All three RCMs use a one-way nesting approach to dynamically downscale the ERA5 reanalysis (Hersbach et al., 2020, described in Sect. 2.2) to a very fine resolution (1 km horizontal grid spacing) with a series of intermediate nests. The MetUM and HCLIM use exactly the same domains, with the 1 km inner domain nested within an intermediate 3 and 12 km domain, as shown in Fig. 1a. Figure 1b shows the 1 km domain, which is focused on the region of interest. Meanwhile, Polar-WRF uses a set-up focused on the same region but with a 9 km outer domain and 3 and 1 km domains with a larger areal extent, as shown in Fig. S1 in the Supplement. The RCMs all have a model top of approximately 40 km, with 52, 65, and 70 vertical levels in Polar-WRF, HCLIM, and the MetUM, respectively, including 16, 20, and 14 levels below 1 km altitude. All three RCMs derive their model surface elevation and land–sea mask from updated high-resolution datasets. For the MetUM and Polar-WRF, this is the REMA dataset at 200 m resolution (Howat et al., 2019), whereas for HCLIM this is v3 of the MEaSUREs BedMachine elevation model (Morlighem et al., 2020; Morlighem, 2022) at 500 m resolution.

The ERA5 reanalysis is used to supply the RCMs with initial and boundary conditions. We use standard atmospheric variables to initialise the model (temperature, pressure, winds, and humidity) as well as sea surface temperature and sea ice. For Polar-WRF and HCLIM, these conditions are applied directly to the edges of the outer nest (9 and 12 km, respectively). For the MetUM, ERA5 is used to force a global model configuration of the MetUM at approximately 40 km horizontal resolution, which then provides boundary conditions for the outermost 12 km nest of the MetUM. Both HCLIM and Polar-WRF are run continuously for each case with boundary conditions supplied at three-hourly intervals and the first 24 h of the model run discarded as spin-up. The MetUM is run in “forecast mode”, which means it is reinitialised at 12-hourly intervals and then run for 24 h forecasts. The initial 12 h of each forecast is discarded as spin-up, while the second halves are kept and concatenated into a continuous time series, as in Gilbert et al. (2022).

The following subsections describe relevant technical differences and parameterisations used in each model.

ERA5 is the latest reanalysis product from the European Centre for Medium-Range Weather Forecasts (ECMWF), spanning the period 1940 to present. The production of ERA5 involves the assimilation of vast quantities of historical observations into the ECMWF's operational forecast model (IFS, at cycle 41r2), which outputs variables hourly on 137 vertical levels at ∼ 31 km horizontal resolution. The model's surface orography is interpolated from numerous elevation datasets, including SRTM30, as described in ECMWF (2016). Its surface scheme, HTESSEL, represents surface fluxes, snow, and SMB-relevant quantities like melt and runoff on surface tiles (Balsamo et al., 2009), while four cloud species (cloud liquid, ice, rain, and snow) and fractional cloud cover are described prognostically by the precipitation and large-scale cloud schemes (ECMWF, 2016). Radiation is computed by the McRad scheme (Morcrette et al., 2008).

Observational data are taken from two instrumented stations on the eastern side of the TG ice shelf: Cavity Camp (75.033° S, 105.617° W) and Channel Camp (75.050° S, 105.4334° W), situated approximately 4 km apart. Cavity Camp is located on a flat, relatively spatially homogeneous area of the ice shelf, whereas Channel Camp is located in the surface expression of a basal melt channel, which can be regions that preferentially develop surface crevasses (Alley et al., 2016, 2021). The Cavity and Channel stations are located approximately 20 km from the grounding line region of the TG ice stream where the terrain begins to slope steeply upwards. The stations are equipped with Automated Meteorology–Ice–Geophysics Observation System (AMIGOS) instruments, including sensors to measure near-surface meteorology like near-surface air temperature, relative humidity, pressure, and wind speed, as well as firn temperature, snow height, and GPS position at hourly resolution. A full description of the AMIGOS set-up is given in Scambos et al. (2013).

We also use output from simulations using the SNOWPACK multi-layer firn model (Wever, 2022) that were performed by Maclennan et al. (2023). In these simulations, the model is used to convert the observed snow height data from the AMIGOS instruments on the TG ice shelf to mass accumulation (Lehning et al., 2002a, b). The SNOWPACK model calculates the surface energy balance based on the observed near-surface meteorological variables, including wind speed, temperature, relative humidity, and incoming short-wave and long-wave radiation. It uses the surface energy balance to diagnose surface height changes from terms like surface melt and sublimation and also calculates snow height decreases from compaction. Then, the calculated snow height is compared to the observed snow height to determine the accumulation terms. The model diagnoses snowfall when all three of the following conditions are met: observed snow height exceeds the simulated height, relative humidity exceeds 50 %, and the difference between air and snow surface temperatures is less than 3 °C (Lehning et al., 1999; Wever et al., 2015). Accumulated snow mass is calculated according to Schmucki et al. (2014) using the density of fresh snow while describing the effect of snow erosion and deposition on snow density using the Redeposit scheme (Wever et al., 2023). This approach provides time-varying new snow density, which, for this area, results in a calculated average surface snow density of approximately 450 kg m⁻³. Rainfall is assumed when the air temperature exceeds 1.2 °C, during periods where the forcing model predicts precipitation. In this study, we compare the local accumulation at the Cavity and Channel stations to modelled snowfall rates. However, it is important to note that the local accumulation can be substantially impacted by net snow erosion if the station is located in a region of drifting snow divergence or by deposition when the station is in a zone of drifting snow convergence. Other sources of uncertainty concern the density of new snow, which can be of the order of 20 %–30 % (Keenan et al., 2021) and directly impacts the estimated accumulated mass from snow height measurements. Our approach also assumes that decreases in snow height are correctly captured by the model. Sublimation generally leads to very small changes in snow depth over the period of the described events, and melt during the events was also estimated to be too small to lead to any substantial surface height decrease. The snow depth measurements show a strong increase in snow depth, suggesting that any possible wind erosion of the snow surface was overshadowed by the net increase in mass in the area.

Both the summer and winter cases examined in this work have a similar synoptic set-up, with cross-terrain flow established by an atmospheric pressure dipole between the northwestern Antarctic Peninsula (high pressure) and Amundsen Sea (low pressure), as shown in Fig. 2. This same synoptic set-up is shown to be responsible for establishing cross-terrain flow and foehn winds over PIG and the wider ASE region (Francis et al., 2023). Both feature extreme precipitation, with SNOWPACK-simulated accumulation totals of 108 and 80 mm w.e. in summer at the Cavity and Channel stations, respectively, and 61 and 113 mm w.e., respectively, during winter. In both cases, observations suggest that near-surface temperatures rose to the melting point of 0 °C, supporting previous studies (e.g. Maclennan et al., 2023; Lu et al., 2023) that extreme precipitation events can have consequences not only for accumulation, but also for melting (see Fig. S2). Both cases are also detected by the Wille et al. (2021) AR-detection algorithm (not shown).

The summertime case consists of a family of three ARs that made landfall over TG and the ASE coast in quick succession, driven by a high–low-pressure couplet (Fig. 2a). A pressure dipole, with low pressure over the Amundsen Sea (reaching a minimum mean sea level pressure, MSLP, value of 951 hPa in ERA5) and high pressure over the eastern tip of South America (maximum MSLP of up to 1035 hPa), drove convergence along a pressure ridge that extended lengthways along the eastern side of the Antarctic Peninsula. The high-pressure system strengthened over the course of the case, enhancing the pressure gradient and driving more intensified convergence. This channelled moist air towards and across the West Antarctic coast, especially over the region of TG–PIG, and prevented the low-pressure systems over the Amundsen Sea from migrating. The air mass picked up moisture as it travelled over the largely sea-ice-free Bellingshausen Sea west of the peninsula and was forced over the topography of the West Antarctic Ice Sheet.

The wintertime case consists of high pressure over the Drake Passage and Bellingshausen Sea (top left of Fig. 2b, with ERA5 MSLP reaching up to 1035 hPa) and low pressure over the Amundsen Sea (bottom left of Fig. 2b, with ERA5 MSLP falling as low as 955 hPa) (Fig. 2b). A chain of low-pressure systems pushed towards the ice sheet from subpolar latitudes all the way to approximately 85° S during the case, driving high temperatures and precipitation over the ASE region.

We have established that liquid precipitation is likely falling during the cases examined over the TG and PIG ice shelves. Next, we explore the origins of this rainfall. To avoid repetition, the following analysis will use data from the 1 km MetUM simulation only, which exhibits some of the highest rainfall totals.

Figure 6 shows the relationship between maximum air temperatures and accumulated rainfall totals during the winter and summer cases in the MetUM. Maximum temperatures reach well above freezing at the lowest elevations in summer, especially over the PIG ice shelf, where maximum temperatures of ∼ 5 °C are simulated across much of the ice shelf nearest to the surface (Fig. 6a). However, maximum temperatures above zero are also simulated during winter over the Abbot, Cosgrove, and PIG ice shelves (Fig. 6d). These above-freezing temperatures are simulated over regions in the immediate lee of steep terrain, suggesting that a foehn effect could be altering the meteorology associated with the ARs, as also shown by Francis et al. (2023) over PIG, by Gehring et al. (2022) over Davis, by Wille et al. (2022) over the Antarctic Peninsula, and Lu et al. (2023) over the South Orkney Islands. A foehn effect could be expected to produce a warm pool near the surface in the lee of steep terrain when air flow is across this topography.

Panels b and c of Fig. 6 show that there is indeed a warm pool of air at the surface over the TG and PIG ice shelves and that the region of above-freezing summer maximum mean air temperatures along the transects through TG and PIG reaches up to above 1 km in altitude and up to nearly 2 km altitude over PIG. These higher temperatures aloft may be partly associated with a foehn effect and/or by the air mass being orographically lifted, resulting in cooling, condensation, and latent heating of the surrounding atmosphere. In winter, regions of above-zero maximum air temperatures are much more spatially confined and are concentrated in small pockets over the PIG ice shelf at altitudes below 750 m (Fig. 6e), while temperatures remain below freezing along the TG transect (Fig. 6f).

Although near-surface temperatures over the PIG and TG ice shelves reach maximum values near or above freezing during both cases (Table 1, Fig. 6), median temperatures are colder, remaining below freezing during the winter case and staying very close to zero in the summer case (Table 1). Further, most of the liquid precipitation falls at times and altitudes when ambient temperatures are not above freezing, suggesting that temperatures above 0 °C are not always required to produce rainfall.

Indeed, the highest rainfall totals over both the TG and PIG ice shelves are simulated at times and altitudes where temperatures are within a few degrees of 0 °C. During the summer case, a near-surface warm pool develops each time the three ARs push air over the steep terrain, particularly over the Abbot, Cosgrove, and PIG ice shelves. When near-surface temperatures are highest (> 2 °C) and the surface warm pool is more vertically extended, rainfall totals are lower, suggesting that the highest temperatures can reduce the amount of liquid precipitation via evaporation and/or sublimation. However, during times when maximum air temperatures reduce to more moderate values of around -2 to +2 °C, rainfall totals become considerable over both the TG and PIG ice shelves. For example, rainfall totals over the TG ice shelf are highest during easterly flow, which advects warm air from regions with the strongest foehn warming, allowing it to mix and cool by the time it reaches TG. Similarly, the highest rainfall totals over PIG are simulated when the warm pool begins to fade. Over both ice shelves, the highest rainfall totals are simulated along the flanks of steep terrain and often at altitudes 500–1500 m above the surface. During the winter case temperatures are cooler, so rainfall only occurs in periods of the strongest foehn warming. However, temperatures during these periods are also within the same range of -2 to +2 °C.

The dashed lines in Fig. 6 indicate the maximum height of the 0 °C air temperature threshold, which can be considered a maximum “melt line” height. Below this threshold, snow and ice might be expected to melt and fall as rain if the melt layer that solid particles fall through is thick enough. However, in the summer case, the highest accumulated rainfall totals are not always concentrated below the melt line, especially in the case of the TG transect. Moreover, in the winter case the region of above-zero temperatures is extremely spatially constrained along the PIG transect and completely absent along the TG transect. Despite this, accumulated rainfall totals in excess of 10 mm w.e. are simulated along the transects through both ice shelves, resulting in median summer surface rainfall totals of 16.9 and 6.2 mm w.e. over the TG and PIG ice shelves, respectively, and 4.5 and 1.1 mm w.e., respectively, in winter (Table 2). This means that supercooled liquid precipitation is falling, even fairly high up in the atmosphere.

Over both ice shelves and in summer especially, accumulated rainfall totals are higher just above the surface than at the surface, indicating that some rain is being lost to evaporation, riming, or freezing as it descends through the warm near-surface layer or that rain droplets are being lofted upwards by updraughts as air is forced to flow up the steep terrain of the ice stream, as described in Silber et al. (2019). Meanwhile in winter, most of the rainfall that reaches the surface is spatially constrained to further along the TG transect, upstream of the grounding zone where the AR intersects with steep topography. Over the PIG ice shelf winter rainfall is more minimal, despite temperatures being warmer than over the TG ice shelf, as indicated in Tables 1 and 2.

As shown in Fig. 6, patterns of accumulated rainfall are relatively consistent between locations and seasons. In both seasons, rain is present at altitudes of up to ∼ 2500 m, although in summer, the height of the 5 mm w.e. accumulated rainfall contour is marginally higher, reaching up to around 3000 m a.s.l. (Fig. 6b, c). This is despite the patterns of maximum air temperature being quite different between the transects. In both seasons, there are also pockets of high accumulated snowfall totals above the surface over the ice shelves (Fig. S5), indicating that in some locations snow is sublimating and/or melting into rain as it descends through the warm layer described above. This same effect (snow transitioning to rain as it falls through a warm layer) has been observed during combined AR–foehn conditions at Davis station in East Antarctica by Gehring et al. (2022).

Liquid precipitation falls from liquid-bearing mixed-phase clouds at temperatures as cold as -11 and -8 °C in winter and summer, respectively (not shown), suggesting that the formation mechanisms are supercooled warm-rain processes. Cloud droplets reach a large enough size to descend through the atmosphere, collecting other droplets on their way and producing supercooled rain through warm-rain processes like collision–coalescence and accretion. Supercooled drizzle usually forms when there are insufficient ice nuclei to allow the atmospherically available water vapour to nucleate and grow into solid precipitation via ice processes. This is also consistent with a pristine region such as the West Antarctic and the northeasterly wind direction, which has advected air from across Ellsworth Land, where it is unlikely to have picked up ice-nucleating particles. These results also suggest that rainfall is being generated via the seeder–feeder mechanism, which increases orographic rainfall rates and can often result in the formation of supercooled drizzle in mixed-phase clouds (Fernández-González et al., 2015; Ramelli et al., 2021). The process involves ice or liquid descending from upper (seeder) levels and enhancing the rate of accretion in lower (feeder) levels, thereby increasing surface rainfall amounts (Lean and Clark, 2003). It frequently occurs in regions with significant updraughts, such as when air is forced up and over steep terrain (Ramelli et al., 2021). Finlon et al. (2020) show that the 0 °C isotherm separates the seeder and feeder levels in such cases, for example in an AR over the Southern Ocean, which is consistent with the results shown here.

In both cases, the region of accumulated rainfall reaches several hundred kilometres inland. Maximum air temperatures are higher over the PIG transect (as shown in Fig. 6a), but in summer, higher accumulated rainfall totals are simulated over the TG ice shelf and upstream (Fig. 6b). This suggests that air temperatures are not the only control on rainfall totals and that factors like large-scale cloud and synoptic patterns, as well as orography, likely influence the phase and quantity of precipitation that falls. Bromwich et al. (2024) and Bozkurt et al. (2024) also highlight the complexity of processes involved in initiating the phase transition from snow to rain in ARs.

This is also supported by Fig. 7, which shows the ratio of rain to snow over the TG and PIG transects in summer and hints at different formation mechanisms of precipitation. Hatched regions in Fig. 7a show that there are several regions where the amount of summertime rainfall exceeds the amount of snowfall: most of these regions are offshore, with some also immediately above the steepest slopes in the lee of airflow over the terrain, for example over the PIG and Abbot ice shelves. Along the TG and PIG transects shown in Fig. 7b and c, summer precipitation is largely mixed phase, and accumulated totals are dominated by snowfall. However, rain / snow ratios of above 0.5 (indicating that at least 33 % of precipitation is falling as rain) highlight regions where rainfall totals are relatively high, and liquid precipitation processes may be comparatively more important. Figure 7b shows that along the TG transect, the highest rain / snow ratio is found above the zero-degree line at higher elevations around 150 km inland above the glacier slope. However at the surface, the highest rain / snow ratios are simulated at the foot of the steep terrain, as indicated by Fig. 7d.

Along the PIG transect, much more snow falls relative to rain (Fig. 7c), except at the lowest elevations and directly above the steepest orographic gradient (Fig. 7e), where the effect of foehn-driven warming is highest (Fig. 6a, b). Over the PIG transect, the highest rain / snow ratios are found in the warmest regions, indicating that some of the rain has formed by the melting of solid precipitation as it descends through the warm near-surface layer (Fig. 6c). Some of the precipitation likely also sublimates and evaporates as it descends through this warm layer, which is illustrated by the fact that the highest rainfall fluxes are simulated in pockets above the surface in both winter and summer.

The importance of supercooled rain formation processes is especially evident in the winter case, where maximum air temperatures above 0 °C are simulated only in extremely confined locations over the PIG ice shelf (Fig. 6e). Supercooled liquid precipitation of up to more than 10 mm w.e. occurs over both ice shelves, especially at higher altitudes. This compares with maximum cumulative totals of ∼ 20 mm w.e. in summer, i.e. double the magnitude of winter.

Liquid precipitation has been shown to occur at very low temperatures in Antarctica. For example, Silber et al. (2019) demonstrated that drizzle can occur at temperatures down to -25 °C near McMurdo. They showed that in the presence of sufficient condensate, liquid precipitation can fall from stratiform clouds like those that were present during the summer case. In the cases studied here, the source of condensate was from the three ARs that made landfall, as well as the orographic lifting and condensation of moisture in the air as it was forced over the steep topography. In the pristine Antarctic atmosphere there are also fewer aerosols that can act as ice-nucleating particles, thus preventing the rapid glaciation of the cloud.

While rainfall does not occur frequently over the ASE in the current climate and is exclusively associated with extreme events like the ARs described here, these conditions may occur more often in future. Antarctic total precipitation amounts will increase with warming (by 5.5 % °C⁻¹ on average), and the number of extreme precipitation events and ARs over the ASE will probably continue to rise with global temperatures (Nicola et al., 2023). There is also some evidence that climate change may cause ARs to shift poleward, along with the storm track (Ma et al., 2020; Chemke, 2022). Although ARs are unlikely themselves to destabilise ice shelves like the TG ice shelf, they affect accumulation and so could change the timing of future mass loss from West Antarctic glaciers (Maclennan et al., 2023). For example, Donat-Magnin et al. (2021) use RCMs to show that ablation will increase over ASE ice shelves such as the PIG and TG ice shelves, primarily driven by increased melt and runoff rates, but that the ratio of melting to snowfall is still low compared to other Antarctic ice shelves like the Ross or Filchner–Ronne. This is corroborated by Gilbert and Kittel (2021), who show that even under a high-emissions scenario, the risk of surface-driven West Antarctic ice shelf collapse is low until the end of the century because the amount of runoff is greatly outweighed by increased snowfall.

However, climate change will also increase the proportion of precipitation falling as rain, as has been shown for the Arctic (McCrystall et al., 2021). For instance, Donat-Magnin et al. (2021) show that rain occurs on 0.5 and 1 d yr⁻¹ over the TG and PIG, respectively, in the current climate but that this may increase to 4.2 and 5.8 yr⁻¹ by 2100. These changes to total precipitation and the phase of precipitation are extremely important for determining ice sheet and ice shelf SMB. For example, rain-on-snow events can cause the surface to darken, reducing surface albedo and enhancing melt via the melt–albedo feedback (Box et al., 2022; Stroeve et al., 2022). As the climate warms, the future of the Antarctic precipitation phase will remain an extremely pertinent subject of research.

As noted in Sect. 1, model resolution can strongly determine the amount of precipitation simulated, especially in regions with steep topography and where precipitation is dominated by orographic processes, such as those we have shown to be occurring there.

As shown in Fig. 3 and Table 2, the RCMs simulate higher accumulated snow and rainfall totals than ERA5, which is likely related to resolution. At ∼ 31 km, the resolution of ERA5 is considerably coarser than the resolution of the 1 km innermost RCM nests. One of the variables for which this is most apparent is air temperature. As shown in Table 2, all three RCMs indicate much warmer median air temperatures over the TG and PIG ice shelves than ERA5, suggesting the lower resolution of ERA5 is not sufficient to capture the topographic modification of the AR. The absolute maximum temperatures are also lower, with median air temperatures over the PIG and TG ice shelves approximately 10 °C lower in ERA5 than the RCMs in both cases (Table 2). The agreement between the RCMs, all of which are forced by ERA5, suggests that the representation of complex topography is the cause of this difference, rather than internal model differences. This indicates that high resolution is necessary to capture the complex interactions between large-scale, mesoscale, and local meteorological processes, which are important during such conditions in this region, and that RCMs are required to reproduce the factors governing the drivers and impacts of AR events in regions with complex topography.

However, there is a trade-off when conducting simulations between resolution and computational efficiency, so it is worth investigating what kind of resolution is sufficient to capture the dynamics of extreme precipitation in this region. To explore this, Fig. 8 shows accumulated snow and rainfall totals during the summer case across the three different-resolution nests of the MetUM: 1, 3, and 12 km grid spacing.

There are few differences in accumulated snowfall amounts between the various resolution domains of the MetUM, indicating that simulated snowfall totals are comparatively insensitive to resolution. Given that snowfall dominates over rainfall, this suggests that the model can capture the dominant drivers of extreme precipitation, even at the relatively coarser resolution of 12 km. However, there is still some advantage to using higher resolution to capture finer-scale spatial variability when concentrating on smaller regions such as a single ice shelf. Furthermore, while total accumulated snowfall in the 12 km simulation is comparable to 1 km totals, the accumulated rainfall totals over TG ice shelf are 16 times lower in the 12 km simulation compared to the 1 km simulations during the summer case and 4 times lower during the winter case (Fig. 8). Across the entire domain, rain is simulated only over the steepest terrain, such as Thurston Island and the King Peninsula. Because considerably less rain falls than snow, this has a fairly minimal effect on total accumulated precipitation. Nonetheless, this suggests that subgrid-scale processes may not be parameterised adequately at coarser resolutions, for example the subgrid phase partitioning of cloud or the microphysical processes controlling the generation of supercooled drizzle. These are known problems with the MetUM (Gilbert et al., 2020; Bodas-Salcedo et al., 2016; Abel et al., 2017) and other RCMs (Hyder et al., 2018; Vergara-Temprado et al., 2018), especially in Antarctica and the Southern Ocean. This suggests that while coarser-resolution configurations of RCMs can capture the large-scale characteristics of precipitation extremes, finer grid scales are required to fully capture the complex interactions producing liquid precipitation in this region.