Introduction

The system of ocean currents collectively known as the Atlantic Meridional Overturning Circulation (AMOC) are a major driver of climate variability. The AMOC is a highly non-linear system, and some climate models suggest that it will cross a tipping point this century due to additional external forcing from increasing greenhouse gas concentrations1,2. A substantial AMOC slowdown would have serious ramifications for regional climates, such as up to ~15 °C mean annual temperature reductions in Northwest Europe3. Thus, understanding its response to climate forcing is critical. One key component of the AMOC is the Gulf Stream (GS), which transports warm water from the Gulf of Mexico northward towards Northwest Europe (Fig. 1). Due in part to the GS, winter surface air temperatures in western Europe are up to 10 °C higher than the zonal mean at comparable latitudes4. GS path and AMOC strength are intrinsically linked. A weaker AMOC reduces the Deep Western Boundary Current (DWBC), weakens the northern recirculation gyre, and shifts the GS separation point northward5. Anthropogenic warming has reduced AMOC strength to perhaps its weakest state over the past millennium, inducing a northward migration of the Gulf Stream North Wall6,7. Temperature records from the Gulf of Maine have traced the effects of GS path migration over the last 100 years8, with northward movement raising regional sea surface temperatures (SSTs). However, on pre-instrumental timescales that are critical for understanding sensitivity to global temperature change, precise GS position remains elusive.

Fig. 1: Schematic of relevant ocean currents in the North Atlantic Ocean.
figure 1

a Idealised modern day ocean currents (modified from ref. 88). The GS North Wall describes the northern extent of the GS and is defined by the strong temperature front. The point at which the path of the GS diverges from the eastern coastline of North America is known as the GS Separation Point. b Proposed change in the ocean currents during the LIA compared to modern day. The size of the arrows relative to those in (a) denote the change in strength. Note that the locations are purely illustrative of the direction of movement and thus are likely exaggerated and idealised. The black box highlights the location of Bermuda. The abbreviations are as follows: DWBC Deep Western Boundary Current, GS Gulf Stream, LC Labrador Current, NRG Northern Recirculation Gyre, SWC Slope Water Current.

Full size image

The Little Ice Age (LIA; 1300-1850 CE) represents a key time interval for understanding how GS positioning and dynamics respond to climate change because it was characterised by extended Northern Hemisphere (NH) cold intervals. Proxy evidence suggests the LIA featured a GS that was approximately 10% weaker and positioned further south, likely due to a stronger Labrador Current (LC) and enhanced recirculation gyre9,10,11. The lack of instrumental records from the LIA means that palaeoclimate reconstructions are crucial for inferring GS state and position.

Stalagmites are increasingly providing important constraints on the interpretation of palaeoclimates, particularly due to their propensity to provide accurate and precise chronologies and potential for continuous, high-resolution records, in some cases at sub-monthly timescales. Strong environmental seasonality can result in visible or geochemical laminations offering high-precision chronological control, complementing any radiometric (e.g., U-Th or 14C) determinations (e.g., refs. 12,13). Magnesium concentrations in stalagmites can exhibit a seasonal signal (e.g., refs. 13,14) that can be employed both as a chronological tool via its seasonal cyclicity and as a climate proxy through its relative values. Stalagmite Mg concentrations can record temperature based on the temperature dependency of Mg’s partition coefficient (DMg) in calcite (e.g., refs. 13,15). However, sea spray contributions to rainfall may also affect dripwater chemistry, especially in island karst systems and, consequently, stalagmite Mg concentrations16,17,18,19.

Bermuda’s position within the recirculation gyres south of the GS makes it an ideal site to capture the effects of a southward GS shift, and the presence of extensive cave systems offers a rich speleothem archive for climate reconstruction. However, while other SST reconstructions for Bermuda exist (e.g., refs. 20,21,22,23), few provide a continuous record spanning more than the last three centuries. Understanding the palaeoclimatic changes experienced in Bermuda on this temporal scale will provide new insights into how the LIA, with its clearly expressed NH climate anomaly, affected GS position and strength9,24,25,26.

Here, we use the Mg concentrations of a Bermudan stalagmite, which are connected to sea spray contributions to dripwater, to construct a monthly palaeotemperature record spanning 564 years (1449–2013 CE). We derived a Mg-SST calibration based on a regression between our record and a previously developed coral Sr/Ca-SST record21. The resulting reconstruction yields SSTs covering most of the LIA and reveals past GS migration patterns at very high temporal resolution and chronological accuracy.

Site description

Leamington Cave in northeastern Bermuda (32° 20′ 31.64″ N, 64° 42′ 30.93″ W; entrance at 18 m a.s.l.) is a privately-owned former show cave connected to the ocean via flooded conduits and a large cave pool (Supplementary Fig. 1a). The cave contains many actively growing speleothems above the high tide line, with cave monitoring data revealing a same-day drip rate response to rainfall that suggests residence times capable of preserving a seasonal signal (Supplementary Fig. 2). The Bermuda Islands consist of a carbonate platform above a volcanic seamount and are comprised almost entirely of aeolianites and palaeosols27. The Walsingham Formation, an aeolianite limestone unit (the oldest limestone unit, deposited approximately 1.1–0.8 Ma), hosts most of Bermuda’s caves (including Leamington Cave) and crops out along the eastern border of the Harrington Sound and the western edge of Castle Harbour28,29.

Bermuda has a humid subtropical climate (Cfa in Köppen–Geiger classification30) and lies within the North Atlantic hurricane belt, with the hurricane season running from June to November each year. Meteorological data31 indicate that mean SSTs are lowest in February (17.8 °C) and peak in August (28.5 °C) (Supplementary Fig. 1c). May historically experiences the lowest mean monthly rainfall (86.7 mm), whereas October has the highest (161.6 mm)31. The hurricane season accounts for a mean of 55% of annual rainfall. Except for September and October, when winds prevail from the east, winds typically come from SW and SSW, reaching a maximum in January (6.01 m/s) and a minimum in September (4.49 m/s)31.

Results and discussion

Bermuda temperature trends

Our SST reconstruction for Bermuda is derived from stalagmite BER-SWI-13 using an age model constructed from a Mg cycle count. The cycle count is modelled to a radiocarbon chronology in intervals where cyclicity is less pronounced and is better constrained and more precise than both the radiocarbon and U-Th chronologies (Fig. 2; full details of our chronology and age model construction are discussed in Materials and Methods). By virtue of the inverse relationship between wind speed and SST, our monthly-resolved Mg record serves as an indirect proxy for SST via sea spray contributions to dripwater (Fig. 3). SST estimates were obtained through calibrating the stalagmite Mg record to a previously published coralline SST record (r = −0.56, p ≪ 0.001; 1782–199821); (Fig. 4a, b). The record presented here cannot be confidently compared to observational SST data (starting in 1950) due to anthropogenic influences on speleothem deposition post-1961, likely as a result of land use change above the site (Supplementary Fig. 5). However, because the coral-derived record calibration is derived from a regression with instrumental data21, the significant correlation with the stalagmite reconstruction likely reinforces the accuracy of our record, despite the errors being amplified by the additional proxy regression. The resultant record spans 564 years (1449 to 2013 CE), making it one of the longest continuous high-resolution records for the region.

Fig. 2: Stalagmite BER-SWI-13 chronology.
figure 2

a Development of the BER-SWI-13 chronology, comparing the preliminary cycle count (green) and the radiocarbon (black) to the final cycle count (blue), which is developed from both. The grey shaded area signifies the 2σ confidence interval of the radiocarbon chronology. The cycle counts plot outside the radiocarbon chronology errors at the top as the radiocarbon model used a rough estimate for the bomb spike, which was updated in the cycle counts. Black areas (left-hand side) mark depths modelled to the radiocarbon mean. b Radiocarbon data recorded in the stalagmite (purple) compared to the U-Th dates (cyan). The black arrow indicates the trajectory of the radiocarbon data if only influenced by the Suess Effect. Note that the radiocarbon ages here have not been corrected for dead carbon fraction and that the U-Th dates use (238U/232Th)initial = 5 ± 2.5. c Mean growth rates based on radiocarbon (black, 0.345 mm yr−1), preliminary (green, 0.373 mm yr−1), and final (blue, 0.360 mm yr−1) chronologies.

Full size image
Fig. 3: Schematic of the proposed Mg-SST proxy mechanisms.
figure 3

a Observational wind speed data (red) plotted against observational SST data (blue)31. b Mg in collected rainwater (magenta)89 plotted against the observational SST data in (a). c BER-SWI-13 Mg concentrations (gold) shown against Bermudan SSTs reconstructed from a coral (green)21. d BER-SWI-13 Mg concentrations (gold) shown against the observational SST data in (a). Stalagmite Mg cyclicity is not directly compared to wind speed or rainwater Mg variability due to potential anthropogenic impacts on the Mg record post-1961 and wind speed and rainwater observations beginning in 1975 and 1989, respectively. Note that the stalagmite Mg and wind speed axes are inverted and that the bold italics in the schematic denote source concentrations (i.e., ocean water is Mg-rich and rainwater has comparatively low Mg concentrations).

Full size image
Fig. 4: The stalagmite-derived SST reconstruction.
figure 4

a The model predictions against the warped coralline SSTs. The linear best-fit for these data with 95% confidence interval (grey shading) is compared to that of an ideal relationship (i.e., SSTfinal = SSTcoral). b The regression model output (1449–2013 CE; 564 years) at both a monthly (light blue) and annual (dark blue) resolution plotted with a coralline SST record (grey; 1782–1998 CE; 217 years) from Bermuda21.

Full size image

The stalagmite-derived SST reconstruction exhibits notable agreement with both six previous coralline palaeotemperature records from offshore Bermuda and reanalysis data (Fig. 5 and Supplementary Fig. 10), illustrating its reconstructive power20,21,22,23,32. The new stalagmite SST record (Fig. 4) suggests that prior to the 18th century, SSTs were relatively stable, fluctuating within a range of 1.9 °C around a mean of 24.6 °C (compared to the 1950–2018 average of 22.7 °C30). A cooling trend lasting ~130 years began in ~1720 CE at a rate of approximately −0.27 °C per decade and ended with the onset of the post-1850 warming. The cooling period exhibits high frequency interannual oscillations with decadal-scale fluctuations of ~1.5 °C (a range of 4.6 °C around a mean of 22.8 °C), compared to fluctuations of less than 0.5 °C over the same timescale before ~1720 CE. A warming trend of ~0.16 °C per decade then starts around 1850 CE and lasts the rest of the record. The interannual variance also decreases after 1961 CE, but this is likely the product of anthropogenic impacts on the site’s hydrology (e.g., the construction of buildings and land use change above the site) and/or a shallower mixed layer depth33,34 rather than changes in climate as the annual signal across that interval is also less pronounced, despite no change in the annual climate signal.

Fig. 5: Stalagmite-based SST reconstruction compared with various other proxy reconstructions of Bermudan SSTs.
figure 5

a–f Comparison of the stalagmite Mg concentration-based record to six coralline palaeo-ocean temperature records for Bermuda20,21,22,23 and (g) a normalised winter North Atlantic Oscillation index dataset with 15-point moving average44. Note that the ref. 22 record was averaged to annual spacing from monthly to reduce noise and that axes have been inverted to match the direction of the SST axis.

Full size image

SST proxy records from four independent archives show stable, likely elevated, SSTs from the mid-1400s to the early 1700s (Fig. 5). However, a coral extension record from the region suggests that SSTs were slightly lower than the stalagmite record indicates and that a brief warming period ensued during the late 1600s (Fig. 5a and c)23. This could be an example of wind speed artificially amplifying the SST signal in the stalagmite as the period coincides with an extended negative NAO interval (Fig. 5g). Similarly, however, the stalagmite record is in better agreement than the coral extension record with the other available records, and thus other forcing and/or chronological uncertainty could explain the inconsistency. Bermudan coral extension rates have been tied to nutritional pulses associated with negative NAO winters35 and hence temperature is likely not the only control on the proxy36. Additionally, the record’s chronology is developed from a layer count which is not supported by other methods and thus could be susceptible to undercounting resulting from low growth rates/hiatuses21. Despite this discrepancy, both records agree that SSTs were uncharacteristically high during the LIA. The records then show a cooling trend lasting until the mid-1800s. The coral extension record suggests this cooling lasted until the early 1900s23. However, two δ18O records from different corals (Fig. 5d)20 support the mid-1800s trend reversal seen in the stalagmite and ref. 21 records. All records then show a strong warming trend lasting until the present. Both the reconstruction presented here and the ref. 21 coralline record show a brief cooling interval starting in the late 19th century, which lasted around two decades before warming resumed (Fig. 5a, b). The relatively cooler mean SSTs of the 19th and 20th century, when compared to those pre-1720 CE are likely explained by wind speed changes artificially increasing estimates and/or aerosol forcing in the NH reducing SSTs37,38.

Influences on reconstructed SST variability

The SST record reconstructed here exhibits three distinct intervals: (1) pre-1700 elevated, stable SSTs; (2) 1700–1850 cooling; and (3) post-1850 warming. Enhanced radiative forcing from additional greenhouse gases and rising total solar irradiance at the end of the LIA are likely to be the cause of the increased temperatures observed after 1850 CE (Fig. 6c and Supplementary Fig. 11). A comparison with NH temperature anomalies, which remained largely low before the mid-1800s (Fig. 6g), highlights that Bermudan SSTs were uncharacteristically high during the LIA, suggesting conditions influenced primarily by natural factors with minimal anthropogenic impact.

Fig. 6: Proxy reconstructions of North Atlantic variability.
figure 6

(top) Map of modern observational North Atlantic regional SSTs (2023 mean; 0.1° × 0.1° spatial resolution)90. Coloured circles on the map denote the sites of the records in the panel below. (bottom) Palaeoclimatic records for the study period. a Percentage abundance of the polar species N. pachyderma (sinistral) in a marine sediment core (OCE326-MC13) from the Northwest Atlantic39. b Chesapeake Bay reconstructed SST data with a 15-point moving average40. c The annual SST model for Bermuda with a 15-point moving average. d Mean sortable silt (\(\overline{{{{\rm{SS}}}}}\)) grain size data from three cores (red, 56JPC; orange, 48JPC; grey, MO217)10,39. e Gulf Stream transport estimates in the Florida Straits9. f Normalised winter North Atlantic Oscillation index data with 15-point moving average44. g Northern Hemisphere temperature anomaly reconstruction with 95% confidence interval for proxy-derived estimates57. The grey shaded area denotes the temporal extent of the LIA.

Full size image

Higher latitude palaeotemperature reconstructions exhibit the opposite trends to those seen in Bermuda. Neogloboquadrina pachyderma abundances in a sediment core from the northwest Atlantic39 suggest a warming trend beginning in the 18th century (Fig. 6a). Similarly, a temperature record from Chesapeake Bay using the Mg/Ca ratios of ostracodes40 exhibits cooler SSTs during the LIA followed by an increase starting in the mid-1700s (Fig. 6b). However, our reconstruction from Bermuda (Fig. 6c) suggests regional ocean temperatures were high during the oldest part of the record (1449–1720 CE) until a cooling trend ensued around 1720 CE, lasting approximately 130 years. The opposing temperature trends in the N. pachyderma and Mg/Ca records, compared to the Bermuda reconstruction, suggest an antiphase thermal gradient between the records, whereby an increase in SSTs at lower latitudes leads to a decrease at higher latitudes.

Due to the extended cold intervals, the LIA is associated with expanded sea ice extent, which would have disrupted North Atlantic ocean currents by changing temperature and salinity profiles26. Two mean sortable silt (\(\overline{{{{\rm{SS}}}}}\)) records from sediment cores off Cape Hatteras (Fig. 6d) suggest that the DWBC was stronger during the LIA before weakened due to enhanced freshwater fluxes39. Similarly, a \(\overline{{{{\rm{SS}}}}}\) record from a core off the southeastern Grand Banks (Fig. 6d) indicates that the LC was strongest during the LIA (over the last 1.45 ka), likely due to elevated freshwater transport from the Arctic10. A stronger DWBC and LC would have increased cold water flow southward into the western North Atlantic Ocean, having consequences for wider ocean circulation patterns. Sediment cores near the Dry Tortugas and Great Bahama Bank (Fig. 6e) are thought to have captured Florida Current transport over the last millennium9. The reconstruction suggests that the GS was systematically weaker during the LIA by approximately 10%. Modelling studies (e.g., refs. 41,42) show that both stronger cold, high latitude currents and weaker GS flow cause the GS to move south and reduce northward heat transport11. As Bermudan SSTs are controlled by GS dynamics, a southerly GS path could increase local SSTs and lead to reduced SSTs at higher latitudes. Therefore, the antiphase relationship could be explained by GS path migration during the LIA as a result of changing ocean circulatory patterns. Warming in several cores south of Newfoundland are thought to reflect northward slope water current migration during the LIA43, but our SST record indicates a broader antiphase system suggesting the northward ocean current shift was more widespread.

The high SSTs in our reconstruction are coeval with the persistent negative NAO index during the LIA observed in many NAO reconstructions44,45,46,47,48. For comparison the ref. 44 winter NAO reconstruction (Fig. 6f) is shown, selected for its wide spatial coverage using numerous sites, high temporal resolution, and extensive use in the literature. The NAO index is defined by the normalised pressure difference between the Icelandic Low and the Azores High and modulates mid-latitude westerly strength as well as heat, freshwater, and momentum fluxes in the North Atlantic (e.g., ref. 49). Because a negative NAO index is associated with decreased westerly wind strength, the high reconstructed SSTs pre-1700 in Bermuda may be attributable to the proxy’s wind speed connection capturing changes in atmospheric phenomena. A weakening of the westerlies at this time could have reduced sea spray over the site and thus led to less Mg incorporation into the stalagmite. These weaker winds are evident in the relatively low implied seawater contribution from the Mg and Sr concentrations at the time (~4.47% compared to the mean ~5.81%; Supplementary Fig. 7). However, considering that a transition in the 18th century can also be seen in sediment core-derived temperature and flow speed reconstructions39,40, there is likely a direct temperature effect in addition to any changes in wind speed. Studies have also found a link between the GS and the NAO50,51. During a more negative NAO phase, decreased heat loss, higher precipitation, and weaker winds reduce convection in the Labrador Sea, corresponding to a weaker AMOC and in turn a more southerly GS path49,52. This is evident in current GS path migration (1993–2016), which may reflect an increasingly negative NAO (e.g., refs. 51,53). Thus, the GS likely did migrate at this time, but the potential for reduced wind speed over Bermuda to exaggerate the reconstructed SST signal complicates the quantification of the exact southward positional extent of the GS movement. Furthermore, higher frequency variability may also be influenced by other atmospheric phenomena, but the significant agreement with oceanographic data suggests that either oceanic forcing is the salient factor or that oceanic and atmospheric forcing are heavily coupled.

Speleothem Mg concentrations are sensitive to the effects of prior calcite precipitation (PCP) and are often used as a proxy for rainfall amount (e.g., refs. 54,55). However, areas with evenly distributed rainfall throughout the year usually do not exhibit a strong PCP effect54,56. Therefore, a change in cave processes (such as the degree of water-rock interactions or PCP) is unlikely to be the primary driver of Mg concentration variability as they are controlled by broader climatic factors, and only the period after 1961 exhibits the effects of PCP following anthropogenic impacts on the cave system (Supplementary Fig. 7). For this reason, although cave processes or the amount of fluid inclusions may reduce the accuracy of the temperature estimates, the trends exhibited here are likely not exclusive to the BER-SWI-13 record and reflect broader regional trends.

Assuming our SST estimates reflect solely temperature change without amplification due to wind speed, the 2–3 °C temperature change over 130 years suggests a moderately fast rate of GS migration when considering the relatively shallow nearby SST gradient in the modern North Atlantic (Fig. 6), implying that GS position is reasonably sensitive to forcing. Observational data tracking the current northward GS migration detected a shift in the GS north wall of ~1.5° in 52 years (1965-2017) between 50°W and 40°W6. Therefore, with the continuation of current trends in the GS and AMOC, the sensitivity of regional temperatures to GS path drift could have serious ramifications for future regional climates, ecosystems, and extreme weather events. Furthermore, the mid-1700s transition seen here agrees with similar trends seen in previous AMOC proxy records (e.g., ref. 39, Supplementary Fig. 11), potentially supporting an earlier onset of AMOC weakening.

Our study reconstructed regional SST using a stalagmite from Bermuda via sea spray contributions to dripwater. A Mg cycle count, modelled to a radiocarbon chronology, yielded a high-precision age model and a monthly-resolved Mg record spanning over five centuries (1449–2013 CE). Owing to the Mg record’s significant agreement with SST records but with the opposite directionality to what is expected from Mg’s partition coefficient into calcite, the record likely captures SST variability via a connection to wind speed. Palaeo-oceanographic temperature estimates were derived from modelling the Mg concentration record to a previously published coralline proxy SST record21. The resultant record indicates temperatures before the early 1700s were consistently high, contrary to more Northerly regional reconstructions and the NH average39,40,57. We propose this antiphase thermal gradient suggests that the GS migrated northwards towards the end of the LIA following a prior period characterised by a more southerly GS path. Flow speed reconstructions for the GS, DWBC and LC support this interpretation9,10,39, providing evidence of weakening southward currents in the 18th century from strong conditions during the LIA. Further research should focus on quantifying the rates of positional change seen here, to gain a better understanding of GS sensitivity to additional forcing as well as its subsequent manifestations and effects. These results provide evidence of a moderately fast natural GS migration during the LIA. Previous research has proposed changes in ocean circulation towards the end of the LIA (e.g., ref. 39), and our high-resolution SST reconstruction from Bermuda reinforces this concept.

Methods

Stalagmite BER-SWI-13

BER-SWI-13 is a 192 mm long stalagmite that grew in Leamington Cave, Bermuda (Supplementary Fig. 1a, b). The stalagmite was selected for its favourable internal structure and active drip hydrology, which is responsive to rain events58 and plots as a near-ideal seasonal flow59. Because cave monitoring data revealed a same-day drip rate response to rainfall events (Supplementary Fig. 2)58, the residence time of the seawater signal is sufficiently low to accurately preserve hydrology-induced seasonality at a monthly-resolution59. The cave pool’s connection to the ocean promotes daily cave ventilation by tidal flushing, suggesting the stalagmite never experienced condensation corrosion due to CO2 buildup58. Additionally, monitoring data shows daily cave temperature is positively correlated to daily mean SST (r = 0.83, p < 0.01) (Supplementary Fig. 3)58.

U-series and radiocarbon dating

Reconnaissance CT scanning determined the position of stalagmite BER-SWI-13’s central growth axis and informed its sectioning. Cut sections were polished and then cleaned in an ultrasonic bath of deionised water (see ref. 58 for full methods). Visual examination after sectioning showed no signs of growth hiatuses (Supplementary Fig. 1b). Sixteen ~400 mg carbonate powder samples (Supplementary Data 1) were milled along the stalagmite growth axis. These were dated using multi-collector inductively coupled plasma mass spectrometric (MC-ICPMS) U-Th methods at the facilities of Bristol Isotope Group, University of Bristol, and University of St. Andrews Isotope Geochemistry (STAiG). U and Th separation and measurement protocols described in ref. 60 were followed, where CRM U112a and an inhouse Th standard (Teddii) was used for bracketing the samples and correcting for instrumental drift in mass bias, SEM yield. We adopt relevant decay constants from ref. 61. U-series data was analysed using the IsoplotR software62.

U-Th ages for material that is both young (i.e., <1 ka) and has low 238U/232Th ratios are poorly constrained because there is no prior knowledge of the 230Th/232Th ratios at the time of calcite precipitation. Because of this, we attempted to assess the initial Th ratio using isochron methods for two of the milled samples (BER-SWI-UTh-13 and −15) by performing replicate analysis (~100 mg sub-samples). However, there was insufficient 238U/232Th ratio variation in the two sets of replicate analysis to determine useful estimates of the initial 230Th/232Th component.

Most researchers use an estimate of initial Th assuming it is equivalent to bulk Earth silicate composition (238U/232Th = 0.8 ± 0.4) as a starting point. However, there are many examples of elevated values of this ratio where either carbonate dust with high U/Th or hydrogenous elements in cave drip waters have probably influenced the isotopic composition: see review by Richards and Dorale63 and recent papers by refs. 64,65. Other speleothem samples from the North Atlantic and Caribbean region, including the Bahamas, Yucatan, and Puerto Rico, have been similarly affected66,67,68,69,70,71.

In the absence of sufficiently robust estimates of the initial Th ratio via isochron methods, we considered the principle of stratigraphic constraint (see ref. 72) and adopted an elevated value ((238U/232Th)initial = 5 ± 2.5) that eliminates the major age inversion at the top of the sample (Supplementary Fig. 4) that is evident when using the bulk Earth initial Th ratios. Using this elevated value also shifts the basal age (BER-SWI-UTh-16) younger and would suggest a growth rate for the oldest material that is closer to that exhibited for the rest of the speleothem. One anomalous age remains for the sample (BER-SWI-UTh-5), which is likely to have been affected by post-depositional alteration.

Additionally, twenty-five 8–9 mg carbonate powder samples were milled for 14C analysis (Supplementary Data 2) using a semi-automatic high-precision drill (Sherline 5400 Deluxe) at ETH Zurich. The 14C samples were milled in increments of 5 mm for the top 50 mm, and 10 mm for the remaining 140 mm58. The top ~0.1 mm was discarded for each 14C sample. Potential contamination during sub-sampling was minimised by cleaning all tools with methanol and drying with compressed air in between samples. Carbonate samples were prepared for accelerator mass spectrometry analysis by graphitisation at the Laboratory for Ion Beam Physics of ETH Zurich. The samples were processed on an automatic graphitisation system coupled to a carbonate-handling system (CHS-AGE, Ionplus, Switzerland) by acidification to CO2 using 2 mL of 85% H3PO4 and followed by graphitisation using excess H2 over Fe catalyst. Radiocarbon contents were measured on a MICADAS accelerator mass spectrometer (Ionplus) following standard protocols for quality and blank assessment: oxalic acid II (NIST SRM 4990 C) was used as the normalising standard, measured to a precision of better than 2‰. We used IAEA-C1 as a blank and IAEA-C2 and a modern coral standard as secondary standards. The procedural blank was established using a 14C-free stalagmite (MAW-1, ~170 kyr old).

Trace element analysis

Trace element concentrations were measured using the prototype RESOlution M-50 excimer (193 nm) laser-ablation system with a two-volume laser-ablation cell coupled to an Agilent 7500ce/cs quadrupole ICPMS at Royal Holloway University, London73,74,75.

Ablation tracks were pre-ablated to remove any superficial contamination and measured using a 140 by 10 μm rectangular laser slit across 50 mm sectioned blocks on the opposite half to the carbonate sampling. Sectioning was slightly off perpendicular to the growth direction so that overlaps in transects could compensate for any missing data between sections. A 15 Hz laser repetition rate and a laser fluence of ~4 J/cm2 with a stage scan speed of 10 µm s−1 of the LA cell were used during the continuous main track measurement73. Speleothem analyses were bracketed by analyses of NIST 612, NIST 610 for quantification using43 Ca as internal standard76, and MACS3 standards for accuracy control (see ref. 77).

The resultant data were then reduced using the lolite software package, using NIST 610/612 standards for external standardisation78. Trace element analyses using similar techniques have achieved accuracies better than 5%74,77.

BER-SWI-13 chronology

The U-Th chronology suggests an unusual range in speleothem growth rate for a humid subtropical cave environment (65 to 1636 μm yr−1; where the 90th percentile of annual layers in a global review12 is 404 μm yr−1). This variability is most pronounced between depths from top of 26 and 35 mm where the growth rate implied from U-Th dating suddenly rises from 233 μm yr−1 to 1636 μm yr−1 for 5.5 years before returning to 196 μm yr−1, despite no hiatuses being observed. Additionally, the carbonate samples returned consistently low U concentrations of 83–136 ng/g, except for two samples that are 299 and 362 ng/g. The low U concentrations and unrealistically high variations in growth rate suggest that a chronology based solely on U-Th dating would yield less accurate results than radiocarbon methods after dead carbon fraction correction. For this reason, we use the U-Th chronology only as confirmation when developing an alternative chronology based on radiocarbon dating and annual Mg cycle counting. First, we developed an ‘average’ chronological model using the radiocarbon-based technique presented in ref. 79 and refined in ref. 80. The model aligns the stalagmite radiocarbon data with known atmospheric radiocarbon levels at the time of carbonate deposition, thus correcting for DCF effects. DCF is assumed to be stationary (i.e., fluctuating around a constant mean value) and trends are detected using hydrology connected proxies (Mg/Ca and δ13C), but small variations around a long-term mean are allowed thus permitting the use of this technique on the BER-SWI-13 record. The chronology produced this way is anchored to a point of known age, which in our case is given by the clear increase in 14C indicating the rise of the atmospheric bomb spike at 1955 CE (Fig. 2b). The bomb spike was identified using the anomalous rise in raw 14C ages at a depth of approximately 8 mm, which can only be related to the rise in atmospheric 14C through nuclear bomb testing in the 1950–1960s, as it results in much too young 14C ages to be reflecting natural variability. To precisely constrain the depth of the bomb spike inflection point (i.e., the point where the bomb spike begins, our chronological marker), we extrapolated the trends of the data points immediately before (4) and after (3) the visible bomb spike rise. The stalagmite collection year (2013 CE) provides a second chronological anchor. However, owing to possible localised anthropogenic influences on the stalagmite’s drip hydrology (e.g., land use change above the cave related to the development of the cave for tourism) starting ~1961 CE, we do not consider data post-1961 in our interpretations (Supplementary Fig. 5).

First the radiocarbon age model defines an initial long-term average growth rate by computing a best fit through all the stalagmite 14C activity measurements, with an assumed constant atmospheric 14C. This enables each sample to be assigned an approximate age. The model then iterates through adjusting the growth rate in line with atmospheric 14C values from an IntCal13 calibration curve81 and updating the assigned ages accordingly until the growth rate term converges. Thus, one of the underlying assumptions of the model is that the stalagmite growth rate remains relatively constant and thus the output chronology is a near perfect straight line. However, the error in this assumption is propagated with the depth uncertainty incurred by sampling. Therefore, most of the uncertainty lies with the growth rate estimate and hence leads to very conservative errors that increase with distance from the anchor point and follow near straight lines. This technique is similar to methods used in other geochronological fields, such as 210Pb dating82 or 230Th excess dating83 of sediments and has been used to accurately date stalagmites in previous studies (e.g., refs. 84,85).

The LA-ICPMS-derived Mg profiles exhibit well-developed cyclicity through much of the record. A comparison between the number of Mg cycles in the record and the U-Th and radiocarbon chronologies strongly suggests that the Mg cycles are annual. We conducted multiple Mg cycle counts with different observers to test their reproducibility, and constructed a chronology based on the annual Mg cycles (Fig. 2). The post-1961 loss of cyclicity prevents counting cycles relative to the collection year and instead cycles were counted relative to the bomb spike depth, which was adjusted to match a local minimum to obtain a sub-annually resolved chronological anchor. This cyclicity is ambiguous for some sections of the record; thus, in the final cycle count these sections were modelled using the radiocarbon chronology mean growth rate (Supplementary Fig. 6).

Mg concentrations in BER-SWI-13 are significantly inversely correlated (r = −0.56, p ≪ 0.0001, n = 217) with a previously developed coral Sr/Ca-derived proxy SST record from Bermuda21, which is contrary to that is expected based on the temperature dependency of DMg. Instead, Mg concentrations appear to be an indirect proxy for SST via sea spray contributions to rainfall (Fig. 3). A clear inverse relationship exists between SST and wind speed (Fig. 3). Colder temperatures are associated with higher wind speeds, which gives rise to higher sea spray contributions to rainfall and vadose dripwater. Seawater has a much higher Mg concentration than normal meteoric water and this influences the concentrations in cave dripwaters (and secondary calcite precipitates such as stalagmites). Additionally, because the Mg and Sr data do not plot along a PCP vector (Supplementary Fig. 7), PCP is unlikely to be a dominant control on the record. The post-1961 CE data appear to show evidence of PCP supporting our decision to exclude these data from our interpretations due to alterations to the stalagmite’s drip hydrology. However, the Mg and Sr data do plot along a mixing curve between the concentrations of the bedrock and marine aerosols (Supplementary Fig. 7), supporting the perspective that sea spray (in turn controlled by SST) was the dominant control on Mg and Sr concentrations.

By virtue of the inverse relationship between speleothem Mg and SST, the Mg cycle count was tuned intra-annually so that the Mg concentration minima were assigned a decimal date corresponding to the 15th of August (midway between the months with the lowest mean wind speed) and maxima (determined by the broad shape of the cycle to avoid individual weather events) were assigned a date corresponding to January 1st (start of month with the highest mean wind speed). The resultant high-precision final cycle count (1449–2013 CE) is consistent with the radiocarbon chronology, with a nearly identical mean growth rate (Fig. 2a), as well as high intra-annual precision. The final cycle count yields markedly younger values than the U-Th data until depths from top greater than ~140 mm, when the U-Th growth rate rapidly increases, whilst also having a more consistent annual growth rate (range in annual cycle length—final cycle count: 865.8 μm yr−1; U-Th: 1571.8 μm yr−1). The COnstructing Proxy Records from Age models (COPRA) algorithm86 was used to develop an age-depth model with cycle count annual maxima and minima as inputs. The resultant COPRA model incorporated 2000 Monte Carlo simulations and piecewise cubic Hermite interpolating polynomial interpolation to assign precise ages to every data point. The COPRA algorithm did not identify any hiatuses in the record. The complete dataset consists of 18,471 data points and creates a record of sub-monthly resolution with an average of 32.8 values per year.

Mg-SST calibration

To derive SST estimates, the speleothem Mg record was interpolated to monthly spacing and passed through a five-point Savitzky–Golay filter to remove the influence of weather extrema and/or areas with anomalously high amounts of fluid inclusion before averaging into annual values. The correlation between the annual stalagmite Mg concentration record and the annual coralline SST record21 over the ~200 years of overlap was then optimised by shifting the SST record within dating uncertainties using dynamic time warping (DTW) with a maximum shift term equal to 25 years to avoid overfitting (see Supplementary Fig. 8). The DTW applied offset was negative (mean of −13.8 years) before 1866 CE, after which it changes to positive (mean of 0.6 years), likely due to the change in trend in both records around this time. It is thought that undercounting is most prevalent in their record before the mid-1800s when growth rates were extremely slow21, supporting the negative DTW offset before 1866. Given the difference in magnitude between the Mg (range: 1759–4769 ppm; mean: 3074 ppm; standard deviation: 632 ppm) and SST (range: 20.3–25.2 °C; mean: 22.7 °C; standard deviation: 1.1 °C) values, both series were standardised prior to regression to ensure numerical stability. An initial ordinary least squares linear regression returned autocorrelated residuals (Durbin–Watson statistic: 0.906). Therefore, both series were detrended for subsequent modelling using a five-point centred moving average, with the window size optimised based on residual autocorrelation. A linear regression of both standardised residual series yields the following equation:

$$\begin{array}{c}{\epsilon }_{{{{\rm{SST}}}}}=-0.0004\left(\pm 0.0448\right)-0.6360(\pm 0.1025)\cdot {\epsilon }_{{{{\rm{Mg}}}}}\\ (2\sigma ,\,95 \% {{{\rm{conf}}}}.,\,{R}^{2}=0.46,{p}\, \ll \, 0.0001,\,{{{\rm{RMSE}}}}=0.3027)\end{array}$$
(1)

To return accurate SST estimates, the inverse of trend of the predictor series was added to the series obtained from the regression, and the result was then transformed back to non-standardized form using the mean and standard deviation of the target series. Thus, the full equation can be written as:

$${{{{\rm{SST}}}}}_{{{{\rm{final}}}}}=\left(\left({\beta }_{0}+{\beta }_{1}\cdot {\epsilon }_{{{{\rm{Mg}}}}}\right)+-{{{{\rm{Mg}}}}}_{{{{\rm{trend}}}}}\right)\cdot {\sigma }_{{{{\rm{CoralSST}}}}}+{\mu }_{{{{\rm{CoralSST}}}}}$$
(2)

where \({\beta }_{0}\) and \({\beta }_{1}\) are the coefficients of the linear regression of the processed data; \({\epsilon }_{{{{\rm{Mg}}}}}\) is the standardised, residual Mg after detrending; \({{{{\rm{Mg}}}}}_{{{{\rm{trend}}}}}\) is the smoothed standardised Mg series; and \({\sigma }_{{{{\rm{CoralSST}}}}}\) and \({\mu }_{{{{\rm{CoralSST}}}}}\) are the standard deviation and mean of the warped SST data, respectively. Substituting in the derived coefficients and propagating the errors to derive an equation in non-standardised space yields:

$$\begin{array}{c}{{{{\rm{SST}}}}}_{{{{\rm{final}}}}}=27.8214 (\pm 0.7783)-0.0011(\pm 0.0002)\cdot {{{\rm{Mg}}}}-0.0006(\pm 0.0002)\cdot {{{{\rm{Mg}}}}}_{{{{\rm{trend}}}}}\\ (2\sigma ,95 \% \, {{{\rm{conf}}}}.,{R}^{2}=0.77,\,p \, \ll \, 0.0001,{{{\rm{RMSE}}}}={0.5070\, }^\circ {{{\rm{C}}}})\end{array}$$
(3)

The modelled stalagmite SST record (Supplementary Data 3) has less variance than the coralline record (Fig. 4a) despite its near-ideal relationship, likely because of aquifer mixing smoothing the signal (potentially the mixing of diffuse and concentrated flow paths). The signal-dampening effect that the strong correlation between cave temperature and SST would have via the temperature dependency of Mg fractionation may also be responsible for this reduced variance. For this reason, the record’s ability to capture variations in seasonal signals is hindered and so we interpret the record at annual resolution. Additionally, the regression’s predictive power was evaluated using split-period calibration/verification tests in both directions (see Supplementary Fig. 9). The regression passed these tests and thus yielded both a monthly and an annual SST record spanning over five centuries years, making it longer and higher resolution than many previous SST records from the region (Fig. 4).