Palaeoclimate simulations improve our understanding of the climate, inform us about the performance of climate models in a different climate scenario, and help to identify robust features of the climate system. Here, we analyse Arctic warming in an ensemble of 16 simulations of the mid-Pliocene Warm Period (mPWP), derived from the Pliocene Model Intercomparison Project Phase 2 (PlioMIP2).
The PlioMIP2 ensemble simulates Arctic (60–90
The simulation of past climates improves our understanding of the climate system, and it provides an opportunity for the evaluation of the performance of climate models beyond the range of present and recent climate variability (Braconnot et al., 2012; Harrison et al., 2014, 2015; Masson-Delmotte et al., 2013; Schmidt et al., 2014). Comparisons of palaeoclimate simulations and palaeoenvironmental reconstructions have been carried out for several decades (Braconnot et al., 2007; Joussaume and Taylor, 1995) and show that while climate models can reproduce the direction and large-scale patterns of changes in climate, they tend to underestimate the magnitude of specific changes in regional climates (Braconnot et al., 2012; Harrison et al., 2015). The comparison of palaeoclimate simulations with future projections has aided in the identification of robust features of the climate system which can help constrain future projections (Harrison et al., 2015; Schmidt et al., 2014), including in the Arctic (Yoshimori and Suzuki, 2019).
One such robust feature is the Arctic amplification of global temperature anomalies (Serreze and Barry, 2011). Increased warming in the Arctic region compared to the global average is a common feature of both palaeoclimate and future climate simulations and is also present in the observational record (Collins et al., 2013; Masson-Delmotte et al., 2013). Arctic warming has a distinct seasonal character, with the largest sea surface temperature (SST) and the smallest surface air temperature (SAT) anomalies occurring in the summer due to enhanced ocean heat uptake following sea ice melt (Serreze et al., 2009; Zheng et al., 2019). It is critical to correctly simulate Arctic amplification as it is shown that projected Arctic warming affects ice sheet stability, global sea-level rise, and carbon cycle feedbacks (e.g. through permafrost melting; Masson-Delmotte et al., 2013). Several multi-model analyses that included palaeoclimate simulations and/or future projections found that changes in northern high-latitude temperatures scale (roughly) linearly with changes in global temperatures (Bracegirdle and Stephenson, 2013; Harrison et al., 2015; Izumi et al., 2013; Masson-Delmotte et al., 2006; Miller et al., 2010; Schmidt et al., 2014; Winton, 2008).
Underestimation of Arctic SAT has been reported for several climates in the Palaeoclimate Modelling Intercomparison Project Phase 3 (PMIP3), including the mid-Pliocene Warm Period (Dowsett et al., 2012; Haywood et al., 2013a; Salzmann et al., 2013), Last Interglacial (LIG; Bakker et al., 2013; Lunt et al., 2013; Otto-Bliesner et al., 2013), and Eocene (Lunt et al., 2012a). PMIP4 simulations, however, of the LIG showed good agreement with SAT reconstructions in the Canadian Arctic, Greenland, and Scandinavia, while showing overestimations in other regions (Otto-Bliesner et al., 2020). PMIP4 simulations of the Eocene were also able to capture the polar amplification indicated by SAT proxies (Lunt et al., 2020).
In the present work, we analyse the simulated Arctic warming in a new
ensemble of 16 simulations in the Pliocene Model Intercomparison Project
Phase 2 (PlioMIP2) (Haywood et al.,
2016). PlioMIP2 is designed to represent a discrete time slice within the
mid-Pliocene Warm Period (mPWP; 3.264–3.025 Ma; sometimes referred to as
mid-Piacenzian Warm Period): Marine Isotope Stage (MIS) KM5c, 3.204–3.207 Ma (Dowsett
et al., 2016, 2013; Haywood et al., 2013b, 2016). The mPWP is the most
recent period in geological history with atmospheric CO
Palaeoenvironmental reconstructions show that the elevated CO
The dominant mechanism for global warming in mid-Pliocene simulations is through changes in radiative forcing following increases in greenhouse gas concentrations (Chandan and Peltier, 2017; Hill et al., 2014; Hunter et al., 2019; Kamae et al., 2016; Lunt et al., 2012b; Stepanek et al., 2020; Tan et al., 2020). Polar warming is also dominated by changes in greenhouse gas emissivity (Hill et al., 2014; Tindall and Haywood, 2020). Apart from the changes in greenhouse gas concentrations, changes in boundary conditions that led to warming in previous simulations of the mPWP included the specified ice sheets, orography, and vegetation (Hill, 2015; Lunt et al., 2012b).
Models participating in PlioMIP2 used in this study.
In PlioMIP1, the previous phase of this project, model simulations underestimated the strong Arctic warming that is inferred from proxy records was found (Dowsett et al., 2012; Haywood et al., 2013a; Salzmann et al., 2013). This data–model discord may have been caused by uncertainties in model physics, boundary conditions, or reconstructions (Haywood et al., 2013a).
Uncertainties in model physics include physical processes that are not incorporated in the models and uncertainties in model parameters. It was found that the inclusion of chemistry–climate feedbacks from vegetation and wildfire changes leads to substantial global warming (Unger and Yue, 2014), while excluding industrial pollutants, explicitly simulating aerosol–cloud interactions (Feng et al., 2019), and decreasing atmospheric dust loading (Sagoo and Storelvmo, 2017) leads to increased Arctic warming in mPWP simulations. Similarly, in simulations of the Eocene, two models that implemented modified aerosols had better skill than other models at representing polar amplification (Lunt et al., 2020). Changes in model parameters, such as the sea ice albedo parameter (Howell et al., 2016b), may provide further opportunities for increasing data–model agreement in the Arctic.
Several studies found changes in boundary conditions that could help resolve some of the data–model discord in the Arctic for PlioMIP1 simulations. The studied changes in boundary conditions include changes in orbital forcing (Feng et al., 2017; Prescott et al., 2014; Salzmann et al., 2013), atmospheric CO2 concentrations (Feng et al., 2017; Howell et al., 2016b; Salzmann et al., 2013), and palaeogeography and bathymetry (Brierley and Fedorov, 2016; Feng et al., 2017; Hill, 2015; Otto-Bliesner et al., 2017; Robinson et al., 2011).
New in the experimental design of PlioMIP2 is a closed Bering Strait and
Canadian Archipelago in the mPWP simulation. The closure of these Arctic
Ocean gateways has been shown to alter oceanic heat transport into the North
Atlantic (Brierley
and Fedorov, 2016; Feng et al., 2017; Otto-Bliesner et al., 2017).
Additionally, the focus on a specific time slice within the mPWP allows for
reduced uncertainties in reconstructions and boundary conditions, in
particular with regards to orbital forcing. These changes have led to an
improved data–model agreement for reconstructions of SST, particularly in
the North Atlantic (Dowsett
et al., 2019; McClymont et al., 2020; (Haywood et al., 2020). Multi-model
mean (MMM) SST anomalies in the North Atlantic deviate less than 3
In the following sections, we first evaluate the simulated Arctic
(60–90
The simulations of the mPWP by 16 models participating in PlioMIP2 were used in this study. The models included in this study are listed in Table 1. A more detailed description of each model's information and experiment setup can be found in Haywood et al. (2020). All model groups incorporated the standardized set of boundary conditions from the PlioMIP2 experimental design in their simulations (Haywood et al., 2016).
For each simulation, the last 100 years of data are used for the analysis.
Individual model results are calculated on the native grid of each model.
MMM results are obtained after regridding each model's output to a
2
To evaluate the ability of climate models to simulate mPWP Arctic warming, we first perform a comparison to SAT estimates from palaeobotanical reconstructions. The data–model comparison is performed using temperature anomalies, calculated by differencing the mPWP and the pre-industrial simulation, to avoid overestimations of agreement due to strong latitudinal effects on temperature (Haywood and Valdes, 2004).
Reconstructed mPWP SATs are taken from Feng et al. (2017), who updated
and combined an earlier compilation made by Salzmann et al. (2013) (Table S1).
Qualitative estimates of confidence levels for each reconstruction were made
by Feng
et al. (2017) and Salzmann et al. (2013). Only reconstructions that are
located at or northward of 60
Simulated global and Arctic
The data–model comparison will be a point-to-point comparison of modelled
and reconstructed temperatures estimated from palaeobotanical proxies, which
initially does not take the uncertainties of the reconstructions (Table S1)
into account. The potential influence of the uncertainties in
reconstructions on the outcomes of the data–model comparison will be
investigated in a later section. The temporal range of the reconstructions
is broad and certainly not resolved to the resolution of the KM5c time
slice, unlike the dataset of SST estimates compiled by Foley and Dowsett (2019) used for PlioMIP2 SST
data–model comparisons by Haywood
et al. (2020) and McClymont et al. (2020). Prescott et al. (2014)
found that peak warmth in the mPWP would be diachronous between different
regions based on simulations with different configurations of orbital
forcing. Orbital forcing is particularly important in the high latitudes and
for proxies that may record seasonal signatures (e.g. due to recording
growing season temperatures). As such, there may be significant biases in
the dataset, as the temporal ranges of the proxies include periods with
substantially different external forcing than during the KM5c time slice for
which the simulations are run. Feng et al. (2017) investigated the effects
of different orbital configurations, as well as elevated atmospheric
CO
Further uncertainties arise due to bioclimatic ranges of fossil assemblages, errors in pre-industrial temperatures from the observational record, potential seasonal biases, and additional unquantifiable factors. Ultimately, the uncertainties constrain our ability to evaluate the Arctic warming in the PlioMIP2 simulations substantially. A more detailed description of the uncertainties in the SAT estimates can be found in the work of Salzmann et al. (2013).
The reconstructed temperatures are differenced with temperatures from the observational record to obtain proxy temperature anomalies. Observational-record temperatures are obtained from the Berkeley Earth monthly land and ocean dataset (Rohde et al., 2013a, 2013b), and the average temperature in the 1870–1899 period was used.
MMM annual temperature anomalies in the Arctic:
Furthermore, the simulation of mPWP SIE will be evaluated using three
palaeoenvironmental reconstructions that indicate whether sea ice was
perennial or seasonal at a specific location. Darby (2008) infers that perennial sea ice was
present at Lomonosov Ridge (87.5
Ratio between the mean Arctic
The PlioMIP2 experiments show substantial increases in global annual mean
SAT (ranging from 1.7 to 5.2
All models show a clear Arctic amplification, with annual mean SAT in the
Arctic (60–90
Annual mean SST in the Arctic increased by 1.3 to 4.6
Mean annual SIE (10
The greatest MMM SAT anomalies in the Arctic are found in the regions with
reduced ice sheet extent on Greenland (Haywood et al., 2016), which
generally show warming of over 10
The distinct seasonality of Arctic amplification (Serreze et al., 2009; Zheng et al., 2019) can be used to identify mechanisms causing Arctic amplification. Figure 3 depicts the seasonality of Arctic warming for each model, with monthly SAT and SST anomalies normalized by the annual mean anomaly for that specific model.
The ensemble simulates a consistent peak in Arctic SST warming between July and September (Fig. 3b). This is consistent with the response that increased seasonal heat storage from incoming heat fluxes would have upon the reduction of SIE (Serreze et al., 2009; Zheng et al., 2019). Minimum SAT warming is expected in the summer because of the increased ocean heat uptake, while maximum SAT warming is expected in the autumn and winter following the release of this heat (Pithan and Mauritsen, 2014; Serreze et al., 2009; Yoshimori and Suzuki, 2019; Zheng et al., 2019). This is not simulated by all models, however (Fig. 3a). COSMOS, GISS-E2-1-G, IPSL-CM6A-LR, and MRI-CGCM2.3 all do show this autumn and winter amplification of annual mean SAT anomalies and decreased warming in the summer. Decreased summer warming is simulated by CCSM4-Utrecht, EC-Earth 3.3, and IPSLCM5A in combination with autumn amplification and by CESM2 and NorESM1-F in combination with winter amplification. All other models in the ensemble do not show an autumn or winter amplification in combination with decreased summer warming, suggesting a more limited role of reductions in SIE underlying the seasonal cycle of Arctic SAT anomalies.
The MMM of Arctic annual SIE (sea ice concentration
The seasonal cycle of SIE anomalies is depicted in Fig. 5a. Reductions in SIE are slightly greater in the autumn (September-November) compared to other seasons for the MMM. There is, however, no consistent response in the seasonal character of SIE anomalies in the PlioMIP2 ensemble. CCSM4-UoT, CESM2, IPSLCM5A, and IPSLCM5A-2.1 simulate the largest reductions in SIE in winter (December–February), while GISS-E2-1-G and HadCM3 simulate the largest SIE reductions in spring. The remaining 10 models simulate the greatest SIE anomalies in autumn.
A more consistent response is observed when comparing monthly mean mPWP SIEs
and pre-industrial SIEs. For each model, the largest reductions in SIE in
terms of percentages occur between August and October (Fig. 5b). This may be
explained by the lesser amount of energy that is needed to melt a given percentage
of the smaller SIE that is present in the summer compared to winter. A total of 11 out
of 16 models simulate sea-ice-free conditions (SIE
There is a strong anti-correlation between annual mean Arctic SAT and SIE
anomalies (
Correlations between annual mean SIE anomalies and
To evaluate the ability of the PlioMIP2 ensemble to simulate Arctic warming,
we perform a data–model comparison with the available SAT reconstructions
for the mPWP. The data–model comparison hints at a substantial mismatch
between models and temperature reconstructions. Mean absolute deviations
(MAD) range from 5.0 to 11.2
Point-to-point comparison of MMM and reconstructed SAT. The size of SAT reconstructions is scaled by qualitatively assessed confidence levels (Salzmann et al., 2013). Data markers for reconstructions in close proximity of each other have been slightly shifted for improved visibility.
The deviation from reconstructions for each model and the PlioMIP2 and
PlioMIP1 MMMs is represented by the box and whisker plots in Fig. 8. A
consistent underestimation of the temperature estimates from SAT
reconstructions is present in the PlioMIP2 ensemble. CESM2 simulates the
smallest deviations from reconstructions in the ensemble, with a MAD of 5.0
Box and whisker plots depicting the distribution of biases (models minus reconstruction) with biases over (under) 0 representing locations where models overestimated (underestimated) reconstructed temperatures. Boxes depict the interquartile ranges (IQRs) of the distribution, whiskers extend to the 2.5th and 97.5th percentiles, the median is displayed by a horizontal line in the boxes, and outliers (outside of the 97.5th percentile) are shown by open circles outside of the whiskers. Given the sample size of 15 reconstructions, the two outer values are depicted as outliers using these definitions.
Some of the data–model discord may be caused by uncertainties in the temperature estimates (Table S1; Salzmann et al., 2013). To investigate how these uncertainties may have affected the outcomes of the data–model comparison, we construct a maximum uncertainty range. This range spans from the highest possible temperature within uncertainty and the lowest possible temperature within uncertainty. The uncertainties for the temperature estimates were taken from the compilation of mPWP Arctic SAT estimates from Feng et al. (2017) (Table S1).
Figure 9 depicts the locations for which at least one model in the ensemble
simulates a temperature within the maximum available uncertainty range of a
reconstruction. For 6 out of the 12 reconstructions that included an
uncertainty estimate, the models in the PlioMIP2 ensemble simulate
temperatures that are within the uncertainty range (Fig. 9). Additionally,
both overestimations and underestimations are present for the Magadan District
reconstruction for which no uncertainty estimate is available (60
Blue circles highlight where at least one model in the ensemble simulates a temperature that falls within the uncertainty range of the reconstruction. The size of SAT reconstructions is scaled by qualitatively assessed confidence levels (Salzmann et al., 2013). Data markers for reconstructions in close proximity of each other have been slightly shifted for improved visibility.
Number of models simulating
Ultimately, when considering the full uncertainty ranges of the reconstructions, it becomes evident that solely reducing potential errors in SAT estimates would not fully resolve the data–model discord for several locations in the Arctic. It is thus likely that other sources of error contribute to the data–model discord, such as uncertainties in model physics (e.g. Feng et al., 2019; Howell et al., 2016b; Lunt et al., 2020; Sagoo and Storelvmo, 2017; Unger and Yue, 2014) and boundary conditions (e.g. Brierley and Fedorov, 2016; Feng et al., 2017; Hill, 2015; Howell et al., 2016b; Otto-Bliesner et al., 2017; Prescott et al., 2014; Robinson et al., 2011; Salzmann et al., 2013). The focus on the KM5c time slice has helped resolve some of the data–model discord that was present in the North Atlantic for SST (Haywood et al., 2020), and similar work for SAT reconstructions may thus be beneficial. However, this may not always be possible given the lack of precise dating and chronologies available. It is at this moment unclear whether the underestimation of Arctic SAT is specific to the mid-Pliocene, through uncertainties in reconstructions or boundary conditions, or an indicator of common errors in model physics.
The limited availability of proxy evidence (three reconstructions) severely limits our ability to evaluate the simulation of mPWP sea ice in PlioMIP2 simulations. Nevertheless, a data–model comparison is still worthwhile, as the few reconstructions that are available may form an interesting out-of-sample test for the simulation of sea ice in the PlioMIP2 models.
Figure 10a depicts the number of models per grid box that simulate perennial
sea ice. Six models simulate the inferred perennial sea ice (mean sea ice
concentration
The uncertainties in both the SAT and SIE reconstructions are large, and it
may not be possible to match both datasets in their current forms. This
would require increased Arctic annual terrestrial warming compared to the
mean model (Sect. 5.1) as well as perennial sea in the summer and a large
SIE in winter (extending at least into the Iceland Sea). Moreover, McClymont et al. (2020) found
that the warmest model values in the PlioMIP2 ensemble tend to align best
with North Atlantic SST reconstructions, further indicating that strong
Arctic warming is required for data–model agreement. If there was no
perennial sea ice in the mPWP like most models in the PlioMIP2 ensemble, the
different proxy records may be more compatible, but this would be in
disagreement with findings from Darby (2008). The CCSM4-Utrecht model,
which simulated a relatively high Arctic SAT anomaly (10.5
Research into the mPWP is often motivated by a desire to understand future climate change (Burke et al., 2018; Haywood et al., 2016; Tierney et al., 2019). Here, we analyse how the mPWP may teach us about future Arctic warming by comparing two climatic features of the mPWP simulations to simulations of future climate. The climatic features include Arctic amplification and a feature for which there is some proxy evidence available that may also aid in model evaluation: the AMOC.
A linear relationship between global and Arctic temperature anomalies is
present in the PlioMIP2 ensemble (
For four ensembles of future climate simulations, from the previous phase of
the Coupled Model Intercomparison Project (CMIP), CMIP5, data for MMM Arctic
(defined there as 67.5–90
The increased Arctic warming per degree of global warming indicates that
apart from warming through changes in atmospheric CO
Using PlioMIP2 simulations for potential lessons about future warming may be improved by isolating the effects of the changes in orograph. Similar changes in ice sheets and vegetation may occur in future equilibrium warm climates, but the changes in orography are definitively non-analogous to future warming. Several groups isolated the effects of the changed orography on global warming in PlioMIP2 simulations and found that it contributes, respectively, around 23 % (IPSL6-CM6A-LR; Tan et al., 2020), 27 % (COSMOS; Stepanek et al., 2020), and 41 % (CCSM4-UoT; Chandan and Peltier, 2018) to the annual mean global warming in the mPWP simulations. Furthermore, this warming was strongest in the high latitudes (Chandan and Peltier, 2018; Tan et al., 2020) indicating that the additional Arctic warming in PlioMIP2 simulations, as compared to future climate simulations, are likely partially caused by changes in orography that are non-analogous with the modern-day orography. These findings highlight the caution that has to be taken when using palaeoclimate simulations as analogues for future climate change.
The AMOC, a major oceanic current transporting heat into the Arctic (Mahajan et al., 2011), is inferred to have been significantly stronger in the mPWP compared to pre-industrial values based on proxy evidence (Dowsett et al., 2009; Frank et al., 2002; Frenz et al., 2006; McKay et al., 2012; Ravelo and Andreasen, 2000; Raymo et al., 1996). An analysis of AMOC changes in PlioMIP2 simulations shows that, indeed, the maximum AMOC strength increases: by 4 % to 53 % (Fig. 12; Table S2: Z. Zhang et al., 2020). The closure of the Arctic Ocean gateways, in particular the Bering Strait, likely contributed to the increase in AMOC strength (Brierley and Fedorov, 2016; Feng et al., 2017; Haywood et al., 2016; Otto-Bliesner et al., 2017).
Strengthening of the AMOC contrasts projections of future changes by CMIP5 models that predict a weakening of the AMOC over the 21st century, with best estimates ranging from 11 % to 34 % depending on the chosen future emission scenario (Collins et al., 2013). These opposing responses may help explain some of the additional Arctic warming that is observed in the PlioMIP2 ensemble compared to the future climate ensembles (Fig. 11b).
The strengthening of the AMOC in the PlioMIP2 ensemble is consistent with
the additional 0.4
Maximum pre-industrial and mPWP AMOC strength (Sv). The black line indicates equal pre-industrial and mPWP maximum AMOC strength.
The PlioMIP2 ensemble simulates substantial Arctic warming and 11 out of 16 models simulate summer sea-ice-free conditions. Comparisons to reconstructions show, however, that the ensemble tends to underestimate the available reconstructions of SAT in the Arctic, although large differences in the degree of underestimation exist between the simulations. The models that simulate the largest Arctic SAT anomalies tend to match the reconstructions better, and investigation into the mechanisms underlying the increased Arctic warming in these simulations may help uncover factors that could contribute to improved data–model agreement. We find that, while some of the SAT data–model discord may be resolved by reducing uncertainties in proxies, additional improvements are likely to be found in reducing uncertainties in boundary conditions or model physics. Furthermore, there is some agreement with reconstructions of sea ice in the ensemble, especially for seasonal sea ice. The limited availability of proxy evidence and the uncertainties associated with them severely constrain the compatibility of the different proxy datasets and our ability to evaluate the Arctic warming in PlioMIP2. Increased proxy evidence of different climatic variables and additional sensitivity experiments, among other goals, are needed for a more robust evaluation of Arctic warming in the mPWP. Lastly, we find differences in Arctic climate features between the PlioMIP2 ensemble and future climate ensembles that include the magnitude of Arctic amplification and changes in AMOC strength. These differences highlight that caution has to be taken when attempting to use simulations of the mPWP to learn about future climate change.
The reconstructions used in this study are available in the Supplement. The model data can be downloaded from PlioMIP2 data server located at the School of Earth and Environment of the University of Leeds, an email can be sent to Alan Haywood (a.m.haywood@leeds.ac.uk) for access. At the time of publication, the data from CESM2, EC-EARTH3.3, GISS-E2-1-G, IPSL-CM6A-LR, and NorESM1-F can be downloaded through CMIP Search Interface at
The supplement related to this article is available online at:
QZ and WdN designed the work. WdN did the analyses and wrote the manuscript under supervision from QZ. QL and QZ performed the simulations with EC-Earth3. XL and ZZ provided input on AMOC analysis. HJD provided the input on reconstructions. All the other co-authors provided the PlioMIP2 model data and commented on the manuscript.
The authors declare that they have no conflict of interest.
This article is part of the special issue “PlioMIP Phase 2: experimental design, implementation and scientific results”. It is not associated with a conference.
The EC-Earth3 simulations are performed on the
Swedish National Infrastructure for Computing (SNIC) at the National
Supercomputer Centre (NSC). COSMOS PlioMIP2 simulations have been conducted
at the Computing and Data Center of the Alfred-Wegener-Institut
Helmholtz-Zentrum für Polar und Meeresforschung on a NEC SX-ACE high-performance vector computer. Gerrit Lohmann and Christian Stepanek acknowledge funding via the
Alfred Wegener Institute's research programme PACES2. Christian Stepanek acknowledges
funding by the Helmholtz Climate Initiative REKLIM. Camille
Contoux and Gilles Ramstein thank ANR
HADOC ANR-17-CE31-0010; the authors were granted access to the HPC resources of
TGCC under the allocations 2016-A0030107732, 2017-R0040110492,
2018-R0040110492 (gencmip6), and 2019-A0050102212 (gen2212) provided by
GENCI. The IPSL-CM6 team of the IPSL Climate Modelling Centre
(
This research has been supported by the Vetenskapsrådet (grant nos. 2013-06476, 2017-04232).The article processing charges for this open-access publication were covered by Stockholm University.
This paper was edited by Alessio Rovere and reviewed by two anonymous referees.