West-warming East-cooling trend over Antarctica reversed since early 21st century driven by large-scale circulation variation

Antarctic climate changes prior to 2000 were characterized by a strong zonally asymmetric pattern. Over 90% of the land ice mass loss occurred around a limited area in West Antarctica, accompanied by a rapid surface warming rate about three times the global mean. In contrast, surface warming and glacier mass loss around East Antarctica are not significant, until the decades since 2000 when several individual stations show that the temperature trends might have reversed. The asymmetric climate changes between East- and West-Antarctica are largely attributable to the inter-decadal variabilities over the Pacific and Atlantic Oceans through tropical–polar teleconnections, leaving open the question of whether the post-2000 phase shift of the lower-latitude decadal variability causes a flip of the asymmetric Antarctic changes. Here, by synthesizing 26 in-situ observations and 6 reanalysis datasets using a statistical method and integrating the results with a series of climate model experiments, we find that the West-warming, East-cooling trend over Antarctica has systematically reversed in austral spring since the early 21st century, largely due to the atmospheric circulation anomaly over the Antarctic Peninsula–Weddell Sea region, which is associated to the teleconnection with Pacific and atmospheric internal variability. This reversal of the temperature seesaw suggests that substantial decadal-scale fluctuations of the Antarctic climate system exist, including for sea-ice and land-ice systems, superimposed on and modifying longer term changes.


Introduction
In the satellite era since 1979, zonally asymmetric climate changes have been observed over the Antarctic continent, the most dramatic being a rapid warming over West Antarctica and the Antarctic Peninsula [1][2][3], accompanied by accelerated land ice mass loss and sea ice melting [4][5][6]. In contrast, a mild cooling trend was observed over the East Antarctic [3,7] accompanied by even glacial mass gain [4]. Even though the current mass loss of Antarctic ice shelves is mainly attributed to their basal melting [8], the Antarctic surface warming plays an important role in promoting the Antarctic surface mass loss in the future [9]. These seesaw-like climate changes between the East and West Antarctic have broad implications for global sea level rise [10], the global energy balance [11], and the global carbon cycle [12].
Since the early 2000s, the warming trend over the North Atlantic slowed down, implying an upcoming reversal of the AMO [19], while the phase of the IPO started to switch after the 2015-2016 El Nino event [19,33,35]. Meanwhile, scattered observations over both the East-and the West-Antarctic reported some reversed surface air temperature (SAT) trend signals [7,16,36]: warming trends at several stations in West Antarctica [36] and the Antarctic Peninsula [7,16] decreased beginning around 2000, whereas a record warming has recently been reported at the South Pole [37,38], although with strong regional and seasonal features [7,16,36,37] and a large uncertainty [1,7,36,37]. This naturally raises questions as to whether a systematic adjustment of the Antarctic climate system started after 2000 and whether changes in decadal-scale tropical variability is contributing to this polar adjustment. Answering these questions requires accurate high-resolution observations for identifying the regionality, seasonality, and decadal variation in Antarctic SAT trends. Unfortunately, reliable observations for explaining this complexity are limited. Surface temperature observations from only about 20 Antarctic stations are available for assessing multi-decadal trends [7], with spatial coverage that is too sparse to represent the entire continent. Further, pronounced biases and spurious trends exist in the satellite observations [39] and reanalysis datasets over Antarctica [40,41]. These disadvantages in existing datasets [39][40][41] hinder the assessment of the extent to which the inferred climate change reversals are robust or systematic and what broader processes underpin the switch.
Nevertheless, while any individual dataset may not represent the complexity of the Antarctic SAT trend, each of the above mentioned datasets provides additional information to reduce the uncertainty, underscoring the need for advanced statistical techniques that combine all available data to obtain more definitive estimates of Antarctic SAT trends. Here, using such an approach, referred to as combined maximum covariance analysis (CMCA) [42], the most coherent signals among different reanalyses and observation datasets are retrieved. We thus extract these coherent SAT modes among 26 surface stations and 6 reanalysis datasets, demonstrating that trends in East Antarctica cooling and West Antarctic warming in austral spring systematically reversed after 2000. This switch is further attributed to the large-scale adjustment of the atmospheric circulation associated with the tropical-polar teleconnection and atmospheric internal variability.

Data
We use observed monthly SAT data from 26 Antarctic stations (see supplementary table 1) collected by the Scientific Committee on Antarctic Research Reference Antarctic Data for Environmental Research project [43]. Monthly SAT observation data for the Byrd Station is obtained from a reconstructed temperature record [2].
We also use monthly SAT and sea level pressure (SLP) data from six reanalysis datasets: (1) the Climate Forecast System Reanalysis (CFSR) [44] and the National Center for Environmental Prediction (NCEP) coupled forecast system model version 2 (CFSv2) [45]; (2) the NCEP-Department of Energy (DOE) Reanalysis version-2 (NCEP2) [46]; (3) the Modern Era Retrospective-Analysis for Research and Applications, version 2 (MERRA2) [47]; (4) the Japanese 55 year Reanalysis (JRA55) [48]; (5) the Interim reanalysis data from the European Centre for Medium-Range Weather Forecasts (ERA-Interim) [49]; and (6) the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) [50]. Because the resolutions of these datasets differ, we interpolated all reanalysis datasets into a 2 • × 2 • grid. These datasets cover a common period from 1979 to 2019, except that MERRA2 started in 1980 and ERA-Interim ended in August 2019. The monthly 500 hPa geopotential height (GPH) data from ERA5 are also used to calculate the SAM, PSA1, and PSA2 patterns and the associated time series, given the superior quality of this dataset over the Southern Hemisphere compared with other reanalysis datasets [51,52]. In addition, the monthly Hadley Centre SST dataset (HadISST) [53] at 1 • × 1 • horizontal resolution has been used to obtain the SST regression map onto time series for each mode of the CMCA (see below).
The ZW3 pattern is identified using the third harmonic of a Fourier transform of ERA5 500 hPa GPH over 1979-2019 averaged across the latitude band of 55-63 • S.

The CMCA method
The CMCA method [42] is an effective statistical tool to extract the most coherent spatial-temporal modes among multiple different datasets with different mappings for the same variable [54]. Here we perform the CMCA method on the detrended monthly SAT anomalies of the seven observational and reanalysis datasets over the Antarctic (90 • S-55 • S) to obtain the most prominent coherent modes of SAT interannual and decadal variability during the period of 1979-2019.
The coherent modes with the largest agreement among different datasets are calculated by maximizing the combined covariance matrix among different datasets, which is defined as: where q is the sample size of each dataset, which refers to the time range. X obs is the state variable (SAT) of all in-situ observations with 26 stations, each with 123 samples for September-October-November (SON), from 1979 to 2019. X rea is a 6m × q matrix (m is the total number of spatial grids of each reanalysis, while q is the total number of time steps), which combines the state variable of all six reanalyses at each time step. The singular value decomposition is then performed on the combined covariance matrix C to retrieve the largest coherent spatial-temporal modes that maximize the cross-covariance among all reanalyses and station observations. Notably, the spatial patterns of these CMCA modes among the six reanalysis datasets show high consistency [54]. Therefore, here we use the ensemble mean of reanalyses to represent a combined reanalysis spatial pattern for each mode.

Other statistical methods
The SAM, PSA1, and PSA2 modes in this study are the first three Empirical Orthogonal Function (EOFs) [55], respectively, retrieved from the monthly ERA5 500 hPa GPH anomalies over the Southern Hemisphere (0 • -90 • S). Spatial correlation coefficients (weighted by the area of each grid cell) have been used to illustrate the coherence between circulation patterns associated with CMCA Mode 1 and PSA2. We use the Sen's slope method [56] to calculate the linear trend, with the significance level calculated through the Mann-Kendall test [57]. Linear regression, Pearson's correlation, and composite analysis are also used in this study to (respectively) investigate the regression maps of the CMCA time series and the correlation between PSA2 and ZW3. The statistical significance is calculated using the Student's t-test, which is undertaken using the effective degree of freedom taking the autocorrelation of the variables into consideration.
In this study, two statistical methods were employed to identify the change points of the time series. The first method utilized was the robust change point detection method developed by Alin et al [58]. This method identifies the robust change points by calculating the sum of squares for error for various regression models. The second method employed was the traditional moving t-test [59], which is based on the 15 year moving trend of the time series. This method detects significant change points by testing whether the trend has undergone a sudden change at a certain point. The change points are determined according to the intersection of these two methods.

Seesaw SAT trends between East and West Antarctica
We calculate the trends of the Antarctic SAT using 26 frequently-used stations for two periods, 1979-2000 and 2001-2019. Although the sparse distribution of these stations may not represent the SAT trends over the entire Antarctic, the distribution of these warming and cooling signals depicts a seesawlike pattern, especially for austral spring (SON). Most of the stations over West Antarctica experienced rapid warming trends before 2000 (figure 1(a)), followed by cooling trends after ( figure 1(b)). This reversal is especially evident at the Faraday/Vernadsky (figure 1(c)) and Rothera (figure 1(d)) stations located at the west coast of the Antarctic Peninsula, as well as the Byrd Station (figure 1(e)) over in-land West Antarctica. In contrast, most of the stations over East Antarctica show mild, insignificant cooling trends before 2000 (figure 1(a)), followed by warming trends after ( figure 1(b)). This switch is clear at the Neumayer Station (figure 1(f)) over Queen Maud Land and the Amundsen-Scott and Vostok stations (figures 1(g) and (h)) in the interior of East Antarctica.
The multi-decadal SAT trends (based on in-situ observations) for the other three seasons are illustrated in supplementary figure 1. Similar seesawlike SAT trends are seen in most of these seasons, especially for austral summer (December-January-February, DJF, see supplementary figures 1(a) and (d)). However, these seesaw signals are either weaker (see supplementary figures 1(a) and (d)), or more complex (see supplementary figures 1(b), (c), (e) and (f)) than those of the austral spring (figure 1). Below we focus exclusively on the seesaw-SAT trend and its reversal in spring (SON).
To further evaluate the spatial distribution of this seesaw-like SAT trend, we illustrate the spring SAT trends before and after 2000 based on three stateof-the-art reanalysis products, namely the MERRA2 (see supplementary figures 2(a) and (d)), the JRA55 (see supplementary figures 2(b) and (e)), and the ERA5 datasets (see supplementary figures 2(c) and (f)). Although these reanalyses show some degree of the West warming-East cooling trend reversal after 2000, there is considerable discrepancy in the spatial extent and magnitude of the west-east trend pattern, which is not surprising given previously reported discrepancies among these reanalyses [40,41]. These discrepancies hinder the further assessment of the west-east pattern, including its robustness or its potential causes, and call for a statistical approach that might yet distill a robust and common signal from these datasets.
Despite limited station observations and significant discrepancies between reanalysis datasets, effective methods can be used to combine all available information and extract reliable patterns of SAT variability. The CMCA method has been demonstrated to be particularly effective for this purpose, providing a means to extract coherent information from multiple datasets pertaining to the same variable [54]. In this study, we therefore apply the CMCA method to 6 reanalysis datasets and 26 in-situ observations to identify the most consistent modes of Antarctic SAT variability, and further investigate the robustness and underlying mechanisms of the reversal of the Antarctic SAT trend in SON. Figure 2 shows the spatial patterns and time series of the first three CMCA modes of SON SAT. The first mode (figure 2(a)) explains 39% of the total covariance, showing a seesaw pattern between West Antarctica (including the Antarctic Peninsula) and a large portion of East Antarctica comprising Queen Maud Land and interior East Antarctica, hereafter referred to as the Antarctic-seesaw SAT pattern. Importantly, the associated time series (figure 2(b)) of Mode 1 shows a significant trend reversal in ∼2000. A significant negative trend before 2000 is observed, which represents substantial warming over West Antarctica and cooling over the Queen Maud Land area. After 2000, a significant positive trend is observed, with warming over most regions of East Antarctica and cooling over West Antarctica. In short, the first mode represents a seesaw multi-decadal SAT trend pattern between West Antarctica and the Queen Maud Land area with a phase reversal in the early 21 st century.
The spatial pattern of CMCA Mode 2 (figure 2(c)) shows a pan-Antarctic cooling at its positive phase (hereafter named as Antarctic-continental SAT pattern), which might be related to the variability of the SAM [60] or radiative forcing associated with changes in greenhouse gas concentrations. Mode 3 (figure 2(e)) depicts a West Antarctic SAT-dipole pattern [61][62][63] between the Antarctic Peninsula and the west of Marie Byrd Land [63]. Unlike Mode 1, the time series of Modes 2 and 3 (figures 2(d) and (f)) show no significant trends. The first three modes of CMCA explain more than 86% of the covariance between the set of 26 stations and the 6 reanalysis datasets, indicating that these three modes represent the majority of SAT variability over the Antarctic in austral spring.
We also conducted a CMCA analysis on monthly temperature anomalies for the other three seasons (see supplementary figures 3 and 4). The first three leading modes show very similar spatial patterns to those for the spring, with the rank of percentage variance explained by Mode 1 (Antarctic-seesaw pattern) and Mode 2 (Antarctic-continental pattern) switched. The similarity of the leading three modes among different seasons reinforces the robustness of these dominant modes of Antarctic SAT variability. Below we discuss the circulation associated with the SON seesaw-SAT trend and its reversal.

Anomalous atmospheric circulation associated with the seesaw SAT trends
The SLP projected onto the time series of the Antarctic-seesaw SAT pattern (CMCA Mode 1) through linear regression ( figure 3(a)) shows a Rossby wave-like pattern, with an anomalous high-pressure center over the Ross Sea and a low-pressure center over the Antarctic Peninsula-Weddell Sea region; the latter extends to inland Antarctica and induces large-scale adjustments in thermal advection. In particular, during its positive phase (as shown in figure 3(a), predominantly post-2000), the deep low pressure center over the Antarctic Peninsula-Weddell Sea drives cold air advection from interior Antarctica to West Antarctica and warm advection from the Southern Ocean to Queen Maud Land. For the negative phase (predominantly pre-2000), the orientation of the above warm-cold advection is reversed (not shown), thus helping explain the switch of the seesaw SAT pattern in austral spring. The anomalous circulation pattern appears linked to the second PSA mode (PSA2); the correlation coefficient between the circulation pattern associated with the seesaw SAT mode (CMCA Mode 1) (i.e. figure 3(a)) and the PSA2 ( figure 3(b)) reaches 0.91, and their temporal correlation coefficient is 0.6 ( figure 3(c)).
The seesaw pattern shows strong signals along the coastal regions of the Queen Maud Land and the West Antarctic, as depicted in figures 2(a) and 3(a). To investigate the potential role of oceanic processes in these coastal signals, we conducted further analysis by calculating the sea ice concentration (SIC) and SST anomalies associated with the seesaw SAT mode in SON through linear regression (see supplementary figure 5). Our findings indicate that the high loadings across the coastal regions (figure 2(a)) are closely linked to the air-sea heat flux associated with the SIC and SST anomalies (see supplementary figure 5). Specifically, the positive phase of the seesaw mode (figure 2(a)) corresponds to a sea ice expansion over the Amundsen-Bellingshausen Sea and a sea ice retreat over the coastal East Antarctic (see supplementary figure 5(a)) caused by thermal advection and mechanical forcing [64,65]. These sea ice anomalies further cool the West Antarctic and heat the East Antarctic by reducing/intensifying the heat flux from the ocean to the atmosphere [66,67]. Additionally, the positive phase of the seesaw SAT mode is also closely associated with sea surface warming over the Atlantic sector of South Ocean and surface cooling over the Amundsen-Bellingshausen Sea (see supplementary figure 5(b)), which may contribute to the coastal warming/cooling over the East/West Antarctic by impacting on the sea ice anomalies [68] and airsea heat flux. Therefore, the oceanic processes related to the SIC and SST anomalies play a crucial role in the high loading over the coastal regions of the seesaw SAT mode.
The circulation pattern associated with the Antarctic-continental mode (CMCA Mode 2) (see supplementary figure 6(a)) is dominated by Recent studies [69,70] indicated that the PSA2 is also associated with atmospheric internal variability including the ZW3 mode over the southern hemisphere. Correlation between the PSA2 and different atmospheric variables, including SAT, 500 hPa GPH, mean SLP, 500 hPa wind speed, and 10 m meridional wind speed, show clear ZW3 imprints in these fields (see supplementary figure 7). Composite anomalies over seven years with the largest (85th percentile) PSA2 years show that the positive phase of PSA2 is associated with a ∼30 • (longitude) eastward shift of the ZW3 pattern (see supplementary figure 8), including a shift of the prominent Amundsen Sea Low (ASL) circulation mode eastward toward the Antarctic Peninsula and Weddell Sea, which here contributes to the deep pressure center over the Antarctic Peninsula-Weddell Sea after 2000.

Conclusion and discussion
Studies investigating Antarctic climate variability and trends are challenged by the limited observational data [3,40]. Available in-situ observations are too sparse to represent variability over the entire Antarctic continent [7]. On the other hand, although we have many high-resolution gridded datasets, including satellite observations and reanalyses, the wide range of observational techniques, accuracies, and methods for combining observations and models necessarily lead to inconsistencies between different datasets [3,40,41]. Our application of the CMCA method extracts dominant modes of temperature variability in the station observations and six different reanalyses, one of which represents the coherent SAT mode featuring the Antarctic-seesaw pattern.
The dynamics of Antarctic climate change and its zonal-asymmetric feature is a fundamental issue in climate science [1,71], especially given its broad implication for global sea level rise [10], global energy balance [11], and the global carbon cycle [12]. Recent studies indicated that the zonally asymmetric Antarctic SAT trends with East-cooling and West-warming during the second half of the 20th century were largely driven by the multi-decadal SST variability over the Pacific and the Atlantic Oceans. The SAM [60], the ASL [72,73], and remote forcing from the tropical oceans [19,22,28] are each a part of the dynamics driving the observed Antarctic climate variability and change [22,23,28,[72][73][74][75], especially the dipole-like sea ice and SAT changes over the West Antarctic between the Ross Sea and Amundsen-Bellingshausen Seas.
Here we reveal a broader, continental-wide, seesaw SAT pattern between East and West Antarctica, which underwent a reversal around 2000 ( figure 4). These reversed trends show strong warming over Queen Maud Land and inland East Antarctica and cooling over West Antarctica during 2000-2019. In contrast to previous studies that attributed the zonal asymmetry of the Antarctic SAT trend to the ASL [19,20,28,76], we find that this continental-wide seesaw SAT change is related to thermal advection caused by an anomalous circulation center over the Antarctic Peninsula-Weddell Sea region. Before 2000, there was an anomalous high-pressure center over the Peninsula-Weddell Sea region, leading to a warm air advection from lower latitudes to West Antarctica, and a cold air advection from the South Pole to Queen Maud Land ( figure 4(a)). After 2000, an anomalous low-pressure center was present over the Peninsula-Weddell Sea region [5], reversing the patterns of thermal advection prior to 2000 ( figure 4(b)). This anomalous circulation center is in part a result of forcing from the tropical Pacific, and in part an eastward shift of the ZW3 projected onto PSA2 variability, although the ultimate cause for the changes involving ZW3 and PSA2 requires further investigation, giving the dynamical complexity of the PSA2 [55,69,70,77].
During the satellite era since 1979, the Antarctic has experienced a series of asymmetric climate changes, with the strongest changes typically appearing over the West Antarctic [2][3][4]73]. The reversal of the Antarctic seesaw SAT trend pattern and the associated adjustment of the large-scale atmospheric circulation over the Antarctic we reported have broad implications for projecting long-term changes of the Antarctic sea-ice and land-ice [4][5][6]9], in that these changes might not appear unidirectional and can be masked by the adjustment we documented. The Antarctic climate adjustment revealed by our study potentially has broad implications for the future projections of sea level rise, the energy budget of the earth system, and the global carbon cycle.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files).