Ice samples were obtained by drilling a shallow (20 m, 0.24 m diameter) ice core at Taylor Glacier, Antarctica, a blue ice area located in the McMurdo Dry Valleys (Baggenstos et al., 2017). The core contains ice spanning gas ages from ∼58 ka near the surface to 74 ka at 20 m depth (Menking et al., 2019). We excluded samples above 4 m depth to avoid alteration/contamination due to near-surface thermal fractures (Baggenstos et al., 2017). A total of 56 samples (including 11 replicate samples from identical depths) from Taylor Glacier were measured, with an average sample weight of 806 g and a mean record temporal resolution of 330 years. In addition, four WAIS (West Antarctic Ice Sheet) Divide samples from late MIS 4 (∼66–64 ka) were measured to replicate the Taylor Glacier results using samples from a different ice core. All ice core samples were analyzed for Kr/N2, Xe/N2, and Xe/Kr using the method described by Bereiter et al. (2018b). The average of the three noble gas ratios was used to determine the final MOT following procedures in Shackleton et al. (2019). For brevity, we will refer to the MOT reconstructed from measured noble gases as “MOT data” in this work.

We apply the ice core age model of Menking et al. (2019) with slight modifications for the MOT reconstruction. The age model was developed by matching measured variations in CH4 and δ18Oatm in the Taylor Glacier ice core to deep ice core records on the Antarctic Ice Core Chronology 2012 (AICC2012) timescale (Veres et al., 2013). Tie points were manually selected, and noble gas sample ages were determined from linear interpolation between tie points. For this study, we selected tie points from the higher-resolution North Greenland Ice Core Project (NGRIP; rather than EPICA Dronning Maud Land – EDML) CH4 record on AICC2012, as well as three additional tie points from the EDML CO2 record, also on AICC2012 (Table 1). Tie point uncertainties are reported relative to AICC2012 and do not include age uncertainty of the AICC2012 chronology itself. Tie point uncertainties have a minimal impact on the interpretation of the record.

The noble gas ratios measured in ice cores must be corrected for fractionation that occurs within the firn, which alters the noble gas ratios from their original atmospheric values (Headly and Severinghaus, 2007). We apply the correction approach of Shackleton et al. (2019), which uses a linear least-squares method to solve and correct for gravitational (Schwander, 1989) and thermal (Severinghaus et al., 1998) fractionations using measured isotope ratios of argon, nitrogen, and krypton. Further details on the applied fractionation corrections are included in Appendix A.

Corrections are more robust when calculating relative MOT change, rather than absolute MOT values, because errors in the fractionation corrections may produce a systematic offset in the corrected noble gas ratios, whereas the relative changes in these ratios are minimally influenced (Shackleton et al., 2020 and Appendix A). Therefore, we report MOT relative to Holocene MOT measured in the same ice core with the same applied method of fractionation correction. For the Taylor Glacier samples, we compare our data to five early Holocene (10.6 ka) replicate Taylor Glacier samples from Shackleton et al. (2020). WAIS Divide samples are reported relative to the average of Holocene samples from 11 to 10 ka (n=4) (Bereiter et al., 2018a). WAIS Divide (Bereiter et al., 2018a) and EPICA Dome C (EDC) (Baggenstos et al., 2019) records both suggest that the entire Holocene was a very stable interval for MOT (respective 1σ standard deviations of 0.2 and 0.1 ∘C for all Holocene samples), so the particular choice of reference Holocene interval has minimal impact on the reported results.

We employ the box model of Bereiter et al. (2018a), which calculates the MOT anomaly relative to the modern ocean from the firn-corrected noble gas ratios. Parameterizations of the MOT box model applied in this study are detailed in Baggenstos et al. (2019). In addition, we use the recently published xenon and krypton solubilities from Jenkins et al. (2019). The box model requires input of sea level (Grant et al., 2012) to account for changes in the oceanic reservoir of xenon, krypton, and nitrogen that are unrelated to ocean temperature change. This includes changes in ocean volume, salinity, and sea surface pressure (Headly and Severinghaus, 2007). For Holocene MOT reference data and the MIS 2 data against which the record is compared, we use the sea level record of Lambeck et al. (2014) in the MOT box model. We also reevaluate the WAIS Divide Holocene and MIS 2 MOT record (Bereiter et al., 2018a), applying the same box model parameterizations as applied in this study (and in Shackleton et al., 2020) for a consistent comparison.

To estimate the effect of a cooling ocean on the atmospheric CO2 concentration via the solubility pump, we use a simple carbon cycle model (Bauska et al., 2016) to run a forward scenario of prescribed ocean temperature change from MOT constraints. The model consists of 14 boxes, representing the surface oceans (6 boxes), intermediate oceans (2 boxes), and deep oceans (3 boxes), a well-mixed atmosphere (1 box), and a terrestrial biosphere (2 boxes). The model simulates thermohaline circulation and mixing, air–sea gas exchange, export production, sediment burial/CaCO3 compensation, and exchange of carbon between the atmosphere and terrestrial biosphere. More details on the model can be found in Appendix B.

The error on our MOT reconstruction is estimated by propagating all known uncertainties with 10 000 Monte Carlo simulations of our data. Sources of uncertainty include the analytical uncertainties for the noble gas ratios as well as the isotope ratios used for firn fractionation corrections. Additional uncertainties include the age uncertainty for the Taylor Glacier tie points (Table 1) and temporal and analytical uncertainties in the sea level curve. The calculated uncertainty for an individual sample is 0.2 ∘C (1σ), and the pooled standard deviation of the 11 replicate samples is 0.3 ∘C. Systematic errors (such as changes in ocean saturation state) may cause us to underestimate total uncertainty in our record.

To produce the splined MOT record, we use the 10 000 Monte Carlo iterations of the dataset and randomly sample the 56 individual MOT points via bootstrapping using the “randsample” MATLAB function with replacement. We then fit each of the 10 000 time series using a spline with a 2500-year cut-off period using the “csaps” MATLAB function and average the resulting splines to produce a final, smoothed version of our MOT record including uncertainty estimates.

During the rapid drawdown in atmospheric CO2 (∼72–68 ka, Fig. 2), we observe mean ocean cooling in two phases, with an overall net cooling of 0.9 ± 0.3 ∘C (1σ). In the first phase (72–70 ka), MOT decreased by 0.7 ± 0.3 ∘C over roughly 2 kyr, coincident with Antarctic cooling and Greenland Interstadial 19 (GI19). In the second phase (70–68 ka), MOT stabilized and then decreased by a further 0.2 ± 0.3 ∘C, reaching a minimum around 67.5 ka.

Here, we do not use the intervals identified and defined by benthic δ18O to compare MOT in MIS 4 and MIS 2, as the alignment of ice core and sediment records is uncertain, particularly in MIS 4. Instead, we define MIS 4 as the interval in which CO2 and Antarctic temperature remain low and stable (70.3–63.7 ka, or Greenland Stadial 19 and Greenland Interstadial 18). For Taylor Glacier samples, we compare MIS 4 samples to five replicate MOT samples from MIS 2 (19.9 ka). For WAIS Divide samples, we compare the measured MIS 4 samples to all available, previously published (Bereiter et al., 2018a) MOT data from MIS 2 (24–18 ka) with the fractionation corrections and MOT box model parameterizations used in this study applied. The difference in WAIS Divide MOT results for the full MIS 2 interval (n=11) versus 20–19 ka (n=4) is less than 0.01 ∘C, so the difference in the selected intervals to define MIS 2 for each core should not affect the MIS 4–2 comparison.

The Taylor Glacier data show that MIS 4 MOT was statistically indistinguishable from MIS 2 (-0.1 ± 0.3 ∘C relative to MIS 2 or -2.7 ± 0.3 ∘C relative to the Holocene). The four WAIS Divide MIS 4 samples cover a narrower interval (66.2–63.9 ka) but are consistent with the Taylor Glacier results (+0.1 ± 0.5 ∘C relative to MIS 2 or -2.6 ± 0.5 ∘C relative to the Holocene). However, the WAIS data show more scatter. If we instead correct the WAIS Divide data for thermal fractionation using a firn model (Buizert et al., 2015), as in Bereiter et al. (2018a), and compare the results to MIS 2 data using this method of fractionation correction, we find that the WAIS Divide MIS 4 data are slightly less scattered (Fig. 2). With these corrections applied, the MIS 4 interval in WAIS Divide is 0.2 ∘C warmer than in MIS 2 (-2.5 ± 0.3 ∘C relative to the Holocene). While all results are indistinguishable within error, they emphasize the importance of future work developing further understanding of firn air processes and their influence on MOT results. A detailed discussion of the choice in fractionation corrections and their effect on calculated MOT is included in Appendix A.

While our record may not capture the full transition into MIS 3, we find that there was a substantial increase in MOT towards the end of MIS 4. By 59.7 ka, MOT had reached levels comparable to the MOT peak at the end of MIS 5a at 72 ka (Fig. 2). Because our record does not contain a clear leveling of MOT, it is uncertain if or by how much MIS 3 MOT exceeded levels found at the end MIS 5a.

While the MOT proxy was developed over a decade ago (Headly and Severinghaus, 2007), only in the last few years have high-resolution MOT records become available (Baggenstos et al., 2019; Bereiter et al., 2018a; Shackleton et al., 2019, 2020). With the additional data from this study, we take the opportunity to review available MOT records and their relation to other key climate variables with a particular emphasis on the glacial inception (Fig. 3). While the available MOT records do not cover the entire last glacial cycle, we may still gain insight into climate evolution during (de)glaciations by comparing contemporaneous MOT, benthic δ18O, CO2, and Antarctic δ2H.

Given that global ice volume was greater (Cutler et al., 2003) and CO2 concentrations were lower in MIS 2 than in MIS 4, the comparably cold conditions during these intervals suggested in this study and by recent (Doughty et al., 2021) and previous (Kohfeld and Chase, 2017; Snyder, 2016) work is somewhat puzzling. All else being equal, the 20 ppm higher atmospheric CO2 in MIS 4 would lead to 0.5 ∘C warmer global average surface temperatures than in MIS 2, assuming a climate sensitivity of 3.5 ∘C. Other forcing may be required to resolve this conundrum. While changes in planetary albedo are often assumed to scale with ice volume, this may not be appropriate when comparing MIS 4 and MIS 2. Doughty et al. (2021) suggest that while ice volume was lower, glacier extent may have been greater in MIS 4 than in MIS 2. This may have led to a higher planetary albedo in MIS 4 than in MIS 2. However, the authors suggest that the greater glacial extent was due to the cold conditions in MIS 4, so there is some circularity to this argument.

Another notable difference between MIS 4 and MIS 2 exists for the Saharan region. Proxy records of the Sahara suggest that MIS 4 was a uniquely arid interval within the last glacial cycle (Castañeda et al., 2009; Skonieczny et al., 2019; Tierney et al., 2017), while regions of the eastern Sahara and Sinai Desert during MIS 2 may have been wetter than today (Hamdan and Brook, 2015), suggesting greener Saharan conditions in MIS 2 compared with MIS 4. Climate simulations of a greener Sahara suggest globally warmer temperatures due to lower albedo and higher atmospheric moisture (Tabor et al., 2020). We speculate that more arid Saharan conditions in MIS 4 may have contributed additional cooling of MIS 4 relative to MIS 2 and, in part, countered the warming effect of higher CO2 during this interval. Of course, this new MOT record cannot shed light on the conditions of the low-latitude hydrosphere. However, it adds to a growing body of work suggesting that, despite the smaller ice volume and higher atmospheric CO2, MIS 4 was comparably cold to MIS 2. We believe that this conundrum merits further investigation and may be a valuable target for forthcoming climate modeling efforts.

In addition to their cold temperatures, MIS 4 and MIS 2 also share an absence of millennial-scale variability. Theories explaining the apparent lack of bipolar seesaw behavior during very cold periods (such as MIS 4 and MIS 2) have invoked mechanisms related to thresholds in ice volume (McManus et al., 1999) and Southern Ocean temperature (Buizert and Schmittner, 2015). While ice volume during MIS 2 exceeds that of MIS 4, MOT during MIS 2 and MIS 4 indicate equally cold ocean conditions. This supports the idea that thresholds in ocean temperature, rather than global ice volume, may determine the presence or absence of millennial-scale variability within a glacial.