## Main

The Asian monsoon provides fresh water for a densely populated region and plays a major role in global climate as a conveyor of latent and sensible heat1. Changes in the cross-equatorial ocean energy transport driven by the Atlantic Meridional Overturning Circulation (AMOC) play a fundamental role in the monsoon system through its impact on the interhemispheric temperature gradient and the latitudinal position of the intertropical convergence zone (ITCZ)2. A weaker AMOC also decreases sub-polar temperatures, which affect the course and strength of the westerly winds that interfere with southerly monsoon winds3. Palaeo-climate records offer a unique opportunity to study the links between changes in AMOC strength and the Asian monsoon system during periods of the past when the AMOC was substantially weaker or even completely collapsed4,5, but there is no general consensus about the dominant mechanisms. Understanding the response of the monsoon system to rapid changes in the strength of the AMOC is particularly relevant in the context of recent observational evidence suggesting a weakening of the AMOC over the past decades6,7,8,9.

The oxygen isotopic composition of speleothem calcite from East Asia suggests that rapid changes in the strength of the AMOC and the Asian monsoon were indeed tightly linked3,10,11,12,13,14,15. During the coldest parts of the last glacial, that is, during Heinrich events, a weakening or collapse of the AMOC4 was systematically associated with higher speleothem δ18OCc (refs. 12,13,16). Classically, higher speleothem δ18OCc values have been interpreted as ‘weak monsoon intervals’ and explained by a lower proportion of low δ18OP (Indian summer monsoon (ISM) and East Asian summer monsoon (EASM)) rainfall in annual totals11,16,17 and or decreased upstream depletion of 18O in either the ISM or East Asian monsoon (EAM) regions11,14,17,18, and vice versa for low speleothem δ18OCc values, which are interpreted as ‘strong monsoon intervals’. However, with the available data, it is not possible to distinguish how the different components of the Asian monsoon system (for example, temperature gradients, the position of the ITCZ, the length of the monsoon season or the intensity of the ISM or EASM) are affected by changes in the strength of the AMOC.

## Deconvolving calcite δ18O

Part of the challenge arises from the fact that δ18OCc depends on not only mean annual drip water δ18O but also temperature-controlled isotope fractionation between water and calcite19, and potential kinetic isotope effects20. Here, we use an innovative approach to deconvolve the effects of kinetic isotope fractionation, temperature and drip water δ18O by combining: (i) the TEX86 speleothem thermometer to reconstruct changes in cave air temperatures21,22, (ii) the dual clumped isotope (Δ47 and Δ48) thermometer that allows to assess both changes in cave temperature and the extent of kinetic isotope fractionation23,24 and (iii) the hydrogen and oxygen isotope composition of speleothem fluid inclusions, that is, δ2HFl and δ18OFl, as a proxy for the isotopic composition of precipitation, that is, δ2HP and δ18OP (refs. 25,26). We applied these techniques to two south-west Chinese speleothems from Jiangjun Cave spanning the penultimate deglaciation27 (Fig. 1 and Extended Data Figs. 1 and 2), a time period that was characterized by two successive episodes of vast freshwater discharge into the North Atlantic28 (meltwater pulses (MWPs)) that caused a weakening29 (MWP 2A) and full collapse (MWP 2B) of the AMOC4.

The present-day local climate at Jiangjun Cave is characterized by warm/wet summers and cold/dry winters. The summer monsoon peaks from May to September30, contributing ~73% to annual rainfall (~1,320 mm per year), and the mean annual air temperature is ~23 °C. Wet-season (summer) precipitation in the region of Jiangjun correlates positively with precipitation in north-east India, which, in turn, depends on the strength of the ISM31. Consequently, Jiangjun δ18OP is about 10‰ lower in July–August compared with January–March (Extended Data Fig. 3).

## Structure of Termination II

During the penultimate deglaciation, Jiangjun δ18OCc (Supplementary Data Table 1a) shows trends similar to those found in Sanbao Cave δ18OCc (and other EAM records; Extended Data Fig. 4). From 151 to 130 ka, two positive shifts are observed in all EAM records, superimposed on a gradual trend towards higher δ18OCc (Fig. 2a). The first occurred around ~139 ka (1‰), synchronous with MWP 2A. The second shift occurred at ~136 ka (2‰; Fig. 2), within dating error of the onset of MWP 2B28 (135 ± 1 ka) and Heinrich Stadial (HS) 11. After this event, δ18OCc remains stable with the exception of the small Termination II (TII) interstadial peak at ~134 ka. At the end of the glacial termination (~130 ka), Jiangjun δ18OCc shows a 7‰ shift towards lower δ18OCc, whereas Sanbao shifts by only 4‰. The 7‰ shift at Jiangjun replicates the δ18OCc change at the nearby Xiaobailong Cave31, while the 4‰ shift at Sanbao is consistent with those found in Dongge and Yangkou caves located north-east of Jiangjun, suggesting a greater influence of ISM precipitation in south-west China (Fig. 1 and Extended Data Fig. 4).

The structure of the Jiangjun δ18OCc record is similar to our TEX86 temperatures. We observe a long-term cooling from 144 to 135 ka of about 3 °C. Superimposed on this cooling trend, an additional cooling of around 1 °C occurs during MWP 2A (Fig. 2b and Supplementary Data Table 1b). After MWP 2A, temperatures rise by around 1 °C, coinciding with an increase in δ18OCc, and subsequently decrease by ~2 °C at the beginning of HS11. During the last part of the glacial termination, after ~130 ka, TEX86 temperatures increase abruptly by 4 °C. These temperatures are consistent within uncertainty with the temperature estimates based on dual clumped isotope data (Fig. 2b and Supplementary Data Tables 1c–f). In addition, our dual clumped isotope data show that calcite precipitated indistinguishable from isotope equilibrium (Extended Data Fig. 5). Applying a temperature-dependent water–calcite isotope fractionation of 0.21‰/1 °C (ref. 32), we estimate the contribution of temperature to the changes observed in the δ18OCc record. This calculation reveals that cave air temperature can only explain a small fraction of the variability observed in the δ18OCc record, for example, only about 0.63‰ of the near 4‰ change in δ18OCc from 144 to 135 ka, or about 0.84‰ of the 7‰ change observed during the last part of the deglaciation after ~130 ka. Therefore, our data demonstrate that changes in δ18OCc are mainly dominated by variations in annual mean parent water δ18O (Extended Data Fig. 6).

However, in contrast to δ18OCc, Jiangjun δ18OFl shows no response to MWP 2A. An abrupt shift towards more positive δ18OFl is only observed around 133 ka, coinciding with the peak of MWP 2B (Fig. 2 and Supplementary Data Table 1g). One potential explanation for these differences could be that the fluid inclusions are not well preserved and do not represent palaeo-rainfall. A key criterion to assess this is that, if unaltered, the fluid inclusion isotope data should plot close to the local and global meteoric water line33. For Jiangjun Cave, the most representative Global Network of Isotopes in Precipitation data are from Kunming, which are highly consistent with the fluid inclusion data (Fig. 3a). Furthermore, Jiangjun’s interglacial δ18OFl values are very similar to values calculated for parent drip water based on Jiangjun’s TEX86 and dual clumped isotope temperatures and the δ18O temperature calibration from ref. 34, which bolsters the quality of our dataset (Extended Data Fig. 6). The fluid inclusions are thus well preserved and represent the isotopic composition of the parent water.

Closer examination of our dataset revealed a ~1–3‰ offset between parent water δ18O (calculated by subtracting the temperature contribution from calcite δ18O) and δ18OFl from ~139 to ~128 ka, with δ18OFl being lower than calculated parent water values (Extended Data Fig. 6). Since this offset occurs in the overlap zone of the two speleothems where δ18OCc, δ18OFl and TEX86 proxy records replicate well, we do not believe that this reflects kinetic fractionation or diagenetic processes. We propose instead that, in this interval, the difference arises because δ18OFl is seasonally biased, mainly capturing the δ18O of intense ISM rainfall events, while δ18OCc captures the annually averaged signal (Supplementary Information and Fig. S1). Our microscale analysis of the speleothem fabric and δ18OCc indeed reveal an inhomogeneous distribution of fluid inclusions during glacial times, characterized by an alternation of fluid inclusion-poor layers (associated with higher δ18OCc) and fluid inclusion-rich layers (associated with lower δ18OCc) (Fig. 3b). The low δ18OCc values found in the fluid inclusion-rich laminae are associated with the ISM season (which is characterized by low δ18OP). These data suggest that, during the driest part of the glacial period, characterized by low speleothem growth rates (Fig. 2d) and a less developed soil on top of the cave, strong monsoon rainfall events infiltrated rapidly and were captured in fluid inclusion-rich calcite laminae. In the dry season, fluid inclusion-poor laminae were formed, causing a distinct seasonal bias in fluid inclusion incorporation, while the calcite grew in both wet and dry seasons35 (Fig. 3b). This effectively decouples δ18OFl and δ18OCc records, and highlights that δ18OFl predominantly captures the strength or ‘intensity’ of the ISM that integrates the rainfall history along the moisture pathway, while δ18OCc values are more likely to reflect the annually averaged hydroclimate signal.

## Implications for monsoon dynamics

Our data indicate that MWP 2A did not affect δ18OFl but did affect temperature and δ18OCc, so how can this be reconciled? MWP 2A was associated with a peak in ice-rafted debris36 (IRD) in the North Atlantic and caused a weakening of the AMOC as observed in benthic δ13C values (ref. 29) (Fig. 4). Two additional IRD peaks of similar magnitude have been observed before and after MWP 2A, also associated with benthic δ13C minima, suggesting multiple periods of weakening and reinvigoration of the AMOC during this time period. A weakening of the AMOC leads to cooling in the subpolar North Atlantic, which increases the equator-to-pole temperature gradient and triggers a southward shift of the westerly jet, as shown by climate simulations of HS1 (ref. 3). Interestingly, our speleothem TEX86 temperature record shows three cooling events (of ~1 °C) in south-west China that occur within age uncertainties of these North Atlantic IRD events (Fig. 4). However, MWP 2A, as well as the two additional IRD events, were likely too small to induce a complete shutdown of the AMOC. Consequently, ocean-to-continent temperature gradients were likely not affected to the extent that they impacted the strength of the ISM, as suggested by the relatively stable Jiangjun δ18OFl values during this period. This interpretation is supported by the high correlation (r = −0.74, P < 0.001, n = 17) between δ18OFl and independent marine reconstructions of the ISM wind-induced upper ocean stratification in the east equatorial Indian Ocean37, which suggest a weak response of the ISM wind strength to MWP 2A and surrounding MWP events (Fig. 4). Despite this apparent lack of response of the ISM intensity, a clear signal is observed in the δ18OCc records during MWP 2A and the onset of MWP 2B that cannot be explained by the observed temperature change alone, suggesting a change in annual mean δ18OP. We hypothesize that these observations can be reconciled if the timing of the northward retreat of the westerly jet in the Asian monsoon region was delayed in response to the enhanced equator-to-pole temperature gradient, shortening the monsoon season in a similar fashion as during HS1 (ref. 3). In agreement with climate simulations3 and Jiangjun speleothem growth rate (Fig. 2), this scenario would have resulted in a decrease in annual mean precipitation at Jiangjun during this interval (Fig. 2d). The consequences of a delayed northward retreat of the westerly jet would have been similar in northern India and in the Arabian Peninsula, causing also a shortening of the monsoon season and thus more positive annual mean parent water δ18O. More dominant westerlies also brought 18O-enriched water vapour from the Mediterranean to the Arabian Peninsula38, supporting our hypothesis. The small southward shift of the ITCZ as indicated by the Mulu δ18OCc record39 (Fig. 4) is in agreement with a moderate southward shift of the westerly jet. Thus, our interpretation represents a viable mechanism to explain the consistency of speleothem δ18OCc across the Asian monsoon region, as well as their apparent discrepancy with the available ISM records.

In contrast to MWP 2A, MWP 2B was associated with massive iceberg discharge events in the North Atlantic and a full shutdown of the AMOC, as shown by the kinematic AMOC tracer 231Pa/230Th4 (Fig. 4e). Our data reveal that MWP 2B was characterized by colder temperatures and a dramatic decrease in the intensity of the ISM (reflected in a 2–3‰ increase in δ18OFl; Figs. 2 and 4). These changes coincide with a maximum in planktonic foraminifera surface-to-thermocline δ18O gradients in the southernmost Bay of Bengal (Fig. 4b). The role of sea surface temperature in the Bay of Bengal during Heinrich events has not been resolved yet40, but climate model simulations of HS1 suggest that, as opposed to the Arabian Sea, sea surface temperature did not change significantly in that region (see fig. S4 in ref. 15), suggesting that the increase in the planktonic foraminifera gradients reflects a strengthening of ocean stratification in response to a weaker ISM wind strength. Our data provide empirical support for this interpretation. In addition, the timing and structure of our δ18OFl mirror the Mulu cave δ18OCc record, suggesting that the weakening of the ISM during MWP 2B was closely coupled to the migration of the ITCZ towards its southernmost position (Fig. 4a). Thus, our data suggest that the strong Northern Hemisphere cooling associated with the collapse of the AMOC during MWP 2B affected both the ocean-to-continent and the equator-to-pole temperature gradients, impacting both the intensity of the ISM and the length of the monsoon season.

These results underpin the potential of using multiple analytical techniques to quantitatively deconvolve the different components affecting speleothem δ18OCc. Our δ18OFl record suggests that the intensity of the ISM only decreased substantially with the full shutdown of the AMOC during the peak of MWP 2B, which was associated with an extreme southward migration of the ITCZ. In contrast, smaller AMOC reductions associated with earlier North Atlantic IRD events, including MWP 2A, appeared to have impacted atmospheric temperatures and reduced the length of the monsoon rain season (as suggested by δ18OCc records), but did not affect the intensity of the ISM substantially. Thus, our data indicate a distinct response of the Asian monsoon system to changes in the strength of the AMOC.

## Methods

### Stalagmite samples

The speleothems were collected from Jiangjun’s main passage, which has a length of 4 km, height of 1.5–20 m and width of 3–25 m. The speleothems were collected ~2.8 km from the main entrance. The surface above the cave is densely vegetated during the summer season. The main entrance is covered by ~80 m of host rock.

Stalagmite JJ0406 from Jiangjun Cave is 860 mm long and grew 460 mm from 139 to 124 ka, JJ0403 is 485 mm long and grew 348 mm from 147 to 134 ka. JJ0406 shows a pool-type structure at the growth axis, which we avoided when sampling for fluid inclusions and 230Th/U dating. We show a thin section of primary calcite with clear continuous lamination that corresponds to the fabric from which fluid inclusion samples were taken (Extended Data Fig. 1). The speleothem calcite δ18O records were already published27. There is a second Jiangjun Cave in south-west China41, which should not be confused with the Jiangjun Cave from this manuscript.

Uncrushed samples for fluid inclusion isotope analysis encompass ~5–20 mm of stratigraphy at the growth axis and thus represent about 50–200 years with a speleothem growth rate of 100 µm per year, or more with lower growth rates. Therefore, the fluid inclusion isotope record tracks ISM strength on multi-decadal to centennial timescales.

### 230U/Th dating and age–depth modelling

A total of 29 and 11 230Th/U datings were conducted on stalagmite JJ0406 and JJ0403, respectively (Supplementary Data Table 1a). Per sample, between 200 and 300 mg carbonate powder was used. Uranium and Th were chemically separated following a protocol adapted after ref. 42 and were measured using the technique described in ref. 43. We used U decay constants as in refs. 43,44. JJ0406 and JJ0403 contain only 20 ppb U, and 234U/238Uinitial atomic ratios are low. These stalagmites are thus rather difficult to date, which resulted in relatively large age uncertainties (Supplementary Data Table 1a). Part of the U series ages of speleothem JJ0406 were recently published27, being consistent with new U-series ages produced for this study. Age–depth modelling was performed using the StalAge algorithm45 (Extended Data Fig. 2).

### Carbonate oxygen isotope analysis

The oxygen and carbon isotope records of stalagmites JJ0403 and JJ0406 were constructed in the Xi’an Jiaotong Isotope Laboratory. Oxygen isotopes were published in ref. 27. A total of 284 subsamples were drilled from speleothem JJ0403, every 1 mm for the top 190 mm and every 2 mm below 190 mm. Meanwhile, 172 subsamples were drilled from speleothem JJ0406 every 5 mm. The stable isotope analyses were conducted in the Isotope Laboratory of Xi’an Jiaotong University on a Finnigan-MAT 253 mass spectrometer connected with a Kiel Carbonate Device IV. Oxygen isotope ratios are reported relative to the Vienna Pee Dee Belemnite (VPDB) standard. The international standard TTB1 was added to the analysis every 10–20 samples to check the reproducibility of results. Results show analytical errors (1 s.d.) of 0.06‰. The carbon and oxygen isotope data are presented in Supplementary Data Table 1a.

The crushed calcite powders were analysed at the Max Planck Institute for Chemistry in Mainz, on a Thermo Delta V mass spectrometer equipped with a GASBENCH preparation device. A CaCO3 sample (~20–50 μg) in a He-filled 12 ml exetainer vial was digested in >99% H3PO4 at 70 °C. Subsequently the CO2–He gas mixture was transported to the GASBENCH in helium carrier gas. In the GASBENCH, water vapour and various gaseous compounds were separated from the He–CO2 mixture before sending it to the mass spectrometer in nine separate peaks. Isotope values of these individual peaks were averaged and are reported as δ18O relative to VPDB. A total of 20 replicates of two in-house CaCO3 standards were analysed in each run of 55 samples. CaCO3 standard weights were chosen such that they span the entire range of sample weights of the samples. The reproducibility of these standards is typically better than 0.1‰ (1 s.d.) for δ18O. The oxygen isotope data of the crush residues are presented in Supplementary Data Table 1g.

High-resolution sampling on the thick section of sample JJ0406-690 was performed with a micromill at 20 and 40 µm resolution using a pointy-tipped drill bit. Approximately 3–5 µg of carbonate was analysed with a cold-trap method connected to the setup described above (see ref. 46 for more information). The oxygen isotope data are presented in Supplementary Data Table 1a.

### Fluid inclusion isotope analysis

Stalagmites were sampled for fluid inclusion isotope analysis (Supplementary Data Table 1g) using a handheld device (Dremel Micro), equipped with a 20–38 mm diameter diamond-covered circular saw. Depending on the water content of the calcite, we cut small calcite blocks of 0.1–1.1 g per analysis. Fluid inclusion isotope analysis was performed following the protocols adapted after ref. 26 using a Delta V advantage mass spectrometer (MS) coupled to a Conflow IV and a Thermo Conversion Element Analyser (TC-EA) from Thermo Scientific or an online preparation line coupled to a Picarro L2140i cavity ring-down spectrometer (CRDS). The calcite was crushed by a downward rotary movement in a Potsdam-type crusher47 at a temperature of 120 °C. In ref. 26, aliquots of the same samples were analysed with both techniques repeatedly. The results were consistent, proving that the techniques can be used interchangeably. In between samples, the crusher was cleaned with acetic acid and milliQ water and subsequently dried with pressurized air.

For the MS set-up, the water vapour is transported with He carrier gas to a cold trap at −105 °C. Water is trapped for 6 min, then flash-heated and transported to the TC-EA26,48. We used a standard bracketing approach with an evaluation procedure to correct for signal intensity, memory and non-linearity effects26. Memory and non-linearity effects of the set-up were assessed on a weekly basis following ref. 26. By choosing a water standard with an isotope composition that is within a few ‰ of the fluid inclusions, the memory effect was minimized. To correct for signal intensity, a series of at least three standard waters were analysed before the crush, with water amounts ranging from 0.28 to 0.12 µL. After the crush, two or three injections of 0.4 µL of standard water were injected to flush the system, which was followed by one standard water analysis, which was not used for the data evaluation because it may still be affected by a potential memory effect related to the fluid inclusions from the crushed calcite. A minimum of another two water standards with water amounts similar to the crush were analysed to assess the stability of the system and verify that it remained leak-free during and after the crush. We applied two outlier criteria: signal intensity and replication. Only samples with a water yield resulting in a peak area larger than 35 V-s for δ2H were considered for the dataset, while any samples with a lower water yield were discarded. For 22 out of 31 samples, it was possible to do two or more analyses with the TC-EA MS technique. Only one sample did not replicate within 1‰ for δ18O and was discarded (Supplementary Data Table 1g).

For the CRDS set-up, an online preparation line detailed in ref. 26 is coupled to a Picarro L2140i water isotope analyser. A stable water background of known isotope composition is generated and subtracted from the sample peak, in a similar way to refs. 49,50. For correction purposes, we analysed seven water standards of 0.3 µL before the first sample and five water standards with volumes corresponding to the water yield from the crushed calcites after the last sample. Each of the four samples measured with the CRDS technique was measured twice, replicating within 0.1‰ for δ18O (Supplementary Data Table 1g).

For fluid inclusion isotope analysis conducted on both the MS and CRDS systems, the uncertainties are based on the reproducibility reported by ref. 26. In that study, fluid inclusion isotope analysis was conducted on the same TC/EA–MS and CRDS systems that produced the fluid inclusion isotope data for this manuscript. The reproducibility is based on repeated measurements of five samples. For the IRMS system, the reproducibility is 0.4‰ for δ18O and 2‰ for δ2H (1 s.d.). For the CRDS system, the reproducibility is 0.3‰ for δ18O and 1.1‰ for δ2H (1 s.d.). These uncertainties compare well with those from several previous fluid inclusion isotope studies using very similar analytical equipment50,51,52,53. Where available, replicate analyses are shown individually in the data to indicate the extent to which replication remains within the reproducibility uncertainty. All fluid inclusion δ18O and δ2H data were normalized to the Vienna Standard Mean Ocean Water–Standard Light Antarctic Precipitation scale by direct comparison with in-house water standards that were directly calibrated against the international standards Vienna Standard Mean Ocean Water and Standard Light Antarctic Precipitation 2.

### Dual (Δ47, Δ48) clumped isotope analysis of carbonates

Dual clumped isotope analysis of carbonates (Supplementary Data Table 1c–f) was performed at Goethe University Frankfurt using the analytical set-up described in ref. 23. Δ47 and Δ48 data are reported on the Carbon Dioxide Equilibrium Scale at a reaction temperature of 90 °C (CDES 90). For data processing, we followed the CDES 90 correction protocol detailed in ref. 24. It includes (1) correction of m/z 47, 48 and 49 raw intensities for negative backgrounds using scaling factors and intensities measured on the m/z 47.5 cup, (2) correction for scale compression using CO2 gases equilibrated at 25 °C and 1000 °C, respectively, and (3) correction for subtle long-term variations in Δ47 values of reference carbonates ETH 1 and ETH 2. The two samples (JJ0406-312 and JJ0406-690) were analysed between May and August 2019, that is, covering the same period as session 2 in ref. 24. Equilibrated gas data, reference carbonate data, empirically determined scaling factors, empirical transfer functions and temporal Δ47 offset functions relevant for sample data correction are, therefore, identical to those reported for session 2 in ref. 24. Background corrected raw data (versus working gas) and final Δ47 (CDES, 90 °C) and Δ48 (CDES, 90 °C) values of equilibrated gases, reference carbonates and samples are provided in Supplementary Data Table 1d–f. Δ47 (CDES, 90 °C) and Δ48 (CDES, 90 °C) values obtained for the two samples are (within errors of 2 s.e.) indistinguishable from the equilibrium Δ47 (CDES, 90 °C) versus Δ48 (CDES, 90 °C) relationship obtained in ref. 24. Temperatures of formation and associated uncertainties were, therefore, calculated from measured Δ47 (CDES, 90 °C) values, the equilibrium Δ47 (CDES, 90 °C) versus T relationship provided in ref. 24 and 2 s.e. values. All 2 s.e. values are error propagated, considering both the long-term repeatability (2 s.d.) for Δ48 (CDES, 90 °C) and Δ47 (CDES, 90 °C) measurements of 69.9 and 17.9 ppm, respectively, and the 95% uncertainties associated with the Δ48 (CDES, 90 °C) versus T and Δ47 (CDES, 90 °C) versus T calibration regression lines24.

### Analysis of GDGTs

The procedure to analyse the distribution of glycerol dibiphytanyl glycerol tetraether (GDGT) lipids in the stalagmites from Jiangjun Cave was based on an adaptation of the methods proposed by ref. 22. GDGT measurements (Supplementary Data Table 1b) were performed in the same samples used for fluid inclusion measurements, to maximize the comparability of temperature estimates with calcite and fluid inclusion δ18O. After the fluid inclusion measurement, 0.55–1.54 g pulverized material was weighed in a 60 mL cylindrical glass vial. Calcite was subsequently dissolved in 20 mL of 6 N HCl and digested on a heating plate at 100 °C for 2 h. After cooling, the digested sample was extracted three times using 20 mL dichloromethane (DCM) in each extraction. The DCM fraction was collected in a 60 mL vial using a 50 mL separation funnel. After the third extraction, 40 µL of C46 GDGT internal standard was added to the extract for GDGT quantification purposes54. The solvent was evaporated to complete dryness under nitrogen gas on a heating plate at 45 °C. To remove potential traces of acid and other impurities, the extract was re-dissolved in a mixture of DCM–methanol and passed through a 5% deactivated silica column. The eluate was collected in 4 mL vials, dried, and filtered into a 1 mL vial through a 0.2 µm polytetrafluoroethylene membrane filter using a 1.8% mixture of isopropanol in n-hexane. The filtered extracts were analysed using Agilent 1260 Infinity high-performance liquid chromatography (HPLC) coupled to an Agilent 6130 single-quadrupole MS, using a slight adaptation of the method proposed by ref. 55, as described in ref. 56.

Cave temperature estimates were obtained by using the recent calibration proposed by ref. 21

$${{{\mathrm{Cave}}}}\;{{{\mathrm{air}}}}\;{{{{T = - 7}}}}{{{\mathrm{.34 + 34}}}}{{{\mathrm{.64}}}} \, {{{\mathrm{TEX}}}}_{{{{\mathrm{86}}}}}$$

The external analytical precision of the method was estimated by analysing a speleothem standard sample prepared by homogenizing around 1 kg of flowstone material from Scladina Cave, Belgium. This standard was analysed at least in duplicate in each batch of samples analysed, obtaining an average cave T of 14.57 °C with a standard deviation of 0.28 °C (1 s.d., n = 38). Uncertainties shown in Fig. 2 are 2 s.d. based on the external analytical precision derived from repeated measurements of the speleothem standard (0.56 °C). Due to the limited sample availability, full replicate analysis was only possible in one sample. The s.d. obtained (0.20 °C) is consistent with our estimate from our long-term in-house standard. Calibration uncertainty for speleothem TEX86 is ~1 °C (2 s.d.) in the temperature range of the samples presented in this study21. However, it is important to note that calibration uncertainties mainly affect absolute values not sample-to-sample variability.

The potential effect of crushing and/or heating (to 120 °C) during the fluid inclusion measurement and the cleaning procedure of the crusher on estimated GDGT temperatures was evaluated by performing the fluid inclusion measurement protocol on three samples of our standard speleothem sample from Scladina Cave. The analysis of GDGTs after the fluid inclusion treatment yielded an average T of 14.61 °C with a standard deviation of 0.10 °C (1 s.d., n = 3), which is statistically indistinguishable from the T obtained for untreated samples (t test, P < 0.01). These results are consistent with the reported stability of TEX86 estimates in samples subjected to hydrous pyrolysis at temperatures below 220 °C (ref. 57).

### Calculation of δ18OP

Parent water δ18O was calculated using water-to-calcite isotope fractionation factors based on the temperature calibration from ref. 34, our TEX86 temperature data and dual clumped isotope temperatures.

### Precipitation map

The map in Fig. 1 was created ﻿online with climate explorer (https://climexp.knmi.nl/start.cgi) using the CRU TS4.02 dataset58.

### Data smoothing

Jiangjun δ18OFl and TEX86 smoothed loess curve in Fig. 4 was modelled with a smoothing factor of 0.17, and 95% confidence limits were determined by a bootstrapping technique in PAST software59.