The firn cores analysed in this study were obtained during the field campaign of the ASUMA project (Assessing SUrface Mass balance of Antarctica) which aimed at quantifying the recent variations of surface mass balance along a transect in Adelie Land. The field campaign was carried out from December 2016 to January 2017 and was led by an expedition involving the French institutes IGE (Institut des Géosciences de l'Environnement) and IPEV (Institut polaire français Paul-Émile Victor). Twenty-five shallow firn cores, ranging from 20 to 40 m long, were drilled along a 1371 km-long loop (Fig. 1) starting from the Italian–French Robert Guillard-Cap Prud'homme station, located at sea level 5 km away from Dumont D'Urville on the continent and ascending to 2416 m above sea level at Stop0 (385 km from the Dumont D'Urville station).

The present study focuses on 9 firn cores that were drilled at 3 sites: D47, Stop5 and Stop0 (Fig. 1, Table 1). The cores were extracted using a portable electromechanical drilling system (Ginot et al., 2002). After bulk density measurements, the cores were sealed in polyethylene bags, stored in clean isothermal boxes and shipped to IGE cold room facilities in Grenoble, France.

Ground penetrating radar (GPR) data between the sites were collected using a MALÅ^® ProEx system fitted with a 100 MHz rough-terrain antenna, towed by one of the tractors from the traverse. The distance between the transmitter and the receiver was consistently maintained at 2.2 m. This setup was similar to the one employed on a traverse between Dome C and Vostok (Le Meur et al., 2018). Triggering was set at 2 s intervals, resulting in a radar trace every 8 m, based on the convoy's average speed of about 14 km h⁻¹. A time window of roughly 1.5 µs was applied, which allowed for a depth of investigation of about 125 m, given a 128-fold stacking and an average wave velocity of 0.2 m ns⁻¹. The resulting waveform was then sampled at a rate of about 10 times the nominal center frequency of the antenna (i.e 10 × 108 Hz = 1 GHz) thereby providing some 1500 samples per trace. A GNSS receiver mounted on the vehicle recorded the geographic position of each trace along the survey profiles. Pre-processing during acquisition included 128-fold stacking to enhance the signal-to-noise ratio. Post-processing steps involved a time-zero corrections, a zero-phase low-cut filter (devow) to eliminate continuous direct currents, and an “energy decay”; gain to counteract signal attenuation from volumetric spreading. Band-pass filtering was replaced by spatial averaging over 25 to 50 traces, which was deemed more efficient than traditional finite impulse response (FIR) filters due to the large dataset (∼ 150 000 traces). Time-to-depth conversion was performed by migrating radargrams using a vertical velocity profile for the radar wave.

Water stable isotope composition of the core ASUMA2016_7 was measured by laser spectrometry (Picarro L2130-i) on discrete samples at a 4 cm resolution at the Laboratoire des Sciences du Climat et de l'Environnement (LSCE) corresponding to a temporal resolution of 0.9 to 1.4 month. The estimated 1-sigma uncertainties are 0.2 ‰ for δ18O and 0.7 ‰ for δD (Grisart et al., 2022). All the other cores were measured using a Continuous Flow Analysis (CFA) facility paired with a Picarro L2130-i water vapor analyzer and a home-made vaporiser. We used a setup based on a similar principle as the one described in Dallmayr et al. (2016, 2025a) and presented in Petteni et al. (2025) to melt 30 mm × 30 mm sticks at an average rate of 2.5 cm min⁻¹. Only the water from the inner clean 18 mm × 18 mm part of the stick was kept for analysis, and split to feed the Picarro spectrometer on one side and a fraction collector on the other side. Isotopic calibrations were conducted at the beginning and end of each CFA day using 3 in-house standards themselves calibrated against Vienna Standard Mean Ocean Water (VSMOW) and Standard Light Antarctic Precipitation (SLAP): Ross (δ18O = -18.63 ± 0.05 ‰, δD = -144.8 ± 0.7 ‰), NEEM (δ18O = -32.89 ± 0.05 ‰, δD = -254.1 ± 0.7 ‰) and Adélie (δ18O = -40.55 ± 0.05 ‰, δD = -321.0 ± 0.7 ‰). By running check standards in some CFA runs, we concluded that the overall uncertainties given for discrete measurements also hold for CFA isotopic measurements.

The fraction collector was set to collect 3 mL samples, which corresponds to a depth resolution between 20 and 30 mm. Major ions concentrations ([Na⁺], [Cl⁻] and [SO42-]) were measured at LSCE using two Dionex ion chromatography systems working in parallel (ICS-5000+). For cations, separation was obtained using a CG16-4 µm (2 × 50 mm) guard column and a CS16-4 µm (2 × 250 mm) separation column, with isocratic runs of 30 mMol MSA as eluent at 0.17 mL min⁻¹. For anions, the eluent was 55 mMol KOH at 0.3 mL min⁻¹ and the system was equipped with an AG28-4 µm (2 × 30 mm) guard column and AS28-4 µm (2 × 150 mm) separation column. Both systems were running with Dionex DRS600 (2 mm) dynamically regenerated suppressors. By calculating the pooled standard deviation over a panel of 93 samples remeasured with a different set of standards, we estimated a 1σ uncertainty of 1.0, 1.2 and 0.8 ppb for [Na⁺], [Cl⁻] and [SO42-] respectively. The non-sea-salt SO42- (nssSO₄) was calculated following the classical approach: nssSO4= [SO42-] - 0.25[Na⁺] (Abram et al., 2013; Nardin et al., 2021). This relationship assumes a [SO42-]/[Na+] ratio value of 0.25 for marine aerosol and a negligible fractionation.

Additionally, artificial radionuclides resulting from atmospheric thermonuclear tests carried out from the 1950s to 1980s were deposited in Antarctica after being transported in the upper troposphere and stratosphere, creating unambiguous chronostratigraphic markers in the firn cores, peaking at years 1955 ± 1 and 1965 ± 1 AD (Jouzel et al., 1979; Pourchet et al., 1983, 1997, 2003; Le Meur et al., 2018). These layers were identified via the detection of ⁹⁰Sr, ²⁴¹Pu (deduced from ²⁴¹Am analysis) and ¹³⁷Cs. To perform this analysis, firn cores from Stop0 and D47 were processed and cut for artificial radioactivity measurements applying a method developed by Delmas and Pourchet (1977): core sections (100–250 g) were melted, weighted, acidified and filtered through ion exchange resin papers (MN 616 LSA-50 and LSB-50 strongly acidic cation and basic anion exchange resins, respectively) where all the radionuclides are trapped. Filters were first analyzed at IGE semi-underground laboratory by an alpha/beta Berthold LB-790 low-noise counter (Magand, 2009). A more detailed and targeted detection of the ¹³⁷Cs and ²⁴¹Am radionuclides was then performed using very low background gamma spectrometry in the Modane Underground Laboratory (Hodák et al., 2019).

The Paleochrono software serves as a probabilistic dating tool, akin to Datice and Icechrono1, with enhanced mathematical, numerical, and programming capabilities (Lemieux-Dudon et al., 2010; Parrenin et al., 2024). It has been originally designed to establish a unified and optimized chronology for archives from various paleoclimatic sites. It has then largely been used for ice core dating, both for deep ice cores (Oyabu et al., 2022; Bouchet et al., 2023) or for shallow firn cores (Oyabu et al., 2023) as also done in the present study. When applied to ice cores, Paleochrono relies on Bayesian inference for shaping the underlying background scenario of accumulation rate (A) and thinning of annual ice layers (τ).

To quantify the credibility of the background scenario, the accumulation function and chronological information (e.g., dated horizons, stratigraphic tie-points between cores) are expressed as a probability density, assumed to be Gaussian. The coherent age scale construction then relies on the Least Square optimization method and results in the best compromise between changes of the background scenarios within their associated uncertainties (imposed by the user) and respect of the chronological observations within their associated uncertainties.

Here we assigned a constant prior accumulation for the period covered by our records (∼ last 40 years) using available information from radionuclides-based dating, GPR-isochrones between Stop5 and D47, surface accumulation stakes or from a regional atmospheric model largely applied for Antarctica (Modèle Atmosphérique Régional, hereafter MAR, Gallée and Schayes, 1994; Agosta et al., 2019). Although some of these methods (beta counting and gamma spectrometry, stake measurements) provide accurate long-term average estimates of accumulation rate, we deliberately associate them with a relatively high uncertainty (50 % to 200 %) to allow Paleochrono to produce a variable accumulation rate in time based on the alternative chronological and stratigraphic constraints. Given that the present study focuses on firn cores, ice thinning is not a relevant parameter. Instead, the density evolution vs. depth was measured on each firn core and is implemented in the background scenario with negligible associated uncertainty.

The chronological constraints for our study are the followings. When the data are non-ambiguous, the beta counting and gamma spectroscopy provides dated horizons. We also incorporated dated intervals for each core, assessed by counting the number of seasonal peaks over regular 2 m intervals. Peaks and cycles were identified synchronously in both water isotopes and impurity records, which supports recording of annual layers in the firn cores (see Sect. 3). Note that for some sections only water isotopes were used for layer counting since there are some gaps in the impurity records especially for Stop0 (even no data on one of the 3 cores). When counting annual layers, we computed the uncertainty by adding 0.5 years each time the identification of a seasonal layer was doubtful. This approach follows the one used in Rasmussen et al. (2006) to build the GICC05 chronology and applied in the common Antarctic chronology exercise by Lemieux-Dudon et al. (2015).

Finally, we identified multiple stratigraphic tie-points in pairs of ice cores at each site which permit aligning the three records using both isotopic and chemical data. The selection of the stratigraphic tie-points between cores is based on the visual identification of common patterns or peaks by comparison of the chemistry and water isotope records. Depending on the confidence we have in the pattern correspondence between different cores, we assigned an uncertainty from 0.2 years (unambiguous matching) up to 2 years (when matching is disputable) for each tie-point. These choices are reasonable but subjective and constitute a documented limitation of the Paleochrono approach (Parrenin et al., 2024).

To help with the interpretation of our water isotopes records, we compared them to a simulated VFC δ18O signal inferred from isotope-enabled AGCM outputs. It follows the general approach of Sime et al. (2011) revisited by Casado et al. (2020). We use δ18O in precipitation and precipitation amount timeseries outputs of two isotope-enabled AGCMs, ECHAM6-wiso and LMDZ6iso extracted from the grid cells of our sites of interest. The precipitation amounts of ECHAM6-wiso and LMDZ6iso were scaled so that the average of the modeled accumulation rate is equal to the mean accumulation inferred at each drilling site. The effect of diffusion on the firn isotopic profile is calculated using the classical diffusion model after Johnsen et al. (2000) with a diffusion length calculated from the density profile, the accumulation rate and the mean annual surface temperature.

ECHAM6-wiso is the isotopic version of the ECHAM6 AGCM (Stevens et al., 2013), developed to explicitly simulate isotopic variations in the atmospheric hydrological cycle. The implementation of water isotopes in ECHAM6-wiso has been previously described by Cauquoin et al. (2019), and has since undergone several updates to improve its consistency with recent observations of water isotope behavior (Cauquoin and Werner, 2021). These updates include the incorporation of the isotopic composition of snow over sea ice, a revised supersaturation equation, and the assumption of wind-speed independence for kinetic fractionation factors during oceanic evaporation. The orbital parameters and greenhouse gases concentrations used in the ECHAM6-wiso simulation were set to the values corresponding to the modeled year. We have used daily values of ECHAM6-wiso model outputs from a simulation at high spatial resolution (0.9° horizontal resolution and 95 vertical levels) nudged to European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis (Hersbach et al., 2020). The ECHAM6-wiso 3D fields of temperature, vorticity and divergence as well as the surface pressure field were nudged toward the ERA5 reanalysis data every 6 h. A detailed description and evaluation of the ECHAM6-wiso simulation can be found in Cauquoin and Werner (2021).

LMDZ6iso (Risi et al., 2010) is the isotopic version of the AGCM LMDZ6 (Hourdin et al., 2020). We used LMDZ6iso version 20231022.trunk with the physical package NPv6.3, nearly identical to the atmospheric setup of IPSL-CM6A (Boucher et al., 2020) used for phase 6 of the Coupled Model Intercomparison Project (CMIP6, Eyring et al., 2016). The simulation was performed with the standard low resolution (LR) horizontal grid (2.0° in longitude and 1.67° in latitude, 144 × 142 grid cells) and 79 vertical levels, with the first atmospheric level located around 10 m a.g.l. The LMDZ6iso 3D fields of temperature and wind are nudged toward the 6-hourly ERA5 meteorological reanalysis data (Hersbach et al., 2020) with a relaxation time of 3 h, except below the sigma-pressure level equivalent to 850 hPa a.s.l., where no nudging is applied in order to let the model's physics and dynamics to be expressed in the boundary layer. Surface ocean boundary conditions are derived from monthly mean sea surface temperature and sea-ice concentration fields from the ERA5 reanalysis. The simulation, identical as in Dutrievoz et al. (2025), covers the period 1980–2017, from which we use daily outputs.

Leroy-Dos Santos et al. (2023) demonstrated that ECHAM6-wiso δ18O outputs exhibit a clear correlation with the isotopic composition of precipitation and water vapor at Dumont d'Urville station. The same performances have been made for LMDZ6iso by Dutrievoz et al. (2025). These findings highlight the model capacity to help in the interpretation of shallow firn cores.

For each given site and as expected, the average δ18O and δD values are similar for the 3 cores (Figs. 2 and S1 in the Supplement, Table 2). Yet, average values of δ18O and δD display different orders of magnitude between sites ranging from approximately -27 ‰ (D47) to -37 ‰ (Stop0) for mean δ18O and -209 ‰ (D47) to -290 ‰ (Stop0) for mean δD. The decreasing average δ18O and δD values from D47 to Stop0 through Stop5 were expected due to the gradual rise in elevation, decrease in temperature and the increasing distance from the coast (resulting in greater atmospheric distillation).

The observed cyclicity suggests that water isotopes record seasonal cycles with interannual variability in the amplitude. When comparing shallow cores from Stop5 or from Stop0, the isotopic records display similar variability in the seasonal variations at neighboring depths. This is not the case at D47, where the three isotopic records display only few common patterns (Fig. 2).

For each core, the variability recorded in the nssSO₄ profiles (Figs. 2 and S1) share strong similarities with those in Na⁺ or Cl⁻ profiles (not shown, data available in open repository). As we aimed at using these records to provide stratigraphic tie-points, we only concentrated on nssSO₄ records in the following. The mean nssSO₄ values are 17.3, 20.0 and 21.8 ppb at Stop0, Stop5 and D47 respectively (Table 2). The nssSO₄ records show variations at the same cyclicity as that observed in the water isotopes records. In particular, we observe a clear correspondence in the nssSO₄ and water isotopes cycles at Stop5 and Stop0 (Fig. 3). These cycles have a periodicity compatible with the annual amount of accumulated snow (Table 3) and despite δ18O profile being smoothed by diffusion, both water isotopes and nssSO₄ show concomitant maxima supporting the fact that seasonal cycles are well recorded at Stop0 and Stop5 in both nssSO₄ and water isotopes. Such annual cycles in both water isotope and impurity records have been identified in other sites with accumulation rates in the same order of magnitude like WAIS Divide ice core in West Antarctica (Jones et al., 2023). At D47, cycles are also observed in nssSO₄ but the correspondence between nssSO₄ and water isotopes is less clear, which do not support a good record of the seasonal cycles at this site. Finally, neither the Pinatubo nor the Agung eruptions can be unambiguously identified in the nssSO₄ records. A slight increase in the average value of nssSO₄ can still be observed between 10 and 12 m at Stop0 and Stop5 and between 7 to 12 m at D47 and may include the signature of the Pinatubo eruption which is visible over years 1992 and 1993 in some Antarctic ice cores (Cole-Dai and Mosley-Thompson, 1999; Sigl et al., 2013; Emanuelsson et al., 2022). Because we do not find an unambiguous nssSO₄ signal to attribute to the Pinatubo eruption, no volcanic horizon was included in the Paleochrono runs.

Paleochrono requires a prior estimate of the accumulation rate to derive a chronology at each site. Table 3 compiles all the accumulation data available derived from beta counting and gamma spectrometry, stakes measurements, and simulation outputs from MAR forced by the ERA5 reanalysis covering the period 1979–2020 (same simulation as in Davrinche et al., 2024). At Stop5 and Stop0, the different estimates are close to each other, within 10 %. Our results neither depend on the exact value of the Paleochrono prior accumulation rate among the estimates mentioned above nor on the value of the uncertainty attached to this prior accumulation rate (a final value of 50 % has been used for the final chronology presented below for Stop0 and Stop5 while at D47 the uncertainty was set to 200 % due to the high variability of the different estimates).

At D47, we observe strong differences among the estimates of accumulation rate. The beta counting and gamma spectrometry data suggest depth of 19.10 ± 0.5 and 16.10 ± 0.5 m for years 1955 and 1965, but the ages obtained from these techniques are less reliable than at Stop0 due to uncomplete gamma spectrometry data caused by technical issues during data treatment. The two closest stake measurements differ by a factor of more than 2, despite being only 2 km away from one another. This reflects the strong short scale variability of accumulation rates in this region. We thus believe that the MAR output at the D47 grid point is not a faithful estimate of the accumulation at the exact location of the drilling site.

We thus explore an alternative way to estimate the accumulation rate at D47. Two GPR reflectors identified between Stop5 and D47 were identified at 17.90 and 20.44 m at Stop5 and at 10.89 and 13.38 m at D47. Using the ages obtained at Stop5 from Paleochrono with much more faithful chronological constraints than at D47 (40.2 and 46.1 years), we deduce accumulation rates of 157 ± 21 and 159 ± 21 mm w.e. yr⁻¹, with uncertainties derived from isochrones calculated as in Le Meur et al. (2018) (Table 3). We used the average of 158 ± 21 mm w.e. yr⁻¹ as input for Paleochrono. These estimates are close to the estimates inferred from the beta counting and gamma spectrometry. For this site, there is a huge difference between the values of the accumulation rate obtained as output of paleochrono for sensitivity tests performed with uncertainties of 50 % or 200 % in the background accumulation rate (Table 3).

For each drilling site, we construct stack series on a coherent age scale using the Paleochrono tool together with dating and stratigraphic constraints following the method described in Sect. 2.3. In addition to the absolute dating constraints from beta counting and gamma spectrometry on Stop0, dating constraints relies on annual layer counting in 2 m-intervals. As discussed above, the annual layers are clearly visible in both the nssSO₄ and water isotopes records at Stop0 and Stop5, whereas it is more ambiguous at D47 hence leading to larger associated uncertainties in the Paleochrono input files (Table S1). For combining the different records, we use stratigraphic tie-points based on the resemblance of the patterns in water isotopes (Table S2) and nssSO₄ (Table S3). Because this approach is subjective, we performed several sensitivity tests using either only the nssSO₄ tie-points, only the water isotopes tie-points, or both the nssSO₄ and water isotopes tie-points. We then compared at each site the stacks obtained for different configurations (Fig. 4): (1) on the original depth scale (i.e. without any dating constraint), (2) on an age scale without any stratigraphic tie-points between the cores, (3) on an age scale with only nssSO₄ tie-points, (4) on an age scale with only water isotope tie-points and (5) on an age scale with both nssSO₄ and water isotope tie-points.

For the different sites, we do not observe large differences in the stacked water isotope records after matching records of nssSO₄, water isotopes or both nssSO₄ and water isotopes. As expected, the uncertainty calculated by Paleochrono is the largest when there are no stratigraphic tie-points and is the largest for D47 where the confidence on layer counting is low (Fig. S2). At Stop5, the 1σ dating uncertainty decreases from ∼ 2.5 years (no tie-point) to ∼ 1.5 year (only nssSO₄ tie-points) or slightly less (water stable isotopes and nssSO₄ tie-points).

We have a few ways to check the validity of the chronology calculated by Paleochrono. First, we find a good agreement between the a priori and output accumulation rate at Stop5 and Stop0. It means that the layer counting is in good agreement with the value of the background accumulation rate. Second, at Stop0, the chronology also uses two tie-points from beta counting and gamma spectroscopy which are hence in agreement with annual layer counting and background accumulation rate. For Stop5, no beta counting and gamma spectroscopy has been used in the chronology construction but, as for Stop0, the annual layer counting is associated with small uncertainty and there are numerous tie-points (nssSO₄ and water isotopes) between the 3 cores. A third support for the chronology of the stack records come from the nssSO₄ increase which could be related to the Pinatubo eruption. Figure 2 shows the depth range corresponding to the period 1992–1993 on each core. At both Stop0 and Stop5, it corresponds to a depth range with a level of nssSO₄ which is relatively high.

At D47, there is a very poor agreement between the a priori and output accumulation rate when a 200 % uncertainty is used for the background scenario. This reflects a disagreement between layer counting and background accumulation rate. We have however very low confidence in layer counting at D47 (see above). Note that we find correct dates for the beta counting and gamma spectroscopy peaks not used in the chronology construction when we use the chronology built using an uncertainty of 30 % on the background accumulation rate. This is however not enough to produce a robust chronology for this site.

Finally, Paleochrono calculates an uncertainty (1σ) for the chronology of the three stacked records (Fig. 5). This uncertainty is relatively small; 1σ is less than 1 year for Stop0 and less than 1.5 year at Stop5 which reflects the good confidence in the layer counting and agrees with variability of the chronology when running sensitivity tests with Paleochrono (choice of tie-points, uncertainty on the background accumulation rate). At D47, 1σ is less than 6 years which seems underestimated when comparing the chronology obtained with an uncertainty of 50 % or 200 % on the background accumulation rate (Fig. S3). This underestimation is linked to effects not yet implemented in Paleochrono such as the fact that the biases in background accumulation rate or layer counting for the neighboring cores are not independent.

Figure 5 presents the δ18O signal at the three sites as obtained from the virtual firn core calculation using either the ECHAM6-wiso or the LMDZ6iso model. In both cases, the virtual water isotopic records display clear seasonal cycles which persist despite diffusion in the firn.

In general, there is an excellent agreement in the variability of the water isotopic records of the VFC obtained using either the ECHAM6-wiso or LMDZ6iso output even if the mean δ18O value is 5 ‰ lower using LMDZ6iso than ECHAM6-wiso at Stop0 (Table S4). Finally, we observe that the seasonal cycles are less visible in the D47 VFC when compared to the VFC for Stop0 and Stop5. This is due to the significantly lower accumulation rate at D47.

We observe (Table S4) similar mean δ18O values at D47 and Stop5 for both models (D47: -27.2 ‰ and -27.4 ‰; Stop5: -32.5 ‰ and -33.7 ‰ for ECHAM6-wiso and LMDZ6iso, respectively) compared to the stack of measured δ18O records (D47: -26.9 ‰; Stop5: -34.1 ‰). The agreement is less good at Stop0 where we find significant differences between the mean δ18O of the stacked data (-36.7 ‰) and the modeled values (-34.2 ‰ for ECHAM6-wiso and -39.5 ‰ for LMDZ6iso). The model-data discrepancies can at least partly be related to the difference in temperature between model outputs and measurements obtained from the automated weather data or borehole temperature (Table S4).

The amplitude of the seasonal cycles is in general lower in the stack than in the VFC, which is expected since stacking series tends to decrease the variability in case of slight misalignments. At Stop5, the mean seasonal δ18O amplitude is larger in the modeled series (5.1 ‰ and 5.8 ‰ for respectively ECHAM6-wiso and LMDZ6iso) than in the stacked series (3.2 ‰). The same is observed for Stop0 with values of respectively 5.4 ‰ and 9.9 ‰ for the mean seasonal cycle in the VFC based on the ECHAM6-wiso and LMDZ6iso models, and 3.6 ‰ for the stacked series. At D47, the amplitude of the variability is comparable between the modeled series (2.0 ‰ and 2.2 ‰ for ECHAM6-wiso and LMDZ6iso respectively) and the stacked series (2.7 ‰). This result is surprising as D47 features strong katabatic winds that induce erosion and upper mixing not included in the VFC but which are expected to decrease the amplitude of the isotopic variability.

Despite the general good visual agreement between model and stacked ice core variability on the whole series, some discrepancies are visible over some sections. As an example on Stop5, we see that while the stacked data displays a clear maximum around year 1989, the modeled signal is much more muted. The strongest disagreement between modeled and stacked isotopic records is visible on the upper section. The upper part of the records appears indeed much smoother in the stacked data than in the modeled VFC isotopic records. This cannot be attributed to the stacking itself removing some of the real variability as individual cores also feature a reduced variability compared to the VFC (Figs. S3, S4). The measurement technique is also not responsible for such smoothing effect because similar effects are seen when the measurements are performed through continuous flow analysis or on discrete samples. In the VFC isotopic records, the simulated diffusion follows the approach of Johnsen et al. (2000) and remains weak for the upper part of the core. However, surface firn is fragile and porous making it more susceptible to further diffusion during transport and storage (Dallmayr et al., 2025a) not considered in the VFC construction. Finally, the upper part of the firn is also very sensitive to wind pumping, snow mixing and redistribution, snow redistribution expected to be particularly strong at windy sites. It is difficult to discriminate between these three processes from the time series of isotopic composition alone, so we have decided not to account for them in the construction of the VFC signals. Still, they can explain the rapid loss of variability of the isotopic records at the top of the shallow cores (Casado et al., 2020).

At D47, we observe only 31 maximums in the stacked δ18O data while the record is expected to last 39 to 62 years from our chronology construction (Table 4), whereas on Stop5 and Stop0, we see one clear δ18O maximum every year except for a few years where we better see a shoulder in the δ18O record. The difference in the number of peaks between Stop0, Stop5 and D47 is the combination of the significantly lower amount of accumulation at D47 compared to Stop0 and Stop5, as well as large amount of erosion or mixing of surface snow by wind (Picard et al., 2019; Zuhr et al., 2023) at D47 which is one of the places with the strongest katabatic winds in Antarctica (Kodama et al., 1985; Davrinche et al., 2024). This effect can be seen through the difference of the local roughness between the 3 sites (Table 4): the surface roughness is documented by calculating the root mean square of surface deviation on a circle of 300 m around the drilling site as indication of the surface roughness obtained from the Reference Elevation Model of Antarctica (Howat et al., 2019). The root mean square of surface deviation is 2.7 m at D47 compared to values of 0.06 and 0.1 m respectively for Stop0 and Stop5. Such topographic effects were already shown to create δ18O variations which cannot be related to climatic variability (Dallmayr et al., 2025b). We thus conclude that D47 is not a good site to provide annually resolved climate and/or atmospheric water cycle reconstructions from water isotopic records in firn or ice cores using only a stacking of 3 neighboring cores. In the following, our analysis focuses on the records obtained at Stop0 and Stop5 only.

At Stop5 and especially at Stop0, there is a strong resemblance between the stack and the VFC records even at the seasonal scale. This good agreement is a strong validation for the performance of the ECHAM6-wiso and LMDZ6iso in this region. Still, we observe small temporal shifts (< 1 year), especially since our stack maximums do not always coincide with summer. This mismatch is within the uncertainty of the chronology and is inherently linked to the method used here for the chronology construction which does not impose any tie-points between peaks in δ18O or nssSO₄ and summer periods. Numerous studies however already used the nssSO₄ peaks as markers for austral summer when dating Antarctic ice cores (Emanuelsson et al., 2022; Steig et al., 2005). Indeed, the main source of nssSO₄ is dimethyl sulfide (DMS) produced by phytoplankton blooms starting in December with the break-up of sea ice (Abram et al., 2013). Maxima of δ18O are also routinely used as markers for summer periods for firn and ice cores when annual layers are clearly recorded (e.g. Emanuelsson et al., 2022; Vega et al., 2016). Here, we follow these approaches and modify the chronologies obtained from Paleochrono by forcing the nssSO₄ peaks and/or the δ18O maxima to coincide with summer periods. At Stop5, because the nssSO₄ profiles are complete, the alignment can be performed through matching the nssSO₄ peaks from the stack record with the summer periods (1 January) starting from the most recent period (Fig. S5). When plotting the stack of water isotopes on this new chronology, we find a good match with the VFC δ18O record with both maxima in the stack δ18O and maxima in the VFC δ18O records occurring during the austral summers. At Stop0, because of the lack of nssSO₄ data, we matched the δ18O maxima with maxima in the VFC δ18O records (list of tie points in Table S5) resulting in a modification of the stack chronology by less than 1 year which does not significantly modify the uncertainty obtained from layer counting. Note that when we do the matching through δ18O variations at Stop5, we obtain similar results than for the matching through nssSO₄ maximums (Fig. 5).

Table S6 compiles the correlation coefficients between individual cores or stacked data at each site and VFC δ18O series after modification of the chronology to have nssSO₄ and δ18O peaks during summer. We find correlation coefficients r between the stack and the VFC of 0.58 and 0.54 (ECHAM6-wiso) and 0.55 and 0.60 (LMDZ6-iso), at Stop0 and Stop5 respectively, which further motivates using such stacks to try to recover climatic information in the following. Because both LMDZ6iso and ECHAM6-wiso VFC models display similar variability in the VFC, we only show analyses performed with ECHAM6-wiso VFC in the following section. Same conclusions can be drawn from LMDZ6iso VFC.

In a previous study performed on a single firn core in coastal Adélie Land, Goursaud et al. (2019) found no clear correlation between δ18O and temperature at the seasonal scale. Later, Leroy-Dos Santos et al. (2023) provided 2 years of atmospheric monitoring of water vapor δ18O at Dumont d'Urville station and showed that if δ18O of water vapor is not strongly correlated to temperature, there is an excellent agreement between the δ18O of water vapor modeled by the ECHAM6-wiso model and the one measured on-site. This result, together with the good performance of the ECHAM6-wiso model when simulating climatic parameters, shows that the ECHAM6-wiso model properly reproduces the climatic signature of the water vapor isotopic signal. The agreement was less good for the precipitation δ18O. When using the VFC δ18O signal calculated from ECHAM6-wiso outputs, the agreement was poor with the δ18O signal measured in a single firn core (Leroy-Dos Santos et al., 2023). One of the reasons invoked for this mismatch was stratigraphic noise. In our study, we can test this hypothesis since we removed at least part of the stratigraphic noise by stacking three records at each site. In the following, we thus test how our stacks can be used to get representative insights of the regional climate over the past decades.

Since both the VFC signals and stack records display clear seasonal cycles and interannual variability, we concentrate on the possible identification of extreme warm summer in our records. First, we address the possible link between the extreme values in δ18O and in temperature. To infer how the stacks can be used to identify years with particular/extreme climatic patterns, we compare how the 12 warmest summers of the period 1979–2016 can be identified in the stacks. This comparison is done for Stop0 where we find the best visual agreement (and the highest correlation) between the stack and VFC δ18O records. Table 5 (right part) indicates the 12 years with the warmest summer periods according to the ERA5-derived temperature. From these, only 6 are also within the 12 highest summer peaks in the stack δ18O curve. Even considering a possible uncertainty of our chronology by 1 or even 2 years, we cannot match the 12 warmest summer periods with the 12 highest summer peaks in the stack δ18O curve. This confirms that the stack δ18O curve cannot be used to robustly identify the extreme years in terms of temperature variations. This is not unexpected since, as mentioned above, δ18O is not a direct indicator of temperature and cannot directly be used to identify maximum or minimum temperature levels; it is also influenced by the seasonality of the precipitation, the precipitation intermittency, the climate at the evaporative source and the trajectory of the vapor precipitated.

To address the issue of the intermittency of the precipitation, we then compare the 12 highest peaks in precipitation-weighted monthly temperatures with the peaks in the stack δ18O and only find 7 peaks in common (Table 5, left part). This means that precipitation intermittency is not the only reason why δ18O archived in snow and ice in coastal Adélie Land is not a perfect indicator of past temperature. A second limitation is linked to the influence of atmospheric processes within the water cycle and atmospheric transport on the isotopic composition of precipitation. To account for this effect, we compare the VFC δ18O record with the stack δ18O record. In this case, 8 of the 12 most prominent δ18O peaks in the VFC isotopic curve are among the 12 most prominent δ18O summer peak in the stack record (Table S7). Again, it is not possible to match the 12 highest δ18O maxima of the VFC and the stack within the dating uncertainty.

Another way to compare the stacked δ18O record to the different series (VFC δ18O, precipitation weighted temperature, temperature) over the summer peaks is to look at the correlation between the values of the summer peaks between the different series. The correlation is the highest (r= 0.66) when comparing the stacked δ18O record and the VFC δ18O and the lowest (r= 0.49) when comparing the stacked δ18O record and the temperature variations (Table 6). We again conclude that using the model derived VFC δ18O series improve the agreement with the stacked δ18O, which is expected since it considers both the intermittency of the precipitation and the effects of source origin and air transportation on the δ18O record. Still, a correlation of r= 0.66 shows that part of the δ18O variations are not captured by the VFC calculations.

If stacking 3 cores at one site is expected to remove the stratigraphic noise and increase the resemblance between the isotopic records and the VFC δ18O, discrepancies remain which can be attributed, at least partly, to the variability from one core to the other (to be attributed to deposition or post-deposition effects). Except for four very prominent summer periods identified in Table 5 for stack δ18O, temperature, weighted temperature and VFC δ18O, the variability in the seasonal amplitude of the δ18O record among individual firn core leads to a muted stack signal with a relatively large uncertainty envelope. Stacking more than 3 cores at one site may help reducing this envelope and better resolve the interannual variability. Another reason for the discrepancy between stacked δ18O and VFC δ18O series could also be that the models do not faithfully reproduce the interannual variability in δ18O of precipitation as observed in Leroy-Dos Santos et al. (2023). In particular, the AGCM may not fully reproduce the amplitude of the synoptic events which strongly contribute to the δ18O signal in coastal Adélie Land (Leroy-Dos Santos et al., 2023).

Despite these limitations in faithfully inferring the interannual variability of the δ18O from a stack of 3 cores at the same site, the overall good resemblance between the stack and the VFC isotopic curves support the use of such stacked cores to reconstruct first order trends in δ18O over the last decades or centuries. At Stop5 and Stop0, the overall good agreement between the stack and the VFC (Fig. 5), in spite of the unavoidable dating uncertainty, is a promising result. At coastal sites, where pre- and post-depositional processes significantly amplify δ18O variability, reconstructions of δ18O series from stacking of several records are thus crucial. They enhance the representativeness of the site, may improve our understanding of the processes driving non-climatic δ18O variability, and serve as valuable validation data for climate models. Such results open the door toward model-data comparison in coastal Antarctica to better understand how much climate signal is preserved and archived in coastal ice and firn cores.