The global and regional climate changed dramatically with the
expansion of the Antarctic Ice Sheet at the Eocene–Oligocene transition
(EOT). These large-scale changes are generally linked to declining
atmospheric
Global cooling and the significant expansion of glacial ice over Antarctica at
the Eocene–Oligocene transition (EOT)
While global circulation models (GCMs) are useful tools for testing our understanding of the Earth system, their inherent uncertainty within this region shows that it is necessary to integrate proxy evidence to build up a more robust picture of Southern Ocean changes across the EOT. To this end, here we compile a large multi-proxy database of temperature for the high-latitude Southern Hemisphere, incorporating a multitude of different proxy records in terms of methods, sites and temporal coverage. Despite sometimes not being directly comparable, the inclusion of very different kinds of proxy evidence provides both qualitative and quantitative measures against which model simulations can be compared and evaluated. The quantitative elements of the dataset can also be used to describe general temperature patterns (e.g. in terms of the regional mean or latitudinal gradient), and model simulations that perform relatively well can then be used in conjunction with the proxy dataset to start to explain what changes may have occurred in this region across the EOT.
Proxy records of past climate and the “equilibrium” climate simulations
generally performed for the EOT both have strengths and weaknesses. Proxy
records, specifically sediment cores, are particularly good for
reconstructing the temporal domain of past climate, showing changes through
long time periods at a particular point in space (e.g. Zachos et al., 2001).
By contrast, complex fully coupled climate models generally cannot be run
for long (
Equilibrium climate simulations also simplify orbital variations that would
have occurred on timescales ranging from 10
The aim of this data synthesis is to create proxy datasets that are comparable to model simulations, i.e. can be used to validate the models in the spatial domain. This necessitates reducing the temporal variability of the proxy data into broad time slices, which was done for late Eocene absolute conditions (generally 36.4–34.0 Myr), relative changes across the EOT and early Oligocene absolute conditions (generally 33.2–32.0 Myr). Once time averaged, it is assumed that the records should be more representative of the longer-term climate state at their location. Dictated by the nature and inherent uncertainties in the age models associated with the proxy data, the definition of the time slices remains reasonably crude. Indeed, proxy records used will be on different age models at each locality and cover somewhat different periods and lengths of time. This introduces an element of uncertainty. In addition to and possibly driven by orbital variability, it has been shown that there was variability in the few million years either side of the Eocene–Oligocene boundary (E–O; e.g. Coxall and Pearson, 2007; Scher et al., 2014; Galeotti et al., 2016). Time averaging approximately 2 Myr prior to and after the E–O will, however, average out this temporal variability (if a long record for a particular location is available) or potentially skew results (if, for example, a short-term excursion is captured in the record). To an extent these uncertainties are unavoidable and must be considered when interpreting the results presented here.
Two specific research questions are addressed in this paper. Firstly, what are the spatial patterns of temperature change inferred from proxy records for the high-latitude Southern Hemisphere before, after and across the EOT? Secondly, which GCM boundary conditions give the best fit to the range of qualitative and quantitative proxy records of temperature before, after and across the EOT?
A brief overview of the data synthesis, model simulation details and evaluation methods follow in Sect. 2. Section 3 presents the results of the model–data comparison. Finally, Sects. 4 and 5 discuss the significance of the results and the potential scope of future research, respectively.
Many different proxy records for in situ sea surface temperature (SST) are
available. These include quantitative records using stable isotopes and
trace metals (
Some terrestrial surface air temperature (SAT) records are also available, such as those derived from clay weathering products (S-index; e.g. Passchier et al., 2013) and from vegetation reconstructions (based on the nearest living relative, NLR, e.g. Francis et al., 2009, or the coexistence approach, e.g. Pound and Salzmann, 2017). These records may or may not be in situ (in time or space), with clay weathering products, for example, having been exported from terrestrial regions to where they are deposited in ocean sediment cores.
Values and data are compiled from a range of sources within published material. Ideally, the data are taken from the supplementary material of the related papers. In other cases, mean values might be quoted in tables, figures or in the text of papers; however, it can be unclear over what time period these means are taken or how uncertainty values are calculated. Although this is not the most accurate way of obtaining data, in some cases this might provide the only data available and so still warrants inclusion. The sources of all data points used are outlined in detail in the Supplement (Tables S1–S3), a digital version of which can also be accessed through the Open Science Framework (Kennedy-Asser, 2019).
These proxies respond to the climate system in different ways and all rely on various assumptions, resulting in uncertainty ranges that can be incorporated into the model–data comparison. Uncertainty in the proxy data records could arise due to calibration uncertainties or could be particularly due to temporal variability in the record (as noted in Sect. 1). These various aspects make it challenging to rigorously define and quantify uncertainty. Generally, uncertainty is taken as the published values where available. Alternatively, generalized calibration uncertainty for a given proxy (if known) or 2 standard deviations of the temporal variability in the records can be taken as the uncertainty. Some records are presented in terms of annual temperature range, and these limits can be taken as the uncertainty around the annual mean (assumed to be the mean of the maximum and minimum of the range). The sources of the uncertainty ranges used are also detailed in the Supplement of this paper (Tables S1–S3) and Kennedy-Asser (2019).
It is likely that some seasonal (summer) bias is incorporated into marine
proxy records, particularly at high latitudes where light and temperature
may become limiting in certain periods. In contrast, for SAT estimates based
on vegetation, other conditions such as high atmospheric
Some studies (e.g. Waelbroeck et al., 2009; Dowsett et al., 2012; Pound and Salzmann, 2017) devise semi-quantitative metrics for the quality of proxy records based upon factors such as preservation, dating quality and calibration errors when compiling their datasets. Here, there is no formal assessment of the quality of individual proxies or records, nor is there any reinterpretation or recalculation of existing datasets, as this would be beyond the scope of the paper. Instead, here the dataset integrates as many independent proxies as possible for each site, and all are used to evaluate the model simulations. It is important to note that the same proxy is only used in the compilation once per site per time slice. If two or more records using the same proxy at the same site are available, generally the most recent value in the literature is used (e.g. Passchier et al., 2013 and 2016 both provide estimates for temperature using the S-index in Prydz Bay, so the 2016 value is used). Different proxies are weighted equally in the model evaluation, with sites where there are multiple records therefore being weighted more strongly for the purpose of model–data comparison.
In total, data were taken from 14 sites (10 ocean and 4 terrestrial),
ranging in palaeolatitude from 53 to 77
Mean annual temperature (
Changes in mean annual temperature (
The proxy datasets compiled here are compared to the fully spun-up HadCM3BL-M2.1aE (HadCM3BL henceforth) simulations outlined in Kennedy-Asser et al. (2019) and simulations from FOAM outlined in Ladant et al. (2014). An overview of the simulations used is provided in Table 1. A detailed description of the model setup and the simulation details can be found in the respective references.
Brief overview of climate models and the boundary conditions varied for each.
These models are all relatively low resolution and less complex than some others that have been used in recent studies (e.g. Hutchinson et al., 2018; Baatsen et al., 2018); however, they are still regularly used in palaeoclimate research (e.g. Goddéris et al., 2017; Farnsworth et al., 2019; Saupe et al., 2019). For the present day, HadCM3BL is shown to perform comparably to CMIP5 models in terms of a number of global mean variables, although it produces a moderate cold bias globally, with high (northern) latitudes being too cold because of an exaggerated seasonal cycle and overly cold winter (Valdes et al., 2017). This bias is similar to other higher-resolution variants of the model (Valdes et al., 2017). FOAM has been shown to capture most of the major characteristics of present-day climatology (Jacob, 1997; Liu et al., 2003) as well as reasonable climate variability (Wu and Liu, 2005). As HadCM3BL, FOAM exhibits a cold high-latitude bias in the Northern Hemisphere, in particular in winter (Gallimore et al., 2005).
It should be noted that for both models the Antarctic ice sheets are
prescribed and cannot expand or contract through the simulations and also
that, along with the palaeogeographies, the ice sheets used are different
between the original studies. Orbital variability was accounted for in the
FOAM simulations (Ladant et al., 2014), having both warm summer and cold
summer orbital variants available for comparison. Given that it is not
possible to definitively show if proxy records are capturing extreme cases
of orbital variability, these simulations are used to inform the potential
magnitude of uncertainty this might introduce. The model spin-up period also
differs between the two studies, with the HadCM3BL simulations being
significantly longer. The HadCM3BL simulations were selected from
Kennedy-Asser et al. (2019) based upon their extended spin-up, meaning the
modelled results are expected to be highly robust with negligible trends to
bias the conclusions. FOAM simulations have been integrated for 2000 years
and are in equilibrium in the upper ocean. Small cooling trends exist in the
deep ocean, but the rates of temperature change are smaller than
0.1
In order to evaluate against the proxy dataset of relative changes across
the EOT, pairs of model simulations can be selected that represent the
forcing changes occurring across the EOT. These pairs of model simulations
represent a before and an after state, with the difference in the boundary
conditions between the pairs described as the forcing and the difference in
the modelled climate representing the change across the EOT. Given that the
vast majority of glaciological proxy data give evidence of glacial
expansion, here the modelled forcing must include some sort of ice expansion
(i.e. the early Oligocene simulation must contain an ice sheet, and the late
Eocene simulation must contain no ice sheet). The simulation pairs may
additionally include other forcing changes that are potentially relevant to
describe the state of the Earth system before and after the EOT, namely
the an expansion of ice over Antarctica from an ice-free state to either an EAIS
or full AIS, with all other boundary conditions remaining the same; a similar expansion of ice over Antarctica but also combined with a
simultaneous drop in a similar expansion of ice over Antarctica but also combined with a
simultaneous change in palaeogeography (an opening of the Drake Passage),
with a similar expansion of ice over Antarctica but also combined with a
simultaneous change in palaeogeography (an opening of the Drake Passage) and
a drop in
This produced 9 pairs of simulations from HadCM3BL and 18 pairs from FOAM.
FOAM simulations were always compared with the same orbital variability
before and after the EOT. A detailed description of all simulations and
simulation pairs used is included in Table S4.
Most proxies in this compilation provide continuous quantitative data that can be directly compared to models or other records (e.g. absolute temperature estimates from geochemical proxies). Other proxies may provide ordinal (qualitative) data; that is, data that can be ranked into an order of greater or lesser magnitude but from which absolute values are not attainable (e.g. dinoflagellate species assemblage). Both of these kinds of data can be used to evaluate the palaeoclimate model simulations.
At each site where proxy data are available, the modelled annual mean air or
water temperature is taken as the mean over a three by three grid cell area
surrounding each proxy location, with the maximum and minimum modelled
temperature also taken from these nine grid cells as the modelled
uncertainty. Given the relatively coarse resolution of these models, this
represents a very large area (ranging 2.25–
Firstly, the “standard” RMSE, defined in Eq. (1), is calculated from the
maximum or minimum of the uncertainty range of the proxy data to the minimum
or maximum of the uncertainty range in the model (if the model is too warm
or cold, respectively):
Secondly, the RMSE is calculated once the mean temperature of all data
points and sites (either in the proxy dataset or for a given model simulation)
has been removed. The purpose of removing the mean is so the model
performance is not primarily judged against systematic warm or cold biases,
the latter of which are typical at high latitudes in palaeoclimate
simulations of past warm climates (Huber and Caballero, 2011; Lunt et al.,
2012). This “normalized” RMSE, defined in Eq. (3), instead evaluates the
spatial pattern of temperature in the Southern Ocean. This metric is used
with continuous data when a mean value is available, again with the error
taken between the ranges of the proxy and model uncertainty:
“Count metrics” can also be used for the absolute and relative change data comparisons, allowing a large range of proxy records to be incorporated. These metrics count how many of the data points the model is consistent with in terms of magnitude (i.e. within the error bars) and, for the change across the EOT, the number of records for which the model simulations correctly predict the direction of change. This can allow ordinal data (such as increasing cold water taxa) to contribute to the comparison.
In order to assess the simulations across multiple criteria, metric scores that have comparable units (e.g. the two RMSE metrics) can simply be summed or averaged. Additionally, to further expand upon the idealized model–data comparison of Kennedy-Asser et al. (2019), it is important to consider not just if the simulated change across the EOT is realistic, but also if the starting and ending state are realistic compared to the late Eocene and early Oligocene datasets. This is done by combining the metric scores for a pair of simulations that describe the change across the EOT (compared to the EOT dataset) with metric scores for the pre- and post-EOT simulations that make up that pairing (compared to the late Eocene and early Oligocene datasets). If the datasets had a consistent spatial coverage for each of the time slices, the difference between the late Eocene and early Oligocene absolute datasets would be the same as the EOT relative change dataset. However, because there are some sites with records available only before or after the EOT, and some relative changes for which absolute values are not available, the pair of simulations that gives the best fit before and after the EOT is not necessarily the same pair that gives the best fit for the observed change across the EOT. Which metric is used to evaluate across the time slices and if there is any weighting put on the absolute or relative change datasets is subjective. Although the count metrics are shown for reference, here the combined rank score for each time slice is based upon only the two RMSE metrics, and the three time slices (late Eocene, early Oligocene and EOT) are weighted equally.
For the model simulations to be described as performing particularly “well”
or “poorly”, it is necessary to have some sort of benchmark to compare the
model performance against. For the three time slices, two benchmarks are
used: these can be thought of as hypothetical generalizations of the whole
regional high-latitude Southern Hemisphere climate based only upon proxy
data. First, the mean temperature (or temperature change) of all sites and
proxies is taken as a homogeneous value at all sites. Second, the ordinary
least squares linear fit through the mean temperatures (or temperature
change) with palaeolatitude from all proxies and sites, shown in Fig. 3,
is taken to produce a synthetic, latitudinally varying temperature field for
the region. If model simulations perform better than both benchmarks, they
can be described as showing
The regional means of the proxy records and latitudinal temperature gradient
benchmarks are shown in Fig. 3 along with the best HadCM3BL and FOAM
simulations identified in Sect. 3.5. In addition to the latitudinal
gradient calculated for the full dataset for each time slice, an uncertainty
range for the gradient was calculated by systematically omitting single
points from the regression to test for potential bias in the proxy record
compilation. The absolute temperature profiles in the late Eocene and early
Oligocene proxy datasets show colder temperatures at higher latitudes than
mid-latitudes, as would be expected. The latitudinal gradient is comparable
between the early Oligocene (0.54
Latitudinal profiles of
Standard RMSE (
The standard RMSE, normalized RMSE and count metric for all of the ice-free
simulations and the benchmarks in comparison to the late Eocene dataset are
shown in Fig. 4a for the annual mean temperature. Equivalent simulations
for the summer mean temperatures are shown in Fig. S2a. The
standard RMSE scores show that absolute temperature biases are generally
large compared with the benchmarks. The standard RMSE scores are better for
simulations at higher
When the mean temperature bias is removed for the normalized RMSE, more of
the simulations outperform the constant mean benchmark and some outperform
the latitudinal gradient benchmark. For FOAM, with the cold temperature bias
removed, the lower
No simulations outperform both benchmarks for both RMSE metrics, so none can
be described as good by our definition. However, the HadCM3BL simulation at 3
Figure 4b shows the standard RMSE, normalized RMSE and count metric for all
glaciated simulations against the early Oligocene dataset for the annual
mean temperature (with summer temperatures shown in Fig. S2b). Again, there is a general cold bias indicated by the poorer standard
RMSE scores for the lower
For the normalized RMSE, all simulations outperform at least one benchmark.
The HadCM3BL simulations with an open Drake Passage at either
Like for the late Eocene, no simulation can be described as good for both RMSE
metrics; however, the glaciated HadCM3BL simulation at 3
All pairs of model simulations representing the change in annual mean temperature that occurred across the EOT are shown in Fig. 5. A comparable plot using summer mean temperatures is shown in Fig. S3. It is important to note that, generally, the uncertainties in the EOT dataset are much greater relative to the magnitude of change compared to the uncertainties relative to the absolute values in the late Eocene and early Oligocene datasets. As a result, the latitudinal gradient benchmark provides a remarkably good fit for the data covering the EOT, lying almost entirely within the data uncertainty. No model simulations perform as well as this benchmark, but again, because the uncertainty in the change relative to its magnitude is greater than in the absolute datasets, generally the model RMSE scores are lower for this dataset than the late Eocene or early Oligocene datasets. For this dataset, there is not a clear picture of modelled changes being over- or underestimated relative to the proxy records. The largest error for all models is in representing the large cooling shown at the Falklands Plateau.
Standard RMSE (
Three HadCM3BL simulation pairs (Fig. 5a) outperform the constant mean
change benchmark for the standard RMSE metric: those with an open Drake
Passage in response to AIS growth and a
Similar to what was shown for the late Oligocene, the FOAM simulations
(Fig. 5b) generally fit the dataset best in terms of the standard and
normalized RMSE when they have smaller ice sheets added. Although it makes
little difference for the normalized RMSE scores, FOAM simulations which
combine a
Generally, in terms of the forcings across all model simulation pairs, the
AIS growth forcing in isolation produces the best normalized RMSE and
performs comparably to the combined AIS growth and
No simulations from any model perform better than either benchmark for both
RMSE metrics, with the best HadCM3BL simulation pairing (with an open Drake
Passage in response to both AIS growth and
As noted in Sect. 2.3, it is possible to evaluate the model simulations and model simulation pairs across various metrics. The five best simulations (or simulation pairs) for the late Eocene, early Oligocene and for the change across the EOT based on the mean of their two RMSE metrics are shown in Table 2, along with the mean of the two RMSE metrics for each of the benchmarks for comparison. As well as taking the average RMSE for each time slice, the average RMSE can be taken across all three time slices. It is not always the case that simulation pairs that perform well for the observed EOT change also perform well when the late Eocene and early Oligocene data are incorporated. As was noted in Sect. 3.1 and 3.2, for the absolute temperatures, simulations with a closed Drake Passage perform relatively poorly. As a result, when the combined ranked performance score is calculated across all three time slices, the pairings with a closed Drake Passage are not found to perform as well, highlighting the importance of incorporating the absolute values into this model–data comparison. Again, this suggests that the Drake Passage was open prior to the EOT and the late Eocene. The five best simulations in terms of the mean standard RMSE and normalized RMSE across all three time slices are also listed in Table 2, along with the benchmarks for comparison.
The five highest ranked simulations (or simulation pairs) in terms of mean standard and normalized RMSE for each time slice and across all three time slices.
This model–data comparison shows that the most realistic representation of
the high-latitude Southern Hemisphere climate before, after and across the
EOT would be simulated by the expansion of an AIS, possibly with some
combination of atmospheric
The marked reduction in performance by HadCM3BL when the Drake Passage is either closed before and after the EOT or closed before but opens across the EOT supports the conclusions of Goldner et al. (2014) that changes in ocean gateways around the EOT are not the best way to model the changes observed in the proxy record. This is in support of the general shift in consensus away from the gateway hypothesis as the sole cause of the changes at the EOT, at least in terms of the direct climatic implications (DeConto and Pollard, 2003; Huber and Nof, 2006; Sijp et al., 2011; Ladant et al., 2014 etc.). However, a preconditioning by gateway deepening and invigorated Antarctic Circumpolar Current is still plausible based on SST proxy data and microfossil distribution from directly prior to the EOT (Houben et al., 2019). There is inconclusive evidence in the literature for fundamental changes in the Drake Passage around the EOT (e.g. Lagabrielle et al., 2009, and references therein), in agreement with the Getech palaeogeographic reconstructions, which have the gateway open throughout the period (see the Lunt et al., 2016, Fig. S1; Kennedy-Asser et al., 2019, Fig. S1). However, it should be noted that proxy evidence and reconstructions suggest the Tasman Seaway deepened close to, but probably prior to, the EOT (e.g. Stickley et al., 2004; Scher et al., 2015; Houben et al., 2019) and this could have different implications for the climate (which are potentially more consistent with the temperature proxy records compiled here) from the results shown for Drake Passage opening. The preconditioning effects of widening and deepening the Tasman Seaway could therefore be of interest in future model comparisons.
It is important to bear in mind that this result was obtained from a relatively low-resolution model. With higher-resolution models, it is possible that changes in modelled ocean circulation and atmospheric response could be very different, particularly given that much smaller changes in the Southern Ocean gateways than were modelled here could have occurred across the EOT (e.g. Viebahn et al., 2016). For this paper, it was not feasible to use higher-resolution models for such a range of boundary conditions and length of simulations, and this should remain an important priority in future research.
The better fit with proxy data by FOAM when the AIS is not at its full
extent would also be consistent with the other glaciological evidence.
Various sites around the Ross Sea showed the maximum AIS expansion occurring
around
HadCM3BL simulations with differing AIS extent boundary conditions (those with the Getech palaeogeographic reconstructions from Kennedy-Asser et al., 2019) also show a similar result, with simulations with a smaller EAIS fitting the data better (figure not shown). However, as was discussed in Kennedy-Asser et al., these simulations are potentially not fully spun up, so they are not included in the analysis of this paper. It should be noted that the FOAM simulations have a relatively short spin-up of 2000 years (Table 1), and without deeper investigation into the time series of the model spin-up, it is not possible to say if this model is fully in equilibrium yet.
Although this model–data comparison provides some interesting results, there
is still clear room to improve model performance and reduce
discrepancies with the data. The zonal mean temperature for each of the best
pairs of simulations from HadCM3BL and FOAM (across all three time slices)
are shown in Fig. 3 along with the proxy records and their uncertainty.
For the late Eocene and early Oligocene, the latitudinal gradients produced
by the models are reasonably similar to the gradients shown in the proxy
records, although the models generally have a cold bias of around 5–10
Although there could be an element of seasonal bias in some of the proxy
records (Hollis et al., 2019) that could explain absolute temperature biases
before and after the EOT, the supplementary results presented here show that
using modelled summer temperatures generally results in worse model
performance for the relative change across the EOT and for the normalized
RMSE scores. Higher-resolution modelling and better representation of
climate feedbacks offer some potential improvements in this regard (Huber and Caballero, 2011; Baatsen et al., 2018), and the current DeepMIP
modelling effort (Lunt et al., 2017) might provide further insights into the
causes of this common model bias. It should also be noted that these
simulations were run with relatively arbitrary
A major concern identified in the model–data comparison is that even the
best simulation pairs for both models do a poor job at recreating the change
across the EOT compared to the latitudinal gradient benchmark or even the
constant mean benchmark (Fig. 3c). From 55 to 65
Critically assessing the proxy records that are included in the compilation could explain some of the differences between the records and the models. For example, it can be unclear as to what area the terrestrial proxies such as the S-index represent or to what extent this record is affected by reworking. The S-index, like any detrital-based proxy, will suffer to some extent from the reworking of older material (Passchier et al., 2013). This residual signal, primarily built up in warmer periods, implies that a warm bias is likely. Additionally, although the dataset used here was as large as could be compiled at the time of writing, there are still large data gaps spatially and temporally. It is possible the sites around the Ross Sea are part of very localized microclimates, which may not align with the average climate of the large areas covered by a model grid cell.
A second option that could partly explain the model–data discrepancy is that
local- to regional-scale warming signals in response to Antarctic glaciation due
to enhanced circulation, deepwater formation and sea ice feedbacks (as
identified in models by Goldner et al., 2014; Knorr and Lohmann, 2014;
Kennedy et al., 2015; Kennedy-Asser et al., 2019; and some of the FOAM
simulations used here from Ladant et al., 2014; figure not shown) could be
compensating for some of the cooling. When this warming is combined with a
Another significant model–data discrepancy is the strong cooling indicated
by
A final important consideration is that the temporal averaging of the
dataset carried out here could be inappropriate. A number of studies have
suggested there was cooling in the several million years prior to the EOT,
particularly at high latitudes (e.g. Raine and Askin, 2001; Petersen and Schrag, 2015;
Passchier et al., 2016; Carter et al., 2017; Pound and Salzmann, 2017). Even in the high Northern Hemisphere changes have been
identified occurring prior to the EOT (e.g. Coxall et al., 2018). These
changes could all have a range of different forcings; however, it is
possible that some of them are related. Even a global forcing such as
atmospheric
Regardless of whether late Eocene cooling was earlier or simply amplified at
higher latitudes, in both cases it is likely that the Ross Sea site
experienced significant cooling prior to the EOT. This would support
evidence of some tundra vegetation in the region recorded prior to the EOT
(Raine and Askin, 2001). It therefore might be necessary to include older
records and further split the dataset into additional time slices to capture
the climate of Antarctica before any cooling occurred. The only record of
this age included in the current dataset is the McMurdo erratic, which
suggested temperatures less than 13
Currently, the data compilation is not big enough to allow for such an analysis to be carried out; however, this could potentially offer a more appropriate comparison with the equilibrium climate model simulations used here, which are broadly “warm and ice-free” or “cool and glaciated”. If this hypothesis is correct and if more comparable records were included for the period pre-cooling and glaciation (e.g. dating from 40 Myr), it is possible that the high-latitude change from the middle to late Eocene through to the Oligocene would be greater than that shown in Fig. 3, closer in line with the model simulations.
An extensive review of temperature proxy records for the high-latitude
Southern Hemisphere region before, after and across the EOT was presented
and used to evaluate model simulations of the EOT. These simulations came
from two different GCMs with different sets of boundary conditions. The best
simulations were able to capture spatial patterning of absolute temperature
recorded in the late Eocene and early Oligocene proxy datasets. The
performances were not as good for the dataset of relative changes across the
EOT due to the models inadequately capturing changes in the latitudinal
gradient shown by the data. The latitudinal gradient discrepancy is possibly
related to the paucity of data in certain regions (particularly at very high
latitude), the time averaging of the proxy records into time slices (with
some of the higher-latitude changes possibly occurring prior to the EOT),
localized climatic effects (e.g. ocean upwelling or ice free coastal
microclimates), or the glaciation of Antarctica resulting in
some localized warming through changes in atmospheric or oceanic circulation
that approximately balances the general cooling across the EOT (e.g. due to
The best pairs of simulations for modelling the absolute temperatures and
relative changes were found by assessing the individual simulation
performances across all time slices for various metrics. This suggests that
the best simulations for representing the EOT were by HadCM3BL with an open
Drake Passage, AIS expansion and possibly a drop in atmospheric
The performance of FOAM for the early Oligocene time slice was generally better with smaller ice sheet configurations over Antarctica, potentially in agreement with proxy records of ice volume and extent (e.g. Bohaty et al., 2012; Huang et al., 2014; Galeotti et al., 2016). A similar finding is also seen in the HadCM3BL simulations using the Getech palaeogeographies (not shown; Kennedy-Asser et al., 2019); however, as these simulations could be affected by a lack of spin-up, they were not included in the analysis. Further spinning up those HadCM3BL simulations with multiple ice sheet sizes could provide some interesting insights into whether this climatic fingerprint of a smaller AIS is robust.
These results point towards some interesting conclusions about how the Earth system changed across the EOT; however, this work remains a first step upon which further research should be built. An important consideration in interpreting this model–data comparison is the relative paucity of data available for the region during the EOT (only 14 sites), in combination with records generally showing heterogeneous temperature patterns. Particularly for the normalized RMSE, an important measure for determining if the model is showing the correct spatial patterns, there are only a handful of sites which can be used across all sectors of the Southern Ocean. With the relatively limited data coverage available here, it is possible that these latitudinal profiles could be biased by anomalous values. However, as noted in Sect. 2.4, even with the most extreme points omitted for the calculation of the latitudinal gradients for each time slice, no gradient fundamentally changed. Expanding the datasets in the future as more data points become available is a more appropriate method for testing if points used here are anomalous and if the latitudinal profiles are robust.
Future research by the palaeoclimate community will inevitably produce new records in new locations, potentially refining or even correcting older, spurious results or having an impact on the inferred spatial patterns shown in the proxy record. Future work on this research could improve the consistency of the data used, for example in terms of using the same proxy calibrations, age models and definitions of uncertainty, as well as fully accounting for uncertainty in seasonal biases and orbital variations, but that is currently beyond the scope of this paper. To this end, the datasets used here have been uploaded to the Open Science Framework (Kennedy-Asser, 2019) to aid in the continuation of this research and the expansion of this analysis in the future.
Additionally, future work can also expand upon this analysis by including more model simulations and trialling other metrics and scoring techniques, as palaeoclimate modelling results are often model dependent (Lunt et al., 2012). It is also important to note that the models used here are of relatively low spatial resolution, meaning the spatial averaging of temperature is taken over a very large area, and potential smaller-scale ocean changes resulting from changes in ocean gateways may be poorly represented. Therefore, although these simulations are likely to capture large-scale climate phenomena, clearly much could be learnt in future research from using higher-resolution models.
The challenge in synthesizing the many changes that occurred in this large and heterogeneous region across the EOT is huge, but this research shows that with increased modelling and proxy data results, some convergence of ideas within the palaeoclimate community appears possible.
All of the model data presented in this research along with MATLAB scripts
used to carry out the analysis are available via the Open Science Framework
(Kennedy-Asser, 2019). Further HadCM3BL variables from these simulations
(which were not used in this research) are freely available through the
University of Bristol's BRIDGE server (
The supplement related to this article is available online at:
ATKA carried out the analysis and compiled the proxy datasets. DJL helped in the experimental design. PJV and JBL provided the climate model data. JF and VL provided guidance in the interpretation and compilation of proxy data. ATKA wrote the paper with contributions from all authors.
The authors declare that they have no conflict of interest.
HadCM3BL climate simulations were
carried out using the computational facilities of the Advanced Computing
Research Centre, University of Bristol (
This research has been supported by NERC (grant no. NE/L002434/1).
This paper was edited by Yannick Donnadieu and reviewed by two anonymous referees.