Regional climate models (RCMs) and reanalysis datasets provide valuable information for assessing the vulnerability of ice shelves to collapse over Antarctica, which is important for future global sea level rise estimates. Within this context, this paper examines variability in snowfall, near-surface air temperature and melt across products from the Met Office Unified Model (MetUM), Regional Atmospheric Climate Model (RACMO) and Modèle Atmosphérique Régional (MAR) RCMs, as well as the ERA-Interim and ERA5 reanalysis datasets. Seasonal and trend decomposition using LOESS (STL) is applied to split the monthly time series at each model grid cell into trend, seasonal and residual components. Significant systematic differences between outputs are shown for all variables in the mean and in the seasonal and residual standard deviations, occurring at both large and fine spatial scales across Antarctica. Results imply that differences in the atmospheric dynamics, parametrisation, tuning and surface schemes between models together contribute more significantly to large-scale variability than differences in the driving data, resolution, domain specification, ice sheet mask, digital elevation model and boundary conditions. Despite significant systematic differences, high temporal correlations are found for snowfall and near-surface air temperature across all products at fine spatial scales. For melt, only moderate correlation exists at fine spatial scales between different RCMs and low correlation between RCM and reanalysis outputs. Root mean square deviations (RMSDs) between all outputs in the monthly time series for each variable are shown to be significant at fine spatial scales relative to the magnitude of annual deviations. Correcting for systematic differences results in significant reductions in RMSDs, suggesting the importance of observations and further development of bias-correction techniques.

The largest source of uncertainty in 2100 sea level rise (SLR) projections, for a given representative concentration pathway (RCP), is from the contribution of ice sheets

The primary method of ice shelf retreat, when considered across the entire ice sheet, is currently through oceanic basal melting

RCMs are limited-area, physically based, nested models driven at the boundaries by lower-resolution global climate models (GCMs) or reanalysis datasets. The high resolution available from RCMs is important for capturing fine-scale climatic processes in regions of complex topography, such as föhn winds that occur over ice shelves on the Antarctic Peninsula

The atmospheric model dynamics, surface scheme, parametrisation, driving data, boundary conditions, domain, resolution and orography are all examples of components that contribute to systematic error

The two reanalysis datasets and six RCM simulation outputs compared in the paper. The label with which each simulation is referred to in the paper is given.

Six Antarctic-wide RCM simulations are compared, two from each of the Met Office Unified Model version 11.1 (MetUMv11.1), the Modèle Atmosphérique Régional version 3.10 (MARv3.10) and the Regional Atmospheric Climate Model version 2.3p2 (RACMOv2.3p2). Comparisons are also made to the reanalysis driving data of ERA-Interim and ERA5. The resulting eight Antarctic-wide datasets analysed in this paper are given in Table

Historic evaluation simulations are chosen to remove dependency on emission scenarios, which have been shown to introduce divergent trajectories of variables such as melt

The ensemble of Antarctic-wide RCM simulations examined in this paper is part of the Coordinated Regional Climate Downscaling Experiment (CORDEX;

ERA-Interim, produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), is a global reanalysis dataset spanning 1979–2019 with 6 h temporal resolution and approximately uniform horizontal resolution of 79 km spacing and 60 vertical levels up to 10 Pa

MAR is a hydrostatic RCM, specifically developed for the polar areas

RACMO is a hydrostatic RCM with a polar version developed to represent the climate specifically over ice sheets

The MetUM is a non-hydrostatic climate model, not specifically developed or optimised for use over the polar regions but adapted in these simulations for use over Antarctica

The RCM simulations examined in this paper all use an equatorial rotated coordinate system, where a quasi-uniform horizontal-resolution grid is defined over the region by first specifying the grid over the Equator with constant latitude and longitude spacing between each grid cell and then applying a rotation that takes the domain over the region of interest, for example Antarctica. Direct comparisons between the model output are made by regridding onto a common grid, with a common domain and spatiotemporal coordinates. Cubic precision Clough–Tocher interpolation

Map of Antarctica with some of the main regions and ice shelves labelled, made using the Quantarctica mapping environment

To study annual, seasonal and monthly variability separately, seasonal and trend decomposition using LOESS (STL)

Basic time series decomposition involves first approximating the trend component by applying a polynomial fit through the data. Subtracting this component gives the detrended data that are then split into seasonal sub-series (e.g. January, February), and an average of each sub-series gives the seasonal component of the data. Subtracting both the trend and seasonal components then gives the residual component of the series. STL is a more sophisticated procedure that allows options such as robust fitting (where the influence of outliers is limited) and also a time-varying seasonal component. The algorithm is iterative and involves two loops: the outer loop reduces the influence of outliers by assigning weights based on the magnitude of the remainder term; the inner loop involves estimation of the trend and seasonal components through iterative feedback

The seasonal component is allowed to vary smoothly over the time series, which is done by applying a LOESS (local regression) smoothing to the monthly sub-series with window length

The trend component is estimated using LOESS with a window of default size

In this paper temporal variability between the ensemble of Antarctic-wide datasets is assessed in several ways, including calculating the Pearson linear correlation coefficient between the outputs for each component of the time series and each variable of interest, quantifying differences in the mean of the time series as well as in the standard deviation of the seasonal and residual components, and calculating the root mean square deviation (RMSD) between the outputs for each variable of interest. Each metric is calculated for every grid cell in the domain, with Antarctic-wide plots showing spatial patterns. Differences in the monthly mean and standard deviation of the components are calculated over the 37-year 1981–2018 period. For snowfall and melt, differences at each grid cell are expressed as a proportion of the respective inter-annual deviations, providing some measure of the relative significance of differences at each location. The impact of systematic differences in snowfall and melt on estimates of ice shelf stability depends not only on absolute magnitudes but also on the relative magnitude against a baseline variance. The inter-annual baseline deviation at each grid cell is approximated as the ensemble average standard deviation in the trend component of the time series. Results presented in spatial maps then show the relative significance of systematic differences and are not simply dominated by the sites with the highest-magnitude snowfall/melt.

Variability in the ensemble of Antarctic-wide outputs (Table

The correlation for snowfall

The median correlation by grid cell in the residual component of the monthly time series between the 28 unique model pairs for snowfall

Results are presented for the correlation in the deseasonalised and detrended residual component of the time series between each of the 28 unique model output pairs. The correlation is computed at every grid cell, and for melt, grid cells where the ensemble 40-year average monthly melt is less than 1 mm water equivalent per month (mm w.e. per month) are masked as these regions only experience sporadic and insignificant-magnitude melt events, essentially equating to numerical noise in the simulations. The average grid-cell correlation across the entire ice sheet is then taken, and the results are given in Fig.

A spatial map of the median correlation in the residual component across the 28 unique model output pairs is plotted in Fig.

The 1981–2018 mean and standard deviation for each component of the monthly time series of the ice sheet total snowfall, average near-surface air temperature and total melt are displayed in Table

To understand how systematic differences vary spatially, the 1981–2018 mean and seasonal and residual standard deviations for the monthly time series of each variable are also computed at a 12 km grid-cell level. Since it is found that systematic differences in the mean and standard deviations are most pronounced between different models in the ensemble, results presented in Figs.

After aggregating across the ice sheet, the mean and standard deviation for each component of the monthly time series for total snowfall, average near-surface air temperature and total melt are given. Values for snowfall and melt are expressed in units of gigatonnes, while values for temperature are expressed in kelvin.

The difference to the ensemble average (model-ensemble avg) for the 1981–2018 time series of snowfall, in the mean

The difference to the ensemble average (model-ensemble avg) for the 1981–2018 time series of near-surface air temperature, in the mean

The difference to the ensemble average (model-ensemble avg) for the 1981–2018 time series of melt, in the mean

In Fig.

It can be seen that for snowfall the difference present in the mean of the time series has a similar spatial signature and sign to the difference in the standard deviation of the seasonal and residual components (e.g. Fig.

As with snowfall, there exist significant differences over both the ocean and the land for near-surface air temperature between the models, again particularly in the mean of the time series (Fig.

A land-only mask has been applied for melt in Fig.

The RMSD of the monthly time series is evaluated at each grid cell for each of the 28 unique output pairs of the ensemble. For snowfall and melt, the metric is scaled at each grid cell by the ensemble average inter-annual standard deviation, described here as the proportional RMSD value. The average is then taken across the ice sheet for each variable, and results are given in Fig.

The RMSD/proportional RMSD for snowfall

From Fig.

The percentage change in RMSD/proportional RMSD after adjusting for equal means as well as seasonal and residual standard deviations is significant for all variables, as shown in Fig.

The results presented in this paper show that for all variables studied, when considered across the entire ice sheet, the outputs that came from the same model (MetUM(011/044), MAR(ERAI/ERA5), RACMO(ERAI/ERA5)) exhibit the highest correlations in the time series as well as the smallest systematic differences and RMSDs. This is despite significant differences in resolution between the MetUM runs, which span the highest- and lowest-resolution RCM simulations made available from the Antarctic CORDEX project, as well as significant differences in the driving data for the two MAR and RACMO runs. Note that, although ERA5 is an update to ERA-Interim, the results in Table

Results therefore suggest that differing resolution and driving data are not primary contributors to large-scale spatial variability across the ensemble. Similarity in the spatial and temporal patterns between Antarctic-wide outputs of the same RCM with different driving data agrees with findings from

The magnitude of differences in snowfall and near-surface air temperature due to resolution is greatest over regions of sharply varying topography, such as the Transantarctic Mountains, the coastal slopes of the ice sheet and the Antarctic Peninsula. The representation of atmospheric processes occurring over mountainous regions including föhn winds that occur over the Antarctic Peninsula and katabatic winds occurring over the coastal slopes of East Antarctica is known to be resolution dependent

The same-model RCM simulations in the ensemble, as well as having identical model physics, parametrisation and tuning, also share factors such as the domain specification, ice mask applied, digital elevation model and boundary conditions. The relative contribution of these additional factors is explored in Sect.

The exact spatial domains differ between the RCM simulations as shown in Fig.

As well as having differences in the outer domain boundaries, the different models also have slight differences in the specified boundaries of the ice sheet due to different coordinates and ice masks used. This creates edge effects at the periphery of the ice sheet, which are particularly noticeable for melt in for example Fig.

Another important consideration when comparing RCM simulations is how the method of applying boundary conditions varies across the ensemble. In particular, although all RCMs examined are nudged at the boundaries within buffer zones, MAR and RACMO also use spectral nudging that constrains the large-scale circulation in the interior of the domain, while the MetUM instead uses a re-initialisation procedure. Spectral nudging involves applying the relaxation technique throughout the interior of the domain to the long-wavelength components of the climate model fields

The differences between DEMs used across the ensemble are plotted in Fig.

In this section, features including the domain specification, ice mask applied, digital elevation model and boundary conditions applied are argued to not be the primary contributors responsible for the large-scale systematic differences within the ensemble of model outputs. This result, in addition to the previously discussed secondary contributions of resolution and driving data towards large-scale differences, by way of elimination indicates that the joint influence of choices in model physics, parametrisation and tuning is the primary factor influencing large-scale systematic differences across the ensemble.

Specific features in the variability, identified and mentioned in Sect.

In Sect.

As with for correlation, the systematic differences shown between the outputs in the ensemble vary depending on the region and topography; see Sect.

In Sect.

The spatial nature and magnitude of variability present in an ensemble of current, state-of-the-art Antarctic-wide RCM outputs and global reanalysis datasets are examined for snowfall, near-surface air temperature and melt. This is done at a 12 km grid level, rather than across elevation bins, which reveals significant spatial patterns in correlation and systematic differences in the mean as well as the seasonal and residual standard deviation. Time series decomposition is used to split comparisons across an approximately inter-annual trend component, a periodic seasonal component and a monthly residual component, which is useful for impact assessments where knowledge of variability in the climate data across different timescales and climate drivers is needed.

It is found that the RCM outputs and reanalysis datasets show high correlation for the monthly time series of snowfall and surface temperature across the majority of Antarctica and the bounding Southern Ocean. Despite this, there exist significant differences, with respect to both magnitude and spatial scale, in the mean as well as the seasonal and residual standard deviations of the time series. In addition, high RMSD between the outputs is found for all variables and is particularly significant for melt with respect to the proportional values relative to annual fluctuations. The primary sources of large-scale, systematic differences between the simulations, for all variables and components, are identified as deriving from differences in the model dynamical core, the surface scheme, and the parametrisation and tuning. Differences in driving data, resolution, domains, ice masks, DEMs and boundary conditions are identified as having a secondary contribution. On local, fine spatial scales the relative contribution from different factors is more complex and differences in for example resolution are shown to have a more significant impact.

The variability in snowfall, near-surface air temperature and melt shown is expected to introduce significant uncertainty in estimates of the ice shelf stability with regard to collapse events, which as discussed may have an important contribution to future SLR estimates. It is suggested that the magnitude and scale of systematic differences across the ensemble preclude the direct use and interpretation of individual outputs in impact assessments regarding ice shelf collapse. Results show that removing systematic differences in the ensemble of outputs significantly reduces the average RMSD. Therefore, as concluded in

Figure

An example of STL decomposition applied to the monthly time series of snowfall

The difference for the 1981–2018 time series of snowfall, in the mean

The difference for the 1981–2018 time series of near-surface air temperature, in the mean

The difference for the 1981–2018 time series of melt, in the mean

The difference in the DEM used by each climate model relative to the ensemble average is plotted:

A density scatter plot showing the correlation between the difference in elevation for each model relative to the ensemble and the difference for near-surface temperature in the mean of the time series

The monthly output from all RCM simulations examined in this paper, as well as the processed data used for figures and tables, is available at

JC contributed in terms of conceptualisation, methodology, software, validation, formal analysis and writing – original draft. AL contributed in terms of conceptualisation, writing – review and editing, and supervision. AO contributed in terms of conceptualisation, data curation, writing – review and editing, and supervision. CK contributed in terms of data curation and writing – review and editing. JMvW contributed in terms of data curation and writing – review and editing.

The contact author has declared that none of the authors has any competing interests.

Publisher’s note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

J. Melchior van Wessem was partly funded by the NWO (Dutch Research Council) VENI grant VI.Veni.192.083. Computational resources for MAR simulations have been provided by the Consortium des Équipements de Calcul Intensif (CÉCI), funded by the Fonds de la Recherche Scientifique de Belgique (F.R.S.-FNRS) under grant no. 2.5020.11 and the Tier-1 supercomputer (Zenobe) of the Fédération Wallonie-Bruxelles infrastructure funded by the Walloon Region under grant agreement no. 1117545. Christoph Kittel was supported by the Fonds de la Recherche Scientifique – FNRS under grant no. T.0002.16 and by H2020 CRiceS. Andrew Orr was supported by the European Union’s Horizon 2020 research and innovation framework programme under grant agreement no. 101003590 (PolarRES). The code for analysis is written in Python 3.8.12 and makes extensive use of the following libraries: Iris

This research has been supported by the Engineering and Physical Sciences Research Council (grant no. EP/R01860X/1).

This paper was edited by Thomas Mölg and reviewed by Rajashree Datta and one anonymous referee.