1. Introduction
The future contribution of Antarctica to sea level rise is uncertain, with models disagreeing on the sign of the contribution over the next century (Edwards and others, Reference Edwards2021; Payne and others, Reference Payne2021; Seroussi and others, Reference Seroussi2024). Similarly, early 21st-century observations find that Antarctica can gain (Smith and others, Reference Smith2020) or lose (Rignot and others, Reference Rignot, Mouginot, Scheuchl, van den Broeke, van Wessem and Morlighem2019; Velicogna and others, Reference Velicogna2020) mass depending on the time period studied. Gains in ice sheet mass from increased snowfall can mitigate the sea level contribution due to dynamic ice loss (Medley and Thomas, Reference Medley and Thomas2019). The combination of spatial and temporal variability can have compounding impacts on estimating Antarctic-wide climate averages. While climate models project snowfall to increase as a response to warming temperatures (Frieler and others, Reference Frieler2015; Lenaerts and others, Reference Lenaerts, Vizcaino, Fyke, van Kampenhout and van den Broeke2016), empirical studies indicate high uncertainty in the sensitivity of accumulation to temperature (Monnin and others, Reference Monnin2004; Fudge and others, Reference Fudge2016; Nicola and others, Reference Nicola, Notz and Winkelmann2023). Satellite observations indicate that the ice sheet has gained mass since 2020 due to increased snowfall from East Antarctica (Wang and others, Reference Wang, Shen, Chen, Wang and Yu2025), but long-term dynamic ice loss is also projected to continue and likely increase (O’Neill and others, Reference O’Neill2025). Similar uncertainty surrounds the historical variability of Antarctic temperature, which is important to constrain the sensitivity of Antarctic climate to natural and anthropogenic forcings (Stenni and others, Reference Stenni2017).
Understanding modern climate change in Antarctica is challenging due to a deficiency of reliable observations. The number of Antarctic weather stations is limited (Hakim and others, Reference Hakim, Bumbaco, Tardif and Powers2020), satellite observations for gravimetry and altimetry only span recent decades (e.g. Rignot and others, Reference Rignot, Mouginot, Scheuchl, van den Broeke, van Wessem and Morlighem2019), and instrumental reanalyses are reliable only since the beginning of the satellite era (i.e. 1979; Schneider and Fogt, Reference Schneider and Fogt2018) and have significant limitations in accuracy (e.g. Dale and others, Reference Dale, McDonald, Coggins and Rack2017; Bromwich and others, Reference Bromwich, Ensign, Wang and Zou2024). A greater understanding of past trends is necessary to place modern changes in Antarctic accumulation and temperature variability in a historical context. Ice core records offer a means to reconstruct Antarctic climate variability (e.g. Kaspari and others, Reference Kaspari2004; Monaghan and others, Reference Monaghan2006; Schneider and others, Reference Schneider2006; Goosse and others, Reference Goosse2012; Steig and others, Reference Steig2013; Steiger and others, Reference Steiger, Steig, Dee, Roe and Hakim2017).
The spatial distribution of annually resolved ice core records across the Antarctic Ice Sheet (AIS) is uneven (Fig. 1a); there is a relatively high density of cores located on the West Antarctic Ice Sheet (WAIS) and Antarctic Peninsula and a relative sparsity on the East Antarctic Ice Sheet (EAIS). Low accumulation rates across the interior EAIS often preclude accurate annual dating and necessitate drilling in coastal areas, which are often more logistically and topographically complex (Vance and others, Reference Vance2016). The resulting uncertainty in the EAIS proxy record may limit our ability to quantify both historical Antarctic temperature variability and sea level contribution (Klein and others, Reference Klein2019; Medley and Thomas, Reference Medley and Thomas2019; Wang and Xiao, Reference Wang and Xiao2023; Eswaran and others, Reference Eswaran, Truax and Fudge2024). The EAIS is the largest contributor to Antarctic surface mass balance (Mottram and others, Reference Mottram2021), and future changes in EAIS mass are a major source of uncertainty in projections of future sea level rise (Seroussi and others, Reference Seroussi2024). In addition to spatial limitations in the ice core record, there are also temporal limitations; the majority of Antarctic ice cores were drilled in the 1990s and 2000s, and fewer records cover the 21st century than any other time period since 1800 CE (Stenni and others, Reference Stenni2017; Thomas and others, Reference Thomas2017). This limits our ability to place 21st-century Antarctic climate variability in a historical context, and to understand spatial variability by comparing ice core records with modern surface and satellite observations.
(a) Locations of 110 existing annually resolved ice core accumulation and water isotope records across the Antarctic Ice Sheet (Stenni and others, Reference Stenni2017; Thomas and others, Reference Thomas2017). (b) Locations of national research bases in Antarctica. Cores drilled as part of the US component of the ITASE program (Steig and others, Reference Steig2013) are indicated in (a) with cross-hatching.
Figure 1 Long description
A) The map of Antarctica displays locations of ice core records. Symbols indicate accumulation only, water isotope only, both records and future records. Accumulation only sites are marked with circles, water isotope only with squares, both records with triangles and future records with diamonds. The distribution is uneven, with a higher density in the West Antarctic Ice Sheet and Antarctic Peninsula and fewer in the East Antarctic Ice Sheet. B) The map shows national research bases across Antarctica, marked by colored dots representing different countries. The bases are spread across the continent, with clusters in coastal areas. The legend lists countries such as Argentina, Australia, Brazil, Chile, China, France, Germany, India, Italy, Japan, Korea, New Zealand, Norway, Russia, South Africa, Spain, United Kingdom, United States and others, each associated with a specific color. The spatial distribution highlights the international presence in Antarctica, with bases concentrated in accessible regions.
Ice cores have often been drilled to reconstruct local and regional climate signals (e.g. Curran and others, Reference Curran, Van Ommen and Morgan1998; Criscitiello and others, Reference Criscitiello2013; Philippe and others, Reference Philippe2016; Thomas and Abram, Reference Thomas and Abram2016). Vance and others (Reference Vance2016) provide a detailed description of many of the factors typically evaluated when selecting a site, targeting a regional rather than continental-scale signal. For the Mount Brown South ice core, Vance and others (Reference Vance2016) assess the glaciological characteristics (accumulation rate, timescale, windborne redistribution of snow and ice-flow velocity) and atmospheric circulation criteria (strong mid-latitude teleconnection and complementary relationship to other ice cores) focused on the Indo-Pacific. One challenge of climate reconstructions is how to extract a continental-scale signal from the noisy and conflicting regional signals. The PAGES Antarctica 2k working group has compiled annually resolved ice core-based reconstructions of Antarctic accumulation (Thomas and others, Reference Thomas2017) and temperature (Stenni and others, Reference Stenni2017) for the past 1000 years. They create regional composites for seven different Antarctic regions, which are then spatially averaged into a single continental-scale reconstruction. Accumulation and temperature trends vary greatly between regions; for example, snow accumulation is increasing in the Antarctic Peninsula but decreasing in Dronning Maud Land (DML) (Thomas and others, Reference Thomas2017), and the Antarctic Peninsula and West Antarctica are warming while the East Antarctic Plateau shows little significant warming (Steig and others, Reference Steig, Schneider, Rutherford, Mann, Comiso and Shindell2009; Stenni and others, Reference Stenni2017). In this paper, we seek an answer to the question: where should new ice core records be collected to maximize the potential continental-scale information gain?
Here, we present an objective, data-driven method for selecting shallow ice core drilling sites (spanning the past ∼200 years) to optimize reconstruction of any climate metric of interest at annual resolution. Our paper presents an optimization of two climate metrics: annual snow accumulation (ice equivalent) and average annual surface air temperature. These metrics were chosen to follow the PAGES Antarctica 2k studies (Stenni and others, Reference Stenni2017; Thomas and others, Reference Thomas2017), and are taken as spatially integrated across the entire AIS, as well as WAIS and EAIS separately. Increased snow accumulation over Antarctica has a mitigating effect on sea level rise; spatially-integrated accumulation quantifies the extent of this sea level mitigation, which constrains the contribution of the AIS to past sea level change (Medley and Thomas, Reference Medley and Thomas2019). Average Antarctic surface air temperature can constrain the Antarctic temperature-accumulation relationship by quantifying the Clausius-Clapeyron sensitivity (Nicola and others, Reference Nicola, Notz and Winkelmann2023). More generally, it constrains the response of Antarctic climate to natural and anthropogenic forcings (Stenni and others, Reference Stenni2017), such as climate sensitivity to CO2 (e.g. Uemura and others, Reference Uemura2018). We use the terms ‘annual Antarctic accumulation’ and ‘annual Antarctic temperature’ moving forward to refer to these metrics.
We use an optimal sensor analysis, an established technique that has gained increasing use in paleoclimatology and meteorology for improving the selection of observational sites. Previous studies have performed optimal sensor analyses to evaluate and design networks of weather stations in the Pacific Northwest (Mauger and others, Reference Mauger, Bumbaco, Hakim and Mote2013) and Antarctica (Bumbaco and others, Reference Bumbaco, Hakim, Mauger, Hryniw and Steig2014; Hakim and others, Reference Hakim, Bumbaco, Tardif and Powers2020), as well as coral δ18O proxy records (Comboul and others, Reference Comboul, Emile-Geay, Hakim and Evans2015). An equivalent analysis does not exist for the Antarctic ice core proxy network. We use the ensemble Kalman filter-based optimal sensor methodology implemented by King and others (Reference King, Tierney, Osman, Judd and Anchukaitis2023a), which follows that of Comboul and others (Reference Comboul, Emile-Geay, Hakim and Evans2015). We additionally consider the physical and logistical constraints that are critical for any rigorous evaluation of potential ice core drilling sites (e.g. Vance and others, Reference Vance2016). We emphasize that our methodology is generalizable to any climate metric. The underlying method can be used to optimize ice core site selection for a broad array of research questions, including different climate variables and smaller spatial scales.
We evaluate optimal Antarctic ice core drilling locations to reconstruct annual Antarctic accumulation and annual Antarctic temperature from 1801 to 2000 CE. To do so, we evaluate the statistical covariance and variance relationships from two different climate ensembles, which are constructed with either nine CMIP5 Last Millennium simulations or four instrumental reanalyses. We also account for proxy uncertainty by weighting against low-accumulation locations. An efficient search algorithm (Comboul and others, Reference Comboul, Emile-Geay, Hakim and Evans2015; King and others, Reference King, Tierney, Osman, Judd and Anchukaitis2023a) allows us to incorporate this information into a quantitative optimal ranking of potential ice core sites. We then qualitatively consider physical and logistical factors, such as ice flow and topography, to ground our analysis with real-world constraints.
2. Methods
2.1 Methods overview
We treat ice core site selection as an optimal sensor problem, where we utilize the ensemble Kalman filter-based algorithm implemented by King and others (Reference King, Tierney, Osman, Judd and Anchukaitis2023a), which follows that of Comboul and others (Reference Comboul, Emile-Geay, Hakim and Evans2015). This method iteratively designs an optimal Antarctic ice core proxy network by ranking the ability of potential proxy sites to constrain the variance of a given climate metric sampled across an ensemble of climate estimates.
The optimal sensor algorithm requires three inputs:
1) the climate metric to be reconstructed
2) the proxy estimates representing each potential ice core location
3) the proxy uncertainty for each proxy estimate
The climate metric and proxy estimates are sourced from either an ensemble of CMIP5 simulations or an ensemble of instrumental reanalyses. Metrics are calculated as spatially-integrated averages across either AIS, WAIS or EAIS. The proxy uncertainty is estimated by scaling with the inverse of accumulation rate (more detail in 2.5). The following sections explain these methods in further detail.
We run three different experiments.
1) The ‘blank slate’ experiment designs an optimal ice core network from scratch (i.e., assuming no ice core records exist).
2) The ‘augmented network’ experiment incorporates information from existing ice cores and finds sites to complement the existing network.
3) The ‘redrilling’ experiment finds which existing ice cores would provide the greatest information gain with an extension to the present, drawing only from existing ice core sites.
2.2 Optimal sensor algorithm
The optimal sensor algorithm iteratively uses an ensemble Kalman filter and an ensemble of climate simulations to rank potential proxy data sites.
Given n potential proxy records (potential ice core sites) and m ensemble members, the algorithm requires three inputs:
1) Climate metric: J is the climate metric we wish to reconstruct. For our experiments, it is either annual Antarctic accumulation or average annual Antarctic temperature. J is a scalar metric distributed across m ensemble members (1 × m row vector).
2) Proxy estimates: Ŷ is the synthetic proxy estimates, which are designed to simulate hypothetical accumulation or δ18O records from all grid cells across Antarctica. Ŷ is the proxy estimates for n proxy sites, distributed across m ensemble members (n × m matrix)
3) Proxy uncertainties: R is the uncertainty assigned to each hypothetical ice core record, and scales inversely with accumulation rate. R is the estimated uncertainty in each of n proxy records (n × n diagonal matrix)
A n × 1 vector Δσ is computed according to the following equation:
\begin{equation}{\boldsymbol{\Delta }}{\sigma _{\mathbf{i}}} = {\mathbf{cov}}{\left( {{\mathbf{\unicode{x0176}}_{{\mathbf{i}},}}{\mathbf{J}}} \right)^{\mathbf{2}}}{\left[ {{\mathbf{var}}\left( {{\mathbf{\unicode{x0176}}_{\mathbf{i}}}} \right) + {{\mathbf{R}}_{{\mathbf{ii}}}}} \right]^{ - {\mathbf{1}}}}\end{equation}where the scalar Δσi is the amount of variance in J that can be reduced by the ith proxy record.
The proxy Ŷi with the largest corresponding value of Δσi is selected as the optimal sensor and is assimilated using an ensemble Kalman Filter to update the climate metric J.
First, the Kalman gain matrix K is computed from the covariance across the ensemble:
\begin{equation}\mathbf{K} \text{} = \text{}\mathbf{c} \mathbf{o} \mathbf{v} \left( {\mathbf{\unicode{x0176}}_{\mathbf{i} ,}}{\mathbf{J}_\mathbf{i}} \right){\left[ \mathbf{c} \mathbf{o} \mathbf{v} \left( {\mathbf{\unicode{x0176}}_\mathbf{i}} \right)\text{}+\text{}\mathbf{R} \right]^{-\mathbf{1}}}\end{equation}This is subsequently used to update J, by accounting for the reduction in variance due to the assimilation of proxy Ŷi:
\begin{equation}{\mathbf{J}_{\mathbf{n} \mathbf{e} \mathbf{w}}} = \text{}\mathbf{J} \text{}-\text{}\mathbf{K} \hat{\text{Y}}\end{equation}Proxy Ŷi is then removed from the network of potential proxy sites, with Δσi set equal to 0. The process is then iteratively continued, returning to Eqn (1), until all potential proxy sites are ranked.
Hakim and others (Reference Hakim, Bumbaco, Tardif and Powers2020) refer to the process described above as a ‘blank slate’ test. The blank slate experiment assumes no prior information, as it does not account for the influence of existing ice core records. In contrast, the augmented network experiment begins by applying the Kalman update equations (Eqns 2 and 3) to existing proxy sites to account for the variance constrained by existing ice core records; only after this is the optimal sensor algorithm (Eqns 1–3) applied.
2.3 Climate metric
The optimal sensor analysis is a scalar optimization problem; the scalar in question is the climate metric J, which is distributed across an ensemble. We test four possible values of J, evaluating two different climate variables and two different climate state ensembles.
The climate variables in question are two continental-scale metrics: 1) annual Antarctic accumulation, which can be used to calculate sea level mitigation, and 2) annual Antarctic temperature, which provides insight into the temperature-accumulation relationship. Notably, our metrics for the ‘representativeness’ of an ice core site reflect variability on a continental, rather than a regional scale.
Separate experiments with each metric are performed for each of AIS, WAIS and EAIS. The metrics are calculated as either a spatial integration (for accumulation) or a spatial average (for temperature), across the region in question. The results can be interpreted as finding the locations that best constrain the variability in the climate metric across all of Antarctica, West Antarctica or East Antarctica, respectively. We highlight that we are targeting annual resolution to simulate high-resolution shallow ice cores spanning the past ∼200 years.
We evaluate two different climate ensembles: a Multi-Model Ensemble (‘MME’) of nine different CMIP5 Last Millennium Simulations and a Multi-Reanalysis Ensemble (‘MRE’) of four different atmospheric reanalyses. The inclusion of multiple models with different spatial covariance structures minimizes structural uncertainty, and has been shown to increase climate reconstruction skill (Goosse and others, Reference Goosse, Crespin, de Montety, Mann, Renssen and Timmermann2010; Steiger and others, Reference Steiger, Hakim, Steig, Battisti and Roe2014; Parsons and others, Reference Parsons, Amrhein, Sanchez, Tardif, Brennan and Hakim2021).
The MME is constructed by randomly drawing 100 years from each of nine different CMIP5 Last Millennium Simulations (850–1850 CE), following Eswaran and others (Reference Eswaran, Truax and Fudge2024). All models are regridded to the individual model (CCSM4) with the highest resolution (0.9375° × 1.25°, latitude-longitude), using bilinear interpolation. The nine models included are CCSM4 (Gent and others, Reference Gent2011), iCESM (Brady and others, Reference Brady2019), BCC-CSM1.1 (Wu and others, Reference Wu2014), CSIRO Mk3L (Phipps and others, Reference Phipps, Rotstayn, Gordon, Roberts, Hirst and Budd2012), FGOALS-s2 (Bao and others, Reference Bao2013), GISS-E2-R (Schmidt and others, Reference Schmidt2014), HadCM3 (Collins and others, Reference Collins, Tett and Cooper2001), IPSL-CM5A-LR (Dufresne and others, Reference Dufresne2013) and MPI-ESM (Gutjahr and others, Reference Gutjahr2019).
The MRE is constructed by drawing 40 years (1980–2019 CE) from each of four different atmospheric reanalyses, following a similar procedure to Wang and Xiao (Reference Wang and Xiao2023). This period is chosen because reanalysis products are of limited reliability prior to the satellite era (i.e., before 1979; Bromwich and others, Reference Bromwich, Fogt, Hodges and Walsh2007; Schneider and Fogt, Reference Schneider and Fogt2018; Bromwich and others, Reference Bromwich, Ensign, Wang and Zou2024). The four reanalyses included are ERA5 (Hersbach and others, Reference Hersbach2020), MERRA-2 (Gelaro and others, Reference Gelaro2017), CFSR (Saha and others, Reference Saha2014) and JRA-55 (Kobayashi and others, Reference Kobayashi2015). Reanalyses are regridded to the MME resolution to allow for a consistent comparison.
The MRE and MME offer complementary strengths and weaknesses. The MME is a far larger ensemble and samples a broader range of plausible climate states; nevertheless, individual ensemble members of the MRE are theoretically a better representation of the ‘true’ climate state, as they are constrained by instrumental observations. The CMIP5 simulations that make up the MME have been shown to consistently overestimate Antarctic precipitation (Palerme and others, Reference Palerme, Genthon, Claud, Kay, Wood and L’ecuyer2017); furthermore, their coarse resolution renders them unable to fully capture the orographic spatial variability of temperature and precipitation (Tang and others, Reference Tang, Chenoli, Samah and Hai2018). The reanalyses that make up the MRE are a generally more skillful representation of precipitation and temperature spatial covariance than the MME models (Tang and others, Reference Tang, Chenoli, Samah and Hai2018); they have been found to present spurious climatic trends (Bromwich and others, Reference Bromwich, Nicolas and Monaghan2011), but with a notable improvement since 1979 (Bromwich and others, Reference Bromwich, Ensign, Wang and Zou2024). The structural uncertainty is relatively high in the MRE, with differing levels of agreement with observations (Boisvert and others, Reference Boisvert, Webster, Petty, Markus, Bromwich and Cullather2018); the highest spread and bias occur in the low-accumulation interior of East Antarctica (Palerme and others, Reference Palerme, Genthon, Claud, Kay, Wood and L’ecuyer2017).
2.4 Proxy estimates
The optimal sensor algorithm requires estimates of potential ice core proxy records (or ‘proxy estimates’): Ŷ. We test four possible values of Ŷ, evaluating two different climate variables and two different climate state ensembles.
The climate state ensembles are the same as those used to construct J: namely, the MME and the MRE. In both cases, we evaluate 5794 potential proxy sites in total, corresponding to the gridscale resolution of CCSM4. By evaluating each Antarctic grid cell, we can assess how well individual ice core sites can constrain the variance in J.
The proxy estimates (Ŷ) simulate either ice core accumulation or water isotope records, but are drawn from the climate state ensemble rather than from the actual proxy records. If J is a time series of annual Antarctic accumulation, proxy estimates simulate accumulation by drawing from the annual precipitation for each grid cell from the climate state ensemble. Because the variability between accumulation and precipitation in Antarctica is highly correlated and precipitation is the dominant term in Antarctic surface mass balance (Bromwich, Reference Bromwich1988; Lenaerts and others, Reference Lenaerts, Medley, van den Broeke and Wouters2019), we treat the two as equivalent. If J is instead temperature, proxy estimates simulate δ18O by drawing from the average annual surface air temperature for each grid cell. This approach assumes that temperature can be related to δ18O, as commonly done in proxy system models (e.g. Dee and others, Reference Dee, Steiger, Emile-Geay and Hakim2016). While this simplification enables practical assessment of each site’s potential to constrain J, it also requires the interpretation of results within the context of these assumptions.
Our method does not account for spatial uncertainty in the δ18O-temperature relationship. Analyses of isotope-enabled climate models and ice cores indicate significant regional differences in this relationship, with the relationship being weaker in East Antarctica (Goursaud Oger and others, Reference Goursaud Oger, Sime and Holloway2024). Although nonlinear proxy system models (PSMs) can be used to estimate spatial variability in isotope fractionation (e.g. Markle and Steig, Reference Markle and Steig2022), the nonlinear PSMs also have uncertainties that can erode reconstruction skill (Dee and others, Reference Dee, Steiger, Emile-Geay and Hakim2016; Steiger and others, Reference Steiger, Steig, Dee, Roe and Hakim2017). Paleoclimate data assimilation studies show that nonlinear PSMs may not yield improvements when compared to simpler linear calibrations, especially for ice core δ18O (Dee and others, Reference Dee, Steiger, Emile-Geay and Hakim2016), except in cases where an isotope-enabled model is used (Steiger and others, Reference Steiger, Steig, Dee, Roe and Hakim2017).
Given these challenges, we do not employ nonlinear PSMs to model spatial variability and uncertainty of the δ18O–temperature relationship; instead, we derive δ18O proxy estimates from surface air temperature, assuming that the uncertainty is spatially uniform. While evaluating the influence of spatial variability in our ability to relate δ18O data and temperature on optimal ice core site selection is beyond the scope of this study, it represents an important direction for future investigation. Another avenue for further work is to explore the calibration of accumulation records to temperature, or the inclusion of an isotope-enabled model as a source of proxy estimates (e.g. Dalaiden and others, Reference Dalaiden2020). We emphasize that our study assumes linearity between temperature and δ18O, and our temperature results should be taken with this assumption in mind.
2.5 Proxy uncertainties
Numerous sources of uncertainty impede the ability to extract a meaningful and accurate signal from an ice core. We express the uncertainty with the error matrix R. We scale the uncertainty to the inverse accumulation rate, as the accumulation rate affects both the timescale accuracy and spatial representativeness, elaborated upon below. The diagonal elements (hereafter referred to as the ‘R values’) vary with the inverse of accumulation rate:
\begin{equation}{\mathbf{R}} = \boldsymbol{\alpha} {{\mathbf{P}}^{ - {\mathbf{1}}}}\end{equation}Here, P is an n × n diagonal matrix, with the nth diagonal representing the average (1981–2020) total annual precipitation at the nth evaluated ice core site, drawn from ERA5 reanalysis. The scaling constant α = 104 is empirically derived from data assimilation methods using the error variance in linear calibration to estimate R values (e.g. Hakim and others, Reference Hakim2016; Tardif and others, Reference Tardif2019; Eswaran and others, Reference Eswaran, Truax and Fudge2024).
We now elaborate on our discussion of proxy (ice core) uncertainty. We consider two primary sources of ice core uncertainty: timescale uncertainty and spatial representation uncertainty. We again note that this study targets high-resolution (i.e. annually resolved) ice cores, which renders the impact of diffusion insignificant in all but very low-accumulation records (Cuffey and Steig, Reference Cuffey and Steig1998).
Annually resolved ice core records are dated through identification of seasonal variations in trace chemistry impurities and stable water isotope concentrations. When the absolute accumulation rate is low (e.g. on the East Antarctic Plateau), the accurate identification of annual layers becomes increasingly challenging (e.g. Alley and others, Reference Alley1997). This indicates that ice core uncertainty is dependent on accumulation rate, with low-accumulation sites having high uncertainty and vice versa (Casado and others, Reference Casado, Münch and Laepple2020; Servettaz and others, Reference Servettaz2020). Nevertheless, accumulation rate is not the only factor in annual layer identification. For instance, high marine sulfate deposition at coastal sites can obscure the seasonal cycle and volcanic markers (Dixon and others, Reference Dixon2005; Steig and others, Reference Steig2005; Winstrup and others, Reference Winstrup2019; Hoffmann and others, Reference Hoffmann2022). Another exception is surface melt at high-temperature sites (e.g. the Antarctic Peninsula), which can bias the seasonal cycle and smooth the δ18O record (e.g. Orsi and others, Reference Orsi, Kawamura, Fegyveresi, Headly, Alley and Severinghaus2015; Hoffmann-Abdi and others, Reference Hoffmann-Abdi2023). These site-specific features must be evaluated, yet the general trend of uncertainty decreasing with accumulation is empirically supported (Fig. S2). Thus, we choose that R values vary inversely with accumulation rate (Eqn 4); using our method, low-accumulation proxy sites (e.g. on the East Antarctic Plateau) are penalized, while high-accumulation sites (e.g. in coastal West Antarctica) are rewarded (Fig. 2a). Generally speaking, our method serves to filter out extremely low-accumulation sites (Fig. 2b).
Accumulation-based R values, calculated using Eqn (4). R values are shown in (a). across the Antarctic Ice Sheet for all potential ice-core sites, with the accumulation estimates sourced from ERA5 reanalysis, and in (b) as a function of accumulation rate.
Another source of uncertainty in ice core records is that of spatial representation: is the ice core truly representative of the area for which it is supposed to provide a climate history (e.g. the relevant grid cell in the model or reanalysis)? As ice core length scales are on the order of centimeters while grid cells are on the order of kilometers, no ice core site can be fully representative of its grid cell. We identify three considerations in this question of spatial uncertainty: the ice velocity, orography and wind.
Except at ice divides, ice with increasing depth has been deposited at increasing distance upstream from the drilling site due to ice flow. We create a simple heuristic to assess the spatial representativeness of an ice core due to ice flow by comparing the size of our grid cells with the ice displacement for a representative reconstruction timescale. As we target annually resolved reconstructions from high-resolution records, we set the representative timescale to 200 years (i.e. for a reconstruction from 1800 to 2000 CE). Given a CMIP grid resolution of approximately 60 km and a 200-year reconstruction period, we find that an upper bound on the spatial representativeness of ice core records will be when ice velocity exceeds 300 m a−1. This threshold is only exceeded within ice shelves and terminal portions of ice streams (Rignot and others, Reference Rignot, Mouginot, Scheuchl and Jeong2022); while sections of fast-moving ice certainly exist elsewhere on the continent behind the grounding line, nearby slow-moving regions (e.g. ice domes and ice rises) can still be found and are much better drilling locations. Therefore, we do not consider ice shelves when assessing potential ice core sites, but we do consider ice rises which are surrounded by ice shelves. We note that the 300 m a−1 value is in practice a significant overestimate; the vast majority of existing ice cores have been drilled at sites with a velocity of over an order of magnitude lower. Furthermore, significant climatic variability can occur within a single grid cell due to the local orography.
While orography and wind certainly influence the spatial representation uncertainty, we do not find that it can be quantified in a non-arbitrary fashion; therefore, we do not incorporate a quantitative measure of orography into R. Generally, climatological variability within a grid cell will vary with orographic variability, due to elevation-dependent changes (e.g. the atmospheric lapse rate) and wind-based snow redistribution; thus, as orographic complexity increases, individual ice core sites are less likely to be spatially representative of the surrounding area (Cavitte and others, Reference Cavitte2023). Furthermore, it is more logistically difficult to drill in mountainous and other topographically complex regions (e.g. coastal regions and ice rises within or bounding ice shelves). A comparison of ice core accumulation records at ice rises to ice-penetrating radar data reveals that the spatial representativeness of ice cores tends to vary with orographic complexity and orography-wind interactions (Cavitte and others, Reference Cavitte2023; Wauthy and Dalaiden, Reference Wauthy and Dalaiden2026). However, these factors are challenging to quantify and ice cores have still been drilled at mountainous locations in Antarctica (e.g. Stenni and others, Reference Stenni1999; Goodwin and others, Reference Goodwin, Mosley-Thompson, Wilson, Porter and Sierra-Hernandez2016) and coastal ice rises (e.g. Philippe and others, Reference Philippe2016; Winstrup and others, Reference Winstrup2019; Neff, Reference Neff2020; Hoffmann and others, Reference Hoffmann2022), which show good agreement with ERA5 (Fig. S2a and b). Nevertheless, we qualitatively consider these factors when we present our drilling case studies below.
2.6 Existing ice core records
In order to conduct the ‘augmented network’ and ‘redrilling’ experiments, we first account for the influence of the existing ice core proxy network. We note that the proxy records themselves are not assimilated; instead, we assimilate synthetic ‘proxy estimates’ (Ŷ) derived from climate models or reanalyses at each existing ice core location. To represent a given water isotope or accumulation record, we create a synthetic record from the precipitation or surface air temperature time series for the relevant grid cell; this is distributed across the ensemble. This allows us to analyze the relationship between a given site in the ensemble and the climate metric calculated from the same ensemble.
We assimilate proxy estimates (Ŷ) from 56 sites with accumulation records and 54 sites with water isotope records in total (Fig. 1a). These sites include 54 accumulation sites (Thomas and others, Reference Thomas2017) and 49 water isotope sites (Stenni and others, Reference Stenni2017) from the PAGES 2k databases. We additionally include estimates of accumulation from the B40 site (Medley and others, Reference Medley2018), accumulation and δ18O from the South Pole ice core (Kahle and others, Reference Kahle2021), and δ18O from Derwael Ice Rise (Philippe and others, Reference Philippe2016), PIG2010, DIV2010 and THW2010 (Criscitiello and others, Reference Criscitiello2013; Medley and others, Reference Medley2013).
3. Results
3.1 Blank slate tests
We run the standard optimal sensor algorithm, assuming no prior influence (i.e. no existing ice cores) and with all grid cells within the grounding line (Zwally and others, Reference Zwally, Giovinetto, Beckley and Saba2012) as potential proxy sites. Figure 3 shows the resulting calculated optimal network of 10 ice cores, and the percentage of variance in the climate metric that is theoretically constrained by each potential ice core site. We note that comparing the values of constrained variance between independent experiments can be misleading. The choice of 10 cores approximates how many ice cores could be expected to be drilled over a period of several years.
‘Blank slate’ optimal Antarctic ice-core accumulation (a, b) and δ18O (c, d) networks for reconstructing annual Antarctic accumulation (a, b) and temperature (c, d), respectively. Climate estimates are sourced from the MME (a, c) and MRE (b, d). The optimal 10-site network is shown, with green shading indicating the normalized variance in the climate metric that is constrained across the ensemble by an ice-core record from each possible site. Gray dots indicate existing ice-core locations.
Figure 3 Long description
A) The map titled ′MME Accumulation′ displays the optimal ice core network for reconstructing annual Antarctic accumulation. Green shading indicates the percent variance constrained, ranging from 0 to 100 percent. Purple and blue dots represent proxy ranks from 1 to 10. B) ′MRE Accumulation′ map shows similar data with different proxy site distributions. C) ′MME Temperature′ map illustrates the optimal network for temperature reconstruction, with green shading and proxy ranks marked by colored dots. D) ′MRE Temperature′ map presents the same temperature data using MRE estimates. Each map includes gray dots indicating existing ice core locations. The legend explains the percent variance constrained and proxy rank colors.
The results indicate that the variance in annual Antarctic accumulation can be best constrained by sites around the WAIS Divide and the southern Ross Ice Shelf (Fig. 3a and b), likely due to high covariance in this region with Antarctic accumulation (Figs S1a and b). Consequently, the calculated accumulation optimal networks are extremely clustered. Results from both the MME and MRE calculations strongly favor sites in the Transantarctic Mountains at the southern edge of the Ross Ice Shelf, selecting four and three sites in this region, respectively. Records from these regions show high potential to constrain variability in both WAIS-specific and EAIS-specific accumulation (Fig. S3). Agreement between the MME and MRE is a sign of confidence in the results because they have complementary strengths and weaknesses: the larger size of the MME accounts for a broader range of plausible climate states, but the covariance structures of the MRE are supported by instrumental observations.
Additionally, for the AIS accumulation experiments, the MME selects two sites near the existing South Pole Ice Core (Kahle and others, Reference Kahle2021; Fig. 1a), while the MRE selects three sites slightly north of the proposed Hercules Dome Ice Core (Fudge and others, Reference Fudge2023; Fig. 1b). The optimal sensor algorithm selects ‘clusters’ of multiple sites when these locations have strong covariance with the continental climate metric. As all ice core records have inherent variability and noise, a stronger signal can be extracted from the noise with multiple adjacent records (Comboul and others, Reference Comboul, Emile-Geay, Hakim and Evans2015).
Sites at the southern edge of the Ross Ice Shelf show high potential in both the MME and MRE to constrain annual Antarctic temperature (Fig. 3c and d, Fig. S4). The MME additionally selects a cluster of sites near the Brunt Ice Shelf in DML, a region which has a particularly high potential to constrain temperature variability in EAIS specifically (Fig. S4). The MRE network is much more dispersed, but both the MRE and MME select sites in the Transantarctic Mountains and near Troll Station in DML. Coastal sites along the Amundsen and Bellingshausen Sea (e.g. Neff, Reference Neff2020; Thomas and others, Reference Thomas2024) show high potential in both the MME and MRE to constrain WAIS-averaged temperature variability (Fig. S4).
3.2 Augmented network tests
While the blank slate tests provide insight into which regions of Antarctica best constrain the variance in annual Antarctic temperature and accumulation, a network of records already exists (Fig. 1a). We again run the optimal sensor algorithm with all grid cells treated as potential proxy sites, but now we first account for the variance constrained by the 110 existing ice core records of accumulation or δ18O.
The influence of the existing proxy network is readily apparent with the accumulation experiments. In the blank slate tests, sites across much of WAIS have high potential to constrain the variance in annual Antarctic accumulation (Fig. 3a and b, Fig. S3). Once the influence of existing records is accounted for, this potential dramatically drops (Fig. 4a and b, Fig. S3). This is a positive sign, implying that current ice cores in WAIS are well-placed to constrain the variance in annual Antarctic accumulation. For new accumulation sites, the MME strongly favors a small region at the southern end of the Ross Ice Shelf, in the Transantarctic Mountains. The MRE additionally finds that sites around the Brunt Ice Shelf in western DML have high potential to constrain variance in annual EAIS accumulation (Fig. S5).
‘Augmented network’ optimal Antarctic ice core accumulation (a, b) and δ18O (c, d) networks for reconstructing Antarctic accumulation (a, b) and temperature (c, d), respectively. Climate estimates are sourced from the MME (a, c) and MRE (b, d). The optimal 10-site network is shown, with green shading indicating the normalized variance in the climate metric that is constrained across the ensemble by an ice-core record from each possible site. Gray dots indicate existing ice-core locations.
Figure 4 Long description
A) The map titled ′MME Accumulation′ shows Antarctica with green shading indicating percent variance constrained, ranging from 0 to 100 percent. Purple and blue dots represent proxy ranks from 1 to 10. High variance is concentrated in the southern Ross Ice Shelf region. B) The ′MRE Accumulation′ map displays similar shading and proxy ranks, with notable variance around the Brunt Ice Shelf in western DML. C) The ′MME Temperature′ map shows variance constrained across Antarctica, with proxy ranks distributed throughout. D) The ′MRE Temperature′ map presents a detailed variance pattern, with proxy ranks along the coastlines. Each map uses a legend to indicate percent variance constrained and proxy rank, with green for variance and a gradient from purple to blue for proxy rank.
For the temperature experiments, results from both the MME and MRE calculations select a cluster of sites at the southern end of the Ross Ice Shelf. The MME selects a second cluster further north in the Transantarctic Mountains, in the vicinity of McMurdo Station. It selects a third cluster on the edge of the Brunt Ice Shelf near Halley Research Station, which is particularly effective in constraining annual EAIS temperature (Fig. S6). Meanwhile, an MRE cluster is selected on ice rises in the Ronne Ice Shelf, near the existing Berkner Island ice core (Mulvaney and others, Reference Mulvaney2002; Fig. 1a), which also has good potential to constrain annual EAIS temperature. According to the MRE calculations, annual WAIS temperature can be well constrained by a cluster of sites near the Ellsworth mountains (Fig. S6). Finally, for all experiments except the MME temperature test, various miscellaneous sites are selected at coastal locations in East Antarctica, particularly near the existing Mawson and Dumont d’Urville stations (Fig. 1c); these primarily provide insight into annual EAIS temperature (Fig. S6).
3.3 Redrilling tests
The augmented network tests provide insight into where to drill a new ice core. However, they are predicated on the assumption that a continuous chronology exists for the entire existing ice core network throughout the period of interest. In reality, most existing cores were drilled more than two decades ago through efforts such as the International Trans-Antarctic Scientific Expedition (ITASE; Mayewski and others, Reference Mayewski2005; Steig and others, Reference Steig2005; Fig. 1b) and the European Project for Ice Coring in Antarctica (Stauffer and others, Reference Stauffer, Flückiger, Wolff and Barnes2004). Indeed, of the 110 ice core records we evaluate, only 48 have a chronology that extends past 2000 CE and only 14 extend past 2010 CE; this is why many Antarctic climate reconstructions choose to terminate in 2000 (e.g. Medley and Thomas, Reference Medley and Thomas2019; Dalaiden and others, Reference Dalaiden, Goosse, Rezsöhazy and Thomas2021). Additionally, these chronologies begin at different times; the 110 records we evaluate only have a complete overlap between 1962 and 1983 CE, which is mostly prior to the satellite era.
Thus, the effective variance constrained by the existing proxy network is a function of time, as some years will contain more information than others. We evaluate the variance constrained by the ice core network that exists for each individual year from 1800 to 2010 CE. Both climate ensembles and variables show a consistent pattern: starting from 1800 CE, our ability to reconstruct Antarctic climate steadily increases until the mid-to-late 20th-century, and then rapidly plummets starting in the mid-1990s (Fig. 5). These results indicate that our ability to reconstruct 21st-century Antarctic climate, using ice cores, is less than 10% that of the mid-20th century. This motivates the choice of our study, and others (e.g. Medley and Thomas, Reference Medley and Thomas2019), to terminate the period of reconstructive interest in 2000 CE. Thus, new ice core records for the 21st-century are strongly desired, in order to extend the presently limited overlap period with satellite-era reanalysis and contextualize modern changes.
Percentage of the maximum possible normalized variance constrained by the existing Antarctic ice-core proxy network from 1800 to 2010 CE, evaluated for two climate variables (accumulation and temperature), two climate state ensembles (MME and MRE), and three regions: AIS, WAIS and EAIS. The maximum variance is constrained from 1962 to 1983; during this period, the chronologies of all evaluated ice core records overlap.
Figure 5 Long description
Three multi line graphs titled, AIS, WAIS and EAIS. Common elements across all three graphs. Horizontal axis label, Year. Horizontal axis range, 1800 to 2010. Vertical axis label, Normalized variance. Vertical axis range, 0 to 1. Legend entries, MME Accumulation, MRE Accumulation, MME Temperature, MRE Temperature. The four series are distinguished by four different line appearances. AIS. The four lines rise in steps from the early years toward values near 1 by the late 1900s. A steep drop occurs near the end of the time range, with all four lines falling from near 1 to values near 0 by 2010. WAIS. The four lines rise from mid range values in the early years toward values near 1 by the late 1900s. A steep drop occurs near the end of the time range, with all four lines falling from near 1 to values near 0 by 2010. EAIS. The four lines rise from mid range values in the early years toward values near 1 by the late 1900s. A steep drop occurs near the end of the time range, with all four lines falling from near 1 to values near 0 by 2010.
Selecting and drilling at novel ice core sites is one method of gaining information about recent Antarctic climate variability, but another option is to redrill at existing ice core locations, thereby making their chronologies up to date. Our ‘redrilling’ tests address this latter question: which existing ice cores would be best to redrill and bring up to date? Fig. 6 shows the potential variance in annual Antarctic accumulation and temperature constrained by accumulation and δ18O proxy estimates from all existing ice core sites.
Normalized percent of variance in annual Antarctic accumulation (a, b) and annual Antarctic temperature (c, d) constrained by accumulation (a, b) and δ18O (c, d) at existing ice-core sites. Proxy estimates (Ŷ) are sourced from the MME (a, c) and MRE (b, d).
Figure 6 Long description
Four polar maps of Antarctica display percent variance constrained by accumulation and temperature. Each map is labeled: a) MME Accumulation, b) MRE Accumulation, c) MME Temperature, d) MRE Temperature. The maps use circular markers to indicate variance levels at different locations across Antarctica. Darker markers represent higher percent variance constrained, while lighter markers indicate lower variance. The color scale ranges from 0 to 100 percent. In maps a and b, accumulation variance is concentrated in West Antarctica and the Antarctic Peninsula, with darker markers showing higher variance. Maps c and d depict temperature variance, with notable concentrations in East Antarctica and coastal regions. Key values approach 80 to 100 percent in darker areas, while lighter areas are closer to 0 to 20 percent. The maps illustrate differences between MME and MRE models, highlighting spatial patterns in climate data across Antarctica.
The accumulation tests (Fig. 6a and b) contain a significant regional disparity. Results from both the MME and MRE experiments indicate that WAIS sites are by far the most optimal to redrill. This is likely because this region has the highest local covariance with continental-scale precipitation (Fig. S1). We note that many of the ITASE sites in WAIS (Fig. 1a) are chosen by both ensembles as being highly optimal to redrill. From now on, we use ‘ITASE’ as a shorthand for the West Antarctic records obtained through the U.S. component of the International Transantarctic Scientific Expedition.
By contrast, the δ18O tests (Fig. 6c and d) offer a less immediately obvious conclusion, with the MME and MRE experiments showing divergent results across a variety of regions. The MRE results suggest that redrilling one of the numerous ice cores in interior DML (Oerter and others, Reference Oerter, Graf, Wilhelms, Minikin and Miller1999), as done by Winton and others (Reference Winton, Mulvaney, Savarino, Clem and Frey2024), would be advantageous. Other sites include the Mount Erebus Saddle core (Rhodes and others, Reference Rhodes2012) and the Bruce Plateau core (Goodwin and others, Reference Goodwin, Mosley-Thompson, Wilson, Porter and Sierra-Hernandez2016) on the Antarctic Peninsula. The MME instead favors sites in coastal, rather than interior, DML, such as IND-22/B4 (Thamban and others, Reference Thamban, Laluraj, Naik and Chaturvedi2011); it is also more favorable to sites in WAIS than the MRE. This divergence can likely be attributed to the fact that the spatial temperature covariance patterns between the MME and MRE are quite dissimilar (Fig. S1c and d), indicating the challenges in accurately constraining temperature over the AIS.
4. Discussion
Our results identify potential locations for new ice core drilling that are best for the purposes of reconstructing changes in Antarctic-wide annual temperature and accumulation. The results do not evaluate site-specific limitations such as ice velocity and topography, or local influences on accumulation such as precipitation intermittency, seasonality and wind redistribution. We describe the following three case studies to illustrate the importance of additional site-specific understanding, and do not explicitly propose them as future ice coring sites. Instead, our goal is to illustrate that despite the relatively large spatial scale of our analysis, there are potential drilling sites that could be targets. In all cases, proposals to drill specific locations will benefit from in situ data collected at the site, including, but not limited to, conducting ice-penetrating radar surveys to understand the glaciological setting and drilling short firn cores covering the satellite era to establish the representativeness of preserved climate records. We conclude the discussion by addressing limitations of the current data and methodology and suggesting avenues for future work.
4.1 Case studies
Here, we present three case studies of potential new shallow ice coring projects: 1) drilling a new core at Crary Ice Rise (CIR); 2) drilling a new core at Lyddan Ice Rise; and 3) redrilling the US-ITASE cores collected in the early 2000s in WAIS (Table 1). These case studies were chosen to combine both climatic reconstructive ability (Fig. 7) and logistical feasibility. We emphasize that our study identifies ice core sites to constrain our two chosen climate metrics: annual accumulation and annual temperature. As noted above, our methodology does not quantify all the relevant complexities of ice core site selection, and our case studies should not be taken as a formal proposal to drill at these locations. Nevertheless, these case studies represent promising targets for future drilling and illustrate how our data-driven optimization approach can be adapted to account for real-world site selection and logistical constraints.
‘Optimality scores’ for each potential Antarctic ice-core site, calculated by taking the root mean square of the normalized percent constrained variance for each grid cell across (a) all four ‘blank slate’ experiments (‘Total’), (b) the two accumulation experiments, and (c) the two temperature experiments (Figure 3). The sites of three drilling case studies: Crary Ice Rise, Lyddan Ice Rise and ITASE core redrilling at ITASE-99-01, ITASE-01-02 and WAIS Divide are displayed.
Figure 7 Long description
Three maps of Antarctica display optimality scores for potential ice core sites. A) The map labeled ′Total′ shows overall scores with a gradient from 0 to 1. B) The ′Accumulation′ map highlights scores based on accumulation experiments. C) The ′Temperature′ map focuses on temperature experiments. Each map includes markers for specific sites: Crary (yellow circle), Lyddan (orange circle), ITASE-99-01 (red triangle), ITASE-01-02 (purple triangle) and WAIS Divide (pink triangle). The reconstruction score is represented by a gradient bar ranging from 0 to 1, indicating varying levels of suitability across the continent.
Physical characteristics of five potential ice-core sites: Crary Ice Rise, Lyddan Ice Rise, ITASE-99-01, ITASE-01-02 and WAIS Divide. We include topographic information from Bedmachine (Morlighem and others, Reference Morlighem2020), ice velocity (MacAyeal and others, Reference MacAyeal, Bindschadler, Shabtaie, Stephenson and Bentley1987; Kaspari and others, Reference Kaspari2004; King and others, Reference King, Anderson, Vaughan, Mann, Mobbs and Vosper2004; Koutnik and others, Reference Koutnik2016) and accumulation rate from surface observations (Bindschadler and others, Reference Bindschadler, Roberts and Iken1990; Kaspari and others, Reference Kaspari2004; King and others, Reference King, Anderson, Vaughan, Mann, Mobbs and Vosper2004; Orsi and others, Reference Orsi, Cornuelle and Severinghaus2012).
Table 1 Long description
Columns list ice core site, latitude, longitude, region, surface elevation in meters above sea level, bed elevation in meters above sea level, ice thickness in meters, ice velocity in meters per year, and accumulation rate in meters per year ice equivalent. Crary Ice Rise at latitude minus 83.05 degrees and longitude minus 172.50 degrees in the Ross Ice Shelf has surface elevation 130, bed elevation minus 400, thickness 530, ice velocity negligible, and accumulation rate 0.11. Lyddan Ice Rise at latitude minus 74.05 degrees and longitude minus 21.00 degrees in the Brunt and Riiser Larsen Ice Shelves has surface elevation 260, bed elevation 10, thickness 240, ice velocity negligible, and accumulation rate 0.30. ITASE 99 01 at latitude minus 80.62 degrees and longitude minus 122.63 degrees in the West Antarctic Ice Sheet has surface elevation 1300, bed elevation minus 1110, thickness 2410, ice velocity 20, and accumulation rate 0.15. ITASE 01 02 at latitude minus 82.00 degrees and longitude minus 110.01 degrees in the West Antarctic Ice Sheet has surface elevation 1760, bed elevation minus 180, thickness 1940, ice velocity 10, and accumulation rate 0.19. WAIS Divide at latitude minus 79.46 degrees and longitude minus 112.09 degrees in the West Antarctic Ice Sheet has surface elevation 1780, bed elevation minus 1680, thickness 3460, ice velocity 3, and accumulation rate 0.22.
4.2 Crary Ice Rise
Locations at the southern edge of the Ross Ice Shelf and within the Transantarctic Range are consistently identified as good drilling sites (Figs 3a–d, 4a–d, Figs S3–4). Climate records from this region can constrain both annual Antarctic accumulation and temperature when averaged across each of AIS, WAIS and EAIS, yet no ice cores have been drilled in this region to date (Figure 1a and 8a), excepting a single 8-year firn core (Markle and others, Reference Markle, Bertler, Sinclair and Sneed2012). The Ross Ice Shelf and Transantarctic Range regions both present challenges for obtaining ice core records. The Ross Ice Shelf and associated ice streams have velocities of over 100 m a−1 (Fig. 8b), compared to velocities of under 10 m a−1 at most existing ice core sites on the AIS (Rignot and others, Reference Rignot, Mouginot, Scheuchl and Jeong2022); ice core sites on ice shelves thus may not accurately reflect the site history. Drilling in the Transantarctic Mountains generates logistical difficulties as well as raising questions about spatial representativeness, where high orographic variability (Figure 8c) combined with wind-borne snow transport leads to high climatic variance within a small geographic area (King and others, Reference King, Anderson, Vaughan, Mann, Mobbs and Vosper2004; Cavitte and others, Reference Cavitte2023). Altogether, this begs the question: is there a viable path between the Scylla of rapid ice flow and the Charybdis of mountainous topography?
(a) ‘Optimality scores’ (an inset of Figure 7a); (b) ice velocity (Rignot and others, Reference Rignot, Mouginot, Scheuchl and Jeong2022); and (c) standard deviation of ERA5 grid-cell orography (Hersbach and others, Reference Hersbach2020), shown for the Ross Ice Shelf region. Crary Ice Rise is indicated with a star; also displayed are existing ice cores, national bases and field camps and the South Pole Overland Traverse route between McMurdo and Amundsen-Scott stations.
Figure 8 Long description
A) The map shows reconstruction scores for the Ross Ice Shelf region, ranging from 0 to 1. Key locations include Amundsen-Scott Base, McMurdo/Scott Base, Camp 20 and Crary Ice Rise marked with a star. National bases are indicated by squares, field camps by triangles and existing ice cores by circles. The South Pole Overland Traverse route is marked with a red dashed line. B) The map displays ice velocity in meters per year, with a gradient indicating varying speeds. Crary Ice Rise is marked with a star, showing higher velocities near the ice streams. C) The map illustrates the standard deviation of ERA5 grid-cell orography, with a gradient indicating variability. Crary Ice Rise is marked with a star, showing higher variability along the Transantarctic Range.
CIR is a small ice rise close to the southern edge of the Ross Ice Shelf, which formed ∼1100 years ago when the Ice Shelf re-grounded on marine sediment (Catania and others, Reference Catania, Hulbe, Conway, Scambos and Raymond2012). At a pinning point of the Ice Shelf, ice flow at CIR reduces to near-zero and the accumulation rate is about 0.11 m a−1 ice equivalent (Table 1). Ice rises provide a solution to the problem of drilling on ice shelves, as they show far lower rates of ice flow (e.g. Matsuoka and others, Reference Matsuoka2015); this is why numerous ice cores have been drilled on Antarctic ice rises, such as Derwael Ice Rise (Philippe and others, Reference Philippe2016), Siple Dome (Taylor and others, Reference Taylor2004), Roosevelt Island (Winstrup and others, Reference Winstrup2019), Skytrain Ice Rise (Hoffmann and others, Reference Hoffmann2022) and the Amundsen Sea Coast (Neff, Reference Neff2020). We note that there are challenges associated with drilling on coastal ice rises, but these can be mitigated. High marine sulfate deposition can impede accurate dating, but alternate dating mechanisms such as methane matching can be utilized (Winstrup and others, Reference Winstrup2019) and CIR is far from the coast. Ice rises show high local accumulation variability due to blowing snow and wind-orography interactions (King and others, Reference King, Anderson, Vaughan, Mann, Mobbs and Vosper2004; Cavitte and others, Reference Cavitte2023), indicating that care should be taken to choose a site that is spatially representative of the surrounding area. Any drilling proposal at CIR should additionally account for the impact of extreme precipitation events (Turner and others, Reference Turner2019). The drilling and analysis of a shallow core (∼20 years) would provide an effective way to assess the spatial representativeness of the ice core site.
No ice core has been drilled yet at CIR; however, it has been an active site for glaciological fieldwork, including a recent study of the evolution of the ice rise using radar transects (Hillebrand and others, Reference Hillebrand, Conway, Koutnik, Martín, Paden and Winberry2021). This existing ice-penetrating radar data will significantly aid site selection efforts. Hillebrand and others (Reference Hillebrand, Conway, Koutnik, Martín, Paden and Winberry2021) find discontinuous layers at 250–300 m above the bed, which are composed of former Ross Ice Shelf marine ice. However, the uppermost ∼200 m contains well-preserved and spatially continuous internal layering from local meteoric accumulation that is more than thick enough to cover the past 200 years and likely considerably more. Coring to recover both this meteoric snow accumulation as well as access former Ross Ice Shelf marine ice would yield valuable insight not only into Ross Sea regional climate but also the former sub-ice shelf marine environment.
CIR is additionally an active site of interest for recent mass change, motivated by the effects of ice stream deceleration on the embayment buttressing regime of the Ross Ice Shelf (Verboncoeur and others, Reference Verboncoeur2025). Previous work at CIR shows that logistical support is feasible; it is located just off the South Pole Overland Traverse (SPOT) between McMurdo and Amundsen-Scott stations and close to the existing Camp 20 field camp (Fig. 8a). As part of the SWAIS 2C project (Patterson and others, Reference Patterson2022), an international team will drill a sediment core at CIR, with logistical support being provided by SPOT.
4.3 Lyddan Ice Rise
Our analysis suggests that the best location for reconstructing annual Antarctic temperature, particularly in EAIS, is coastal western DML in the vicinity of the Brunt and Riiser-Larsen ice shelves (Figure 7c and 9, Fig. S4). Our results also indicate that coastal western DML is a good location to complement existing ice cores to improve reconstructions of 19th–20th century annual EAIS accumulation (Figure S5) by drilling at high accumulation coastal sites in EAIS (e.g. Jackson and others, Reference Jackson, Vance, Crockart, Moy, Plummer and Abram2023; Vance and others, Reference Vance2024).
(a) Temperature ‘optimality scores’ (an inset of Figure 7c), and (b) ice velocity (Rignot and others, Reference Rignot, Mouginot, Scheuchl and Jeong2022) shown for the western Dronning Maud Land region. Lyddan Ice Rise is indicated with a star; existing ice cores and national bases are also displayed.
Figure 9 Long description
A) The map shows reconstruction scores for western Dronning Maud Land, with scores ranging from 0 to 1. Lyddan Ice Rise is marked with a star. National bases are indicated by red squares and existing ice cores by blue circles. Higher scores are concentrated in coastal areas. B) The map displays ice velocity in meters per year, ranging from 0 to 100. The Brunt Ice Shelf and Riiser-Larsen Ice Shelf are labeled. Velocity is highest near the coast and decreases inland.
Numerous ice cores have been drilled in the interior DML (e.g. Oerter and others, Reference Oerter, Graf, Wilhelms, Minikin and Miller1999; Medley and others, Reference Medley2018; Winton and others, Reference Winton, Mulvaney, Savarino, Clem and Frey2024) yet are primarily outside the coastal regions we identify as highly correlated with annual Antarctic-wide temperature. The Derwael, Lokeryggen and Hammarryggen Ice Rise records (Philippe and others, Reference Philippe2016; Wauthy and others, Reference Wauthy2024; Wauthy and Dalaiden, Reference Wauthy and Dalaiden2026) are the only ice cores drilled at coastal ice rises in this region, but they are substantially farther east than our optimal sites. Goel and others (Reference Goel, Matsuoka, Berger, Lee, Dall and Forsberg2020) evaluated the stability of ice rises in DML, which are promising ice core targets owing to their low ice velocity (Fig. 9b). Lyddan Ice Rise, located between the Brunt Ice Shelf and Riiser-Larsen Ice Shelf, stands out as the site which has the potential to constrain the most variance in annual Antarctic-wide temperature (Fig. 9a). Additionally, the ‘double lineations’, or double-peaked Raymond arches (Raymond, Reference Raymond1983), observed on Lyddan are an indicator of long-term stability (Goel and others, Reference Goel, Matsuoka, Berger, Lee, Dall and Forsberg2020). The Raymond arch suggests stability for the past ∼3 ka and suggests deeper ice coring could also be of interest for longer temporal records. Furthermore, the regional accumulation signal from western DML is also of significant scientific interest, as increased snowfall in this region has been driving EAIS mass gain (Medley and others, Reference Medley2018; Zuhr and others, Reference Zuhr2026).
From a logistical perspective, Lyddan is also a feasible site. The accumulation rate is about 0.30 m a−1 ice equivalent (Table 1). Ice-penetrating radar data across Lyddan shows an accumulation gradient, with accumulation decreasing in the predominantly downwind (westward) direction (King and others, Reference King, Anderson, Vaughan, Mann, Mobbs and Vosper2004), indicating snow redistribution by wind. Annual accumulation variations may be more difficult to distinguish from spatial variability because of the gradient, although temperature reconstructions should be less affected than accumulation. The existing radar profiles can be studied in detail for the purposes of site selection. Additionally, any proposal to drill at Lyddan should carefully analyze the potential for surface melting, which has been observed at nearby ice rises (Dey and others, Reference Dey2023), as well as the impact of extreme precipitation events (Turner and others, Reference Turner2019; Wauthy and Dalaiden, Reference Wauthy and Dalaiden2026). The British, Norwegian, Indian and Russian Antarctic programs are showing active interest in coastal DML ice drilling (e.g. Goel and others, Reference Goel2025), and thus may be the best options for logistical support.
4.4 ITASE redrilling
The ITASE (Mayewski and others, Reference Mayewski2005) collected a network of shallow (∼50–300 m) ice cores from locations throughout Antarctica. The US-led portion collected many cores in the interior WAIS from 1999 to 2002 CE (Kaspari and others, Reference Kaspari2004; Dixon and others, Reference Dixon2005; Steig and others, Reference Steig2005, Reference Steig2013; Banta and others, Reference Banta, McConnell, Frey, Bales and Taylor2008; Schneider and Steig, Reference Schneider and Steig2008). These cores, and others in WAIS, have been highly effective at constraining 20th-century variability in annual West Antarctic accumulation (Fig. 3a and b, Fig. 4a and b). However, most of these records—with certain exceptions (e.g. Morris and others, Reference Morris2017)—were drilled over two decades ago, which renders them unable to capture recent changes.
Initial efforts to drill shallow ice cores from coastal ice domes along the Amundsen Sea coast have been successful (Neff, Reference Neff2020 and personal communication) and have the potential to improve projections of modern sea level change. However, the full potential of these new coastal records can only be realized if they are compared to records from inland WAIS, which are absent for the 21st-century. Updating the ITASE network would provide a continuous, high-resolution climate record that bridges the gap between past ice core data and ongoing satellite observations. This extended overlap would improve calibration of remote sensing datasets, enhance detection of recent shifts in temperature and accumulation patterns, and refine projections of future ice sheet behavior under continued climate forcing.
In our redrilling tests, US-ITASE records in inland WAIS are consistently identified as optimal sites for constraining annual Antarctic-wide accumulation (Fig. 6a and b). In both climate models and reanalyses, accumulation from locations across western WAIS has high interannual covariance with Antarctic-wide accumulation (Fig. S1). The WAIS-EAIS, while geographic, is not necessarily climatological. Previous studies have found a climatological transition at approximately the East Antarctic Ice divide (∼78° S–80° S) between predominantly Pacific-sourced and Indian-sourced air masses, where sites grid-west of the divide (e.g. South Pole, Kohnen) have a mean moisture source longitude of ∼60° W to 180° W, associated with transport over WAIS and the Transantarctics (Sodemann and Stohl, Reference Sodemann and Stohl2009; Petteni and others, Reference Petteni2026). Thus, high-quality records from WAIS have the ability to provide information about accumulation variability in parts of EAIS.
In addition, sites in WAIS have a higher absolute accumulation rate, which we demonstrate generally correlates with higher annually resolved record quality due to lower timescale uncertainty in annual layer counting (Fig. 2). West Antarctic ice core records have also been found to show strong linkages to large-scale circulation signals (Küttel and others, Reference Küttel, Steig, Ding, Monaghan and Battisti2012; Man and others, Reference Man2025), rendering them less susceptible to being dominated by stratigraphic noise and localized processes (e.g. wind-borne redistribution of snow). We further note that our results are reflective of our focus on interannual variability (more discussion in Section 4.5).
We seek to identify the most optimal drilling sites in WAIS. Similar to Fig. 7, we calculate an ‘optimality score’ for each US-ITASE site by adding in quadrature the normalized percent constrained variance from both the MME and MRE simulations (Fig. 10a). Two particular sites are identified as having the highest optimality scores: these are ITASE-1999-1 (‘99-1’) and ITASE-2001-2 (‘01-2’). 99-1 is located in the vicinity of the Byrd camp, while 01-2 is about halfway between WAIS Divide and Hercules Dome.
(a) Normalized percent of variance in annual Antarctic accumulation constrained by proxy estimates at ITASE coring sites sourced from the MME and MRE, combined in quadrature. (b) The percent variance constrained in annual Antarctic accumulation as a function of the number of hypothetically redrilled ITASE cores.
Figure 10 Long description
Multiple circular markers and text labels. A horizontal scale bar labeled 0, 0.2, 0.4, 0.6, 0.8, 1.0. Text below the scale bar reads, Normalized Percent Constrained Variance. Visible marker labels include 01-05, 01-04, 01-03, 01-01, 01-02, 02-04, 00-03, 00-04, 00-05, WAIS Divide and 99-01 Byrd. The image B showing a line graph. The horizontal axis label reads, Number of Redrilled ITASE Cores, with tick labels 0, 2, 4, 6, 8, 10, 12. The vertical axis label reads, Percent Variance Constrained, with tick labels 0, 10, 20, 30, 40, 50. The plotted line starts at 0 cores and 0 percent, rises to about 45 percent near 1 core, rises to about 50 percent near 2 cores and stays at about 50 percent from 2 to 12 cores.
Our analysis indicates that the information gained about annual Antarctic accumulation has diminishing returns after the first three records in West Antarctica are drilled (Fig. 10c). The 02-4 record (Jacobel and others, Reference Jacobel, Welch, Steig and Schneider2005) will be effectively updated with the uppermost part of the upcoming Hercules Dome ice core drilling (Fudge and others, Reference Fudge2023; Hills and others, Reference Hills2025). Drilling three ∼10–15-meter (∼20–30 years) firn cores to update WAIS Divide, 99-1, and 01-2 would maximize the scientific return for the logistical effort. The drilling could be accomplished with day-long ground traverses or fixed-wing trips out of WAIS Divide. Both 99-1 and 01-2 are located approximately 250 km from WAIS Divide, with 99-1 being located close to the site of the former Byrd Camp, which has been proposed to (re-)replace WAIS Divide as the major West Antarctic logistical base for the United States Antarctic Program.
4.5 Outlook and future work
Our present ability to constrain the climate history of Antarctica is hampered by spatial and temporal limitations in the existing ice core network. The relative dearth of records in EAIS (Fig. 1a) has led to high uncertainties around historical Antarctic sea level contribution (Medley and Thomas, Reference Medley and Thomas2019; Eswaran and others, Reference Eswaran, Truax and Fudge2024); the slowed pace of ice core drilling since the 2000s means that our ability to reconstruct 21st-century annual temperature and accumulation is less than 10% that of the mid-20th century (Fig. 5).
Our data-driven optimal sensor methodology, already a well-established technique to identify paleoclimate observation sites (e.g. Comboul and others, Reference Comboul, Emile-Geay, Hakim and Evans2015; Hakim and others, Reference Hakim, Bumbaco, Tardif and Powers2020), offers an objective way to evaluate future ice core sites. The three case studies above are intended as a demonstration of how our optimal sensor methodology can be combined with an evaluation of surface conditions and logistics to produce community-relevant drilling recommendations. Our study should not be taken as a formal recommendation to drill at these specific sites; rather, we use these case studies as illustrations that specific sites can be found within regions identified by the optimal sensor method. Here, we discuss the limitations in our specific implementation of the optimal sensor methodology, and potential avenues for future investigation. We emphasize that these limitations are not intrinsic to the optimal sensor approach as a concept, and can be mitigated by future work.
First, our temperature optimizations rely on the assumption that temperature scales linearly with δ18O. In reality, the relationship can be nonlinear, with dependencies on condensation and evaporation site temperatures and the deuterium excess parameter (Markle and Steig, Reference Markle and Steig2022). In Antarctica, significant spatial and temporal variability has been observed in this relationship (Klein and others, Reference Klein2019; Goursaud Oger and others, Reference Goursaud Oger, Sime and Holloway2024). Thus, our proxy estimates are not perfect representations of δ18O records, which lowers confidence in our identified optimal temperature sites. Nevertheless, we note that existing data assimilation-based polar climate reconstructions have predominantly used linear temperature-δ18O calibration (e.g. Badgeley and others, Reference Badgeley, Steig, Hakim and Fudge2020; O’Connor and others, Reference O’Connor, Steig and Hakim2021); furthermore, investigating the spatial correlations of reanalysis temperature fields is a well-established method of selecting ice core sites for water isotope records (e.g. Vance and others, Reference Vance2016; Bertler and others, Reference Bertler2018). Future studies can use an isotope-enabled climate model (e.g. iCESM; Brady and others, Reference Brady2019) or atmospheric reanalysis (e.g. WisoMIP; Bong and others, Reference Bong2026) as the source ensemble of the proxy estimates; this would allow for direct assimilation of δ18O records, which holds promise for skillful reconstruction of Antarctic temperature (Dalaiden and others, Reference Dalaiden2020). The climate ensemble can alternately be taken from high-resolution polar climate products (e.g. RACMO2; van Wessem and others, Reference van Wessem2018). These regional climate products skillfully capture small-scale climate variations and surface processes, and would allow for a more targeted and localized selection of ice core sites.
Furthermore, our methodology relies on the assumption of stationarity; i.e. that the spatial covariance structure of the MME and MRE holds for the target reconstruction period. This assumption is not unique to our study, as it applies to any spatially complete Antarctic climate reconstruction using a model or reanalysis prior (e.g. Medley and Thomas, Reference Medley and Thomas2019). The MRE is particularly sensitive to shifts in atmospheric circulation induced by anthropogenic climate change; due to its short and modern timespan (1980–2019 CE), it may not effectively capture atmospheric circulation variability on decadal timescales. Recent reconstructions have indicated that dynamic, rather than thermodynamic, change is the primary driver of accumulation variability, which underscores the importance of atmospheric circulation (Medley and Thomas, Reference Medley and Thomas2019). Nevertheless, the limitations of the MRE are counterbalanced by the MME, which is drawn from pre-anthropogenic (850–1850 CE) Last Millennium Simulations.
A separate direction for future investigation can be a more detailed examination of the spatial correlation length scale. Our results frequently find ‘clusters’ of adjacent optimal sites (Figs 3 and 4). It is true that a climate signal from a given region is certainly more robust given multiple records, as there always exists a timescale uncertainty in each record (quantified by the R value). Nevertheless, our methodology does not account for the fact that climate variability at adjacent sites will be dominated by the same synoptic-scale events (e.g. Vance and others, Reference Vance2016), nor the spatial correlation in timescale uncertainty between adjacent records. Future work can adapt the optimal sensor algorithm to penalize the selection of closely adjacent sites, e.g. by adding a spatial correlation length scale cutoff or a different metric of distance.
Future studies can seek to better quantify the uncertainty in ice core records. Our accumulation rate-based estimate is a reasonable first-order approximation for the timescale uncertainty, but it does not objectively account for the numerous sources of spatial uncertainty (e.g. blowing snow, ice velocity and orographic variability) nor the impact of precipitation intermittency. Even at high-accumulation ice rise sites, these variables can create significant uncertainty in the record (e.g. Münch and Laepple, Reference Münch and Laepple2018). Incorporating these variables in an objective and quantitative fashion would lead to a stronger estimate of proxy uncertainty. Furthermore, hand-augured shallow cores offer a means of assessing the spatial representativeness of a given site.
In our study, we evaluated only the two climate metrics of annual continent-wide accumulation and temperature; future work can investigate different metrics for pertinent research questions. Other climate variables can be chosen, such as sea ice extent or atmospheric circulation indices including the Amundsen Sea Low (e.g. Raphael and others, Reference Raphael2016; O’Connor and others, Reference O’Connor2025) and Southern Annual Mode (e.g. Marshall, Reference Marshall2003; O’Connor, Steig, Hakim, Reference O’Connor, Steig and Hakim2021; King and others, Reference King, Anchukaitis, Allen, Vance and Hessl2023b; Dalaiden and others, Reference Dalaiden, Goosse, Holland and Barthelemy2025). Alternatively, the optimal sensor algorithm can be adapted to optimize a vector, rather than a scalar metric (Hakim and others, Reference Hakim, Bumbaco, Tardif and Powers2020). This allows the algorithm to constrain spatial variability, rather than purely temporal variability, in the metric of interest. While our study primarily examines the variance constrained as a normalized value (Figs 3 and 4), the raw variance values can provide insight into reconstruction skill over time (Fig. 5) and into the optimal number of cores to drill (Fig. 10c).
Additionally, the climate metric can be optimized for sub-regions of Antarctica (e.g. the Ross Sea region), rather than for the entire AIS, WAIS or EAIS (e.g. Figs S3–S6). Our results show that the disaggregation of climate metrics into sub-regional estimates can lead to differing results, in addition to adding greater nuance to existing results. For example, the CIR region can constrain annual accumulation variability in both WAIS and EAIS (Fig. S3); the Lyddan Ice Rise region offers little insight into WAIS temperature, but is extremely well-placed to constrain EAIS temperature (Fig. S4). Regional disaggregation can identify new drilling locations that are not optimal for reconstructing a continental average; for example, coastal sites along the Amundsen and Bellingshausen Seas are not particularly optimal for the purposes of reconstructing annual AIS temperature, but are very well-placed to reconstruct annual WAIS temperature (Fig. S4). Performing additional experiments where the climate metric is disaggregated into smaller spatial regions, such as individual drainage basins, is a promising direction for future work.
While this study focuses on a specific temporal resolution, i.e. annually-resolved ice cores to reconstruct interannual variability for the past ∼200 years, the presented methodology is equally applicable to other timescales of variability (e.g. decadal or centennial). We focus on interannual variability because it is well characterized in modern climatology; however, targeting a longer timescale could better utilize East Antarctic records, which do not resolve interannual variability due to low accumulation rates. Using longer timescales will necessitate a new metric of proxy uncertainty; the uncertainty in annual layering becomes less important on lower-resolution timescales, while the spatial representation uncertainty due to ice flow becomes more important with longer timescales. This will increase the optimality of low-accumulation, low-ice velocity sites (e.g. ice domes on the East Antarctic Plateau). A longer-timescale optimal sensor analysis would rely solely on the MME, because the 40-year-long MRE is too short in duration to serve as a source of proxy estimates at these timescales.
An additional future application is to specifically reconstruct a continental metric rather than reconstructing histories for each grid cell and aggregating them. For past annual Antarctic accumulation, existing climate reconstructions have chosen to reconstruct each grid cell (e.g. Medley and Thomas, Reference Medley and Thomas2019; Dalaiden and others, Reference Dalaiden, Goosse, Rezsöhazy and Thomas2021; Eswaran and others, Reference Eswaran, Truax and Fudge2024); continental-scale metrics are then calculated by spatially integrating across the grid cells. If the goal is to reconstruct the continental value, it may be more accurate and efficient to directly reconstruct the scalar metric, as this does not rely on the assumption of spatially homogeneous reconstruction skill (King and others, Reference King, Tierney, Osman, Judd and Anchukaitis2023a).
5. Conclusion
We present a novel data-driven approach to optimize the selection of Antarctic shallow ice-core sites for their ability to reconstruct annual Antarctic accumulation and average temperature at annual resolution since 1800. We choose two Antarctic-wide climate metrics to illustrate how the optimal sensor method can be used to guide ice-core site selection. The method is equally applicable for many other climate metrics, including regional reconstructions and other climate parameters. We account for dating and preservation uncertainty by scaling the uncertainty to the inverse of the accumulation rate. This enables us to produce a quantitative ranking of potential ice-core sites, to answer the pertinent questions of both 1) where to drill new records (to better constrain 19th–20th century climate variability) and 2) where to redrill old records (to extend their chronologies into the 21st-century).
We find that the southern end of the Ross Ice Shelf is the best drilling region for reconstructing annual Antarctic accumulation (Fig. 7b), while the west coast of DML is the best region to reconstruct annual Antarctic temperature (Fig. 7c). We recognize that the optimal sensor algorithm cannot capture all the relevant complexities of ice-core site selection, so we present three ‘case studies’ to investigate the feasibility of these sites: 1) new drilling at CIR, 2) new drilling at Lyddan Ice Rise, and 3) redrilling the US-ITASE cores. Crary and Lyddan Ice Rises are logistically feasible to access and have existing ice-penetrating radar profiles, which allow for more precise selection of drilling sites. For the redrilling of the US-ITASE cores, we find that redrilling three records is the best trade-off between maximizing climate reconstruction skill and minimizing the logistical burden. In a broader sense, there is a compelling need to update the Antarctic ice-core network, particularly as the pace of drilling has slowed in recent decades. New records would place recent mass change in a historical context, and would extend the overlap period with satellite-era reanalysis. This paper offers a data-driven approach to optimize site selection and provides an objective means to answer the question of where to drill ice cores for reconstructing any climate metric of choice.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/jog.2026.10218.
Data availability statement
The results of our experiments can be found on Zenodo at https://doi.org/10.5281/zenodo.21523198. The code used to reproduce our results can be found at https://github.com/advikeswaran/optimal-sensor. The optimal sensor analyses were conducted using the DASH open-source data assimilation framework (King and others, Reference King, Tierney, Osman, Judd and Anchukaitis2023a). Positions of the ice-core sites were taken from the PAGES Antarctica 2k records (Stenni and others, Reference Stenni2017; Thomas and others, Reference Thomas2017), with some additions (Criscitiello and others, Reference Criscitiello2013; Medley and others, Reference Medley2013, Reference Medley2018; Philippe and others, Reference Philippe2016; Kahle and others, Reference Kahle2021). The MME was constructed using CMIP5 Last Millennium simulations (Collins and others, Reference Collins, Tett and Cooper2001; Gent and others, Reference Gent2011; Phipps and others, Reference Phipps, Rotstayn, Gordon, Roberts, Hirst and Budd2012; Bao and others, Reference Bao2013; Dufresne and others, Reference Dufresne2013; Schmidt and others, Reference Schmidt2014; Wu and others, Reference Wu2014; Brady and others, Reference Brady2019; Gutjahr and others, Reference Gutjahr2019). The MRE was constructed using four reanalyses: ERA5 (Hersbach and others, Reference Hersbach2020), MERRA-2 (Gelaro and others, Reference Gelaro2017), CFSR (Saha and others, Reference Saha2014) and JRA-55 (Kobayashi and others, Reference Kobayashi2015). For the case studies, data are qualitatively analyzed from ice-penetrating radar at Crary Ice Rise (Hillebrand and others, Reference Hillebrand, Conway, Koutnik, Martín, Paden and Winberry2021) and Lyddan Ice Rise (King and others, Reference King, Anderson, Vaughan, Mann, Mobbs and Vosper2004), as well as the US-ITASE ice cores in West Antarctica (Kaspari and others, Reference Kaspari2004; Dixon and others, Reference Dixon2005; Banta and others, Reference Banta, McConnell, Frey, Bales and Taylor2008; Steig and others, Reference Steig2013).
Acknowledgements
We thank Mathieu Casado (scientific editor) and four anonymous reviewers for their constructive feedback and improvements to the manuscript. We thank Jonathan King for providing the open-source code framework (DASH) used to implement the optimal sensor algorithm. The PAGES Antarctica2k and CMIP initiatives, along with reanalyses developed by ECMWF, NASA, NOAA and JMA, provided the compilation of original data sources for this study. We thank the ITASE initiative for originally identifying, drilling and analyzing cores from many of the sites examined in this study. We thank Michael Town and members of the University of Washington Glaciology group for fruitful discussions. A.E. was supported by funding from the University of Washington Program on Climate Change. Travel funding for A.E. was supported by the Princeton University Office of Undergraduate Research. O.J.T., S.M. and V.H.L.W. were supported by the New Zealand Ministry of Business, Innovation and Employment through the Antarctic Science Platform (ANTARCTICANZ2504). V.H.L.W. was supported by a Rutherford Discovery Fellowship awarded by the Royal Society Te Apārangi from New Zealand Government funding (RDF-VUW2203).
Author contributions
A.E., O.J.T. and T.J.F. conceptualized the project. A.E. performed all the experiments and data analysis in this study, and prepared the original draft of the manuscript. O.J.T. and T.J.F. were the primary supervisors of this study and guided the design of the methodology and the interpretation of the results. O.J.T., S.M., K.M., P.D.N., E.J.S., V.H.L.W. and T.J.F. provided feedback on the draft and guided the formulation of the case studies.
Open access
