We use the polar-oriented regional atmospheric model MAR version 3.12 to assess the variability of snowfall and temperature in Antarctica. Version 3.11 of this model has shown good performance when representing the surface mass balance (defined as the total precipitation minus sublimation and runoff; Agosta et al., 2019), temperature (Kittel et al., 2021), and cloud (Kittel et al., 2022) variability. The version used here (3.12) differs mainly in the temperature transition of rainfall to snowfall (both can now co-occur in the model) and the correction of the snowpack temperature (Antwerpen et al., 2022) as well as albedo tuning for the dense Antarctic snow. The simulation domain is on an Antarctic polar stereographic projection (EPSG:3031) with a horizontal resolution of 35×35 km. The MAR model is nudged to ERA5, the latest re-analysis product of the European Center for Medium-Range Weather Forecasts, which gives the most accurate representation of both temperature and surface mass balance in Antarctica among climate re-analyses (Gossart et al., 2019). The use of the MAR regional model provides further detail in the near-surface atmospheric layers and spatial scale compared to ERA5. The simulation used in this study covers the period 1979–2020 (42 years). There is a notable lack of direct snowfall observations in the low-accumulation regions of Antarctica, which hinders our ability to directly evaluate the model for this parameter. The modelled surface mass balance can be evaluated against the observed accumulation, but this includes several processes, ranging from snowfall to snow drift, the evaporation of falling snow, and evaporation–condensation on the snow surface, each of which comes with its own uncertainties. We evaluate the performance of MARv3.12 in representing the temperature and surface mass balance in Figs. A1–A4.

Although the temperature recorded in water isotopes is imprinted at the condensation level (Jouzel and Merlivat, 1984), we chose to use the 2 m air temperature for simplicity because condensation levels change both spatially and temporally. Studies using water isotopes usually bypass condensation to surface temperature changes by directly calibrating the isotope–temperature slope with the 2 m temperature in most cases (e.g. Jouzel et al., 2007; Stenni et al., 2017) or applying a ratio of temperature changes that would be amplified at the surface (e.g. Jouzel et al., 2003). If we used the condensation-level temperature, the difference from the climate normal would depend on the level of precipitation formation that is vertically spread on the atmospheric column, making the comparison more complex. With the condensation temperature, we would expect weaker seasonal cycles because winter surface cooling is amplified by a strong inversion, but the long-term temperature variability may not change much, as implied by deglaciation simulations (Liu et al., 2023). Choosing the 2 m temperature also enables comparison with available observations, and this is the level considered in many paleotemperature reconstructions.

In this study, we use the average temperature calculated with the arithmetic mean, 1T=∑day=1ndaysTdayndays, and the snowfall-weighted temperature, defined as the weighted average of the temperature with the daily snowfall rate as the weighting coefficient: 2Tw=∑day=1ndaysTday×SFday∑day=1ndaysSFday. In both cases, Tday refers to the temperature on a given day, and SF_day is the snowfall on the same day. Temperature averages can be computed for the entire study period (ndays=15341), a year (ndays=366), or on a given day of year (ndays=42, used for the climate normals T‾ and Tw‾). We define the climate normal temperature for each model grid cell as the average for the same day of year across the 42 years, and we subsequently apply a 30 d rolling mean to smooth the signal.

To quantify the difference in temperature associated with snowfall, we define the snowfall-weighted temperature difference as 3ΔT=Tw-T. This metric was previously described as “precipitation-weighted biasing” in Sime et al. (2008), but we choose not to call it a bias to avoid confusion with the modelling temperature bias, which refers to the difference between the modelled and observed temperatures here.

Decomposition of the daily temperature Tday into the climate normal temperature on that day (T‾day) and the daily anomaly with respect to the climate normal (T′day) allows us to separate the seasonal and non-seasonal effects of snowfall weighting: 4ΔT=∑day=1ndaysT‾day+Tday′×SFday∑day=1ndaysSFday-∑day=1ndaysT‾day+Tday′ndays5ΔT=+∑day=1ndaysT‾day×SFday∑day=1ndaysSFday-∑day=1ndaysT‾dayndays︸seasonal+∑day=1ndaysTday′×SFday∑day=1ndaysSFday-∑day=1ndaysTday′ndays︸=0︸non-seasonal. A summary of the abbreviations used is given in Table 1.

The temperature during snowfall is statistically higher than the average temperature on the same day, as shown by the mostly positive temperature anomalies (Fig. 1). Despite a wide distribution of temperature anomalies at any given snowfall rate, the average temperature anomaly increases with snowfall. Over Antarctica, there is a +5 °C increase between snowfall rates of 0 (no snowfall) to 1 kg m⁻² d⁻¹, and a gradual increase of another +5 °C as snowfall rates increase from 1 up to 100 kg m⁻² d⁻¹. Conversely, days without snowfall are 2 °C cooler than average. For major regions of Antarctica, similar patterns are modelled, with negative temperature anomalies on days without snowfall and increasing temperature anomalies of up to +10 °C for snowfall exceeding 50 kg m⁻² d⁻¹. The main difference for East Antarctic high elevations (Fig. 1f) is that the temperature anomalies reach the +5 °C threshold for snowfall rates of less than 1 kg m⁻² d⁻¹.

The positive anomaly associated with snowfall affects all of the Antarctic continent, although with varying intensity, as shown by the map of differences between the snowfall-weighted temperature and average temperature (denoted ΔT, Fig. 2; abbreviations listed in Table 1). Over the entire Antarctic continent and ice shelves, ΔT averages 5.4 °C. While coastal regions and West Antarctica show ΔT values of 0 to 5 °C, the East Antarctic Plateau and ice shelves reach ΔT values of up to 10 °C. Interestingly, ΔT is highest in large topographical depressions such as the Recovery (20° W, 80° S), Aurora (115° E, 75° S), and Wilkes (150° E, 70° S) basins or the Byrd glacier catchment inland of the Transantarctic Mountains (150° E, 80° S). On the other hand, over steep slopes and ridges, the snowfall-weighted temperature is relatively close to the average temperature. Note that hoar frost is computed separately from snowfall in the model and occurs in cold conditions (Schlosser et al., 2016). Therefore, ice accumulated by hoar frost can mitigate the warm conditions associated with snowfall, but this is not depicted by ΔT, which accounts only for snowfall. Another modelling study by Sime et al. (2008) showed ΔT values of up to 10 °C in East Antarctica and lower values of about 5 °C in West Antarctica for the present day, consistent with the results presented here. Our results mostly differ in the coastal regions, which may relate to the increased resolution used in this study or differences in modelling the physical processes of the katabatic-affected Antarctic slopes. In this work we focus on the quantitative temperature increase, but the degradation of the climatic signal due to the loss of correlation induced by precipitation intermittency has been treated in similar studies (Sime et al., 2011; Casado et al., 2020).

To better understand the temperature anomaly associated with snowfall at the Antarctic scale, we analyse 10 sites where the impact of extreme precipitation events on the total accumulation has previously been discussed (Turner et al., 2019), and we show the temperature–snowfall relationship at these locations (insets in Fig. 2). There are strong differences between sites located near the coast (Law Dome, East and West Peninsula, Gomez) compared to high-elevation sites on the East Antarctic Plateau (High Plateau, Dronning Maud Land, Dome C): high-elevation sites are characterized by low snowfall rates, but events causing a snowfall larger than 1 kg m⁻² d⁻¹ are accompanied by a temperature increase of more than 10 °C on average, and commonly up to 20 °C (Table 2). These high-elevation sites with low snowfall and large temperature anomalies are responsible for the sharp increase in temperature associated with low snowfall rates (Fig. 1f). Most locations reach temperature anomalies close to 10 °C at their respective maximum snowfalls (WAIS Divide, Gomez, Law Dome), except for sites where dry warming usually occurs, driven by foehn (East Peninsula) or katabatic adiabatic compression (coastal slopes). Each site shows an increasing temperature trend with snowfall rate, with steeper slopes for sites with low accumulation (Fig. 2). Overall, days with snowfall are statistically warmer than average conditions, and they are increasingly so for higher snowfall rates at a given location.

The analysis of yearly snowfall-weighted temperature (yTw) and “true” yearly temperature (yT, Fig. 3) further supports that the effect of snowfall weighting is not constant and may depend on local parameters including the temperature, but also probably the precipitation regimes. Importantly, yTw is not linearly related with yT, suggesting that changes in the annual temperature are not matched by proportional changes in the snowfall-weighted temperature. This relationship may also change depending on whether we average annually or at other time resolutions. Besides, any given yT is matched with a large distribution of yTw, which means that snowfall weighting induces variability in the temperature.

Snowfall events, in particular large precipitation events, are an important source of variability in the Antarctic climate (Turner et al., 2019). As there is a clear link between snowfall intensity and the temperature anomaly, the variability of snowfall translates into a variability of temperature. We first investigate the seasonality of temperature anomalies associated with snowfall (Fig. 4) by considering the climate normal snowfall-weighted temperature (Tw‾, see Sect. 2 for details on computation) and the climate normal temperature (T‾). Tw‾ differs from T‾ by 3 °C in summer and up to 8 °C in winter on average on the Antarctic ice sheet. The larger difference in winter results from the attenuation of near-surface temperature inversion during the passage of precipitating atmospheric systems. Indeed, T‾ reaches extremely low temperatures in winter, driven by the strong surface radiative cooling (Connolley, 1996; Hudson and Brandt, 2005; Genthon et al., 2021). While it contributes to large variability in winter temperatures compared to the summer season (Ricaud et al., 2017), snowfall consistently occurs under warm conditions. The seasonal amplitude of Tw‾ is thus 20 % lower than that of T‾ on average in Antarctica (Fig. 5). The reduction in seasonal amplitude occurs consistently over the Antarctic continent and is strongest on coastal slopes.

In winter, cyclogenesis is slightly higher (Uotila et al., 2011) and atmospheric blocking is more frequent (Wille et al., 2021; Scarchilli et al., 2011), increasing the probability of poleward moisture transport and Antarctic snowfall. This results in higher snowfall in the winter months at the Antarctic scale (Fig. 4b) and causes the snowfall-weighted temperature to be influenced more by the winter season, when the snowfall-related warming is the strongest.

We decompose the contributions of the seasonal distribution of snowfall (Fig. 6a) and the event-related daily temperature anomaly (Fig. 6b) to ΔT. As most of Antarctica receives more snowfall in winter (Palerme et al., 2017), the difference induced by seasonality averages -0.7 °C over the ice sheet but rarely exceeds -3 °C. On the contrary, snowfall-event-related warming causes a difference of +6.1 °C and dominates the difference between snowfall-weighted and all-day temperatures. These results are also in good agreement with the frequency decomposition of Sime et al. (2008), who showed that most of the ΔT signal was in the synoptic signal, comparable to the daily anomaly of the temperature used here. Although the seasonal signal is mostly negative in Fig. 6a, we note weakly positive ΔT in Victoria Land, where Sime et al. (2008) also found positive ΔT for their seasonally band-passed signal. The extent of this positive region is greater in Sime et al. (2008), extending well within continental East Antarctica, but this may be related to the discrepancy in modelled seasonal precipitation for the dry East Antarctic plateau, with the summer precipitation maximum causing positive ΔT in Sime et al. (2008) as opposed to the winter maximum causing negative ΔT here (Figs. 5 and 6, High Plateau site). In another study using the same method, Masson-Delmotte et al. (2011) find much stronger ΔT over the East Antarctic plateau, which is linked to seasonal effects on temperature. However, this difference is likely to have emerged from the ERA40 re-analysis used, which was documented as having a lack of winter precipitation and cyclone intensity in winter in the driest regions of Antarctica (Bromwich et al., 2007; Marshall, 2009), which leads to unrealistically large seasonal effects of precipitation weighting.

The dampened seasonal amplitude of the snowfall-weighted temperature results from averaging across 42 years, smoothing out interannual variability. While the yearly averaged temperature (yT) is relatively stable over time (Fig. 7a), the interannual snowfall is highly variable, especially for the winter season (Casado et al., 2020; Turner et al., 2019). This causes the yearly snowfall-weighted temperature (yTw) to vary significantly from one year to another, with the standard deviation increased by +80 % on average over the ice sheet compared to yT (Fig. 7d). The variability is especially increased in Dronning Maud Land and the eastern part of West Antarctica facing the Ronne Ice Shelf, where the interannual variability of yTw can be 200 % larger than the variability of yT. Previous studies highlighted that the large variability of winter temperatures causes the winter season to be dominant in the interannual temperature variability, as a warm year is usually due to a warm winter, which is often accompanied by one or multiple snowfall events in winter (Persson et al., 2011; Casado et al., 2020; Servettaz et al., 2020). Despite the reduced seasonality of the snowfall-weighted temperature and the tendency to oversample warm winters, its interannual variability is increased. In other words, the temperature averaged equally over all days of a year is more stable than the temperature taken during snowfall only. This is because of the sporadic nature of snowfall, which causes the temperature to be subsampled on a limited number of days (Fujita and Abe, 2006; Turner et al., 2019) at random times of the year and with a temperature bias which depends on precipitation intensity. On the interannual scale, the variability of yearly averaged temperature is thus enhanced when weighting with snowfall.

Water isotopes are used in Antarctic paleoclimate studies as a proxy for temperature due to the relationship between air temperature at condensation and the isotopic ratio in precipitation (Stenni et al., 2017; Dansgaard, 1964). Ice cores retrieve material accumulated over time onto the ice sheet, where the ice mass contribution depends on the snowfall. Therefore, snowfall-weighted temperatures provide an analogue to the isotopes we can expect to measure in an ice core. We thus discuss how the temperature during snowfall may affect δ18O at the time of snow deposition, prior to post-deposition effects that occur later and further modify snow isotopes. Although we mainly discuss water isotopes due to their preponderance in paleotemperature studies in Antarctica, these effects would theoretically apply to any snowfall-dependent temperature proxy.

Water isotopes in snow are deposited under warmer-than-normal conditions (Fig. 2), which leads to a higher-than-expected δ18O. Some climatic information is lost for δ18O as cold days are not recorded or they are recorded with lower weight. Previous works suggested that δ18O could be seasonally biased due to the annual cycle of snowfall (Markle and Steig, 2022; Werner et al., 2000; Persson et al., 2011), but here we show that the temperature increase associated with precipitation events is clearly the main factor controlling the snowfall-weighted temperature in Antarctica. The variety of ΔT values across modern Antarctica suggest that it depends on precipitation regimes and the amplitude of temperature change during precipitation at a given site. The stability of ΔT over time is thus decisive for temperature reconstructions based on isotopes because temporal changes in ΔT could accompany temperature and precipitation changes and hinder the δ18O-based reconstructions. Here, the short study period does not allow us to evaluate temporal changes in ΔT, but such changes may be responsible for modifications of δ18O–temperature slopes at longer timescales, as was suggested for the glacial-interglacial range (Buizert et al., 2021). Previous studies also highlighted that, despite being weaker than non-seasonal effects in absolute value, seasonal effects on ΔT are more likely to vary with climate as the seasonality of precipitation changes (Sime et al., 2008) in response to sea ice and moisture source changes (Holloway et al., 2016). Given the spatial variability of ΔT, we advise against the use of spatial gradients to define isotope–temperature slopes for temporal reconstructions.

Moreover, the reduced annual cycle of Tw‾ relative to T‾ may reflect an annual cycle of δ18O with attenuated amplitude, which can explain why the seasonal δ18O–temperature slope appears to be lower in precipitation δ18O studies than in simple isotopic models (Casado et al., 2018). On the other hand, averaging temperature and isotopes yearly to define the δ18O–temperature slope may increase the slope value because of the higher interannual variability of δ18O induced by the irregularity of the snowfall distribution, similarly to the snowfall-weighted temperature. Links were found between Antarctic temperature and large-scale atmospheric circulation patterns in the Southern Hemisphere such as the Southern Annular Mode (abbreviated to SAM, Marshall and Thompson, 2016), which possibly influence the δ18O (Abram et al., 2014; Kino et al., 2021). Nevertheless, we did not find any significant correlation between the SAM and the yearly or monthly snowfall-weighted temperature difference. Detecting possible links between the SAM, or other climate modes, and the precipitation-weighted temperature (or δ18O) would require a more detailed investigation and may be explored in a different study.

Slopes of snowfall-weighted δ18O of precipitation against snowfall-weighted temperature have previously been suggested (Fujita and Abe, 2006; Sturm et al., 2010; Servettaz et al., 2020). The precipitation-weighted δ18O makes physical sense because it mimics the ice core signal (omitting the post-deposition effects), but paleoclimate reconstructions seek the temperature rather than the snowfall-weighted temperature. This work highlights the critical importance of event-related warming for the temperature during snowfall, as it reduces the seasonal amplitude, while the irregular snowfall distribution enhances the interannual variability of the temperature possibly recorded in water isotopes. This explains, at least partly, the higher interannual variability of precipitation-weighted δ18O, which causes an increased δ18O–temperature slope across most of Antarctica at the interannual scale compared to the seasonal scale (Goursaud et al., 2018), and low correlations between modelled δ18O and temperature at the annual scale (Münch et al., 2021). Simulation of δ18O signals that would be recorded in Antarctic Peninsula ice cores also revealed that the interannual variability in δ18O may show a poor correlation to temperature variability, even in high-accumulation regions (Sime et al., 2009a). Non-linearities in the snowfall-weighted temperature as the temperature and climate change (Fig. 3) may be responsible for the non-linear responses of isotopes to temperature and the underestimation of the temperature maximum in warm periods through increased winter precipitation (Sime et al., 2009b).

Understanding the effect of snowfall weighting on the average temperature at ice coring sites will help us to reconstruct paleotemperatures more accurately. Depending on the targeted time frame for temperature reconstruction from isotopes, be it seasonal (Jones et al., 2023) or pluriannual (Stenni et al., 2017), the reconstructed temperature range may be lessened or increased. Moreover, using slopes that are variable over time would result in better temperature quantification because the slope depends on the temperature range and the location (Sime et al., 2009b) and may vary over time (Klein et al., 2019).

Quantifying the local effect of snowfall weighting on the temperature range can help refine the temperature–isotope slopes for more accurate estimation, and it should be done for different settings ranging from glacial to warmer-than-present interglacial climates. Future temperature reconstructions could consider proceeding in two steps: (1) determine the snowfall-weighted temperature from water isotopes, for which the correlation is generally good and can be determined by Rayleigh-type models (e.g. Markle and Steig, 2022), and then (2) determine the average (non-weighted) temperature through the site-calibrated Tw–T slope, calculated for the matching temporal resolution (similarly to Fig. 3, but here we only show the yTw–yT slope computed with yearly averages and include all of Antarctica), while accounting for the difference in temperature between the condensation level and surface, often dictated by inversion strength. Greater snowfall-weighted temperature differences at low-accumulation sites suggest that changes in snowfall regimes could impact the temperature difference and thus bias the reconstructions from isotopes. Further work is necessary to fully understand how changes in snowfall dynamics may influence temperature reconstructions from isotopes, which may be facilitated by atmospheric models equipped with isotopes.