CCC occurrences were mapped and samples were collected from five distinct heaps labelled A to E (Supplement Fig. S1). In addition, a small in situ stalagmite was taken from the same chamber. Cave air temperature was recorded on an hourly basis using a Hobo Temp Pro v2 logger (Onset Computer Corp.) between August 2016 and July 2017.

CCC samples were cleaned in an ultrasonic bath prior to documentation and laboratory analyses. Individual morphologies were examined using a Keyence VHX-6000 digital microscope.

CCC samples were analysed for their stable oxygen and carbon isotope composition using isotope ratio mass spectrometry (Spötl and Vennemann, 2003). In addition, two larger CCC particles were cut in half and micromilled at 0.1 to 0.3 mm resolution. The results are reported relative to the VPDB standard with a long-term precision better than ± 0.08 ‰ (1σ) for both δ13C and δ18O.

A total of 19 individual CCC particles were selected for 230Th dating; 15–20 mg of calcite was drilled using a handheld drill from the centre of 17 crystals in a laminar flow hood. Two skeletal CCC crystals were analysed as a whole as they were too small for aliquots to be drilled from them. Growth layers of a stalagmite (Cioc1) collected next to the CCC spots were drilled at three discrete horizons (2, 24 and 50 mm from the top) and prepared for analyses.

Ages were determined by measuring U and Th isotope ratios on a multi-collector inductively coupled mass spectrometer after their chemical separation following Edwards et al. (1987) and Cheng et al. (2013). Analyses were performed at Xian Jiaotong University (China); 2σ uncertainties for U and Th isotopic measurements include corrections for blanks, multiplier dark noise, abundance sensitivity and contents of the same nuclides in the spike solution. Decay constants for 230Th and 234U were reported by Cheng et al. (2013). Corrected 230Th ages assume an initial 230Th / 232Th atomic ratio of (4.4 ± 2.2) × 10-6 and a 232Th / 238U value of 3.8 as the value for material at secular equilibrium with the bulk earth. Final ages are given in years before present (before CE 1950).

Heat conduction from the surface to 70 m depths was modelled using a 1D heat-flow model (10.5281/zenodo.3982221) that considers conductive heat transfer only and solves the heat equation (Eq. 1) utilising finite differences as space discretisations alongside a forward Euler time-stepping scheme, stabilised with a diffusion-type Courant–Friedrichs–Lewy condition. The heat equation (Eq. 1) is described as 1∂T∂t=α∂2T∂z2, where α is the thermal diffusivity, T is temperature and z is the vertical coordinate. The model assumes a homogenous host rock with no internal heat generation. Thermal diffusivity (α) of the limestone was set to 1.2×10-6 m2 s-1 (Hanley et al., 1978) to account for some air-filled porosity.

As the lower boundary condition for the model, the ground heat flux was set to 0 to account for the presence of the shaft in the entrance zone of the cave, allowing exchange with the ambient air in this upper part of the cave above the CCC-bearing gallery and suppressing the minor effect of the geothermal heat flux. Modelling results are shown as MAAT against depth below the surface (Supplement Figs. S2–S4).

CCCs occur as loose crystals and crystal aggregates in small heaps on and partly underneath five breakdown blocks (Supplement Fig. S1). Individual crystals and aggregates thereof come in a variety of shapes and sizes, the largest reaching 1.4 cm in length. Morphologies include amber-coloured crystals and crystal aggregates of rhombic, raft, beak-like and split crystal habits (Supplement Fig. S1). Split and beak-like crystals are most abundant (heaps A, B and D). Translucent skeletal crystals are rarely present in heaps B and D but are dominant in C and E.

CCCs show high δ13C values varying from 1.3 ‰ to 5.4 ‰ and low δ18O values ranging from -21.8 ‰ to -10.1 ‰ (Fig. 2, Supplement Table S1). These values fall within the compositional range characteristic of CCCs (Žák et al., 2018, and references therein), confirming their precipitation in slowly freezing pockets of water enclosed in cave ice. A beak-like crystal 4.5 mm in diameter revealed ≤ 1.3 ‰ and ≤ 1.2 ‰ of internal variability in δ13C and δ18O, respectively, while a 5 mm rhombohedral crystal shows ≤ 0.4 ‰ and ≤ 0.2 ‰, respectively. Holocene stalagmites from the same chamber exhibit distinctly different stable isotope values, with δ13C and δ18O values ranging from -7.6 ‰ to -2.0 ‰ and from -9.1 ‰ to -7.4 ‰, respectively (Fig. 2).

The 238U concentration of CCC samples varies from 0.8 to 2.2 ppm (Table 2). Four samples yielded 230Th / 232Th atomic ratios less than 100 × 10-6, indicating significant detrital contamination (Table 2), and the corresponding ages were excluded. The 2σ precision of the remaining 13 ages ranges from 0.4 % to 2.7 %.

All CCC ages fall within the YD (as defined by Rasmussen et al., 2014), whereby individual ages show a spread from 12.61 ± 0.22 to 12.08 ± 0.19 ka. Although most of the ages overlap within their 2σ errors, some 230Th ages suggest that CCC formation commenced at the beginning of the YD (i.e. 12.61 ± 0.22 ka). On the other hand, the error-weighted mean of all ages is 12.19 ± 0.06 ka, and 90 % percent of 230Th dates cluster at ∼ 12.2 ka. There is no systematic age difference between samples from individual heaps or between different CCC morphologies.

In contrast to CCC, the 238U concentration of stalagmite Cioc1 is much lower (∼ 0.5 ppm). The resulting 230Th ages demonstrate that stalagmite growth commenced during the late glacial at 14.98 ± 0.14 ka (Table 2). Petrographic analysis provides strong evidence for a growth interruption after which calcite deposition restarted in the mid-Holocene at 5.88 ± 0.15 ka and continued until 1.32 ± 0.27 ka.

We performed a series of model runs covering possible climate scenarios for the YD (Table 1) to explore the relationship between atmospheric temperature changes and the temperature 50 m below the surface at this high-elevation site. We define the thermal boundary conditions for CCC formation as -1 to 0 ∘C (the “CCC window” – see the Discussion section) whereby the likelihood of CCC formation is highest between -0.5 and 0 ∘C. Palaeotemperature estimates used in these computations are based on regional annual and summer air temperature reconstructions.

CCCs form in slowly freezing water pockets enclosed in cave ice when the cave interior temperature is slightly below 0 ∘C (e.g. Žák et al., 2018). Although the deposition of CCCs in many cases marks climate transitions (e.g. Luetscher et al., 2013; Richter and Riechelmann, 2012; Spötl and Cheng, 2014; Spötl et al., 2021; Žák et al., 2012), the large size of some CCCs (up to 50 mm in caves elsewhere; unpublished data from our group) and 230Th ages from their central and rim areas (unpublished data from our group) argue for very stable cave microclimate conditions for at least several years.

CCCs (Table 2) provide unequivocal evidence that perennial ice was present in the lower descending gallery of Cioccherloch during the first part of the YD (Fig. 5). To evaluate the likelihood of CCC formation, the kernel density of all 230Th ages was calculated. As the majority of 230Th ages overlap within their 2σ errors, it is not possible to determine whether CCC formation took place semi-continuously for 400–600 years or if they represent two different generations clustering at ∼ 12.6 and ∼ 12.2 ka (Fig. 5). Overall, CCCs in Cioccherloch record negative interior cave air temperatures very close to 0 ∘C from ∼ 12.6 to ∼ 12.2 ka, initiating progressive freezing of meltwater pockets in perennial ice which were created by drip water.

The air temperature in the homothermic zone of caves reflects the MAAT of the outside atmosphere. Ice-bearing caves, however, commonly represent an exception to this rule (e.g. Luetscher and Jeannin, 2004; Perşoiu, 2018, and references therein). Due to its descending geometry lacking a lower entrance, Cioccherloch acts as a cold trap, as evidenced by the snow and ice cone at the base of its entrance shaft. The resulting negative thermal anomaly, however, is mostly restricted to the upper cave level close to this snow cone, and cave air temperature data show that this influence gets attenuated towards the inner parts of the cave. Today, the descending gallery with the CCC occurrences experiences a small intra-annual temperature variation (∼ 0.5 ∘C), and the MAAT in this gallery matches the ambient MAAT (2.5 ∘C). If the CCC-bearing chamber had been ventilated in the past at times when the outside air temperature was lower than the cave air temperature (i.e. during the winter season), the fine crystalline variety of CCCs would have formed by rather rapid freezing of water films (e.g. Žák et al., 2008) instead of coarse crystalline CCCs that require stable thermal conditions. Today this gallery is connected to the cave's upper level via a narrow squeeze, but prior to cave exploration in the 1980s and 1990s, this squeeze was closed by rubble. Therefore, we presume that this descending gallery was essentially thermally decoupled from the upper cave level in the past and hence unaffected by the negative thermal anomaly, especially when the snow and ice cover was larger than today and buried the entrance of the squeeze. As a result, CCCs in the lower gallery record changes in the thermal state of the subsurface as a consequence of atmospheric temperature changes. Given the lack of ventilated shafts connecting this gallery to the surface, we argue that heat exchange between the latter and the gallery occurred primarily via conduction. Additional heat transfer occurred via drip water, and minor air advection via small fissures in the ceiling cannot be ruled out. Seepage water leading to meltwater pockets in the cave ice, however, is likely thermally equilibrated with the ca. 50 m thick rock above the cave and given its very low discharge carries comparably little heat from the surface. Advective processes are difficult to quantify for any time in the past, and no attempts were made to include them in the thermal model. Qualitatively, both processes would increase the rate of temperature change in this gallery as a response to atmospheric change above the cave. By considering heat conduction only, our simulations yield quantitative constraints on the maximum duration of temperature changes propagated into the shallow subsurface.

The formation of CCCs near the lower end of the descending gallery requires the 0 ∘C isotherm to be located about 50 m below the ground surface at this high-elevation site in the Dolomites during the YD. On the other hand, stalagmite Cioc1 provides strong evidence that the air temperature in this gallery was constantly above 0 ∘C during the Bølling–Allerød interstadial. The δ13C values of Cioc1 decrease from -2.5 ‰ to -6.6 ‰, suggesting gradual soil development above the cave (unpublished data by the authors). In the early YD, atmospheric cooling led to the aggradation of cave ice. During the mild YD summers (from ∼ 12.6 to ∼ 12.2 ka) drip water sourced from snowmelt and/or rain created meltwater pools on the ice that subsequently underwent slow freezing at cave air temperatures ∼ 2–3 ∘C (i.e. 2.5 ± 0.5 ∘C) lower than today.

CCC deposition in caves has traditionally been attributed to near-surface permafrost degradation in response to atmospheric warming (e.g. Žák et al., 2018, and references therein). In this case, however, stalagmite growth (Cioc1) in Cioccherloch indicates cave air temperatures above 0 ∘C and the presence of dripping water in the cave at the Bølling–Allerød interstadial. Thus, CCC formation starting at 12.6 ka cannot represent a delayed response to the Bølling–Allerød interstadial warming. Instead our data suggest that CCCs in Cioccherloch formed during a cold climate associated with permafrost aggradation. Therefore, we hypothesise that analogous to climate warmings possible CCC windows opened during the transition into stadials and remained open for a variably long period of time depending on the local thermal conditions in the subsurface.

Our thermal model provides quantitative constraints on how the cave air temperature evolved in response to different climate conditions. The model uses two input parameters: MAAT and the insulating effect of winter snowpack. A recent study by Schenk et al. (2018) suggests that YD summers remained relatively warm, with temperature decreases of 4.3 ∘C in NW Europe and 0.3 ∘C in E Europe relative to the preceding Bølling interstadial. We therefore kept the July temperatures 3 to 4 ∘C lower than modern values (Table 1) and attributed most of the MAAT change to winter cooling. As the buffering effect of the snow is set to its maximum value to counteract the winter chill, this approach allows the reconstruction of the possible maximum amplitude of MAAT (winter) cooling of the surface.

CCCs dated to the first and second half of the YD indicate conditions very close to 0 ∘C for an extended period of time during this stadial. Our thermal model shows that CCC formation starting at 12.6 ± 0.2 ka at this sensitive mountain site requires a moderate atmospheric cooling at the Allerød–YD transition of -4.5 to -5.0 ∘C relative to today (Fig. 5). Without winter snow cover (scenarios 2b and 2d) the CCC window would open too early and close quickly afterwards, inconsistent with the CCC ages (Fig. 5). Scenarios 2c and 2e show that if the YD climate was characterised by a -5.0 to -4.5 ∘C drop in MAAT relative to today (i.e. ΔMAATAllerød-YD=-3.0 to -2.5 ∘C), stable snow cover during winter is needed to prevent the cave from freezing, effectively shielding the ground from the cold stadial winters (see Zhang, 2005). Scenario 2e includes winter snow cover that provides the best fit with the CCC data, giving rise to a 400-year-long period characterised by a very slow cooling of the subsurface with cave air temperatures near -0.8 ∘C.

Our data do not support the notion of a very cold YD in the Alps (ΔMAATAllerød-YD=-7 to -8 ∘C; scenario 2a) as suggested by speleothem data from low-lying caves in northern Switzerland (Affolter et al., 2019; Ghadiri et al., 2018). Such a drastic lowering of the MAAT would freeze Cioccherloch rapidly, even if a thin winter snowpack was present, and would result in rather abrupt development and deepening of permafrost, preventing CCC formation (Fig. 3). Such a stark cooling would in fact lead to climate conditions similar to the Last Glacial Maximum (LGM), for which noble gas data from groundwater studies around the Alps suggest 7–10 ∘C lower temperature compared to the Holocene (e.g. Šafanda and Rajver, 2001; Stute and Deák, 1989; Varsányi et al., 2011) and which would inevitably lead to the build-up of glaciers at this elevation in the Dolomites.

Our data, however, are consistent with observations from rock glaciers (Frauenfelder et al., 2001) and lake sediments in the Swiss Alps (von Grafenstein et al., 2000) suggesting a moderate cooling at the Allerød–YD transition. A fluid-inclusion-based palaeotemperature reconstruction using stalagmites from Bärenhöhle in western Austria also indicates a maximum temperature drop of about 5.5 ∘C, supporting our interpretation (Luetscher et al., 2016).

CCCs provide a uniquely robust control on cave air temperatures and consequently on the MAAT above the cave. Our data argue for a ≤ 3 ∘C drop in MAAT at the Allerød–YD transition but provide no direct information on the seasonal cycle of ambient atmospheric temperatures. A recent multi-proxy model comparison by Schenk et al. (2018) suggests persistently warm summers during the YD with a median regional cooling of 3 and 0.3 ∘C over NW and E Europe, respectively, compared to Bølling–Allerød summers. These authors also argue that previous studies using chironomids overestimated the YD cooling signal. A similar amplitude of change is suggested for July air temperatures by lake records from the western Alps. Pollen and Cladocera-inferred temperature reconstructions indicate a summer cooling of 2–4 ∘C at the Allerød–YD transition at Gerzensee (Swiss Plateau; e.g. Lotter et al., 2000), consistent with a 3.5 ∘C drop in July air temperatures reported from the Maloja Pass (Fig. 6) in eastern Switzerland (Ilyashuk et al., 2009). We therefore consider 0.3 and 4 ∘C as minimum and maximum estimates of YD summer cooling, respectively, relative to the Bølling–Allerød. As our CCC data in conjunction with thermal modelling constrain the drop in MAAT at the Allerød–YD transition to ≤ 3 ∘C, they allow us to calculate the maximum possible increase in seasonality at the Allerød–YD transition. If we assume that YD summers were only 0.3 ∘C colder than in the Allerød by using the lowest summer temperature change proposed by Schenk et al. (2018), we find that early YD winters at 2270 m a.s.l. were no colder than -13.7 ∘C (mean January temperature). This argues for an enhanced seasonality in the Dolomites, whereby the winter–summer temperature difference increased by up to 5.4 ∘C at the Allerød–YD transition (i.e. ≤ 22.4 ∘C in the YD).

Thermal modelling shows that a winter snowpack effectively shielding the subsurface during the cold winters is needed to account for CCC formation commencing at 12.6 ± 0.2 ka at Cioccherloch. Studies in modern permafrost areas suggest that stable winter snow cover of only ∼ 35 cm results in a positive shift (i.e. snow ΔT) of up to 5.5 ∘C in the mean annual ground surface temperature (Zhang, 2005). Changes in the timing, duration, thickness and density of the snow cover may promote either the development or the degradation of permafrost (Zhang, 2005). While a snowpack in the cold season leads to a positive ground temperature anomaly, summer snow cover insulates the ground from warm air and hence facilitates the development of permafrost. In a study of Arctic permafrost, Park et al. (2014) found that the thermal state of the underlying soil is more affected by early winter than peak winter snowfall. Therefore, we argue for a moderately humid early YD with snowfall during fall and early winter.

Proxy data show a heterogeneous picture with respect to precipitation in the Alps during the YD (Belli et al., 2017; Kerschner et al., 2016; Kerschner and Ivy-Ochs, 2008). The distribution of rock glaciers in the eastern Swiss Alps suggests a 30 %–40 % reduction in precipitation compared to today (Frauenfelder et al., 2001). Palaeoglacier records for the central Alps point to a similar reduction, mostly due to a reduction in winter precipitation. The few palaeoglacier records in the southern Alps suggest that precipitation sums were similar to modern-day values, but winter precipitation was probably reduced compared to the modern day, leading to enhanced seasonal contrasts (Kerschner et al., 2016). A recent study compiling data from 122 palaeoglaciers shows that the southerly displacement of the polar frontal jet stream led to cold air outbreaks and increased cyclogenesis over the Mediterranean and negative precipitation anomalies over the Alps compared to the modern day (Rea et al., 2020). The precipitation pattern during the early YD closely resembled the modern Scandinavian circulation patterns, resulting in an autumn to early winter (September to November) and midwinter (December to February) precipitation increase over the Mediterranean (Rea et al., 2020). This is consistent with our findings, which argue that YD autumns and early winters remained relatively snow-rich in the southern Alps, although the amount of annual precipitation may have been lower than today.

High-resolution speleothem (Baldini et al., 2015; Bartolomé et al., 2015; Rossi et al., 2018) and lake records (Bakke et al., 2009; Brauer et al., 2008; Lane et al., 2013) from W and N Europe suggest a time-transgressive change in atmospheric circulation during the YD associated with a warming of parts of Europe due to a retreat of winter sea ice and a northward migration of the polar front (e.g. Baldini et al., 2015; Bartolomé et al., 2015). At Meerfelder Maar, Germany, this so-called mid-YD transition occurred at 12.24 ± 0.04 ka (Lane et al., 2013). This timing is strikingly similar to the error-weighted mean of the CCC dates from Cioccherloch (12.19 ± 0.06 ka), suggesting that the main phase of CCC formation at this site may have been related to supra-regional climate change. Our thermal simulations provide important constraints on the type and magnitude of climate change during the mid-YD transition at this high-Alpine site and suggest a slight warming by up to 1 ∘C (MAAT), with a slight reduction in precipitation.

A change from moderately snow-rich to snow-poor autumns and early winters from the early to the late YD combined with a small atmospheric warming (scenario 3c) is consistent with the main advance of Alpine ice glaciers (Fig. 6) during the first few centuries of the YD (Baroni et al., 2017; Heiri et al., 2014b; Ivy-Ochs et al., 2009), followed by a glacier reduction and the parallel increase in rock glacier activity (Ivy-Ochs et al., 2009; Kerschner and Ivy-Ochs, 2008; Sailer and Kerschner, 1999).

Benthic ostracod δ18O records from Ammersee (von Grafenstein et al., 1999) and Mondsee (Lauterbach et al., 2011) on the northern fringe of the Alps show a gradual increase of ca. 1 ‰ across the YD (Fig. 6). This slight increase is compatible with our interpretation of a small decrease in autumn precipitation in the northern Alpine catchment areas of these lakes in the second half of the YD coupled with a small (≤ 1 ∘C) warming. High-resolution δ18O speleothem records from Hölloch Cave west of the Ammersee catchment (Li et al., 2021) and from Milandre Cave in the Swiss Jura Mountains (Affolter et al., 2019) exhibit a gradual increase in δ18O across the YD similar to the northern Alpine lake records (Fig. 6). While our data and those from palaeoglaciers and lake sediments are consistent with a slight warming at the mid-YD transition, they argue for a reduction in fall and winter precipitation. This suggests that the popular model of a south–north migration of the polar front and a concomitant increase in westerly-driven precipitation (e.g. Lane et al., 2013) is too simplistic, at least for the greater Alpine realm, underscoring the need for regionally resolved palaeoclimate models. In fact, even at the key site of Meerfelder Maar, the proposed increase in (winter) precipitation is poorly captured by most proxy data except for the abundance of Ti, which is attributed to spring snowmelt (Lane et al., 2013).