a,¹⁴C and²¹⁰Pb dating. The¹⁴C ages (with 1σ ranges; grey, rejected dates) from planktic foraminifera yield a modern core-top age and indicate an average sedimentation rate over the past 1,000 years of 320 cm kyr⁻¹ (dashed line). The presence throughout the core of abundant lithogenic grains in the >150-μm fraction—along with the coarse sortable-silt mean grain size values—suggests that some reworking of foraminifera has probably occurred, resulting in average¹⁴C ages that may be slightly (around 50 years) older than their final depositional age, consistent with the fact that the²¹⁰Pb dates do not splice smoothly into the¹⁴C ages (the¹⁴C ages appear slightly too old). The final age model was therefore based on the²¹⁰Pb ages for the past century, and was then simply extrapolated back in time using the linear sedimentation rate of 320 cm kyr⁻¹. Given that none of our findings depend on close age control in the older section of this core (that is, beforead 1880), this uncertainty (with converted¹⁴C ages being about 50 years older than the extrapolated linear age model) does not affect our conclusions.b, Left, the age model for the top 80 cm of core 56JPC is based on²¹⁰Pb dating of bulk sediment, using the constant initial concentration (CIC) method (rejecting the date at 47 cm, which probably indicates a burrow). A simple two-segment linear fit to the²¹⁰Pb dates is adopted (rather than point-to-point interpolation or a spline) because sedimentological evidence—an abrupt increase in the percentage of coarse fraction at 23 cm depth, not observed elsewhere in the core—is indicative of a step change in the sedimentation rate. Horizontal dashed lines denote the depths of the segments at which the sedimentation rate is inferred to change. Centre, further support for the age model of 56JPC over the past century comes from the down-core abundance profile of spheroidal carbonaceous particles (SCPs, derived from high-temperature fossil fuel combustion, counted as described³⁹), which ramped up from the mid to late 1800s and peaked in the 1950s to 1970s (40 cm to 25 cm) before declining over recent decades, consistent with the²¹⁰Pb-based age model. Right, the occurrence of¹³⁷Cs in the top 40 cm or so of the core is also consistent with the²¹⁰Pb-based age of around 1950 at 40 cm. The age uncertainty (1σ) for the past 60 years of the core is estimated at ±2–3 years. We note that the sediment core top is at 3 cm depth in the core-liner.

Source data

a,¹⁴C-based age model, derived from linear interpolation of¹⁴C-dated planktic foraminifera (with 1σ ranges) in sediment core KNR-178-48JPC (used for the DWBC_LSW sortable-silt reconstruction), yielding a modern core-top age and an average sedimentation rate of around 50 cm kyr⁻¹. We note that the core top is at 3 cm depth in the core-liner. The inset shows the SCP profile for 48JPC on the basis of the¹⁴C age model, confirming the modern age of the top sediments, with SCPs showing the expected profile—increasing in concentration from the late 1800s onwards, peaking at around 1950 to 1970, and declining afterwards.b, Updated age model for core KNR-158-10MC (after ref.⁴⁷; used in Extended Data Fig. 5 examining regional near-surface temperature trends in the Northwest Atlantic during the industrial era), using new²¹⁰Pb dating (CIC method) for the top 7 cm and rejecting the anomalously old¹⁴C age at 4 cm depth; the inset shows²¹⁰Pb age constraints in the top 8 cm. A single detectable occurrence of¹³⁷Cs at 2–2.5 cm (equivalent to 1957 on the²¹⁰Pb-based age model) can be linked to the bomb peak at 1963, supporting the age model. Also, SCPs were found in the top 5 cm of this core, confirming the industrial-era age for the top 5 cm; however, the low concentrations of SCPs prevent meaningful interpretation of the down-core trends and are not shown.c, Age model for core OCE-326-MC29B (used forT_sub reconstruction of the Northwest Atlantic shelf):¹⁴C ages of planktic foraminifera (with 1σ ranges), from ref.⁴⁸. Support for this age model is provided by the SCP concentrations (inset; this study), which show the expected down-core profile³⁹ when plotted using the¹⁴C ages.²¹⁰Pb dating⁴⁸ also suggests a sedimentation rate of around 120 cm kyr⁻¹ for the uppermost sediments, consistent with the¹⁴C ages and SCP profile.

Source data

Locations are shown in Fig. 2b.a–c, Temperature proxy records^48,49,50 used for the Northwest Atlantic stack (Emerald Basin, Laurentian Fan and Gulf of St Lawrence), where model studies^11,12 indicate that AMOC weakening results in warming of surface and subsurface waters.d–g, Records used to reconstruct Northeast Atlantic SPG subsurface temperatures:d, Gardar drift⁵¹;e, combined South Iceland data (Bjorn drift)^52,53;f, Feni drift⁵⁴;g, Eastern North Atlantic Central Water (ENACW), largely composed of waters formed in the eastern SPG^55,56.h, The high-resolution alkenone sea-surface temperature (SST) record from the North Iceland shelf⁵⁷ was not included because it is not located within the open North Atlantic SPG (although it does also show, like the other Northeast Atlantic records, that the lowest temperature of the past 1,600 years occurred during the most recent century). Also shown for reference is the Rahmstorf central SPG SST reconstruction (based largely on terrestrial proxies)⁶.

a,b, Stacked, normalized proxy temperature data (T_sub) from the Northwest Atlantic shelf/slope (a) and Northeast Atlantic SPG (b).c, The derivedT_sub AMOC proxy, calculated as the numerical difference between the stacks shown ina andb.d, The residual temperature variability in stacksa andb that is not described by the (anti-phased dipole)T_sub AMOC proxy shown inc—that is, the in-phase temperature variability common to both stacks, calculated as the numerical sum of the two stacks (if divided by two, this would be the numerical mean). This represents the inferred non-AMOC-related temperature variability common to both regions, and broadly resembles Northern Hemisphere temperature reconstructions, most notably colder residual temperatures during the LIA, around 1350 to 1850. Fora–d, black solid lines and squares represent preferred binning (50 years for 1800–2000; 100 years for 400–1800); green line and symbols, as for preferred binning, but with stacks produced by first binning the proxy data at each site and then averaging these binned site values, as opposed to binning all the proxy data together in one step (the former ensures equal weighting for each site, the latter biases the final result to the higher-resolution records); black dashed lines and symbols, 100-year bins offset by 50 years from the preferred bins; grey lines and symbols, 50-year bins (not shown forc andd); blue lines and symbols; 30-year bins for 1790–2000. Error bars fora–d are ±2 s.e.e–g, As fora–c, except using a Monte Carlo approach and published uncertainties for age assignment and temperature reconstructions; light and dark grey shading represent ±1σ and ±2σ, respectively.h, Jackknife version ofc, with each line representing theT_sub AMOC proxy but leaving out one of the individual proxy records each time.

Source data

a, Modelled SST difference between a weak (negative) and strong (positive) AMOC⁵⁸. This pattern is model-dependent, with the study cited here⁵⁸ chosen because of its good agreement with observations of Gulf Stream variability. The locations of cores used for panelb are shown by black stars.b, Percentage abundances of the polar speciesN. pachyderma (sinistral) in marine sediment cores from the Northwest Atlantic, as an indicator of near-surface (around 75 m) temperatures. A 15% increase indicates around 1 °C of cooling (we note the reversedy axes). The opposing trends over the past 200 years are consistent with the SST pattern modelled for a weakening of the AMOC, as shown in panela. Data and age models for the cores are: OCE326-MC29⁴⁸ using the original¹⁴C dating and as shown in Extended Data Fig. 2; OCE326-MC13 and OCE326-MC25⁴⁹ using the original¹⁴C age ties at the top and bottom of the core and scaling the intervening sedimentation rate to the percentage of CaCO₃ content^49,59,60; KNR158-MC10, this study, using the age model in Extended Data Fig. 2.

Source data

a, Top,T_sub AMOC fingerprint¹¹ obtained using empirical orthogonal function (EOF) analysis of the EN4 dataset (light green, the leading mode (EOF1) ofT_sub variability from 1993–2003, as defined by Zhang¹¹, applied to the EN4 data; dark green, the second mode ofT_sub variability (EOF2) of the North Atlantic for 1900–2015, equivalent to the EOF1 defined for 1993–2003). No substantial twentieth-century AMOC decline is seen in this observation-based reconstruction. Bottom, instrument-based reanalysis of the ‘cold blob’ central SPG region (red; 3-year (thin line) and 11-year (thick line) smoothing; 47° N to 57° N, 30° W to 45° W) used in the Rahmstorf SST AMOC proxy⁶. The data are from the HadISST project. The reconstructed central SPG SST bears some resemblance to theT_sub AMOC fingerprint record, which is not unexpected given that the central SPG forms a substantial spatial component of theT_sub fingerprint. No clear decrease is shown in the central SPG SST, and the equivalent Rahmstorf AMOC proxy⁶ (blue; central SPG minus the Northern Hemisphere (NH) temperature) declines during the twentieth century because of the subtraction of the NH warming trend.b, Reconstructed (predominantly terrestrial-based) AMOC proxy (orange; the temperature difference between the central SPG and the NH) and the central SPG SST reconstruction⁶ (blue). There is a two-step decline in the AMOC proxy, at 1850–1900 and 1950–2000—the former being mainly the result of a strong cooling of the SPG (which probably weakened northward heat transport, paralleling the weakening shown by our DWBC proxy), and the latter being due mainly to subtraction of the strong NH warming trend, rather than a persistent SPG cooling.

Source data

a,b, Climatological surface current direction (in arrows) and speed (shaded, m s⁻¹) obtained from the control simulation with HadGEM3-GC2 and the satellite product OSCAR.

a, Correlation (colour bar) of the vertically averaged ocean density (at 1,000–2,500 m) with the DLSD index (as defined in ref.⁴; green box, 1,000–2,500 m average) in a 340-year present-day control run of the HiGEM model (see ref.³⁶).b, Climatology of the modelled meridional ocean velocity (in m s⁻¹) averaged between 30° N and 35° N, illustrating the modelled position of the DWBC. They axis shows the water depth in metres.c, Cross-correlations between the modelled average DWBC flow speed in the pink box in panelb and indices of DLSD and AMOC at 45° N (the dashed line omits the Ekman component). We note that the box over which the DWBC flow index in panelc is averaged has changed with respect to Fig. 1, in order to take into account of the fact that the return flow is deeper in the HiGEM model than in HadGEM3-GC2.

The model-based DLSD parameter—proposed in ref.⁴ and using the EN4 reanalysis dataset—incorporates a larger area and greater depth range than do instrumental-data-only studies, such as ref.⁵, which examines past variability in Labrador Sea convection and focuses on the central Labrador Sea and on depths less than 2,000 m, where most observational data are available. The comparison here of DLSD (purple line, three-year mean) from the EN4 dataset with instrumental data on density changes in the central Labrador Sea at 1,500–1,900 m depth (grey line, annual averages; black line, three-year mean) illustrates that the two parameters show very similar variability. Both are dominated by the density changes caused by deep convection in the Labrador Sea, which can reach down to around 2,000 m. Estimates of uncertainty are discussed in ref.⁶¹.

Source data

A direct influence of the changing position of the Gulf Stream on the grain size of our core sites can be ruled out by comparing instrumental records of the Gulf Stream position (red, GSI⁵⁸) with the down-core sortable-silt (SS) mean grain size data in 56JPC (blue; thicker line is three-point smoothed). There is no clear correlation between these two proxies (bottom). However, there is a coupling between our SS data (which represent inferred DWBC_LSW flow speed) and density changes in the deep Labrador Sea (grey, annual; black, three-point smoothed; top panel). The 2σ SS error bar (n = 30) is for the three-point mean.