ODP Site 959 was drilled during Leg 159 in the Gulf of Guinea, ∼160 km offshore Ghana and Côte D'Ivoire (3.62∘ N, 2.73∘ W; 2090 m depth; Mascle et al., 1996). It is presently located below the eastward-flowing Guinea Current, which originates from the North Equatorial Counter Current and the Canary Current (Norris, 1998a). SST currently varies between 25.3 and 28.6 ∘C on a seasonal timescale (Locarnini et al., 2013), and surface salinity ranges between 34.6 and 35.0 (Zweng et al., 2013). The water column is characterized by a shallow thermocline, with the 20 ∘C isotherm depth annually varying between ∼ 40 and 60 m in response to coastal upwelling (Locarnini et al., 2013; van der Weijst et al., 2022). A minor upwelling event occurs in boreal winter, and a longer and stronger event occurs in boreal summer (Djakouré et al., 2017; Verstraete, 1992; Wiafe and Nyadjro, 2015).

We studied the 0–160 m interval of Hole C and 162–194 m interval of Hole A. The lithology gradually changes from nannofossil/foraminifer chalk at the bottom to nannofossil/foraminifer ooze at the top (Mascle et al., 1996). The age model is based on nannofossil biostratigraphy (Shin et al., 1998) and benthic δ18O stratigraphy (Wagner, 1998) between 0 and 23 meters below seafloor (m b.s.f.), benthic δ18O stratigraphy between 33 and 46 mbsf (van der Weijst et al., 2020), astronomical tuning of high-resolution X-ray fluorescence data (Vallé et al., 2016) between 51 and 100 mbsf, and a spline function through planktic foraminifer and nannofossil biostratigraphy between 162 and 194 mbsf (Norris, 1998b; Shafik et al., 1998). We plot the data between 21.35 and 97.43 mbsf on the revised meters composite depth (rMCD) scale of Vallé et al. (2016). Because outside this interval a splice is unavailable, we constructed a revised meters below sea floor (rMBSF) scale in coherence with Vallé et al. (2016), assuming ∼5 % loss of sediment between consecutive cores (Supplement).

The core was sampled at a relatively low resolution between 0 and 15 Ma for biomarker analysis and at a higher resolution in the late Pliocene (2.7–3.5 Ma) interval for multi-proxy reconstructions. A total of 219 samples were extracted for TEX86, of which 105 are closely spaced in the 2.7–3.5 Ma interval. U37k′ data were generated for 79 of these samples, of which 52 are from the 2.7–3.5 Ma interval. Additionally, we performed palynological assessments on 48 late Pliocene splits of samples that were also used for biomarker analysis.

At Utrecht University, sediment samples were freeze-dried, and the outer surfaces were removed using a clean knife to prevent contamination. The sediment was finely ground using a mortar and pestle. In the process, a split of coarsely ground material was taken for palynology. Biomarkers were extracted at 100 ∘C and 7.6×106 Pa using a dichloromethane (DCM) : methanol (9:1, v/v) solvent mixture in an accelerated solvent extractor (ASE 350, Dionex). The total lipid extract was separated over an activated Al2O3 column into apolar, ketone and polar fractions using hexane : DCM (9:1, v/v), hexane : DCM (1:1, v/v) and DCM : methanol (1:1, v/v) as eluents, respectively. For GDGT analysis, a C46 GDGT (99 ng) standard was added to the polar fraction before it was filtered over a 0.45 µm PTFE filter and dissolved in hexane : isopropanol 99:1 (v/v). Measurements were performed on an ultrahigh-performance liquid chromatograph–mass spectrometer (UHPLC-MS) using the method of Hopmans et al. (2016). Samples with a branched to isoprenoid tetraether (BIT) index of >0.3 could be influenced by terrestrially produced GDGTs (Hopmans et al., 2004; Weijers et al., 2006) and were excluded from further analysis (n=3). All samples were within the reliable range of the methane index (MI; Zhang et al., 2011) and ring index (RI; Zhang et al., 2016), suggesting no appreciable inputs from methanogenic archaea or other non-thermal factors. We will present our results using various calibrations, including the conventional approach of calibrating TEX86 to SST, but also using various models using the relation between TEX86 and various depth integrations.

The ketone fraction was dissolved in hexane and analyzed on a gas chromatograph (GC) coupled to a flame ionization detector (GC-FID, Hewlett Packard 6890 series) equipped with a CP-Sil 5 fused silica capillary column (25m×0.32mm; film thickness 0.12 µm) and a 0.53 mm precolumn. Samples were injected on-column at 70 ∘C with helium as a carrier gas and a flow rate of 2 mL min-1. The oven program was as follows: 70 ∘C for 1 min, then ramped to 130 ∘C at 20 ∘C min-1, then to 320 ∘C at 4 ∘C min-1, and then held isothermal for 10 min. U37k′ values were calculated from the fractional abundances of the C37:2 and C37:3 alkenones following Prahl and Wakeham (1987) and calibrated to SST using the calibration of Müller et al. (1998). We chose not to adopt the Bayspline approach (Tierney and Tingley, 2018) or other tropical calibrations (e.g., Sonzogni et al., 1997, DSR II) as they may amplify noise into signal because of difficulty in accurately determining the very small C37:3 peak areas at the high range of SST (Herbert et al., 2020).

Coarsely crushed freeze-dried samples were spiked with a known amount of Lycopodium clavatum spores to quantify absolute palynomorph (Stockmarr, 1972) and treated with 30 % HCl (2×) and 40 % cold HF (2×) to remove carbonates and silicates. The residue was sieved over nylon mesh, and from the 10–250 µm fraction, the lighter organic fraction was separated from the heavy mineral fraction by suspension. The remaining palynological residue was mounted on glass microscopic slides with glycerine jelly. Samples were analyzed using an optical microscope under 400× magnification and approximately 300 dinoflagellate cysts (dinocysts) were counted per sample. Taxonomy follows that of Williams et al. (2017). Capisocysta was only identified at the genus level, because known Capisocysta species are best identified by the number of antapical plates (Head, 1998), which is complicated by the common dissociation of the hypocystal plates. We calculate the relative abundances of a taxon based on the total dinocyst sum. Dinocysts with affinities for cold-temperate waters in the modern ocean (Boessenkool et al., 2001; Zonneveld et al., 2013) were grouped as “cool species” and used for estimating surface water temperature. This species are Ataxodinium choane, Corrudinium devernaliae, Corrudinium harlandii, Impagidinium pallidum, Nematosphaeropsis labyrinthus and Pyxidinopsis reticulata. We also use the P/(P+G) ratio (Versteegh, 1994) as a paleoproductivity index. The P/(P+G) ratio quantifies the number of cysts produced by heterotrophic dinoflagellates, which are characterized by peridinioid-type tabulation (P) over the total number of dinocysts, including those produced by autotrophic or mixotrophic species with a gonyaulacoid-type tabulation (G).

We explore our TEX86 results using several calibration approaches in Fig. 2, including an SST calibration using the exponential TEX86H (Kim et al., 2010). It should be noted that this calibration yields significant residuals in tropical regions and suffers from regression dilution (Tierney and Tingley, 2014). We also use a linear calibration model using a Bayesian method, a model specific for the tropical Atlantic and models assuming contributions from GDGTs produced below the mixed layer. Below, we first explore our TEX86 record in the context of other GDGT-based indicators and other data using the TEX86H calibration for SST (Figs. 3–7), while subsurface calibrations are discussed in Figs. 7 and 8.

Between 15 and 11 Ma, TEX86H-based SST values fluctuate around an average value of ∼29 ∘C (Fig. 3), similar to the present-day local annual maximum surface temperature of 28.6 ∘C. A long-term gradual cooling starts around 11 Ma and is punctuated by two major cooling steps of ∼2 ∘C each, around 4.9 and 3.4 Ma, in which TEX86H drops well below the present-day local annual minimum surface temperature of 25.3 ∘C. The cooling trend seems to end in the late Pleistocene, where TEX86H records a ∼4 ∘C warming between 0.5 Ma and present, with a core top TEX86H of 24.0 ∘C. U37k′ is at saturation (∼28 ∘C) for most of the interval but registers some colder temperatures in the late Pliocene (27–28 ∘C) and drops below 27 ∘C after 1.8 Ma. The core top sample registers a temperature of 26.6 ∘C, which is within calibration error (1.5 ∘C; Müller et al., 1998) of the 27.5 ∘C modern mean annual SST (Locarnini et al., 2013). The GDGT[2/3] ratio is 7.0 in the core top. In the oldest part of the studied interval, GDGT[2/3] values vary between 4 and 5.5. Following an abrupt increase at ∼13.6 Ma, GDGT[2/3] varies largely between 5 and 8. After 1.7 Ma, GDGT[2/3] persistently increases to >7.

Palynological assemblages are dominated by dinocysts with only minor amounts of terrestrially derived pollen and spores. Dinocyst preservation is generally good. The assemblages are dominated by Brigantedinium spp. (19 %–60 %), closely followed by Spiniferites spp. (8 %–36 %) (Fig. 4; see the Supplement for a complete overview of the assemblages). The interval between 3.33 and 3.16 Ma was sampled at a higher resolution (∼5 kyr on average) and displays strong variability of the negatively correlated Brigantedinium spp. and Spiniferites spp. on timescales <10 kyr. Other relatively abundant and consistently present taxa are Operculodinium spp., Pentapharsodinium dalei and Impagidinium spp. Peak abundances of Lingulodinium spp. were recorded at ∼3.26 Ma (Lingulodinium machaerophorum) and at 2.77 Ma (Lingulodinium hemicystum). Capisocysta sp. is consistently present in the lower part of the interval, with peak abundances (>20 %) at 3.32 and 3.20 Ma, but has its last occurrence at 3.17 Ma. There is general coherence between TEX86H-SST variability and the relative abundance of the “cool species” group, which is dominated by N. labyrinthus. On average, the dinocysts with affinities for colder-temperate waters make up 6 % of the total assemblage, but peak abundances of 10 %–13 % occur during the M2 glacial. The P/(P+G) paleoproductivity index is 0.48 on average, with minimum and maximum values of 0.23 and 0.70, respectively (Fig. 4). P/(P+G) values are primarily driven by the abundance of Brigantedinium spp. The P/(P+G) ratio shows no significant relationship with SST as derived from TEX86H (R2=0.05, p=0.14 in linear regression analysis).

The 15-million-year TEX86H record at Site 959 shows a cooling from >30 ∘C between 11 and 15 Ma to <20 ∘C between 0 and 3.3 Ma (Fig. 3). The U37k′ record, on the other hand, is near or at proxy saturation (>27 ∘C) for most of the interval, with some lower values after 1.8 Ma. Assuming that both proxies reflect SST, the relatively small offset between U37k′ and TEX86H-SST offset in the core top sample (2.7 ∘C) could be explained by a combination of calibration errors and seasonality. However, the average downcore offset between TEX86H and U37k′ is 5.4 ∘C and peaks at 7–8 ∘C during the late Pliocene interglacials (Figs. 5 and 6), which demonstrates that TEX86 underestimates past SST at this site, assuming that U37k′ is a more accurate representation of SST. This could potentially be resolved by using an alternative regional SST calibration. The BAYSPAR-SST calibration (Tierney and Tingley, 2014) accounts for spatial variability of the TEX86-SST relationship and is the only available calibration that raises absolute SST estimates. However, similar to TEX86H (Kim et al., 2010) and low-latitude Atlantic (Zhang et al., 2018) calibrations, it likely overestimates SST variability (Fig. 2), indicating that the calibration slope is too steep. This is potentially a problem inherent to the calibration of TEX86 to SST and could reflect downcore TEX86 sensitivity to SubST instead of SST variability (Ho and Laepple, 2016). In the following sections, we explore multi-proxy data from Site 959 in search of indications for the water depth the downcore TEX86 signal is reflecting.

Several TEX86 calibrations for subsurface temperature (TEX86-SubST) are available, with each assuming, among other things, a different depth integration of the water column (Fig. 8). The BAYSPAR-SubST (Tierney and Tingley, 2014) and TEX86H-SubST (Kim et al., 2012) calibrations both target the upper 200 m of the water column, but the weight of the BAYSPAR-SubST calibration is centered at relatively shallow depths compared to the linearly weighted TEX86H-SubST calibration (Fig. 8). The HL16-SubST (Ho and Laepple, 2016) calibration targets the upper 950 m. The validity of this depth interval was questioned (Tierney et al., 2017), but with peak calibration weight at 100–350 m (Fig. 8), it is compatible with the documented vertical distribution of Thaumarcheotal cell counts and GDGT concentrations (e.g., Hernández-Sánchez et al., 2014; Schouten et al., 2012; Wuchter et al., 2005; Sintes et al., 2016), leaving export efficiency as the major uncertainty. Moreover, based on the large proportion of core tops with higher (>5) GDGT[2/3] values in the global calibration set (Fig. 1), a slightly deeper calibration target (including depths below 200 m) is arguably preferable. This deeper calibration target lowers the slope of the calibration, which dampens reconstructed Miocene–Pliocene temperature variability at Site 959 (Fig. 2).

The ratio between temperature change in the surface and subsurface ocean is 1:1 when averaged across many sites and on longer timescales (Ho and Laepple, 2016). Under the assumption that the (integrated) depth of GDGT sourcing and the structure of the water column was stable, a depth-integrated TEX86 record might still be of value to assess SST variability. Depending on the calibration, the magnitude of Late Miocene to Pleistocene cooling at Site 959 is 10 ∘C (BAYSPAR-SubST), 7 ∘C (TEX86H-SubST) or 5 ∘C (HL16-SubST; Fig. 8). Assuming a 1:1 ratio, the long-term SST trend should be of a similar magnitude. Currently available tropical SST estimates on the studied timescale are either compromised by saturation of the U37k′ proxy or based on TEX86 and are therefore not suitable for independent comparison. However, global mean surface temperature estimates based on benthic δ18O data suggest a ∼5 ∘C cooling across the same interval (Hansen et al., 2013; Tierney et al., 2020). In the absence of major oceanographic changes, it is unlikely that the local cooling trend at Site 959 was larger, because temperature variability is generally lower in the tropics compared to high latitudes where deep-ocean waters derive from (i.e., polar amplification). This suggests that the BAYSPAR-SubST and TEX86H-SubST calibrations may overestimate the magnitude of long-term cooling (Fig. 8). From the currently available calibrations, HL16-SubST best approximates both expected glacial–interglacial variability and multi-million year trends (Figs. 2 and 8) and could potentially be used in combination with other SST records and/or modern water column data to correct for the local SST-SubST offset (Fig. 2). It is possible that the 1:1 ratio between SST and SubST did not persist across the entire record, e.g., due to changes in SACW production. To further improve tropical SST estimates in warmer climates, it should be further explored under which conditions a TEX86 record that is driven by subsurface variability can be used to reconstruct SST variability and how to correct for the surface–subsurface temperature offset.

It is currently unclear what caused M2 glacial inception, but hypothesized mechanisms include declining atmospheric CO2 levels (Berends et al., 2019; Dolan et al., 2015) and reduced latitudinal heat transport by the North Atlantic Current (De Schepper et al., 2009, 2013) in the Northern Hemisphere (NH) and by Indonesian throughflow (De Vleeschouwer et al., 2018) in the Southern Hemisphere (SH). Multi-proxy data from Site 999 in the Caribbean Sea show that declining CO2 levels lagged the benthic δ18O glacial expression at the onset of M2 (de la Vega et al., 2020). This indicates that declining CO2 levels did not initiate the M2 glacial, although they might have affected its intensity and duration.

Because the TEX86 record at Site 959 is affected by SACW, it can potentially be used to study the connection between our low-latitude site and higher latitudes. The TEX86 and benthic δ18O records at Site 959 correspond closely but are out of sync during the M2 glacial (Fig. 6). Both records, generated on the same samples, display a double peak around M2, but TEX86 leads δ18O by ∼5 kyr. Whereas the depth-integrated TEX86 signal at Site 959 is sensitive to SH conditions through SACW, bottom waters at Site 959 consisted predominantly of NADW sourced from the NH during the late Pliocene (van der Weijst et al., 2020). The TEX86 lead therefore indicates a lead of SH over NH cooling/glaciation at the onset of M2, in agreement with transient ice sheet simulation (Berends et al., 2019).

We further explore the relative chronology around M2 in Fig. 9. The age models were aligned with peak glacial δ18O values, and the same records are plotted in the original age models in Fig. S1 in the Supplement. Although it is difficult to discriminate between signal and noise at this resolution, some patterns seem to emerge from this compilation. Directly before M2, benthic δ18O, SubST and SST records from the eastern equatorial Atlantic (Sites 959 and 662) and Caribbean Sea (Site 999) indicate gradual, widespread cooling, while CO2 levels were rising. A sharp drop in TEX86H-SubST indicates intermediate ocean cooling at the onset of M2, leading pronounced glacial cooling (benthic δ18O) of the deep ocean and tropical surface ocean (best observed at Site 662). According to the chronology in Fig. 9, this intermediate ocean cooling occurred during a SH summer insolation minimum, suggesting that astronomical forcing may have amplified Southern Ocean cooling and tipped the system to glacial conditions. The sensitivity of the Antarctic ice sheet to astronomical forcing is also reflected by the 3.3 Ma onset of ∼20 kyr cyclicity in ice-rafted debris (IRD) at East Antarctic IODP Site U1361, with peak IRD mass accumulation rates at SH summer insolation minima (Patterson et al., 2014). Based on the chronology of Fig. 9, SH cooling led atmospheric CO2 drawdown by ∼ 15–20 kyr. Lags of a similar magnitude were not unusual during Pleistocene glacial inceptions (Petit et al., 1999). Marginal Antarctic diatom assemblages and lithology suggest that considerable sea ice expansion occurred during M2, which could have inhibited CO2 outgassing from the deep ocean (Ishino and Suto, 2020; McKay et al., 2012). Moreover, a slight increase in dust-mediated iron fertilization of the Southern Ocean may have facilitated export productivity (Martínez-Garcia et al., 2011). In combination with reduced mixing between Northern Component Water and Southern Component Water (van der Weijst et al., 2020), these processes would have led to increased carbon storage in the deep ocean in response to M2 cooling. Higher-resolution records and correlations would be required to fully test this hypothesis.