Model performance and surface impacts of atmospheric river events in Antarctica

Abstract

There is increasing evidence that atmospheric rivers (ARs) drive extreme precipitation and melt events across Antarctica and that these impacts are more accurately captured in high-resolution models. However, a comprehensive evaluation of AR impacts, comparing the performance of models with varying resolutions and physics across multiple AR events, has not yet been conducted. In this study, we simulate four recent AR events using the regional climate model HCLIM43 in its ALADIN (11 km) and AROME (11 km and 2.5 km) configurations, as well as ERA5 (31 km) and MERRA-2 (50 km), to analyze the dominant factors driving melt and precipitation and how spatial resolution and model physics affect surface impacts compared to observations. The events include intense snowfall and longwave radiation (Jun 2019), surface melt from foehn winds (Feb 2020), a large-scale heat anomaly driven by radiative and turbulent processes (Mar 2022), and inland surface warming after moisture is released by sea ice and ice shelves (Dec 2023). While all reanalyses and models underestimate surface warming and melt during these events, the high-resolution 2.5 km AROME configuration tends to simulate the most realistic precipitation and melt extents, largely due to its improved representation of foehn effects and reduced cloud biases. Longwave radiation generally dominates AR-induced warming, particularly over wider inland regions, while sensible heat fluxes are dominant in coastal and foehn-prone regions. Lastly, substantial differences among models/reanalyses in cloud phase and total cloud water paths underscore the need for improved cloud parameterizations and surface energy budget calculations.

1 Introduction

Atmospheric rivers (ARs) carry large amounts of moisture in narrow bands across great distances, and have been increasingly recognized for their important role in dominating precipitation patterns and inducing melt episodes across various regions of Antarctica [1,2,3,4,5,6,7]. There is observation- and model-based evidence that the relative contribution of ARs to Antarctic precipitation (on average 10-30%) has increased in recent years and will likely further increase with future warming [8, 9]. Recent studies have emphasized that ARs not only influence coastal areas but that their impact often extends further inland especially when the atmospheric dynamics support the continued poleward advection of warm, moist air masses [1, 5, 6]. These impacts include not only precipitation, but also significant melt events that cause runoff and hydrofracturing, especially in West Antarctica, where between 40% and 80% of the total summer melt of ice shelves was found to be AR-induced [10, 11]. That the safety band of Antarctic ice shelves may be vulnerable to ARs [12], with a concomitant effect on the Antarctic contribution to sea level rise, is an important justification for understanding the causes, processes and consequences of ARs in Antarctica. Another indirect, cumulative effect of ARs may be meltwater-induced doline fractures, such as those observed in 2020 on the George VI Ice Shelf [13] following extended periods of warm air advection from the surrounding ocean [14]. Furthermore, increased precipitation from ARs has also been flagged as a potential balancing factor for the ice sheet mass budget, balancing increased ice loss driven by ocean melting and calving and further being affected by reduced sea ice cover that permits enhanced accumulation to penetrate further inland [7, 15, 16].

However, conclusively attributing AR events to melt and/or accumulation episodes remains challenging due to the often concurrent influences of phenomena such as foehn winds, cyclones, or atmospheric blocking which contribute to the overall impact [17,18,19,20,21]. Additionally, observations are very limited, and most models rely on simplified parameterizations of key physical processes like convection or rain heat flux. To better understand the characteristics and surface impacts of Antarctic ARs, recent studies have analyzed intense AR events in selected regions using reanalysis, observations, and/or model simulations [10, 11, 22,23,24]. These analyses suggest that meteorological background conditions, such as atmospheric blocking and foehn winds, largely influence the lifetime and impact of ARs, which are otherwise determined by the season and region of each AR event. There is consensus that AR-induced precipitation and surface impacts as well as the associated horizontal and vertical AR structure are most accurately captured in high resolution (HR) settings due to more detailed topography and improved model physics [24, 25], which however presents computational challenges at a pan-Antarctic scale. In this study, we negate this issue by analyzing four unique recent AR events (some of which have been addressed in earlier studies) that occur in various Antarctic sectors and seasons with varying impacts on the surface. We are motivated to assess the range of possible surface impacts of ARs and how these vary by season, location and atmospheric structure.

2 Approach and methodology

2.1 Background on AR events

After detecting ARs from 01/2015 to 02/2024 using ERA5 and MERRA-2, we selected four events based on intensity, representing different seasons, regions, and impacts, such as snowfall versus melt events. By comparing reanalysis and observations with three different high-resolution simulations of the regional climate model (RCM) HCLIM-43, we examine the physical impacts of the four events and evaluate the advantages of high-resolution RCMs in more accurately capturing local effects.

Table 1 Four anomalous AR events investigated in this study, with their landfall locations, literature references, and availabla observations to evaluate model performance. S represent surface observations based on Antarctic weather stations, and R represent radionsonde observations sourced from the IGRA2 archive

Event 1 (Jun 2019, East Antarctica) refers to a mid-winter AR event that reached the West Ice Shelf (WeIS) and Amery Ice Shelf (AmIS) on Jun 9, 2019, delivering substantial snowfall to the coastal region. This AR event occurred just before the sudden drainage of 0.75 km$^{3}$ of water from an englacial lake beneath the AmIS between Jun 9-11 [26]. Although these two events are likely not directly related, as no rainfall or surface melt coincided with the AR, it notably influenced snow accumulation, altering snowpack properties, albedo, and surface mass balance (SMB) [27, 28]. This makes it an important case for evaluating model performance in this sensitive region. Notably, [29] found that 50% of the annual precipitation on the AmIS is delivered by just the most intense 10 days of the year. If this AR event had occurred during the summer, it could have induced considerable surface melt and rainfall, both of which are known to trigger hydrofracturing and subsequent drainage events [10, 11, 30,31,32]. The WeIS is also vulnerable to calving events and surface melt, with significant warming and melt trends already observed [33, 34].

Event 2 (Feb 2020, Antarctic Peninsula) emerged from moist Pacific air masses being pushed to the AP, resulting in the highest surface temperature ever recorded in Antarctica (18.3$^{\circ }$C at Esperanza station on Feb 6, 2020) [35, 36]. [37] also reported an unprecedented regional mean temperature anomaly of +4.5$^{\circ }$C, averaged across several AP stations, mainly driven by warm horizontal advection and a high-pressure system that diabatically and adiabatically warm the surface. Meanwhile, the local record temperatures observed at stations on the eastern flank of the AP (e.g. Esperanza) were primarily attributed to foehn mechanisms, but with significant differences regarding the dominant contributor to the foehn effect: [35] found that vertical diabatic processes (sensible heat and radiation) dominated surface warming at Esperanza station, while [18] showed that at Marambio station (further south and higher up) adiabatic warming, i.e. the vertical isentropic drawdown, generated the foehn effect. The heatwave also resulted in a record extent and duration of melt observed on Larsen C ice shelf [38], located downstream of the AR on Feb 6 that initiated the heatwave (which is the focus of our Feb 2020 analysis).

Event 3 (Mar 2022, East Antarctica) was extensively discussed in [39, 21], and [40], where the uniqueness of the event lay in the +38$^{\circ }$C daily mean surface temperature anomaly at Concordia Station – the largest ever recorded globally. [39] attributed this anomaly primarily to cloud-induced downward longwave radiation and found that models tend to underestimate surface temperatures by 5-10$^{\circ }$C. These studies have further shown that the peak surface warming occurred on Mar 18 and resulted in surface melt and rainfall near the coast. The sudden loss of the Conger ice shelf during this event, likely triggered by surface melt and hydrofracture lends further urgency to the question of surface impacts of ARs. The ice shelf had been getting progressively smaller and thinner, due to ocean driven warming, but the short term shock applied by the high temperatures associated with the AR demonstrates how short term weather variability can have a long-term effect on ice sheet dynamics [41].

Event 4 (Dec 2023, West Antarctica) occurred in the austral summer of 2023/2024 (Dec 28 to Jan 1) and has not yet been covered by any literature. We found that it ranked among the strongest AR events in the region based on its IVT intensity, characterized by a direct northerly flow over the Bellingshausen Sea before the AR reached the Abbot Ice Shelf and the Transantarctic Mountains bringing significant amounts of snowfall.

The net impact on the SMB of the four events depends on the spatial extent considered but is likely significantly positive over the wider AR domain, as melt is constrained to coastal areas and runoff is usually low. For example, while most of the studies describing the above events did not calculate SMB, [40] confirmed a substantial Antarctic SMB increase for the March 2022 heatwave due to intense, large-scale precipitation. Regarding data sources, most studies relied on ERA5 or localized surface observations, which in some cases exhibited significant deviations. [21] analyzed the Mar 2022 event with the RCM Modèle Atmosphérique Régional (MAR) at 35 km resolution with ERA5 as boundary and forcing data, but did not compare the RCM results to ERA5. None of the four events have yet been analyzed using a convection-permitting RCM like the non-hydrostatic AROME configuration. Although convection is generally limited in cold Antarctic environments, it can occur over ocean areas and may play a role during extreme ARs through processes like latent heat release [42]. More importantly, precipitation is highly determined by the resolution of the model with high resolution models better able to resolve orographic effects in precipitation patterns that are likely underestimated in the reanalysis products.

2.2 Data framework

For each event we analyze a 10-day simulation over the respective region using 3 different configurations of the RCM HCLIM, version 43 [43]: HCLIM43-ALADIN at 11 km (ALADIN11 hereafter) and HCLIM43-AROME at 11 km and 2.5km (AROME11 and AROME2.5 hereafter). A brief overview of the physics and resolution of the models is presented in Table 2 and the respective model boundaries are shown in Fig. 1. Details about the model components, forcing, and run configurations are covered in Sect. 2.3. Additionally, we used data from ERA5 (25 km) [44] and MERRA-2 (50 km) [45] to detect the corresponding ARs for these events (which requires a larger spatial and temporal window, and thus cannot be achieved with the 10-day HCLIM runs) and to compare their performance to the higher resolution HCLIM runs.

Table 2 Overview of the resolution, convection physics, albedo scheme and snow scheme of MERRA-2, ERA5, and HCLIM configurations (ALADIN and AROME)

Fig. 1

a-d: Four AR events which were simulated with 3 different configurations of HCLIM in this study (here plots are based on ERA5, see 19 for MERRA-2 and the HCLIM configurations). For each event, the 6-hour time step when IVT was highest within the AR land domain (yellow outline; 6-8$^{\circ }$ spherical cap around the central AR axis in black) is shown. The red window represents the lateral boundaries of the HCLIM-AROME (2.5km) domain. Light blue line marks the sea ice extent (15 SIC%). e-h: Precipitation for the same events, zoomed in on the AROME2.5 domain windows. Green marked stations represent IGRA2 radiosonde profile locations, and red marked stations represent weather stations used for evaluation. Station names: Amery G3 (AMG3), Panda (PAN), Davis (DAV), Robertson Island (ROB), Hugo Island (HUG), Prospect Point (PRO), Rothera (ROT), Dome-A (DOM-A), Aurora (AUR), Concordia (CON), Casey (CAS), Dumont-Durville (DDU) and Thurston Island (THU). i-l: Extended time series of the weighted mean surface temperature of the AR domain (yellow outlines in upper panel) for each event based on ERA5. Yellow shadings mark the investigated time span of each event, where IVT was above average

We also retrieved data from [46] who employed a melt detection algorithm on passive microwave radiometer data (AMSR2). The authors provided a binary melt mask that we use to evaluate model performance regarding melt extent. The AMSR2-based data shows relatively high accuracy, with 93% agreement and a Kappa coefficient of 0.79, validated against in situ air temperature measurements [47]. Lastly, we retrieved surface observations from Antarctic weather stations (details below) as well as radiosonde observation measurements from IGRA2 [48] to evaluate vertical profiles (of e.g. temperature and humidity) in the AR regions.

2.3 HCLIM simulations using ALADIN and AROME

We used the HCLIM modelling system [43, 49, 50], cycle 43, to conduct high-resolution simulations of the selected events. It is based on the ALADIN - High Resolution Limited Area Model (HIRLAM) numerical weather prediction system, with two different atmospheric physics packages, namely HCLIM-AROME and HCLIM-ALADIN. The physics and parametrizations for both model configurations are presented in Table 3.

Table 3 Parametrization and dynamics of the ALADIN and AROME configurations used in HCLIM simulations

Each event was simulated with both physics packages but to different resolutions. For HCLIM-ALADIN, the events were simulated only at 11 km of horizontal resolution with 65 vertical levels. For HCLIM-AROME, the chosen domains run in a one-way nested simulation at 11 km and 2.5km of horizontal resolution, both with 65 vertical levels.

2.4 IVT definition and AR algorithm

IVT is defined as

$$\begin{aligned} \begin{aligned} IVT \hspace{0.2cm} =&\hspace{0.4cm} \sqrt{\left( \frac{1}{g}\int _{p=1000}^{p=250} qu\,dp\right) ^2} + \hspace{0.3cm} \sqrt{\left( \frac{1}{g}\int _{p=1000}^{p=250} qv\,dp \right) ^2} \end{aligned} \end{aligned}$$

(1)

where g is the approximate gravitational acceleration (9.81 m s$^{-2}$), q is the specific humidity, u and v are the eastward and northward wind components, and dp is the difference of each two adjacent pressure levels from 1000 to 250 hPa.

ARs were identified from 2015-2024 using MERRA-2 and ERA5 data (6-hourly resolution) with the following criteria: First, we identify regions where IVT surpasses the 98_th percentile of IVT_u,v from monthly climatologies at each grid cell, with an additional fixed minimum threshold of 40 kg m$^{-1}$ s$^{-1}$ IVT_u,v (slightly lower than in [69] to include inland ARs). If the size of this region exceeds the minimum length-to-width ratio of 2 and the minimum length of 1700 km [70], we classify the area as an AR. In comparison to conventional AR algorithms, we include an additional step to address ARs with a width that exceeds the defined AR criterion. This scenario often arises in extreme ARs where the IVT threshold is easily surpassed in adjacent regions (e.g. during the Mar 2022 heatwave). When the length is sufficient, but the width is too large, we incrementally increase the IVT threshold by 0.03% and repeat the calculation of the length-to-width ratio. This process typically narrows the AR region, enhancing the likelihood of detecting the AR within the broader region, and thus provides a more robust and less threshold-conditional detection. Although these details are less relevant for this study, as we focus on four specific events rather than broad AR catalogue metrics, we suggest this method better aligns with the physical characteristics of ARs. Finally, only ARs that reach land are considered as we focus on land surface impacts.

2.5 Observations: weather stations, radiosondes, and microwave sensors

Surface data (temperature, relative humidity and pressure) of several Automatic Weather Station (AWS) were sourced from the following institutions or organized databases:

• Australian Antarctic Data Centre (AADC) [71]: Amery G3, Aurora-BN, Panda, Dome-A

• Antarctic Meteorological Research centre (AMRC) [72]: Thurston Island

• Antarctic automatic weather stations (AntAWS) [73]: Hugo Island, Prospect Point, Robertson Island

All sources offer quality-controlled data (for details refer to the listed references above), but caution is advised particularly under extreme conditions.

Next to surface data from AWS, we retrieved vertical profiles of temperature, humidity and wind speeds from the IGRA2 archive [48], providing multiple quality-controlled measurements distributed across the Antarctic continent (of which we selected those located in the vicinity of the ARs; Fig. 1h-k). Technical details about IGRA2 measurements and quality controls can be found in [74, 75] and [76]. All stations (AWS and radiosonde profile stations) are listed in Table 1.

The large-scale IVT patterns of the four AR events are comparably represented across all five models (Fig. 19). In the following, we (a) describe the surface impacts over the AR land domain over the 10-day simulations, before we (b) evaluate the models’ performance in representing point observations of surface variables and vertical profiles with weather stations and radiosonde observations, and lastly (c) analyze differences in vertical cloud cross-sections and surface maps of precipitation and melt. In the models, melt is estimated by designating grid points as melt regions when the skin temperature is at or above 0$^{\circ }$C (we also include sensitivity tests for this threshold).

3 Results

Most of the radiosonde stations are assimilated in MERRA-2 and ERA5 (used to force HCLIM), hence all models showed very similar temporal and vertical patterns compared to the radiosondes (Figs. 23-27). We still highlight instances where models deviate from observed profiles and find that these deviations differ across events, with no model consistently performing better or worse than the others. In Figs. 23-27 we compare the time series of the approximate upper air AR flow (800 hPa) and vertical profiles at the time of the temperature maximum. Greater emphasis is placed on differences from surface data from AWS, which are typically not assimilated in reanalyses and thus offer an independent benchmark for model evaluation [44, 45, 77, 78]. For both radiosonde and AWS observations, data were collected for 40 days before and after each event to assess deviations in temperature and wind speed relative to the 3-month average.

When presenting spatially averaged AR metrics (e.g. Figure 7), such as mean IVT or the sum of precipitation, we refer to the land domain considered to be impacted by the AR (yellow outlines in Fig. 1a-d) that represent a 6-8$^{\circ }$ spherical cap around the central AR axis over land). We chose to focus on domains slightly wider than just the detected AR band because the impacted regions often extend beyond the AR itself, and ARs tend to move and change shape over time.

3.1 June 2019: East Antarctic Winter AR reaching the Amery Ice Shelf

• (a) Event Evolution and Surface Impact

The mid-winter AR that reached the East Antarctic coast in the vicinity of the AmIS resulted in a sudden increase of precipitation from 0 to $\sim$1.3 mm 6 h$^{-1}$ averaged over the AR land domain on Jun 9, 2019 (Fig. 2c). The skin and surface temperature maxima at most sites occur approximately one day later than IVT and precipitation (here on Jun 10; Fig. 2a), due to a delayed surface response caused by the higher thermal inertia of snow and ice.

Fig. 2

10-day time series of skin temperature, surface melt extent, precipitation, rainfall, SEB and longwave radiation of the AR event in Jun 2019, as the weighted sum or mean over the AR land domains (yellow outline in Fig. 1a), simulated by ERA5, MERRA-2 and the three HCLIM configurations

Although most precipitation during this event is concentrated near the eastern part of the ice shelf edge (Fig. 1e), more southerly locations such as the AmIS lake drainage collapse site (72$^{\circ }$S/69$^{\circ }$E) also experienced peaks in snowfall, IVT and temperature on Jun 9/10 (Fig. 20), although these maxima lie within normal variability at the site. In the absence of any rainfall or surface melt, we do not suspect a direct link between the AR and the lake drainage event. However, the formation of englacial lakes, their drainage, and many other ice shelf dynamics result from cumulative surface melt, rainfall, and snow deposition [79], all of which are strongly influenced by the frequency and intensity of ARs [7, 11, 80]. Compared to the other AR events, this event was relatively short-lived, likely due to the low winter temperatures, which reduce the potential for high humidity and sustained warmth typically associated with ARs. It is noteworthy that the water vapour penetrated so far inland after crossing a large region of sea ice, where much of the heat and moisture from the AR was lost before reaching the coast. As discussed in the following, we suggest that the existence of coastal polynyas has contributed to additional oceanic moisture uptake, which contributed to the intense precipitation on Jun 9. The brief but intense nature of the event is not only reflected in the sudden spike in precipitation but also in a rapid increase in downward longwave radiation over the AR land domain (from -50 to almost 0 W m$^{-2}$ in 12 h; Fig. 2f). This positive longwave radiation anomaly is the dominant factor in generating the sudden increase (+15$^{\circ }$C) in skin temperature (Fig. 2a). However, the net SEB over land (including the ice shelves) is highest in ERA5, not due to higher longwave radiation (Fig. 2f), but higher turbulent heat fluxes (Fig. 29b,c). Averaged over the AR land domain, all other models simulate a pronounced decrease in both downward sensible heat flux (up to -25 W m$^{-2}$) and latent heat flux (up to -3 W m$^{-2}$) from June 9 to June 11. These negative averages result from negative latent heat fluxes over the AmIS due to strong katabatic winds [81] and also negative sensible heat fluxes at the ice shelf edge, likely due to nearby oceanic heat loss in the absence of coastal sea ice (Figs. 22 & 21). Over the other areas of the AR land domain, all models simulate positive turbulent heat fluxes, which are highest in ERA5. The secondary but significant influence of sensible and latent heat fluxes on the net SEB is common across all events, where their net impact is often minor because latent and sensible heat anomalies tend to cancel each other out, as further discussed in the Feb 2020 event below. We find that the models tend to diverge more in the turbulent fluxes, while the radiative budgets show similar patterns with recurring differences (e.g. MERRA-2 consistently simulates lower longwave radiation, while ERA5 ’underestimates’ shortwave radiation; Figs. 2,7,11,15 and Figs. 29-32). Based on calculated monthly climatologies (1994-2023) of SEB components in MERRA-2 and ERA5, we found that these radiative biases also persist under non-AR conditions, but are of reduced magnitude ($\sim$-10 W m$^{-2}$ lower).

• (b) Evaluation with Station Data

While the highest temperatures of the event at the AWS sites at Panda and Amery-G3 were fairly well captured by the models, the temperatures on days before and after were underestimated by up to 15$^{\circ }$C by all models (Fig. 3a,d). At Panda, which is located further inland, all models also exhibit a one-day delay in the timing of the temperature peak, and MERRA-2 and ERA5 are too cold during the entire 10 days. The delayed temperature increase may be caused by too low skin temperatures and inaccurate snowpack properties in the models/reanalyses, delaying the surface air warming response to advected heat. All models accurately simulate relative humidity at Panda ($\sim$70%), except for too high values in MERRA-2, possibly affected by the cold temperature bias. At Amery-G3 however, all models fail to capture the consistently high relative humidity of $\sim$90% over the 10-day period. If this location represents conditions at nearby sites, it suggests that the models may also underestimate precipitation near the WeIS and AmIS, including the lake drainage area. HCLIM simulates the mean and variability of temperature and relative humidity slightly better than the reanalyses at both stations. We performed a sensitivity analysis by interpolating all models to the lowest resolution of MERRA-2 (0.5$^{\circ }$) at both stations and noted negligible differences. For all AWS comparisons, we however acknowledge potential errors arising from the mismatch between the larger grid cells and point-based AWS observations.

Fig. 3

10-day time series of surface temperature, surface relative humidity and surface wind speed at Amery G3 station and Panda station (red dots in Fig. 1e). Weather station data in green, other colours represent the different models at the nearest grid point. Wind speed data was not available at the two sites

To also evaluate the local vertical atmospheric structure, we compared closest-grid point model profiles at Davis station (located on the coast where the AR makes landfall) with radiosonde observations during the event. The station data shows that the air temperature at 800 hPa peaked on Jun 9, recording -10$^{\circ }$C (Fig. 23a), marking 9$^{\circ }$C higher than the average temperature recorded in the 3-month event time window (Fig. 23g). All models and configurations accurately capture the 800 hPa temperature and strong northeasterly wind speeds (over 20 m s$^{-1}$ higher than the 3-month average), while HCLIM slightly underestimates the wind speeds above and below 800 hPa by 5-10 m s$^{-1}$ (Fig. 24d), suggesting inaccuracies in the adjustments of mid-tropospheric and katabatic wind speeds introduced by the downscaling process.

• (c) Clouds and Precipitation

Compared to the other events, most of the moisture was transported near the surface above 800 hPa, owing to the relatively flat coastal orography. Initially low-level clouds formed over the ocean and sea ice, with cloud water content extending to higher altitudes as they developed near the coast (Fig. 4). At this point, models diverge significantly in their simulation of vertical cloud extents, snowfall, and cloud phase (Fig. 34). The two AROME configurations simulate fewer clouds than the hydrostatic ALADIN11, which exclusively generates ice-phase clouds (not only in the AR cross-section but also averaged over the land domain; Fig. 29d-f). This difference is consistent across all events, where ALADIN11 simulates more ice-phase clouds at the expense of liquid-phase clouds at low levels, and also more upper-level clouds overall. The two reanalyses tend to fall in between, with ERA5 simulating more clouds (primarily ice-phase) over land than MERRA-2 (mainly liquid-phase) across all events. The bias towards producing more ice- instead of liquid-phase clouds in ERA5 was also found to occur at West Antarctic sites [82], who confirmed this underestimation by comparing the reanalysis to various ground-based instruments and radiosonde measurements. The authors found that the liquid water path was often severely underestimated, and that the ice-phase cloud bias was strongest in cold and clear-sky conditions as well as during cases where mixed-phase layered clouds were observed. This suggests that the ERA5 bias persist (and may even be more pronounced) in non-AR conditions, as may similarly be the case for ALADIN11. Both ERA5 and ALADIN11 use simplified microphysics schemes [57, 58, 83], which appear to be too efficient in promoting ice crystal formation and glaciation processes within clouds over Antarctic land, whereas the ICE3-OCND2 microphysics used in AROME [59, 60] include more stringent conditions for ice initiation in clouds [84]. We also examined temperature cross-sections to determine whether vertical air temperature gradients could explain these differences, but did not find that ALADIN11 or ERA5 is consistently colder aloft compared to MERRA-2 and HCLIM.

Despite these ice-phase cloud biases in AROME11 and ERA5, the downward longwave radiation anomalies are fairly similar across all models (Figs. 2-15f). Assuming that the clear-sky longwave radiation in ALADIN and AROME is comparable, this implies that the total emissivity of the predominantly higher-level clouds in ALADIN11 remains comparable to that of the lower liquid-phase clouds in AROME, likely because the greater ice water content and total cloud cover in ALADIN11 balance out the fact that liquid-phase clouds typically emit more longwave radiation. Based on clear-sky longwave radiation from ERA5 and MERRA-2, the increase in longwave radiation over the Jun 2019 and Mar 2022 events was entirely due to the cloud radiative effect in both reanalyses (Figs. 6, 28 and 33). During both events, the water vapour radiative effect was even slightly reduced after the main AR day, likely due to air drying following intense precipitation.

Fig. 4

Cross-section of the total cloud water content along the central axis of the AR on the peak day of the event.

Fig. 5

a-e: As in Fig. 1e but zoomed in on the AR land domain window.

Fig. 6

Total downward longwave radiation (purple) and clear-sky longwave radiation (coral) based on ERA5 during the four AR events, averaged over the respective AR land domains (see Fig. 1). The same figure based on MERRA-2 is shown in Fig. 28. Regional variability of the longwave component anomalies are shown in Fig. 33

Shortly after the AR reaches sea ice on Jun 9, ALADIN11 begins simulating clouds and continuous snowfall, whereas in AROME, snowfall does not start until just before the coast (Fig. 4; on the following days all models simulate snowfall on sea ice as shown in Fig. 5). Likewise, the two reanalyses differ in their representation of clouds and snowfall: ERA5 primarily simulates ice-phase clouds (Fig. 34d) and moderate local snowfall at the coast, while MERRA-2 predominantly simulates a shallower, lower layer of liquid-phase clouds over sea, with more gradual snowfall rates along the AR path (Fig. 13d,e). The pronounced temperature drop at the centre of the cross-section in MERRA-2 (Fig. 4e), and local reduction in 4-day accumulated precipitation in all models (Fig. 5a-e) marks the area of the landfast ice and grounded icebergs west of the WeIS (Fig. 5f,g), where sea ice concentrations (SIC) are near 100% in all models (Fig. 21). Downstream of the AR, to the southwest, are large coastal polynyas [85, 86] and the climatologically mild and ice-free Vestfold and Larsemann Hills (Fig. 5g). In MERRA-2, SIC are especially low up- and downstream of the landfast ice region, causing the cold and dry anomalies at the centre of the cross-section. The average wind speeds over the AR land domain do not differ greatly across models ($\sim$15 m s$^{-1}$ on event days; Fig. 29h). The non-hydrostatic AROME configurations simulate more snowfall over the broader land domain (Fig. 2c), particularly at the coast (Fig. 5). Between AROME11 and AROME2.5 there are no significant precipitation differences (also not at most station sites or during the other events), except from slightly higher local precipitation maxima in AROME2.5, likely due to better-resolved orography. Generally we note a near-linear decrease in maximum local precipitation values from higher to lower resolution models during all events (except for Mar 2022, where maximum precipitation in AROME11 is higher than that in AROME2.5).

3.2 February 2020: melt event on the antarctic peninsula

• (a) Event Evolution and Surface Impact

The AR that initiated the Feb 2020 heatwave reached the centre of the AP on Feb 6 and was followed by additional moisture intrusions from the north and east that prolonged the heatwave, leading to peak temperatures, maximum melt extent, and highest precipitation over the AR land domain on Feb 10 (Fig. 7). We focus the atmospheric analysis on the AR on Feb 6, while also discussing the models’ simulations of the event-long accumulated melt and precipitation.

Fig. 7

Time series as in Fig. 2. Here for the Feb 2020 event, as the weighted sum or mean over the AR land domains (yellow outline in Fig. 1b)

Although the area-averaged skin temperatures of the models are comparable (Fig. 7a), close to 0$^{\circ }$C when averaged over the entire AR land domain, there is considerable variability in the extent of skin temperature-based surface melt (Fig. 7b). All models significantly underestimate melt extents across all observed melt events. Comparing the 3-month satellite data surrounding each event to ERA5, we find that melt in the Feb 2020 and Dec 2023 events is consistently underestimated or absent, also during non-AR conditions. AROME2.5 and ERA5 ’simulate’ the maximum melt extent of 100,000 km$^{2}$ on Feb 10 reasonably well. However, melt largely decreased after Feb 11 in all models, while satellite data indicate a gradual decline that extended until the end of March (not shown). MERRA-2 underestimates the maximum melt extent by 50% due to low skin temperatures, preventing any melt initiation before Feb 8 in melt-prone regions, which are offset by overestimated temperatures elsewhere resulting in overall comparable skin temperatures. Satellite data show that melt extent had already exceeded 50,000 km$^{2}$ in the weeks preceding the event, whereas ERA5 (driving HCLIM) recorded less than 20,000 km$^{2}$ on pre-simulation days. Lowering the melt threshold to -1$^{\circ }$C (Fig. 42) produced more comparable results with satellite data – also for the Mar 2022 event – although substantial model and temporal differences persisted (and pre- and post-AR melt during Feb 2020 was not simulated even with a -2$^{\circ }$C threshold). In section b, we show that surface temperatures at all three AWS sites exceed those in the models from Feb 10 onward, suggesting that the surface response to atmospheric warming may be a key source of error. We further discuss regional variability in melt days in section c.

A substantial portion of the precipitation during the heatwave fell as rain ($\sim$50% on Feb 10; Fig. 7d), with notable model variability on Feb 6 due to differences in the regional distribution and total precipitation magnitude. For example, the AR in ALADIN11 precipitates most of its moisture as snow before reaching the coast, leaving a reduced amount to fall mainly as rain over land (Fig. 4c). The daily amount of total precipitation from Feb 4 to Feb 13 also varied significantly across models, without a time-consistent bias, i.e. different models exhibited more or less precipitation depending on the day (Fig. 7c).

The net downward SEB is strongly influenced by shortwave radiation due to the northern location and summer season, leading to significant diurnal variations in the net SEB. Underlying this diurnal variability, the models show a slight decrease in shortwave radiation between 20 and 50 W m$^{-2}$ due to AR-related cloud development (Fig. 30a), which was likely moderated by the persistent high-pressure system during the heatwave, as well as foehn-induced cloud thinning [35, 37]. Still, all models simulate an increase in the net SEB during the peak days of the event, primarily driven by a rise of $\sim$40 W m$^{-2}$ in longwave radiation on Feb 6 and 10 (Fig. 7f) and a 20-50 W m$^{-2}$ increase in the downward sensible heat flux depending on the model (Fig. 30b). When examining 2D field anomalies of the four SEB components for all events (Figs. 381-414), we note a familiar pattern where the components within radiative and turbulent fluxes counteract each other: areas with increased longwave radiation typically coincide with reduced shortwave radiation due to cloud formation, while areas with high sensible heat flux anomalies (often linked to foehn winds) show negative latent heat flux anomalies from sublimation in response to the drier air masses [87,88,89]. Unlike the other events where the local turbulent flux components tend to balance each other, the negative latent heat flux anomalies on the leeward side during the Feb 2020 event were too low to offset the strong sensible heating, possibly due to limited snow/ice for sublimation. These results are in line with [35, 37] and [18], who concluded that both sensible heating (on the eastern AP) and radiation (mainly on the western AP) were responsible for the surface warming and melt intensity.

• (b) Evaluation with Station Data

Because the AWS used to evaluate the Feb 2020 event are located either on islands or along the coast, the nearest grid points in all models were situated over sea, except for AROME2.5 at Prospect Point. We attempted to evaluate the models using the nearest grid points over land, which in some cases positioned the closest point more than 100 km away from some AWS locations, resulting in larger mean errors, especially in lower-resolution models. We therefore disregard the land/sea issue to ensure a fair comparison of the models, but emphasize the importance of topographical details alongside general errors introduced by the larger grid cells compared to point-based AWS observations.

The relative humidity of the models at Robertson Island and Prospect Point is at most times over 10% too high in all models, however this is likely a result of the grid points being over sea, as relative humidity in AROME2.5 at Prospect Point compares well with the AWS (Fig. 8b,h). Temperatures on the other hand are between 2 and 12$^{\circ }$C underestimated, especially at Robertson Island (located slightly southeast of Esparanza base), where 11.5$^{\circ }$C and 12.5$^{\circ }$C were observed on Feb 6 and Feb 10 (Fig. 8a). Conditions might have been even warmer on Feb 6 or on Feb 8/9, when data at all three stations were very sparse, possibly due to quality control removals. At Robertson Island and Prospect Point, all models reached 0$^{\circ }$C on Feb 6 and showed minimal variability afterward (except for diurnal variations in ERA5 and AROME2.5 respectively), as all available energy was used for melting. At Hugo Island, variability was also underestimated but models simulate constant temperatures near 2.4$^{\circ }$C, which is still 2 to 5$^{\circ }$C too low during the event days (Fig. 8d). While all models thus simulate melt conditions at the three AWS sites, the consistently underestimated temperatures suggest that the melt volume is likely also too low. The cold model bias across all three stations is amplified/initiated shortly after the first AR reached land on Feb 6, suggesting that models had difficulties to capture surface responses to the extreme conditions. However, we note that while the observed data from the event are partially quality-controlled, they remain temporally incomplete and may still contain errors, particularly in wind speed (Fig. 8c,f,i). Here the models simulate sharp increases up to 25 m s$^{-1}$ on Feb 6 and 20 m s$^{-1}$ on Feb 10, whereas the AWS show little response to the AR event, likely due to wind being particularly influenced by local orography, which the models cannot resolve. The three stations are in close proximity to the AR cross-section, which we discuss in section c regarding cloud dynamics, precipitation and the radiation budget.

Fig. 8

10-day time series of surface temperature, surface relative humidity and surface wind speed at Robertson Island, Hugo Island and Prospect Point (red dots in Fig. 1f). Weather station data in green, other colours represent the different models at the nearest grid point. Wind speed data was not available at Robertson Island

While the surface warming peaked on Feb 10, the highest temperatures and wind speeds at 800 hPa in Rothera were recorded on Feb 6 (Fig. 24), validating that this AR was the strongest among the following intrusions occurring until Feb 10 (at least at Rothera). All models accurately capture the well-mixed boundary and relatively unstable conditions throughout the lower, near-surface troposphere on Feb 6. At all radiosonde stations for the Feb 2020 and Mar 2022 events, AROME2.5 tends to simulate higher wind speeds at the surface but lower wind speeds aloft, while the opposite is observed for ERA5 (Fig. 24-27). The Feb 2020 event in particular exemplifies how ARs can trigger intense heatwaves with the strongest effects occurring several days after the initial AR if post-AR conditions remain favourable to surface warming. This is particularly true for summer ARs linked to melt-induced feedbacks in contrast to shorter-lived ARs like the Jun 2019 event in austral winter.

• (III) Clouds, Precipitation and Melt Extent

We selected a cross-section that shifts slightly south of the AR’s strongest central axis as it reaches the AP on Feb 6 (Fig. 1b), allowing us to capture the ARs impact and vertical structure over Larsen C rather than just a narrow land area. In AROME2.5, clouds and precipitation on Feb 6 are more concentrated around the coastline due to the models’ more precise orography and non-hydrostatic configuration (Fig. 9).

While sea ice covered the entire Weddell Sea to the east of the AP, the Bellingshausen Sea to the west was sea ice-free (Fig. 1b), allowing for additional oceanic moisture uptake over waters with temperatures of 2-3$^{\circ }$C. Clouds west of the AP are hence positioned at low levels around 800 hPa in all models and are predominantly in the liquid phase, with ALADIN11 also simulating some upper-level ice-phase clouds (Fig. 9 and 35). Although the total cloud water content near the land edge is nearly twice as high in the two AROME configurations compared to the other models, the local precipitation amounts across all models are quite similar. This is likely because ice clouds (which AROME lacks) are more effective in producing precipitation, as ice crystals can grow more quickly (by direct water vapor deposition; Bergeron process) compared to liquid cloud droplets which depend on processes like collision-coalescence and strong updrafts to grow and precipitate [90,91,92]. Therefore, the amount of precipitation in the models can be similar despite different processes driving it, similar to SEB influences on surface melt. Here the precipitation mainly falls as rain before, and snow after the AP in all models (Fig. 9). However, day-to-day variations in total precipitation over the AR land domain during the 10 simulation days are significant (Fig. 7c). The intense rain- and snowfall was confined to the west coast, with local accumulations in AROME2.5 exceeding 500 mm over the 7-day event (Fig. 10a). Due to the significant portion that fell as rain, it is possible that the rain heat flux (which is not included in models) amplified the observed melting on Alexander Island and George VI Ice Shelf, where models showed rainfall but still strongly underestimated melt compared to AMSR2 (Fig. 10f-l).

We find that the sensible heat flux along the east coast and on Larsen C is strongly enhanced, while the latent heat flux is reduced compared to pre-event conditions, indicating stronger foehn conditions that also result in increased shortwave radiation. The increase in foehn is most pronounced in the two AROME configurations, both locally (Fig. 392) and regionally-averaged (Fig. 30b,c). As longwave radiation anomalies over the eastern AP and Larsen C are of comparable magnitude across models (except for higher values in ERA5), we suggest that the underestimated foehn effect is the primary reason for the lack of melt areas in ALADIN11 and the reanalyses (Fig. 10).

Fig. 9

As in Fig. 4 but for the AR in Feb 2020 (see cross-section in Fig. 1b)

Fig. 10

As in Fig. 1f but zoomed in on the AR land domain window (yellow outline) for accumulated precipitation (a-e) and melt days (f-k) over the AR event days (Feb 5 to Feb 11), for each individual model (configuration). Central axis of the AR on Feb 6 is shown in black. l: Zoomed-in view based on REMA including the Larsen C, Larsen D, and George VI ice shelves as well as Alexander Island

In ERA5 the melt magnitude was as high as in AROME2.5 due to increased cloud coverage (Fig. 9d) causing stronger longwave radiation (Fig. 392i) even downwind after the mountain range, which compensated for the lack of foehn-related sensible and shortwave heating. According to AMSR2, melt over Larsen C had already occurred prior to the event and continued throughout the entire 7-day span, while in our models melt was only initiated after the first AR on Feb 6 and continued for up to 5 days over Larsen C and up to 2 days on land and Larsen D (Fig. 10). This example demonstrates that both longwave and sensible heat-driven surface warming can strongly contribute to increased surface melt, and that models can simulate comparable melt extents but different physical mechanisms. Following observation-based studies on the event that attributed the melt on the eastern AP and Larsen C primarily to foehn conditions [18, 35, 38], we suggest that ERA5 underestimates the turbulent fluxes, and overestimates cloud coverage and longwave heating over land and Larsen C. Thereby it not only simulates incorrect drivers, but also melt in the ’wrong’ regions: much of the melt in ERA5 occurs over the western land region of our AR domain, where no melt was observed by AMSR2 (Fig. 10f,j), while ERA5 fails to simulate melt over the foehn-prone region near Larsen D. Large persistent biases in ERA5’s turbulent fluxes and surface temperatures near West Antarctic ice shelves were also recently documented in [93], who mainly attributed to its low resolution and wrong topography. We find that, although driven by ERA5 at the boundaries, the regional melt distribution is more accurate in HCLIM (and even in ALADIN11, which despite a too low melt extent captures melt over Larsen D).

3.3 March 2022: heatwave in east antarctica

• (a) Event Evolution and Surface Impact

In only five days from Mar 13 to Mar 18 2022, the average temperature over the AR land domain rose from roughly -45 to -15$^{\circ }$C, with notable differences between the models, with ERA5 and MERRA-2 being $\sim$5$^{\circ }$C colder than HCLIM (Fig. 11a). Mar 18 also represents an exceptional peak in area-wide precipitation (Fig. 11c), air temperature, wind speeds and IVT in all models (Fig. 31g-i). As much as 10% of the total precipitation fell as rain in HCLIM (Fig. 11d), which is significant given the large extent of the considered AR land domain, where precipitation occurred over vast areas of the plateau, even south of 70$^{\circ }$S (Fig. 14). Rainfall was confined to the coast (not shown) and is likely underestimated in the reanalyses. For instance, [40] reported rainfall at Dumont-Durville (DDU) on Mar 17 using a Micro Rain Radar (MRR), which here is only captured by ALADIN11 (Fig. 43b). The other two HCLIM configurations instead simulate small amounts of rainfall on Mar 19 at DDU, even though the MRR only observed snow on that day. While the amount of observed rainfall is unknown, [40] noted 9 mm of snowfall from Mar 15 to 20, which most models underestimate by about half, with the highest accumulated precipitation (including rainfall) reaching 5.3 mm in the two AROME configurations (Fig. 43a).

Fig. 11

Time series as in Fig. 11. Here for the Mar 2022 event, as the weighted sum or mean over the AR land domains (yellow outline in Fig. 1c)

Melt was also underestimated in the models, as it only occurred on Mar 17, with no melt at all simulated in MERRA-2 (Fig. 11b). Setting the melting threshold to -1$^{\circ }$C resulted in overestimated melt extents in the HCLIM configurations ($\sim$35,000 km$^{2}$ on Mar 17), highlighting the high threshold sensitivity. However, melt extents in ERA5 and MERRA-2 are still underestimated, with minimal melt on Mar 18 across all models, even with a -2$^{\circ }$C threshold. This suggests that melt feedback effects in the models were insufficient to sustain or amplify surface melt, and/or indicates a lack of model physics for triggers such as latent heating from rain.

• (b) Evaluation with Station Data

At both stations (Aurora-BN at 71$^{\circ }$S and Dome-A at 80$^{\circ }$S), surface temperatures in all models were fairly well simulated (but on average 3-5$^{\circ }$C too cold in MERRA-2, and 2-3$^{\circ }$C too warm in AROME; Fig. 12). Wind speeds at Dome-A are also fairly well captured, while relative humidity was overestimated by 15% on average. The temperature profiles at the three radiosonde locations were fairly similar across models and sounding observations, all showing a deepening of the mixed layer, where even at 400 hPa temperatures were over 15$^{\circ }$C higher than on Mar 13 (Fig. 25-27; pre-event profiles are not shown). The observed surface wind directions indicate a shift from southerly to north/northeasterly winds, where wind speeds in the boundary layer were very variable across models, with AROME2.5 simulating the strongest surface winds on the main event days, which are overestimated by 5-10 m s$^{-1}$ compared to the radiosonde soundings. At Casey station, the highest temperature was reached as early as on Mar 16 representing the start of the AR intrusion, followed by Concordia further inland on Mar 17, and lastly DDU on Mar 18 due to the melting conditions which prolonged the surface warming. While the observed data at Concordia were too sparse to draw conclusions, the conditions at the coastal stations were better captured in ERA5 and HCLIM than in MERRA-2.

Fig. 12

10-day time series of surface temperature, surface relative humidity and surface wind speed (at 600 hPa) at Dome-A station and Aurora Basin North station (red dots in Fig. 1f). At Aurora-BN, MERRA-2 did not have data at the 600 hPa level, only at 500 hPa, where wind speeds in all models were significantly larger, hence we omitted MERRA-2 in this case. Weather station data in green, other colours represent the different models at the nearest grid point. Relative humidity and wind speed data was not available at this site

• (c) Clouds, precipitation and melt extent

Following [40], we also find that the extensive temperature records set across East Antarctica were primarily driven by enhanced longwave warming from the radiative effects of liquid clouds: the clear-sky radiation anomalies in ERA5 and MERRA-2 show that the water vapour radiative effect only intensified after the air masses moved inland from the coastline, while the cloud radiative effect is significantly positive over the entire domain, up to 70 W m$^{-2}$ higher on Mar 17 compared to pre-event days in both reanalyses (Fig. 6c, 28c and 33). Over the entire AR domain, the turbulent fluxes played a minor role and are only slightly elevated in ERA5 and MERRA-2 (Fig. 31b,c) due to less negative latent heat flux anomalies in coastal, foehn-prone regions (Fig. 403p-t). However, all models show that the areas where extensive melt occurred coincided with regions where the longwave anomaly is strongly reduced or even negative, and instead the sensible heat anomaly dominates the positive SEB (Fig. 403). The negative shortwave anomaly at these coastal sites is also weaker, while in the inland AR domain shortwave radiation is strongly reduced (Fig. 403k-o), with area-averaged shortwave radiation reduced to half of pre-event levels (Fig. 31a), consistent with [21]. Nonetheless, even the net mean shortwave radiation remained above 10 W m$^{-2}$, and thus contributed to coastal melting and surface warming farther south. We find that clouds are in the liquid or mixed phase even $\sim$1,000 km inland (except in ALADIN11; Fig. 36), with yet again large differences across models regarding cloud content, precipitation amount and type, as well as surface temperatures across the AR cross-section (Fig. 13).

Fig. 13

As in Fig. 4 but for the AR in Mar 2022 (see cross-section in Fig. 1c)

Fig. 14

As in Fig. 1g but zoomed in on the AR land domain window (yellow outline) for accumulated precipitation (a-e) and melt days (f-k) over the AR event days (Mar 14 to Mar 21) for each individual model (configuration). Central axis of the AR on Mar 17 is shown in black

As in the other events, ALADIN11 and ERA5 primarily simulate upper-level ice-phase clouds, and ALADIN11, in particular, shows no liquid-phase clouds over land (Fig. 36c). The lack of low-level clouds in ALADIN11 is also evident in reduced shortwave radiation values during all events (Figs. 29a-32a). However, since the total water content of ALADIN’s ice-phase clouds is higher than that of the liquid-phase clouds in the other models (also shown in Fig. 31d-f), the longwave radiation is almost identical to that in AROME and MERRA-2, both in area-averaged values (Fig. 11f) and cross-sections (not shown). Longwave radiation is highest in ERA5, which simulates a thicker layer of mixed-phase clouds, even south of 75$^{\circ }$S. However, the upper cloud layer in ERA5 decouples from the surface further inland, as indicated by a strong, gradually lifting inversion (Fig. 13d). This decoupling results in significantly colder surface temperatures in ERA5, especially compared to HCLIM, which is 5-10$^{\circ }$C warmer at the end of the cross-section.

Both the heavy precipitation and melt areas are constrained to the coast and are highest in AROME configurations (Fig. 14). AROME2.5 and ALADIN11 simulate a larger melt area compared to the other models but still fail to capture the full extent and duration of the melt event. In the AR cross-section shown in Fig. 13d,e (which traverses the extensive surface melt area observed by AMSR2), temperatures remain just below 0$^{\circ }$C in ERA5 and MERRA-2. Coastal precipitation is higher in MERRA-2 than in ERA5 (Fig. 14e) and predominantly composed of snow instead of rain, possibly explaining why surface temperatures do not surpass 0$^{\circ }$C.

3.4 December 2023: mid-summer AR over the west antarctic ice sheet

• (a) Event Evolution and Surface Impact

The West Antarctic AR event in the summer of 2023/2024 triggered a sharp skin and surface temperature increase from Dec 28 to Dec 30 of +7$^{\circ }$C over the AR land domain (Fig. 15a). Surface winds and IVT increased from 3 to 10 m s$^{-1}$, and from 25 to 75-100 kg m$^{2}$, respectively, while none of the models showed significant changes in the area-averaged turbulent fluxes (Fig. 32). However, in section c, we show that this is a result of compensating regional positive and negative turbulent anomalies. The primary energy change contributing to the warming was the longwave radiation response to intense cloud formation (predominantly ice-phase in ALADIN11, mixed-phase in ERA5, and liquid-phase in MERRA-2 and HCLIM; Fig. 32d-f). This longwave increase outweighed the shortwave blocking induced by clouds, leading to a moderate and not significant rise in the total surface energy balance (SEB), with ERA5 showing the highest increase on the main event days due to higher net turbulent fluxes (Fig. 15e). This again highlights that although large-scale AR-related warming is typically driven by longwave radiation, the greatest discrepancies between reanalyses/models often lie in the turbulent heat flux components.

Fig. 15

Time series as in Fig. 2. Here for the Dec 2023 event, as the weighted sum or mean over the AR land domains (yellow outline in Fig. 1d)

While summer surface melt is common on the western and eastern sides of the AR land domain, it is rare in the shielded regions east of the Abbot and west of the Stange Ice Shelf (Fig. 18g) [94]. Consequently, the only area where AMSR2 detects melt is at the western part of the AR land domain, for a period of two days beginning on Dec 30 (Fig. 18f), whereas the models only indicated melt when the threshold was lowered to at least -1$^{\circ }$C (Fig. 42f,i,l).

• (b) Evaluation with Station Data

The nearest AWS with available data for the December 2023 event is located on Thurston Island, $\sim$10$^{\circ }$ west of the AR cross-section. Although precipitation on the day of the AR was minimal at Thurston Island station, it is situated within the intense AR moisture plume (Fig. 1d,h). While the models indicate increased wind speeds and higher temperatures at Thurston Island starting on Dec 27, the AWS recorded a significant temperature increase only on Dec 30, reaching 2$^{\circ }$C, which none of the models capture. Given the station’s location in the AMSR2-observed melt region, this agrees with the AMSR2-detected melt beginning on Dec 30, which the models did not simulate. AROME2.5 is the only model that does not consistently underestimate the surface temperature during the event, except on Dec 30, when the observed temperature reached 2$^{\circ }$C (Fig. 16a). It also compares best against relative humidity observations, which are overestimated by 5% on average in the other models. We tested the location of the closest grid points of all models and confirmed that all points are assigned to land and very close to the original AWS coordinate. This suggests that the better performance of AROME2.5 is due to more detailed orography and model physics (Table 3) rather than proximity.

Fig. 16

10-day time series of surface temperature, surface relative humidity and surface wind speed at Thurston Island station (red dot in Fig. 1g). Weather station data in green, other colours represent the different models at the nearest grid point. Wind speed data was not available at this site

• (c) Clouds, Precipitation and Melt Extent

Similar to the previous events, AROME configurations simulate more concentrated surface liquid-phase clouds over the open ocean, with minimal cloud formation downstream of the coast (Figs. 17 and 37). The pronounced  50 W m$^{-2}$ longwave radiation anomaly observed in AROME over land (Figs. 414f,g) thus primarily arises from the radiative water vapour effect. We also tested whether a steeper vertical temperature gradient in AROME may have contributed to the elevated longwave flux, but did not find that the upper air was warmer (or the lower air cooler) in AROME compared to the other models. In ALADIN11, the cloud radiative effect due to high-altitude ice-phase clouds further increases net longwave radiation (Figs. 15f and 414 h). Despite increased cloud cover in ALADIN11, solar blocking across HCLIM was comparable (Figs. 32a and 414k-m), likely due to the higher opacity of low-level liquid clouds compared to high-level ice clouds. The less negative shortwave anomaly in ERA5 and MERRA-2 compared to pre-event conditions results from generally lower absolute shortwave radiation values, causing smaller differences during main event days (Fig. 32a).

In AROME and ERA5, the AR resulted in rainfall on sea ice, likely contributing to its melt at $\sim$0$^{\circ }$C (Fig. 17a,b,d), followed by heavy snowfall over the Abbot Ice Shelf and coastal regions, peaking at 97 mm in 6 days in AROME2.5 (Fig. 18). Beyond the initial coastal orographic precipitation, most models (except MERRA-2) also simulate cloud formation and increased snowfall near the Jones Mountains (Fig. 18g) at 700 hPa (Figs. 13 and 18). The region south of the mountain range, which has been identified as favorable for foehn effects [24], showed pronounced positive sensible heat flux anomalies in all models (Fig. 414u-y) that extended to the Ellsworth Mountains, followed by a similar yet weaker sensible heat flux anomaly pattern. These local turbulent heat flux anomalies are approximately half the magnitude of the longwave radiation anomalies, highlighting their considerable contribution to the net SEB at a local scale, particularly downstream of mountain ranges within AR pathways. Similar to the Mar 2022 heatwave, the precipitation and SEB impacts of this AR reached extensive inland areas, including Pine Island and Thwaites Glacier. Even further downstream at the Transantarctic Mountains, local snowfall exceeded 70 mm in 6 days in both AROME configurations (Fig. 18).

Fig. 17

As in Fig. 4 but for the AR in Dec 2023 (see cross-section in Fig. 1c)

Fig. 18

As in Fig. 1h but zoomed in on the AR land domain window (yellow outline) for accumulated precipitation (a-e) over the AR event days (Dec 28 to Jan 2), for each individual model (configuration). f: Surface melt days, only shown for the observed AMSR2 data as the models do not simulate melt (i.e. skin temperatures are always below 0$^{\circ }$C). Central axis of the AR on Dec 29 is shown in black. g: Zoomed-in view based on REMA including the Stange and Abbot ice shelves, as well as Pine Island and Thwaites Glacier, and the Jones, Ellsworth, and Transantarctic Mountains

We suggest that the coastal sea ice (although modest in extent) may have mitigated more intense coastal melt and rainfall, as warm, moist air masses were lifted and weakened at the sea ice edge following moist isentropes (Fig. 17). This buffering effect is becoming increasingly relevant in light of the record-low Antarctic sea ice extents observed in recent years.

4 Discussion

We evaluated and analyzed the surface impacts and atmospheric structures of four recent AR events across different Antarctic regions using ERA5, MERRA-2, and three HCLIM simulations driven by ERA5. Our analysis shows that different ARs can have very different surface expression and highlights the critical role of atmosphere-surface coupling in Antarctica on ARs when determining their impacts, particularly on ice sheet mass budget.

ARs are distinctive events with their own individual characteristics, determined by spatial and temporal location as well as air mass. Our analysis also identifies strong temporal and spatial differences in the impact of ARs on the surface of Antarctica. Given the importance of ARs to AIS mass budget, both as a source of accumulation, and potentially ice shelf destabilisation, distinguishing factors that determine exact surface impacts is crucial. We find a temporal variability in atmosphere and surface responses as well as effects from the presence or absence of sea ice, and proximity to the coast as well as high relief topography all play important roles in determining the exact impact of an AR event.

The selection of these four unique events highlights the diverse surface impacts of ARs, which are influenced by location, meteorological conditions, seasonality, and AR characteristics such as wind speeds and humidity content. For example, Event 1 (Jun 2019) led to significant snow accumulation and cloud-induced longwave radiation in East Antarctica during mid-winter, while Event 2 (Feb 2020) was marked by extreme surface temperatures and surface melt driven by foehn winds and warm advection over the AP and Larsen C Ice Shelf. The prominent East Antarctic spring heatwave (Event 3; Mar 2022) exhibited an unprecedented +38$^{\circ }$C surface temperature anomaly due to large-scale longwave radiation and local sensible heating in coastal areas, where increased melt was observed. Lastly, Event 4 (Dec 2023) highlights how sea ice and ice shelves can dissipate much of the moisture and strength of ARs before they reach the coastline, which weakens inland impacts; but in this summer case can still cause anomalous surface warming (via longwave and sensible heating) up to the Transantarctic Mountains.

These differences emphasize the importance of regional factors, such as sea ice extent, topography, and seasonal conditions, in modulating AR impacts. Understanding these variations is crucial for accurately assessing the Antarctic Ice Sheet’s surface mass balance and the potential for ice shelf destabilization under changing climate conditions.

The dominant driver of increased large-scale warming and surface melt during all AR events was longwave radiation, which was significantly enhanced over the entire AR land domains, except for the strongly foehn-impacted east coast of the AP during the Feb 2020 event, where sensible heating was dominant. In the Mar 2022 and Dec 2023 events, foehn-prone regions also showed enhanced sensible heat fluxes and less negative shortwave radiation anomalies which were usually higher than longwave anomalies. Only during the Jun 2019 event did longwave radiation alone cause the temperature increase, due to the flat coastline/ice shelf and the absence of solar radiation. The increased longwave radiation in the Jun 2019 and Mar 2022 events was almost entirely driven by cloud formation, while during the two summer events (where the total longwave anomaly was smaller) the water vapour radiative effect contributed at least half (based on ERA5 but this also applies to MERRA-2). We suggest that the smaller area-averaged impact of the cloud radiative effect (and the higher impact of the water vapour radiative effect) during the Feb 2020 and Dec 2023 events is due to 1) foehn-induced cloud thinning over large regions and 2) steeper vertical temperature gradients during summer ARs, where higher upper-air temperatures contrast with a surface that remains close to the melting point. Also, if increased longwave radiation anomalies in summer are primarily cloud-driven, the resulting negative shortwave anomaly would offset some of the longwave-induced warming, thereby reducing the total SEB anomaly.

Foehn-prone regions during summer ARs are most susceptible to AR-induced melting, as sensible heating, solar radiation, and clear-sky longwave anomalies combine to cause extensive melt, as observed during the Feb 2020 heatwave, where melt was primarily observed in foehn-affected regions. Over larger areas (i.e. our AR land domains), AR-induced turbulent flux anomalies are often not significant, as regions of enhanced sensible heat flux usually overlap with reduced latent heat flux areas. These detailed regional foehn-related patterns are poorly represented in the two reanalyses, which explains why ERA5 and MERRA-2 either simulate insufficient or incorrect melt regions (e.g. during the Feb 2020 event, where ERA5 overestimates melt on the western AP due to cloud-induced longwave radiation).

Key processes that determine how ARs evolve and affect the surface in models also include realistic snow pack properties that allow surface net energy fluxes to evolve during ARs. Similarly, the presence or absence of sea ice and coastal polynyas is important both for the vapour transport and energy fluxes of ARs but also in terms of orographic precipitation. We highlight critical areas for model improvement: all models underestimated observed melt durations and extents during and after ARs, with MERRA-2 performing the worst due to persistent cold biases, most likely due to inadequate surface schemes in these versions of the models. The reported melt extents of models in this study are based on skin surface temperature and sensitive to the melting threshold, with a -1$^{\circ }$C threshold resulting in more comparable melt extents to satellite-based AMSR2 data, although the melt duration is still often underestimated. The low melt extents align with model underestimations of surface temperatures at six out of eight investigated AWS sites (excluding the Mar 2022 event), where MERRA-2 again showed the strongest deviations. Consistent with the cold surface biases, models overestimated relative humidity at all stations except for Amery-G3. The representation of surface conditions was especially poor during the Jun 2019 event, where models underestimated temperatures at the AmIS and the inland Panda station by 10$^{\circ }$C on average.

Upper-air temperatures at the five radiosonde profile sites matched observations well (which however are assimilated in reanalyses), which indicates that the problem primarily lies in the models’ surface response during extreme AR conditions. As some of the temperature and moisture anomalies were already present before the AR events, our findings would benefit from a long-term climatological bias analysis of the reanalyses and model configurations.

Overall, the considerable discrepancies across models argues for considerably upgraded and systematised observations in Antarctica, particularly those including radiation and enabling surface energy budget calculations, to assist in evaluating and developing models. The poor performance in simulating melt in these model versions also highlights the need for improved representations of cloud processes, energy budgets, and surface snow and ice schemes to better assess AR impacts on Antarctic SMB, ice shelf stability, and the ice sheet contribution to global sea-level rise. Addressing these model discrepancies in a climatological analysis would provide insights into how much they persist under non-AR conditions, helping to clarify the uniqueness of AR impacts on Antarctica.

5 Conclusions

We compare the performance of three different model set-ups and two reanalysis products in simulating four different AR events in Antarctica. The ARs were selected as representative of different types of AR both in terms of season, location, and melt and precipitation extremes. We found some systematic biases in the surface impacts of ARs, particularly related to atmosphere-surface coupling and (likely) poorly resolved surface snow schemes. In particular, there is a high threshold sensitivity to surface melt and melt feedbacks in summer are insufficient to sustain melt to the extent shown in satellite imagery. Large-scale surface warming is generally driven by longwave fluxes, indicating the importance of clouds. However, melt primarily occurs over foehn regions, where sensible heat flux and shortwave anomalies are larger than the net downwelling longwave radiation anomaly. To improve estimates of future AR impacts, addressing model discrepancies related to turbulent heat components will be essential. Optimizing the SEB, including all its terms, is critical not only for accurately capturing surface melt and precipitation but also for preceding processes over the adjacent ocean. For example, our findings suggest that coastal polynyas can enhance oceanic moisture uptake and subsequent precipitation over land and ice shelves, which highlights the need for model improvements across both atmospheric and oceanic domains.

Lastly, the season, region, and orography significantly influence the surface impacts of ARs: downward shortwave fluxes are more critical in summer and northern locations, longwave fluxes are important year-round but particularly further inland, while turbulent fluxes dominate key melt regions near mountain ranges. Likewise, differences in the partitioning of snow and rain depend on the season but also on model physics. The high-resolution (2.5 km) AROME simulations suggest that rain may play a significant role in surface energy budgets, though this remains hypothetical due to the lack of observational data for validation. Similarly, foehn effects are highly sensitive to model resolution and cloud physics, with lower-resolution models generally underestimating the magnitude of energy fluxes and associated melt extents. These findings have implications for projecting future ice loss rates, as commonly used global climate models often have resolutions lower than the reanalyses examined here.