Warm Middle Miocene Indian Ocean Bottom Water Temperatures: Comparison of Clumped Isotope and Mg/Ca‐Based Estimates

The middle Miocene is an important analogue for potential future warm climates. However, few independent deep ocean temperature records exist, though these are important for climate model validation and estimates of changes in ice volume. Existing records, all based on the foraminiferal Mg/Ca proxy, suggest that bottom water temperatures were 5–8°C warmer than present. In order to improve confidence in these bottom water temperature reconstructions, we generated a new record using carbonate clumped isotopes (Δ47) and compared our results with Mg/Ca‐based estimates for the Indian Ocean at ODP Site 761. Our results indicate temperatures of 11.0 ± 1.7°C during the middle Miocene Climatic Optimum (MCO, 14.7–17 Ma) and 8.1 ± 1.9°C after the middle Miocene Climate Transition (MCT, 13.0–14.7 Ma), values 6 to 9°C warmer than present. Our record also indicates cooling across the MCT of 2.9 ± 2.5°C (uncertainties 95% confidence level). The Mg/Ca record derived from the same samples indicates temperatures well within uncertainty of Δ47. As the two proxies are affected by different non‐thermal biases, the good agreement provides confidence in these reconstructed temperatures. Our Δ47 temperature record implies a ~0.6‰ seawater δ18O change over the MCT, in good agreement with previously published values from other sites. Our data furthermore confirm overall high seawater δ18O values across the middle Miocene, at face value suggesting ice volumes exceeding present‐day despite the warm bottom water temperatures. This finding suggests previously underappreciated additional influences on seawater δ18O and/or a decoupling of ice volume and ocean temperature.


Introduction
The middle Miocene is a period of strategic importance as a future warm climate analogue: It is the most recent period in geological history with atmospheric CO 2 concentrations relevant to elevated future CO 2 scenarios. The warmer conditions of the middle Miocene Climatic Optimum (MCO;~17-15 Ma) are associated with CO 2 concentrations ranging from~400 to 600 ppmv (Badger et al., 2013;Greenop et al., 2014;Ji et al., 2018;Kürschner et al., 2008;Sosdian et al., 2018;You et al., 2009;Zhang et al., 2013) and global temperatures were significantly warmer than present. Higher temperatures are well documented in terrestrial environments (e.g., Pound et al., 2012), including significantly warmer than modern conditions on the Antarctic continent (Feakins et al., 2012;Lewis et al., 2008;Warny et al., 2009).
Direct geological, faunal, and floral evidence of AIS retreat and expansion during the MCO and MCT, respectively, comes from the ANDRILL project and Ocean Drilling Program Sites 1165 and U1356 (e.g., Fielding et al., 2011;Hauptvogel & Passchier, 2012;Levy et al., 2016;Passchier et al., 2011;Pierce et al., 2017;Sangiorgi et al., 2018;Warny et al., 2009). This evidence, including sedimentological evidence of significant changes in ice extent, suggests globally observed changes in δ 18 O b over the middle Miocene (Cramer et al., 2009;Mudelsee et al., 2014) were in large part driven by ice volume. It is unclear whether the changes in CO 2 were a driver or consequence (or both) of cooling and increased glaciation. Langebroek et al. (2009) found that there is likely a CO 2 threshold that must be met for initiation of AIS expansion during the middle Miocene. The geological evidence recovered at the ANDRILL sites has also led to the intriguing suggestion that Antarctic ice volume could have been larger than today at times during the middle Miocene (Passchier et al., 2011), a result with significant implications for our understanding of ice sheet behavior.
Ideally, independent temperature estimates are combined with δ 18 O b records to deconvolve the contribution of temperature from the δ 18 O of deep ocean water (δ 18 O sw ), in order to estimate changes in global ice volume. For time periods as old as the middle Miocene, there are currently only two proxies independent of δ 18 O which can be used to estimate deep sea temperatures: the Mg/Ca and carbonate clumped isotope (Δ 47 ) paleothermometers. To date, only Mg/Ca-based bottom water temperatures have been published for the middle Miocene. Lear et al. (2000) estimated middle Miocene deep sea temperatures to be~5 to 6°C warmer than modern, taking into consideration multiple sites. Using the revised Mg/Ca temperature calibration of Lear et al. (2002), Billups and Schrag (2003) found slightly warmer temperatures: about 9°C (MCO) and 8°C (post-MCT) in the Indian Ocean, suggesting bottom waters were about 6 to 7°C warmer than modern, and 7.5°C (MCO) and 6°C (post-MCT) in the Southern Ocean, which is about 5 to 7°C warmer than today. These authors also recalculated the Lear et al. (2000) record and derived an approximately 6°C warmer-thanmodern multi-site estimate for the middle Miocene. At the high southern latitude Ocean Drilling Program (ODP) Site 1171 (Tasmanian Gateway), Shevenell et al. (2008) estimated maximum temperatures of 7 to 8°C for the MCO and~6.5°C after the MCT (compared to 0.5°C at present). Those authors also estimated an overall cooling across the MCT of 2 ± 2°C. More recently, applying an updated Mg/Ca calibration taking changes in seawater Mg/Ca into account, Lear et al. (2015) found temperatures at tropical ODP Site 806 (Ontong Java Plateau, Pacific) on average~8°C warmer than present during the MCO and~6°C warmer than present after the MCT. Mg/Ca temperature calibrations have evolved to include more complete information, and the effects of non-thermal factors (e.g., seawater Mg/Ca, carbonate saturation state, diagenesis) are better understood (Hollis et al., 2019), but a complete understanding leading to reliable absolute Mg/Ca temperatures has not yet been achieved. The Δ 47 paleothermometer requires fewer assumptions in comparison Evans, Sagoo, et al., 2018) and thus may be an ideal proxy to assess the accuracy of Mg/Ca bottom water temperatures. Using benthic foraminifera from ODP Site 761, this paper compares new Δ 47 based temperatures with previously published Mg/Ca records from the same site in order to assess the potential of both proxies for reconstructing middle Miocene bottom water temperatures and changes in seawater δ 18 O.
Every temperature proxy has benefits and drawbacks (Hollis et al., 2019). Mg/Ca analyses can achieve relatively high precision with relatively small amounts of sample material, enabling generation of high temporal resolution records. However, the Mg/Ca ratio of foraminiferal calcite is dependent on several factors other than temperature: Past seawater Mg/Ca, salinity, carbonate ion saturation state, and pH have all been shown to have an effect on Mg/Ca in foraminifera (e.g., Elderfield et al., 2006;Evans et al., 2016;Gray et al., 2018;Marchitto et al., 2007). Mg/Ca temperatures also require species-specific calibrations (e.g., Evans & Müller, 2012), at least in part due to differing sensitivities to non-thermal factors (e.g., Evans et al., 2016). Fortunately, the majority of cosmopolitan deep-dwelling benthic foraminifera used for geochemical reconstructions of the middle Miocene are extant today, including all species used in the present study.
Direct evidence for past seawater Mg/Ca is relatively sparse. Changes through the Cenozoic are still debated (Broecker & Yu, 2011;Coggon et al., 2010;Evans, Sagoo, et al., 2018;Horita et al., 2002;Lowenstein et al., 2001), and there are particularly few constraints for the middle Miocene (see, for example, Figure 5 of Lear et al., 2015, and Figure 2 of Evans, Sagoo, et al., 2018). Furthermore, changes in sea level due to changes in ice volume and atmospheric CO 2 concentrations were most likely accompanied by changes in the saturation state of deep waters. Carbonate saturation state affects the biological uptake of Mg into benthic foraminiferal calcite (Elderfield et al., 2006) and may have varied, driven largely by variations in seawater [CO 3 2− ] (Kender et al., 2014). To counter both issues, Lear et al. (2010Lear et al. ( , 2015 used a benthic foraminifer (Oridorsalis umbonatus) which they determined to be less sensitive than other species to changes in seawater Mg/Ca. O. umbonatus is also a shallow infaunal species; infaunal foraminifera may be less susceptible to changes in saturation state in certain settings, since they live primarily within pore waters which may maintain a saturation state closer to equilibrium in carbonate rich sediments (Elderfield et al., 2006).
Preservation is also an important consideration for any foraminifera-based proxy, including Mg/Ca. Postdepositional incorporation of inorganic calcite crystals has been argued to incorporate calcite with higher Mg/Ca than primary foraminiferal calcite, which could bias temperature (Bryan & Marchitto, 2008). Ferromanganese (FeMn) oxyhydroxide coatings may contribute excess Mg and increase apparent Mg/Ca (Nairn et al., 2020), while dissolution may preferentially remove Mg from foraminifer tests and decrease apparent Mg/Ca. Well-established cleaning methods (Barker et al., 2003) aim to minimize the impact of trace metal contamination. Nevertheless, specimens should be examined to avoid those showing signs of dissolution or inorganic overgrowths. Diagenesis in the form of closed-system recrystallization is more complex, but Mg/Ca is generally thought to be relatively robust to early-stage recrystallization (Sexton et al., 2006). Thus, preservation is important, but sample screening will alleviate most diagenesis-related problems. The most serious issues to affect Mg/Ca are therefore those which require independent information, such as changes in seawater chemistry affecting the primary uptake of Mg in foraminifera.
More recently, the carbonate clumped isotope (Δ 47 ) paleothermometer has been applied to foraminifera for paleoclimate reconstructions (e.g., Rodríguez-Sanz et al., 2017;Tripati et al., 2014). This proxy involves measurement of the "clumping" or bonding of the heavy rare isotopes 13 C and 18 O within the same carbonate ion (Eiler, 2007(Eiler, , 2011. Bonds between the heavy rare isotopes are thermodynamically favored at colder temperatures, with no dependence on the fluid isotopic composition in which carbonates form (Eiler, 2011;Schauble et al., 2006).
The abundance of the 13 C-18 O isotopologue is very low; thus, measurement of m/z 47 relative to m/z 44 as in the case of carbonate-derived CO 2 is imprecise. Consequently, large analytical uncertainties are typically associated with single replicate analyses, and multiple replicate analyses are commonly performed. Compared to Mg/Ca, the application of Δ 47 requires considerably larger sample sizes, and therefore when working with foraminifera is much more time consuming.
Similarly to Mg/Ca, preservation is also important for Δ 47 . Recrystallization, precipitation of inorganic calcite, and adherence of carbonate formed in different parts of the water column (such as surface-dwelling coccoliths onto benthic foraminifera) could bias results. Nevertheless, the effect of recrystallization of primary calcite on Δ 47 is thought to be mitigated in benthic foraminifera Stolper et al., 2018) since this process often occurs relatively early in their burial history (Edgar et al., 2013;Voigt et al., 2015Voigt et al., , 2016 and thus at similar temperatures as the primary calcification temperature. To date, it is unknown whether dissolution might affect Δ 47 . Thorough cleaning is also required for the Δ 47 proxy to eliminate other adhered carbonates, and a representative number of samples should be examined for preservation and the effectiveness of the cleaning procedure. The advantages of Δ 47 lie in the fact that it is independent of seawater chemistry. Past seawater pH, dissolved inorganic carbon, and carbonate saturation state have been shown to have no discernible effect outside analytical uncertainty in the range of pH expected in average marine environments (Eagle et al., 2013;Hill et al., 2014;Kelson et al., 2017;Tang et al., 2014;Tripati et al., 2015;Watkins & Hunt, 2015). Moreover, no species-specific effects have been found for foraminifera (Evans, Badger, et al., 2018;Grauel et al., 2013;Meinicke et al., 2020;Peral et al., 2018;Piasecki et al., 2019;Tripati et al., 2010). In fact, nearly all carbonates are within error of the same calibration line Bonifacie et al., 2017;Kele et al., 2015;Kelson et al., 2017;Tripati et al., 2010), with very few exceptions (e.g., some corals and speleothems; Affek et al., 2008;Eagle et al., 2013;Saenger et al., 2012). Thus, the Δ 47 paleothermometer is an ideal proxy to compare with Mg/Ca in order to assess the accuracy of the more established Mg/Ca proxy . Furthermore, similarly to Mg/Ca, Δ 47 is independent of the δ 18 O of the fluid which the carbonate precipitated from (Eiler, 2011;Schauble et al., 2006), and thus, seawater δ 18 O can be estimated using the Δ 47 and δ 18 O b generated from the same measurements. Finally, analytical advances decreasing sample mass requirements (Hu et al., 2014;Meckler et al., 2014;Müller et al., 2017) have made it possible to perform Δ 47 analyses on downcore foraminifera samples, although more material is still required compared to Mg/Ca. In this regard, foraminiferal Mg/Ca analyses require much less material per analysis, and the analyses themselves are much higher precision; thus, the Mg/Ca proxy can be used to generate much higher resolution temperature records.

Setting
Ocean Drilling Program (ODP) Site 761 is located off northwest Australia on the Wombat Plateau (16°44.23′ S, 115°32.10′E) at 2,179 m water depth ( Figure 1). The carbonate-rich sediments of Site 761 have been investigated extensively using the Mg/Ca paleothermometer, making this a suitable location to compare the two proxies while examining the thermal history of the deep Indian Ocean. Wombat Plateau sedimentation is characterized as "mature ocean" pelagic since the early Miocene von Rad et al., 1992). Benthic foraminifera at this site appear relatively well preserved (Lear et al., 2010). Paleobathymetric indicators suggest that no significant depth changes have occurred since the middle Miocene (Holbourn et al., 2004).

Sampling
The samples used for this study are the same as those described in Lear et al. (2010), from the middle Miocene section of ODP Site 761B. This allowed for the most direct comparison possible between Mg/Ca and Δ 47 temperatures, although we could not use the same species of foraminifera (all Oridorsalis umbonatus were previously removed for Mg/Ca analysis). The age model is based on biostratigraphy following Holbourn et al. (2004) with ages updated to GTS 2012 (Hilgen et al., 2012), and additional isotope stratigraphy (supporting information Table S2 of Leutert et al., 2020).
After collecting foraminifera from either a single sample or two combined adjacent samples, the specimens were gently cracked between glass plates to open all chambers. With the exception of a few samples where Cibicidoides spp. or Cibicidoides wuellerstorfi and Cibicidoides mundulus were grouped together due to low abundances, aliquots for clumped isotope measurement were monospecific. Cracked specimens were transferred to 0.5 ml polypropylene microcentrifuge vials and irrigated with deionized water (~13 MΩ). Samples were cleaned via repeated ultrasonication cycles: the material was irrigated, vortexed, the majority of liquid removed by pipette, and the sample was ultrasonicated. Methanol (ACS reagent grade, Merck) was used for the second cleaning cycle; samples were then cleaned for at least three more cycles with deionized water and ultrasonication, until the supernatant was no longer cloudy. Samples were dried at 50°C in a standard drying oven.
After preparation, a small amount of material (from widely spaced but randomly selected core samples) was transferred to carbon adhesive covered stubs, ensuring different examples of each species employed were selected. Stubs were Au/Pd coated and imaged with a Zeiss Supra 55VP scanning electron microscope (SEM) to check for cleanliness and preservation.

Mass Spectrometry
Our analytical approach generally follows previous work (Meckler et al., 2014;Müller et al., 2017). We analyzed Δ 47 on small carbonate samples using two identical instruments, Thermo Fisher Scientific Kiel IV preparation devices coupled to MAT253Plus gas source isotope ratio mass spectrometers. Individual aliquot mass was 85-105 μg (both machines). The long-integration dual-inlet (LIDI) method (Hu et al., 2014) was employed following Müller et al. (2017). Carbonate digestion was performed using~104% phosphoric acid at 70°C. CO 2 was purified by diffusion through silver wool and 1.5 cm of PoraPak-Q (ethylvinylbenzene and divinylbenzene copolymer bead) held at −20°C within a sulfinert coated stainless steel tube. Sample gas was passively expanded through a sulfinert coated, stainless steel capillary and measured for 400 s, with signals typically decreasing from 15 to 10 V on m/z 44. Reference gas, starting at matching initial pressure, was measured subsequently from a separate static microvolume. The PoraPak-Q column was heated to 120°C and exposed to high vacuum for at least 1 hr daily for cleaning. The pressure baseline effect He et al., 2012) was corrected daily using peak scans, employing a flat region of background as close as possible to the peaks selected from scans at five different intensities (5, 10, 15, 20, and 25 V;Meckler et al., 2014). Sample measurements were corrected following Meckler et al. (2014), with some modifications. We used four carbonate standards of differing δ 13 C, δ 18 O, and Δ 47 ; samples and standards were measured in a 1:1 ratio. Three of the standards (ETH1, ETH2, and ETH3) were used to normalize the samples to the absolute reference frame (Dennis et al., 2011), correct for drift in δ 13 C and δ 18 O, and correct scale compression in δ 18 O. The fourth (ETH4) was used as a consistency standard to monitor instrument performance. The number of replicates of each standard was chosen to improve the accuracy and precision of the empirical transfer function (Kocken et al., 2019). The conversion into the absolute reference frame was performed in a single step using ETH1, ETH2, and ETH3 (without initial offset correction), using recalculated standard values from Bernasconi et al. (2018). A block of standard measurements, half on either side of any given sample measurement, was used to convert sample data into the absolute reference frame. The number of standards used ranges from 40 to 80, depending on instrument performance during the relevant correction interval. Correction intervals are discrete blocks of time (ranging from weeks to months), demarcated by major maintenance events such as changing reference gas or source tuning. For the correction intervals of this study, the reproducibility of the standards was on average~0.038‰ (1 SD) for Δ 47 (reproducibility per correction interval per standard is listed in the supporting information, Tables S1 and S2). Average reproducibility of secondary standard ETH4 was 0.037‰ in Δ 47 , 0.03‰ in δ 13 C, and 0.06‰ in δ 18 O (all 1 SD). All data were processed using the software package Easotope (John & Bowen, 2016) employing IUPAC correction parameters (Brand et al., 2010;Daëron et al., 2016). Further details are published elsewhere Meinicke et al., 2020;Piasecki et al., 2019).

Δ 47 Temperature, Seawater δ 18 O, and Uncertainty Calculations
The small carbonate sample method employed in this study reduces the sample mass needed compared to conventional Δ 47 measurements. However, many replicate measurements are still required to obtain a reasonably precise estimate of temperature. Based on our current reproducibility and calibration error, a precision of ≤2°C (68% CI) requires samples to be replicated at least 30 times. Consequently, >3 mg of sample are required for a reasonably well-constrained temperature, making it difficult to obtain a temperature value from each individual downcore sample (~30 cm 3 of sediment) when targeting benthic foraminifera. Site 761 was no exception, particularly since we employed samples which had previously been used for Mg/Ca and stable isotope analyses, and thus, the numbers of benthic foraminifera were already somewhat depleted. Although individual samples were replicated more than once where material availability allowed, speciesspecific replication per sample typically ranged from only one to four measurements (with the exception of two subsamples which generated six and seven species-specific replicates). Furthermore, not all species were abundant enough in every sample to generate any replicate measurements. The solution employed here is to group replicates from different species and adjacent samples together for Δ 47 (Grauel et al., 2013), acknowledging that this may introduce additional variability. We performed this averaging in two ways: using discrete binning as well as a moving average smoothing function. Discrete sample bins were grouped according to the following criteria, in order of importance: (1) avoiding extended temporal gaps; (2) keeping replicates from the same sample together; and (3) grouping over the shortest interval possible. For this dataset, the maximum bin size possible while maintaining the above criteria is~30 (measurements); alternatively, smaller bins were tested for comparison. Age ranges, the span between the age of the first and last aliquot within a bin, are indicated by horizontal bars in our figures. We provide confidence intervals (CI) rather than standard error to improve comparability to other records , including the propagated calibration error following Huntington et al. (2009).
The smoothing function employed here is the probabilistic Gaussian window filter (GWF) described in Rodríguez-Sanz et al. (2017), which gives a larger weight to the values nearest the center of the window following a Gaussian probability distribution. We used a 1 Myr window size and a step of 100 kyr. Following those authors, the 5,000 Monte Carlo simulations were performed using a mean equal to the observed Δ 47 values and an analytical uncertainty equal to the average long-term reproducibility of ETH4. A GWF was then applied to these data, and uncertainties in the final Δ 47 time series were calculated from the percentiles of the 5,000 simulations. Finally, calibration uncertainty was propagated into the GWF uncertainty again as per Huntington et al. (2009).
Since the relationship between Δ 47 and temperature is nonlinear, mean Δ 47 values are determined for each discrete bin or window step prior to calculating temperature. For this study, we employed the Kele et al.
Temperatures calculated using this travertine-based calibration are indistinguishable within uncertainty from benthic and planktic foraminifera calibrations performed in our laboratory (Meinicke et al., 2020;Piasecki et al., 2019) though the travertine calibration is based on a wider temperature range. Δ 47 values are provided with four decimals to avoid rounding errors in subsequent calculations (Tables S1 and S2 and archived raw data, see Acknowledgments). Two replicate analyses, out of a total of 428, were determined to be outliers and removed from the dataset using a 4σ outlier test against the mean of all sample data. No further outlier removal was performed regardless of the method of temperature calculation.
Species-specific δ 18 O offsets must be accounted for before calculating δ 18 O sw , since each temperature is the result of a group of Δ 47 measurements, and each group includes several species of foraminifera. Offsets were adjusted based on Appendix A of Shackleton et al. (1984), correcting all species to be equivalent to C. mundulus. Benthic foraminiferal δ 18 O was then averaged for each group, and δ 18 O sw was calculated using Equation 9 of Marchitto et al. (2014): where BWT is bottom water temperature. Temperatures at 68% and 95% CI were used to estimate δ 18 O sw uncertainty in order to focus on the uncertainty derived from the precision of the Δ 47 temperatures.

Mg/Ca Temperatures
All O. umbonatus Mg/Ca data used to calculate temperatures for Site 761 were originally published in Lear et al. (2010). That work presented two Mg/Ca temperature records by (1) (Hall & Chan, 2004;Lear & Rosenthal, 2006;Marriott et al., 2004). However, these calibrations do not take into account different seawater Mg/Ca. For this reason, the records derived in Lear et al. (2010) were originally interpreted in terms of relative downcore temperature variations and not absolute temperatures. For that record, the relative temperature variations were considered robust because seawater Mg/Ca (Mg/Ca sw ) is thought to have changed relatively little over the middle Miocene period investigated (Evans, Sagoo, et al., 2018;Lear et al., 2015). Subsequently, Lear et al. (2015) presented a new calibration describing the sensitivity of O. umbonatus to both temperature (via an exponential relationship) and Mg/Ca sw (via a power law relationship) (hereafter L2015): The constants in the equation (A, H, and B) were determined by examining a suite of samples from the early Eocene, when it can be reasonably assumed there is little to no water storage in ice and therefore the temperature component could be reasonably independently estimated using benthic foraminiferal δ 18 O (Lear et al., 2015). Here we recalculated Mg/Ca temperatures for Site 761 using the L2015 calibration as it allows incorporating information about Mg/Ca sw yielding absolute temperature estimates. We used the same values as Lear et al. (2015) for A, B, and H (0.66 ± 0.08, 0.114 ± 0.02 and 0.27 ± 0.06, respectively) and applied a 5-point smoothing to the data consistent with the original publication. A disadvantage of the L2015 calibration is that it cannot mathematically take into consideration changes in carbonate saturation state and should therefore only be applied to O. umbonatus Mg/Ca records where porewaters were well buffered (Lear et al., 2015). At Site 761, however, the downcore changes in Li/Ca ratios (Lear et al., 2010) suggest that the O. umbonatus specimens experienced changing Δ[CO 3 2− ], likely due to slow sedimentation rates resulting in porewaters on the Wombat Plateau not being buffered to the same extent as elsewhere (Lear et al., 2015). For this reason, Lear et al. (2015) did not apply their calibration to the Mg/Ca data from Site 761. To circumvent this issue while employing the L2015 calibration, and thus a different seawater Mg/Ca value, we estimated the maximum likely bias caused by changes in porewater Δ[CO 3 2− ]. While core-top data confirm temperature as the dominant control on benthic foraminiferal Mg/Ca, they also estimate the offset caused by carbonate saturation state (Elderfield et al., 2006;Yu & Elderfield, 2008). Benthic foraminifera can be found in undersaturated bottom waters with Δ[CO 3 2− ] as low as −20 μmol kg −1 (Elderfield et al., 2006). Above +25 μmol kg −1 any further change in Δ[CO 3 2− ] has a negligible impact on Mg/Ca (Yu & Elderfield, 2008 ] at this site was generally lower in the Miocene Climatic Optimum than after the middle Miocene Climate Transition, and a marked minima in the sediment % coarse fraction suggests carbonate dissolution during the Miocene Climatic Optimum (Lear et al., 2010). We therefore suggest that the Mg/Ca temperatures calculated with the L2015 calibration might be biased towards cooler temperatures, particularly in the older portion of the record, by up to~1.4°C.
Mg/Ca-based seawater δ 18 O was calculated in the same manner as for Δ 47 , using the δ 18 O b from Lear et al. (2010) of the same samples as the Mg/Ca analyses. The uncertainty of the Mg/Ca temperature was estimated following Lear et al. (2015). Those authors calculated a set of BWT time series, incorporating differences between Mg/Ca calibrations, a ±0.5 mol/mol uncertainty in Mg/Ca sw , and the range of values for B and H derived from the range of Mg/Ca sw assumed for the early Eocene period that their calibration is tied to (analytical uncertainty is very small in comparison, thus negligible). This set of temperature calculations results in a temperature envelope of approximately BWT − 1.5°C and BWT + 2.5°C for the middle Miocene.
Uncertainty in Δ 47 temperature is dominated by counting statistics associated with the analysis, with predictions of uncertainty based on mass spectrometer shot noise coming very close to our average analytical reproducibility (see Huntington et al., 2009, for further discussion). The latest improvements on Δ 47 standardization routines have greatly improved inter-laboratory comparability and yield robust calibrations Meinicke et al., 2020). Thus, the major sources of error between the two proxies used here are very different, with uncertainty in Mg/Ca potentially controlled by systematic biases whereas Δ 47 uncertainty is dominated by random error.

Species-Specific Effects on Δ 47
Grouping Δ 47 values from multiple species together to generate one temperature estimate assumes the absence of species-specific effects. In modern core top samples, no species-specific effects have been found for benthic foraminifera outside the precision of the Δ 47 thermometer (Grauel et al., 2013;Peral et al., 2018;Piasecki et al., 2019;Tripati et al., 2010). Here we compare multiple species downcore, for samples where temperatures are unknown, to assess whether our middle Miocene benthic foraminifera display the same lack of species-specific effect on Δ 47 .
A one-way ANOVA was performed on our Δ 47 data after dividing them into two groups: (1) older than 14.5 Ma and (2) younger than 13.5 Ma. This division is required to group the data into pools large enough to achieve a meaningful statistical power, and has the additional benefit that the shifting temperatures of the MCT cooling event, and less well replicated Δ 47 values in that interval, are avoided. The results show no species-specific offsets (Figure 2), although an important caveat is that some species have not been analyzed in sufficient numbers (n < 10; only results for species with n ≥ 10 are shown in Figure 2). Repeating the ANOVA with all data pooled together, we again observe no species-specific offsets, with the caveat that the data spread per species is slightly larger (not shown). Within current precision, our downcore comparison agrees with core top studies on the absence of species-specific effects on Δ 47 . The lack of species-specific offsets on Δ 47 indicates that pooling calcite from different species into single replicate measurements should not cause biases, which is particularly useful when material is sparse. However, nonlinear mixing, which occurs when stable isotope offsets between different constituents mixed into a single sample are large (Defliese & Lohmann, 2015), should still be considered; this effect is explained further in section 3.3.

Δ 47 Temperature Results: Comparison of Averaging Methods
Figure 3 compares methods of calculating and presenting Δ 47 temperatures using the described discrete binning or GWF (section 2.4) and compares these to the Mg/Ca-based temperature record. The results of the binning and smoothing methods are slightly different, although within uncertainty of each other.
For the GWF, window size and increment must be selected, both with units of time . A window as long as possible will reduce uncertainty, but the trade-off is that climate events can be smoothed out. Since the primary climate event of interest in this interval, the MCT, lasted for around 2 Ma, half this time (1 Myr) was selected. This selection is further justified by the fact that it ensures no fewer than 60 measurements are used per window. The step size of 100 ka was selected on the basis that, for the majority of the record, there are more than 10 new replicates per new temperature "window." An exception is the mid-MCT, where benthic foraminifera were scarce; uncertainty increases proportionally to reflect this scarcity of data (temporal distribution of Δ 47 replicate measurements is shown in upper panels, Figure 3). The initial and final parts of the GWF curve were removed over the first/last 10 replicate measurements to avoid "edge effects." Other smoothing filters, using functions such as a cubed spline or locally weighted regression (LOESS/LOWESS), show similar but more pronounced edge effects; this is because the GWF results do not depend as strongly on data near the temporal end points of the record until much nearer those end points. Furthermore, other smoothing filters require selection of parameters such as degrees of freedom; in the case of this dataset, using an objective method such as generalized cross-validation to calculate degrees of freedom for a LOESS fit results in a nearly flat line across the record (not shown).
For the binning approach, the calculated mean temperatures are sensitive to the number of replicates used (Figures 3a-3c). As expected, the scatter of binned mean temperatures decreases with increasing replicates, and the bin means move closer to the values obtained with the smoothing function. The version we consider most robust is that with~30 measurements per bin (n = 27-34), which yields 14 discreet temperature estimates (Figure 3c and supporting information). Due to the fact that individual temperature data points are each based on a large number of measurements, we calculate uncertainty from the sampling statistics of these data pools themselves (SE = 0.005 to 0.008‰) rather than using the long-term reproducibility of the secondary standard measurements (SE = 0.003 to 0.005‰).
The highest Δ 47 temperature of 13.9 ± 3.8°C (95% CI) is observed at 16.2 Ma. While this apparently high temperature occurred close to the time of maximum CO 2 emissions from the Columbia River Basalt (Kasbohm & Schoene, 2018), its mean value falls outside the 68% CI uncertainty of the highest temperature calculated using the GWF, 11.7 ± 2.3°C (95% CI; Figures 3c and 3f; Table S2). Considering the uncertainty associated with our Δ 47 temperatures, it would be prudent to attempt to replicate this temperature separately before this value is used as a constraint for past bottom water temperatures. The rest of the Δ 47 temperatures calculated with the largest discrete bin size do not differ greatly from those calculated using the GWF; therefore, we focus on the GWF-based temperature record when comparing with the Mg/Ca-based temperatures in the following sections.

10.1029/2020PA003927
Paleoceanography and Paleoclimatology inflection points on the composite dataset to define the intervals. The MCT is defined as the region between 13.0 and 14.7 Ma; all data older than this interval are considered to fall into the MCO, and all data younger than this interval are classified as post-MCT. These ages bracket the major MCT ice sheet advances between 14.7 and 13.8 Ma . Bracketing the MCT in this manner also allows us to consider climate mean states before and after the MCT.
Our GWF-calculated Δ 47 temperature record indicates a range of bottom water temperatures over the MCO from 10.0 ± 3.6°C to 11.7 ± 2.3°C, with an overall average of 11.0°C (Figures 3 and 4 and supporting information). Averaging all individual Δ 47 measurements across the entire MCO gives a value of 11.0 ± 1.7°C (Table 1 and Figure 3f), identical to the average GWF result. In the post-MCT interval, the GWF-calculated Δ 47 temperatures range between 7.3 ± 2.9°C at 12.4 Ma to 9.3 ± 3.1°C at 12.9 Ma, with an overall average of 8.2 ± 2.8°C. The GWF-based average again compares very well to the discrete bin average of all post-MCT measurements of 8.1 ± 1.9°C. These values suggest the overall cooling over the transition was 2.9 ± 2.5°C. A two-sample t test shows that the long-term discrete bin average Δ 47 values of the post-MCT and MCO periods are significantly different (p < 0.005), lending confidence to this estimate.
The Δ 47 temperatures agree very well with the Mg/Ca temperatures, with the L2015 based record and the version of this record including the potential Δ[CO 3 2− ] offset both being within the 68% CI band over the majority of the record (Figures 3e and 4a). This result is unexpected considering the uncertainty relating to the important seawater chemistry parameters necessary to calculate absolute temperature from foraminiferal Mg/Ca such as seawater Mg/Ca and the uncertainties associated with the Mg/Ca sensitivity to Δ[CO 3 2− ] discussed above (section 2.5). Though still within uncertainty, the Mg/Ca-based ΔT across the MCT is low relative to the mean Δ 47 -based ΔT for the same interval. As stated in section 2.5, it is likely that the Δ[CO 3 2− ] effect on Mg/Ca was higher in the MCO, suggesting the change in BWT of 0.8°C is an underestimate. The fact that the offset between the proxies is most pronounced during the MCO, a time when Δ[CO 3 2− ] is expected to have been most impacted by relatively high atmospheric CO 2 , indicates that changes in carbonate saturation state are the most likely explanation of the offset. The direction of the offset suggests that the Mg/Ca temperatures could be depressed relative to the rest of the record during that time.
The remaining question to explore is whether these relatively warm temperatures, as estimated by both proxies, are realistic. They seem hard to reconcile with evidence for significant ice volume on the Antarctic continent and atmospheric CO 2 concentrations between modern and preindustrial levels (Sosdian et al., 2018;Super et al., 2018) after the MCT. These temperatures also lead to the calculation of

10.1029/2020PA003927
Paleoceanography and Paleoclimatology heavy seawater δ 18 O (discussed further in section 3.4). For Δ 47 , biases may be caused by thermal and nonthermal factors, including incorporation of calcite formed at the sea surface (e.g., coccoliths) and recrystallization of primary calcite at warmer temperatures. Recrystallization of primary calcite to highly offset stable isotope compositions, even if it occurred at the same temperature as primary calcification, could also cause bias due to nonlinear mixing effects.
While a few nannofossil remnants are visible in some SEM images, a conservative estimate is that they make up far less than 1% of the material analyzed, and they should therefore not impart a measurable bias. However, most specimens seem to display some degree of recrystallization (Figures 5 and 6); thus, the effects of this process must be considered. The main observations from our SEM images are (1) exposed surfaces of  test walls tend to exhibit blocky, euhedral inorganic calcite, varying from a few crystals across an entire test to large areas of the surfaces being recrystallized, and (2) cross sections through freshly broken test walls exhibit what appear to be primary textures alongside textures associated with inorganic crystallization in highly varying proportions. These observations indicate that the primary calcite is mixed with recrystallized calcite to varying degrees in our samples. However, benthic foraminifera are thought to undergo most recrystallization relatively early during burial (Edgar et al., 2013;Leutert et al., 2019;Voigt et al., 2015Voigt et al., , 2016. Even if some recrystallization occurred at a later stage during burial, our samples were relatively shallowly buried (~35 to 52 mbsf), leading to negligible geothermal heating at burial depth. The geothermal gradient effect would furthermore have been counteracted by deep-water cooling over the period since deposition. Consequently, it is unlikely that recrystallization occurred under a significantly different temperature regime. If recrystallization did occur much later than deposition, due to the shallow burial depth, the Δ 47 compositions would most likely be biased towards lower temperatures, but this cannot explain the offset to slightly warmer temperatures compared to the Mg/Ca values.
Nonlinear mixing effects can cause a warm or cold bias beyond analytical precision if both the δ 13 C and δ 18 O compositions of the recrystallized material are more than 2‰ offset from the primary calcite (the required difference increases to~15‰ if only one of δ 13 C or δ 18 O is offset; Defliese & Lohmann, 2015). We consider such different composition of the diagenetic calcite phase as highly unlikely. The recrystallized portion of material could potentially have much different δ 13 C or δ 18 O if calcification occurred in a fluid influenced by dissolved inorganic carbon (DIC) or other sediment components with very different isotopic compositions. However, Site 761 sediments are carbonate-rich and low in total organic carbon (TOC) in the studied interval (~81% to 92% CaCO 3 , ≤0.02% TOC; Shipboard Scientific Party, 1990), indicating no significant influence on pore fluid DIC from organic material (normally characterized by isotopically light carbon). Similarly carbonate-rich sites show minimal isotopic offsets in benthic foraminifera regardless of the amount of recrystallization (Voigt et al., 2016). Furthermore, the δ 13 C and δ 18 O values of C. mundulus and C. wuellerstorfi from Site 761 are very similar in absolute values, trends, and magnitude of change to records produced from similar benthic species at a wide array of pelagic sites (including, but not limited to, Eastern Pacific Sites 1236 and 1237, Holbourn et al., 2013, and Sites U1335, U1336, U1337, and U1338, Kochhann et al., 2016; Southern Ocean/Tasmanian Gateway Site 1171, Shevenell & Kennett, 2004;and Southwest Pacific Site 588A, Flower & Kennett, 1995). If post-depositional diagenetic overprinting due to the local carbon and oxygen pools had caused offsets in the stable isotope composition of the recrystallized material, it is unlikely that our stable isotope data would agree so well with such a far-reaching array of co-eval benthic foraminifera records. Our findings are consistent with the recent study of Leutert et al. (2019), which determined that benthic foraminiferal Δ 47 compositions are largely robust to diagenesis in typical deep-sea settings.
For Δ 47 compositions, a final mechanism to consider is the effect of dissolution; unfortunately, its effects have not yet been tested. Although we observe some dissolution of foraminiferal calcite in the SEM images ( Figure 6), this does not appear to be particularly extensive, nor does it seem to occur to different extents in different time intervals. Changes in the percent coarse fraction and changes in sedimentation rate, two variables which could be related to dissolution, do not appear related to the Δ 47 temperatures ( Figure S2). Thus, while a dissolution effect cannot be ruled out, dissolution itself appears to have occurred to a minor degree in our samples.
The primary potential causes for biases in Mg/Ca temperatures include incorrect assumptions regarding the species-specific temperature calibration or seawater Mg/Ca values used, changes in carbonate saturation state, and diagenetic addition of Mg. Salinity and pH effects are not discussed here as these have not been shown to have an effect on benthic foraminifera. Middle Miocene seawater Mg/Ca composition is fairly uncertain (see Figure 2  Compared to a coeval section at Pacific Ocean ODP Site 806, Lear et al. (2015) noted that Site 761 samples had elevated Mn/Ca and higher Mg/Ca ratios. Those authors concluded that the observed Mg/Ca offset was caused either by porewater carbonate saturation effects or diagenetic contamination, the latter implying a warm bias in the Mg/Ca temperatures at ODP Site 761. This is important because, as indicated above, Site 761 exhibits unusual porewater chemistry, with possible consequences for Mg/Ca in the shallow infaunal O. umbonatus. The [Mg] porewater profile at Site 761 is monotonous over the upper 300 m of sediment, rather than decreasing as observed at other pelagic sites (De Carlo, 1992). It is thought that this unusual profile is due to the shallow burial and low level of lithification allowing seawater circulation through the pore space to a higher degree than at other pelagic marine sites (De Carlo, 1992), leading to a lack of effective buffering capacity. Though the two proxies used here produce temperatures generally within uncertainty, the longterm trend of the Mg/Ca temperatures follows a much shallower slope (Figures 3 and 4), also illustrated by the smaller ΔT between the MCO and post-MCT periods compared to the Δ 47 -based temperatures ( Table 1). The difference in ΔT, and apparently larger difference between Δ 47 and Mg/Ca based temperatures in the MCO, implies that carbonate saturation effects may bias the Mg/Ca record towards colder temperatures in this interval, supporting the interpretation of Lear et al. (2010). While the uncertainty of Δ 47 does not allow us to conclude with confidence, the Δ[CO 3 2− ] corrected Mg/Ca temperatures are closer to the average Δ 47 temperatures in the MCO. Changes in carbonate saturation state are not thought to affect carbonate Δ 47 compositions to a measurable extent (Eagle et al., 2013;Hill et al., 2014;Kelson et al., 2017;Tang et al., 2014;Tripati et al., 2015).
When comparing the absolute temperatures, the effect of diagenesis on benthic foraminiferal Mg/Ca should also be considered. If the inorganic Mg partition coefficient is higher than the foraminiferal Mg partition coefficient, recrystallisation of primary calcite or addition of inorganic calcite could be expected to result in a bias to warmer temperatures. Though this argument has been invoked previously (e.g., Bryan & Marchitto, 2008), electron microprobe analyses have recently shown that the difference between the Mg partition coefficients for benthic foraminiferal calcite and inorganic calcite is likely orders of magnitude smaller than previously thought, at least for the natural range of ocean bottom water temperatures (Lammers & Mitnick, 2019). Nevertheless, it is thought that biological control during calcification limits Mg 2+ incorporation into foraminiferal calcite (e.g., Bentov & Erez, 2006). High-resolution secondary ion mass spectrometry results from Kozdon et al. (2013) seem to support a middle ground between these arguments, showing an approximate doubling of the Mg/Ca ratio in inorganic overgrowths on planktic foraminifera compared to primary material. While even a doubling of Mg/Ca in any inorganic calcite overgrowths would undoubtedly impact the calculated temperatures, we do not observe significant inorganic calcite overgrowths on our specimens of the type investigated by Kozdon et al. (2013) (Figures 5 and 6). If our samples were affected by significant proportions of calcite overgrowths, we would also expect the Mg/Ca and Δ 47 temperatures to diverge more, since the effect on each proxy would be opposite (warmer apparent temperatures in the case of Mg/Ca, the same or cooler in the case of Δ 47 ).
While we have discussed potential biases at length, it is important to highlight how well the two proxy temperature records agree overall. The Mg/Ca and Δ 47 thermometers suffer from different potential biases, often working in opposite directions. This is particularly important considering the relatively poor constraints on Miocene seawater Mg/Ca and carbonate saturation state changes and their effect on Mg/Ca temperatures, as well as any as-yet unknown effects due to dissolution or unaccounted for diagenetic effects which could bias the Δ 47 temperature record. Consequently, the similarity between the Mg/Ca and Δ 47 temperature estimates at ODP Site 761 lends credibility to the finding that deep ocean temperatures here were significantly warmer than today across the entire interval from 16.5-11.5 Ma. The Δ 47 -based temperatures of 11°C during the MCO and 8°C post-MCT are about 9°C and 6°C warmer than present, respectively, for bottom waters at Site 761. Mg/Ca-based temperatures, whether corrected for the carbonate saturation state effect or not, agree very well, particularly in the younger part of our record. While high CO 2 concentrations are invoked for the relative warmth of the MCO, the post-MCT period is characterized by atmospheric CO 2 concentrations on average somewhere between pre-industrial and modern (between 280 and 400 ppm; Sosdian et al., 2018;Super et al., 2018), suggesting a significantly different climate state for similar greenhouse forcing.

Seawater δ 18 O
The GWF Δ 47 -based δ 18 O sw record shows relatively heavy values, averaging approximately 0.5‰ during the MCO and approximately 1‰ after the MCT. Before and during the MCT, the Δ 47 -based δ 18 O sw values are slightly higher than those derived using Mg/Ca temperatures (Figure 4d), a result of the combination of slightly warmer mean temperatures, particularly when not correcting for the carbonate saturation state effect, and essentially identical δ 18 O b (Figure 4c). Mg/Ca-based δ 18 O sw was calculated using the δ 18 O b from the same samples as presented in Lear et al. (2010), whereas Δ 47 -based δ 18 O sw is calculated using our new δ 18 O b measurements. As our new stable isotope compositions do not show any offsets compared to two independently generated records (Figures 4b and 4c; Holbourn et al., 2004;Lear et al., 2010), the two δ 18 O sw records are comparable.
To estimate the change in δ 18 O sw across the MCT, we use the δ 18 O sw values from GWF windows centered around time slices at the onset and end of the MCT (13.0 and 14.7 Ma; values marked with red arrows in Figure 4d) rather than the long-term average values discussed above for temperature. This approach allows us to avoid the less-well replicated section in the center of the MCT and be more comparable to other records focusing solely on the transition. Δ 47 -based δ 18 O sw values at the beginning and end of the MCT (0.54 ± 0.59‰ at 14.7 Ma and 1.18 ± 0.65‰ at 13.0 Ma) suggest an increase of 0.64 ± 0.88‰ across the transition. These estimates are only very weakly dependent on the poorly replicated region between 13.4 and 13.9 Ma, and the Δ 47 -based δ 18 O sw record is fairly consistent across the early and late parts of the MCT interval, indicating this value of relative change over the climate transition is robust. The Mg/Ca-based change in δ 18 O sw across the MCT of about 0.8‰ is within uncertainty of that derived from Δ 47 , although this is likely overestimated considering the Li/Ca evidence for changes in saturation state. This estimate does not take into consideration that a Δ[CO 3 2− ] bias correction is likely appropriate in the older part of the Mg/Ca record, particularly the MCO, whereas it may be unnecessary to make this correction in the youngest part of the record. Lear et al. (2010) suggested that a change in δ 18 O sw of 0.6‰ was most accurate based on the combined Mg/Ca and Li/Ca data, which is nearly identical to our Δ 47 -based seawater δ 18 O estimate.
Taken at face value, the Δ 47 δ 18 O sw record suggests~73% of the change in δ 18 O b at Site 761 is due to increased ice volume. This estimate is consistent with Mg/Ca-based estimates at other sites (e.g.,~70% at ODP Site 1171, Shevenell et al., 2008; 60 ± 14% from a global meta-analysis by Mudelsee et al., 2014). It is also consistent with the fact that the δ 18 O b record of Site 761 closely follows the global sea level curve between~15.1 and 11.5 Ma (Holbourn et al., 2004), which implies that the deep water at ODP Site 761 tracked global climate signals.
Compared to modern δ 18 O sw , both Δ 47 and Mg/Ca-based δ 18 O sw estimates suggest that δ 18 O sw values were relatively high across the record and significantly higher than present after the MCT, reaching similar values to those observed for the last glacial maximum (LGM). This observation, if interpreted in terms of ice volume, is difficult to reconcile with the significantly warmer than modern bottom water temperatures suggested by both proxies, as well as our current theories of global mean sea level for the middle Miocene. Our assessment of potential nonthermal biases of the proxies does not support a diagenetic control on temperature and thus δ 18 O sw , although we acknowledge that there may be other unknown biases that we have not accounted for.
If these high δ 18 O sw values are accurate, they may imply either a larger than modern global ice volume after the MCT, a strongly different isotopic composition of the ice sheet during this time period, and/or that high salinity (and correspondingly high δ 18 O sw ) bottom waters were more persistent in the deep oceans, as suggested for tropical Pacific ODP Site 806 (Lear et al., 2015). Our heavy values could also reflect the effects of different seawater pH on δ 18 O b (Zeebe, 1999) or possibly different species-specific δ 18 O fractionation during calcification (vital effects) compared to today. The latter effects are difficult to test quantitatively. Although distinct relationships between calcite δ 18 O and pH have been found for planktic species (Spero et al., 1997;Uchikawa & Zeebe, 2010), Marchitto et al. (2014) did not find a clear relationship between pH and the calcite δ 18 O of the epifaunal benthic species that make up a large part of our record. It is also impossible to directly test living specimens to examine whether vital effects were different during the middle Miocene compared to today.
Storage of water on the poles, in ice or otherwise, could explain a large part of the difference between our expectations and our results if conditions were different to today. Evidence is insufficient to support significant ice in the Northern Hemisphere during the middle Miocene (Larsen et al., 1994); thus, storage within ice sheets was likely primarily restricted to Antarctica. It is possible that the topography of Antarctica was different in the middle Miocene, such that the area available for ice growth could have been larger (e.g., Gasson et al., 2016), and there is geomorphological evidence for ice sheets which were significantly thicker than today (Denton et al., 1984;Denton & Sugden, 2005). Ice grounding lines extending significantly further than modern also suggest larger than modern ice sheet extent (e.g., Passchier et al., 2011). Recently, it has also been suggested that during deeper geological time periods such as the Cenomanian and Turonian hothouse periods, groundwater storage must be invoked to account for sea level cycles in an ice-free world (Sames et al., 2020).
In addition to a different volume of water stored on land, the isotopic difference between the deep ocean and terrestrial water storage reservoirs could have been greater than today. Thicker ice sheets covering a smaller area, positioned far from the source of evaporation, could at least partly address the issue, since both longer distances and higher elevations would drive the necessary fractionation. However, it is highly improbable that this latter possibility is the sole explanation for the heavy δ 18 O sw ; assuming the same sea level and ice volume as today, an average "stored water" δ 18 O of −70‰ would only explain a 0.3‰ difference in seawater Finally, our heavy δ 18 O sw results could be explained by a different source of deep waters, specifically a process which produced more saline and correspondingly heavier δ 18 O deep waters. Evaporation-dominated deep water formation would provide the necessary mechanism, though there is little evidence available for where and how this mechanism might have operated for the middle Miocene. Relatively heavy δ 18 O b values from a suite of ODP Sites have previously been explained as resulting from evaporation-controlled saline deep water derived from the Tethys Ocean, although the water mass responsible is thought to have ceased by~14.5 Ma (Ramsay et al., 1998). Woodruff and Savin (1989) envisioned a scenario in which this water mass could have been refrigerated as it moved southwards and was re-entrained in high latitude deep water formation; this might explain the temperatures we have reconstructed being inconsistent with tropical or subtropical latitude deep water formation.
The discussion above indicates a middle Miocene δ 18 O sw paradox, suggesting that deep water δ 18 O sw values cannot be explained with our current understanding of the primary controls on seawater δ 18 O. The very good agreement between the Mg/Ca-and Δ 47 -derived temperatures at Indian Ocean Site 761 provides confidence that these high seawater δ 18 O values are accurate. Benthic records from western equatorial Pacific ODP Site 806 and Southern Ocean ODP Site 1171 have independently produced similarly high bottom water δ 18 O sw values for the post-MCT period (Lear et al., 2015;Shevenell et al., 2008), consistent with our assessment. These records are based on benthic foraminiferal Mg/Ca, but use different species (epifaunal C. mundulus in the case of Site 1171) or are based on the same species (O. umbonatus) but from porewater which was likely buffered against changes in the carbonate saturation state (Pacific ODP Site 806). The consistency of these findings taken together suggests gaps in our understanding of the warmer middle Miocene climate system and should be explored further in future work. In particular, oxygen isotope-enabled modeling could greatly enable more insight into this apparent paradox.

Conclusions
We applied clumped isotope thermometry to benthic foraminifera across the middle Miocene climate transition. Reconstructed temperatures are based on a large number of measurements on multiple species from multiple downcore samples. In agreement with previous core top studies, our downcore Δ 47 results show no indication of species-specific effects on benthic foraminiferal Δ 47 compositions. A moving average smoothing function such as the probabilistic Gaussian Window Filter (GWF) first proposed for Δ 47 data by Rodríguez-Sanz et al. (2017) yields a robust temperature record for Site ODP 761.
The long-term discrete average Δ 47 temperatures calculated for the MCO (11.0 ± 1.7°C) and post-MCT (8.1 ± 1.9°C) are statistically distinguishable and suggest a cooling across the MCT of approximately 2.9 ± 2.5°C.
Our bottom water Δ 47 temperature record agrees well with temperatures based on previously published Mg/Ca ratios, whether these are corrected for an offset caused by the carbonate saturation state effect on benthic foraminiferal Mg/Ca or not. Larger apparent offsets between the two proxies during the earlier (MCO) part of the record confirm previously suggested saturation state effects on Mg/Ca in that interval. Although we cannot completely rule out diagenesis-related biases in Δ 47 or Mg/Ca, the most likely effects of post-depositional recrystallisation would work in opposite directions for the two proxies and are not supported by the slight positive offset of Δ 47 compared to the Mg/Ca temperatures in parts of the record. Considering that seawater Mg/Ca and changes in saturation state are poorly known quantities which may affect Mg/Ca based temperatures, while the effects of dissolution on Δ 47 are currently unconstrained, the agreement between the two proxies is remarkably good.
The Δ 47 thermometer estimates a change in δ 18 O sw of~0.6‰ across the MCT, in excellent agreement with the Mg/Ca paleothermometer (Lear et al., 2010). The Δ 47 -based estimate is also consistent with other estimates of δ 18 O sw change over this interval. As previously observed, δ 18 O sw values are relatively heavy throughout the record, resulting in post-MCT values similar to those reconstructed for the last glacial maximum. For the reconstructed values to be realistic, we discuss several options, which in part appear to be contradictory to the warm bottom water temperatures we observe for the middle Miocene. This suggests that the middle Miocene relationship between δ 18 O sw and ice volume was different from the present or more recent glacial periods and requires further research.
While currently the Δ 47 proxy requires substantial amounts of sample material (2-3 mg), it is not affected by changes in seawater chemistry and species-specific offsets. Mg/Ca is capable of providing temperature information at significantly higher temporal resolution due to the smaller sample mass requirements, but can be biased by a range of factors which become more poorly constrained further back in time. Thus, these two proxies are highly complementary and lend themselves well to a multiproxy approach. In this case, the comparison of Δ 47and Mg/Ca-based records has provided more confidence in the warm bottom water temperatures estimated at ODP Site 761 during the middle Miocene.

Data Availability Statement
All data related to this study can be found in the supporting information (standard results and final calculated temperatures) as well as in two online repositories. Raw data are available in the EarthChem Database (https://doi.org/10.26022/IEDA/111523); final calculated temperatures are available at Pangaea (https://doi.pangaea.de/10.1594/PANGAEA.914305).