We have compiled a global network of published quantified SSTs using faunal and floral plankton assemblages, Mg / Ca and Sr / Ca ratios of calcareous organisms, and U37K′ estimates across the period of record interpreted as representing the LIG. In many instances, we used the period represented by low 18O values in benthic foraminifera shells (the lightest isotopic values during 90–150 ka representing minimum global ice volume). In some sequences, δ18O values were not reported, and therefore we relied on other complementary proxies, for instance, the CaCO3 content of sediments as a measure of glacial–interglacial variability (Turney and Jones, 2010; Cortese et al., 2013) (Figs. 1 and 2). Whilst the age control points defining the plateaus in δ18O and other proxies are not absolutely dated with chronological uncertainties of 1 to 2 millennia (Martinson et al., 1987; Lisiecki and Raymo, 2005), it is important to note that we are not aiming to resolve centennial- and millennial-scale variability through the interglacial. We acknowledge that some individual SST estimates may not fall within the LIG or have been excluded (due to these chronological uncertainties), but we consider the averaging of values across the full interglacial to provide a robust value for each record and ultimately the regional and global reconstructions.

We have therefore not attempted to generate a time series of SSTs through the LIG. Previous studies have highlighted that individual site δ18O changes in benthic foraminifera (for instance, during deglaciation) may be offset by several millennia as a result of local deep-water temperature and δ18O seawater variations (Govin et al., 2015; Waelbroeck et al., 2008) (Fig. 2). In an attempt to bypass some of these issues, other studies have attempted alignment of marine records to speleothem-dated ice core reconstructions (Hoffman et al., 2017), but modelled age uncertainties can be in the order of millennia (e.g. Hoffman et al., 2017, Fig. S7), while the assumed synchroneity of extra-regional changes has challenges; for instance, more than half of reported Pacific marine cores (those from the Northern Hemisphere) in a recent study were correlated to the Antarctic EPICA Dome C δD (Hoffman et al., 2017), with warming in the south known to lead the north by 1 to 2 millennia (Hayes et al., 2014; NEEM Community Members, 2013; Kim et al., 1998; Rohling et al., 2019). The development of accurate and precise age estimates for LIG records is urgently needed to resolve the timing of global climate change but will require a considerable future international effort (Govin et al., 2015). Given the relatively large chronological uncertainties associated with comparing global SST time series (Hoffman et al., 2017; Govin et al., 2015; Capron et al., 2017), we have therefore not attempted to generate a time series of changes within the LIG but instead determine average temperatures as a robust estimate of mean climatic conditions. Whilst not offering precisely dated geochronological frameworks, the global ice volume minima as represented by the δ18O plateau and/or associated proxy measures of interglacial conditions are sufficiently well-defined in all marine records to accommodate local deep-water temperature and δ18O variations, sampling resolution, and/or sedimentation rates to identify the LIG, thereby maximizing the number of records that have reported quantified SSTs across the interglacial (Cortese et al., 2013; Govin et al., 2015); a minimum of three SST values across the LIG in each record was required for inclusion in our dataset. This is not to downplay the significance of millennial-scale climate variability across the LIG (Galaasen et al., 2014; Rohling et al., 2002; Tzedakis et al., 2018; Jones et al., 2017), but our approach does provide some benefits. Whilst our approach sacrifices temporal control, it does minimize the uncertainty of zonal and global temperature averages.

To quantify the temperature difference between the LIG and present day, we do not compare the LIG estimates to the relatively poor observational coverage of earlier periods, including the 19th century (pre-industrial) (Hoffman et al., 2017) or the long-term annual means calculated from 1900 to 1997 (Capron et al., 2014), both of which have considerable uncertainties given the limited network of “observations” prior to the satellite era (Brohan et al., 2006; Huang et al., 2020). Here instead we report SSTs expressed as anomalies relative to global “modern” instrumental and satellite observations across the period 1981–2010 obtained from HadISST (Rayner et al., 2003). Each LIG temperature record is linked to at least one literature source, the citation of which includes author(s), year of publication, and typical archiving information (e.g. journal, volume, issue, pages, publisher, and place of publication). Where multiple temperature estimates have been published over time from the same site, we chose the most recent publication for inclusion in the database (so long as the data were not flagged as erroneous) (Fig. 3). Note that alkenone proxies are interpreted as providing annual SST estimates.

Here we use the mean temperature estimates to constrain the role of thermal expansion in global sea level rise across the LIG and provide boundary conditions for future modelling studies investigating the impact of warming on polar ice sheets. To determine the greatest possible contribution of warming to ocean thermal expansion and ice sheet melt, we used the published age models to identify the maximum annual SST within the first 5 kyr of the LIG (i.e. 129–124 ka). For the purpose of this sensitivity analysis, the maximum temperatures were assumed to be synchronous globally, a scenario we recognize as unlikely but which does provide an upper limit for warming in the “early” LIG. To provide an upper estimate on the magnitude of warming in polar waters over the deglaciation, we also report here the difference between late Marine Isotope Stage 6 mean SSTs (∼ 140–135 ka) and the maximum early LIG SSTs for marine cores in the mid- to high-latitudes. To calculate the anomaly relative to present day, we utilize SSTs from the nearest 0.5∘ latitude × 0.5∘ longitude averaged across the period 1981–2010 (Rayner et al., 2003). For the uncertainties calculated for the regional and global SST anomalies, we incorporate the uncertainties from the proxies (reported in the database), and the uncertainties associated with estimating regional and global temperatures from limited spatial coverage. To achieve this we propagated the SST uncertainties for each measurement through each of the averaging steps (i.e. temporal to grid cell to zonal to area-weighted global) in our ocean-area-weighted average (McKay et al., 2011). We used quoted uncertainty estimates for each study where reported; if not available, we applied proxy-specific uncertainty estimates. Although the impact of the spatial coverage was not explored in this study, it has been previously estimated using the same approach (McKay et al., 2011). In that study, the uncertainty associated with the limited spatial range of the oceanographic proxies was estimated by calculating 1000 random 1-year global SST anomalies over the 20th century and comparing them to averages derived using only the palaeoceanographic network. No systematic biases were identified, with a 1σ uncertainty estimated to be < 0.1 ∘C. In this study, we have expanded the spatial network and consider ±0.1 ∘C to be a reasonable, high-end uncertainty estimate.

The database comprises six worksheets of data comprising maximum annual temperatures during the early LIG (defined here as the maximum temperature reported within the first 5 millennia of the LIG; 129–125 ka), mean annual temperature, the Marine Isotope Stage 6/5 SST difference, December to February temperature (DJF; Northern Hemisphere winter and Southern Hemisphere summer), June to August temperature (JJA; Northern Hemisphere summer and Southern Hemisphere winter), and summary statistics (see the Supplement).

The early maximum and mean annual SST dataset comprises 189 marine sediment and coral records from latitudes spanning from 55.55∘ S (radiolaria assemblage transfer function reconstruction obtained from site V18-68) (CLIMAP, 1984) to 72.18∘ N (planktonic foraminifera assemblage modern analogue technique from site V27-60) (Vogelsang et al., 2001).

Crucially, modern calibration relationships are an average developed using a selected number of locations that will not necessarily capture the range of “signal drift”. This drift is caused by the fact that planktic SST recorders can be transported over considerable distances in the water column before being deposited, which particularly applies to all those sites that lie under strong boundary currents or near major ocean fronts (van Sebille et al., 2015). Unfortunately, Ocean General Circulation Models (OGCMs) typically have insufficient spatial resolution to capture mesoscale features that are critical for modelling the lateral drift of particles (Nooteboom et al., 2020). To investigate the impact of drift on SST reconstructions, we therefore used contemporary ocean circulation as a first-order approximation for the LIG. Whilst we acknowledge that there was likely a weakening of the Atlantic Meridional Overturning Circulation (AMOC) during the early LIG (Shackleton et al., 2020; Turney et al., 2020a; Thomas et al., 2020; Jones et al., 2017), subsequent recovery after 127 ka appears to have established a global circulation comparable to present day as suggested by recent ocean δ13C modelling results across the mid-interglacial (Bengtson et al., 2020). We performed an experiment with virtual particles in an eddy-resolving ocean model (the Japanese Ocean model For the Earth Simulator, OFES) (Masumoto et al., 2004), which has a 1/10∘ horizontal resolution and near-global coverage between 75∘ S and 75∘ N (van Sebille et al., 2012). Utilizing the 3D velocity field of the model, we used the Parcels code (http://oceanparcels.org/, last access: 4 December 2020; Lange and van Sebille, 2017) to compute the trajectories of more than 170 000 virtual planktic particles that end up at each of the sites by tracking them backwards in time, first simulating the sinking to these sites at 200 m d-1 and subsequently the advection at 30 m depth for a lifespan of 30 d; coral SSTs were not corrected for drift. Given the lifespan of most organisms that have been used to generate a temperature signal (Jonkers et al., 2015; Bijma et al., 1990), we consider a 30 d drift provides a reasonable estimate of the drift distance. Previous work has demonstrated comparable uncertainties between different models (van Sebille et al., 2015), providing confidence in the use of the OFES for the purposes of this study.

During the 30 d lifespans, we recorded the temperatures along the trajectories and compared those to the local temperature at 30 m water depth at the site where the particles would end up on the ocean floor. This resulted in daily temperature anomalies along the trajectories, which were averaged through the lifespan and over the 840 virtual particles that ended up at each site and then subtracted from the reported LIG estimates (Fig. 1 and Database). With the recent recognition that core-top calibrations may be incorrect given historic changes in marine communities that have accompanied anthropogenic warming (Jonkers et al., 2019), it should be noted that SST proxy calibrations based on regional core-top calibrations may give an incorrect absolute value, an aspect that will form the focus of future work.

Global mean SST anomalies were calculated by averaging anomalies in a 10∘ latitude × 10∘ longitude grid and then averaging them globally after weighting for the area of ocean in each grid cell (Fig. 5). The uncertainty calculated for global SST anomalies incorporates uncertainties in the SST proxies as reported in the original studies, which typically range from 1 to 2 ∘C, and is then propagated through subsequent steps in the analysis. Additional uncertainty associated with estimating global anomalies from limited spatial coverage and the potential impacts of age uncertainty or averaging non-synchronous data are not considered here. Consequently, the derived estimates do not capture all of the uncertainty in global and zonal SST anomalies; however, the zonal consistency of the results suggests that the signal is large enough to overcome these unquantified sources of uncertainty. Furthermore, whilst some regions may exhibit substantial differences arising from drift (Fig. 4), globally the mean annual temperature estimates are comparable (Fig. 5). The new LIG SST dataset allows us to report the estimated thermosteric contribution to LIG sea levels using the method reported by (McKay et al., 2011). We use the above temperature changes to calculate the thermosteric contribution to LIG sea levels by using the Thermodynamic Equation of Seawater 2010 (TEOS-10). To provide an estimate of thermosteric sea level rise, we explored a range of scenarios where warming penetrated different ocean depths: 700 m, 2000 m (approximately the upper half of the ocean), and 3500 m (the whole ocean). We determined the change in the specific volume of the warmed water column of each 10∘ latitude × 10∘ grid cell while holding the salinity constant and neglecting changes in ocean area. Here absolute temperature is considered, as specific volume is more sensitive to temperature changes at warmer temperatures.

The Last Interglacial SST database is derived from published articles that have already been peer-reviewed. To generate the database, we undertook a comprehensive check to remove duplicate records, erroneous location information, and other errors. In addition, to ensure consistency of data processing and any recalculations (for instance, sea surface temperature anomalies relative to the period 1981–2010 CE), we also checked uncertain metadata reported for individual sites, and directly communicated with selected article authors and/or other experts as part of the record-validation process.

A challenge for the Last Interglacial is determining what influence (if any) ocean circulation had on the temperatures experienced (and reconstructed) by organisms that are used to generate SST reconstructions. Addressing this issue is an important objective of the current study, but we found the magnitude of temperature offset (bias) is limited to only a few key locations (Fig. 1), with similar final reconstructions for individual sites, and latitudinally averaged and globally averaged temperatures (Figs. 4 and 5, and Table 1). This provides an important check of our temperature recalculations. As a sensitivity test, we therefore explored virtual planktic particles that “live” for 30 d to investigate whether a prolonged period of drift made a discernible difference (data not reported here). Only a few species have been suggested as living for a longer period of time. For instance, in laboratory experiments the planktic foraminifer Neogloboquadrina pachyderma sinistral has been shown to survive up to 230 d (Spindler, 1996), but this species may be an exception due to its ability to survive in sea ice (Dieckmann et al., 1991).

Using 30 d drift to simulate the travelling time and lifespans of virtual planktic particles in the upper part of the water column, we quantified the inherited temperature signal of flora and fauna at each site in the database. The virtual microorganisms with a 30 d “lifespan” travelled from a few dozen to a few hundred kilometres. The temperature offsets are almost all positive in the tropical eastern Pacific, the North Atlantic, and the South China Sea, meaning that the planktic particles originated from warmer climates and hence record a higher temperature estimate than local conditions would suggest; the opposite effect was observed in the tropical western Pacific and Southern Ocean (Fig. 1). The offset can be substantial – with values ranging from -6.9 ∘C for site MD98-2162 at 4.7∘ S in the tropical western Pacific (Visser et al., 2003) and up to 3.5 ∘C for site RC13-110 on the Equator (Pisias and Mix, 1997) – with the largest changes associated with boundary currents and major ocean fronts. Intriguingly, these values are comparable to the difference previously reported for Mg / Ca foraminifera core-top calibration with those obtained from laboratory-cultured Mg / Ca calibrations (Elderfield and Ganssen, 2000; Hönisch et al., 2013). Both the uncorrected and 30 d drift temperatures are provided in the database. These temperature reconstructions led to statistically indistinguishable global temperature (and thermosteric sea level change; Fig. 5). Users of the database are therefore able to use either the authors' original sea surface temperature determinations or our drift-corrected estimates, as required.

To evaluate potential biases in our analysis, we further subsampled our database by proxy type (Fig. 4). The large network of sites and proxies does not appear to demonstrate any significant offset in annual reconstructions (at least within the uncertainty of the reconstructions), although there is a tendency for alkenone temperatures to be at the upper end of the range, implying there may be a seasonal bias, as reported previously (Hoffman et al., 2017). Importantly, we also compiled seasonal quantified temperature estimates that have been reported as the seasonally warmest or coolest months in the year (taken here to represent June–August and December–February depending on the hemisphere being considered). Our result suggests that any bias, if real, is smaller than the uncertainties at the global or zonal level reported here. Intriguingly, the warmest month estimates for the high latitudes in both hemispheres show more muted warming than the mean annual estimates, while the low to mid-latitudes exhibit considerably cooler estimates (Table 1). In contrast to the alkenone estimates for the annual estimates, the more muted response of foraminifera, radiolaria and diatoms for the seasonal reconstructions implies that they are influenced by a larger part of the seasonal cycle. We therefore consider that seasonal reconstructions should be treated as conservative estimates of temperature for the LIG.

We find global average annual temperatures across the full duration of the LIG were only marginally warmer than present day. We derive a global mean annual temperature anomaly of 0.2 ± 0.1 ∘C, the same value obtained after correcting for drift (Table 1). These values, however, mask considerable zonal differences, with significantly cooler mean annual uncorrected temperatures (i.e. not corrected for drift) within 23.5∘ of the Equator (-0.3 ± 0.2 ∘C) and amplified warming polewards (Fig. 5). Ideally, we would have a dense network of records in the mid-latitudes to high latitudes for investigating the impact of warming surrounding polar ice sheets, but unfortunately the number of sites and their spatial distribution does appear to have an impact on the reconstructed values. Comparison of the SST anomalies poleward of 45 and 50∘ latitude (Table 1) shows substantial differences, most notably in the Southern Hemisphere where a large increase in zonally averaged SST occurs alongside a decrease in the number of records polewards of 50∘ S (Table 1). For instance, the drift-corrected SSTs for the LIG are 0.8 ± 0.3 ∘C (n=13) and 2.7 ± 1.1 ∘C (n=3) polewards of 45 and 50∘ S, respectively. It should also be noted that while the Northern Hemisphere polar estimates are similar for both latitudinal ranges, the majority of sites are in the North Atlantic, with limited representation in the Pacific Ocean (Fig. 1). We therefore recommend that when considering mid- to high-latitude zonal SST averages, the values derived from records polewards of 45∘ latitude are more likely robust but acknowledge these may be conservative estimates (with considerably larger warming further to the south). We therefore estimate uncorrected “polar” warming in the Northern Hemisphere to be 2.0 ± 0.4 ∘C and 0.2 ± 0.3 ∘C in the Southern Hemisphere (Table 1). Correcting for drift decreased the northern estimate to 1.5 ± 0.4 ∘C and increased in the south to a mean annual SST to 0.8 ± 0.3  ∘C.

The maximum temperatures of the early LIG were up to 0.9 ± 0.1 ∘C warmer than 1981–2010, regardless of whether the values were corrected for drift or not (Table 1 and Fig. 6). Similar to the mean SSTs of the LIG, there were apparently considerable zonal differences in the uncorrected values: 0.1 ± 0.2 ∘C within 23.5∘ of the Equator, 3.2 ± 0.4 ∘C polewards of 45∘ N, and 1.5 ± 1.1 ∘C polewards of 45∘ S. After correcting for drift, the estimated SST in the north changed to 2.8 ± 0.4 ∘C and in the south it changed to 2.1 ± 1.1 ∘C. The latter estimate from the Southern Hemisphere is ∼ 2 ∘C (relative to 1981–2010), potentially providing an important constraint for future Antarctic ice sheet model simulations for the LIG (Turney et al., 2020a; Golledge et al., 2015). These data support previous work which reported substantial polar temperature amplification during the LIG, particularly in the Northern Hemisphere (described in the literature as “Arctic amplification”) (Overpeck et al., 2006; Mercer, 1978; Mercer and Emiliani, 1970; Thomas et al., 2020; Miller et al., 2010). The global temperature pattern closely follows insolation changes across this period, during which the Earth's greater eccentricity led to reduced radiation over the Equator and more intense high-latitude spring–summer insolation (Fig. 2) (Overpeck et al., 2006; Hoffman et al., 2017). Comparison to Marine Isotope Stage 6 SSTs appears to show the greatest warming in the northeastern Atlantic and southern Atlantic (Fig. 7), suggesting that the Greenlandic and the West Antarctic ice sheets would have been particularly vulnerable to warming in the early interglacial (Clark et al., 2020; Turney et al., 2020a; Dutton et al., 2015; Mercer, 1978) though we cannot resolve the relative timing of mass loss in this analysis (Rohling et al., 2019; Hayes et al., 2014). Recent work suggests the earliest warming took place in the Atlantic (and Indian) Ocean sectors of the Southern Ocean (Chadwick et al., 2020), consistent with our findings. However, our observed polar warming is larger than in some climate model simulations, implying the latter fail to capture one or more key feedbacks (e.g. carbon, sea ice and ice sheet feedbacks) in the climate system (Bakker et al., 2013; Otto-Bliesner et al., 2013; Thomas et al., 2020; Clark et al., 2020; Fogwill et al., 2015).

The LIG is characterized by higher GMSL than present day (+6.6 to +11.4 m) (Grant et al., 2014; Dutton et al., 2015; Turney and Jones, 2010; Rohling et al., 2017, 2019). Here we quantified the contribution of the relatively high temperatures to global sea levels through ocean thermal expansion for warming down to 2000 m water depth (Table 2). We find that through the LIG, the average SST contribution to thermosteric sea level was negligible, approximately 0.05 ± 0.10 m uncorrected for ocean drift and 0.08 ± 0.10 m corrected for drift, consistent with a recent reconstruction of near-modern global ocean heat content and negligible thermosteric sea level rise (Shackleton et al., 2020). But for the early LIG (129–124 ka), using our maximum SST estimate, we obtained high-end estimates of thermal expansion to GMSL of 0.36 ± 0.10 m (uncorrected) and 0.39 ± 0.10 m (drift-corrected). These quantified estimates are comparable to a previously reported value of 0.4 ± 0.3 m (McKay et al., 2011), which used the same methodology as here but a smaller network of SST records. However, we should recognize that the depth of ocean warming is uncertain, and could have extended deeper than 2000 m. If we assume warming penetrated the full ocean depth (down to 3500 m), we obtained a maximum early LIG thermosteric sea level rise of 0.67 ± 0.10 m (uncorrected) and 0.72 ± 0.10 m (drift-corrected) (Table 2). The recently reported early LIG (∼ 129 ka) peak in global ocean heat content reconstructed from isotopic ratios of atmospheric trace gases has determined a maximum thermal expansion of 0.7 ± 0.3 m (Shackleton et al., 2020). To achieve ∼ 0.7 m of thermosteric sea level rise during the early interglacial peak in temperatures, we have to use both our maximum estimate of temperature rise and our maximum estimate of the depth of warming. A recent modelling-proxy estimate proposed a range of 0.08 to 0.51 m for peak LIG warmth centred at 125 ka (Hoffman et al., 2017), which is more consistent with our results. Even though 125 ka is later than the peak in global ocean heat content, this is effectively the same event but represents the age uncertainties in the marine records. Although some uncertainty remains in the amplitude of thermal expansion between these studies, it is clear that the sustained high global sea levels across the LIG and the limited role of warming on thermal expansion implies a greater contribution from ice sheets, mountain glaciers, permafrost, and hydrological change. With the greatest warming relative to Marine Isotope Stage 6 in the Atlantic basin (Fig. 7), our results are consistent with previous studies suggesting substantial mass loss from the Greenland and the West Antarctic ice sheets early in the Last Interglacial (Clark et al., 2020; Turney et al., 2020a; Dutton et al., 2015; Mercer, 1978; Hayes et al., 2014; Rohling et al., 2019).