Southern Ocean surface pressure and winds during the 20th century from proxy-data assimilation

Winds and pressure over the Southern Ocean are critical to many aspects of the climate system, including ocean circulation and carbon uptake, sea ice extent, and the mass balance of Antarctica. However, reliable climate data around Antarctica begin only in 1979. Here, we reconstruct sea level pressure and zonal surface wind anomalies over the Southern Ocean through the 20th century, using data assimilation with a global database of paleoclimate proxy records. There is very good agreement between the reconstructions and satellite-based reanalysis products both at the large scale and in the smaller Amundsen Sea, a key region of West Antarctica where rapid glacier retreat has occurred in recent decades. The reconstructions show insignicant trends in the zonal-average circumpolar westerlies, but a signicant strengthening in mid-latitude Pacic westerlies, associated with a deepening of the Amundsen Sea Low, beginning well before the satellite era. The mean zonal-wind trend along the continental shelf break in the Amundsen Sea is easterly through the 20th century, contrasting with previous results that have suggested that glacier change in this region can be attributed to strengthening westerlies. Our reconstructions underscore the value of using paleoclimate data assimilation methods in assessing historical changes in Southern Hemisphere climate and suggest that zonally asymmetric features of atmospheric circulation may be key for understanding high-latitude climate and associated ice-sheet changes.


Introduction
Winds over the Southern high latitudes play a pivotal role in the global heat and carbon budget through their in uence on air-sea exchange and vertical convection in the Southern Ocean 1,2 . Winds also play a signi cant role in sea ice variability 3,4 and in modulating the distribution of water masses that affect the delivery of heat to the margin of the Antarctic Ice Sheet, potentially affecting ice sheet stability 5,6 .
Much attention has been paid to changes in the large-scale westerlies, which have strengthened and may have shifted poleward in recent decades 7,8 . Such changes are commonly expressed as a positive trend in the Southern Annular Mode (SAM) index, a measure of the zonal mean difference in high vs. mid-latitude pressure around the Southern Ocean 9 . It is also important to consider the spatial pattern of atmospheric circulation changes 10,11 . For example, it has been suggested that the intensi cation of the large-scale circumpolar winds may explain the widespread increase of ice discharge in Antarctica 12 , because winddriven incursions of Circumpolar Deep Water (CDW) onto the continental shelf enhance the melting of oating ice shelves 5,13 . Yet, the largest changes in climate have occurred in the relatively small region of West Antarctica, where zonally asymmetric features of atmospheric circulation are more important 14 ; wind variability in this region is uncorrelated with the SAM index 15 .
The greatest increases in ice discharge have occurred in the Amundsen Sea region of West Antarctica 6 .
Wind variability in this region is associated with the Amundsen Sea Low (ASL), which is strongly in uenced by the El Niño/ Southern Oscillation (ENSO) in the tropical Paci c 15,16,17 . Large ENSO events have a demonstrated impact on the ow of CDW onto the continental shelf 18 , but it is not clear that tropical forcing can account for the increasing ice discharge that has evidently occurred through the 20th century. This has led to the suggestion, supported by climate-model simulations, that interannual wind variability in the Amundsen Sea region, while dominated by tropical forcing, is superimposed on a longterm westerly trend driven by radiative forcing 19 .
A challenge to understanding how Southern Ocean wind and associated pressure elds have affected the ocean and cryosphere is that neither is well documented prior to the satellite era (i.e., before 1979). Global atmospheric reanalyses and other gridded products of wind and pressure that span the 20th century 20,21,22 show large inconsistencies with one another at high southern latitudes 23 and are known to be particularly uncertain in the critical West Antarctic sector 21 .
Advances in data assimilation methods using paleoclimate proxy records provide an opportunity to improve our understanding of wind and pressure changes at the high latitudes, where instrumental data are often sparser than proxy data. A well-established method, which employs the ensemble Kalman lter, has been used to reconstruct a range of variables, including precipitation, geopotential height, and sea ice, and has proven to be competitive with instrumental reanalyses 24,25,26 .This approach blends temporal information from temperature-sensitive proxy records with spatial covariance information from climate models, used as the prior ensemble. It relies on comparison between the temperature eld in the prior ensemble with the proxy data, which are calibrated against instrumental temperature records, accounting for uncertainties in both. Here, we use this method to reconstruct gridded anomalies of annual-mean sea level pressure (SLP) and zonal surface winds (U S ) from 1900 through 2005, as these are the most relevant variables to the oceanographic and glacier changes occurring around Antarctica (Methods). We also reconstruct surface temperature. We assess the skill of our reconstructions by comparing them to instrumental reanalysis products and other proxy-based reconstructions using standard metrics (Methods). We discuss trends in the data over recent decades, and the full 20th century, in the context of climate and glaciological applications. We consider both the large-scale changes relevant to the response of high latitude climate to radiative forcing, and the regional changes relevant to Antarctica, with particular attention to the Amundsen Sea region.

Veri cation and comparison with instrumental reanalyses
We discuss two main reconstructions generated from assimilation of proxy records with two different last millennium climate model priors, from the Community Earth System Model ("CESM" 27 ) and the Hadley Centre Coupled Model, version 3 ("HadCM3" 28 ). We evaluate our results by comparing SLP and U S reconstructions with the established instrumental reanalysis products ERA5 29 and NCEP2 30 . Maps of the correlation (r) and coe cient of e ciency (CE) between our reconstructions and the instrumental reanalyses during the satellite era (since 1979) indicate very good skill (correlation p-values < 0.05, CE > 0) for SLP and U S over most of the Southern Hemisphere (Figs. 1, S1, S2). For both variables, skill is greatest in the Paci c sector of the Southern Ocean, re ecting the density of ice-core proxy data near this region and the strong teleconnections between the West Antarctic and lower latitudes 17,31 . Correlation and CE tend to be slightly greater in the reconstruction that uses CESM as the prior, than in the reconstruction using HadCM3.
Our reconstructions reproduce both the magnitude and spatial pattern of trends in annual-mean SLP and U S that are observed in ERA5 and NCEP2 (Fig. S3). All show the large-scale pattern characteristic of the well-known increase in the SAM index 32 : increasing SLP at mid latitudes, decreasing SLP poleward of ~ 50°S, and stronger mid-latitude westerlies. Our reconstructions also capture the deepening of the ASL that has been noted in recent decades 33 and is documented in HADSLP2 34 data since 1951 11 (Fig. 2).
HadSLP2, which uses surface observations only (i.e., no satellite data), was produced with an interpolation procedure independent of the reanalyses products. As in HADSLP2, ASL deepening since the 1950s in our reconstructions is accompanied by increasing SLP over the Drake Passage, a pattern that has also been noted in previous work 35 .
Our reconstructions also reproduce the well-established warming trend over the Antarctic Peninsula and the West Antarctic Ice Sheet since the 1950s, and insigni cant trends over East Antarctica (Fig. S4), comparing well with independent statistical reconstructions 14,36 . This pattern of warming and cooling in Antarctica is consistent with the SLP trends 14,37,38 . Reconstructions obtained with and without ice core data show that these temperature and SLP patterns are derived predominantly from the ice-core proxy data, particularly the snow-accumulation records. Previous work has noted that Antarctic accumulation records provide a strong constraint on temperature 39 . Reconstructions conducted with the direct assimilation of δ 18 O data (i.e., independent of calibration against instrumental temperature data) produce similar temperature, SLP and U S patterns, though the trends are comparatively damped (Fig. S4).
Hence, both ice-core accumulation and δ 18 O records play an important role in our data assimilation results. Indeed, reconstructions without ice cores show a trend toward increasing pressure in the Amundsen Sea, opposite to that in the HADSLP2 observations. 2.2 Large-scale SLP and wind trends during the 20th century As noted in the previous section, our reconstructions reproduce the SLP and zonal-wind trends during the satellite era that are commonly associated with an increase in the SAM index. Here, we consider trends over the full 20th century by calculating the SAM index from our reconstructions using the difference in zonal average SLP between 65°S and 40°S (Fig. S5). Our results compare well (correlation p-values < 0.05, Table S2) with the instrumental Marshall SAM index 9 as well as with a previously published 1000- year-long reconstruction of the SAM index 40 (calculated with a weighted composite of Antarctic ice-core and South American tree-ring data). Consistent with previous work, our results show a weak positive trend in the SAM Index earlier in the 20th century and a stronger positive trend late in the 20th century (Table  S2). Correlations with the Marshall SAM index and the Abram index are poor if ice-core data are excluded (Table S2), again highlighting the important contribution from ice cores.
An increase in the SAM index is associated in the literature with a combination of the strengthening and poleward shift of the polar jet and surface westerlies 41 . We quantify the Southern Hemisphere (SH) surface westerly position as the latitude of the center of mass of U s between 20°S and 70°S (Methods).
The SH surface westerly strength is the wind speed at that same latitude. We perform these calculations for both the circumpolar average and by sector (Paci c, Atlantic, and Indian; Fig. 3). We nd excellent agreement in interannual variability between the reconstructions and results of the same calculations for instrumental reanalysis products for the period 1979-2005, for both circumpolar position and strength (Table S2). We note that the reconstructions may inherit mean biases from their priors, but we nd only small biases in the reconstructions relative to ERA5 and NCEP2 (Methods , Table S4).
In our reconstructions of SH westerly wind position from 1979-2005, there is a more equatorward trend in the Paci c sector than in the zonal average, though associated with large uncertainties (Fig. 3).
Similarly, previous work shows inconclusive trends in annual mean SH westerly position since 1979 8 , with contrasting regional trends 42 . Our results show that there is no signi cant trend (Methods) in annual mean position for the full 20th century, and that contrasting regional trends started well before the satellite era (Table S3). This indicates that SH westerly position is subject to large variability, and trends associated with short datasets should be interpreted with caution. We also compare our results with two additional instrumental reanalysis products, 20CRv3 22 and ERA20C 20 , which extend through the full 20th century. Both show a signi cant poleward shift and strengthening of the zonal surface winds during the full 20th century, with much larger trends than our reconstructions indicate (Table S4). The greatest differences in these 20th century instrumental reanalyses lie in the rst half of the century (Fig. S6), which is not surprising given that both 20CRv3 and ERA20C are largely unconstrained by high-latitude observations prior to the satellite era. Our resultswhich are constrained by high-latitude observations throughout the entire 20th century --thus support previous work suggesting that the zonal wind and pressure trends in 20th century instrumental reanalysis products are spuriously large 11 .
2.3. Wind and pressure trends during the 20th century in the Amundsen Sea region As noted above, a distinct feature of our reconstructions is deepening of the Amundsen Sea Low and increasing pressure over the Drake Passage (Fig. 2). This zonally asymmetric pattern persists through the 20th century in our reconstructions (Fig. 4), contrasting with the 20CRv3 and ERA20C instrumental reanalysis products, as well as with climate-model simulations, which tend to show more annular, zonally symmetric strengthening of the circumpolar westerlies (Fig. S8). Both of our reconstructions show similar patterns in the Amundsen Sea region; areas of disagreement are largely restricted to the East Antarctic sector, where veri cation skill is lowest and there are the fewest observational constraints (Figs. 1, S1, S2). The pressure pattern in the Amundsen Sea is associated and physically consistent with the westerly wind trend evident in the mid-latitude Paci c, as well as an easterly trend along the Antarctic coastline.
Pressure and wind changes in the small region of the Amundsen Sea Embayment, close to the Antarctic continent, are of particular interest. The zonal winds along the continental shelf break (70-72°S, 102-114°W) provide a useful index of the processes relevant to wind-driven changes associated with glacier retreat 5,18 . Our reconstructions of SLP in this region show excellent agreement with instrumental reanalyses (Figs. 5, S9, Table S2). For winds, agreement with ERA5 is good, but is poor with NCEP2; we note that independent assessments have concluded that ERA Interim (the previous generation of the ERA5 product), is the most accurate product for Amundsen Sea winds 43 . For our reconstructions, the mean zonal wind trend for the period 1900 to 2005 is easterly; this is accompanied by a negative pressure trend. These trends are all statistically signi cant (Table S2), and the sign of these trends is robust to the use of different proxy types (e.g., with and without the ice core data) (Fig. S10).

Discussion
The reconstructions presented here extend the record of sea level pressure and surface winds over the Southern Ocean through the 20th century, constrained by high-latitude proxy observations. There is very good agreement with instrumental reanalyses during the period of overlap, 1979-2005, at both large and regional scales. Trends in the SH surface westerlies during the satellite era are similar to satellite-based reanalyses, within uncertainties, and the patterns of longitudinal variation have been seen in previous work 42 and other reanalysis products, such as MERRA 44 and JRA55 45 . We emphasize that while our reconstructions are limited to annual-mean only, instrumental reanalyses and climate models show statistically signi cant poleward shifts since 1979 only during the austral summer, counteracted by equatorward shifts during the austral spring 8 . Thus, the annual mean trend is weak, and generally not statistically signi cant, as our reconstructions show. Our results do not support the strong 20th century trends suggested by the 20CRv3 or ERA20C instrumental reanalyses in either SH westerly wind position or strength. We do, however, nd signi cant strengthening of the surface westerlies in the Paci c sector throughout the 20th century. This strengthening is closely related to the deepening of the ASL, which is statistically signi cant for the full period of our reconstruction (1900-2005, Table S2), and has been observed in independent analyses since the 1950s 11 .
Our results have implications for the role of radiative forcing in driving atmospheric circulation changes at high southern latitudes over the 20th century. While strengthening and poleward-shifting SH westerlies during the austral summer are associated with increasing greenhouse gases and stratospheric ozone depletion 46,47,42 , longitudinal variations in position can be attributed to natural variability 42 , suggesting that the insigni cant trends in 20th century position observed here may be a result of large internal variability. However, the strengthening mid-latitude westerlies in the Paci c sector may be associated with increased greenhouse gases, as this is a feature found in historical simulations with increased CO 2 forcing 42 .
Our results also have implications for our understanding of the relationship between climate and glacier change in West Antarctica. In particular, a recent study suggested that there has been a mean westerly trend in zonal winds along the continental shelf break in the Amundsen Sea through the 20th century 19 , which could explain the increased ice discharge in this region. That study 19 used climate simulations from a "tropical pacemaker" climate-model ensemble ("PACE"), in which fully coupled-runs of CESM1.1 were constrained to follow observed tropical sea surface temperatures 48 , but are unconstrained in the extratropics. In our reconstructions, the mean 20th century trend in this region is easterly. This difference is not an artifact of our reconstruction method: we conducted a "pseudoproxy" experiment, in which we sample the PACE temperature eld at the locations of real proxy data and use these to reconstruct the known climate model history (Methods). Both the interannual variability and trends in the Amundsen Sea are faithfully reproduced (Figs. S11, S12). Thus, the discrepancy between our ndings and those in PACE for winds in the Amundsen Sea may be related to the simulation of the ASL in climate models. Historical climate model simulations for the 20th century, including PACE, generally do not show the ASL deepening seen in our reconstructions (Figs. S8, S12). Previous work has noted that the ASL is poorly represented in climate models, probably because of inadequate resolution of the Antarctic topography 49,50 . Because the ASL is associated with meridional ow that affects winds around West Antarctica 51,10 , it is thus not surprising that climate models may not adequately simulate historical wind changes at the regional scale in the Amundsen Sea.
Changing winds have almost certainly played a leading role in the enhanced intrusion of CDW onto the continental shelf in the Amundsen Sea, and elsewhere in Antarctica, driving Antarctic ice-sheet mass loss 12 , but our results show that relatively simple mechanisms like increasing westerlies along the continental shelf break 19 do not provide an adequate explanation. Instead, glacier change in Antarctica may have been initiated by speci c westerly wind events, such as are associated with the very strong 1939-1942 El Niño 52,15,53 , or by changes in event frequency associated with changes in the intensity of ENSO activity in recent decades 54,55 . Alternatively, the properties of CDW on the continental shelf may be more strongly affected by changes in the large-scale circumpolar westerlies, well to the north of the continental shelf 56 than has been suggested by previous work focused primarily at the regional scale 13 .
Our reconstructions underscore the value of using paleoclimate data assimilation methods in assessing historical changes in climate, especially in the high latitudes of Southern Hemisphere where instrumental data constraints are sparse. Our results also support other work that suggests that climate models may not adequately simulate higher-order patterns of atmospheric circulation variability such as the Amundsen Sea Low 50 , and suggest that simulations such as the PACE ensemble may produce a simpler, more annular circumpolar westerly strengthening through the 20th century than has occurred in reality.
Such limitations should be considered when considering projections of high latitude climate, and associated ice-sheet changes, in the coming centuries.

Methods
Data assimilation method. The reconstructions are generated using the ensemble Kalman lter to spatially spread information from global paleoclimate proxy data using teleconnection patterns captured by covariance structures in climate models, following the Last Millennium Reanalysis framework 25,26 . The ensemble approach generates a distribution of climate states, allowing us to estimate the most probable state as well as quantify associated errors. Additionally, this approach can generate multiple climate elds at once, including those not typically derived from proxy data. We use an o ine data assimilation method in which we randomly draw annually-averaged anomaly states from climate model output to form a prior ensemble with 100 members, which we use for every year reconstructed. This approach ensures that the temporal variance and trends in the nal reconstruction, or "posterior" ensemble, are generated from the proxies rather than variability in the model. This approach also leaves us with a posterior ensemble of 100 members, which we use to calculate uncertainty. Additionally, the o ine approach allows us to test the sensitivity of the reconstruction to the choice of climate model used for the prior.
Our ensemble data assimilation method follows these equations: The forward model used to map the prior estimate to proxy space is a seasonal linear regression between instrumental temperature anomalies and each proxy record for the period of overlap. For tree-ring records, we use a seasonal bilinear forward model calibrated to both instrumental temperature and precipitation data with proxy-speci c seasonality 26 . Our calibration data for temperature and precipitation are GISTEMP 60 and Global Precipitation Climatology Centre 61 (GPCC), respectively. For ice core records from Antarctica, we calibrate the data to temperatures from Nicolas and Bromwich, 2014 rather than GISTEMP. Nicolas and Bromwich, 2014 is the most up to date and accurate spatial temperature reconstruction for Antarctica, though we note that the reconstructions show only very minor differences if we calibrate the ice cores to GISTEMP. The ice core records we use comprise both water-isotope (δ 18 O) records and annual-accumulation records; both are calibrated to temperature, following a study 39 that found that using both accumulation and δ 18 O yields better performance for temperature than reconstructions using δ 18 O alone. GPCC is not reliable over Antarctica, and while other precipitation data sets are available, climate models generally show signi cant bias in accumulation; we therefore do not use an accumulation prior.
The anomaly reference period in the reconstructions is 1961-1990 unless otherwise noted (e.g., in comparisons to satellite-based reanalysis products that start in 1979, in which we shift the anomaly reference period in the reconstructions by moving them by the difference in reconstruction mean from 1961-1990 and 1979-2005). The reconstructions are regridded to 1° resolution.
Correlation and coe cient of e ciency calculations. To assess reconstruction skill, we calculate correlation to measure signal timing (after accounting for autocorrelation) and coe cient of e ciency to quantify errors in signal amplitude or bias. We use the Pearson correlation coe cient to calculate correlation on the ensemble mean time series. Correlation signi cance is calculated with the 2-tailed student t-test with 95% con dence, using the number of independent samples (N) that accounts for autocorrelation 62 : Calculations of trend magnitude and uncertainty. Trends are calculated as the slope of a linear regression for the ensemble mean of the reconstruction. To calculate uncertainty on the magnitude of the trend, we randomly draw from the 100 ensemble members for each year to generate a random ensemble of time series (i.e., the ensemble indices are randomly labeled in time). We perform this step 100 times and calculate the uncertainty as twice the standard deviation of the 100 linear ts associated with these random draws. We note that 100 is su cient for the statistics to converge. We use 2σ as the uncertainty bounds on the magnitude of the trend, as the data these are performed on are approximately normal (by failing to reject a Kolmogorov-Smirnov test for normality with 95% con dence). Thus, 2 standard deviations from the mean is equivalent to the middle 95% of the 100-member ensemble trends.
Trend signi cance. Trends are considered statistically signi cant if two criteria are met: (1) the trend magnitudes and their uncertainties for both reconstructions overlap with each other and do not overlap with 0 (i.e. they are consistent in sign and magnitude), and (2) if the 95% con dence intervals of the ensemble mean trend do not overlap with 0, after accounting for autocorrelation, in both reconstructions (i.e. how meaningful the linear trend is for how noisy the data are). The 95% con dence interval of the linear t (for the second criteria) is calculated as follows: Southern hemisphere surface westerly position and strength calculations. We calculate the circumpolar Southern Hemisphere (SH) westerly wind position as the center of mass latitude for zonally averaged westerly winds between the latitudes of 20°S and 70°S. The SH surface wind strength is the wind speed at the center of mass latitude. To perform these calculations, we add the mean of the prior back into the respective reconstruction so that we are working with the full climate eld. Trends are calculated with 100 randomly generated time series drawn from ensemble members following the steps outlined above. We perform this calculation on the zonal average (full circumpolar) and by regions 64 (Table S3).
Model bias in the reconstructions. The trends in the reconstructions are generated by information from the proxies, but the mean wind speeds and wind positions are largely inherited from the models that the priors are drawn from (Fig. S7, Table S4). As a result, the CESM reconstruction shows a positive bias in wind strength (1.19 m/s) relative to ERA5, while the HadCM3 reconstruction shows a negative bias relative to NCEP2 (-1.19 m/s). The reconstructions show minor position biases relative to both ERA5 and NCEP2 (-0.43° to 0.86°). These biases are small compared to biases seen across CMIP5 and CMIP6 models 50 . Wind position biases in models have been shown to affect trends in the circumpolar westerly wind position 64,65,66 , so trends subject to biases must be treated with caution. However, even with contrasting biases in strength and small biases in position, both reconstructions show insigni cant trends in position and weak strengthening of the SH westerlies during the 20th century, with a large strengthening component from the Paci c region.
Pseudoproxy experiments. We use the same data assimilation method as the real-proxy reconstructions to generate pseudoproxy reconstructions, with a few changes. We use a prior from a historical model so that we can generate pseudoproxies from the 20th century. We use a PACE prior composed of ensemble members 2-20, concatenated together to form a su ciently long list of climate states to draw from. We generate the pseudoproxies by drawing temperature from the proxy locations in ensemble member 1 of PACE. Since the pseudoproxies are taken directly from the temperature elds, no forward model is needed. We use an observational error of 0.01°C. The ability of the pseudoproxy reconstructions to reproduce the historical climate models allows us to understand the upper bound of how well our method can reproduce the truth, given the proxy locations that we have and the covariance structures in the models. Declarations