A survey of Devaux cave was conducted using a compass and clinometer, as well as a laser distometer (Disto-X; Heeb, 2014). In addition to cave ice, chemical and clastic deposits were mapped in the cave (Fig. 2b). The labelling of the cave chambers (A to K) follows the nomenclature introduced by Devaux (1929) and Rösch and Rösch (1935).

A map of potential solar radiation (RAD) of the MPm was obtained using an algorithm which considers the effects of the surrounding topography on shadowing considering the position of the sun. RAD was calculated for every month and was then averaged to obtain an annual mean. Details of this computation can be found in Pons and Ninyerola (2008).

The cave consists of large rooms (e.g. room F and those located beyond SCAL chatière) connected by small galleries (Fig. 2b), locally with narrow passages (e.g. galleries close to room SPD or SCAL chatière; Fig. 2b). A total of 15 stations were installed in the outmost ∼ 350 m of the cave to monitor air (11 sensors), water (2 sensors), and rock temperature (2 sensors) (Fig. 2b). Cave air temperature variations were recorded using different devices: HOBO Pro v2 U23-001 (accuracy ± 0.25 ∘C, resolution 0.02 ∘C), Tinytag Talk 2 (accuracy ± 0.5 ∘C, resolution 0.04 ∘C), and ELUSB2 (accuracy ± 0.21 ∘C, resolution 0.5 ∘C). The cave river temperature was recorded at two points. The first site (W7) was located close to the Brulle spring (Fig. 2b; HOBO TidbiT V2, accuracy ± 0.21 ∘C, resolution 0.02 ∘C), and the second site (W6) was located in room F (Fig. 2b; HOBO UA-001-08; accuracy ± 0.53 ∘C, resolution 0.4 ∘C). Both sensors were installed at a water depth of 20 cm. Finally, the rock temperature was recorded at two sites (R1 and R2 in room D and K, respectively) using a HOBO U23-003 device (accuracy ± 0.25 ∘C, resolution 0.02 ∘C). Each sensor has two external temperature probes (channels 1 and 2). These temperature probes were installed in two horizontal drill holes of 60 cm depth, ∼ 1.5 to 2 m from each other.

We monitored sporadically the cave during different intervals between 2011 and 2015, while a continuous monitoring was carried out between July 2017 and July 2021. Maximum, minimum, and mean temperatures, as well as the number of frost and/or warm days, were obtained for each sensor and site (Fig. 2b). Changes in the ice morphology were evaluated using wall marks measured at four points since 2013 in room G and using one point during 2020–2021 in room SPD (Fig. 2b) using a digital sliding caliper.

The outside temperature was measured at the “Porche” entrance (∼ 2836 m a.s.l.) and on the southern face of MPm at ∼ 2690 m a.s.l. For comparison, these temperature records were corrected assuming an adiabatic lapse rate of 5.5 ∘C km-1 (López-Moreno et al., 2016; Navarro-Serrano et al., 2018) to an elevation of ∼ 2850 m a.s.l., corresponding approximately to the lower limit of the hydrological catchment area of Devaux. In both cases, the temperature was measured using Tinytag Talk 2 sensors equipped with a radiation shield. These data were compared to the temperature record from the Pic du Midi de Bigorre meteorological station (PMBS; 2011–2020) (2860 m a.s.l., ∼ 28 km north of Devaux) obtained from Météo-France. Moreover, the homogenized data available since 1882 from PMBS (Bücher and Dessens, 1991; Dessens and Bücher, 1995) were used to identify long-term temperature trends.

X-ray diffraction (XRD) analyses were performed on sulfate and carbonate crystals from rooms G, D and K, as well as on sulfide and oxidized crystals thereof from the host rock (Fig. S1 in the Supplement). The analyses were performed at the Geosciences Institute in Barcelona (GEO3-BCN-CSIC) using a Bruker-AXS D5005 powder diffractometer configured in θ/2θ mode (e.g. Rodríguez-Salgado et al., 2021).

Samples of cave drip water, ice, and river water were analysed for major ions by ion chromatography (IC) at the laboratories of the Pyrenean Institute of Ecology (Zaragoza). Carbonate alkalinity was determined by titration within 24 h after sampling.

A total of 16 samples, including sulfate crystals, dissolved sulfate, and pyrite crystals were selected for sulfur isotope analysis at the Godwin Laboratory for Paleoclimate Research of the University of Cambridge (UK), following the methodology of Giesemann et al. (1994). For gypsum samples, ∼ 5 mg of powdered gypsum were dissolved in deionized water at 45 ∘C overnight. Then, a BaCl2 solution (50 g L-1) was added to induce BaSO4 precipitation. In the case of water samples, BaCl2 was added directly to the sample. Subsequently, 6 M HCl was added to remove any co-precipitated carbonate minerals, and the BaSO4 precipitate was rinsed several times with deionized water. Finally, BaSO4 was dried at 45 ∘C overnight. Sulfate dissolved in water was precipitated using the same method.

Isotope measurements were carried out using a Flash Elemental Analyzer (Flash-EA) at 1030 ∘C. The samples were folded in tin capsules. After sample combustion, the generated SO2 was measured by continuous-flow gas source isotope ratio mass spectrometry (Thermo Scientific, Delta V Plus). Samples were run in duplicate and calibration was accomplished using NBS-127. The reproducibility (1σ) of δ34S was better than 0.2 ‰, similar to the long-term reproducibility of the standard over the run (0.2 ‰). δ34S isotope values are reported relative to VCDT (Vienna Canyon Diablo troilite).

Devaux cave is ∼ 2500 m long and comprises three distinct levels (Fig. 2b). The lower and the middle levels correspond to the Brulle spring (0 m) and the “Porche” entrance (∼ +14.5 m), respectively. The third one comprises chambers and galleries +21 to +29 m above the Brulle spring (Fig. 2b). In the inner part of the cave, some unexplored vertical chimneys may connect to sinkholes in the catchment above the cave (Fig. 2a). The main ice deposits are located in rooms D, G, SPD, and K (Fig. 2b). Except for SPD, these chambers located above the Porche entrance (between ∼ +1 and +7 m) can be accessed via ascending passages.

During the cold season, the cave river starts freezing at the spring, and the ice then expands backward into room F (Fig. 2b). The ice totally or partially clogs the main gallery and dams the water inside the cave, forming a small lake (Rösch and Rösch, 1935). This process is important for the seasonal ice extent as the flooding of the cave depends on whether the springs (North 1 and North 2) are frozen or not (e.g. Rösch and Rösch, 1935). Webcam observations (Gavarnie, Oxygène hut) suggest a possible freezing of the Brulle spring from late November to mid-May simultaneous with the freezing of the Gavarnie waterfall. Moreover, historical photos (e.g. Devaux, 1929; Rösch and Rösch, 1935) and our own observations show that snow during winter and spring can reach the Brulle entrance – a situation that also favours the blocking of the springs. As a result of such flooding events, slack water deposits formed in the cave entrance zone but locally also further into the cave (e.g. in rooms I, J, K, and SCAL chatière, along the main gallery; Fig. 2b), while silty sediments are found at elevated positions with respect to the river level (e.g. in rooms D and G). Sandy sediments dominate in the large rooms located beyond the SCAL chatière. Two such successions (∼ 1 m thick) comprising hundreds of rhythmitic fine sand and silt layers are present in elevated areas with respect to the current river, witnessing major events of backflooding.

Observations made during summer show a dominant airflow direction from the inner to the outer parts of the cave, exiting through the Brulle and Porche entrances. Conversely, the opposite is expected for the cold season (chimney effect). When the Brulle spring is partially clogged by ice during early summer forcing the stream to flow below the ice, air flows from room F to C (Fig. 2b) (e.g. summer 2021). The airflow is imperceptible in rooms D, G, and close to K located away from the main cave passages.

The MAAT at the elevation of Devaux cave is ∼ 0 ∘C (-0.04 ∘C; 2017–2021). On the other hand, a positive MAAT (1.8 ∘C) is recorded on the southern side of the MPm at a similar altitude (Fig. 3a). Maximum and minimum air temperatures outside the cave vary between 24.5 and -17.2 ∘C (hourly values, 2017–2021). The PMBS MAAT record (Fig. 3b) shows a warming trend of around +1.5 ∘C since the beginning of the measurements in 1882. Before 1985, temperatures below 0 ∘C dominated the annual cycle, while positive MAATs became more frequent in recent years. Minimum temperatures also show an increasing trend of ∼ +2.5 ∘C, while the maximal annual temperatures do not show a clear trend. The north-facing Gavarnie cirque is associated with a clear RAD anomaly (Fig. 4). Values lower than 215 kWh m-2 are observed at ∼ 2000 m and between ∼ 2800 and 2900 m a.s.l., corresponding to the cirque bottom, the area located behind La Torre peak and the surroundings of Devaux cave. At the cave entrance the RAD value is only 390 kWh m-2, in stark contrast to the summit areas and surroundings where the RAD often exceeds 1500 kWh m-2 (Fig. 4).

While the mean daily air temperature (MDAT) at the cave entrance (purple line in Fig. 5) and the temperature series from PMBS (pink line in Fig. 5) agree in their absolute values, the variability in MDAT at the Devaux entrance is lower than at the PMBS. This pattern could be related to local topographic conditions, leading, for instance, to less RAD, or to the position of the sensor on the cliff (less night emissivity). Given this radiation contrast, warmer temperatures prevail on the southern side of the MPm (Fig. 4), favouring early snowmelt in spring and early summer, while at the same time the temperature stays below 0 ∘C in the cave's surroundings.

The cave can be separated into distinct areas depending on their thermal regime: ventilated galleries (rooms A, B, C, and F and the main gallery from SPD to SCAL chatière) and poorly ventilated parts off the main airflow path (rooms D, G, K; Figs. 2b, 5).

The chemical composition of water in Devaux cave is dominated by calcium and bicarbonate with relatively high Mg concentrations and locally also elevated sulfate concentrations (Table 1). Total dissolved solids (TDSs, n=7) vary from 57 to 315 mg L-1. Devaux's drip water has higher mean sulfate concentrations (65 mg L-1) than the cave river (11 mg L-1) and massive and seasonal ice (2.8–18 mg L-1). The δ34S value of dissolved sulfate in the drip water is -14.4 ‰ (n=1), which is significantly higher than in cave river water (-28.5 ‰ to -27.3 ‰, n=2; Table 2). Gypsum crystals in room D show δ34S values ranging from -15.1 ‰ to -15.8 ‰ (n=7), while in room G they range from -12.3 ‰ to -11.9 ‰ (n=5). A pyrite sample from the host rock yielded a δ34S value of -12.7 ‰ (n=1).

A complex spatial distribution and a high degree of heterogeneity are among the main characteristics of mountain permafrost (Gruber and Haeberli, 2009). In Devaux cave the existence of permafrost can be related to a combination of two processes: (i) cave atmospheric dynamics and (ii) conductive heat transfer through the rock.

Devaux cave is characterized by mean air and rock temperatures lower than the external mean annual temperature (Fig. 5). The low cave temperatures in winter lead to an inward airflow and an associated negative thermal anomaly behind the entrance zone. In contrast, during summer the cold and dense air flows out of the cave due to the temperature difference between outside and inside air. The heat supplied to the cave by the river also influences the cave air temperature by exporting thermal energy from the cave during winter. Similar seasonal ventilation patterns have been observed in ice caves elsewhere (e.g. Luetscher et al., 2008; Colucci and Guglielmin, 2019; Perşoiu et al., 2021).

On the other hand, positive temperatures are observed both in the cave river and in the air at the entrance (Fig. 5), reflecting heat advected by water (river) and the influence of the external temperature (see Luetscher et al., 2008; Badino, 2010). The lack of correlation between the external and internal temperatures and the small temperature variability in rooms D, G, and K reflect their thermal isolation from well-ventilated cave parts. There, the apparent thermal equilibrium between the rock and the cave atmosphere (Trock=Tair) supports the notion that heat exchange is dominated by conduction through the bedrock.

The MAAT at the altitude of the cave is -0.04 ∘C (2017–2021) suggesting that the 0 ∘C isotherm is located close to the cave. Using an array of techniques (geomatic surveys, temperature monitoring, temperature at the base of the snowpack (BTS), and geomorphological and thermal mapping), Serrano et al. (2019) observed mean annual ground temperatures between -1 and -2 ∘C on the northern slope of the MPm, suggesting that discontinuous permafrost is present between 2750–2900 m a.s.l., with more continuous permafrost starting at 2900 m a.s.l. The orientation of the Gavarnie cirque, as well as the high slope angle and shadow from the surrounding peaks, favours the preservation of permafrost at lower elevations (e.g. Gubler et al., 2011).

Given the high thermal inertia of the rock, the permafrost temperature at depth is still under the influence of past climate conditions (e.g. Haeberli et al., 1984; Noetzli and Gruber, 2009), and, therefore, part of the current permafrost in the area could be inherited from previous colder times (e.g. Colucci and Guglielmin, 2019). In particular, the low mean annual temperatures recorded at PMBS in the late 19th century were favourable conditions for permafrost development. We surmise that the current permafrost could be inherited from colder periods of the Little Ice Age.

In well-ventilated ice caves hoarfrost is the most dynamic ice formation on seasonal timescales. The presence of perennial hoarfrost is, however, indicative of a continuously frozen bedrock and thus representative of caves within the permafrost zone (e.g. Luetscher and Jeannin, 2018; Yonge et al., 2018). In Devaux cave, perennial hoarfrost is observed in rooms where the bedrock is surrounded by small ice bodies (e.g. gallery close to room SPD; Fig. 6g). Devaux (1929) indicated the presence of ice crystals on the ceiling at the entrance of room D. In the same way, du Cailar and Dubois (1953) showed a schematic cross-section of room D, where ice crystals are present at the beginning of the room. These historical reports suggest these areas were probably more ventilated in the past, which favoured the hoarfrost formation. On the other hand, seasonal hoarfrost is present in ventilated galleries (A, B, C, and F and between SPD and J). Seasonal hoarfrost in room B and C and in the area between H to J disappears at the end of summer probably because of the heat delivered by the cave river, as recorded by the T5 sensor (Fig. 5).

The presence of permafrost in Devaux's catchment is supported by the absence of drips and/or seepage in the investigated cave passages (e.g. Luetscher and Jeannin, 2018; Vaks et al., 2020). Active drips and seasonal ice formations are limited to the first ∼ 15 m of the cave, as well as to the inner part (beyond room K). Mountain permafrost thus penetrates ∼ 350 m longitudinally from the eastern cliff of the Gavarnie cirque to the southern side of the massif. On the other hand, given the elevation of the cave and the topography above the cave, the current maximum permafrost thickness on the southern side of the MPm is ∼ 200 m (without taking into account the active layer).

The transparent and massive character of Devaux's cave ice, as well as the presence of CCCs, whose formation requires low congelation rates (Žák et al., 2004), suggests that this ice formed by the slow freezing of water dammed by ice at the spring. This model is consistent with the climate of the Gavarnie cirque, cave geomorphological observations, cave air and water temperatures, and historical reports. The cave water level can rise by several metres as indicated by slack water deposits upstream of the Brulle spring.

The distribution and characteristics of ice bodies in Devaux cave indicate that the hydraulic head rose by at least ∼ 15–29 m, which is the elevation of the ice bodies in rooms G, F, and K. This situation requires that all springs (including Porche) are blocked for a sufficiently long time to allow for complete freezing of these cave lakes. The lack of important unconformities in this massive ice (e.g. detrital layers), which are usually related to seasonal ablation (e.g. Luetscher et al., 2007; Stoffel et al., 2009; Hercman et al., 2010; Spötl et al., 2014), suggests that the ice deposit in room G is the result of a single flood event. In contrast, the small unconformities recognized in the ice body in room D suggest that several cycles of damming and subsequent ice formation cannot be discarded in the formation of this ice deposit.

These observations indicate that under the current climate (both in the cave and outside) only part of the water dammed in rooms F and E freezes during winter and spring. This strongly suggests that the ice bodies in Devaux cave must have been associated with colder and/or longer events of ponding and freezing than today, when the cave was effectively sealed from the outside for prolonged times. We hypothesize that the advance of a glacier on the steep slopes of Devaux's surroundings could have contributed to the blockage of the spring, leading to backflooding and the formation of large ice bodies in the cave. In the study area, such periods of glacier growth occurred during the Little Ice Age and/or the Neoglacial advance (González Trueba et al., 2008; García-Ruiz et al., 2014, 2020).

The freezing of a flooded cave passage cannot be explained by the advection of cold air alone. It is thus surmised that heat transfer through the host rock is a more plausible mechanism for the complete freezing of ponded water. The cave ice bodies, as well as the presence of cryogenic minerals, therefore record a long cold period or several shorter episodes. Although cryogenic minerals and in particular CCCcoarse are typically associated with permafrost thawing during warm spells (Žák et al., 2004; Richter et al., 2010; Žák et al., 2012; Luetscher et al., 2013), permafrost conditions prevailed during ice formation in Devaux cave. The water that feeds Devaux's springs infiltrated during late spring and summer from ponors at Lago Helado and/or surrounding poljes. However, the heat supplied by this water may have probably not been enough to thaw the frozen host rock. It is thus very likely that the host rock temperature was lower and/or the outlets remained closed for longer periods than today to allow for the complete slow freezing of the ponded water.

In Devaux cave, CCCs and CCG are still present within the ice (Fig. 9a, e). Worldwide, only very few in situ observations of coarse-grained cryogenic cave minerals are known (e.g. Bartolomé et al., 2015; Colucci et al., 2017). Du Cailar and Dubois (1953) reported the presence of gypsum crystals at ∼ 50 cm depth within the ice in Devaux cave. The first evidence of in situ CCCcoarse in cave ice was reported from Sarrios 6, an ice cave at 2780 m a.s.l. on the southern slope of the MPm (Bartolomé et al., 2015). Colucci et al. (2017) documented the presence of CCCcoarse in a small ice cave in the Italian Alps. Recently, Munroe et al. (2021) found CCCcoarse in ice of Winter Wonderland Cave (Utah, USA). Because of the abundance of cryogenic cave minerals, the size of individual crystals and aggregates thereof, as well as their different mineralogy, Devaux cave provides an additional opportunity for studying the origin of such cryogenic cave minerals.

CCG in Devaux cave represents, to our knowledge, the first occurrence of its kind in a carbonate karst terrain. So far, CCG have only been reported from gypsum karst areas in Russia and Ukraine (Korshunov and Shavrina, 1998; Žák et al., 2018, and references therein). In those caves, tiny gypsum crystals form during the rapid freezing of water. When ice sublimates in winter, these particles are released and accumulate as powdery deposits on the ice surface. Eventually, they partly dissolve during spring and summer due to the increase in cave air humidity, and they later recrystallize, forming a wide variety of crystal morphologies. CCG from Devaux cave shows features that do not correspond to those previously published from gypsum karst caves. In particular, (i) the Devaux cave CCG appears together with CCCcoarse crystals (≥ 5 mm in some cases, in rooms D and G), (ii) the (raft-like) gypsum crystals are large (Fig. 9b) and, in some cases, are still found within the ice (Fig. 9a) and surrounded by milky ice rich in air inclusions (Fig. 9a, e), and (iii) boulders are locally overgrown by gypsum (Fig. 9c).

Coarse-grained cryogenic cave minerals form in a semi-closed system when water freezes very slowly (Žák et al., 2004). Once supersaturation is reached, CCMs start to crystallize. The formation of gypsum crystals requires the presence of elevated concentrations of dissolved sulfate which may relate to (i) sedimentary gypsum deposits intercalated within carbonates (e.g. Sancho et al., 2004), (ii) the presence of hydrothermal water containing H2S related to hydrocarbons (e.g. Hill, 1987), or (iii) the oxidation of sulfides (e.g. pyrite) disseminated in carbonate rocks (e.g. Bottrell, 1991). In the case of Devaux cave marine evaporite rocks (e.g. of the Upper Triassic Keuper facies) and hydrocarbons are absent in the catchment of the cave. The most plausible explanation for the presence of dissolved sulfate in Devaux's water is the oxidation of pyrite present in the limestone (du Cailar and Dubois, 1953; Requirand, 2014).

δ34S values of gypsum (-11.9 ‰ to -15.8 ‰), pyrite (-12.7 ‰), and dissolved sulfate (-14.4 ‰ in drip water and -28.5 ‰ to -27.3 ‰ in Brulle spring water) are within the range of biogenic pyrite and differ notably from values of marine evaporites (10 ‰–35 ‰) (Seal, 2006). Thus, the δ34S values, together with the geological setting of the cave, support the hypothesis that disseminated pyrite in the host limestone is the main source of dissolved sulfate and subsequently of CCG. Only the dissolved sulfate δ34S values of Brulle spring are considerably more negative (-28.5 ‰ and -27.3 ‰). This may be a consequence of microbially mediated redox processes in the karst that discriminate against 34S (Zerkle et al., 2016; Temovski et al., 2018). Further studies on the microbiology of the cave may shed light on these mechanisms and how the local sulfur cycle may have changed in the recent past.

In gypsum caves, dissolved sulfate dominates over the bicarbonate, and the typical crystallization sequence during the freezing of water with high TDSs is gypsum → carbonate (commonly calcite) → celestine (Žák et al., 2018). In Devaux cave, however, bicarbonate dominates over sulfate, and our observations show that gypsum crystals partly nucleated on CCCcoarse. Accordingly, the crystallization sequence at Devaux cave is calcite → gypsum, taking place in a semi-closed system at low freezing rates.

The second aspect that makes the CCG in Devaux cave unique is the size and euhedral shapes of the crystals (Fig. 9b), which differ notably from the much smaller sizes of gypsum crystals (20–200 µm) and gypsum powders (1–30 µm) found in gypsum caves in Russia and Ukraine (Žák et al., 2018, and references therein). Another characteristic of CCC and CCG occurrences in Devaux cave is the presence of milky ice surrounding them (Fig. 9a, e) which seems to be related to the freezing process during the formation of cryogenic minerals in a subaqueous environment. Similar to that, CCCs were found within the ice and surrounded by bubbles in the Sarrios 6 ice cave (Bartolomé et al., 2015). However, the scarce presence of CCCs within the ice today, together with the very few sites where this topic is investigated, leads to a lack of studies about gas inclusions and CO2 degassing during CCC formation.

Finally, the presence of gypsum aggregates overgrowing blocks (Fig. 9c) supports the hypothesis of subaqueous gypsum formation. On the other hand, the absence of gypsum growing on the ceiling or on the walls allows us to discard its formation from seepage water followed by precipitation due to evaporation in the cave (e.g. Gázquez et al., 2017, 2020). In essence, all observations indicate that gypsum precipitated in a semi-closed subaqueous environment and has been preserved from later dissolution by the exceptionally dry environment of this ice cave. Gypsum precipitating from freezing waters has been also documented in the Arctic and the Antarctica (Losiak et al., 2016; Wollenburg et al., 2018) and has been proposed as a mechanism for gypsum formation on Mars (Losiak et al., 2016).