Introduction

During the last 800 thousand years (kyr), the Earth’s climate oscillated between 100-kyr glaciations and relatively brief interglacial intervals of warmth1. If not for anthropogenic influence, the next climate event after the current interglacial would be the transition to a glacial period2. Any changes observed during the past glacial inception offer potential insights into how the transitions between two very distinct climate states, interglacial and glacial periods, took place. Before the LGI, the cryosphere configuration was analogous to that of today, while the sea level was several meters higher than today3,4 and the global mean temperature was likely 1–1.5 °C higher than pre-industrial levels5. After the LGI, ice sheets started to regrow in the Northern Hemisphere, and would eventually lower the sea level by ~120 m6.

Our understanding of glacial inceptions is far inferior to that of deglaciations (Supplementary Fig. 1), and many gaps remain. Recently, a seminal Earth system model study indicated that the AMOC abruptly weakened during the LGI7. However, past studies of the production of North Atlantic Deep Water (NADW) during the LGI have not achieved a consensus, with some indicating active overturning8,9 and others suggesting diminished NADW10,11,12,13. Evidence of polar and subpolar cooling during the LGI, of which the AMOC weakening is a sufficient but not necessary cause, is observed in a Greenland ice core record14 and in some reconstructed North Atlantic sea surface temperature records13,15 but not others16,17,18.

In the North Atlantic, the ratio of radioisotopes 231Pa and 230Th in bulk sediment is a dynamic tracer sensitive to the AMOC strength changes19,20,21,22. Unlike their radioactive decay parents, 238U and 235U, which are homogeneously distributed in seawater, 231Pa and 230Th are readily scavenged by particles raining down through the water column. Because of a difference in their timescale of removal, a strong AMOC preferentially exports 231Pa out of the North Atlantic23,24 and leaves behind a low 231Pa/230Th (hereafter Pa/Th) in underlying sediments19,20,22,23. In turn, a weakened AMOC leads to a high Pa/Th approaching the production ratio (0.093) in sediments deposited in the deep North Atlantic.

Here, we report Pa/Th in three sediment cores collected from sites in the western and central North Atlantic (Fig. 1). IODP Site U1313 (41° 0.081’ N, 32° 57.421’ W, water depth 3413 m) and V30-099 (43°08.9’ N, 32°26.9’ W, water depth 3594 m) sit on the western flank of the Mid-Atlantic Ridge. Core KNR191-CDH19 was retrieved from the Bermuda Rise (33° 41.443’ N; 57° 34.559’ W, water depth 4541 m). These cores were chosen because they are from locations shown to record circulation dynamics and water mass mixing19,22,25.

Fig. 1: Map of the cores used in this study.
figure 1

The red and purple ribbons are the simplified warm surface currents and cold bottom flows, respectively, with the arrows marking the direction of the flows.

Full size image

Results

The three benthic δ18O records show distinctive features in common that facilitated the target alignment to establish a shared temporal framework (Fig. 2B). The variability of the benthic δ18O record from V30-99 suggests the sediment from this core may have experienced post-depositional mixing, possibly as a result of the core’s relatively low sedimentation rate ( ~ 1.5 cm/kyr) and hence susceptibility to bioturbation (Fig. 2B). However, the V30-99 benthic δ18O–based age model is supported by an independent age model based on the constant deposition of 230Th excess and a fixed focusing factor during the study interval26 (Supplementary Fig. 2). The age models created by the two methods differ by at most 847 years. Additionally, we generated the benthic δ18O record for the entire last glacial cycle in V30-99, further bolstering the identification of Marine Isotope Stages (MIS) 5 d and 5e in this core (Supplementary Fig. 3).

Fig. 2: Benthic foraminifera δ18O and Pa/Th results compared to other last glacial inception abrupt changes.
figure 2

A Atmospheric CO255. B Benthic foraminifera δ18O results from this study. The colored dots are individual measurements. The colored lines average multiple data points at the same depth if they are available. The gray line is LR041. C–E Pa/Th results with the 2σ error bars. Notice the y-axes are upside down. The horizontal dashed lines are the Pa/Th production ratio (0.093). B–E are all from this study, except CDH19 benthic δ18O in (B) is from ref. 21 (F) Sortable silt data (dots) and three-point average (line) from ref. 12 (G) Cd/Ca data (dots) and three-point average (line) from ref. 10 (H) The Atlantic Meridional Overturning Circulation intensity, defined as the maximum overturning streamfunction in the North Atlantic7. I The benthic δ13C record from MD02-2448 in the South Indian Ocean11. The line is the three-point moving average of the raw data (dots). J Southern Ocean ice-rafted debris mass accumulation rate record from APcomp39. In (H–J), the numbers mark the corresponding Proposals 1–3 (see text). The blue shading is the last glacial inception (113–119 ka). The pink shading is the last interglacial (119–129 ka).

Full size image

The benthic δ18O-based age models allow us to compare the three Pa/Th records on a common age scale (Fig. 2C–E). A consistent pattern emerges, indicating that Pa/Th increased at the LGI at all three sites, rising close to, but never quite reaching, the production ratio of 0.093. The elevated Pa/Th contrasts with the relatively low Pa/Th at each location during MIS 5e and after the LGI. In V30-99, Pa/Th increased to the production ratio at ~135 thousand years ago (ka), probably a signal of Heinrich event 11 during the penultimate deglaciation13. During the span of MIS 5e, all three Pa/Th records show a decreasing trend starting from an already low Pa/Th, likely indicating a continued AMOC strengthening after the recovery from Heinrich event 11. The 2σ uncertainty of Pa/Th is relatively high in CDH19 because the sediments have a lower percentage of scavenged 230Th relative to the total 230Th measured. The scatter plots between the preserved opal content and the Pa/Th data show a weakly negative relationship in V30-99 and a weakly positive relationship (R2 = 0.03019 and R2 = 0.1892, respectively) in U1313 and CDH19 (Supplementary Fig. 4). A comparison between the Pa/Th and preserved opal time series in V30-99 shows that the opal content stays around 1.5% during 100-135 ka despite Pa/Th increases during the LGI and H11 ( ~ 135 ka) (Supplementary Fig. 5). In U1313, the opal content shows relatively more variability but generally is around 2%, although the opal content measurements missed the Pa/Th increases (Supplementary Fig. 5). In CDH19, the opal content is 2–6 % and does not seem to covary with Pa/Th (Supplementary Fig. 5). Notably, the increase in opal content after 110 ka is not associated with a concurrent increase in Pa/Th. Our comparison of the opal content and Pa/Th data thus indicates that the opal contents in the three cores are generally low, and the small variations in opal content are unlikely to explain the observed changes in Pa/Th, indicating that the observed Pa/Th increases primarily reflect changes in circulation rather than the preferential scavenging of 231Pa by biogenic opal27.

In addition to opal flux, the basin-wide total particle flux is indicated as another factor to potentially bias the Pa/Th proxy28. We calculated the 230Th-normalized particle flux at our three sites (see “Methods”). During the LGI, the particle flux increased at CDH19, was in an increasing trend but did not reach its maximum at V30-99, and did not increase at U1313 (Supplementary Fig. 6). The scatter plots between particle flux and Pa/Th (Supplementary Fig. 7) show a weak relationship between particle flux and Pa/Th in every case. At CDH19, the only core where the relationship is positive, the R2 value is 0.07. Since these are correlations, they do not require any causality, but they provide a maximum estimate of the influence of one variable on the other. That means that particle flux has at most a 7% influence on the variance of sedimentary Pa/Th in those cases, and possibly less, from the perspective of linear modeling with least squares estimation. We infer that 93% or more of the influence on Pa/Th derives from something other than particle flux, which we interpret to be changing ocean circulation. Although we can’t absolutely rule out the possibility that there are other influences, the sedimentary data are inconsistent with a dominant influence of particle flux.

The timing of the Pa/Th increase is roughly in the middle of the transition from MIS 5e to 5 d. In U1313 and CDH19, the Pa/Th increase occurs after the respective increasing trends in benthic δ18O from MIS 5e to 5 d are underway. In V30-99, the Pa/Th increase is nominally spread over a relatively long time ( ~ 8 kyr, compared to the <1-kyr duration in U1313 and CDH18), again a likely sign of the potential influence of post-depositional sediment mixing.

Discussion

Because Pa/Th can be used as a proxy of the overall AMOC strength, the Pa/Th increases at the LGI at all three study sites indicate an extensive AMOC weakening at the time. Since the LGI Pa/Th increases never reach the production ratio, as have been observed within the Heinrich layers19,20,21, the circulation disruption was probably less extreme than during Heinrich events. Nevertheless, the broad geographic range covered by our core sites suggests that the LGI circulation disruption was likely a basin-wide event. Possibly because of the relatively small amplitude of Pa/Th changes compared to Heinrich events as well as the abruptness of the event, previous Pa/Th reconstructions did not observe this episode20,29,30 (Supplementary Fig. 8). The AMOC weakening observed in our Pa/Th records fits the observed slowdown in deep current speed12 and an increase in the proportion of southern sourced waters10 (Fig. 2F, G). We used the accompanying benthic δ18O of these two studies to update the age models of the two records for consistency. The benthic δ18O records have been aligned to LR04 using BIGMACS via the same procedure as the three cores of this study (Supplementary Figs. 9 and 10).

What could have caused the AMOC disruption during the LGI? Here we summarize three leading proposals, one based on modeling and the other two drawn from sedimentary observations (Fig. 3). Proposal 1 focuses on orbital forcing’s influence on the North Atlantic sea ice7 (Fig. 2H). LOVECLIM1.3, an Earth system model of intermediate complexity, was used to demonstrate that the AMOC abruptly declines after an interglacial when northern hemisphere summer insolation dips below a threshold. As insolation decreases, sea ice starts to expand in the subpolar North Atlantic where deep convection takes place. The sea ice cover insulates the surface ocean against heat loss to the atmosphere. The resulting warming and increased buoyancy suppress deep convection31 and lead to an abrupt weakening of the AMOC until insolation bounces back above the threshold (Fig. 2H).

Fig. 3: Schematic north-south transect of the abrupt Atlantic Meridional Overturning Circulation (AMOC) slowdown during the last glacial inception (LGI) and its potential causes and impact.
figure 3

The numbers mark the corresponding Proposals 1 (North Atlantic sea ice), 2 (Southern Ocean sea ice), and 3 (Antarctic iceberg) of AMOC slowdown during the LGI. As a result, the upper overturning cell weakens, and the lower cell sequesters a greater amount of dissolved inorganic carbon due to increased respired carbon accumulation and/or northward Antarctic Bottom Water expansion, leading to the inception of atmospheric CO2 drawdown after the last interglacial.

Full size image

Proposal 2 involves sea-ice formation in the Southern Ocean11,32. The downward trending insolation at the LGI induces an equatorward shift of the westerlies33, misaligning the westerlies with the Antarctic Circumpolar Current, reducing the Ekman transport, and suppressing the upwelling of relatively warm deep waters34. Less upwelling of warm water, together with cooling in the southern hemisphere high latitudes35, allows sea ice to expand. The resulting brine injection intensifies the Antarctic Bottom Water (AABW) production (observation: ref. 11; Fig. 2I; modeling: ref. 36). As the density contrast between the deep and abyssal overturning cells increases, the expansion of a denser AABW forces the AMOC lower limb to shoal and suppresses its overturning32,37,38.

Lastly, an increase in Antarctic icebergs reaching and melting at the Agulhas region was observed during the LGI (Fig. 2J). Proposal 3 points to Antarctic iceberg melting as a mechanism that would advect positive buoyancy flux into the upper limb of the AMOC and therefore cause its disruption39. In a sense, the mechanism of Proposal 3 is the opposite of the “Agulhas leakage” that injects warm and saline water into the Atlantic40. The icebergs reaching the Agulhas region increase the freshwater input to the Atlantic. An increase in sea-ice extent and cooling in the Southern Ocean, attributed to decreased insolation33 (Proposal 2), are suggested to improve the survival of Antarctic icebergs and facilitate a northward shift in iceberg trajectories39.

Our study cannot definitively determine which of the three proposals is the most likely scenario, but we find tentative clues in the particle flux supporting Proposal 1. Specifically, 230Th-normalized particle flux measurements from site CDH19 show an increase during the LGI concurrent with the Pa/Th rise. This particle flux increase could indicate increased export productivity, ice rafting of detrital materials, dust flux, or underwater density flow. These local or regional changes are more likely to be caused by processes that originated in the Northern Hemisphere, which exists in Proposal 1 only. On the other hand, increases in particle flux are much more muted at sites U1313 and V30-99 during the LGI, and the particle flux increases also exist during other periods with little changes in Pa/Th. Therefore, the interpretation that our findings support Proposal 1 is only speculative, and additional future research on the bipolar and subpolar regions during the LGI is required to shed light on the causes of the AMOC weakening.

A common thread among the three proposals is the involvement of orbital forcing in influencing sea-ice formation and, in turn, causing AMOC disruptions. North Atlantic iceberg or meltwater discharge is not implicated in this episode of abrupt AMOC weakening. Indeed, Northern Hemisphere ice sheets were only starting to expand from their nucleation centers41,42, and little evidence exists for a substantial North Atlantic iceberg or meltwater discharge event at the time8,13,43,44. Other instances of abrupt AMOC declines during Heinrich events19,21 and the last interglacial45,46 do not have obvious connections with orbital forcing, highlighting the LGI as a possibly unique instance of orbitally forced abrupt circulation changes.

Since the observed AMOC weakening occurred after benthic δ18O already started increasing from its last interglacial minimum, our results do not indicate that the AMOC slowdown initiated the LGI. Instead, orbital forcing changes alone appears to have been sufficient to initiate the LGI, as has been suggested in modeling studies7,47. Nevertheless, it is possible that the AMOC weakening accelerated the glacial regrowth by curtailing the northward heat transport and cooling North America for the nascent Laurentide Ice Sheet48,49.

Our data present new benchmarks, but not direct challenges, to the “moisture initiators” hypothesis for explaining early ice sheet growth. The “moisture initiators” mechanism states that the supply of moisture towards high-latitude continents is essential for ice-sheet accretion16,17,18. The moisture could then induce snowfall and initiate glaciation during the LGI50,51. Because the upper limb of the AMOC transports warm surface water northward, a vigorous AMOC is argued to supply moisture towards the nucleation sites for the Laurentide Ice Sheet. Other studies emphasize the atmospheric route of moisture transport, which could have been enhanced due to an increased equator-to-pole surface temperature gradient52,53,54. Our observations suggest that the AMOC, the oceanic route of moisture transport, remained strong initially and then weakened within the period of rapid glacial expansion, if benthic δ18O is used as a tracer of ice volume. An increase in subpolar North Atlantic sea-ice formation, as laid out in Proposal 1, could have further suppressed the oceanic moisture supply. Yet, the circulation slowdown did not seem to interrupt the ice volume growth. Our results thus favor the atmospheric route over the oceanic one of moisture transport as a viable explanation for enhancing ice sheet growth under declining insolation.

During the LGI, atmospheric CO2 remained persistently elevated for about four thousand years, even after Antarctic temperature cooled5,55 and global ice volume began to increase1, potentially as a result of declining obliquity56. The first significant drawdown of atmospheric CO2 did not occur until the episode of AMOC weakening at 115 ka (Fig. 2A). This might not have been a coincidence. We propose that increased accumulation of respired carbon and/or northward AABW expansion, linked to AMOC weakening as shown by our results, could lead to CO2 drawdown32,37,38 (Fig. 3). The increased dissolved inorganic carbon (DIC) concentration of AABW would further activate the ocean alkalinity feedback via the lysocline shoaling to amplify CO2 sequestration in the deep ocean57. This process, together with the expansion of Antarctic sea ice that could act as a lid to limit the outgassing of carbon from upwelled deep waters in the Southern Ocean58, could explain the delayed timing of the atmospheric CO2 decrease at the LGI. The LGI may thus exemplify the importance of ocean processes in setting atmospheric CO2 changes in motion.

Methods

The sediment samples from V30-99 and U1313 were freeze-dried, soaked in deionized water, and disaggregated on the Cambridge washing wheel disaggregator for an hour. The wet samples were washed through 63 μm sieves with the help of the Lamont automated sample sieving bench, and the coarse fraction retained in the sieves was dried and transferred to glass vials. The dried >63 μm fraction was again dry sieved at >150 μm and examined under a microscope. In core V30-99, the benthic foraminifera Cibicidoides wuellerstorfi tests were picked. In core U1313, the tests of both Cibicidoides wuellerstorfi and Uvigerina spp. were picked. The ẟ18O measurements on the benthic foraminifera tests were conducted with a Thermo Delta V Plus gas-source isotope-ratio mass spectrometer equipped with a Kiel IV individual acid-bath sample preparation device at the Lamont-Doherty Earth Observatory of Columbia University stable isotope laboratory. The ẟ18O records were corrected to Uvigerina using an offset of 0.64‰ for Cibicidoides59. The long-term standard deviation of ẟ18O measurements made on carbonate standard NBS19 is 0.06 ‰. At depths with abundant benthic foraminifera tests, up to four separate stable isotope analyses were carried out. In core CDH19, a benthic ẟ18O record measured on Cibicidoides wuellerstorfi and Nuttallides umbonifera was previously made public21.

The chronostratigraphies of the three cores were established by aligning the benthic ẟ18O records to the LR04 global stack1 using the open-source BIGMACS software (https://github.com/eilion/BIGMACS)60. While the automated BIGMACS alignment procedure performs generally well, because of how short the alignment period is, we also added two alignment data points using the built-in additional age control function. First, a 172 cm sediment depth in V30-99 is given an age of 118 ka. Second, 3322 cm sediment depth in CDH19 is given an age of 115 ka.

Bulk sediment samples of ~100 mg were spiked with 229Th, 236U, and 233Pa, digested, purified61, and analyzed for uranium, thorium, and protactinium isotopic activities. Isotopes were measured on an Element 2 inductively coupled plasma mass spectrometer (ICP-MS) using either Nickle Jet sample cone and X skimmer cone (for Pa) or Standard sample cone and X skimmer cone (for U and Th), and coupled to either a CETACTM Aridus desolvating nebulizer (For Pa) or an ESI-PC3 Peltier cooled cyclonic spray chamber (for U and Th) at the Lamont-Doherty Earth Observatory of Columbia University. 238U and 232Th were measured in analog mode, and the rest of the isotopes were made in ion counting mode. Tail corrections and mass bias corrections were made, and the analog/counting gain was calculated61. Every batch of 18 samples was accompanied by a procedural blank, an internal standard called the North Atlantic Internal Mega Standard (NAIMS), and a 233Pa/231Pa mixture solution to track the decay of the 233Pa spike since its creation. The procedural blanks from the 11 batches contribute, on average, 3% of the 238U measured from samples, 0.4% of 230Th and 232Th, and 0.9% of 231Pa. The repeated measurements of internal standard NAIMS from the 11 batches determined the 1σ precision to be 8.5% for 238U, 3.4% for 230Th, 8.2% for 232Th, and 9.6% for 231Pa.

The radioactive isotopes 235U (half-life: 704 million years) and 238U (half-life: 4.5 billion years) are highly soluble in seawater and have long residence times ( ~ 400 kyr). In contrast, their decay products, 231Pa (half-life: 32.7 kyr) and 230Th (half-life: 75.584 kyr), are highly insoluble and readily scavenged by sinking particles62. The residence time of 231Pa (100–200 years) is shorter than that of 230Th (20–40 years)63 and approaches the Atlantic deep water transit time. As a result, a vigorous AMOC, such as the condition today, exports approximately half of the 231Pa produced in the Atlantic basin towards the Southern Ocean23,24. A weaker AMOC leads to more 231Pa being deposited in the Atlantic sediments, pushing Pa/Th higher to approach its production ratio (0.093).

The measured bulk sediment concentrations of 230Th and 231Pa include contributions of the detrital (produced from the radioactive decay of U in mineral lattices) and authigenic (from the radioactive decay of U that precipitated from the soluble form U(VI) to its insoluble form U(IV) in anoxic, reducing sediments) fractions. In calculating the scavenged portions of 230Th and 231Pa, these other two sources need to be accounted for and corrected. Detrital 230Th and 231Pa can be estimated from the measured concentration of 232Th, which is entirely of detrital origin64. We apply site-specific lithogenic 238U/232Th activity ratios, using 234U/238U to gauge the presence of authigenic U65. After excluding samples with 234U/238U more than 0.96 to account for the loss of 4% of 234U from the detrital sediments by alpha-recoil65, we also excluded one 238U/232Th outlier data point in V30-99 (encircled in Supplementary Fig. 11B). We additionally found samples with abnormally low 234U/238U in U1313 compared to the adjacent samples, all from a single batch of measurements, which we have excluded as well (encircled in Supplementary Fig. 11C). The average of the remaining 238U/232Th data points in each core is used as the local detrital 238U/232Th. The resulting local detrital 238U/232Th estimates are 0.48 at V30-99, 0.57 at U1313, and 0.52 at CDH19. We additionally apply a lithogenic 230Th/238U activity ratio of 0.8166 and a natural 235U/238U activity ratio of 0.04667. We note that Pa/Th is not as affected by the specific choice of lithogenic 238U/232Th activity ratio as other uranium series proxies68,69. Authigenic 230Th and 231Pa are estimated from the non-detrital portion of 238U and 235U by assuming a seawater 234U/238U activity ratio of 1.146870 and correcting for time passed since uranium precipitation.

The 1σ uncertainty of the measured isotopes is estimated with the standard deviation of the 200 scans of the isotopes by the ICP-MS. We detect and remove outliers of the 200 scans using the modified z-score71: \({M}_{i}=\frac{{x}_{i}-\widetilde{x}}{{MAD}}\), where \({M}_{i}\) is the modified z-score, \({x}_{i}\) is the value to be analyzed, \(\widetilde{x}\) is the median, and MAD is the median absolute deviation. Outliers are defined as values with a modified z-score greater than 2. The uncertainty propagation considers the ICP-MS intensity drift during the 200 scans of each sample and standard (National Institute of Standards and Technology Standard Reference Materials Uranium Standard or SRM). The accepted ratio of the SRM 238U/235U is 137.7145, and we apply a relative standard deviation of 0.5% to the ratio in error propagation. We estimate the relative 1σ uncertainty of the lithogenic 238U/232Th and 230Th/238U activity ratios to be 5%, also propagated while calculating the uncertainty. The conversion from raw protactinium counting data to activities and associated error propagation has been packaged into a Python script named PaxsPy, accessible at https://github.com/yz3062/PaxsPy. The counterpart Python script for thorium has previously been published66 at https://github.com/yz3062/ThxsPy.

To test whether the preferential scavenging of 231Pa by biogenic opal influenced our Pa/Th results27, we measured biogenic opal in V30-99, U1313, and CDH19 following two established methods, one utilizing spectrometry72 and the other with inductively coupled plasma optical emission spectroscopy (ICP-OES)73. At the Bermuda Rise site of CDH19, Pa/Th has previously been shown to be little affected by biogenic opal20. For the spectrometry method, bulk sediment samples of ~100 mg were mixed with sodium carbonate and heated at 85 °C for 5 h to extract opal. Silica concentration was measured with a molybdate-blue spectrophotometry method. For the ICP-OES method, a sub-sample of 0.3 ml from the leachate was added to 10 ml of Milli-Q water in 15 mL centrifuge tubes and neutralized with HNO3 to pH = 6. The final volume was then adjusted to 12 mL with additional Milli-Q water. Silicon concentrations were quantified using an Agilent 720 (axial) ICP-OES housed on the LDEO campus of Columbia University. The standard curve used was brought up in a matching Na2CO3 matrix to minimize any matrix effects between the samples and standards. The standard curve encompassed the expected range of the samples. A drift solution, made of a mixture of samples, was run after every 5th sample to monitor and correct for any changes in sensitivity. Silicon was measured on four different wavelengths (185.005 nm, 250.690 nm, 251.611 nm, and 288.158 nm), all of which yielded good signal intensity. The concentration of Silica was calculated based on each individual wavelength, and they were then averaged together for a final value. Percent opal was estimated from Si concentrations using a formula weight conversion factor of 2.4.

To assess whether particle flux could affect our Pa/Th results, we calculated particle flux F = β * Z / 230Thxs,0, where F is the vertical particle flux, β is the 230Th production rate in the water column, Z is the water depth, and 230Thxs,0 is the excess 230Th corrected for radioactive decay (i.e., the denominator of Pa/Th).