Main

The Atlantic Meridional Overturning Circulation (AMOC) consists of the sinking and southward flow of North Atlantic Deep Water balanced by northward flow of warmer and lighter intermediate water from the Southern Ocean (SO). Upwelling and buoyancy input in the SO converts deep water to intermediate water to close the overturning cell1,2. A second deep overturning cell in the Atlantic consists of northward flow of Antarctic Bottom Water (AABW) balanced by the southward flow of North Atlantic Deep Water. The Atlantic and SO components of the global ocean overturning circulation are therefore intimately linked3.

Southern hemisphere influences hypothesized to affect AMOC strength include the latitude and intensity of the westerly winds over the SO4,5, stratification at the Atlantic–SO boundary and hence meridional density gradients in the Atlantic5,6,7,8, the propagation of Antarctic Intermediate Water (AAIW) salinity anomalies from south to north in the upper branch of the AMOC9,10,11 and the properties and transport of AABW in the deep overturning cell12,13. However, data to test these proposed mechanisms are limited by the short instrumental record and incomplete proxy records, particularly for the SO.

To help fill this gap, we generated a millennium-long coral-based proxy of AAIW to examine the link between SO processes and the strength of the AMOC. Specifically, we radiocarbon dated and analysed geochemical records in the calcite skeletons of 15 colonies of deep-sea bamboo corals (Keratoisididae) in the southwest (SW) Pacific (Supplementary Fig. 1), ranging in age from ~1,300 years BP to modern (that is, collected alive) and collected close to the core of AAIW (~950–1,250 m) in this region14 (Supplementary Fig. 2). Most specimens were collected from seamounts southeast of Australia (see ref. 15 for details of the sites and sampling methods). Temperature, salinity and pH measured close to the seamount surfaces from which the corals were collected match those measured tens of kilometres away16, indicating that environmental conditions experienced by the corals and recorded in their skeletons are representative of water masses over a broader region. To assess the generality of the record, we also examined material collected at AAIW depths off New Zealand (the Wanganella Banks and Chatham Rise) and New Caledonia (the Norfolk Ridge) (Supplementary Fig. 1). Analyses focused initially and primarily on Mg/Ca ratios, which have been widely demonstrated to be temperature-dependent/correlated in abiogenic and biogenic calcites17, including bamboo corals16,18,19,20 and related gorgonians21.

Covariation of AAIW and the AMOC

At the well-sampled seamounts southeast of Australia, Mg/Ca ratios from nine corals are coherent when aligned on the basis of dating (Methods; Fig. 1a). Mg/Ca ratios were consistently high until ca. 1450 CE, declined over the next century, increased in the mid-to-late seventeenth century almost to the earlier levels, then declined again to reach a trough in the mid-1800s before rising slightly and then slowly declined and stabilized from about 1900 onwards. Time series from off New Zealand show similar long-term trends, as does a preliminary dataset from the Norfolk Ridge (Supplementary Fig. 3). On the basis of the regional coherence and studies cited above, we infer that AAIW in the SW Pacific was relatively warm until the mid-fifteenth century and then cooled over the past 500 years. The secular trend in apparent temperatures cannot be attributed to sampling bias in that they are independently evident at several sites and is not a function of the depth or temperature at which a specimen was collected (Supplementary Fig. 2).

Fig. 1: Comparison of millennial North Atlantic and SO time series.
Fig. 1: Comparison of millennial North Atlantic and SO time series.The alternative text for this image may have been generated using AI.
Full size image

a, The Mg/Ca ratios along radial transects across sections of SW Pacific bamboo corals. Colours and labels denote specimens. The MEME (vertical red line), based on the differences between 218 replicate analyses across the radius of K2, was 3.2 mmol mol−1. Dating methods varied among specimens (Methods), but most were determined by radiocarbon dating, adjusted by optimally aligning overlapping signals within the range of date uncertainty. It assumes a constant growth rate, which is approximately true for both the dated samples and for SW Pacific keratoisidids generally18. b, The overlap between the SW Pacific coral signal (from a; thin black lines) and indices of the AMOC rate 46 years later (red line as reconstructed by Rahmstorf et al.9; yellow line instrumental record from Rahmstorf et al.9; blue line AMOC transport from Moat et al.24). The correlation between the coral record, based on annual mean values, and the Rahmstorf et al. reconstruction peaks at a lag of 46 years (F1,719 = 380.6, P < 0.0001).

The implied AAIW temperature record corresponds closely with an index of AMOC strength9 since at least 1200 CE (Fig. 1b). Both the AAIW temperature proxy and the AMOC index are currently the lowest they have been in the past millennium22,23. The AMOC reconstruction and AMOC observations from 2004 to 202224 lag the AAIW record by ~50 years (Fig. 1b). The multicentury coherence and lag between the North Atlantic and SO records are consistent with models that link AMOC strength to AAIW temperature11,25 and salinity23,26, substantiate Rahmstorf et al.’s9 reconstruction of the AMOC (however, see ref. 27 for more details) and imply that large-scale changes in the Atlantic circulation are a delayed response to those in the SO.

Multicentury cooling at intermediate depths over a broad region of the SW Pacific suggests a response to large-scale climate drivers. To help identify these drivers, we complemented the Mg/Ca analyses with data on calcite oxygen and carbon isotope ratios (δ18O and δ13C), reservoir ages and boron isotope ratios (δ11B) (Methods; Fig. 2). The δ18O, measured in three corals that spanned the past ~850 years, can identify trends in salinity after correcting for ‘vital effects’28 resulting from the calcification process; reservoir age, measured in the same three corals, reflects the radiocarbon age of ambient AAIW; and δ11B, measured in a coral (K2) that spanned the past ~250 years, could reflect changes in ambient seawater pH29,30 (see the Methods for discussion on proxies). In all three corals analysed, the corrected δ18O correlates positively with Mg/Ca. The δ18O time series are congruent for the two specimens collected tens of kilometres apart (Keratoisis no. 7 (K7) and K2), whereas values are slightly elevated in the specimen collected further south (K6). Reservoir ages are stationary early in the data series. In the youngest specimen, however, they rise initially but subsequently decline from ~1700 to 1975 (least squares regression, F1,31 = 24.5, P < 0.0001). The δ11B record of K2 covaries antithetically with Mg/Ca and δ18O, rising 1.6‰ as Mg/Ca ratios fell ~13 mmol mol−1 between 1700 and 1900, and then stabilizing over the past ~100 years. The difference before and after 1800 is significant at P < 0.0001 (Mann–Whitney U7,23 of 0). Reservoir ages in this specimen correlate negatively with δ11B (F1,30 = 4.35, P = 0.046) but only weakly track Mg/Ca (F1,38 = 2.78, P = 0.10) and δ18O (F1,38 = 1.68, P = 0.20). Taken together, the Mg/Ca and δ18O records indicate the cooling and freshening of AAIW since the mid-fifteenth century, the reservoir ages show the age of AAIW in the SW Pacific has declined since the early 1700s and the δ11B record suggests a sustained increase in pH over the same period.

Fig. 2: Analysis of multiple environmental proxies in SW Pacific coral.
Fig. 2: Analysis of multiple environmental proxies in SW Pacific coral.The alternative text for this image may have been generated using AI.
Full size image

a–d, Comparisons of isotopic records (adjusted δ18O (b) and δ11B (d)) with each other, with the difference between radiocarbon and optimally aligned (wiggle-matched (WM)) ages (c), and with Mg/Ca records (a) from the same corals. Specimens, identified by colour, were collected south of Tasmania, Australia. Statistics are for least square regressions between Mg/Ca in each specimen and for all three specimens, δ18O and for K2, as well as δ11B and apparent reservoir ages (see text for more details). Red vertical bars indicate measurement uncertainty: for Mg/Ca and δ18O, based on the mean difference between replicate analyses of K2 (n = 218 and 36, respectively) (MEME), and for the age plot and δ11B, calculated instrumental analytical error. The δ18O values are residuals between the within-coral regression between δ13C and δ18O, which were then adjusted for an estimated 2 °C temperature change (Methods). Note the different δ18O scales for K6, K2 and K7. The analyses were each done at different laboratories using adjacent sections cut from the trunks of the corals. Slight temporal offsets between proxies therefore could be real but also could reflect analyses done at different laboratories and using the different coral sections (Methods).

These data rule out two local mechanisms as the primary cause of the observed isobaric temperature decline: displacement of warm, saltier Tasman Sea AAIW by a cooler, fresher southern-sourced variety14,31 and local subsidence of the water mass (that is, ‘heave’ of isopycnals). The first is inconsistent with the geographic extent of the signal. The second, if large enough to match the inferred temperature change (Methods), results in radiocarbon age estimates for the corals that are too old compared with alternative dating methods (Supplementary Fig. 4). Rather, the data are consistent with a hypothesis that the changes in AAIW properties reflect variability in the SO meridional overturning circulation (MOC). The upper cell of the SO MOC consists of the poleward flow of Upper Circumpolar Deep Water (CDW) balanced by the equatorward flow of AAIW and Sub-Antarctic Mode Water derived in part from upwelled CDW1,3. The temperature, salinity, and oxygen content of the CDW can be rapidly reset by air–sea interaction once it is entrained in the surface mixed layer. However, the equilibration times for CO2 (~1 year) and Δ14C (~10 years) are long relative to the residence time of SO surface waters32. Therefore, the signature of the upwelled CDW persists in SO surface waters, which are lower in pH and older (more negative Δ14C) than other surface waters in the global ocean32. AAIW thus retains a signature of the CDW from which it is derived. Less upwelling and/or entrainment of old, relatively saline and carbon-rich CDW would produce AAIW that is cooler, fresher, higher in pH and with a lower reservoir age. The inferred increase in pH and reduction in reservoir age in AAIW after ~1750 therefore suggest a gradual reduction in the volume of CDW incorporated into AAIW. Similarly, the increase in reservoir age between the early 1600s and ~1750, during a period when temperature and salinity of AAIW increased to a peak and then declined, is consistent with greater input of CDW during that period. Less CDW in the AAIW mixture after ~1750 can also account for reduced Southern Hemisphere atmospheric radiocarbon ages in the mid-to-late 1800s33 and variations in Weddell Sea coastal air temperatures over the past half millennium (Supplementary Fig. 5).

AMOC covaries with bottom water properties in the Weddell Sea

If the multicentury cooling and freshening of AAIW results from changes in the SO MOC, we might expect to see a similar signal in the deep overturning cell associated with formation of AABW. We tested this hypothesis by analysing Mg/Ca ratios and δ18O in a ~230-year-old bamboo coral (K22) that was collected alive in 1964 from between 2,119 and 2,592 m depth in the eastern basin of the Bransfield Strait. At these depths, the eastern basin is filled with water derived from the Weddell Sea34. As dense shelf waters derived in part from CDW in the western Weddell Sea are precursors of AABW35, the properties of deep water trapped in the topographically confined deep basins of the Bransfield Strait act as a proxy for bottom water formation in the Weddell Sea as a whole36. The Weddell Sea accounts for ~60% of the circumpolar production of AABW and is the source of bottom waters ventilating the abyssal Atlantic37,38.

Mg/Ca ratios in K22 track those in the pooled SW Pacific corals (Supplementary Fig. 7), being cyclostationary from 1760 to about 1865–1870 and then declining until the coral’s collection in 1964 (Fig. 3 and Supplementary Fig. 6). We tested the fit by comparing K22 with a specimen collected south of Tasmania (K2) that individually spans K22’s date range. Reflecting the long-term downward trend in both records, correlations between Mg/Ca ratios in the two corals are highly significant at all lags tested, between 0 and 48 years (Supplementary Fig. 7). However, the fit is best when the SW Pacific record is lagged by 27 years (correlation between detrended Mg/Ca series after lagging, F1,330 = 12.8, P = 0.004). The lag of several decades is consistent with the transit time of AAIW from its source region in the SE Pacific to the SW Pacific estimated from chlorofluorocarbon data39, suggesting both water mass signals are responding to changes in surface forcing in the SE Pacific/SW Atlantic. K22 δ18O (raw and after adjusting for ‘vital effects’) also declines over the lifetime of the coral (Supplementary Fig. 7b,c), again correlating with K2 and implying that Bransfield Strait bottom water, similar to the AAIW, has freshened as well as cooled since the mid-1800s. Correlations between the K22 trends and Southern Hemisphere climate variables (sea ice extent and the Southern Annular Mode) (Supplementary Fig. 8) are consistent with the hypothesized link to the SO MOC and the coral’s radiocarbon-based age estimate.

Fig. 3: Comparison of a proxy for deep SO processes and the North Atlantic time series.
Fig. 3: Comparison of a proxy for deep SO processes and the North Atlantic time series.The alternative text for this image may have been generated using AI.
Full size image

a, The Mg/Ca transect across the radius of the Bransfield Strait coral (K22), spanning 1760–1964 CE. Raw data, thin grey lines; the heavy black line indicates the 10-year running mean. b, The comparison of reconstructed and RAPID AMOC overturning rates (Moat et al.24; see Fig. 1 for colour codes) with 10-year filtered K22 Mg/Ca (black line) 76 years earlier. Inset: variance accounted for (R2) of linear regressions between AMOC and K22 Mg/Ca at lags ranging from −100 to +100 years.

The Bransfield Strait Mg/Ca record correlates highly with the Rahmstorf et al. AMOC reconstruction (F1,167 = 135.6, P < 0.0001) (Fig. 3a,b) when the latter is lagged by 70–80 years (peak range 74–77 years), as well as with a second, independently derived index of the AMOC40, at a lag of 70 years (Supplementary Fig. 9). The Mg/Ca and Rahmstorf AMOC records remain correlated when both datasets are detrended (F1,167 = 5.12, P < 0.02, lag 77 years), reflecting alignment at decadal scales; there is a similar but weaker trend for the second AMOC record after detrending (F1,144 = 2.45, P = 0.11, peak lag 70 years). At a similar lag, the K22 record also tends to covary with annual mean AMOC transport, as measured from 2004 to 2022 by the RAPID array24 (F1,18 = 2.73, P = 0.11, at lag of 72 years), despite relatively stationary transport in recent decades24. The lag of 70–80 years between the Bransfield Strait signal and AMOC indices is consistent with that predicted by the combination of the ~50-year lag between the AAIW and the AMOC reconstruction and the 20–30-year transit time of AAIW from its source region in the SE Pacific to the SW Pacific39. We also note significant correlations (P < 0.0001 for the raw data and P < 0.01 in detrended data) between the Bransfield Strait record and the AMOC reconstruction at lags of approximately minus (AMOC leading) 15 years and plus (SO MOC leading) 20 years (Fig. 3b, inset). The correlations suggest the link between the AMOC and SO MOC involves two-way interactions between the two hemispheres at relatively short lags.

AMOC, SO links and the role of local forcing

The instrumental record is too short to capture processes linking the northern and southern hemispheres on multidecadal timescales, and the proxy record is of relatively low resolution and incomplete. Model studies of varying complexity suggest the AMOC responds to changes in winds over the SO4,5,41, the properties of SO water masses and associated changes in freshwater transport9,10 and the stratification at the boundary between the SO and the Atlantic basin7,8. Our analyses indicate the dynamical coupling between the SO MOC and the AMOC involves processes acting at different timescales. On a scale of years, the AMOC both lags and leads SO properties, which likely reflects a rapid bidirectional propagation of Kelvin and Rossby waves along the planetary waveguide8,41,42. On a multidecadal timescale, the 70–80-year lag is consistent with the advection of AAIW anomalies from south to north11.

The indices of AMOC strength and AAIW properties diverge during two periods in the 1,260-year record (Fig. 4). Between approximately 1260 and 1450, the AMOC was weaker than expected from the long-term AMOC–AAIW relationship. This interval roughly coincides with the Medieval Warm Period in the northern hemisphere43. The two indices also diverge after the mid-1900s, where the AMOC is again weaker than expected. These two periods of divergence suggest times when forcing local to the North Atlantic exerted a persistent, strong influence on the AMOC44, superposed on remote forcing by southern hemisphere processes and an AMOC that the coral records suggest is preconditioned to remain weak in coming decades.

Fig. 4: Offsets between the Southern Hemisphere and North Atlantic time series.
Fig. 4: Offsets between the Southern Hemisphere and North Atlantic time series.The alternative text for this image may have been generated using AI.
Full size image

a, The comparison of standardized (z-score-normalized) values for the Rahmstorf et al.9 reconstruction of AMOC strength since 1200 (red line) and annual mean Mg/Ca ratios for AAIW-resident bamboo corals in the SW Pacific (black line). The latter have been lagged 46 years (see text for more details). b, Deviations from the predictive AMOC/AAIW relationship since ~1260. Positive (negative) values indicate stronger (weaker)-than-predicted North Atlantic overturning. The comparison suggests weak overturning relative to predictions before ~1450, a weakening beginning possibly as early as ~1940 and an intervening period of stationary but aperiodic variability.

Methods

Specimens

Corals were collected from the Chatham Rise, New Zealand (44.5° S, 179.1° E, 740–800 m), off the Norfolk Ridge, southeast of New Caledonia (24.2° S, 168.0° E, ~700 m), in the eastern basin of Bransfield Strait (61.95° S; 55.88° W, between 2,119 and 2,592 m) and from seamounts at three sites off southeast Australia, the ‘Southern Hills’ (44.1° S, 147.0° E, 987–1,160 m), the South Tasman Rise (46.5° S, 146.5° E, ~1,200 m) and the Cascade Plateau (44.0° S, 150.4° E, ~1,000 m) (Supplementary Fig. 1 and Supplementary Table 1). Specimens were obtained by dredging, in which case, sample depth was determined as the midpoint of the targeted depth range, or by remotely operated vehicle (see ref. 15 for details), in which case, depth was precisely known. The sample of K22, dredged live on 14 March 1964, was obtained from the Smithsonian Institution (United States National Museum specimen number 1019000).

Specimen dating

Dating methods varied among specimens, reflecting the range of studies involved (Supplementary Table 1). Most specimens were dated by radiocarbon analysis of calcite, done at the Australian National University (ANU), following procedures in ref. 45 and using sections adjacent to those geochemically analysed extrapolated to full age. Before analysis, samples spanning each coral’s growth radius were coarsely crushed into ~0.25-mm chunks and subjected to a 0.1 M HCl leach in which ~50% of the outside material was removed. The resulting clean carbonate was rinsed five times in 18MΩ water (Milli-Q) until near-neutral pH was achieved. A subsample of 8–10 mg of the CaCO3 was loaded into a 10-ml serum vial (BD Vacutainers) and evacuated, before CO2 was liberated by the introduction of 85% orthophosphoric acid (Ajax UNIVAR, Analytical Grade). The CO2 was then passed through a cryogenic water trap, measured for percent carbon yield and transferred to an individual graphite reactor assembly. In the presence of hydrogen and using Fe powder as a catalyst and a temperature of 570 °C, the CO2 was converted to graphite, with the resulting water from the reaction trapped using magnesium perchlorate. The graphite was loaded into Al cathode sample holders for 14,13,12C isotope analysis on the single-stage accelerator mass spectrometer located at the ANU Research School of Earth Sciences45. All samples were normalized to oxalic acid I and background-subtracted using 14C-free CaCO3. Uncertainty (1 σ) of the uncalibrated radiocarbon ages was based on the instrumental measurement error of percent modern carbon in each sample and, for calibrated ages, based on ref. 46.

With specific regard to K22, a regression of radiocarbon ages of the inner, middle and outer thirds of the coral’s calcite skeleton against radius is linear (R2 = 0.973) and indicates an age difference of 155 years between the innermost and outermost thirds (mean measurement error ± 43 years). Extrapolating the difference to the full radius of the coral implies a total age of about 227 years and a mean radial growth rate of 14.7 µm year−1.

L11, K9 and K11 were dated by the radiocarbon analysis of surface-derived organic tissue in nodes47,48, L11 and K9 based on the ‘bomb’ radiocarbon signal. As K11 was too old to date against the bomb signal, its maximum age was determined by subtracting an assumed surface reservoir age of 330 years, on the basis of ref. 49, from the radiocarbon date of the centre of the specimen. Subsequent growth and age for K11 were based on increment counts. The organic material was obtained by decalcifying the organic nodes (1 M HCl for 12–24 h), which were then softened in 18MΩ Milli-Q overnight. Concentric bands were hand-sampled using precombusted tweezers and disposable syringe tips under a Nikon inverted microscope (ECLIPSE TE2000-U). Graphite was produced on a dedicated vacuum line to substantially reduce the sample mass required for routine analysis: ~1 mgC to <50 μgC before introduction to the accelerator mass spectrometer. Precision averaged ±4‰ (n = 55)47.

Z10909, live caught, was dated by increment count50 and K19 by aligning the Mg/Ca series of the apparently long dead (surface pitted and eroded) specimen onto that of a radiocarbon-dated specimen (K18) from the same area.

The empirically determined 14C dating error of individual calcite analyses, based on seven replicates spanning two specimens, averaged 51 years (range 10–110 years). Three replicate age determinations for L9 based on node 14C ‘bomb calibration’ ranged from 104 to 117 years (average s.d. ± 27 years)47. All specimens were sampled multiple times for dating (n = 2–55, median 6–7 analyses) (Supplementary Table 1), minimizing the effects of single-point analytical error on a specimen’s overall age estimate. Final dating was based on ‘wiggle matching’51 irregular variability in Mg/Ca ratios, within the range permitted by the uncertainty of the radiocarbon dating, between specimens with overlapping 14C ages. The accuracy of age/date determination was testable directly by comparing that for K22 with instrumental records of the Southern Annular Mode (SAM). Consistent with the instrumental analyses and models52, the decline in Mg/Ca ratios (and inferred changes in upwelling) in K22 correlates significantly with an increasingly positive SAM (Supplementary Fig. 9). The age implied for the coral (~229 years) is in good agreement with its radiocarbon age (~227 years). Reservoir ages for deep water in the SW Pacific are poorly known. For the analysis of the southeast Australian material, we based reservoir ages on refs. 53,54,55, on an alignment between the growth rates of co-located known age orange roughy (Hoplostethus atlanticus) and coral Mg/Ca56 and on the alignments between Mg/Ca trajectories of live-caught and recent subfossil specimens (Fig. 1). The correlation between surface air temperatures in the Weddell Sea, close to the source location of SW Pacific AAIW57, over the past millennium and the AAIW time series (Supplementary Fig. 5) appears to further validate both the wiggle match-adjusted radiocarbon dating of the corals and the assumed reservoir age. For the Norfolk Ridge specimens, dates were assigned on the basis of differences in calcite radiocarbon ages relative to a modern specimen (DW1697).

Analytic procedures

Most specimens were analysed using electron probe microanalyses, following procedures in ref. 58 and, specifically as applied to keratoisids, ref. 59. Microanalysis was performed on basal sections cut from each coral using a geological diamond saw (Supplementary Fig. 10). The surface of the coral was lapped on a glass plate using 6-µm and then 3-µm aluminium oxide paste before final polishing on a pellon plate using 1-µm diamond suspension in an oil slurry. Sections were ultrasonically cleaned in ethanol between all stages and after the final polish. The finished sections were air dried and then placed in a vacuum for a least 1 h before coating to remove surface bound water. Samples were coated before analysis with a 25-nm coat of carbon using a carbon arc. Microanalysis was performed on a JEOL 8900 R Superprobe operated using combined wavelength and energy-dispersive spectrometers. Measurements were taken at 100-µm intervals, beam spot centre-to-centre, using, in all but one case (K6), a circular 50-nA beam accelerated at 15 kV and defocused to 50 µm. K6 was measured using a defocused 100-µm beam; calibration studies59 indicate a negligible difference in estimated weight fractions between defocused 50- and 100-µm beams. At each analysis point, weight fractions (percent dry weight, expressed as percentage or parts per million) were measured simultaneously for Mg using the wavelength-dispersive spectrometers and for Ca using the energy-dispersive spectrometer. Reference materials used for calibration were spinel (MgAl2O4) for Mg and wollastonite (CaSiO3) for Ca. The standards were reverified at the end of each analysis run. The differences between the runs and reference materials were typically 1–3%. In the rare cases where the difference was greater than this, the run was discarded as unduly affected by spectrometer drift.

The measured concentrations (Mg, 1.85–2.47 wt%, Ca, 33.08–39.9 wt%) are similar to those in other biogenic calcites and average more than an order of magnitude higher than the elements’ respective minimum detection limits (170 ppm and 1.5 wt%, respectively). Mean measurement error, based on the comparison of 570 replicate analyses spanning two corals, was ±334 ppm (1.2%) for Mg and ±1,830 ppm (0.5%) for Ca (see ref. 59 for more details). A comparison of 218 replicate analyses across the radius of K2 indicated an average measurement error (mean empirical measurement error (MEME)) for Mg/Ca of 3.19 mmol mol−1. Elevated Mg/Ca around the hollow central core of the coral (Supplementary Fig. 10), which is secondarily infilled54, was excluded from the analysis, as were points at which the apparent weight-fraction of Ca was less than 33%, indicative of low-quality data due to surface irregularities58.

K8 was analysed using laser ablation inductively coupled plasma mass spectrometry, detailed by ref. 47. Carbonate subsamples were collected using a Merchantek New Wave micromill, cleaned (approx. 10% HCl leach) and triple-rinsed with18MΩ Milli-Q. All samples were oven-dried at 50 °C for 3 days and stored in precombusted (400 °C) glass vials at room temperature until they were analysed. Mg and Ca were analysed at the ANU Research School of Earth Sciences (Canberra, Australia) using the ANU HelEx laser ablation system (193-nm ArF Excimer laser and laser ablation cell microsampling system) coupled to a Varian inductively coupled plasma mass spectrometry. Tracks were pre-ablated with a 200-μm diameter laser spot pulsing at 10 Hz, scanning at 200 μm s−1. Analytical passes were conducted with the following laser operation conditions: repetition rate of 5 Hz; spot size of 50 μm; laser output of 50 mJ cm−2; laser scan speed of 20 μm s−1. Laser tracks were bracketed with NIST SRM 614 and a coral pressed powder pellet reference material and sensitivity optimized by monitoring ThO+/Th+ (<0.5%). Laser tracks were run in triplicates and then smoothed using a three point-weighted average. MEME based on 184 replicate analyses across the coral’s radius was 1.32 mmol mol−1.

The preliminary analysis of the Norfolk Ridge samples was done using inductively coupled plasma atomic emission spectroscopy at the Institute de Chimie, Universite de Paris-Saclay. Cross-sections of the coral trunks were made using a diamond saw, then dried before analysis. Powder samples were milled at 0.5-mm intervals along the radius of sections of each coral. Samples were dissolved in 10 ml of 2% nitric acid and sonicated for 10 min to facilitate dissolution before spectrophotometric analysis.

The δ18O and δ13C measurements were done at the New Zealand National Institute for Water and Atmosphere using a Finnigan MAT 252 mass spectrometer, on CO2 liberated from samples micromilled at 25–100-µm intervals, depending on the specimen, along each coral’s radial section. Analytical procedures follow ref. 60. Concurrently run carbonate reference material (NBS19) had an internal precision (1 s.d.) of 0.02–0.08‰ and an external precision (between runs), based on the difference between standard means, of 0.03‰ for δ18O and 0.01–0.06‰ and 0.02‰ for δ13C. Replicate analyses of the outermost 3.7 mm of K2 (n = 36) indicated an average difference between replicates (MEME) of 0.12‰ (range 0.02–0.40) and 0.11‰ (0.00–0.26) for δ18O and δ13C, respectively, and 0.06‰ (0.00–0.23) for the residuals of the regression of δ18O against δ13C. All results are reported relative to vPDB, where δ18O has a value of −2.20‰ and δ13C +1.95‰ for NBS19 calcite.

Sample preparation and measurements of boron isotopes largely followed the protocol of ref. 61, modified slightly to accommodate the larger sample size required for low boron (~10 ppm B) calcite. The K2 radial section was milled along the growth axis in 1-mm increments, to a depth of 4.6 mm and across a (mostly) 5-mm band within the 1-mm growth interval. Each subsample was cleaned in dilute sodium hypochlorite, then sonicated, rinsed in purified (Milli-Q) water and dried before dissolution. Dried powder aliquots of 50 mg were dissolved in nitric acid, with 100 µl of 0.075 N HNO3 then 400 µl of 2.95 N HNO3. The dissolved 500-µ- samples were loaded onto preconditioned ion exchange columns, similar to the 2-tiered cation-anion column setup of ref. 61, a simple protocol that removed the major ions and ensured 100% B recovery. The difference here was using a larger, 10-ml, capacity cation column containing 1.8 ml of AG50W-x8 above the 2.4-ml column with 1 ml of AG1-x8 anion resin. The sample solutions were washed through the columns with 1 ml of 0.075 N HNO3 then the boron collected with three passes of 750 µl of 0.075 N HNO3 to yield a 2.25-ml B solution. Blanks and an in-house modern aragonite coral standard (UWA24.7)61 were also processed through ion exchange chromatography to check reproducibility of the chemical and mass spectrometry procedures. The B sample solutions were analysed using a NU Plasma II multicollector inductively couple plasma mass spectrometer at the University of Western Australia, directly following the procedure of ref. 62, each sample bracketed by the NIST SRM 951 boron standard diluted to 150 ppb. Following convention, the boron isotopic compositions are given in permil (‰) notation. The long-term reproducibility of the coral standard was 25.9 ± 0.3‰ (2 s.d., n = 98)61.

Statistical procedures broadly follow ref. 62. Non-parametric, least square regression analyses and data smoothing were done using Staview and JMP; spectral analysis was done using K-Spectra. P values associated with statistical tests indicate the probability that the results are due to chance; hence, P < 0.01 signifies that a tested association has a probability of less than 1 in 100 of occurring by chance. Relationships are typically considered significant if P is less than 0.05 (5% probability of occurring owing to chance). Gaps in the annually resolved changes in Mg/Ca in the pooled data for the SE Australian specimens (Fig. 1a) were filled by linear interpolation. To test correlations between data sampled at different frequencies or spacing (for example, comparing isotopic and elemental data in the same coral), the higher density dataset was smoothed using a three-point running mean and then resampled at a frequency or spacing equal to that of the lower density set. To minimize the effects of autocorrelation in the non-stationary datasets, for analyses involving K22, which we consider the more robust predictor of AMOC given its proximity to the Weddell Sea, we tested for correlations using both raw and detrended data. For the comparison between the AMOC and the SW Pacific coral dataset, we determined the optimal lag on the basis of raw data only, given pronounced autocorrelations in both datasets that would not have been eliminated by linear detrending.

Elemental and isotope ratios as proxies

In gorgonian calcite, numerous studies have suggested specimen mean Mg/Ca ratios are a proxy for ambient temperature18,19,20,21,63 and growth rate18, which in turn is temperature-correlated54,60. Among specimens, the reported slope of the keratoisidid regression ranges from >0.3 °C (mmol mol−1)−1 of Mg/Ca to 0.05 °C (mmol mol−1)−1 (refs. 16,18). Similar, generally steep slopes have been reported elsewhere for diverse marine biogenic calcites (reviewed by refs. 16,19,20), as well as abiotic calcite17,64. On the basis of a recent analysis65, the decline in AAIW temperature implied the SH corals is ~1.8 °C, but the range of values reported for keratoisidids, although consistently strongly suggesting that falling Mg/Ca ratios in SW Pacific and Bransfield Strait specimens indicate declining temperatures, further suggests that the magnitude of the decline remain uncertain pending additional studies.

Oxygen isotope ratios in seawater reflect salinity and temperature but also correlate highly in corals owing to ‘vital effects’ consequent of the calcification process28,66,67. The effects of salinity can be extracted in keratoisidids68,69 by analysis of the residuals of the regression between δ18O and δ13C—the ‘lines’ method70. A positive offset of δ18O indicates elevated salinity. We extracted residual δ18O from radial transects of three corals that jointly span the last 800 years BP (K6, K7 and K2), plotting individual points against dates assuming stationary radial growth rates. Values were adjusted for temperature-dependent fractionation of δ18O (−0.22‰ per degree Celsius)71 using an estimated long-term change in temperature of 2 °C (the approximate midpoint of the range implied above). The specifics of the method, its validation and calibration are reported by ref. 69. We note that adjusting the residual δ18O time series for temperature effects steepened and increased the significance of the decline across the radius of K2 (F1,217 = 493.0, P < 0.0001), but the decline was significant without the temperature adjustment (F1,217 = 330.3, P < 0.0001). The inferred freshening at AIWW core depths is consistent with regional instrumental analyses that span the past half century72. Similarly, the decline in AAIW reservoir ages we infer from the difference in wiggle match and radiocarbon ages for K2 is consistent with increased 14C values for the water mass in the SW Pacific over recent decades73. For K22, the residual δ18O time series was adjusted for an estimated 0.6 °C temperature decline, on the basis of ref. 71.

The boron isotope composition of marine biogenic carbonates (δ11Bcarb) is controlled by the pH-dependent uptake of borate ions (B(OH)4−) from seawater during biomineralization74,75. Thus, numerous studies have explored δ11B as a proxy for seawater and intracellular pH, largely applied to calcitic foraminifers and aragonitic corals for example76,77,78,79. Although skeletal aragonite δ11B reflects calcifying fluid pH in scleractinians owing to their ability to upregulate intracellular pH during calcification80,81,82, calcite δ11B compositions in foraminifers and octocorals have been shown to reflect near-ambient seawater pH, for example, refs. 29,30,79,80,81,82,83, although inorganic calcite experiments imply that δ11B might be affected by skeletal precipitation rates in some taxa30. In this study, we interpret the measurable increase in K2 δ11B to largely reflect a temporal increase in seawater pH at the sample collection site, hence at AAIW core depth, given its consistency with other proxy data and δ11B-derived pH of the youngest layer with present-day ambient seawater pH.

The δ11B data also argue against an alternative explanation for the long-term decline in Mg/Ca ratios on the basis of the effect of pH on calcification rates. As bamboo corals show little or no offset of the carbonate incorporated borate isotope curve relative to seawater, higher seawater pH, as indicated by the δ11B data, would result in a higher internal pH. The latter, in turn, promotes a higher calcite saturation level, an increased precipitation rate and more rather than less Mg incorporated into the skeleton18,84,85.