Millennial-scale climate variability in the Northern Hemisphere influenced glacier dynamics in the Alps around 250,000 years ago

Abstract

Mountain glaciers are sensitive recorders of natural and human-induced climate change. Therefore, it is imperative to obtain a comprehensive understanding of the interplay between climate and glacier response on both short and long timescales. Here we present an analysis of oxygen and carbon isotope data from speleothems formed mainly below a glacier-covered catchment in the Alps 300,000 to 200,000 years ago. Isotope-enabled climate model simulations reveal that δ¹⁸O of precipitation in the Alps was higher by approximately 1 ‰ during interstadials compared to stadials. This agrees with interstadial-stadial amplitudes of our new speleothem-based estimate after correcting for cave-internal effects. We propose that the variability of these cave-internal effects offers a unique tool for reconstructing long-term dynamics of warm-based Alpine palaeoglaciers. Our data thereby suggests a close link between North Atlantic interstadial-stadial variability and the meltwater dynamics of Alpine glaciers during Marine Isotope Stage 8 and 7d.

Introduction

Caves and speleothems are ubiquitous features in (carbonate-bearing) mountains worldwide, which often harbour glaciers at high elevations and/or during cold climate periods. Mountain glaciers are rapidly vanishing¹ and have been classified as a regional impact tipping element², underscoring the need to improve our understanding of their dynamics on short and long timescales.

Speleothems from the northern rim of the European Alps offer a valuable archive to investigate Northern Hemisphere glacial-interglacial climate variability and provide crucial insights into abrupt climate changes on millennial timescales^{3,4,5,6,7,8,9,10,11,12,13,14,15,16}. The Northern Alps are strongly influenced by the North Atlantic^10,17, a region which is thought to dictate global-scale abrupt climate events during glacial times^{18,19,20,21,22,23}. Speleothems from Alpine caves have been shown to mimic millennial-scale variability and abruptness of climate transitions known from Greenland ice cores^{4,7,9,12,13,16}. The large amplitude δ¹⁸O signals captured in speleothems from the Northern Alps (in contrast to speleothems from low-elevation sites outside the Alps^24,25) have been attributed largely to changes in regional air temperature^7,13, but the detailed processes controlling these isotopic changes remain to be fully understood.

Here, we aim to resolve these processes controlling the millennial-scale isotopic changes in the high-elevation Northern Alps on interstadial-stadial (IS-S) timescales. For this, we produced a high resolution, precisely U-Th-dated speleothem multi-proxy record (δ¹⁸O, δ¹³C, Mg/Ca and Sr/Ca) for the time span ~300 to 206 ka, i.e. covering the end of Marine Isotope Stage (MIS) 9, MIS 8, and the majority of MIS 7, from the northern rim of the Alps in central Switzerland. Speleothem records from this cave site (Melchsee-Frutt) have been found to record interstadial-stadial (IS-S) variability¹⁶ and we identify periods of pronounced IS-S variability during times of intermediate relative sea level. We use speleothem δ¹³C data to identify periods when the cave site was likely overlain by ice in order to provide constraints on the evolution of Alpine glaciers during glacial periods where essentially no other data are available. We propose that the speleothem C isotopic composition and the Mg/Ca ratio record changes in the amount of glacier surface meltwater vs. the amount of basal meltwater occurring on Northern Hemisphere IS-S timescales. Finally, we show that IS-S amplitudes obtained from our speleothem record, after accounting for cave-internal effects, are consistent with those derived from an isotope-enabled fully-coupled ocean-atmosphere model²³ and likely reflect higher δ¹⁸O values of North Atlantic seawater during interstadials compared to stadials.

Results and discussion

Interstadial-stadial variability in the Northern Alps and other European records during MIS 9‒7

We present δ¹⁸O, δ¹³C and Mg/Ca composite records based on three stalagmites retrieved from two high-elevation caves in the Melchsee-Frutt region of central Switzerland (46°47′N, 8°16′E; ~2000 m a.s.l.; composites produced using the iscam age modelling algorithm²⁶, see Supplementary Information and Figs. S1─S7 for details). We detect transitions to interstadial conditions in these records using an automatic algorithm (S-IS transitions, Fig. S8 in Supplementary Information). This algorithm is able to identify transitions based on their higher abruptness relative to other changes in the respective record. We compare our record and detected transitions to other European speleothem and sediment records (Fig. 1). The identified abrupt transitions exclusively occurred during periods of intermediate sea level (below about ‒30 m and above about ‒100 m relative to modern sea level, Fig. S9) as it was common for IS-S variability also during other glacial cycles. Periods of occurrence and absence of abrupt S-IS transitions in the presented speleothem δ¹⁸O record are corresponding to the ones from other European speleothem and sediment records (Fig. 1d–f). Additionally, these periods are also similar for more distant records including a synthetic Greenland ice-core record and a Chinese speleothem composite record (Fig. S10). Common periods of occurrence of S-IS transitions in these records are ~220–230, 240–255, 260–265 and 280–305 ka.

Fig. 1: Comparison of the new Alpine speleothem record with other records from Europe showing interglacial-glacial and millennial-scale (IS-S) variability.

a Melchsee-Frutt δ¹⁸O record (grey and black are original and PCP-corrected records, see Methods and Supplementary Information for details on the PCP correction). b Melchsee-Frutt δ¹³C record (orange) and host rock δ¹³C value indicated by the orange dashed horizontal line together with a global relative sea level (RSL) reconstruction (⁹⁶, black dashed line). c Melchsee-Frutt Mg/Ca record (purple). d Speleothem δ¹⁸O from Spannagel cave, Central Alps^6,97, and Piani Eterni, Italy²⁷. e Speleothem δ¹³C from Ejulve cave, Spain²⁸ and Crovassa Azzurra, Italy²⁹. f Lake Van sediment core reflectance^30,31. g. Iberian Margin alkenone-${U}_{37}^{{k}^{{\prime} }}$ reconstructed sea-surface temperature⁹⁸. Stars mark the detected abrupt interstadial transitions (colours correspond to time series colours, see Supplementary Information for details on transition detection method). Vertical black dashed lines mark MIS substages as indicated by their number on top of the figure (at 204.1, 212, 216.1 and 234.3⁹⁷ for MIS 7, 243, 254, 269, 280, 291 and 309 ka⁹⁹, for MIS 8 and 9).

It is important to note that abrupt climate transitions similar to glacial S-IS transitions occurred also during MIS 7d in the Northern Alps (Melchsee-Frutt speleothems, Fig. 1a–d), in southern Europe (Italy and Spain, recorded by speleothems from Piani Eterni²⁷, Ejulve cave²⁸ and Crovassa Azzurra²⁹, Fig. 1d, e) and southeastern Europe (Lake Van^30,31, Turkey, Fig. 1f). MIS 7d was a cold substage during an interglacial (MIS 7) and demonstrates that relatively short periods of global cooling are sufficient to enable trigger mechanisms for S-IS transitions provided that the conditions are favourable, i.e. relative sea level is at about −60 m for MIS 7d (Fig. 1).

Alpine speleothem δ¹⁸O and δ¹³C data records palaeoglacier dynamics

Melchsee-Frutt speleothem δ¹³C values fall into two populations separated by the mean host rock δ¹³C value of ~2.7 ‰ (Fig. 2). Carbon isotope values close to this value point to conditions where soil respiration was close to zero and hence soil organic matter was largely absent above the cave (cf.³³). Speleothem formation requires that host rock is dissolved by acidified water, in order to enable carbonate precipitation later in the cave. Since we infer from speleothem δ¹³C that speleothems in this cave also grew during times of soil absence, a process other than soil respiration must have led to the acidification of the water. Previous studies suggested that oxidation of pyrite resulting in sulfuric acid can lead to host rock dissolution required for speleothem formation^{32,33,34,35,36,37}. The study area was repeatedly glaciated during the last glacial period^{38,39,40,41,42,43,44} and was probably also covered by a glacier during earlier glacials. Thus, it is likely that during times when signs of biological activity were absent (speleothem δ¹³C higher than mean host rock δ¹³C), a glacier was present above the cave. Conversely, the group showing δ¹³C values lower than the host rock value indicates some input of soil-biogenic ${{p}}{{{\rm{CO}}}}_{2}$ (Fig. 2), suggestive of a climate when thin and patchy Alpine soils and sparse vegetation were present in the catchment, similar to today. Periods with soil coverage above the cave (δ¹³C ≤ 2.7 ‰) are characterised by δ¹⁸O values ranging from –9 to –8‰, while calcite precipitated during periods when no soil was present above the cave shows more depleted δ¹⁸O values (as low as –13.5‰; Fig. 2).

Fig. 2: Scatter plot of δ¹⁸O and δ¹³C values from Melchsee-Frutt stalagmites.

Two groups can be separated by a δ¹³C threshold at ~2.7 ‰ (host rock value). The DBSCAN algorithm included in the Python package scikit-learn¹⁰⁰ was used to cluster the data into groups (parameters: eps = 0.75, minimum number of points in a cluster = 330). Points are automatically assigned by the clustering algorithm where contiguous regions with high density of data points are classified as clusters and points which do not fall into such a region and do not form a contiguous region with high density of data points are not classified. Magenta stars indicate the data characterised by δ¹³C values lower than the threshold while the majority of data is characterised by δ¹³C values above the threshold (blue stars). Black stars mark extreme values which have not been assigned to one of the groups (74 samples).

We find very high δ¹³C values (up to +12‰, Fig. 2), comparable to extreme values previously found in other Alpine speleothems (e.g., up to +20‰ in Spötl & Mangini³⁶, up to +7‰ in Spötl et al. ³³). These anomalously high values cannot be explained solely by the lack of biogenic, low-δ¹³C soil ${{{\rm{CO}}}}_{2}$. Instead, such high δ¹³C values can be attributed to pyrite oxidation and resulting sulfuric acid dissolution followed by prior calcite precipitation (PCP)^33,45 and/or rate-dependent fractionation³³. PCP comprises calcite precipitated before the drip water reaches the top of the stalagmite.

Linear regression slopes between speleothem Mg/Ca and Sr/Ca ratios can be used to detect the influence of PCP on the proxy data^46,47 and we indeed find such evidence (Figs. S11, S12 and S13 in the Supplementary Information). Recent isotope modelling studies have found that PCP does not only affect the δ¹³C value but can also result in changes in δ¹⁸O depending on the conditions for isotopic exchange processes^48,49,50,51. To test this, we conducted isotope simulations using the model by Skiba and Fohlmeister⁴⁹ where we enabled isotopic exchange of ${{p}}{{{\rm{CO}}}}_{2}$ and ${{{\rm{H}}}}_{2}{{{{\rm{O}}}}}$ with the bicarbonate in the drip water. We assumed a cave temperature of 3 °C (Tab. S1) and check if, after some time for PCP, exchange processes are able to bring the δ¹⁸O values of the bicarbonate back to initial δ¹⁸O. We find that exchange processes are too slow (when the PCP occurred within the cave⁴⁸) to re-equilibrate the bicarbonate isotope composition to pre-PCP values when temperatures and cave air ${{p}}{{{\rm{CO}}}}_{2}$ are low as is expected for cold, high-altitude caves (Fig. S14).

We compared the PCP-sensitive Mg/Ca proxy to the δ¹³C and δ¹⁸O records to investigate whether there is a positive relationship between them which would be expected if PCP had a major impact on the isotope records. We applied a low- (>4000 yrs) and high-pass (<4000 yrs) filter on the composite records to distinguish between correlations on IS-S timescales (high-pass filtered data) and longer timescales (low-pass filtered data). Pearson correlation coefficients of low-pass filtered records are negative for δ¹³C vs. δ¹⁸O and δ¹⁸O vs. Mg/Ca but positive for δ¹³C vs. Mg/Ca (Table 1). For the high-pass filtered time series, all correlation coefficients are positive which indicates that PCP is influencing all proxies. Thus, we find that PCP is a dominant process affecting the δ¹³C, δ¹⁸O and Mg/Ca proxies in the Melchsee-Frutt speleothems on IS-S timescales. However, on longer timescales δ¹³C and Mg/Ca are as well dominated by PCP while δ¹⁸O is likely predominantly reflecting broader, regional-scale δ¹⁸O_ppt variability which results in a negative relationship with the PCP-sensitive proxies (δ¹³C and Mg/Ca). Accordingly, we find a high and positive correlation between δ¹⁸O and RSL (r = 0.82, p < 0.0001) on longer timescales (longer than 4000 yrs).

Table. 1 Pearson correlation coefficients between high-pass and low-pass filtered δ¹³C, δ¹⁸O and Mg/Ca.

Enhanced PCP has commonly been attributed to water-limited conditions in the vadose karst system associated with a dry climate^{52,53,54,55,56}. However, climate model simulations suggest that interstadials were characterised by higher precipitation than stadials in the Northern Hemisphere⁵⁷ during both winter and summer, following the Clausius-Clapeyron-relationship (Fig. S15 b and c). Higher precipitation amounts lead to shorter times for PCP due to a higher availability of water in the karst system which could potentially result in decreased C and O isotope values during interstadials, contrary to what our record shows. Therefore, we suggest that the PCP variability in our Alpine speleothems reflects the extent of host rock dissolution where enhanced dissolution and, thus, higher initial ${{Ca}}^{2+}$ concentration in the drip water will lead to increased subsequent calcite precipitation, i.e. higher amounts of PCP (Fig. S14). It is important to note that this is independent of the amount of infiltrating water, which would solely influence the duration of PCP.

In this setting, the extent of host rock dissolution depends on the amount of pyrite oxidation. The oxidation of pyrite releases sulfate and requires the presence of water and dissolved oxygen⁵⁸:

$${{FeS}}_{2}+{H}_{2}O+\frac{7}{2}{O}_{{2}_{{aq}}}\to {{Fe}}^{2+}+{2{SO}}_{4}^{2-}+{2H}^{+}$$

(1)

Studies in the Swiss Alps have found that water originating from snow melt has dissolved oxygen contents of up to 60-80 % of the atmospheric value. On the contrary, water derived from glacier ice melt is usually depleted in oxygen reaching only 35-40 % with respect to the atmospheric value. This can be explained by the closer contact of snow with atmospheric ${{{\rm{O}}}}_{2}$⁵⁹. Furthermore, subglacial meltwater is likely depleted in oxygen because the subglacial environment is known to host aerobic microbial communities^59,60,61. Since the extent of pyrite oxidation (and, thus, drip water ${{Ca}}^{2+}$ concentration and the amount of PCP, Fig. S14) depends on the availability of dissolved oxygen in the water^62,63, we suggest that during interstadials higher amounts of water with higher ${{{\rm{O}}}}_{2}$ content infiltrated into the karst in the caves' catchment, i.e. a higher amount of precipitation and snow melt (Fig. S15) and less meltwater from glacier basal ice. Conversely, a higher proportion of ${{{\rm{O}}}}_{2}$-depleted glacier meltwater entered the karst during stadial times. We reflect on alternative mechanisms and why they are unlikely to explain enhanced PCP observed during interstadials in the Method section.

Two processes could have contributed to a larger proportion of ${{{\rm{O}}}}_{2}$-rich water during interstadials compared to stadials (Fig. 3): (1) higher snow accumulation in winter and more summer surface meltwater due to higher summer temperatures together with higher rainfall in summer (Fig. S15), (2) larger glacier thickness during stadials, thus, lowering the melting point due to increased pressure and resulting in a higher basal meltwater contribution.

Fig. 3: Hypothesised variability of Alpine glaciers and associated input of meltwater into the subglacial karst system on IS-S timescales.

Host rock and cave shown in grey and brown, respectively. During stadials a thicker glacier likely resulted in a higher contribution of ${{{\rm{O}}}}_{2}$-depleted basal meltwater due to the pressure-induced lowering of the melting point. Crevasses might have enabled ${{{\rm{O}}}}_{2}$-saturated surface meltwater to reach the base of the glacier. During interstadials, thinner ice likely increased the melting point and, thus, reduced the amount of basal meltwater contribution. Additionally, the ${{{\rm{O}}}}_{2}$-rich surface water contribution was likely higher due to higher precipitation together with enhanced summer meltwater (Fig. S15).

Northern Alps speleothem δ¹⁸O reflects North Atlantic seawater δ¹⁸O

Previous studies of speleothems of glacial age in the Northern Alps suggested that the speleothem calcite δ¹⁸O reflects δ¹⁸O_ppt^7,13, which today is strongly correlated with air temperature^64,65. Here, we present results from an isotope-enabled climate model, COSMOS-wiso (ref. ²³, see “Methods” section), which confirms higher δ¹⁸O_ppt during interstadials compared to stadials for Western Europe and the North Atlantic (Fig. 4). Positive IS-S differences associated with millennial-scale AMOC oscillations²³ prevail in the North Atlantic which is the main moisture source of the Alps (area between 30 to 60°N)^10,17. Our model simulations suggest δ¹⁸O_ppt values for the Alps which are 0.5 to 1.5‰ higher during interstadials compared to stadials (Fig. 4). We applied a PCP correction on the speleothem δ¹⁸O (Fig. 1a, see “Methods” section and Supplementary Information for details) to obtain an improved estimate of δ¹⁸O_ppt. Indeed, after the PCP correction the amplitude of interstadial-stadial differences in speleothem calcite is close to IS-S differences found for earlier glacials^{4,7,9,12,13,16} and in model simulated δ¹⁸O_ppt over the Alps (Fig. 4).

Fig. 4: Map of simulated interstadial-stadial (IS-S) difference in δ¹⁸O of meteoric precipitation (δ¹⁸O_ppt) and vertically integrated moisture transport for the North Atlantic and Europe.

Colour scale refers to the IS-S mean annual δ¹⁸O_ppt difference and arrows depict the moisture transport, where longer arrows indicate larger IS-S difference. Also shown is the Last Glacial Maximum (LGM) ice extent in the European Alps⁷⁴. Inset shows enlarged Alpine LGM ice sheet and the location of the study site (Melchsee-Frutt, black star). Colour bar refers to both the main map and the inset.

We find that the δ¹⁸O_ppt IS-S difference in the Northern Alps is the result of the IS-S difference of North Atlantic surface seawater δ¹⁸O. Seawater δ¹⁸O of the main moisture source region of the Alps (area between 30 to 60°N) was ~0.5 to 2 ‰ higher during interstadials compared to stadials (Fig. S15g), potentially due to increased evaporation. Thus, it is likely the main driver of the positive IS-S difference in simulated δ¹⁸O_ppt over the Alps as recorded by Northern Alps speleothems. This is corroborated by δ¹⁸O values of planktonic foraminifera from sediment cores in the northern and eastern North Atlantic^66,67. Their δ¹⁸O values indicate a ~1–2‰ IS-S difference in seawater δ¹⁸O when accounting for the ~10 °C calcification temperature difference on IS-S timescales, which is often found in sea surface temperature reconstructions from this region⁶⁸.

A previous study explains speleothem δ¹⁸O variation on glacial and interglacial timescales by zonal shifting of the westerlies leading to enhanced moisture advected to the Northern Alps from the South during the LGM, which is depleted in δ¹⁸O due to orographic rain out¹⁰. However, the COSMOS-wiso model indicates a different pattern on IS-S timescales by simulating a larger contribution of southerly-sourced moisture during warmer compared to colder periods (indicated by IS-S difference in moisture transort, Fig. 4). As this is based on simulations of one single model, further investigations should be conducted to verify this result. However, based on that we conclude that shifts in the position of the westerlies and resulting change in moisture source and pathway are unlikely to explain the higher δ¹⁸O_ppt over the Alps during interstadials relative to stadials.

Conclusions and Implications

The presented δ¹³C record is punctuated with extremely positive values. They occur contemporaneously with δ¹⁸O maxima. We propose that these are the result of an increased proportion of ${{{\rm{O}}}}_{2}$-rich annual precipitation and/or surface summer snow melt compared to a smaller proportion of ${{{\rm{O}}}}_{2}$-depleted glacier meltwater during interstadials relative to stadials. During interstadials, the glacier ice above the Alpine caves was increasingly melting and potentially comparably thin. This could have reduced pressure on the basal ice and thereby the melting point which would lead to less basal meltwater contribution to the karst. Larger contribution of ${{{\rm{O}}}}_{2}$-rich waters then enabled more pyrite to oxidise and carbonate rock to dissolve, leading to higher initial ${{{{\rm{Ca}}}}}^{2+}$ concentrations and, thus, enhanced PCP.

For stadials, speleothem δ¹³C values indicate relatively cold conditions throughout the year resulting most likely in a comparably thick glacier. Thus, annual precipitation and/or summer snow melt amount were supposed to be small, whereas potentially relatively larger amounts of basal meltwater entered the karst (Fig. 3). Larger contribution of ${{{\rm{O}}}}_{2}$-depleted basal meltwater lead to reduced pyrite oxidation, resulting in less carbonate rock dissolution and decreased ${{{\rm{Ca}}}}^{2+}$ concentrations and, hence, reduced PCP compared to interstadials.

According to our interpretation, speleothem δ¹³C values below the host rock δ¹³C value of ~2.7 ‰ provide evidence of carbonate dissolution due to soil ${{{\rm{CO}}}}_{2}$ contribution, and speleothem δ¹³C values above ~2.7 ‰ might indicate the presence of a glacier above the cave site. Thus, the δ¹³C speleothem record would suggest that the catchment of the caves became glaciated already at the end of MIS 9 (Fig. 1b, where δ¹³C values show no isotopic signs of soil influence at the end of MIS 9). However, IS-S variability is only visible in the speleothem δ¹⁸O and not in δ¹³C and Mg/Ca, which suggests that changes in water sources on IS-S timescales are negligible at that time, i.e. potentially indicating the absence of a glacier. Speleothem δ¹³C suggests that the cave site was continuously glaciated throughout MIS 8 with strong changes in meltwater contributions on IS-S timescales. During MIS 7e, c and b the catchment was likely ice-free as suggested by the low δ¹³C values indicating a soil C contribution. During MIS 7d, an interglacial substage, the glacier re-advanced and buried the cave site (indicated by a diminishing soil signal in speleothem δ¹³C) and we find rapid variations in δ¹³C and PCP-corrected δ¹⁸O values during MIS 7d (Fig. 1b) which we consider to result from similar mechanisms as for glacial times.

Our study provides new insights into several relevant and emerging aspects of speleothem science. We show that not only the duration for PCP, determined by the air vs. water content in the karst, but also the amount of PCP (i.e., the difference between initial ${{Ca}}^{2+}$ of the cave drip water, which is determined by the soil ${{p}}{{{\rm{CO}}}}_{2}$ and equilibrium ${{Ca}}^{2+}$ concentrations, when the solution is in equilibrium with cave air ${{p}}{{{\rm{CO}}}}_{2}$) need to be considered when interpreting speleothem stable isotope records. This has important implications for speleothem-based climate proxy interpretation. In the vast majority of studies, PCP-sensitive proxies have been interpreted to reflect the amount of infiltration based on the negative relationship between the duration of PCP and water availability in the karst system. However, the amount of PCP might follow an opposite pattern whereby wetter conditions promote soil bioproductivity⁶⁹ and, thus, increases the initial ${{Ca}}^{2+}$ concentration of drip water with respect to equilibrium ${{Ca}}^{2+}$ concentrations and thereby the amount of PCP.

Our data suggests a strong PCP influence not only on speleothem δ¹³C but also on δ¹⁸O which previously has been thought to be limited due to fast re-equilibration of the dissolved bicarbonate with the drip water as initially suggested by Hendy⁷⁰. Since then, more accurate equilibration times have been reported⁷¹ implying relatively slow equilibration with the drip water as demonstrated also by isotope modelling studies (refs. ^48,49, Fig. S14). Oxygen isotopes are the single most extensively used speleothem-based climate proxy and are of special interest to compare with climate models simulating δ¹⁸O_ppt^72,73. We disentangled the two main signals on IS-S timescales recorded in speleothem δ¹⁸O (δ¹⁸O_ppt vs. PCP) and find a much better match with the isotope-enabled climate model IS-S differences in δ¹⁸O_ppt after correcting for PCP. This also corroborates the isotope-enabled model results that Northern Alps δ¹⁸O_ppt recorded in speleothem archives reflect North Atlantic seawater δ¹⁸O changes on IS-S timescales, but only when accounted for PCP. Thus, it is crucial to focus future research on the impact of PCP on speleothem δ¹⁸O (PCP in the vadose zone vs. on the stalactite, impact of ${{{\rm{CO}}}}_{2}$ exchange) and advancing methods and models to disentangle the PCP component from (isotope) proxies.

If we would have not accounted for PCP in our study, climate-related δ¹⁸O_ppt changes would have been enormously overestimated. This would be difficult to explain with climatic processes alone (e.g., interstadial speleothem δ¹⁸O values as high as interglacial values, Fig. 2). However, it has to be pointed out that the effect of PCP can vary for cave and drip sites and for each case it has to be evaluated how to account for it (e.g., when PCP occurs already in the karst leaving much more time for re-equilibration⁴⁸). PCP corrections involve parameter assumptions which can also introduce errors and such corrections are not always recommended. Furthermore, the strong effect of PCP at our site is also an advantage as it results in more pronounced interstadial-stadial transitions in our speleothem isotope records and thereby improves accuracy when trying to detect them. A major uncertainty of our correction approach arises from the kinetic fractionation effect which we only approximate in our correction. This underscores the need for more laboratory experiments in order to derive robust empirical relationships for quantifying cave-internal processes in speleothem records.

Our study highlights the potential of high-elevation speleothems as an unique archive for studying palaeoglacier dynamics. Such studies of Alpine speleothems which grew during glacials should be extended to other glacials and potentially other mountain regions providing a test bed to validate models of mountain glacier dynamics on millennial to orbital timescales^74,75.

Methods

Stable isotope analyses

Samples for stable C and O isotope measurements were milled along the main growth axis. We hand-milled at 0.2 and 0.5 mm-resolution for M5_14_44B and M5_14_42, respectively. For 0.5 mm resolution we milled samples 0.2 mm wide to maintain a similar length of time over which each sample integrates, but at 0.5 mm steps. Stalagmite M43_401_14 was milled continuously at 0.1 and 0.3 mm-resolution. We obtained 2246, 1103 and 2113 samples for stalagmite M5_14_44B, M5_14_42 and M43_401_14, respectively. Sampling resolution for all stalagmites was guided by the changing growth rate obtained from U-Th ages (mean growth rates are 70, 49 and 42 mm/ka for M5_14_44B, M5_14_42 and M43_401_14, respectively).

Isotopes for stalagmites M5_14_44B and M5_14_42 were measured using an automated carbonate-extraction system (KIEL IV) interfaced with a MAT253 isotope ratio mass spectrometer (IRMS; ThermoFisher Scientific) at the Helmholtz Centre GFZ in Potsdam. Samples of 60-90 µg are dissolved in 103 % H₃PO₄ at 72 °C for 12 minutes and the isotopic composition was measured on the released and cryogenically purified ${{{\rm{CO}}}}_{2}$. Replicate analysis of reference material (NBS19) reported relative to VPDB yielded 1 sigma uncertainties of better than 0.06 ‰ for both δ¹⁸O and δ¹³C. Isotope measurements of M43_401_14 were conducted using a Delta V plus IRMS at the University of Innsbruck. This mass spectrometer is linked to an on-line Gasbench II carbonate preparation system. Details on the measurement procedure can be found in Spötl and Vennemann⁷⁶. Long-term 1 sigma precision is better than 0.08‰ for both δ¹³C and δ¹⁸O. All isotopic ratios are expressed in the delta notation relative to VPDB.

Trace element analyses

The elements Ca, Mg, and Sr were measured along the isotope transects of all three stalagmites using a M4 Tornado micro-XRF scanner at the Helmholtz Centre GFZ in Potsdam. Element intensities were acquired as counts per second [cps], but log-ratios of element intensities are linear functions of element concentrations and provide robust information of relative element concentrations⁷⁷ (Figs. S12 and S13). For comparison with the stable isotope data the trace elemental records were down-sampled using Gaussian Kernels with a width of 0.5 mm.

U/Th dating

Ages for all stalagmites were determined via U/Th disequilibrium dating. Sub-samples of ~200 mg were drilled along growth layers close to the growth axis (Fig. S2). A total of 16 samples were measured for stalagmite M5_14_44B, 18 for M5_14_42, and 25 for M43_401_14. Measurements were conducted with a MC-ICP-MS (ThermoFisher Neptune Plus) at Heidelberg University following the analytical protocol of Wefing et al. ⁷⁸.

Samples are first dissolved, then spiked with ²²⁹Th, ²³³U, and ²³⁶U and subsequently U and Th purified using wet-column chemistry with Eichrom UTEVA resin. Isotope ratios were used to calculate U-series ages according to the decay equations⁷⁹ and error propagation of statistical uncertainty. For this, ²³⁴U and ²³⁰Th half-lives published in Cheng et al. ⁸⁰ were used. Correction for detrital contamination assumes a ²³²Th/²³⁸U weight ratio of 3.8 ± 1.9⁸¹ and ²³⁰Th, ²³⁴U and ²³⁸U in secular equilibrium, and is minor for most samples (Tab. S2). All U and Th isotope measurements have been collected simultaneously and data was corrected for mass fractionation and tailing as mentioned by Wefing et al. ⁷⁸.

Age-depth modelling

Age-depth modelling was performed using iscam²⁶, a programme designed to align records from the same cave (i.e., assuming that a similar signal is recorded) into a master record, which improves the time control provided by individual time series by reducing the age uncertainties within the overlapping periods and increases signal-to-noise ratios of the combined record. The algorithm correlates dated proxy signals from several contemporaneously growing stalagmites by varying their age-depth relation within age uncertainties. It determines the most probable age-depth model and calculates the age uncertainty for the combined record. We run 30,000,000 Monte-Carlo iterations on the time series to find the best option for the age-depth realisation, i.e. the iteration with the highest correlation coefficient between both time series. Significance of the highest correlation coefficient is tested by surrogate time series, which have similar characteristics than the measured isotope records (e.g., same dating information, proxy resolution, memory of the time series²⁶).

For each iteration, we generate 2000 surrogate time series. To calculate the significance limits, we tested each of these 2000 time series in the same way as the measured time series, except that we used only 1000 randomly selected iterations due to computational constraints. For each of the 2000 artificially constructed time series, the highest correlation, found within the 1000 iterations, was used to construct a probability density function, which defines significance limits for the correlation coefficient of the measured time series. All found correlation coefficients of the best-suited age-depth realisations for the individual stacking steps exceed the 95% confidence limits. The resulting isotope records plotted over age agree perfectly with each other (Fig. S3).

Within iscam we applied a pointwise linear interpolation between dated depths to produce the δ¹⁸O composite record. We stacked the individual δ¹³C records to a composite record in the same way as for δ¹⁸O and placed it on the age-model obtained with the δ¹⁸O data. Thus, each data point in the composite records of δ¹⁸O and δ¹³C was generated by the exact same stalagmite(s) samples. The composite records are shown in Figs. S5 and S6.

PCP correction

As explained in the main text, speleothem δ¹⁸O is likely strongly modified by PCP. Thus, we aimed to disentangle the speleothem δ¹⁸O in its PCP and δ¹⁸O_ppt components using the corresponding speleothem δ¹³C as PCP proxy. Since laboratory experiments, isotope modelling (including our sensitivity test using an speleothem isotope model, see Supplementary Information, Fig. S14) and studies investigating speleothem isotope data suggest that kinetic fractionation needs to be accounted for in speleothem isotope studies^{82,83,84,85,86,87}, we utilise an estimate of the kinetic fractionation for δ¹⁸O relative to the fractionation of δ¹³C⁸⁷ (Fig. S16, Supplementary Information for details) to correct our speleothem δ¹⁸O data with an estimate of the δ¹³C-PCP component derived from our speleothem δ¹³C data. We apply the PCP correction on the measured speleothem δ¹⁸O data for each speleothem isotope data point where the δ¹³C value is above the host rock value (2.7 ‰). We use the δ¹³C enrichment above the host rock value (difference to 2.7 ‰) as estimate of δ¹³C increase solely due to PCP and convert it to the expected PCP-driven δ¹⁸O enrichment using the estimate of the kinetic fractionation for δ¹⁸O relative to the fractionation of δ¹³C (Figs. S16 and S17 in the Supplementary Information).

Alternative mechanisms cannot explain enhanced PCP during interstadials

To our knowledge other mechanisms than the one proposed in the main text cannot lead to the larger degree of PCP during interstadials compared to stadials. This section is intended to discuss these other mechanisms and demonstrates that they cannot explain the data.

We expect negligible differences between interstadials and stadials in karst temperature and in infiltrating (melt-) water temperature. When a glacier covered the cave site temperatures below the glacier and in the karst must have been always close to 0 °C. Same can be expected for surface meltwater. Thus, temperature-dependent mechanisms like rate of pyrite oxidation or oxygen saturation in water can be ruled out as explanation.

Carbonate dissolution is comparatively faster than the residence time of water in the karst (saturation achieved in a few hours up to several days^88,89,90), so that equilibrium between ${{{\rm{CO}}}}_{2}$ dissolved in the water and ${{Ca}}^{2+}$ ions should be always attained irrespective of the hydrological conditions (except possibly for drip sites dominated by fracture flow following heavy rainfall).

Since the δ¹³C data indicates no soil and/or vegetation above the cave during glacial times (Fig. 2), also cave air ${{p}}{{{\rm{CO}}}}_{2}$ changes can likely be ruled out as a mechanism for IS-S PCP variability as near-constant atmospheric ${{p}}{{{\rm{CO}}}}_{2}$ values are expected in the cave. As there is only the atmosphere as the main ${{{\rm{CO}}}}_{2}$ reservoir, variability in cave ventilation would not result in substantial varations in cave air ${{p}}{{{\rm{CO}}}}_{2}$.

Climate model simulations

To investigate responses of δ¹⁸O_ppt to S-IS transitions during glacial periods, we used model simulation experiments from Zhang et al. ²³, which successfully simulated glacial millennial-scale climate oscillations using the climate model COSMOS-wiso. The model COSMOS-wiso is a water isotope-enabled fully-coupled atmosphere-ocean general circulation model⁹¹. The atmospheric model is ECHAM5⁹², complemented by the land surface component JSBACH⁹³, and was used at T31 resolution (∼3.75°), with 19 vertical layers. The ocean model is MPI-OM⁹⁴ and includes sea-ice dynamics modelled with viscous-plastic rheology⁹⁵. It has a resolution of GR30 (3° × 1.8°) in the horizontal, with 40 uneven vertical layers.

We chose modelling results from the experiment E40ka_Orb which characterises unforced millennial-scale changes in Atlantic Meridional Overturning Circulation under fixed glacial conditions, resembling glacial S-IS transitions. More details about experimental design can be found in Zhang et al. ²³. We show differences in meteoric precipitation δ¹⁸O (Fig. 4) and seawater δ¹⁸O, precipitation amount and temperature over the eastern North Atlantic and Europe (annual, summer and winter) between the simulated interstadial and stadial states (IS-S, Fig. S15).