For example CMIP historical simulations. I’m fairly sure the models contributing to lig127k all have a historical simulation using the same configuration as PI.

(4) Again in the comparison with reconstructions, the Turney et al 2020 ESSD paper is a surprising omission here. At first it might seem the Turney et al. methodology makes direct comparison with lig127k a bit trickier (Turney et al. report a “LIG average”, not a time slice). However, they also report a 129-124 ka peak warmth. I would suggest using this, since it’s actually very similar to the Capron et al. methodology in practice. This is because Capron et al. aligned their SST records to EDC air temperature, under the assumption of synchronous changes across the entire region. Consequently, their SST peaks are synchronous across all records and so the peak warmth in their study is always at 128 ka. This lies within the 129-124 ka window used by Turney et al. If you prefer to avoid too many comparisons, maybe swap the Capron et al dataset for the newer Turney et al. dataset.

Other points follow below.

L5: are all 99 reconstructions independent, i.e., are they 99 different proxy records?

L10: Again from earlier comment, note that anomalous FW into SO might contribute some cooling (Mackie et al. 2020 & many others)

L30, L88: I’d agree that estimating LIG warmth by compiling peak temperatures is not an appropriate approach. However, the SST records used by Capron et al. 2014/2017 were aligned to the EDC ice core temperature record, such that their synthesis also  implicitly represents a synthesis of ‘peak warmth’ as noted in my earlier comment. Hoffman et al. followed a mixed approach (three key records representing three main ocean basins were aligned to EDC, but other records in each basin aligned to key record by d18O. Specifically from Capron et al: “Marine records are transferred onto AICC2012 by assuming that surface-water temperature changes in the sub-Antarctic zone of the Southern Ocean (respectively in the North Atlantic) occurred simultaneously with air temperature variations above Antarctica (respectively Greenland)”.

L118: Extraction of 127 ka anomaly. Why use 128 ka for the 127 ka anomaly? Surely better to use an average of 128 and 126 ka?

L145: Different characteristics of Chadwick et al 2021 might also reflect that they only used diatoms, whereas other datasets used multiple proxies.

L156: I’d suggest to recap what are the key parameters/modelling choices used by Holloway et al. 2016 – presumably you keep these the same?

Fig 1: Can this be split into four rows, for the four studies, otherwise symbols plot over each other in a jumble.

Table 2 caption: useful to specify here again what is the PI reference used for the temperature anomalies.

L164 ‘afflicted’… ‘affected’?

L181: largest error, not largest bias. RMSE and bias are not the same thing. Which is an important point: you are ranking the datasets on their RMSE, not their bias. I think the bias should also be reported along with the RMSE.

Table 3: what are the reported error bounds in the lig127k vs piControl anomalies?

Table 4 headings: suggest to include RMSE specificically in the headings, i.e, Annual SST RMSE, Summer SST RMSE, etc otherwise this looks like a table of actual temperatures rather than temperature errors.

Figs 3,4,5,6: “We only show differences that are significant at 5% level based on the student’s t-test with Benjamini-Hochberg Procedure controlling false discovery rates (Benjamini and Hochberg, 1995).” What does this refer to? I can’t see where differences are illustrated.

L218 Null scenario: ("To benchmark model performance, we introduce a Null Scenario, where the climate at 127 ka is assumed to be the same as the preindustrial climate. … To demonstrate a better performance than the Null Scenario, the model simulations must have a smaller RMSE when evaluated against the climate syntheses than the Null Scenario"). Can demonstration of the ‘Null Scenario’ be defined more clearly? In particular there could be some confusion here about what is being compared with what. Is the Null Scenario the piControl simulation compared with HadISST1? This approach is problematic because of potentially large errors in HadISST1 1870-1899 PI as noted above for Sec 3.1. Overall this confusion makes subsequent discussion somewhat difficult to evaluate.

Fig 8: vertical error bars for the PMIP3 and PMIP4 ensembles would be useful, e.g. to show the standard deviation.

L272: Again evaluation against 1870-1899 HadISST1 is not a good benchmark given the likely errors in that dataset (particularly lack of data in S Ocean) & I would suggest to evaluate CMIP historical simulations against more recent observations in

L286: “The ocean south of 40S becomes warmer…”. Here and possibly several other places it is better to stick to “SST” rather than “ocean temperature” as it’s only the sea-surface temperature that is being analysed. Suggest: “SST south of 40S becomes warmer…” etc.

L299: Pros/cons of using core-top reconstructed SST or observed SST … Of particular relevance here, not all the LIG SST reconstructions have a core-top sample.

L368: Acknowledgements – as well as acknowledging PMIP contributors, maybe nice to also acknowledge the many authors who have made their SST reconstructions available on public repositories.

Refs

Barnett et al. 2023, Science, https://doi.org/10.1126/sciadv.adf0198.

Capron et al. 2017, QSR, https://doi.org/10.1016/j.quascirev.2017.04.019.

Chandler & Langebroek (2021a), QSR, https://doi.org/10.1016/j.quascirev.2021.107190.

Chandler & Langebroek (2021b), QSR https://doi.org/10.1016/j.quascirev.2021.107191.  

Mackie et al. 2020, J. Clim, https://doi.org/10.1175/JCLI-D-19-0881.1.  

Turney et al. 2020, ESSD, https://doi.org/10.5194/essd-12-3341-2020.

Zhang et al. 2023, Nat Geos, https://doi.org/10.1038/s41561-023-01153-y.

Gao et al compared PMIP4 simulations for the Last Interglacial (LIG, 127 ka) with existing paleoclimate syntheses of sea and air temperatures, and sea ice concentration. The authors found that the warming recorded in the paleoclimate data cannot be captured by LIG model simulations. Aiming to explain the large model-data discrepancy, the authors also performed a North Atlantic freshwater hosing simulation on 128 ka and found that this simulation better agrees with the paleoclimate data syntheses.

The authors are trying to reconcile the data-model discrepancy for the LIG, which is suitable for Climate of the Past. However, both parts of the study have some major issues that need to be addressed before being considered for publication.

Data syntheses

The authors acknowledge some limitations of using different LIG data synthesis (Sec 4.1). However, the quality of the employed data syntheses (mainly SSTs) needs to be substantially improved to make the model-data comparison meaningful. In the supplementary data table, 127-ka annual SST anomaly appears to be unrealistic (11.5 C at site ODP 1089, >5 C at E49-17 and MD02-2588, -6.8 C at MD73-025, 7.7C at MD97-2120). The SST anomalies at the same site based on different studies are drastically different (e.g., DSDP 594, MD88-770, MD97-2120, etc).

The quality of the data syntheses can be significantly improved by revisiting the original data to resolve potential issues associated with inconsistent age models, different proxy calibrations, and the core-top SST values. These issues, mentioned by the authors, needed to be addressed.

The SST reconstruction at any point e.g., at 127 ka can be subject to uncertainties associated with measurements. Using the average over a period (e.g., 125-128 ka) can reduce the impact of such an influence.

To derive the 127-ka SST anomaly compared to pre-industrial from paleo data, Holocene SST changes need to be considered too. This is because the core-top ages at many sites are not late Holocene.

For SSTs, I suggest focusing on 1 data synthesis taking some of the above issues into account, and adding more sites following the same criteria.

These are additional work but are necessary to make the model-data comparison meaningful.

Hosing model simulation

The hosing experiment indeed reduced the RMSE between model and data. However, this improvement cannot be attributed to the 3,000-year hosing, without comparing 128 ka simulation without hosing with paleo data. And the length of the hosing period cannot be justified without comparing the 1,600-year hosing experiment (Holloway et al 2018 from the same group) to the same paleo data.

Additionally, I doubt if RMSE is suitable for evaluating model-data agreements. It appears that the larger RMSE is driven by the systematic offset between data and model. Investigating RMSE ignores spatial patterns of warming in different regions.

Some minor points

Line 76: more details are needed. How do these parameters differ from pi control?

Line 79: why mention the CNRM model specifically here?

Line 89: Annual SST and summer SST in the paleo data syntheses are often derived from the same dataset but with different calibrations (e.g., alkenone, see Chandler and Langebroek 2021). Therefore, these two SSTs in paleo data sets are not independent. This point should be made in the methods.

Line 149: Forcing parameters for the 128 ka simulation need to be described in detail for comparison with the 127 ka simulation

Table 3: Good to add mean deference between pi control and HadSST1. From Fig. 2, the mean difference can be large for some models. How does this contribute to offset between the model (comparing 127 ka with pi) and data ( comparing 127 ka with HadlSST1)?

Line 201: if you mean statistically significant, show the statistics.

Line 215: How many is “a few”