Since the start of the industrial age, atmospheric greenhouse gases have increased to levels that the Earth has not experienced for over at least 800 000 years. These increases have changed the Earth's energy balance with an estimated radiative forcing of 1.6 W m−2 [1]. Environmental changes associated with these changes in forcing are detectable and numerous, including a global warming of 0.74°C from 1906 to 2005 and a rate of sea-level rise averaging 1.8 mm yr−1 for 1961–2003 [1]. Projections of future surface temperature changes by AD 2100 range from 1.1 to 6.4°C depending on the emission scenario pathway followed. Uncertainties include the degree of polar amplification of temperatures, which vary among the models used for projections [2]. Yet, warming of the atmosphere, oceans and land in the polar regions has important implications for stability of the Greenland and Antarctic ice sheets, permafrost degradation and associated methane release, and sustainability of the biological diversity. Show Over the last million years, the Earth's climate has oscillated between colder, glacial climates and warmer, interglacial climates. These changes were driven by the well-known orbital periods [3], which altered the latitudinal and seasonal incoming solar radiation, with resulting feedbacks of greenhouse gases and ice sheets amplifying the orbital forcing. Some of the past interglacials over the last 500 kyr may have been warmer than today [4–6]. Understanding the forcings and feedbacks that produced interglacial warmth and the outcomes from it can help us better project the future climate of our planet. Climate model simulations allow an assessment of how well models used for projecting future climate can reproduce the evidence from past interglacials. The Last Interglacial (LIG, from approx. 130 to 116 ka) is a useful target for model–data comparisons for several reasons. Because it is the penultimate interglacial before the Holocene, our present interglacial, more data are available to compare to models than for earlier interglacials. Ocean drilling has retrieved a large number of sediment cores for the LIG [6,8,9]. Ice cores provide records of temperature derived from stable water isotopes that extend through the LIG in Greenland [9] and the last 800 kyr in Antarctica [10,11]. The availability of pollen records from the LIG, particularly at extratropical latitudes in the Northern Hemisphere (NH), allows estimates of temperature owing to the close relationship of plants and climate [12]. These palaeoclimatic records provide geographical patterns of temperature change for comparison to climate model simulations. Early simulations with atmospheric general circulation models and idealized orbital configurations of maximum tilt and eccentricity and perihelion at 21 June simulated significantly warmer NH mid- and high-latitude summers [13]. High-latitude winters were also simulated to be warmer than modern owing to sea ice feedbacks. Atmosphere ocean general circulation models (AOGCMs) with orbital forcings of 125 and 130 ka confirm these patterns of seasonal warmth [14–17]. AOGCMs though have not been able to produce the warming indicated by the East Antarctic ice cores when forced by orbital forcing changes only [17,18]. Transient simulations for the LIG with intermediate complexity models [19] suggest that Arctic warming peaked early in the interglaciation because obliquity peaked earlier than precession [20], while meltwater forcing introduced to the North Atlantic can generate an early Antarctic warming [17]. In this paper, we compare the CCSM3 climate model simulations for 125 ka to two recent data syntheses, assess the parallels and explore the differences. We use the model simulation to inform an interpretation of data seasonality [21–23]. Owing to limitations on the absolute dating of proxy records during the LIG, both data syntheses assessed maximum warmth in the period 135–118 ka and assumed this warmth was broadly synchronous in time. We include comparisons to CCSM3 simulations for 130 and 120 ka to evaluate this assumption. We also evaluate a sensitivity simulation with the West Antarctic Ice Sheet (WAIS) removed to test the impact of its extent being much reduced during the LIG [18,24]. We conclude with a comparison to a HadCM3 125 ka simulation, which gives an indication of the robustness of the temperature responses across models, as well as a comparison to the warmth projected for the end of this century. We use simulations from a fully coupled, global atmosphere–land surface–ocean–sea ice general circulation model: the Community Climate System Model, v. 3 (CCSM3). Future climate predictions from this model are presented in the IPCC AR4 report [1]. The model has also been used in the Palaeoclimate Modelling Intercomparison Project (PMIP) to simulate Last Glacial Maximum and Mid-Holocene climates [25,26]. The CCSM3 was developed by the US modelling community and is maintained at the National Center for Atmospheric Research (NCAR). We use the T85×1 version of CCSM3 [27] with no flux adjustments. The atmosphere model CAM3 is a three-dimensional primitive equation model solved with the spectral method in the horizontal and with 26 hybrid coordinate levels in the vertical [28]. The T85 spectral resolution corresponds to an equivalent grid spacing of approximately 1.4° in latitude and longitude. The land model CLM3 uses the same grid as the atmospheric model and includes specified but multiple land covers and plant functional types within a grid cell [29]. The ocean model is the NCAR implementation of the Parallel Ocean Program (POP), a three-dimensional primitive equation model in vertical z-coordinate [30]. The ×1 ocean grid has 320×384 points with poles located in Greenland and Antarctica and 40 levels extending to 5.5 km depth. The ocean horizontal resolution corresponds to a nominal grid spacing of approximately 1° in latitude and longitude with greater resolution in the tropics and North Atlantic. The sea ice model uses the same horizontal grid as the ocean model. It is a dynamic–thermodynamic formulation, which includes a subgrid-scale ice thickness distribution and elastic–viscous–plastic rheology [31]. Coupled pre-PMIP3 climate simulations have been completed with the CCSM3 for pre-industrial (PI) conditions and for 125 ka. For the results in this paper, the CCSM3 statistics are calculated from the last 30 years of a 950-year PI simulation (started from an AD 1990 control simulation) and the last 30 years of a 350-year 125 ka simulation (started from a previous LIG simulation with an earlier version of the model CCSM2). Differences between the LIG and PI simulations are evaluated with the Student t-test. The CCSM3 125 ka and PI simulations were run sufficiently long to minimize surface trends. Small trends, less than 0.1°C per century, are still present in the Southern Ocean. The deep ocean is still not in equilibrium with the temperatures at 2.6 km depth, cooling approximately 0.1°C per century in both the 125 ka and PI simulations. For the 125 ka and PI simulations, we assume present-day geography, Greenland and Antarctic ice sheets and vegetation (table 1). The greenhouse gas concentrations for 125 ka, carbon dioxide (CO2) and methane (CH4), are estimated from ice core measurements [32]. Those for PI are set appropriate for AD 1870 (table 1). The greenhouse gas changes result in a radiative forcing (defined as 125 ka versus PI, and calculated using formulae from the 2001 IPCC report [33]) of −0.36 W m−2. The solar constant in the 125 ka simulation was set to 1367 W m−2, the value used in the CCSM3 present-day simulation to allow comparison to previous simulations done with the earlier CCSM2 model [16]. The solar constant in the PI simulation, on the other hand, was set to 1365 W m−2, as in the CCSM3 PMIP2 simulations. The net effect of the differences in the solar constant and greenhouse gas concentrations results in a net radiative forcing of −0.02 W m−2.
The Earth's orbital configuration during the LIG was different from what it is today and this constitutes the dominant forcing change for 125 ka as compared to modern. These orbital changes are well understood and can be calculated from astronomical equations [3]. The obliquity (tilt of the Earth's axis) with an approximately 41 kyr quasi-periodicity was larger at 125 ka (23.80°) than today (23.44° for AD 1990). Obliquity modulates solar insolation at high latitudes of both hemispheres. Eccentricity, with dominant periodicities of approximately 100 and 400 kyr, was also much larger during the LIG (0.040 at 125 ka as compared to 0.0167 for AD 1990). It serves to modulate the 20 kyr precessional cycle. Perihelion (the closest distance of the Earth to the Sun) took place in late July (boreal summer) at 125 ka but occurs in early January (boreal winter) in AD 1990. The orbital forcing modifies the incoming solar insolation at the top of the atmosphere (figure 1). The annual changes at 125 ka as compared to today are small, less than a few W m−2. Seasonal changes, on the other hand, are large, with anomalies as big as ±10–15% of average insolation. At 125 ka, June insolation increases by more than 55 W m−2 at high northern latitudes and mean May–June–July insolation anomalies at 65°N are approximately 20% greater than in the Early Holocene [23]. This increase is partially compensated during the boreal winter such that annual changes are less than 3 W m−2 at these latitudes. Summer insolation in the Southern Hemisphere (SH) was reduced relative to PI.  Figure 1. Latitude–month insolation anomalies (W m−2) for (a) 130, (b) 125 and (c) 120 ka as compared to PI, assuming a fixed-day calendar.
We compare the climate model results to two recent LIG data syntheses [7,8]. Both compilations are based on published records with quantitative estimates of mean annual surface temperature change. In the marine-only dataset of McKay et al. [8], seasonal anomalies are also available with the overall pattern similar to the annual anomalies. The only available synthesis over land is for annual anomalies [7]. The Turney & Jones [7] global dataset is made up of 263 published records that span the LIG and have quantitative estimates of mean annual surface temperature. Three ice core temperature estimates from δ18O are included from Greenland and four from East Antarctica. Marine records of mean annual sea surface temperature (SST) include those obtained from foraminifera, radiolarian and diatom transfer functions and calibrations using Mg/Ca and Sr/Ca ratios and alkenone unsaturation indices (i.e. The McKay et al. [8] global dataset is a compilation of 76 records from published palaeoceanographic sites. The annual SST estimates are from Mg/Ca in foraminifera, alkenone unsaturation ratios The combined datasets, which include some overlap over the oceans, give a broadly consistent global synthesis of global MAT change (figure 2). Annual surface temperatures were warmer than modern at mid- and high latitudes of both hemispheres. The data indicate strong warming in the northern and southern polar regions, generally greater than 4–5°C north of 60°N over land and ocean, and 2–5°C for the Antarctic ice cores. The very large warming over northern Asia and Alaska is based on pollen and plant macrofossils and may reflect a bias towards summer warming [23]. A less consistent picture emerges for temperature changes in the tropics. Coastal upwelling regions, particularly along the California coast and the African coast south of the equator, indicate warming. The rest of the tropical Atlantic Ocean was cooler during the LIG, whereas the eastern tropical Pacific Ocean has adjacent cores suggesting warming and cooling. The data coverage is good for the Atlantic Ocean and eastern Pacific basin, but poor over the rest of the Pacific Ocean, the Indian and Southern Oceans and in the continental interiors. Analysis of McKay et al. suggests a peak LIG global annual SST warming of 0.7±0.6°C as compared to the Late Holocene. Turney & Jones suggest a peak LIG annual surface temperature warming (land + ocean) of 1.5±0.1°C compared to an AD 1961–1990 baseline or approximately 1.9°C warmer than PI [35]. Figure 2. Reconstructed mean annual surface temperature (MAT) change for LIG from modern as reconstructed by Turney & Jones [7] and McKay et al. [8]. See text for description of methods.
The simulated MAT change at 125 ka relative to PI shows significant warming at high and mid-latitudes of the NH (figure 3). This warming is greatest over the North Atlantic south of Greenland (in excess of 4°C) and in agreement with proxy estimates of SST anomalies ranging from 2.7 to 5°C. The model also simulates warming farther north in the Greenland–Iceland–Norwegian Seas, though a comparison with the data is less straightforward because of wide disparity among the proxy estimates of SST change. Warming over North America and Eurasia is greatest in the western portions of these continents and decreases eastward to slight warming or even slight cooling. The model simulates warmer SSTs off the coasts of California and Spain, in good agreement with the data. Figure 3. CCSM3-simulated mean annual surface temperature change for 125 ka minus PI overlain by reconstructed MAT changes. The surface temperatures are limited to the minimum ocean freezing temperature of −1.8°C over the ocean and sea ice covered regions to compare with the SST proxy data. White regions indicate sea ice in both the 125 ka and PI simulations. Differences significant at less than 95% using the Student t-test are stippled.
The model underestimates the proxy indications of warming for the far northern coastal regions of Alaska (simulated: 0.5–1°C versus data: 5.5–6.7°C) and Siberia (simulated: 1–2°C versus data: 7.1–14.8°C). A possible reason is that the models do not include vegetation feedbacks [16, supplementary material, 36], which could be important in these high-latitude regions (see §5). The simulated surface temperature anomalies at the summit of Greenland are 2.2°C as compared to NGRIP and GRIP stable water isotope estimates of annual warming of 5°C [37,38]. Some of the observed warming may be associated with reduced elevation of the central Greenland ice sheet [9,39]. CCSM3 also simulates mean annual warming over the SH subtropical landmasses of South America, Africa and Australia, but lack of terrestrial proxies in the data syntheses for these regions does not allow evaluation of the model. Slightly cooler MATs are simulated for Antarctica in contrast to East Antarctic ice core records that indicate substantial warming of 1.5–4.5°C [10,32,40,41]. Simulated cooler MATs over North Africa and southern Asia are consistent with the simulated enhanced summer monsoons. Previous modelling studies [42] and palaeorecords [43–45] also confirm the relationship between increased seasonality of the insolation in the NH and expansion/intensification of these monsoon systems. Simulated SST changes south of 30°N are small, generally within ±1°C of the PI control simulation. The data, on the other hand, show more regionally variable anomalies. Coastal upwelling regions record much warmer SSTs not simulated by the model. Coastal upwelling regions are difficult for climate models to resolve and simulate well [46]. The reconstructions show SST cooling in excess of 2°C over the tropical Atlantic and Indian Oceans, though with limited data coverage in the latter. Simulated changes in these ocean basins are of the correct sign but underestimate the magnitude of the observed cooling. For the Southern Ocean, the data suggest larger SST anomalies than the model, although with significant regional heterogeneity. CCSM3 shows strong polar amplification of MAT north of approximately 45°N with approximately similar warming over the ice-free oceans and continents (figure 4). This is associated with a seasonal memory of sea ice retreat in CCSM3 that extends the effects of positive summer insolation anomalies on the high-latitude oceans to winter months, affecting the North Atlantic, Arctic Ocean and adjacent continents. Over mid-latitude continental regions, the strong summer warming dominates the annual average. South of 45°N, CCSM3 simulates zonally averaged MAT close to zero or slightly negative, with stronger continental cooling at subtropical northern latitudes associated with the African and Asian monsoons (figure 4). Figure 4. Zonal-average plots for CCSM3-simulated annual surface temperature changes over the oceans (SST), land (TS, surface temperature) and land plus ocean (MAT) for 125 ka minus PI.
While the proxy data suggest significant regional heterogeneity in the MAT change in the tropics and SH extratropics, the simulated response in CCSM3 is relatively uniform with no change to weak annual surface cooling at the locations of the data (figure 5). The proxy data suggest an average annual warming in the tropics (30° N–30° S) of approximately 0.3°C and at SH extratropical latitudes (30–90° S) of approximately 1.2°C, while CCSM3 simulates weak annual cooling of approximately 0.3–0.5°C in both regions (table 2). The proxy data also indicate that the average annual warming over the continents was about twice that over the oceans in the tropics, a feature that CCSM3 does not capture (table 2).
Figure 5. Scatter plots for CCSM3-simulated versus reconstructed MAT change for 125 ka minus PI for (a) SH extratropics, (b) tropics and (c) NH extratropics. Blue dots denote ocean points and green dots denote land points.
CCSM3 warms the NH extratropical (30–90°N) continents more than oceans, with simulated warming at the palaeo-data locations approximately 1.0°C over the continents and approximately 0.5°C over the oceans (table 2), though with much more regional heterogeneity in the simulated SST anomalies than terrestrial MAT anomalies (figure 5). The proxy data, on the other hand, suggest similar and larger annual warming of approximately 1.7°C for the terrestrial and ocean proxies. The full-grid averages from CCSM3 also show approximately 0.5°C more warming over the NH extratropical continents than oceans, suggesting that the lack of contrast in the data is not a reflection of the uneven data distribution. Globally, a simple point-by-point average of the combined proxy reconstructions indicates an annual warming of 0.98°C, with the continents warming by 1.67°C and the oceans warming by 0.76°C. CCSM3 simulates no change of mean annual surface temperatures at 125 ka as compared to PI (table 2). Simulated terrestrial temperatures are globally warmer by 0.92°C when sampled at the proxy locations but show no change when averaged over all model land points. The terrestrial reconstructions are strongly biased towards middle and high latitudes, with no data included in these reconstructions in the monsoon regions in which CCSM3 simulates cooling. These results suggest that the spatial sampling of the reconstructions could introduce a strong bias in our perception of the LIG: the model can simulate warming at terrestrial proxy sites, on average, and at the same time simulate no average change in the ‘true’ model global temperatures. Uncertainties exist in the reconstructions in the seasonality of biological proxies, which may be biased systematically towards specific seasons. Terrestrial quantitative reconstructions are often derived from sediments indicating a change in the geographical ranges of specific plants. The biotic dominance of high-latitude vegetation is influenced by temperature, growing season length and moisture availability [47]. Seasonality is well constrained for Europe but less so for other regions [15,23]. The primary SST proxies are also sensitive to changes in seasonality [48,49]. SSTs derived from foraminifera Mg/Ca are known to reflect the calcification temperature of the species, which is best represented by the warm season conditions [50]. Environmental preferences of alkenone-producing algae may bias their proxy data estimates towards warmer temperatures, particularly in regions affected by both upwelling and open ocean conditions [51]. Locally, their signal can also be strongly seasonal [52,53]. The simulated surface temperatures for 125 ka show a strong seasonality of the surface temperature response (figure 6) consistent with the solar insolation anomalies (figure 1). The boreal summer (JJA) warming is larger and more extensive over the NH landmasses than in the annual mean. Poleward of approximately 30° N, surface temperature anomalies are greater than 2°C over almost all of North America, Eurasia and Greenland, and exceed 5°C over the interiors of Europe, Asia and North America. In Siberia along the Arctic coast, simulated JJA surface temperature anomalies are 2–3°C greater than the annual anomalies but only 0.5 to 1°C greater in northern Alaska. In both regions, the JJA simulation shows better correspondence with the MAT reconstruction but still underestimates the proxy indications of very strong LIG MAT warming (figure 7). The SH continental regions located at tropical and subtropical latitudes also show enhanced warming in JJA. This is not unexpected as the positive JJA insolation anomalies extend into the tropics and SH. Antarctica shows warming in JJA in CCSM3. The simulated JJA surface temperature anomalies compare much better to the calibrated MAT data estimates over the NH extratropics than do the simulated MAT anomalies though CCSM3 cannot capture the large warming indicated by some terrestrial proxies (figure 7). Over the SH extratropics and tropics, the simulated JJA surface temperature changes are only modestly warmer than the simulated MAT changes. Figure 6. As in figure 3 for CCSM3-simulated (a) DJF and (b) JJA surface temperature changes for 125 ka minus PI. White regions and shading as in figure 3. Seasonal averages use a fixed-day calendar.
Figure 7. Scatter plots for CCSM3-simulated annual (ANN), JJA and DJF surface temperature change for 125 ka minus PI versus reconstructed MAT for NH extratropics. Seasonal averages use a fixed-day calendar.
The austral summer (DJF) simulated surface temperature anomalies are cooler in many regions, with less correspondence with the MAT-calibrated data (figures 6 and 7). Only over high northern latitudes are there substantial warm anomalies. These warm anomalies occur over the North Atlantic and to a lesser extent over Canada and Europe. They cannot be explained by local insolation; the DJF insolation anomalies at 125 ka as compared to PI are negative at these latitudes. Rather they reflect the seasonal memory of surface temperature to the reduced sea ice and snow extent associated with the boreal summer insolation anomalies. Over northern Eurasia the simulated warming (though not significant) agrees with the expectation from proxy records of warmer winters at LIG than PI [12] but is underestimated in terms of magnitude or eastward extent. The colder SSTs simulated in the subtropical Indian Ocean in austral summer agree better with the proxy records. The lack of any improved agreement between the simulations and the data in upwelling regions suggests that seasonality is not an important consideration for the data–model mismatch. Because of difficulties in absolute dating for the LIG, both data syntheses assess maximum warmth (135–118 ka) and assumed this warmth was broadly synchronous in time. The seasonal and latitudinal natures of the insolation anomalies, though, change throughout the LIG. The obliquity reached a maximum of 24.24° at 130 ka, remained relatively high at 23.80° at 125 ka and decreased to 23.01° at 120 ka. Obliquity affects both seasonal contrasts and annual solar insolation, with the largest anomalies at high latitudes of both hemispheres. Annual insolation anomalies as compared to present reach close to 6 W m−2 at the poles for 130 ka, remain positive but decrease to approximately 3 W m−2 for 125 ka and become a negative forcing annually at high latitudes for 120 ka. Perihelion shifted from early May at 130 ka to late July at 125 ka to mid-September at 120 ka, with important impacts on NH insolation. Seasonal changes in insolation associated with precessional forcing are strong during the entire LIG because of the large eccentricity throughout the LIG. The orbital changes result in seasonal and latitudinal changes in insolation that evolve over the LIG (figure 1). At 130 ka, the largest positive insolation anomalies in the NH occur in May and June and exceed 50 W m−2 from the North Pole to NH subtropical latitudes. By 120 ka, the positive NH insolation anomalies have shifted to August and September and are much weaker. Greenhouse gas concentrations used in the CCSM3 simulations are different for the three LIG time periods, contributing an additional radiative forcing of about 0.6 W m−2 at 130 ka as compared to that at 125 ka, or potentially approximately 0.4°C of global annual mean warming. It should be noted that the greenhouse gas concentrations used for 130 ka are higher than the Dome C reconstruction followed in the PMIP3 LIG protocols [54]. Rather the CCSM3 130 ka simulation is more appropriate for 128 ka, which had June insolation anomalies similar to 130 ka and atmospheric CO2 and CH4 measurements suggesting concentrations of 287 ppmv and 724 ppbv, respectively. Differences in the radiative forcings associated with modest changes in the greenhouse gas concentrations between 125 and 120 ka are small. Globally CCSM3 is warmer at 130 ka than at 125 ka (tables 2 and 3). CCSM3 exhibits larger warming at 130 ka than at 125 ka over the North Atlantic Ocean, Ellesmere Island and Greenland (figures 3 and 8). The summit of Greenland warms by 3 to 3.5°C at 130 ka as compared to PI, but still less than the 5°C warming indicated by the GRIP and NGRIP ice cores. The simulated greater warming at 130 ka than at 125 ka for the eastern Canadian Arctic is consistent with evidence that the transition into the LIG was rapid, with peak summer warmth early in the LIG [23]. Warming extends farther eastward over the mid- and high-latitude continental regions of North America and Eurasia at 130 ka than at 125 ka, improving but still underestimating the proxy evidence of surface temperature changes. Overall, the NH extratropics warm by 1.04°C at 130 ka (as compared to 0.76°C at 125 ka) in CCSM3 when sampled at the proxy locations and 0.71°C at 130 ka (as compared to 0.27°C at 125 ka) when averaged for the full grid (tables 2 and 3).
Figure 8. As in figure 3 for CCSM3 (a) 130 and (b) 120 ka simulations as compared to PI simulation. White regions and shading as in figure 3.
The Antarctic surface temperature anomalies change from a slight cooling at 125 ka to a slight warming at 130 ka but still cannot explain the observed warming of the ice cores. Early LIG warmth in the Southern Ocean and Antarctic is consistent with East Antarctic ice core and Southern Ocean marine records [10,55]. SH extratropical and tropical simulated cooling is significantly reduced in the 130 ka simulation as compared to the 125 ka simulation, but both regions remain cooler than PI. Late in the interglaciation at 120 ka, the solar insolation anomalies associated with orbital changes result in simulated MAT similar to PI, with anomalies generally less than ±1°C and not significantly different from the PI control. Globally, the MAT change simulated by CCSM3 is +0.10°C at 130 ka with significant warming at mid- and high latitudes in the NH and some warming over Antarctica, −0.16°C at 125 ka with significant warming at mid- and high latitudes in the NH but strong cooling in NH monsoon regions and −0.14°C at 120 ka with little significant temperature change anywhere as compared to PI. A possible explanation of model–data mismatches in some regions is that not all appropriate changes in the boundary conditions have been considered in the design of the experiments. The LIG simulations with CCSM3 assumed modern vegetation. Northern Alaska and northern Siberia are covered with tundra in these simulations. The pollen and macrofossil evidence for the LIG indicates boreal forests extending to the Arctic Ocean coastline everywhere except in Alaska and central Canada [56,57]. Boreal forests have a lower albedo than tundra and can also partially mask snow allowing for more absorption of incoming solar radiation and warming. Deciduous broad-leaf forests can also contribute to greenhouse warming owing to enhanced transpiration of water vapour to the atmosphere [58]. Some of the underestimation of proxy-inferred warming may therefore be a consequence of not including vegetation feedbacks. Previous modelling of the LIG has shown that the feedback between vegetation and climate can enhance the warming at high latitudes [36,59]. Global sea level was likely 4–9 m higher during the LIG relative to present day [16,18,60,61]. A sea-level rise of up to 4 m during this time interval has been attributed to some Greenland ice sheet and other Arctic ice fields melting in combination with ocean thermal expansion [8,16]. A Greenland ice sheet contribution above 4 m is discredited by evidence of the presence of ice in the summit cores dating to before the LIG as well as Greenland ice sheet simulations that show if the ice sheet is completely removed, warming is an additional 10°C in disagreement with Greenland ice records [16]. Any sea-level rise above approximately 4 m during the LIG must then have come from the WAIS [18] and possibly the East Antarctic Ice Sheet. WAIS is largely grounded below sea level. The WAIS plays a role in buttressing the ice sheet and is particularly sensitive to ocean temperatures and circulation. Large WAIS glaciers near the grounding line show large basal melt rates as ocean waters bathing the edges of these glaciers warm [62]. Ice sheet/ice shelf models for Antarctica demonstrate brief but dramatic interglacial retreats of the WAIS, taking one to a few thousand years, for sub-ice melting of 2 m yr−1 under the shelf interior [63]. Benthic foraminifera data from cores in the North Atlantic and Southern Ocean and climate simulations indicate that warm SST in the North Atlantic could have been transported to circumpolar deep water around Antarctica [64]. Direct geological evidence for a retreat of this ice sheet though remains equivocal. A possible explanation then for the LIG simulation mismatches as compared to the Antarctic ice cores is the assumption of the presence of the WAIS in the model design. To test this sensitivity, we removed the WAIS in the CCSM3 130 ka simulation replacing it with ocean of depths up to 2000 m. Additional significant warming is restricted to Antarctica in this simulation. CCSM3 warms the region of the WAIS by more than 10°C (figure 9). It also enhances the MAT warming over East Antarctica but only by a few tenths °C, still greatly underestimating the warming indicated by the ice cores. Simulations with HadCM3 [17] suggest that in addition to the WAIS retreat, freshwater input to the North Atlantic from the Laurentide and Eurasian ice sheets during the termination is needed to produce MAT warming that is consistent with the ice core data. This is consistent with evidence of persistent iceberg melting at high northern latitudes at the beginning of the LIG [55]. Figure 9. As in figure 3 for 130 ka with the removal of the WAIS (replaced by ocean with depths up to 2000 m). White regions and shading as in figure 3.
Simulations for the LIG are now part of PMIP3 [54] (https://pmip3.lsce.ipsl.fr/wiki/doku.php/pmip3:design:li:final). A simulation for 125 ka has been completed with HadCM3 allowing us to compare the response between CCSM3 and HadCM3. The HadCM3 model was developed at the Hadley Centre for Climate Prediction and Research [65] and, similar to CCSM3, it does not include flux corrections [66]. The atmosphere and ocean models are grid-point, primitive equation models. The resolutions of the atmospheric and land components are 3.75° in longitude by 2.5° in latitude, with 19 vertical levels in the atmosphere. The resolution of the ocean model is 1.25° by 1.25° with 20 levels in the vertical. The land surface scheme is MOSES 2.1, and the sea ice model uses a simple thermodynamic scheme and contains parametrizations of ice concentration [67] and ice drift and leads [68]. The CCSM3 and HadCM3 simulations were performed independently by each group so the experimental designs are similar but not identical, as they will be in the more formal model intercomparison project. For the 125 ka and PI simulations, both models assume 125 ka orbital forcing and present-day geography, Greenland and Antarctic ice sheets and vegetation. The solar constant and greenhouse gas concentrations are slightly different for each model but result in very similar net radiative forcing (−0.02 W m−2 for CCSM3 and +0.01 W m−2 for HadCM3). The climate sensitivities for a doubling of CO2 are roughly comparable, being 2.7°C for CCSM3 and 3.2°C for HadCM3. The HadCM3 statistics are calculated from the last 50 years of a 550-year 125 ka simulation, itself started from a PI simulation of over 1000 years in length. HadCM3 was run sufficiently long to eliminate significant surface trends though trends in the deep ocean are still present. Globally, HadCM3 simulates a warmer mean annual climate than CCSM3, with global MAT change as compared to the PI of +0.14°C for HadCM3 and −0.16°C for CCSM3, as calculated on the model grids (tables 2 and 4). Calculated only at the data locations, HadCM3 simulates a global warming for 125 ka of 0.27°C versus 0.10°C for CCSM3. The annual warming in both models is significantly less than the value of 0.98°C calculated from the combined reconstructions.
Robust features simulated by both of the models are the warming over central North America, Europe, South Africa and Australia, and the cooling extending from North Africa to India and Southeast Asia (figures 3 and 10). The models disagree in their responses over Greenland and the North Atlantic, with the HadCM3 model simulating little MAT change and CCSM3 better reproducing the large warming indicated by the reconstructions. Neither model can simulate the strong warming suggested by terrestrial data along the coastal Arctic Ocean. For the SH extratropics, HadCM3 simulates MAT approximately 0.5–1°C warmer than CCSM3, with the warming over East Antarctica simulated by HadCM3 agreeing in sign with the ice core data. However, HadCM3 cannot reproduce the 1.5–4.5°C MAT warming reconstructed from the East Antarctic ice cores. In the tropics, both models simulate only small MAT changes and neither can reproduce the much warmer SSTs in the coastal upwelling regions. Figure 10. HadCM3-simulated mean annual surface temperature change for 125 ka minus PI overlain by reconstructed MAT changes. White regions indicate sea ice in both the 125 ka and PI simulations. Differences significant at less than 95% using the Student t-test are stippled.
CCSM3 and HadCM3 differ markedly in the magnitude and even sign of polar amplification of MAT. CCSM3 shows strong polar amplification of MAT north of approximately 45°N over the oceans and continents (figure 4). South of 45° N, CCSM3 simulates zonally averaged MAT close to zero or slightly negative. HadCM3 exhibits much reduced polar amplification of the MAT and a more symmetric response with small warming at both poles. Differences between the two models may be related to their sea ice sensitivities (figures 3 and 10). Both models simulate strong warming at mid- to high latitudes of the NH in JJA (not shown) consistent with the positive insolation anomalies but only CCSM3 retains memory of this warming in the North Atlantic and adjacent land regions. HadCM3 with less winter (DJF) sea ice south of Greenland at PI shows less sensitivity. In the SH, HadCM3 has less sea ice off East Antarctica than CCSM3 (white areas in figures 3 and 10) allowing the nearby Southern Ocean and East Antarctica to warm in response to the positive anomalies in austral winter and spring. These results are consistent with the contrasting Arctic and Antarctic sea ice extent sensitivities in the late twentieth century simulations in these two models [69]. CCSM3 simulations for the LIG (approx. 130–116 ka) are compared to the two recent data syntheses of Turney & Jones [7] and McKay et al. [8]. The dominant forcing change for the LIG is the large seasonal changes in the incoming solar radiation associated with orbital forcing. The model responds with substantial annual warming at 125 ka for high and mid-latitudes of the NH, although underestimates the proxy indications of this warming. Over Antarctica, the model simulates cooling as compared to large LIG warming recorded by the Antarctic ice cores. CCSM3 simulates an enhanced seasonal cycle over the high-latitude continents of both hemispheres with simulated JJA warming in better correspondence with the reconstructed annual temperature changes. The observed warming was not likely globally synchronous, with the Antarctic ice cores and records from the Southern Ocean indicating early LIG warmth [10,55]. CCSM3 shows better agreement with the SH records when forced with 130 ka insolation anomalies than the simulation for 125 ka, though the simulated warming is small. Further assessment of the seasonality of the proxy records and improvements in dating of proxy records, when possible, will be important for further assessing how well models can simulate the feedbacks in response to orbital forcing. Some of the model–data discrepancy may be associated with the experimental design, with CCSM3 adopting present-day vegetation and polar ice sheets, boundary conditions that data and previous sensitivity simulations indicate may explain a portion of the data–model differences. A CCSM3 sensitivity simulation with the removal of the WAIS provides additional local warming over Antarctica but still not enough to explain the ice core records. Simulations with HadCM3 [17] suggest that in addition to the WAIS retreat, freshwater input to the North Atlantic from the Laurentide and Eurasian ice sheets during the termination is needed to produce Antarctic warming that is consistent with the ice core data. A bipolar response is consistent with evidence of persistent iceberg melting at high northern latitudes at the beginning of the LIG [55]. The main series of simulations presented in this paper are from one model, CCSM3. A similar experiment for 125 ka has been completed with HadCM3. This allows an assessment of robustness of the CCSM3 simulation. Both model simulations suggest little global annual surface temperature change at 125 ka as compared to PI, when averages are computed for the full model grids. On the other hand, both show small global warming when averaged over the data locations. Neither simulates a global warming of approximately 1°C suggested by a simple averaging of the proxy data, although this proxy estimate may be influenced by the lack of good spatial coverage of the data. CCSM3 and HadCM3 differ markedly in the magnitude and even sign of polar amplification of MAT, a feature influenced by their sea ice sensitivities. CCSM3 with a strong sensitivity of NH sea ice [69] simulates strong Arctic warming. HadCM3 exhibits much reduced polar amplification of the Arctic MAT and a more symmetric response with small warming at both poles. The similarities and differences in the responses in CCSM3 and HadCM3 point to the need for the LIG model intercomparison project now implemented within PMIP3 [54]. It is interesting to consider the polar warmth indicated by Turney & Jones [7] and McKay et al. [8] reconstructions in comparison to future projection simulations completed by CCSM3 and included in the IPCC AR4 WG1 [70]. The global surface annual warmings projected by CCSM3 for the first few decades of the twenty-first century for the different SRES scenarios track each other closely. The low estimate climate change scenario (SRES B1) peaks in emissions in the mid-twenty-first century with greenhouse gas concentrations starting to level off in the second half of this century. By the end of the twenty-first century, CCSM3 projects a global mean annual warming of 0.9°C for the SRES B1 scenario [70], comparable to that of the LIG reconstructions, with greater warming at high latitudes than at low latitudes (figure 11). East Antarctica and much of the North Atlantic have warmed up to surface temperatures comparable to that of the LIG proxy records. The SRES B1 warming over Greenland is less than the LIG warmth, though, as for the CCSM3 LIG simulations, the CCSM3 future projections fix the Greenland ice sheet heights and extent at present day. Differences in the primary forcings, orbital insolation changes for the LIG versus greenhouse gas concentrations in the SRES projections, give different seasonal responses. While the 125 ka CCSM3 simulation shows the largest warming in JJA, the SRES B1 future projection simulation has the greatest high-latitude warming during the respective winters of the two hemispheres. As such, it would be wrong to consider the LIG as an exact ‘analogue’ for future climate change. Figure 11. CCSM3-simulated mean (a) annual, (b) December–January–February, and (c) June–July–August surface temperature change for 2080–2099 minus 1980–1999 in the SRES B1 low estimate climate change scenario. The annual surface changes in (a) are overlain with the reconstructed LIG MAT changes. White regions indicate sea ice in both the simulated results for the last two decades of the twentieth and twenty-first centuries.
In summary, there is no clear answer to the question posed in the title ‘How warm was the Last Interglacial?’ Our results show model–data inconsistencies that are not fully understood. The models' sensitivity to the forcings may be too small. Our interpretation of the reconstructed climate parameters may not adequately incorporate seasonal and depth effects and age uncertainties. The implications for future warming scenarios require progress in resolving these inconsistencies between the model simulations and data reconstructions of past interglacials. This research was enabled by the computing resources of the Climate Simulation Laboratory. The CCSM project is supported by the US National Science Foundation (NSF) and the Office of Science (BER) of the US Department of Energy. NCAR is sponsored by NSF. E.J.S. is funded by the EU project Past4Future. This is Past4Future contribution no. 33. The research leading to these results has received funding from the European Union's Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 243908, ‘Past4Future. Climate change: learning from the past climate’. FootnotesOne contribution of 11 to a Discussion Meeting Issue ‘Warm climates of the past—a lesson for the future?’. © 2013 The Author(s) Published by the Royal Society. All rights reserved. References
Page 2Simulating warm climates of the Earth's past has been a challenge for the climate modelling community for decades [1,2]. Using proxy estimates of atmospheric greenhouse gas concentrations in fully coupled climate models yields simulated polar temperatures that are often too cold, see, for example, Lunt et al. [3] for a multi-model comparison with proxy data. An additional challenge is to warm the polar regions without excessively warming the equatorial region [2], a problem that exists in spite of polar feedback processes operating in the climate system. Simulating past warm climates with a fully coupled general circulation model (GCM) is of great importance given the current and projected rise in atmospheric carbon dioxide (CO2). Projections indicate that if humans continue to burn fossil fuels at the current rate, then atmospheric CO2 levels will reach 800–1100 ppmv by the year 2100 [4,5]. It has been tens of millions of years since these concentrations of CO2 have existed in the Earth's atmosphere. There are certainly clear differences between the Palaeocene–Eocene Thermal Maximum (PETM) climate state and the present and projected near-future climate warming. Despite these significant differences, there is still a need to better understand how the Earth's climate processes function in differing climate regimes. Thus, studying the Earth's warm past climates such as the PETM provides rich observational and modelling opportunities to better understand how the Earth operates in a warm climate regime [6,7]. Over the years, many physical mechanisms have been proposed to solve the low Equator-to-Pole thermal gradient problem in climate models. The basic challenge has been to find means of warming the polar regions more than warming the tropics under Eocene conditions (i.e. in the absence of strong snow and sea-ice and terrestrial ice feedbacks), since solely increasing greenhouse gases significantly warms both equatorial and the polar regions (e.g. [1])—. Past proposed climate mechanisms include: increased ocean heat transport [8], polar stratospheric clouds related to enhanced atmospheric methane (CH4) [9], increased deep cloud convection at high latitudes with associated longwave cloud radiative forcing [10] and opening passageways in the Arctic [11], to name a few. It has also been argued that tropical temperatures may have been higher than previously considered, which would allow for a purely enhanced greenhouse gas explanation for warmer climates [2,12,13]. This study explores the role of another mechanism that may have operated in the deep past. It has been pointed out that aerosol properties were no doubt significantly different in deep time [14]. Specifically, Kump & Pollard [14] considered the role of reduced cloud condensation nuclei (CCN) for the warm equable climate of the Cretaceous. They found that lower levels of CCN led to a considerable warming of polar regions relative to warming in the tropics owing to associated changes in cloud properties. This study presents results from a coupled climate model that explores the possible roles of enhanced greenhouse gas concentrations and sensitivity to reduced CCN for pre-PETM and PETM climates. The exact properties of aerosols in the Eocene relative to the present are unknown. However, given that the climate of the Eocene was very different from that of today, it is probable that aerosol properties were different, since aerosol properties are tied to phenomena such as: vegetation type and distribution (organic aerosols, biomass burning), desert regions (dust aerosols), surface wind patterns (sea-salt aerosols) and the ocean productivity (emissions of dimethylsulfide). Thus, the motivation here is to perform a sensitivity study to see whether aerosol–cloud effects could play an important role for the climate of the Eocene. Here, present-day observations from pristine regions (i.e. regions far from human pollution, but still affected by natural sources) are used to constrain cloud properties, given that currently there are no observations to accurately constrain Eocene aerosol properties. The study also explores the sensitivity of the results to the assumed drop number concentration over continents, since, even in present-day conditions, there is an observed difference in natural aerosols from marine to continental regions. Note that we are not suggesting that the Eocene had aerosol properties identical to the modern pristine values. We are using the modern pristine aerosol properties as a sensible starting point for this sensitivity study. Additionally, reduction in low-level cloud cover because of lower CCN leads to increased shortwave heating of the Earth's surface, which has important implications for estimates of global carbon cycle budgets for past warm climates. Presently, models assume high CO2 concentrations that enhance the Earth's greenhouse effect and warm the climate system. If aerosol–cloud effects were significantly different in the past, then less CO2 would be required to create a similar warm climate state. This is significant, since assumed CO2 levels for past warm climates are often higher than those estimated by carbon cycle modelling [15]. This study uses the fully coupled atmosphere–ocean–land–sea ice Community Climate System Model (CCSM) configured for the Early Eocene palaeoclimatic conditions. Simulations are presented for climate conditions representative of both pre-PETM and PETM time periods. The PETM simulation represents climatic conditions for the peak of the warming event, while we use the term pre-PETM for the simulation representing conditions prior to the warming event. The present study does not attempt to simulate the temporal transition across the event owing to computational limitations in carrying out a simulation extending for tens to hundreds of thousands of simulated years. The benefit of considering the Early Eocene compared with periods in the deeper past is that palaeoproxy data for temperatures for this time period cover a significant latitudinal range. These data provide extensive evidence for extreme warmth at high latitudes [16], with a very active hydrological cycle [15]. All of these reconstructions are signatures of a very warm climate regime due to elevated greenhouse gases. The data are also suggestive of mechanisms—positive feedbacks—that amplify the initial greenhouse radiative forcing [17,18]. The study is organized as follows: §2 describes the model configuration and experimental design of the PETM and pre-PETM simulations, §3 presents the results from these simulations and compares simulated surface temperatures with various palaeo proxy data, §4 explores the implications of the present work for understanding the Earth's sensitivity to increased levels of CO2 and §5 summarizes the findings of the present study. This work employs CCSM v. 3 (CCSM3) [19], a fully coupled climate model with active atmosphere, ocean, sea-ice and land components. Lengthy near-steady-state simulations require considerable computational resources so the low-resolution version of CCSM3 [20] is used for all simulations. The atmospheric and land components of the CCSM3 employ an Eulerian spectral dynamical core of T31 (implying an equivalent horizontal resolution of 3.75° × 3.75°) with 26 vertical levels in the atmosphere. The ocean and ice components use a nominal 3° horizontal resolution with 25 oceanic depth levels. Further modifications include a marginal sea parametrization over the Arctic Ocean basin to ensure reasonable salinity values over long equilibrium runs (see [21] for further details). The equilibrium climate sensitivity of this version of the model is 2.5°C warming for a doubling of CO2 from present-day CO2 concentrations. This sensitivity is at the lower end of the canonical range of 2.1–4.5°C [4] in climate sensitivity. Using a model with higher climate sensitivity would require lower greenhouse concentrations to arrive at a similar climate simulation. The model configuration employs recent reconstructions [21,22] of the Middle Eocene palaeogeography, palaeotopography and ocean bathymetry. Specification of the spatial distribution of vegetation follows [22]. The vegetation specification is the same for both PRE-PETM and PETM simulations, i.e. there is no vegetation feedback across the PETM event. Recent studies suggest CO2 levels for the PETM that may lie in a range from approximately 1700 to 2250 ppmv [23,24]. But it is fair to state that wide uncertainty exists in the actual CO2 concentrations during both pre-PETM and PETM climate states. Increases in atmospheric CH4 concentrations have also been proposed for the PETM because of the observed negative carbon isotope excursion in δC13 (see fig. 3 in [25])—. The assumed PETM atmospheric concentrations of CO2, CH4 and N2O are 2250 ppmv, 16 ppmv and 275 ppbv, respectively. The level of atmospheric CO2 is based on the work by Panchuk et al. [24] (L. Kump 2009, personal communication), which used a geochemical model of intermediate complexity to infer atmospheric CO2 levels consistent with geochemical markers. The level of CH4 employed for the present work comes from a modelling study on the effects of CH4 release during the PETM [26]. The pre-PETM atmospheric concentrations of CO2, CH4 and N2O are 1375 ppmv, 760 ppbv and 275 ppbv, respectively. The pre-PETM level of CO2 was obtained by taking the PETM simulation and reducing CO2 levels until the global annual mean temperature was reduced by approximately 5°C in order to agree with the global estimate of observed temperature change. Here the use of a pre-industrial level of atmospheric CH4 is no doubt low for the warm Early Eocene, given the moist environment that would have allowed for more wetland regions. Given that there are no observational data to constrain CH4 concentrations for this time period, a conservative assumption is made concerning the pre-PETM CH4 levels. Note that sustained levels of CH4 require a continual release source of CH4 into the atmosphere given the relatively short CH4 lifetime of approximately 12 years. This issue is addressed in §4. The change in CCN was incorporated into the atmospheric model by altering both the liquid cloud drop number and the effective cloud drop radii. As noted, there are no observations of CCN or cloud microphysical properties for deep time periods. Thus, the present study should be viewed as a sensitivity study with regards to the effects of cloud microphysical properties on past climates. Framed as a sensitivity study, this work will make simple assumptions about aerosol and cloud properties for the Eocene. Given this assumption, an observational composite (see fig. 5 in [27])— of cloud drop number for present-day remote pristine regions is used to set the cloud drop number in the simulations. For present-day pristine regions, the observed cloud drop number concentration is around 50 drops per cm3 for liquid water clouds. Lower CCN leads to fewer cloud drops that grow to larger sizes. Observations indicate that effective cloud drop radii for pristine clouds are approximately 17 μm. For the sensitivity studies, the model configurations assume that all liquid clouds have present-day pristine cloud drop properties. Once these properties are prescribed the cloud microphysical and radiative parametrizations in the atmospheric model respond to these cloud drop properties, i.e. cloud rainout processes and shortwave absorption change owing to the change in cloud drop properties. Decreasing cloud drop number leads to increased precipitation rate and shorter cloud lifetime, which in the time mean implies a reduction in cloud cover. Increased cloud drop size leads to more shortwave absorption in clouds, which dissipates clouds. Fewer low-level clouds results in more shortwave radiation reaching the surface. Since low-level liquid water stratiform clouds predominate at high latitudes, the reduction in these types of clouds leads to a preferential warming of polar regions. To summarize, for pre-PETM and PETM simulations, liquid water cloud properties are changed as follows: the cloud drop density is set to 50 cm−3 everywhere, as compared with the present-day prescription of 400 cm−3 for continental regions, 150 cm−3 for ocean regions and 75 cm−3 over sea-ice and snow-covered regions. The effective liquid cloud drop radius is set to 17 μm everywhere, as compared with the present-day assumed values of 8 μm over land and 14 μm over ocean, sea-ice and snow-covered regions. The role of continental versus marine cloud drop differences is explored in a companion PETM simulation in which the cloud drop density is set to 400 cm−3 and the cloud drop effective radius to 10 μm over continental regions. All simulations employ a constant uniform pre-industrial aerosol optical depth representing a general background aerosol concentration. The simulations also assume a fixed geothermal heat flux into the oceans of 0.088 W m−2. All simulations assume a 0.487% reduction in solar luminosity for the Early Eocene time period and the orbital parameters are those used in [21]. Based on these modelling assumptions, four factors account for the warm simulated climate of the PETM: enhanced CO2 concentrations, enhanced CH4 concentrations, the absence of ice sheets and a reduction in CCN, i.e. lower liquid cloud drop number and larger cloud drop effective radius. In order to assess the relative warming contribution from three of these factors—CO2, CH4 and CCN effects—sensitivity climate simulations were carried out with a version of the CCSM3 that uses a slab mixed layer ocean component in place of the full dynamical ocean. Note that all of these simulations assume the absence of terrestrial ice sheets and use the palaeogeography of the Middle Eocene. In this version of the model the ocean heat transport for the mixed layer model is based on the fully coupled PETM simulation. The advantage of the slab ocean model for sensitivity studies is that it can be run to a steady state within only 40 simulated years. In the first simulation, only CO2 levels were decreased to a pre-industrial value of 280 ppmv. In the second simulation, only CH4 levels were decreased to a pre-industrial level of 760 ppbv. In the third calculation, cloud drop number and effective drop radii were changed to present-day (i.e. polluted) values. The change in annual zonal mean surface air temperature from these three simulations (figure 1a) indicates that the largest warming effect is because of CO2 with polar warming of 12°C. The second largest warming is because of changes in CCN-induced liquid water cloud properties and yields a 7–9°C warming at the Poles, while the third largest contributor to warming arises from increased CH4 with a modest 4°C warming at the Poles. In order to eliminate the effects of a different base state for these results, the normalized change in zonal mean surface air temperature is shown (figure 1b), in which the zonal mean changes are normalized by their respective global mean changes. If similar amplification processes are present in all three simulations, then the three curves should cluster together. However, in the Northern Hemisphere polar regions, there are still significant differences between the CCN simulation and CO2 and CH4 simulations. These differences are due to the inherent differences in radiative forcing for the CCN simulation compared with the greenhouse gases simulations. The radiative forcing from the CCN sensitivity simulation arises from two factors: (i) changing the cloud drop number to 50 cm−3 lowers the liquid water path in the clouds and decreases cloud area (this effect arises from an increase in precipitation efficiency and hence a decrease in cloud lifetime) and (ii) increasing the cloud drop size to 17 μm decreases the single scattering albedo of the clouds, which leads to more shortwave absorption in the clouds. This, in turn, leads to a burn-off of low cloud cover. Both of these effects result in more shortwave radiation reaching the surface. Additional simulations have been performed to isolate these two cloud effects to see which dominates high-latitude warming. These simulations show that the change in cloud drop size—enhanced shortwave cloud absorption—dominates the CCN forcing simulations. At high latitudes, the CCN effect will play a major warming role during late spring to early autumn, i.e. when shortwave radiative forcing is high. However, at high latitudes during local winter conditions, the shortwave CCN effect will not be active. Analysis of the simulations indicates that high-latitude local winter warming is due to three factors: (i) an overall increase in tropospheric water vapour leading to an enhanced greenhouse warming, (ii) an increase in upper tropospheric cloud cover leading to an increase in longwave cloud forcing, and (iii) an open Arctic ocean basin that stores more energy through the winter than an ice-covered Arctic, which does not store energy through local winter. These results indicate that there are differences in response from the CCN effect compared with the standard greenhouse effect. This is essentially because the CCN effect is affecting low-level stratiform cloud cover, which predominates the high latitudes coupled with the seasonal asymmetry in solar radiation reaching high latitudes. These two factors lead to higher forcing at high latitudes than what is obtained from greenhouse forcing. Figure 1. Zonal mean change in: (a) PETM annual surface air temperature (°C) owing to changes in CO2 from 280 ppmv to 2250 ppmv (dashed line), changes in CH4 from 760 ppbv to 16 ppmv (dotted line) and changes in cloud properties (solid line); (b) annual surface air temperature normalized by respective global mean change in surface air temperature due to changes in CO2 from 280 ppmv to 2250 ppmv (dashed line), changes in CH4 from 760 ppbv to 16 ppmv (dotted line) and changes in cloud properties (solid line). Normalized value = ([Ts]exp− [Ts]control)|/〈Ts〉control, where [ ] indicates zonal mean and 〈 〉 indicates global mean.
The fully coupled PETM simulations involved a multi-stage spin-up process. First, an initial PETM simulation was carried out by running the fully coupled CCSM3 for 600 years from an ocean-only simulation that had been run in an accelerated mode for 8300 years. This was done in order to obtain a more representative deep ocean state. The pre-PETM fully coupled model was run from this simulation with the pre-PETM CO2 and CH4 concentrations and pristine CCN conditions. The pre-PETM state was run for another 1400 years. The ocean state for this simulation has vigorous ventilation, which means the deep ocean comes into an approximate steady state. CO2 and CH4 concentrations were then set to their PETM levels and the PETM simulation was initialized from the end of pre-PETM simulation and run for another 1660 years, after which the net energy imbalance of the coupled PETM climate system was less than 0.4 W m−2. Thus, the deepest layers of the ocean in this simulation are still not in a steady state. However, the surface temperatures at the end of this simulation have reached near-steady-state conditions with trends in zonal mean surface temperature less than 10−4 °C per year. All results are based on 50 year averages from the end of the pre-PETM and PETM CCSM3 simulations. The global annual mean sea surface temperature (SST) for the fully coupled PETM simulation is 32.3°C. To obtain an estimate of the observed PETM global mean SST, it is assumed that the zonal surface temperature can be represented by the function Zonal annual mean SSTs from the PETM and pre-PETM model simulations exhibit a reduced Equator-to-Pole temperature gradient compared with the present-day simulation (figure 2). Note that the present-day CCSM3 simulation is in good agreement with observed SSTs with a slight cold bias in the polar regions. As noted from the slab ocean sensitivity simulations, the dominant factors contributing to the amplified polar warmth are increased atmospheric CO2 concentration and the change in cloud properties associated with reduced CCN (see figure 1). Increased levels of atmospheric CH4 contribute one-third of the warming relative to warming from increased CO2. Zonal mean PETM SSTs at approximately 70° north and south are approximately 20°C, in good agreement with the proxy estimates at these latitudes (see table 1 for point-wise comparisons with proxy data). Tropical temperatures for the PETM are approximately 40°C. Currently, there are no proxy estimates of SSTs for the deep equatorial marine environment, but tropical and subtropical temperatures approaching the PETM indicate a warm climate with surface temperatures in the range 35–41°C (see figure S1 in [2]). Figure 2. Zonal annual mean SSTs (°C) from the modern (black line), PETM (red line) and pre-PETM (blue line) CCSM3 simulations.
The simulated spatial distribution of PETM tropical and sub-tropical SSTs (figure 3a) is similar to the modern-day pattern with a warm pool of water in the Indian and western Pacific Oceans and a cold tongue of water in the eastern Pacific. Warm pool temperatures in excess of 40°C exist in the palaeo Tethys region on either side of India, indicating the lack of an ocean thermostat to keep these waters close to present-day values. Warm waters extend far into the extra-tropics with 32°C water off the coast of present-day New Jersey, in good agreement with the proxy data for this region [28] (table 1). At higher latitudes, the PETM Arctic is slightly cold by approximately 8°C compared with the observational value for this region [29]. Note that simulated PETM summer temperatures are in better agreement with the observational estimate (table 1). The largest discrepancy between simulated SSTs and reconstructions occurs at 55° S latitude [30]. Here the model is colder than the proxy data, but again the simulated summer temperature is closer to the observations. Further south at 65° latitude the model agrees with the proxies to within 2°C [31]. Figure 3. Geographical distribution simulated: (a) PETM SSTs (°C) and (b) change in SST (°C) from pre-PETM to PETM climate. Numbers in boxes are observed range in temperature changes [33].
The change in SST (figure 3b) (PETM minus pre-PETM) is in good agreement with observed changes at all latitudes (see table 1 in [33] for a compilation of proxy changes in SSTs). Given that the pre-PETM model atmospheric CO2 level of 1375 ppmv was chosen to ensure an approximately 5°C global annual mean change in surface temperature (see §2), this agreement may seem unsurprising. However, the choice of this CO2 level does not guarantee that the spatial distribution of change in temperature will agree with reconstructions at specific geographical locations. The sea surface salinity distribution for the PETM simulation (figure 4a) shows extremely fresh waters in the Arctic basin, which is in agreement with recently published proxies [15]. One important regional feature of the PETM simulation is high salinity located in the Turgay strait between present-day Europe and Asia. As will be shown below, this feature increases surface water density relative to the pre-PETM simulation, causing sinking in this region. The overall salinity distribution—determined by local water balance between evaporation minus precipitation and run-off (figure 4b,c)—exhibits enhanced net fresh water input at high latitudes compared with the present-day climate and enhanced tropical precipitation. Continental run-off plays an important role in determining salinity levels in coastal regions (compare figure 4a,c), where the largest run-off occurs into the Arctic basin and in the tropical regions. Figure 4. PETM simulated (a) sea surface salinity (practical salinity units), (b) evaporation minus precipitation (mm day−1) and (c) surface run-off (mm day−1).
The location and strength of deep water formation is important for understanding the deep ocean circulation. In particular, changes in deep water formation due to increasing levels of greenhouse gases may have important implications for carbon cycle processes [34,35]. The seasonal cycle of maximum mixed layer depth (figure 5) is an informative measure of the location of deep water formation. Other measures were also used to identify regions of deep water formation occurring in the pre-PETM and PETM simulations, including the seasonal cycle of surface potential density and isopleths of zonal potential temperature and salinity. The conclusions concerning deep water formation using maximum mixed layer depth are supported by all of the above metrics. Two winter months of maximum mixed layer depth for each hemisphere are shown (figure 5), since these seasons are when the densest waters sink to maximum depth. Northern Hemisphere winter pre-PETM maximum mixing (figure 5a,b) occurs in the north Pacific, where water formed in this region penetrates to 4000 m depth. Mixing is very vigorous, as indicated by ideal water ages of less than 50 years at these depths, where ideal age measures the time in years that water at a specific depth was last exposed to the ocean surface. In the Southern Hemisphere winter pre-PETM (figure 5c,d), maximum mixing occurs off the coasts of Australia and Antarctica. Again, ocean ventilation is very efficient for this region. This efficient ventilation for the pre-PETM climate means that the deep ocean is strongly coupled to the ocean surface in high-latitude regions. It also means that the equilibrium time scale of ocean circulation for the pre-PETM ocean state is much shorter than that of the PETM world. The maximum mixed layer depths for the PETM simulation (figure 5e,h) exhibit a very different configuration for deep water formation. The high-latitude Pacific formation sites for both hemispheres no longer exist. Surface warming and fresh water input at high latitudes have significantly reduced the specific density gradients and stratification has essentially shut off high-latitude deep water pathways. However, one region of maximum mixed layer depth remains in the Turgay straight. Here, high surface salinities (figure 4a) cause surface waters to sink to approximately 1000 m depth. This water then spreads out from the Tethys region, forming intermediate waters into the wider Pacific region. Thus, the source location of deep water formation switches in going from the pre-PETM into the PETM climate state, as first suggested by the modelling results of Bice & Marotzke [36] and proxy studies [37–39]. Figure 5. Simulated maximum mixed layer depths (m) for (a) January pre-PETM, (b) February pre-PETM, (c) July pre-PETM, (d) August pre-PETM, (e) January PETM, (f) February PETM, (g) July PETM and (h) August PETM.
The effects of this shift in water mass formation is reflected in the ocean circulation at 1000 m depth between the PETM state (figure 6a) and the pre-PETM (figure 6b). These figures show a shift from a circulation with stronger and larger gyres to a weaker circulation for the warmer climate. Note that the gyre circulation is also due to a shift in the atmospheric circulation in moving into a warmer climatic state. Figure 6a also exhibits evidence for the flow of warmer water from the Turgay straight out to the Pacific in the PETM state compared with the pre-PETM. The near-surface PETM circulation (figure 7) exhibits surface flow from the Atlantic into the Pacific through the open Panama strait. The simulation also shows signatures of both Kurishio- and Gulf-like currents along the eastern boundaries of present-day Asian and North American continents, respectively. Coastal upwelling driven by along-shore atmospheric circulations is also apparent. Figure 6. Simulated 1000 m potential temperature (°C) and ocean currents (cm s−1) for (a) PETM and (b) pre-PETM simulations.
Figure 7. PETM simulated 100 m vertical velocity to denote regions of upwelling (positive) and 100 m surface ocean currents.
One proposed mechanism for maintaining a low Equator-to-Pole thermal gradient was related to an increase in ocean heat transport [8]. The present simulations find no evidence for this hypothesis (figure 8), in agreement with other modelling studies that have used fully coupled GCMs to study the warm Eocene [40]. Indeed, the warmer climate state of the PETM exhibits less Northern Hemisphere ocean heat transport than either the pre-PETM climate or the present-day simulated climate. Peak PETM ocean heat transport at 20–30° N is approximately 30% less than pre-PETM transport and approximately 45% less than present-day peak transport. Note, however, that ocean heat transport in the Southern Hemisphere is greater than the present-day simulated transport and may contribute to warmer high-latitude temperatures in this region. Figure 8. Zonal annual mean of ocean poleward heat transport (PW) for the modern (solid line), pre-PETM (dashed line) and PETM (dotted line) simulations.
Finally, the zonal annual mean vertical thermal structure of oceans for both PETM and pre-PETM simulations (figure 9a,b) is significantly different from the present-day thermal structure (figure 9c). In general, the warm Eocene simulations are more stratified than the present-day ocean structure, in agreement with previous studies (e.g. [34,40]). This stratification is more evident when considering vertical profiles of potential temperature (figure 10a–d) for tropical and northern and southern high-latitude regions. For Northern Hemisphere winter conditions (figure 10a,b), the pre-PETM simulation shows a region in the north Pacific of near-constant temperatures through the depth of the ocean column, indicative of vigorous mixing in this region. In the PETM state, this region of mixing is suppressed. Similarly, for the Southern Hemisphere winter (figure 10c,d), a region of well-mixed water exists in the southern Pacific, which has been suppressed in the PETM climate state. The PETM ocean temperatures at depths ranging from approximately 1300 to 3400 m for specific locations are generally in good agreement with proxy data (table 2). However, it must be remembered that at these depths the ocean state is still warming in the PETM simulation.
Figure 9. Zonal annual mean ocean potential temperature (°C) for simulated: (a) PETM, (b) pre-PETM and (c) present climate simulations.
Figure 10. Vertical profiles of potential temperature (°C) for December–January–February average: (a) tropical, (b) northern high latitudes, and for June–July–August average (c) tropical and (d) southern high latitudes. PETM case is solid line; pre-PETM case is dashed line.
The simulated high-latitude terrestrial annual mean surface air temperatures for the PETM (figure 11a) are 15–20°C, in reasonable agreement with the limited proxy data. For example, temperatures for the Wyoming Bighorn Basin are in excellent agreement with reconstructions (table 1) [42]. Simulated surface temperatures in the tropical regions of South America and Africa are excessively high, in excess of 48°C at some points. Such high temperatures would imply severe conditions for the existence of life in these regions. The change in surface air temperature from pre-PETM to PETM conditions (figure 11b) shows warming of 5–10°C for most regions. This 5–10°C warming in middle North America is in very good agreement with data [6]. For the northernmost region of North America, minimum January temperatures are 8°C, which is supported by the limited palaeobotanical evidence of palm trees at these high latitudes. In general, for the PETM simulation cold month mean surface air temperatures are above freezing at all locations (figure 12a), while for the pre-PETM intercontinental region cold month mean temperatures are below freezing (figure 12b). A comparison of simulated terrestrial surface temperatures with the proxy compilation of Huber & Caballero [12] (figure 13) indicates general agreement across a range of locations. Given that the reconstructions span a wide range of Early Eocene time periods, with few data representative of the PETM event, agreement is expected to be better for the pre-PETM simulation than for the PETM and indeed this is the case (compare figure 13a,b, respectively). Figure 11. Simulated annual mean: (a) PETM surface air temperature (°C) and (b) change in surface air temperature (°C) from pre-PETM to PETM climate.
Figure 12. Simulated cold month mean (°C): (a) PETM and (b) pre-PETM.
Figure 13. Comparison of simulated terrestrial surface air temperatures (°C) with the Huber & Caballero [12] proxy database for the: (a) PETM and (b) pre-PETM climates.
As noted in the Introduction, specification of a continental cloud drop size density of 50 cm−3 may be too low, given the diversity of aerosol sources from vegetation (secondary organics, biomass burning) and changes in surface conditions, e.g. dust loading. An additional sensitivity study was carried out to test how the PETM results depend on this assumption. Using a cloud drop size of 400 cm−3 and drop effective radius of 10 μm over continental regions (figure 14) leads to a 2–3°C reduction in continental surface temperature. Thus, inclusion of a difference between continental and marine CCN conditions slightly cools the continents relative to the simpler approach of uniform CCN or cloud drop properties. Figure 14. Difference in annual mean PETM surface temperature (°C) between a simulation assuming a continental cloud drop density of 400 cm−3 and drop effective radius of 10 μm and a simulation assuming a continental cloud drop density of 50 cm−3 and effective radius of 17 μm.
In general, the hydrological cycle over land is enhanced in the pre-PETM and PETM simulations compared with the present-day simulation as reconstructions suggest [43]. Run-off in the Arctic also increases from the pre-PETM to the PETM, resulting in very fresh water for the entire basin—a feature of the Arctic supported by proxy data [15]. Overall the hydrological cycle is more vigorous for the warm Eocene simulations than for the present-day simulation. The change in precipitation over land (figure 15) from the pre-PETM to the PETM climate state indicates that most regions experience an increase in rainfall. In the northern part of North America, precipitation increases by 20–50%. However, there are locations that experience a slight reduction in precipitation, e.g. the North American southwest. Reconstructions of changes in precipitation [42] in North America offer a complex picture of regions of decreased precipitation early after the event, but a general increase in precipitation at the peak of the warm PETM. The seemingly large percentage change in precipitation in the central African region is somewhat misleading given that both the pre-PETM and PETM simulated precipitation in this region is very low (less than 1 mm d−1). The percentage change in this general region of low precipitation appears to be related to the poleward shift in the zonal mean circulation, which affects moisture transport into this region. Figure 15. Per cent change in terrestrial precipitation between the PETM and pre-PETM climate simulations.
Considerable attention has recently been given to what deep time climates tell us about the Earth's climate sensitivity [5,44–46]. The traditional or Charney climate sensitivity is defined as the equilibrium warming due to a doubling of CO2 about the modern climate state and assumes that the factors amplifying the initial greenhouse radiative forcing take place on time scales of days to decades. Thus, this sensitivity includes feedback mechanisms involving changes in processes such as water vapour, lapse rate, clouds and sea ice. One measure of the strength of these feedback processes is the so-called climate feedback factor defined as the ratio of the doubled CO2 equilibrium warming (approx. 2.5°C) to the radiative forcing due to a doubling of CO2 (approx. 3.7 W m−2), i.e. approximately 0.68°C (W m−2)|−1. As discussed in §3a, the present study arrives at a proxy estimate for the global mean PETM surface temperature of approximately 33°C. Given that the Earth's pre-industrial temperature was approximately 15°C, the PETM was warmer by 18°C compared with the pre-industrial time period. Assuming that the PETM CO2 concentration (2250 ppmv) was eight times larger than pre-industrial levels implies a CO2 radiative forcing of 14.5 W m−2, where the CO2 forcing is obtained by the method described in [47] employing the atmospheric version of CCSM3 in a fixed SST configuration. The shortwave forcing from pre-industrial to the PETM time period (55 Ma), owing to the change in solar luminosity, is a forcing of −1.2 W m−2, assuming a planetary albedo of 0.27 derived from the PETM simulation. Thus, the net forcing (CO2 greenhouse+solar luminosity) from pre-industrial to the PETM is approximately 13 W m−2. Using these estimates for the PETM implies a feedback factor of approximately 18°C/(13 W m−2) or approximately 1.4°C (W m−2)|−1. Thus, the climate sensitivity considering a change in state from the modern to the PETM is two times larger than the modern Charney sensitivity. This larger climate sensitivity has been called the Earth system sensitivity (ESS) [45] and includes feedback processes operating on time scales of decades to many millennia, e.g. ice sheet destabilization, possible methane hydrate release, changes to vegetation and alterations to the global carbon cycle. Note that if a lower PETM CO2 concentration were assumed, then the deduced ESS would be even higher. What do the palaeoclimate simulations of the pre-PETM and PETM in this study have to say concerning the issue of an enhanced ESS over the Charney sensitivity? In addition to CO2 forcing, the absence of ice sheets and palaeogeography, this study considers two other factors to explain the warm simulated climates of the pre-PETM and the PETM relative to the modern climate state: (i) increased CH4 concentrations and (ii) lower CCN. This sensitivity study suggests that these two additional processes may have played a critical role in enhancing climate sensitivity on long time scales relative to the present-day climate. Note that this argument assumes that CH4 release and changes in aerosol–cloud interactions operate as feedback mechanisms and not forcing factors on these long time scales. With regards to CH4, Dickens [18] has argued that release of large reserves of CH4 is quite feasible for the warm Eocene. Continued release of CH4 from expanded wetland regions containing bacteria would also have led to sustained levels of atmospheric CH4 [48]. Sustained emissions of CH4 would be required to maintain elevated CH4 levels in the presence of an otherwise short lifetime. With regards to changes in CCN, this study uses present-day observations for very pristine regions to constrain the model cloud microphysical properties. This assumption is clearly simplistic given the temporal, spatial and chemical variability that exists in real aerosols. Thus, what is presented here is a sensitivity study, which uses present-day knowledge to link CCN and cloud drop number density. Kump & Pollard [14] provided arguments for why CCN would have been lower during the warm Cretaceous. There may be other reasons for lower CCN during past warm climate states, some of which may be linked to atmospheric chemistry. A recent study [49] found that the production of aerosols actually decreases in the presence of certain types of vegetation. This effect is actually opposite to what occurs where the production of secondary organic aerosols increases as a result of the emission of certain biogenic precursors from vegetation. The explanation of these new findings involves the effects of isoprene emissions from vegetation on the concentration of the atmospheric hydroxyl radical, which plays an important role in aerosol formation. In this case, increased warming initiated by increased greenhouse gases would lead to the migration of forests to higher latitudes [50] accompanied by a reduction in CCN with additional warming. This proposal is hypothetical, as such it would be of value to look for particular proxies that could either validate or invalidate this hypothetical biophysical feedback. Note that the conclusions presented here do not depend on this particular proposed mechanism, since this study considers the general sensitivity of the warm climate state to changes in cloud microphysical properties derived from present-day pristine conditions. To date, modelling past warm climates, such as the Eocene, has focused mainly on two climate factors: greenhouse forcing and enhanced climate sensitivity. For sufficiently high CO2 concentrations, enhanced greenhouse forcing yields a warm climate approximating reconstructions. However, the assumed level of CO2 may be too high compared with palaeo pCO2 proxies. Another solution to this dilemma is to use a model with a higher climate sensitivity, which means that a lower CO2 concentration yields similar agreement with the reconstructions. This work explores another factor that is important to the climate system, i.e. shortwave feedback. This sensitivity study shows that including changes in shortwave forcing—via CCN–cloud interactions—results in additional heating of the climate system, which, in turn, implies that less CO2 is required to produce a similar warm climate state. In light of these results, a CCN–cloud mechanism could help alleviate current disparities between assumed CO2 levels in climate models and those estimated from carbon cycle budget models for past warm climates. This study finds that CCSM3 simulations of the PETM climate that includes enhanced levels of CO2, CH4 and lower cloud drop numbers are in good agreement with a wide range of palaeo temperature records. Along with the studies in [12,13], this is one of the few coupled climate model simulations that agrees with much of the proxy data, including polar regions. The simulations show that the climate system is sensitive to the specification of reduced levels of CCN and associated changes in cloud properties. Hence, this may be an important climate factor that needs to be accounted for in simulating past climates. Given these findings, it would be of great value to find means to quantify the aerosol properties that may have existed in past climate states. In support of previous studies and reconstructions, the study finds that the ocean general circulation shifted between the pre-PETM climate to that of the PETM state. In particular, the sites of deep water formation shifted from the high polar regions in both hemispheres to the mid-latitudes upon entering the warmer PETM climate. Simulated deep ocean temperatures agree well with the limited available data for the PETM. In terms of the terrestrial sites, the PETM simulation is in good agreement with much of the proxy data. Surface temperatures are not excessively cold, with cold month mean temperatures staying above freezing. The simulations also indicate that the hydrological cycles of the PETM and pre-PETM were far more active than present day. In addition, the more active high-latitude hydrological cycle led to an increase in fresh water input into the Arctic basin. Remaining questions are related to the magnitude of surface warming for marine and terrestrial regions in the tropics between 30° S and 30° N. Few data exist to constrain surface temperatures in this region. The model simulations suggest surface temperatures in excess of 45°C for certain regions. It is important to find stricter proxy constraints for these tropical regions. Finally, this study supports the finding of others that the Earth's climate system is more sensitive to greenhouse forcing on longer time scales (e.g. [44–46]). In particular, this work finds that the ESS for the PETM or pre-PETM relative to the modern climate is twice as large as the traditional Charney climate sensitivity. We would like to thank Cindy Shellito (University of Northern Colorado) for allowing us to begin our simulations from her PETM Eocene configuration of CCSM3. Conversations with Matt Huber (Purdue University) over the years have greatly benefited our work. We thank Garland Upchurch (Texas State) for stimulating conversations regarding palaeobotanical issues for warm climate conditions. We thank Bill Large and Gokhan Danabasoglu (N.C.A.R.) for helpful conversations regarding the ocean simulations. Data from all simulations cited in this study are available on request to J. Kiehl ([email protected]). N.C.A.R. is sponsored by the National Science Foundation. C.A.S. was supported through a grant from the NSF EAR Sedimentary Geology and Paleoebiology program. FootnotesOne contribution of 11 to a Discussion Meeting Issue ‘Warm climates of the past—a lesson for the future?’. © 2013 The Author(s) Published by the Royal Society. All rights reserved. References
Page 3The term equable climate has been used to describe the global extent of warmth in past climates, which have a reduced equator-to-pole temperature difference (EPTD), warm polar regions with a reduced seasonality and ice free conditions at both poles [1,2]. The extent of this warming is supported by a wide range of data. Recent syntheses of terrestrial [3] and marine [4] proxy climate data for the Early Eocene suggest that the polar temperatures were 15°C or more but that the tropics were only slightly warmer than modern. Moreover, palaeobotanical data also suggest that the high latitudes were above freezing throughout the year [5], which is a major change over present conditions despite the fact that the continents are not that different from the modern. The Early Eocene equable climate problem relates to differences between climate model simulations and proxy reconstructions of the Early Eocene and the climate inferred from climate proxies. The modern generation of climate models has managed to capture much of this warmth from the proxy data in the low and mid-latitudes by forcing the climate with very high concentrations of CO2, 16 times pre-industrial concentrations of CO2 [3,6], but simulating above freezing temperatures at the poles all year round is difficult. The assumption of a strong seasonal bias in the proxy data must currently be assumed in order to reconcile proxy polar temperatures with climate model output [4]. Estimates of Early Eocene temperatures include annual sea surface temperatures (SSTs) of up to 27°C [7] and terrestrial mean annual temperatures (MATs) of up to 18°C [8] at palaeolatitudes greater than 80°N. In the Southern Hemisphere (SH), SSTs between 17°C and 32°C [9–11] have been reconstructed at palaeolatitudes greater than 60°S, while terrestrial MATs between 12°C and 18.8°C have been reconstructed at similar latitudes [12–14]. These high latitude temperatures are likely to have been sufficient to prevent any significant permanent ice cover. While there is reasonable data coverage for the mid- and high latitudes, data from the low latitudes are scarce. Tropical SST data are available from the Tanzania drilling project, which indicate that SSTs at a palaeolatitude of 18°S were approximately 33°C [15]. One of the features inferred from this distribution of temperatures is that the temperature difference between the pole and the equator was much reduced compared to the modern day. There is also evidence of an enhanced hydrological cycle in the high latitudes during the Early Eocene [16–18]. Water vapour has an impact on the radiation balance of the planet through the water vapour greenhouse effect, the cloud greenhouse effect and via reflection of shortwave radiation from clouds and ice [19]. Understanding what role an intensified hydrological cycle may play in developing and maintaining an equable climate is therefore also of interest. The first paper on the Early Eocene equable climate problem was published over 30 years ago [20] and substantial modelling efforts have been undertaken since then in order to simulate the Early Eocene climate. Many advances in model development have also been made, and whereas the earliest Eocene models were limited to either energy balance models (EBMs) or early general circulation models (GCMs) with fixed seasons, the current generation of models simulate the dynamics of the atmospheres and oceans and in some cases vegetation, and are of higher resolution and have improved and revised physics. This has consequently improved our ability to simulate the Eocene climate. Meanwhile, advances in existing proxy methods, the development of new methods and the acquisition of additional proxy data have led to the warmer, revised temperatures for the tropical marine realm [21] and terrestrial realms at all latitudes [22–25]. Thus, the equable climate problem is still apparent in the proxy datasets. While modelling studies have improved in their simulation of the Early Eocene, the processes that contribute to the amplification of polar temperatures during the Early Eocene are difficult to accurately model (e.g. clouds) and are not well understood. However, the model–data discrepancy persists in the high latitudes. The aim of this paper is to understand whether perturbing uncertain climate model parameters can offer insight into the climate processes involved in developing and maintaining the equable Early Eocene climate. Many of the earliest model experiments of the Early Eocene, run with increased CO2 concentrations compared to the modern, simulated high latitudes and continental interiors that were warmer than the modern, but not warm enough compared to the proxy climate evidence of the Early Eocene. Sloan & Barron [26,27] simulated high latitudes and continental interiors that are warmer than the modern, but still cooler than the proxy climate evidence. This model–data mismatch generated a range of possible explanations including missing components and processes in the models such as polar stratospheric clouds (PSCs) [27,28] and tropical cyclones [29–31] to approximations in the boundary conditions associated with coarse model resolution, for example, the presence of large lakes (e.g. [32,33]) altered orbits (e.g. [34]) and the role of heat transport [35,36]. A recurring theme in Early Eocene modelling studies is the contribution of clouds in equable climates. Sloan et al. [37], Kirk-Davidoff et al. [28] and Kirk-Davidoff & Lamarque [38] have all investigated the role of PSCs in equable climates in response to elevated concentrations of CO2 and CH4. Sloan et al. [37] included idealized prescribed PSCs in the GCM Genesis v. 2, which resulted in up to 20°C warming in oceanic regions where sea-ice was reduced. This warming was still insufficient to account for warming seen in the proxy data available at the time, but compared to more recent proxy data these simulations were approximately 10°C too cool at latitudes of around 60°. Kirk-Davidoff et al. [28] and Kirk-Davidoff & Lamarque [38] investigated the mechanisms that led to the formation of PSCs and the response of the climate. PSCs were found to warm in response to higher CO2 via changes in stratospheric circulation and water content, but the large radiative effects required to warm the polar regions were found to be related to ice crystal number density in the PSCs, and a lack of theoretical knowledge may have prevented these hypotheses from being developed further. Abbot & Tziperman [39] identified a high latitude cloud radiative forcing feedback using a simple column model. They found that increased extra tropical surface temperatures led to the initiation of strong atmospheric convection and the convective clouds led to additional warming of the high latitudes. The radiative effect of the resulting convective clouds reduced the EPTD by 8–10°C. Further work using this column model investigated the constraints atmospheric and oceanic heat transport and CO2 concentration had on the convective cloud feedback [40]. This feedback was found to be present in modern model simulations forced with CO2=2240 ppmv, and for the Eocene with CO2=560 ppmv. An alternative solution to the equable climate problem was suggested by Kump & Pollard [41] and Kiehl et al. [42]. Cloud condensation nuclei (CCN) play an important role in cloud properties such as cloud water content, cloud opacity and cloud lifetime. In the past, the distribution of CCN was probably different from today because the distribution and composition of atmospheric aerosols were different [43]. Based on this, Kump & Pollard [41] increased CCN radii in a Cretaceous climate simulation using the Genesis (v. 3.0) GCM. This resulted in a decrease in cloud amount and cloud albedo leading to a dramatic warming, both globally and at the poles, and a decrease in the EPTD. Lunt et al. [4] have recently published a review on Eocene modelling termed EoMIP (the Eocene modelling intercomparison project) in which they compare five recent modelling studies for the Early Eocene. The modelling studies have all been run with different objectives; different boundary conditions and multiple values of CO2 have been used in some studies. The models used were: HadCM3L, the sister model of the FAMOUS model, which is used in this study [4]; ECHAM5/MPI-OM [44]; the GISS model [45]; and the two versions of the model CCSM—CCSM_H [3,46] that has no aerosol load following the approach of Andreae [43] and Kump & Pollard [41,43] and CCSM_W [6,47] that has a modern aerosol load. At a given CO2 concentration CCSM_H and CCSM_W give different global means; for instance, there is a 3°C difference in mean surface air temperature (SAT) between CCSM_H and CCSM_W at 16 times pre-industrial CO2 concentrations, the level at which the best match to Eocene proxy data was found for that model. The range of CO2 concentrations resulting in the best Eocene simulation between the models varied between 2 times and 16 times pre-industrial concentrations, demonstrating the need for better constraints on actual CO2 concentration during the Early Eocene. A comprehensive comparison of model results with recent syntheses of proxy data was made [3,4] as part of the EoMIP and a one-dimensional EBM [44] was used to investigate, identify and understand the inter-model variability. The ACEX data points from the Arctic Ocean [7] indicate SSTs of approximately 13°C for the Ypresian (56.0–47.8 Ma) and an SST of approximately 22°C recorded during the Early Eocene Climatic Optimum (EECO, 53.1–49 Ma). Few of the models manage to simulate these temperatures. In the SH, SSTs greater than 25°C are measured from both EECO and Ypresian material from ODP 1172D in the Pacific Ocean [11] and Waipara River of the coast from New Zealand [48]. CCSM_H is the only model that managed to intersect the lower error bars of these temperature estimates. In summary, there is considerable inter-model variability between the models in the EoMIP. The variability is considerable larger than present day (PD) inter-model differences, with very different CO2 concentrations giving the best fit to data. The differences between models have been attributed to a combination of greenhouse effect and surface albedo feedbacks rather than differences in cloud feedbacks or heat transport [4]. Differences in the climates of CCSM_H and CCSM_W are related to differences in the assumed aerosol loads used. Despite the variation in boundary conditions between these five models, only CCSM_H manages to simulate temperatures within the lower boundaries of the estimates at the warmest high latitude locations (i.e. ACEX, ODP 1172D and Waipara River), thus demonstrating the need for alternative solutions to the equable climate problems. The EoMIP study highlights the inter-model variability between the models studied. Climate models are constructed by discretizing and then solving equations that represent the basic laws that govern the behaviour of the atmosphere, ocean and land surface [49] and many approximations are required in order to solve the nonlinear system of partial differential equations. Note that the solution of a partial differential equation depends on (i) the initial conditions, (ii) forcing boundary conditions (focus of the previous palaeoclimate studies) and (iii) approximations in the form of climate parametrizations (this study). Parameter uncertainty stems from the fact that small-scale processes in all components of the climate system cannot be resolved explicitly in the climate system. This is the case in cloud processes for example [50,51]. Parametrization of sub-grid scale processes is a major source of uncertainty in climate prediction [52], and while in some parametrizations the processes, observational evidence or theoretical knowledge is well understood, where this information is scarce the values chosen for a parametrization may simply be because they appear to work [51]. Future climate change studies have recently focused on quantifying the uncertainty arising from these parameters using Monte Carlo-type techniques [53]. This type of work is referred to as perturbed physics ensembles (PPEs) because suites of simulations are generated by perturbing climate-sensitive model parameters. The resulting spread in predictions is quantified, leading to model-dependent probabilistic estimates of the distribution of future climate, warming and climate sensitivity. In a few cases, the ensembles are very large (i.e. a thousand member ensemble) [53,54] but in most cases the number of simulations is limited by the computational cost of complex climate models to a few tens or a hundred simulations as is the case in [50,55]. Ensembles with perturbed climate-sensitive model parameters have begun to be used in palaeoclimate research, primarily for the Late Quaternary and particularly on the issue of climate sensitivity and El Niño Southern Oscillation (e.g. [56–60]). Ensembles with perturbed climate-sensitive model parameters have also been used to ‘tune’ the climate model to proxy data for the last glacial maximum (LGM) [61]. However, few studies have investigated older time periods apart from a small set of simulations for the Pliocene [62]. In practice, there are several hundreds of parameters that are poorly constrained in climate models and it is impossible to vary all of them. Gregoire et al. [61] identified a total of 10 parameters to be varied in FAMOUS, of which six parameters had been tuned in a previous study [63] and recognized as having a high impact on the climate of HadCM3 [50]. The study by Murphy et al. [50] identified key parameters that had a major impact on Charney climate sensitivity (the global average temperature increase associated with a doubling in CO2 and including a specific set of feedbacks). This paper investigates the effect of parametric uncertainty on the Early Eocene equable climate problem using the model FAMOUS. The motivation of this study is to attempt to detect ensemble simulations that match the proxy data available for the Early Eocene and to understand how processes in these simulations vary from the rest of the ensemble. We deliberately do not limit the parameter set perturbations to only those sets that perform well for modern conditions because we wish to explore if any combination of parameters are able to simulate the Early Eocene equable climates. FAMOUS (Fast Met Office/UK Universities Simulator) is an atmosphere and ocean GCM (AOGCM) that is based on HadCM3 (Hadley Centre Model v. 3) [64]. While its parametrizations of physical and dynamical processes are almost identical to those of HadCM3, FAMOUS has a reduced resolution in both the atmosphere and ocean, and a longer time-step that reduces the computational resources required to run FAMOUS to 10% of that required by HadCM3 [65]. This favours the use of FAMOUS in experiments where large amounts of computational resources are required. The atmosphere component of FAMOUS is based on the Hadley Centre atmosphere model (HadAM3) (see [66] for full details). The atmosphere resolution in FAMOUS is 7.5° longitude×5° latitude grid, with 11 levels in the vertical. The ocean model in FAMOUS is the Hadley Centre ocean model (HadOM3) (see [64] for full details), which is a rigid lid model. The ocean resolution in FAMOUS is 3.75° longitude×2.5° latitude grid with 20 levels and a 12 h time step (using a distorted momentum equation), which is the same resolution as the model HadOM3L, and is lower than the resolution of HadOM3 (1.25° longitude×1.25° latitude grid). Since the resolution of the ocean model is greater than the atmosphere, FAMOUS uses a coastal tiling scheme that combines the properties of land and sea in coastal grid boxes. The ocean model can then use the more detailed coastline allowed by its higher resolution grid while conserving coupled quantities [65]. FAMOUS does not use flux adjustments. Land processes are modelled with the UK Meteorological Office's land surface scheme, MOSES 1 [67]. Smith et al. [65] give a detailed description of FAMOUS and highlight the major differences between FAMOUS and HadCM3. The version of FAMOUS used in this work is identical to that of Gregoire et al. [61] and slightly differs from Smith et al. [65] as described in Gregoire et al. [61]. The resolution of FAMOUS is not as high as the models used to investigate future climate change; horizontal resolution of the order of 1° to 2° is now commonly used in the ocean component of most climate models [68]. However, FAMOUS fills an important niche in the current generation of models sitting between the higher resolution AOGCMs and the lower resolution, highly parametrized Earth system models of intermediate complexity. The reduced resolution allows us to fully spin-up the ocean, with all of the simulations presented in this work extended to at least 8000 years. This would be impossible with the higher resolution models but is essential since the time scale for ocean equilibration is measured in thousands of years. In the original tuning of FAMOUS, Jones et al. [63] systematically tuned the model to reproduce both the equilibrium climate and climate sensitivity of HadCM3. Smith et al. [65] then undertook manual tuning to reduce a cold bias in the northern high latitudes, which led to the removal of Iceland. Gregoire et al. [61] conducted ensembles with perturbed climate-sensitive model parameters for the PD and LGM climates. Building on this work, we use the PD control parameter values in Gregoire et al.'s [61] configuration as our control PD simulation. The PD version of FAMOUS uses the following concentrations of greenhouse gases: CO2, −280 ppmv; CH4, 760 ppmv; N2O, 270 ppmv. The orography is derived from the US Navy 10 min resolution dataset, with some small additional smoothing at latitudes poleward of 60° (see [65] for full details). The ocean resolution of FAMOUS does not allow for flow between the Atlantic and the Mediterranean. Instead a simple mixing has been parametrized for this region in an area that extends from the surface to a depth of 1300 m. An artificial island is used at the North Pole to avoid the problem of converging meridians [65]. Palaeogeographic reconstruction is a critical boundary condition in palaeoclimate modelling, and reconstructing continental interiors, dimensions of palaeo-orography, palaeo-shorelines of ancient lakes and the widths of epicontinental seaways is challenging as the geological evidence left by these features can be minimal [69–72]. Modelling experiments have been used to explore the impact on climate for some of these poorly constrained variables. For example, experiments have investigated the impact of the inclusion of a large lake in western North America [32], the opening and closing of the Arctic seaways during the Early Palaeogene [45] and the impact of uncertain orography [27,73–76] on Eocene climate. Results suggest that uncertain palaeogeography tends to increase regional uncertainty in modelled climate, with some potential for climatic tele-connections and modification of global climate. The Early Eocene simulations presented here use a palaeogeography created using similar methods to Markwick & Valdes [72]. The palaeogeography is similar to the HadCM3 Early Eocene simulations conducted by Tindall et al. [77] but at the resolution of FAMOUS. There is no flow between the global oceans and the Arctic Ocean in these simulations although opening these gateways could impact climate [45] and FAMOUS does not explicitly represent lakes. Early Eocene atmospheric CO2 concentration is an important but uncertain boundary condition. Proxy measurements indicate that CO2 in the Early Eocene was higher than present and estimates range from as low as 300 ppmv to more than 4400 ppmv [7,15,78–82]. Climate modelling studies of the Early Eocene have used different CO2 values that span the entire proxy range. For these Early Eocene simulations, CO2 was set at 560 ppmv (twice pre-industrial concentrations). While this is at the lower end of the range of predicted CO2 values for the Early Eocene, it has been used because Early Eocene sensitivity simulations (P. J. Valdes 2010, unpublished data) showed that the Eocene configuration of FAMOUS is relatively sensitive to CO2. All other greenhouse gases (CH4, N2O) were set to pre-industrial values. Indirect evidence indicates that during the Early Cenozoic methane concentrations of these other greenhouse gases could have been much higher due to the expansion of peat lands and the consequent increase in methanogenesis for instance [83–86]. However, in the absence of suitable proxy data to quantify this increase we use PD values. Orbital changes have been calculated for the past 250 Myr [87] and studies have identified a strong eccentricity and precession signal from Early Eocene sediments [85–87]. We attempt to simulate a very long multi-million year interval in which many orbital configurations would have occurred. While the role of orbital forcing may be a driver for short-term hyperthermal events [84,88], we are interested in simulating the overall warmth of this period and thus have used a modern orbital configuration. Modelling studies that have investigated the impact of orbital forcing on the Early Eocene climate have improved the model–data fit if specific orbits are chosen. Sloan & Morrill [34] showed that extreme orbital values from the calculated Pleistocene range could reduce temperatures in the Northern Hemisphere (NH) continental interiors compared to the orbital configuration for the PD. Sloan & Huber [33] showed that between precessional end members for an Eocene greenhouse world widespread regional variation occurred, including: SST variation of 5°C in the high northern latitudes; up to a twofold variation in upwelling strength in tropical regions; and changes in net surface moisture balance (precipitation–evaporation) of up to 3 mm d−1 in the tropics. Uncertainty in orbital forcing has a limited impact on global mean climate values and a larger impact on regional and seasonal climate, in particular at high latitudes. In the studies referenced here [33,34], uncertainty was more pronounced in high latitude terrestrial realms and in the low latitude marine realm. There is very little data available for vegetation reconstruction of past climates and the data that do exist may not be fully representative of the diversity of the area they come from. Numerous modelling studies have investigated the impact of vegetation on palaeoclimate [89–92] and several studies have looked specifically at Early Eocene modelled vegetation [73,93,94]. While the impact on global climate has been noted to be small, changes to regional climate can be distinct [73,93,94]. Vegetation in model simulations can either remain static and unchanging through time or dynamic and responding to the changing climatic conditions. Both approaches have advantages and disadvantages, as reviewed in Peng [95]; for example, dynamic vegetation may increase precipitation and reduce temperature extremes [96]. The work presented here used a static and uniform vegetation configuration of shrub-like plants everywhere as we consider the effect of vegetation feedbacks to be secondary compared to the parameter perturbations. Future work will examine the impact of vegetation change. Table 1 gives a description of each parameter perturbed in this work. We perturb 10 parameters within their upper and lower bounds. The uncertainty bounds were based on previous studies [50]. The uncertainty arises because of the large spatial and temporal variation of many of these processes.
We have run two sets of perturbed physics simulations. In the first set all 10 groups of uncertain parameters are perturbed simultaneously and at 10 equal intervals between the lower and upper boundaries of their uncertain range; we refer to these simulations as the multiple parameter perturbations (MPPs). In order to facilitate the best use of computing time and the greatest coverage of different parameter sets a statistical method of Latin hypercube sampling (LHS) is used to define the parameter values for the MPP simulations [97]. Using LHS with 10 parameters requires of the order of one hundred simulations to obtain a reasonable coverage of the parameter space [98]. We therefore generated one hundred unique parameter sets, maximizing the parameter space that is sampled for a finite number of simulations in a statistically robust way. Full details of the LHS methodology are available in Gregoire et al. [61] who originally ran PD simulations with the same MPP sets. The MPP simulations were initially set up to run for 6000 years, though runs of particular interest were integrated for 10 000 years. This length of the runs is required in order to achieve full equilibrium in both the surface and deep ocean in the Early Eocene climate. In order to understand some of the causes of the changes in climate, we selected a simulation with a promising Early Eocene climate based on the 6000 year results (from herein referred to as E6000). The climate in E6000 exhibited global warmth (SAT more than 30°C) and polar regions with SAT more than 10°C. We used the 10 groups of perturbed parameter values in E6000 to set up a further set of simulations in order to investigate the response of the climate to changes in one parameter at a time. This second group of experiments was termed the single parameter perturbations (SPPs). We ran 15 SPP simulations in total from the original 10 parameter groups by separating the parameters in CW (threshold value of cloud liquid water at which precipitation commences) into land and sea components; the four parameters in the OCN_DIFF_H group, horizontal ocean diffusivity, were also split into three separate experiments. Finally, OCN_DIFF_V, vertical ocean diffusivity and ATM_DIFF, horizontal atmosphere diffusion parameter groups were sampled twice: once using the values in E6000 and then a second set of simulations were conducted reducing the values even further than in E6000. These simulations are run for up to 9000 years. A summary of the different sets of simulations and criteria used to assess them is shown in table 2. Although E6000 does not make it into the final Eocene simulations, the parameter values of E6000 are shown in table 3 for reference.
PD simulations and with twice pre-industrial CO2 concentrations (560 ppmv) were used to calculate Charney climate sensitivity values for the same MPP sets that were used in the Early Eocene simulations (E_MPP). The PD configuration is identical to that described in §2b with the exception that CO2 concentrations of 560 ppmv were used. These simulations were run for 200 years. Model output is compared to published multi-proxy datasets that have undergone comprehensive selection and standardization. We use the terrestrial dataset first compiled in Huber & Calallero [3] and also used in Lunt et al. [4]. Our marine dataset is also from Lunt et al. [4]. An outline of the proxy data and consideration of the uncertainty associated with this data is given below. The terrestrial proxy dataset compiled by Huber & Cabellero [3] contains fifty Early Eocene data of Ypresian (56.0–47.8 Ma) and Lower Lutetian age. The Lutetian occurred between 47.8 and 41.3 Ma; however Lu1, the first global section of the Lutetian, is dated at 47.47 Ma thus we take the age span of the terrestrial data to be between 56.0 and 47.47 Ma. PETM (Palaeocene–Eocene Thermal Maximum) and other hyperthermal events were excluded in the compilation of the dataset by Huber & Caballero [3]. One Middle Eocene data point approximately 45 Ma from the tropics is included in the absence of any tropical data from the Early Eocene [100,101]. There is no data coverage at latitudes greater than 65°S and coverage is highest in the NH particularly over North America. In order to account properly for systematic bias and spatio-temporal sampling uncertainty, the authors have reconstructed MAT based on leaf-margin analysis (LMA) where possible. CLAMP (physiognomic analysis of leaf fossils) is used for MAT reconstruction when LMA is not available. MATs are calculated using the Kowalski & Dilcher [24] calibration when feasible as this offsets the well-established cool biases that may have been incorporated in the original calibrations (see [3] and references therein). Error bars are included in the terrestrial dataset to encompass the uncertainty introduced from the age of the material, topographic uncertainty and from the calibration method. All palaeolatitudes are adjusted to 55 Myr plate configuration using the Gplates software (www.gplates.org) and the plate model of Muller et al. [102]. Palaeo-elevation uncertainty is quantified by calculating the standard deviation of PD topography at elevations greater than 1500 m, and then applying this to Eocene data to calculate the uncertainty in temperature as a result of lapse rate (±2.4°C), based on the work by Hren et al. [103]. The marine dataset used in this work was compiled by Lunt et al. [4] and includes data from 13 locations. The age range of the data spans the ages of approximately 55.0–49.0 Ma. Data are grouped into three categories by Lunt et al. [4] and include (i) data aged approximately 55 Ma that is termed Late Palaeocene data but excluding the PETM in [4]; (ii) well-constrained EECO data between 53.1 and 49 Myr; and (iii) Early Eocene data which is constrained to the Ypresian, but not thought to be a representative of the EECO. This final dataset is referred to as the background Ypresian. Given the recent new boundaries of the Ypresian (56–47.8 Myr) (www.stratigraphy.org) the Late Palaeocene data referred to in Lunt et al. [4] now are categorized as the earliest Eocene; thus, we term this dataset the earliest Eocene. Multiple data are available at several locations where either two proxy methods have been used or data of different ages are available and our final marine dataset contains 15 data points in total. Data are generally well constrained with the exception of the data from Seymour Island in the Antarctic Ocean [9], which is provisionally classed as background Early Eocene, although this may potentially be Middle Eocene in age. Climate data are included from a range of proxies; δ18O (planktic foraminifera), δ18O (benthic dwelling molluscs), Mg/Ca (planktic foraminifera), clumped isotopes and Tex86. The authors have calculated three temperatures for the δ18O data [77,104,105] in order to capture the upper and lower bounds of temperature estimates. Similarly three assumed values of Mg/Casw values (3, 4 and 5 mol mol−1) are used to calculate Mg/Ca temperatures. There are now several published calibrations available for Tex86 and the ‘high’, ‘low’ and ‘inverse’ calibrations are all used. In addition, samples with a branched versus isoprenoidal tetraether (BIT) index greater than 0.3 are excluded where possible as this is now accepted as good practice (see [106] for further details). However, Ypresian samples from Tanzania [15] and Hatchetigbee Bluff, coastal North America [107] were included by Lunt et al. [4] despite higher BIT indices (0.3–0.5), in order to include more Early Eocene data points. We have averaged proxy temperatures calculated with different methods at the same location but we have not averaged data of different ages. As a result our dataset contains 15 points that we use to compare to model output. The temperature ranges of these data points are summarized in table 4. Minimum and maximum temperature estimates from the multiple proxy methods and calibration errors are plotted in all our estimates. The terrestrial dataset spans the ages of 56.0–47.47 Ma, and no divisions are specified. The marine dataset spans a slightly narrower age range (55.0–49.0 Ma), which is encompassed by the Ypresian, but has been subdivided into three categories: earliest Eocene, EECO and background Ypresian. Non-EECO data (i.e. earliest Eocene and Ypresian) is referred to as the background Early Eocene as it does not include the peak EECO temperatures.
A wide range of proxy data using different methods have been used in these datasets, which introduces uncertainty from numerous sources. For example, uncertainties are associated with reconstructing palaeolocation and depositional environments [73], age control and diagenesis and alteration [115]. The geochemical effects on biological material are another source of considerable uncertainty, for example, while the effects of temperature and seawater δ18O on foraminiferal δ18O have been recognized for a long time [116], the effects of seawater CO2 chemistry on foraminiferal δ18O were only recognized through culturing experiments in the late 1990s [117,118]. This led to the realization that foraminifera δ18O-based temperature estimates may be too low for periods of the past where atmospheric CO2 was high, such as the Early Eocene [78,119,120,121]. Better constraints on Early Eocene CO2 will also help improve temperature estimates from foraminiferal δ18O; however, other ‘unknown’ or currently unquantified factors which affect foraminiferal δ18O may not have been recognized yet. Similarly, Tex86 is a relatively new palaeothermometer [122] and understanding the environmental signal recorded by Tex86 for the Early Eocene is exacerbated by use of this proxy outside of its calibration conditions. High latitude areas from which very warm Early Eocene temperatures have been recorded by Tex86 (for example, the Arctic Ocean) would have undergone several months of darkness due to the boreal winter; the lack of these organisms in the modern high latitude oceans makes use of this proxy method problematic in polar regions [123]. Incubation experiments are required to calibrate the Tex86 palaeothermometer for tropical SSTs as the PD ocean is simply not hot enough [124]. Proxy data are compared point by point with model output at grid box resolution and with zonal mean values. Where the same land surface type is not present in the model as in the proxy data the nearest matching land surface location is used along a band of longitude. Terrestrial data are compared with the surface air temperature at 1.5 m in the model over terrestrial surfaces while marine data are compared to the ocean temperature at a depth of 5 m. Some combinations of model parameters generated by our sampling technique result in climates which are far from realistic, for either a modern climate or a palaeoclimate [61]. Moreover, in the extreme conditions of the Early Eocene, 82 out of 100 MPP simulations fail to complete due to the model generating very extreme climates (e.g. tropical temperature in excess of 50°C) resulting in numerical instabilities in the model. Eocene MPP simulations were required to run for in excess of 10 000 years and Eocene SPPs ran for up to 9000 years. A summary of the initial number of simulations, the selection criteria and the final number of simulations for each set of experiments conducted is given in table 2. It should be noted that in all Eocene simulations, we needed to perform multi-millennial runs in order to reach near equilibrium in both the surface and deep ocean. In some cases, initial results from the first 1000 years of the Eocene MPP simulations gave significantly different results. For instance, some simulations showed an 8°C change of global SAT between the end of 1000 years runtime and the end of 10 000 years. The latitudinal gradients were also impacted such that in some simulations the EPTD changed by more than 15°C from 1000 to 10 000 years. Even between 4000 and 10 000 years, the gradient changed in some simulations by up to 5°C. The changes seemed to be strongly linked with the effects of ocean overturning and the time scales are consistent with this. These results highlight the potential for misinterpretation of the climatic effects of model changes (either parameter or boundary conditions) if the simulations are less than a few thousand years in duration and justify the use of a relatively fast but comprehensive model such as FAMOUS. In order to verify the stability of the Eocene simulations that completed 10 000 year runs, the time series of the global mean top-of-atmosphere (toa) net energy balance and global surface air temperatures were plotted against each other [125]. In three simulations global surface air temperature appeared to be in equilibrium but the toa net energy was not tending to zero and so these simulations were discarded. In the remaining 17 Eocene simulations (15 MPP and two SPP) the global mean net toa energy balance is less than 0.3 W m−2 (and in most cases less than 0.1 W m−2) indicating that the simulations were in radiative balance. Trends in time-series plots of global mean annual surface air temperature are small in the final simulation set with most simulations varying less than 2°C over the final 1000 years of simulation. In initial condition ensembles, model parameters and forcings are identical throughout the ensemble but each simulation has a different starting state. In these ensembles, the natural variability in the system is of interest and thus an ensemble mean value is a useful measure. In PPEs such as the work described here, model parameters and forcings have been changed while the initial conditions are identical. The value of PPEs is in understanding where and how the climate converges and diverges within the ensemble. We therefore describe the range of climates simulated but do not present the ensemble mean. The parameter values of the final simulations and of the control parameter set are given in table 5. The simulations are ranked in the order of ascending global SAT and this ranking is used to identify the different simulations, i.e. the simulation with the lowest SAT is termed E1 and simulation with the highest SAT is referred to as E17. We performed simple regression analysis of each parameter against a number of global annual climate values (i.e. SAT, mean annual precipitation (MAP), tropical SSTs, polar SSTs, EPTD planetary albedo, low cloud and high cloud global values) but the resulting R2 correlation coefficients were all below 0.5 indicating that direct correlation between these variables is not strong and that it is the combination of changes which is key.
Table 3 summarizes some climate variables for the final 17 simulations. SAT in our ‘final’ simulations ranges from 12°C to 32°C, MAP ranges from 2.7 to 4.1 mm d−1 and there is a strong positive correlation between SAT and MAP, with an R2 of 0.97 and a slope equivalent to a 0.76 mm d−1 (approx. 25%) increase per 10°C. This strong relationship also holds for the land and ocean precipitation, i.e. the fraction that falls over land versus ocean (approx. 30% of total precipitation falls over land) remains approximately constant across the range of simulations. Figure 1 shows the SAT averaged from years 9900 to 10 000 for two example runs: simulation E1, which has the coldest global mean temperature of 12.3°C; and the warmest model, E17, with a much higher global mean temperature of 31.8°C. Not surprisingly, the basic spatial patterns are quite similar between the two simulations but with a large offset of approximately 15°C. In E1 (figure 1a), the SATs are significantly below zero at high latitudes in both the north and south. These cold temperatures are even more pronounced seasonally (not shown) as temperatures decrease below −20°C in large parts of the high latitude continents. By contrast, annual mean temperatures in the warmest models remain above freezing for almost the whole globe. Simulations E16 and E17 have no annual mean temperatures below zero and E15 has a small area of sub-freezing temperatures (reaching −10°C) in the very heart of Antarctica, although the coastal regions of Antarctica remain above freezing. Seasonally, there are still some sub-freezing temperatures but in the warmest models, these are confined to very small regions in the heart of the continents polewards of about 60° and in regions where there are no proxy data to evaluate such values. Figure 1. Annual surface air temperature for (a) the coolest simulation E1 and (b) the warmest simulation E17 from the final Eocene ensemble. (Online version in colour.)
The spatial patterns of precipitation for simulations E1 and E17 are shown in figure 2 for both summer and winter seasons. The patterns over land are broadly consistent between warm and cold models but with some of the most marked differences in precipitation occurring at high latitudes. In E1, the North Pole is a ‘polar desert’ (shown clearly in figure 2a), whereas in E17 the poles are relatively moist. This is unsurprising given the much warmer and sea-ice free polar regions in the warm model. In the tropics, there are some important differences particularly over the ocean where the cold model shows a distinct split intertropical convergence zone (also clear in figure 2a) whereas the warmer model has a much broader feature and centred on the equator. However, over land there are somewhat smaller differences in the patterns of precipitation. In both simulations, the sub-tropics are seasonally dry but annual averages reveal only very small areas which are dry throughout the year. Figure 2. Seasonal (DJF, JJA) precipitation for (a) and (b) the coolest simulation E1 and (c) and (d) the warmest simulation E17 from the final Eocene ensemble. (Online version in colour.)
Although some simulations achieve very warm polar temperatures, they also have very warm tropical temperatures so that the resulting EPTDs are generally very similar to present. EPTDs are calculated for all simulations by subtracting mean polar temperatures (70°N to 90°N and 70°S to 90°S) from the mean equatorial temperatures (10°N to 10°S). We have calculated the marine EPTD for the PD control simulation (table 5), and the NH marine EPTD is 27.2°C. Marine EPTDs in the coldest and warmest Eocene simulations E1 and E17 are both 25.1°C. Intermediate models (E7–E15) have a greater NH marine EPTD of up to 31.9°C. In these intermediate simulations, sea-ice acts as a buffer, keeping the marine temperatures at high latitudes around 0°C. Once this reduces, polar oceans begin to warm which then reduces the gradient. Examination of the equivalent terrestrial gradients helps confirm this as it shows a simpler gradual reduction in temperature gradient between the coldest and warmest models. The NH terrestrial EPTD for the PD is 39.2°C. The NH terrestrial EPTD for simulations E1–E15 ranges between 38.2°C and 43.5°C, whereas E16 and E17 have a NH terrestrial EPTD of approximately 32°C, a 6–7°C reduction compared to the PD. The SH terrestrial EPTD for the PD is very large (58.4°C) due to the ice covered Antarctic. During the Early Eocene, with no ice and no circumpolar current, the largest SH terrestrial EPTD for the Eocene is 42.7°C. However, the warmest Eocene models have a terrestrial EPTD of approximately 30.5°C which is a notable reduction. The SH marine EPTD in the PD simulation is 27.8°C, and the majority of the Eocene simulations have an SH marine EPTD between 24.6°C and 28.9°C. Three simulations have a smaller SH marine EPTD: these are E14 (20.9°C), E16 (22.9°C) and E17 (20.1°C). Thus, in our two warmest Eocene simulations (E16 and E17) terrestrial EPTD in both hemispheres and SH marine EPTD do show a small reduction in temperature gradients in both hemispheres, which are compatible with the reduced EPTD suggested by the sparse data available for the Eocene. The 6 times pre-industrial CO2 HadCM3L simulation in the EoMIP [123] had the least polar amplification of temperature from the five models compared. HadCM3L is an intermediate model (resolution) between HadCM3 and FAMOUS. As it is part of the same family of models as FAMOUS we compare the EPTD as calculated using the method above with these simulations. The NH SST EPTD is 29.7°C and the SH SST EPTD is 26.3°C. These EPTDs are larger than those in our PD control but are well within the EPTD range simulated by the Eocene ensemble, maximum values of which are 32.2°C for the NH and 29.0°C for the SH. The sensitivity of the Early Eocene, proto-Atlantic Ocean meridional overturning circulation to changes in the concentration of CO2 (which changed the warmth and the presence of sea-ice) was described in Lunt et al. [4], for HadCM3L. We find similar results for the Atlantic overturning circulation in our suite of simulations. The warmer the simulation, the stronger the Atlantic intermediate-water formation, with a jump in the strength between simulations E6 and E7 related to a loss of year round sea-ice in the North Atlantic. Further increase in Atlantic intermediate-water formation in the very warm simulations (E14–E17) is associated with the almost complete loss of seasonal sea-ice. However, the location of oceanic convection, as indicated by the mixed layer depth, remains quite similar in all models. An intermediate to deep anticlockwise flow develops in the models where sea-ice disappears in the South Pacific (e.g. simulations E9, E14, E16 and E17). The centre of the cell is between 1000 and 2000 m with the bottom of the cell extending to 4000 m in E16 and up to 3000 m in E9, E14, E16 and E17. This replaces a deeper, small bottom water cell in the cooler models that have year round South Pacific sea-ice. In addition to the high latitude sources, there is also a source of intermediate water within the Tethys seaway. The relatively enclosed basin is very warm and experiences high evaporation. As a consequence, the surface waters are sufficiently saline to sink and these then spread out at about 2 km depth (figure 3). Figure 3. Atlantic meridional streamfunction (Sv) for the present day (PD) and final ensemble of 17 Eocene models. Positive values indicate clockwise motion and negative values indicate anticlockwise motion. (Online version in colour.)
A substantial increase in tropospheric mid-latitude westerlies which increase by more than 25% between coldest and warmest simulation is observed. Moreover, the strength of the tropical easterlies weakens considerably but they do not transition to westerlies as seen in [124]. It is probable that some of this difference is related to the resolution of FAMOUS which does not represent the atmospheric wave dynamics (particularly, the Madden Julian Oscillation) reported in [126]. The strengthening of the westerlies in our simulations seems to be strongly linked to a much intensified Hadley cell. Figures 4 and 5 show, respectively, the zonal means for the mean annual SSTs and terrestrial SATs for all 17 final simulations. The marine and terrestrial proxy datasets [3,4] are overlaid in these plots along with the lower and upper temperature bounds and the calibration errors for each data point. No simulation has a SAT or SST zonal mean that intersects all the proxy points (including the error bars). Figure 4. Early Eocene SSTs as compiled in Lunt et al. [4] shown as solid black circles. Upper and lower temperature error bars are shown in black. Model simulated zonal SSTs are plotted over the top. The four warmest simulations, E14 (dotted line), E15 (dashed double-dotted line), E16 (dashed single-dotted line) and E17 (dashed line), are highlighted with thicker lines for clarity. (Online version in colour.)
Figure 5. Early Eocene terrestrial SATs as compiled in Huber & Cabellero [3] shown as solid black circles. Upper and lower temperature error bars are shown in black and calibration errors are plotted in grey. Model simulated terrestrial zonal SATs are plotted over the top. The four warmest simulations, E14 (dotted line), E15 (dashed double-dotted line), E16 (dashed single-dotted line) and E17 (dashed line), are highlighted with thicker lines for clarity. (Online version in colour.)
In order to assess rigorously how well each simulation matches the proxy temperature estimates, we calculate the root mean square error (r.m.s.e.) for the difference between the simulation temperature predictions and the proxy data temperature estimates for all marine and terrestrial data (table 6). The r.m.s.e.'s for the different time periods included in the marine dataset have also been calculated (i.e. Ypresian, earliest Eocene and EECO) as have r.m.s.e.'s for mid- and high latitudes in the terrestrial dataset. Low r.m.s.e. values indicate that there is a better model–data fit than large r.m.s.e.'s. Simulations E1–E10 have greater terrestrial r.m.s.e.'s than marine r.m.s.e.'s, whereas simulations E12–E17 have greater marine r.m.s.e.'s than terrestrial r.m.s.e.'s indicating that above 22.1°C (SAT of E11) an improved fit with terrestrial proxy data is at the expense of the fit with the marine dataset.
Differences between the proxy temperature and simulation temperature estimates have been calculated for the marine and terrestrial datasets and are shown in figures 6 and 7. These are used to assess how well the simulations match the proxy data and to visualize any bias in the simulations. The simulation errors in the terrestrial data (figure 6) have an ‘approximately’ normal distribution. Simulations E1–E15 consistently underpredict terrestrial temperatures (e.g. the distribution is centred below zero). Simulations E16 and E17 over- and underpredict an equal number of terrestrial data points by up to ±10°C (e.g. the distribution is centred about zero). Figure 7 shows the differences for the marine dataset. There are not enough marine data points to assess the distribution of the data. Many of the simulations are skewed to the right indicating an overprediction of SSTs. Simulations E14 and E15 are centred near zero and overpredict SSTs in half the data points by up to 5°C but underpredict the remaining SSTs by between 10°C and 20°C. Simulations E16 and E17 are also centred near to zero; both simulations overpredict SSTs in half the dataset by up to 10°C and underpredict SSTs by the same amount (E17) or slightly more, up 15°C (E16). Figure 6. Histograms showing error (simulation temperature estimate minus proxy data temperature) for all terrestrial data points. Note that 0 is not in the centre of the x-axis. Numbers above the graphs denote the rank of simulation in terms of SAT. (Online version in colour.)
Figure 7. Histograms showing error (simulation temperature estimate minus proxy data temperature) for all marine data points. Note that 0 is not in the centre of the x-axis. Numbers above the graphs denote rank of simulation in terms of SAT. (Online version in colour.)
The four warmest simulations (E14–E17) all consistently overpredict SSTs at four locations. These are the Ypresian age data recorded at Tanzania and Hatchetigbee Bluff and the earliest Eocene data from ODP 865 and ODP 1209 and Seymour Island. E14 has the smallest error and E17 has the largest error in all these locations, with the error varying between 1°C and 11°C at these locations. Three of these locations have specific uncertainties associated with them, uncommon with the other marine data points. The data at Tanzania and Hatchetigbee Bluff were included in the marine data compilation despite having BIT indices between 0.3 and 0.5, indicating a large terrestrial organic matter component in the data signal. This has increased uncertainty in the SST estimate [106,127] but to what degree is not stated. Similarly, the data from Seymour Island have provisionally been aged as earliest Eocene; however, the possibility of these data being Middle Eocene age has been raised [128]. If these data are re-assigned to a Middle Eocene age it may be assumed that Early Eocene temperatures at this location would be higher. The overestimate at Seymour Island is the greatest for simulations E14–E17 from the marine dataset and this is possibly the marine data point with the largest age uncertainty. Better age constraints at Seymour Island will allow this uncertainty to be resolved in the future. In contrast the Middle Eocene age tropical data point included in the terrestrial dataset actually compared reasonably well with the warmest simulations: temperatures are approximately 1.2°C and 2.9°C warmer than the proxy temperatures in simulations E16 and E17, which are well within the published error bars for this data point. Three marine locations are consistently too cold in the Eocene simulations, all of which are high latitude EECO aged data points. The data are from the ACEX core in the Arctic Ocean at a latitude of approximately 83°N [7], Waipara river off New Zealand at a latitude of approximately 54°S [48,129] and at ODP 1172D in the southwest Pacific at a latitude of approximately 64°S [11]. The palaeo-reconstruction of all three of these locations is for a shallow marine environment, with Waipara river and ODP 1172D being coastal and from restricted environments. The ACEX data point, although in a restricted basin, is in the most open setting. There are two factors that may contribute to the underestimation of temperatures at these locations, these are the bathymetry in the model and the use of a static orbital configuration. The original references (table 4) for these proxy data identify these locations as shallow water or restricted environments with water depths of up to approximately 2000 m; however, the bathymetric reconstruction in our model is between 2000 and 3000 m at all these locations. There is evidence of orbital forcing pacing the EECO climate [89,130,131]. Previous modelling studies that have investigated orbital forcing during the Early Eocene identified the climate of the high latitude terrestrial realm as being sensitive [33,34]. Further work with a dynamic orbital configuration may reduce the model–data discrepancy with the EECO proxy data. Overall, differences between the terrestrial dataset and the simulations are much smaller than with the marine dataset, particularly in simulations E16 and E17. Temperatures at three locations in North America are consistently overpredicted by approximately 10°C by simulations E16 and E17 compared to proxy temperatures. These locations are all along the south or west coast of North America, which was mountainous during the Early Eocene. The uncertainty in orographic reconstruction in these particular locations is high and close to ±1000 m [3]. Huber & Cabellero [3] calculate the temperature uncertainty associated with orographic uncertainty in the terrestrial dataset as ±2.4°C for an uncertainty of ±450 m based on the environmental lapse rate of 5.2°C km−1 [103]. Given the larger orographic uncertainty at these locations and the coarser resolution of the land surface in our model than CCSM_H, the model this dataset was prepared for comparison with, a larger temperature error of at least ±5.2°C may be more representative here, and which provides a much improved fit between simulations and proxy data. Taking the marine and terrestrial comparisons together, of the 17 final simulations, two simulations have a more optimal fit with the Early Eocene proxy data; these simulations are E16 and E17 which are the simulations with the highest SAT. The SAT of the best performing simulations for each model in the EoMIP study range between 24.0°C and 29.5°C, with the ECHAM model (2× CO2) and HadCM3L (6× CO2) at the bottom end of this range and CCSM_H (16× CO2) at the upper end. Our two best simulations have higher SATs than the EoMIP models. The two optimal Eocene simulations are described below.
These two simulations also have marine and terrestrial EPTDs which are at the lowest end of the simulations. These simulations are much better at reproducing high latitude SH warmth than NH warmth. While neither simulation manages to replicate the high temperatures recorded in the marine EECO proxy data, the global warmth of the Early Eocene is captured. E16 and E17 do have limitations and neither fit the proxy data perfectly; however, investigating how climate processes and heat transport differ in these simulations may give us insights into understanding low polar seasonality and continental warmth during the Eocene. It should perhaps also be noted that the tropical SSTs in the warmest models are very warm. The zonal mean SST is almost 40°C and in places within the tropics it even exceeds 42°C. Such high temperatures exceed the optimum for many modern day species of ocean biological processes [132,133] such as growth. Thresholds in foraminifera with symbiotic algae have also been linked to enzyme inactivation at temperatures more than 35°C [132]. However, it should be noted that these temperatures decrease away from the equator so that by about 15°N and 15°S they are nearer 35°C. Similarly, at a depth of 50 m the temperatures have decreased to 36°C. Temperature is a strong biogeographic control on ocean biota and reduced zonation of foraminifera and poleward migration of foraminifera has been shown during the Early Cenozoic [134,135]. Similarly, the selective extinction of warm water ocean taxa during subsequent climatic cooling events such as that at the Eocene–Oligocene transition [135] indicates that modern foraminifera are not representative of greenhouse climates such as the Eocene and the possibility that species can adapt to the extreme conditions these temperatures indicate cannot be ruled out [136]. Conversely, for the EECO marine data points, the data may not be hot enough. The uncertainty associated with biological proxy data from past warm periods continues to be problematic and the omission of strong orbital forcing in our model may preclude these temperatures from being simulated in this ensemble. While it is relatively easy to analyse the reasons for the warmth in these simulations relative to the PD control climate it is more difficult to analyse the causes of the warmth between the two warmest models. If we compare simulations E16 and E17 to the PD control simulation we see that there are a number of drivers of change beyond the increase in CO2. Firstly, the relative humidity within the simulations remains relatively constant (albeit with some small decreases at high latitudes in the mid-troposphere) so that the specific humidity increases at all levels and latitudes in the warmer simulations (E16 and E17) compared with the colder simulations (E1–E15) and the PD simulation (PI), resulting in a strong positive feedback from water vapour. Secondly, the removal of land ice greatly decreases the surface albedo. However, this is not a straightforward feedback. In the colder runs, the land ice is largely replaced by heavy snowfall so that the global mean surface albedo does not change appreciably (table 5). However, in the warmer climate simulations there is a major decrease in snow cover and hence we have a strong positive feedback. Sea-ice also experiences major decreases in the warmest simulations. The planetary albedo follows a similar relationship as surface albedo, with decreased albedo with warmer temperatures. As SAT increases in the ensemble, there is also an increase in net solar radiation at the top of the atmosphere (toa) indicating increased radiative forcing. However, there are some more complicated variations from this simple pattern. Specifically simulations E12 and E15 increase their planetary albedo compared to the overall downward trend and subsequently reduce the net solar radiation toa relative to the remaining simulations. This appears to be strongly linked to changes in cloud cover. Overall, the warmer models generally have less total cloud cover which is consistent with the idea that clouds are acting as a positive feedback in these simulations. Moreover, the total cloud cover is strongly correlated with the planetary albedo (table 5). However, the patterns are quite complicated. Low clouds have a tendency to cool the climate system (through their impact on albedo) and hence the large reductions in this type of cloud in our simulations are producing a positive feedback. However, high clouds also decrease which moderates this somewhat. At higher latitudes, all types of clouds act to warm the climate system and in most of the simulations we have an increase in high latitude cloudiness. The ratio of low clouds to high clouds decreases as SAT throughout the ensemble; in E1 this ratio is 1.1 and in E16 and E17 this ratio is 0.8 and 0.9, respectively. Further complicating matters, the parameters perturbed in these simulations impact cloud physical properties such as cloud water content, cloud ice content and subsequently cloud albedo. These variables were not output in these simulations and would need to be assessed alongside any changes in cloud amounts before any definitive conclusions on the radiative balance can be drawn, particularly in relation to the processes suggested in previous studies such as PSCs [28,37,38,137] and high latitude convective cloud feedbacks [39,40,138,139]. In terms of changes in EPTD, it is also useful to examine the poleward heat transport in the simulations. Peak values of heat transport (HT) occur at approximately 40° latitude in the Eocene ensemble and in the PD control. In the PD simulation peak values of HT are 5.3 PW in the NH and 4.9 PW in the SH. In the Eocene ensemble, peak HT ranges between 5.1 and 6.0 PW in the NH and between 5.0 and 5.7 PW in the SH. At the latitude of peak HT, atmospheric heat transport (AHT) accounts for between 89 and 94% of heat transport in the NH and between 85 and 93% of heat transport in the SH. For the modern, peak values of HT are approximately 5 PW at 35° latitude, with AHT comprising 78% and 92% of the total heat transport in the NH and SH in good agreement with [140]. Ocean heat transport (OHT) peaks much closer to the equator and can be important at those latitudes but is relatively unimportant further polewards. Figure 8 shows the distribution of HT for the PD and for the warmest Eocene simulations. The ranges of OHT in the SH and of AHT in the NH are particularly large but the total variation is always dominated by the atmosphere. Figure 8. Ocean, atmosphere and total heat transport in four warmest simulations (E14–E17) and PD control simulation for each hemisphere plotted against the latitude. (Online version in colour.)
The OHT acts to transfer heat from the tropics to higher latitudes and to weaken the latitudinal temperature gradient. The correlation between tropical ocean temperatures and OHT is clearly shown in figure 9a. However, the link between OHT and EPTD is less clear (figure 9b). This is because of two reasons. Firstly, the OHT is not strong, and is almost negligible beyond about 45°N and S, and hence has its strongest effect on mid-latitudes. Most of the heat transport further polewards is performed by the atmosphere. Secondly, the link between total heat transport and EPTD is also complicated because the albedo varies between the simulations. This implies that the total heat transport required to maintain the gradient will also vary [141]. Figure 9. OHT calculated as a per cent of total heat transport in each hemisphere for the Eocene simulations and plotted against (a) tropical SSTs and (b) EPTD. NH data plotted as diamonds and SH data plotted as squares. R2 correlation coefficients are also shown. (Online version in colour.)
Of the 15 MPP simulations in the final 17 Eocene simulations, only one parameter set (from E17) surpasses the control parameter set for a PD simulation in the PD (see table 2 for criteria used to assess PD simulations). The PD equivalent of the Eocene MPP simulation E11 is ranked only one place behind the control simulation in the PD ensemble. The control parameter set, however, does not make it into the final ensemble of Eocene simulations. For all of the 15 final Eocene MPP simulations, we have an equivalent PD control and 2× CO2 concentration simulation. Thus, it is possible to calculate the Charney climate sensitivity for these parameter sets. The Charney climate sensitivity is broadly defined as the equilibrium global mean surface temperature change following a doubling of CO2 concentration. The mean ± 1 s.d. value for Charney sensitivity for the 18 models assessed in the Fourth Assessment Report (AR4) of the IPCC [142] was 3.26°C±0.69°C. Of the subset of our 15 simulations that are run at 2×CO2 concentration the mean ± 1 s.d. value for the Charney sensitivity is 3.25°C±0.58°C, which is very similar to the AR4 mean value. The Charney sensitivity for our best model, E17, is calculated to be 2.7°C, which is below this mean estimate. CCSM3, which was used for the Eocene simulations by [3,46], has a PD climate sensitivity of 2.7°C, while HadCM3 the sister model of FAMOUS has a Charney sensitivity of 3.3°C. Thus, our best performing parameter set for the Eocene, and which was able to simulate the extreme warmth of the Eocene, actually has a reduced climate sensitivity compared to the control parameter set, and a very similar climate sensitivity to CCSM. Moreover, Gregoire et al. [61] use the same 100 MPP parameter sets in their tuning study which focused on the PD and the LGM. Simulation S4 which is highlighted in their study as having a favourable fit has identical parameters to our Eocene simulation E17. Our work is the first attempt at a comprehensive ensemble with perturbed climate-sensitive model parameters for the Early Eocene. The results show that we can get a large diversity in response, with global mean temperature changes which vary considerably, from temperatures that are slightly cooler than the modern to temperatures that are extremely warm. We have managed to simulate levels of warmth comparable to that of the Early Eocene at only twice pre-industrial CO2 which is a much lower concentration than used by many other models. Although many aspects contribute to this warmth, a strong sensitivity to albedo changes associated with cloud cover was apparent. Clouds remain one of the most uncertain aspects of climate modelling with little consensus over the sign of the cloud feedback. In this work, the choice of perturbed parameters affected the physical properties of the clouds. The physical properties of the clouds and the effect on radiative balance will be examined in future work. Within the ensemble, as SAT increases OHT decreases in both the NH and SH. In the SH as tropical SSTs increase and polar SSTs increase this also correlates to a reduction in OHT. However, this relationship is not apparent when OHT and the EPTD are compared across the ensemble. This implies that OHT is not a major control on the EPTD. If OHT is not a major part of the EPTD, AHT and local radiative effects are likely to be involved in driving changes in the EPTD. Proxy–model discrepancies are larger in the marine dataset than the terrestrial dataset. Simulation of the marine EECO temperatures is the most problematic, with the warmest simulation still 12°C too cool compared to the proxy Tex86 temperature estimated. Some of this temperature difference may be attributable to the use of a modern orbital forcing in these simulations. There is evidence for a strong precessional and eccentricity signal pacing the EECO; all the EECO data used in this study are from the high latitudes and previous studies indicate that the high latitudes are most sensitive to orbital forcing during the Eocene [33,34] and other periods [143,144]. It has been known for some time that perturbing the parameters of models can result in a wide spread of results. However, one of the most exciting aspects of our results is that the ‘best’ climate simulation for the Early Eocene was also one of the best simulations for the PD and LGM. For the Early Eocene, our results have to be partly tempered by the uncertainty in boundary conditions, particularly the lack of a precise indicator of past greenhouse gas concentrations. Therefore, we may be obtaining a good comparison to data for the wrong reasons. When we apply this parameter set to a future climate change simulation, we find that the resulting temperature increase due to an instantaneous doubling of CO2 (the so-called Charney climate sensitivity) is 2.7°C. This value is slightly below the mean estimates of Charney sensitivity from the AR4 [142]. This is perhaps surprising since there have been indications [145] that palaeoclimate data would imply that models were under sensitive. Our new results show that it is possible that a model can respond strongly to past changes without it necessarily resulting in a high sensitivity to future changes. Palaeoclimate research focused on comparing proxy data to models will never be able to ‘prove’ that climate models work. However, it does provide a unique test of models’ ability to simulate climates different to present. It is worth bearing in mind that even with an optimal choice of parameters there will be irreducible structural deficiencies in the model that cannot be mitigated. However, it is still very encouraging that a single model parameter set exists that results in a model that simulates well the PD, LGM, and Early Eocene. The authors would like to thank the two anonymous reviewers for particularly helpful suggestions and comments. This work was carried out using the computational facilities of the Advanced Computing Research Centre, University of Bristol, http://www.bris.ac.uk/acrc. The Eocene palaeogeography was created by Fugro-Roberstson. The simulations described in this work are available at the following website: http://www.bridge.bris.ac.uk/resources/simulations. FootnotesOne contribution of 11 to a Discussion Meeting Issue ‘Warm climates of the past—a lesson for the future?’. © 2013 The Author(s) Published by the Royal Society. All rights reserved. References
Page 4Since the beginning of the industrial era, anthropogenic emissions of carbon dioxide (CO2) from fossil fuel burning and, to a lesser extent, land-use change and cement manufacturing have increased the concentration of CO2 in the Earth's atmosphere by approximately 40%. The combined fossil fuel and cement emissions reached a record high in 2010 of 9.1 Pg C yr−1 (1 Pg=1015 g) [1], higher than predicted 20 years ago under business-as-usual scenarios for the year 2010 (8.7 Pg C yr−1, IS92a scenario) [2]. The rapidly rising levels of CO2 in the atmosphere are altering the radiative forcing of the Earth's climate, which, until recently, has been the sole focus of the scientific and public discussion. A second impact of anthropogenic CO2 emissions is ocean acidification, which refers to the ongoing decline in ocean pH and the reduction in the ocean's carbonate mineral saturation state, with possible negative consequences for marine life [3–5]. Other geochemical and physical consequences of an increasingly acidic ocean include effects on metal speciation, reduced NH3/NH4+ ratios (probably affecting ammonia oxidation rates), the marine source of atmospherically active trace gases and alteration of underwater sound absorption [6–9]. Projections of future CO2 emissions and attendant modifications of climate and ocean chemistry have typically focused on the century time scale, most notably until the year 2100 [10]. However, from a geological perspective, the longer term consequences of the carbon released by human activities may be considered equally, if not more, important. For instance, on millennial time scales, total emissions of 5000 Pg C are projected to increase the Earth's global surface temperature by more than 8°C and drop surface ocean pH by approximately 0.7 units (figure 1). A carbon release of this rate and magnitude represents a massive perturbation to the Earth system, most probably unprecedented during the past 56 million years [12–14]. The climatic and geochemical recovery will take tens to hundreds of thousands of years well after emissions have ceased [15]. Biotic recovery in terms of diversity and ecosystem functioning may take millions of years [16]. However, owing to the complexity of the Earth system, particularly involving the contribution of physical and biogeochemical feedbacks, the precise details of the future response is difficult to predict. In this regard, the geological record may provide foresight to what the future will hold for the Earth's climate, ocean chemistry and ecosystems. Figure 1. Consequences of anthropogenic carbon release for various CO2 emission scenarios [4]; tR=release time. Simulations were performed with the LOSCAR model [11]. (Online version in colour.)
The closest analogue for a massive carbon release in the past is the Palaeocene–Eocene Thermal Maximum (PETM, approx. 56 Ma). This event is characterized by a transient global warming of 6°C, with a relatively rapid onset and gradual recovery over 150 kyr [12,17–20]. The onset was accompanied by intense dissolution of carbonate sediments throughout the deep sea as well as an anomalous excursion in the ratio 13C/12C of the surficial carbon reservoir, i.e. ocean, atmosphere and biosphere [17,21]—phenomena which could have been generated only by a rapid and massive release of carbon, causing ocean acidification. Although the surface ocean appears to have remained oversaturated, communities of marine calcifiers, primarily coralgals, phyto- and zooplankton and benthic foraminifera, experienced changes in both diversity and abundances [22–28]. While many species ultimately survived, the community perturbations persisted for tens of thousands of years, recovering only as carbon levels abated and the planet cooled. Numerical models demonstrate that the scale of seafloor carbonate dissolution and 13C/12C excursion can be simulated only with the release of thousands of Pg C, and most of it in less than 5–10 kyr [29–31]. These simulations also show that the long tail of the atmospheric lifetime of this carbon should have exceeded 150 kyr, a result that is consistent with the actual duration of the PETM and ocean acidification. However, using first-order assumptions, the models predict that the main phase of high pCO2 and intense warming should have faded after a few ten thousand years (first-order assumptions signify a simple, single carbon input pulse over a few thousand years). In order to explain the prolonged warming over a time scale of hundred thousand years, additional assumptions are necessary such as continuous, prolonged carbon input over tens of thousands of years [32]. In this paper, we discuss the long-term legacy of massive carbon input to the Earth system, mainly focusing on the Anthropocene and the Early Eocene, and implications for the future. Our aim is not to constrain the PETM carbon input mass, which is discussed elsewhere [29–31,33–35], but to study the long-term legacy of massive carbon input. To this end, we focus on a limited number of carbon input scenarios [31] and employ the LOSCAR (Long-term Ocean–atmosphere–Sediment CArbon cycle Reservoir) model as a tool to illustrate various carbon cycle processes. The LOSCAR model is described in detail in the study of Zeebe [11]. The known total fossil fuel reserves (currently available for combustion) have been estimated at several thousands of Pg C. These figures do not include potential contributions from other fossil resources such as methane hydrates. For total carbon emissions of 3500 and 5000 Pg C over 500 years, the Earth's surface temperature would rise by more than 6°C and 8°C, respectively, during the next few centuries (figure 1). This estimate assumes a climate sensitivity of 3°C per doubling of CO2, which includes only fast feedback processes [10]. However, over millennial time scales, additional, slower feedbacks could become active, which would exacerbate the warming [36–38]. The projected consequences for ocean chemistry are equally severe, with a decline in ocean pH by up to approximately 0.7 units (from approx. 8.2 to 7.5, a fivefold increase in acidity or H+ concentration) and a two- to threefold reduction in the carbonate mineral saturation state (figure 1) [4]. To place this in a geological perspective, surface ocean pH has probably not been below approximately 8.1 during the past 2 Myr [39]. A range of simulations show that, in order to avoid large changes in the Earth's climate and ocean chemistry, drastic and immediate reductions in CO2 emissions would be necessary (figure 1). For instance, in order to limit the total carbon input to 1000 Pg C and stretch emissions over 500 years, global carbon emissions would need to be cut in half over the next 30 years, starting tomorrow. Projections of future changes in ocean carbonate chemistry are relatively robust and largely model-independent on a time scale of a few centuries, mainly because the chemistry of CO2 in seawater is well known and because changes in surface ocean carbonate chemistry closely track changes in atmospheric CO2 [15,40,41]. However, the climatic and biotic response is far more difficult to forecast because of the complexity of the climate system, ecosystem dynamics and biogeochemical feedbacks [42]. One way to improve our predictions of the Earth-system response to massive and rapid carbon release is to look to the past. The PETM as an extreme and transient event that caused widespread environmental change is probably the best analogue for a massive carbon release in the geological past, for which a sufficient number of widely distributed sediment records are available [18–20,43]. One critical element for a comparison between the Anthropocene and the PETM is the time scale of carbon input. While it is clear that the carbon input during the PETM was rapid on geological time scales (a few thousand years), establishing the approximate rate of emissions has proved difficult using conventional stratigraphic methods [33,34]. Given the limitations of stratigraphy, numerical tools are required to provide additional constraints on the time scale of the PETM carbon release, for example by using carbon cycle models that include a sediment component [11,31]. Simulations of the carbon release with a single input of 3000 Pg C (source δ13C=−50‰) indicate that the release time was probably much shorter than 20 kyr, otherwise the shoaling of the calcite compensation depth (CCD) in the deep Atlantic would be too muted (figure 2). Observations across the Palaeocene–Eocene boundary (PEB) have shown that the CCD shoaled substantially in the Atlantic and by at least 2.0 km in the South Atlantic [13,44–46]. Hence the simulations suggest that the release time was approximately 6 kyr or less for an initial input of 3000 Pg C (figure 2). Note that the simulations assume a 40% carbon release directly into the deep Atlantic from the possible dissociation of methane hydrates [47]. If the carbon was injected entirely into the model's atmosphere, the Atlantic CCD shoaling would be less, calling for an even shorter release time [48]. Note also that the CCD shoaling in the Pacific was less pronounced than in the Atlantic [31,34,49]. At input rates over periods approaching approximately 1 kyr, the model predicts a large but short-lived total carbon isotope excursion (CIE) in the surface ocean of up to −6‰. However, this anomaly quickly returns to the long-lived CIE, which slowly decays from a peak value of about −3.5‰ at approximately 3 kyr after the PEB (figure 2). The reason for the short-lived δ13C anomaly is that, on time scales shorter than approximately 1 kyr, the source carbon has not yet been mixed throughout the entire deep ocean, which leaves the atmosphere and surface ocean disproportionately depleted in 13C (relative to the total exogenic carbon pool). So far, such an anomaly has not been found in sediment records [50], which would argue against a release time shorter than approximately 1 kyr. However, at this stage, it is not clear whether it is even possible to observe such an anomaly given the fidelity of even the highest resolution marine/terrestrial sediment records. Figure 2. Effect of releasing 3000 Pg C over various time intervals during the PETM [11,31]. The source carbon has a δ13C value of −50‰; 40% of the carbon was injected into the deep Atlantic. Note that the Pacific CCD shoaling was much less pronounced than in the Atlantic [31]. (Online version in colour.)
All deep-sea carbonate PEB sections are condensed to varying degrees as a consequence of the acidification/carbonate dissolution pulse. A variety of conventional and unconventional strategies have been applied to estimate the duration of the condensed intervals in pelagic sections including orbital stratigraphy and relative abundances of extraterrestrial 3He, a constant flux proxy (figure 3) [51–53]. While the overall duration of the excursion and recovery have been well constrained, both approaches lack the precision to unambiguously constrain the duration of the onset in these condensed sequences to ±10 kyr. Alternatively, carbon isotope data for populations of individual shells from closely spaced samples across the boundary throughout the ocean yield clear bimodal distributions of shells recording pre-excursion and excursion carbon isotope values, but with no transitional values, suggestive of an abrupt shift in surface water δ13C [22,54], though this also could be an artefact of dissolution. Expanded shallow marine, siliciclastic sections, on the other hand, lack the needed stratigraphic control to constrain rapid changes in sediment accumulation, and thus yield conflicting results for the initial onset of the CIE, with estimates ranging from just a few thousand years to as long as 20 kyr [19,33,50,55]. In sum, the rate of carbon release is still insufficiently constrained to eliminate the possibility of a relatively fast release, of the order of a few thousand years, or for a more gradual release, interrupted by one or more rapid pulses. Figure 3. Two estimates of the duration of the CIE and CaCO3 dissolution event at Ocean Drilling Program (ODP) Site 1266, Walvis Ridge, in the South Atlantic [13]. One estimate is based on orbital cycle stratigraphy [51], the other on extraterrestrial 3HeET concentrations [52]. The latter assigns a greater duration to the dissolution interval and a shorter duration to the recovery interval. The lower two panels show the changes in carbonate and non-soluble fractions as measured by Murphy et al. [52] using just the 3HeET age constraints. We note that an undetermined portion of the clay layer (0% CaCO3) represents Upper Palaeocene material deposited prior to the PETM/CIE and thus adds (10–30 kyr) to the total duration of the event. (Online version in colour.)
Carbon cycle models that include a weathering feedback predict an ‘overshoot’ of the CCD in the aftermath of the carbon release. That is, a few ten thousand years after the carbon input has stopped, the position of the CCD is deeper than its position before the event and remains suppressed on a time scale of 100 kyr or more (figure 2d). Note that, while figure 2d shows examples for the Atlantic CCD, the model-predicted CCD overshoot is global [11,31]. The cause for the CCD overshoot can be traced back to the weathering feedback. Immediately after the carbon input has ceased, atmospheric pCO2 is still elevated over the initial pCO2 (figure 2b), which causes enhanced weathering of carbonate and silicate rocks on the continents. The enhanced weathering produces an influx of calcium and carbonate ions to the ocean that exceeds the removal of these ions as CaCO3 because the burial is reduced at that point owing to the diminished carbonate mineral saturation state of the ocean. As a result, the excess weathering flux subsequently begins to raise the ocean's saturation state and deepens the CCD until a quasi-steady state of riverine flux and burial has been established. The quasi-steady state on a hundred thousand year time scale must be maintained at a deeper CCD than initially (because of enhanced influx and burial) until atmospheric pCO2 and weathering fluxes return to their initial steady-state values on a million year time scale. This process slowly removes the excess carbon from the system via silicate weathering. In general, the model-predicted oversaturation and CCD overshoot are in agreement with observations [13,49,52,53,56,57]. The observations include an unusual transient pulse (20–40 kyr) in carbonate accumulation rates during the recovery phase, roughly 100 kyr after peak acidification (figure 3) [52,53] as well as enhanced preservation of plankton shells (figure 4; phase II) [57]. These observations, recorded in all ocean basins and at all depths, indicate that over much of the ocean the entire water column was highly oversaturated. The highly oversaturated surface waters might have contributed to blooms of coccolithophores dominated by just a few species documented at a number of locations [58,59]. Unfortunately, attempts to locate deep-sea sections that were positioned just below the CCD prior to the PETM, and thus might have recorded the transient overshoot, have yet to be successful. Figure 4. The pelagic sediment evidence for ocean acidification during the PETM. (a) Per cent calcite (%CaCO3) showing the dissolution horizon and (b) weight-per cent coarse fraction (wt% CF) records for three sections from Walvis Ridge (Sites 1262, 1263 and 1266) and one from the Weddell Sea (Site 690) [57]. The age model is based on cycle (orbital) stratigraphy [51]. The CF primarily comprises planktonic foraminifera shells, which are highly susceptible to solution, and thus wt% CF represents a qualitative indicator of deep-sea saturation state. The acidification phase is represented in the lower most part of the CIE by the minima in both %CaCO3 and %CF. The period of oversaturation is represented by the relatively uniform %CaCO3 and CF values in phase II of the recovery, as well as by the overall low %CF which is a consequence of enhanced production and flux of coccoliths which are predominantly in the less than 30 μm fraction, thus diluting the more than 63 μm fraction. (Online version in colour.)
The magnitude of the CCD overshoot at The term ocean acidification commonly refers to the ongoing decrease in ocean pH owing to the ocean's uptake of anthropogenic CO2. Over the period from 1750 to 2000, the oceans have absorbed approximately one-third of the CO2 emitted by humans; this absorption has already caused a decrease in surface ocean pH by approximately 0.1 units from approximately 8.2 to 8.1 [61]. In a more general sense, ocean acidification may also refer to a decrease in ocean pH owing to other causes and to time scales that are not limited to the present or near future. However, the phrase ocean ‘acidification event’ should be used in the context of the Earth's history to describe an episode that involved geologically rapid changes of ocean carbonate chemistry on time scales of less than 10 000 years [62,63]. For instance, the decline in surface ocean pH and CaCO3 saturation state (Ω) is coupled on these time scales in response to carbon input. By contrast, on long time scales (greater than 10 000 years), the saturation state of the ocean is generally well regulated by the requirement that CaCO3 sources (weathering) and sinks (shallow- and deep-water CaCO3 burial) must balance [64,65]. The present acidification of the oceans due to anthropogenic CO2 emissions is expected to have negative consequences for a variety of marine organisms [3–5]. For example, a decline in carbonate saturation state will affect stability and production rates of CaCO3 minerals, which constitute the building blocks of coral reefs and the shells and skeletons of other marine calcifying groups. Laboratory and mesocosm studies indicate that a decrease of 0.2–0.3 units in seawater pH inhibits or slows calcification in many marine organisms, including corals, foraminifera and some calcareous plankton. Note that a drop of 0.3 pH units corresponds to a doubling of the hydrogen ion concentration ( For the PETM, a number of carbon input scenarios have been proposed with masses ranging from 1100 to more than 10 000 Pg C over durations of a few thousand to tens of thousands of years [30,31,33,47]. However, initial estimates with very low carbon input mass may have underestimated the magnitude of the CIE and hence the total carbon input [47]. The high-end scenarios with very large carbon input mass require certain assumptions about the CCD before the event and/or predict deep-sea carbonate dissolution patterns during the event that seem difficult to reconcile with the sediment record [31,34,35,43]. Moreover, the mechanism (i.e. source) for such a large and rapid carbon emission is problematic. The scenario that we favour requires an initial carbon pulse of about 3000 Pg C over approximately 6 kyr in order to be consistent with the timing and magnitude of stable carbon isotope records and deep-sea dissolution patterns [31]. We have compared this PETM scenario with a business-as-usual scenario of fossil fuel emissions of 5000 Pg C over approximately 500 years (figure 5). Our results show that if the proposed PETM scenario roughly resembles the actual conditions during the onset of the event, then the effects on ocean chemistry, including surface ocean saturation state, were less severe during the PETM than expected for the future [65,66]. As shown by Zeebe et al. [4], not only the magnitude but also the time scale of the carbon input is critical for its effect on ocean carbonate chemistry. The time scale of the anthropogenic carbon input is so short that the natural capacity of the surface reservoirs to absorb carbon is overwhelmed (figure 1). As a result of a 5000 Pg C input over approximately 500 years, the surface ocean saturation state of calcite (Ωc) would drop from about 5.4 to less than 2 within a few hundred years. By contrast, the PETM scenario suggests a corresponding decline of Ωc from 5.5 to only about 4 within a few thousand years. Note, however, that the PETM scenario may be subject to revision, depending on the outcome of future studies that will help to better constrain the time scale of the carbon input. Figure 5. Comparison of the effects of anthropogenic business-as-usual emissions (total of 5000 Pg C over 500 years) and PETM carbon release (3000 Pg C over 6 kyr) on the surface ocean saturation state of calcite. (Online version in colour.)
The premise that the PETM carbon input had only a moderate long-term impact on the surface ocean saturation state is consistent with the findings on nannoplankton origination and extinction during the PETM, which indicate that the perturbation of the surface water saturation state across the PETM was not detrimental to the long-term survival of most species of calcareous nannoplankton taxa [67,68]. Still, transient anomalies in coccolithophore diversity and abundances have been documented globally at the onset of the event and have been attributed to factors such as reduced fertility and warming, while the contribution of acidification remains unclear [23,28,58,59,67–70]. Similarly, planktonic foraminifer communities at low and high latitudes show reductions in diversity, invasions of warmer water or excursion taxa, but no obvious evidence of severe undersaturation [22,71]. One shallow-water carbonate record from a Pacific Ocean guyot shows no major evidence for a permanent carbonate production crisis after the PETM, indicating that the effects of any changes in temperatures or surface ocean pH may have been relatively short-lived or relatively minor [72]. For calcifiers residing deeper in the ocean, the impact of the PETM was much more severe, for example, with a major extinction event of benthic foraminifera, affecting 30–50% of species globally [25]. It is not clear, however, whether the benthic extinction was caused by changes in oxygenation, bottom water temperatures, carbonate undersaturation as a result of the carbon input, and/or other factors [25,65]. Finally, a growing body of evidence suggests that coastal coral reef and ostracode communities experienced a significant reduction in diversity at the end of the Palaeocene [27,73], though the exact role of acidification has yet to be firmly established. In sum, it appears that the direct effects of ocean acidification on marine planktonic calcifiers during the PETM may have been limited because of a relatively ‘slow’ carbon input rate (slow on human time scales, rapid on geological time scales). However, conclusions are premature at this stage as the number of studies addressing acidification effects on pelagic calcifiers during the PETM is still very limited. The impacts on coastal marine calcifiers, on the other hand, might have been fairly significant. Yet, additional studies are also desirable in this area for a more comprehensive analysis of ocean acidification effects on marine organisms during the PETM. The lifetime of fossil fuel CO2 in the atmosphere has been inadequately addressed by many studies and reports, including the Intergovernmental Panel on Climate Change [74]. The fundamental difference between CO2 and other greenhouse gases such as methane is that the decrease in atmospheric CO2 over time does not follow a simple decay pattern of a single exponential—even after several millennia, a substantial fraction of the CO2 remains in the atmosphere [11,15]. Fossil fuel neutralization involves various processes that operate on different time scales. The steps include ocean uptake, mixing with surface waters and reaction with dissolved carbonate ions (10–102 years), transport and mixing throughout the deep ocean (102–103 years), reaction of CO2 with deep-sea carbonate sediments (102–104 years) and long-term neutralization via weathering of carbonate and silicate minerals on the continents (104–106 years). For example, for a rapid pulse of 1000 and 5000 Pg C injected into the atmosphere, the airborne fraction as calculated by various models is still approximately 20% and 50%, respectively, after 1000 years, and approximately 15% and 20%, respectively, after 10 000 years [15]. Very similar results have been obtained with the LOSCAR model used in this study, where we use the LOSCAR model as a tool to illustrate carbon cycle processes; for a detailed model description, see [11]. For anthropogenic emissions of 5000 Pg C stretched over 500 years (rather than a pulse; see figure 1), the LOSCAR model predicts a maximum pCO2 of approximately 1900 μatm, which declines to approximately 600 μatm after 10 000 years (t=0 here refers to the onset of industrialization; see figure 6). Given a pre-industrial initial pCO2 of 280 μatm, the airborne fraction is hence 20% after 10 000 years, in agreement with the suite of models tested by Archer et al. [15]. After 50 kyr, atmospheric CO2 has dropped below approximately 500 μatm (airborne fraction less than 14%). This number is somewhat sensitive to the choice of parameter values used in the weathering parametrization [75]. However, the LOSCAR model's standard configuration uses a relatively weak weathering feedback. A stronger weathering feedback would produce a smaller airborne fraction after 50 kyr. In summary, state-of-the-art carbon cycle models predict that the long tail of the atmospheric lifetime of fossil fuel CO2 is tens to hundreds of thousands of years. However, the airborne fraction of the initial carbon input should drop substantially over a period of 10–20 kyr. Figure 6. Long-term legacy of massive carbon input to the Earth system: Anthropocene versus PETM. (a) Fossil fuel emissions: total of 5000 Pg C over 500 years. (b) PETM carbon release: 3000 Pg C over 6 kyr plus approximately 1500 Pg over more than 50 kyr. Note different y-axis scales in (a,b). (c) Simulated evolution of atmospheric CO2 in response to the carbon input using the LOSCAR model [11,31]. (Online version in colour.)
On the contrary, PETM records indicate little if any decline in, for instance, δ13C values after 50 kyr (figure 3). Similar durations of the PETM main phase can be inferred from δ18O records (indicating temperature) and surface ocean carbonate chemistry proxies [76]. The inferred main phase duration of greater than 50 kyr is also independent of the age model applied (figure 3). One age model is based on orbital cycle stratigraphy [51], the other on extraterrestrial 3HeET concentrations [52]. An undetermined portion of the clay layer represents Upper Palaeocene material deposited prior to the PETM/CIE and thus adds to the total duration of the event (10–30 kyr). Nevertheless, the two age models agree that the duration of the PETM main phase lasted for at least 50 kyr, a duration that is also consistent with observations from the most expanded terrestrial sequences [77,78]. Based on first-order assumptions of a single carbon input over several thousand years, carbon cycle models predict that the main phase of high pCO2 and intense warming should have faded after a few ten thousand years (compare 6 kyr scenario in figure 2). This behaviour is consistent with the results of the fossil fuel experiments, but inconsistent with the PETM reconstructions. Hence, additional assumptions are required to explain the observed greater than 50 kyr duration of the PETM main phase. For example, we have proposed a PETM scenario that assumes an additional, continuous carbon input of approximately 1500 Pg C over 70 kyr with a δ13C value of −50‰ (figure 6) [31,32]. While the total amount of the additional carbon ‘bleeding’ is significant, the annual rate of approximately 0.02 Pg C yr−1 is modest. For comparison, natural long-term weathering fluxes are of the order of 0.2 Pg C yr−1; fossil fuel carbon emissions in 2010 were 9.1 Pg C yr−1 [1]. Possible causes for the prolonged carbon input could include additional slow dissociation of clathrates in response to continued warming of subsurface sediments 32 and/or terrestrial carbon feedbacks that release carbon under intense greenhouse conditions (i.e. shrinking of soil organic carbon reservoirs). As of yet, most of these feedbacks are not well understood. It seems imperative to identify and thoroughly understand these feedbacks as similar processes could lead to unpleasant surprises in the future 38. The fossil record indicates that recovery of biotic diversity after mass extinctions generally takes several million years. For example, biotic diversity after major extinction events throughout the Phanerozoic required of the order of 5 Myr to rebound [16,79–82]. These events include the Late Ordovician approximately 450 Ma, Late Devonian approximately 370 Ma, End-Permian approximately 250 Ma, End-Triassic approximately 200 Ma and End-Cretaceous approximately 65 Ma, which have traditionally been labelled the ‘big five’ extinctions. However, more recent studies point out that perhaps only three events qualify as true global mass extinctions, among them the End-Permian and End-Cretaceous [16,81]. It took 10–15 Myr after the End-Permian for coral reefs to recover and approximately 2 Myr after the Cretaceous–Tertiary (K-T) boundary for corals to leave a trace in the fossil record [82,83]. Pre-existing levels of coral diversity were established only about 10 Myr after the K-T boundary. Geochemical evidence such as surface-to-deep gradients in δ13C suggests that marine export production was severely suppressed after the K-T event for approximately 0.5 Myr, most probably because of the extinction of grazers [84,85]. Yet, there is little evidence that the K-T impact led to a sterile ocean devoid of life, commonly termed ‘strangelove ocean’ in the literature [64,86]. While the PEB marks a major extinction event of benthic foraminifera, affecting 30–50% of species globally, and the decline of coralgal reefs [25,27,73], most species of calcareous nannoplankton and zooplankton taxa appear to have survived the PEB (see discussion above). Also, terrestrial species experienced only minor extinction [87]. However, the PETM triggered major reorganization and dispersal of animals, particularly in mammals [21,88,89], which also experienced a reduction in mean body size, probably in response to warming or less nutritious vegetation [90]. Plants experienced a major, but temporary, reorganization and drop in diversity related to changes in climate, particularly precipitation [91,92]. In essence, the impacts on biota were largely transient in nature on geological time scales, but long on human time scales. As discussed earlier, parallels exist between the Anthropocene and the PETM in terms of carbon input and climate change. Does this also imply similar impacts in terms of species extinction and recovery? We argue that the Anthropocene will more likely resemble the End-Permian and End-Cretaceous catastrophes, rather than the PETM. First, the present extinction rate of the Anthropocene is more than 100 species per million species per year, while the fossil record indicates background extinction rates of marine life and mammals of 0.1–1 and 0.2–0.5 species per million species per year, respectively [93]. In other words, the current rate of species extinction is already 100–1000 times higher than would be considered natural. The causes for the current extinctions are diverse, including factors such as changes in land use and fresh water, pollution, exploitation of natural resources, etc. Second, with respect to ocean acidification and impacts on marine calcifiers, the anthropogenic carbon input rate is most probably greater than during the PETM, causing a more severe decline in ocean pH and saturation state (figure 5). In addition, changes in ocean chemistry and sea surface temperature will be imposed on ecosystems that are already affected by other environmental factors. Analysis of the marine fossil record suggests that, if the Anthropocene mass extinction rivals the K-T or End-Permian disasters, recovery will take tens of millions of years [16]. At this point, there are obviously large uncertainties regarding the progression of the rate of extinction and origination, dispersal and success of species in the future. However, if the current trend of species extinction continues, the geological record tells us that humans will have a major and long-lasting impact on the evolution of species on this planet for millions of years to come. We have discussed the long-term legacy of massive carbon release into the Earth's surface reservoirs, focusing on the Anthropocene and the PETM. The comparison of the rate of carbon release suggests that the ensuing effects on ocean acidification and marine calcifying organisms will probably be more severe in the future than during the PETM. However, firm conclusions are difficult to draw at this stage because (i) current research shows mixed responses to acidification in some calcifying taxa and (ii) the number of studies addressing acidification effects on pelagic calcifiers during the PETM is still very limited. The observed duration of the PETM appears to be much longer than predicted by models using first-order assumptions, which poses a conundrum. One explanation involves prolonged, additional carbon release, for instance from marine gas hydrate systems [32]. In this regard, additional observational constraints on the CCD before, during and after the PETM main phase are required in the South Pacific, Indian and North Atlantic Ocean. To be of more practical use, these observational constraints should be placed within a robust chronostratigraphic framework that includes, if possible, the long-term background variability (on orbital time scales) immediately preceding and following the PETM. Ultimately, such a framework will help to constrain the carbon release during the PETM. One important task for the modelling community is to focus on simulating carbonate sediment accumulation profiles across the PEB, including carbon isotopes and other sediment/porewater tracers (e.g. calcium, boron). Among other things, this will help to account for the effects of dissolution and sediment mixing on carbon isotope profiles. It is also important to recognize that the PETM is part of a series of hyperthermals superimposed on a long-term warming trend from the Late Palaeocene to the Early Eocene Climatic Optimum. Throughout this interval, carbon isotope ratios gradually drop by about 2‰, while deep-sea carbonate records indicate a long-term deepening of the CCD. Reconciling the character and origin of the multi-million year trend in both the climate system and carbon cycle will aid with setting the baseline state (boundary conditions) for the hyperthermals in models, and thus in identifying potential triggers and feedbacks. In terms of past and future mass extinctions and recovery times of biotic diversity, we have argued that the Anthropocene will more likely resemble the End-Permian and End-Cretaceous disasters, rather than the PETM. If civilization is to avoid such a fate, carbon emission rates must reverse within the next few decades in order to keep total emissions below a certain limit. Note that, while the short-term effects of massive carbon release are modulated by the release time, the long-term legacy is primarily determined by the total integrated emissions. Yet, if the current trend in carbon emissions continues, humans will—given sufficient fossil fuel reserves—release several thousand Pg of carbon, with severe consequences for climate, ocean chemistry, biota, etc., as discussed above. This underlines the urgency for immediate action on global carbon emission reductions and sequestration. We thank the organizers (Dan Lunt, Harry Elderfield, Andy Ridgwell and Rich Pancost) and the Royal Society for hosting a great meeting on ‘Warm climates of the past: a lesson for the future?’ in October 2011 in London. Editor Andy Ridgwell, reviewer Jerry Dickens and one anonymous reviewer provided comments that improved the manuscript. Am 30. Mai ist derWeltuntergang. This research was supported by NSF grant nos. OCE09-02869 to J.C.Z. and R.E.Z. FootnotesOne contribution of 11 to a Discussion Meeting Issue ‘Warm climates of the past—a lesson for the future?’. © 2013 The Author(s) Published by the Royal Society. All rights reserved. References
Page 5Humanity is now the dominant force driving changes in the Earth's atmospheric composition and climate [1]. The largest climate forcing today, i.e. the greatest imposed perturbation of the planet's energy balance [1,2], is the human-made increase in atmospheric greenhouse gases (GHGs), especially CO2 from the burning of fossil fuels. Earth's response to climate forcings is slowed by the inertia of the global ocean and the great ice sheets on Greenland and Antarctica, which require centuries, millennia or longer to approach their full response to a climate forcing. This long response time makes the task of avoiding dangerous human alteration of climate particularly difficult, because the human-made climate forcing is being imposed rapidly, with most of the current forcing having been added in just the past several decades. Thus, observed climate changes are only a partial response to the current climate forcing, with further response still ‘in the pipeline’ [3]. Climate models, numerical climate simulations, provide one way to estimate the climate response to forcings, but it is difficult to include realistically all real-world processes. Earth's palaeoclimate history allows empirical assessment of climate sensitivity, but the data have large uncertainties. These approaches are usually not fully independent, and the most realistic eventual assessments will be ones combining their greatest strengths. We use the rich climate history of the Cenozoic era in the oxygen isotope record of ocean sediments to explore the relation of climate change with sea level and atmospheric CO2, inferring climate sensitivity empirically. We use isotope data from Zachos et al. [4], which are improved over data used in our earlier study [5], and we improve our prescription for separating the effects of deep ocean temperature and ice volume in the oxygen isotope record as well as our prescription for relating deep ocean temperature to surface air temperature. Finally, we use an efficient climate model to expand our estimated climate sensitivities beyond the Cenozoic climate range to snowball Earth and runaway greenhouse conditions. The Cenozoic era, the past 65.5 million years (Myr), provides a valuable perspective on climate [5,6] and sea-level change [7], and Cenozoic data help clarify our analysis approach. The principal dataset we use is the temporal variation of the oxygen isotope ratio (δ18O relative to δ16O; figure 1a right-hand scale) in the shells of deep-ocean-dwelling microscopic shelled animals (foraminifera) in a near-global compilation of ocean sediment cores [4]. δ18O yields an estimate of the deep ocean temperature (figure 1b), as discussed in §3. Note that coarse temporal resolution of δ18O data in the intervals 7–17, 35–42 and 44–65 Myr reduces the apparent amplitude of glacial–interglacial climate fluctuations (see electronic supplementary material, figure S1). We use additional proxy measures of climate change to supplement the δ18O data in our quantitative analyses. Figure 1. (a) Global deep ocean δ18O from Zachos et al. [4] and (b) estimated deep ocean temperature based on the prescription in our present paper. Black data points are five-point running means of the original temporal resolution; red and blue curves have a 500 kyr resolution. Coarse temporal sampling reduces the amplitude of glacial–interglacial oscillations in the intervals 7–17, 35–42 and 44–65 Myr BP.
Carbon dioxide is involved in climate change throughout the Cenozoic era, both as a climate forcing and as a climate feedback. Long-term Cenozoic temperature trends, the warming up to about 50 Myr before present (BP) and subsequent long-term cooling, are likely to be, at least in large part, a result of the changing natural source of atmospheric CO2, which is volcanic emissions that occur mainly at continental margins due to plate tectonics (popularly ‘continental drift’); tectonic activity also affects the weathering sink for CO2 by exposing fresh rock. The CO2 tectonic source grew from 60 to 50 Myr BP as India subducted carbonate-rich ocean crust while moving through the present Indian Ocean prior to its collision with Asia about 50 Myr BP [8], causing atmospheric CO2 to reach levels of the order of 1000 ppm at 50 Myr BP [9]. Since then, atmospheric CO2 declined as the Indian and Atlantic Oceans have been major depocentres for carbonate and organic sediments while subduction of carbonate-rich crust has been limited mainly to small regions near Indonesia and Central America [10], thus allowing CO2 to decline to levels as low as 170 ppm during recent glacial periods [11]. A climate forcing due to a CO2 change from 1000 to 170 ppm is more than 10 W m−2, which compares with forcings of the order of 1 W m−2 for competing climate forcings during the Cenozoic era [5], specifically long-term change of solar irradiance and change of planetary albedo (reflectance) owing to the overall minor displacement of continents in that era. Superimposed on the long-term trends are occasional global warming spikes, ‘hyperthermals’, most prominently the Palaeocene–Eocene Thermal Maximum (PETM) at approximately 56 Myr BP [12] and the Mid-Eocene Climatic Optimum at approximately 42 Myr BP [13], coincident with large temporary increases of atmospheric CO2. The most studied hyperthermal, the PETM, caused global warming of at least 5°C coincident with injection of a likely 4000–7000 Gt of isotopically light carbon into the atmosphere and ocean [14]. The size of the carbon injection is estimated from changes in the stable carbon isotope ratio 13C/12C in sediments and from ocean acidification implied by changes in the ocean depth below which carbonate dissolution occurred. The potential carbon source for hyperthermal warming that received most initial attention was methane hydrates on continental shelves, which could be destabilized by sea floor warming [15]. Alternative sources include release of carbon from Antarctic permafrost and peat [16]. Regardless of the carbon source(s), it has been shown that the hyperthermals were astronomically paced, spurred by coincident maxima in the Earth's orbit eccentricity and spin axis tilt [17], which increased high-latitude insolation and warming. The PETM was followed by successively weaker astronomically paced hyperthermals, suggesting that the carbon source(s) partially recharged in the interim [18]. A high temporal resolution sediment core from the New Jersey continental shelf [19] reveals that PETM warming in at least that region began about 3000 years prior to a massive release of isotopically light carbon. This lag and climate simulations [20] that produce large warming at intermediate ocean depths in response to initial surface warming are consistent with the concept of a methane hydrate role in hyperthermal events. The hyperthermals confirm understanding about the long recovery time of the Earth's carbon cycle [21] and reveal the potential for threshold or ‘tipping point’ behaviour with large amplifying climate feedback in response to warming [22]. One implication is that if humans burn most of the fossil fuels, thus injecting into the atmosphere an amount of CO2 at least comparable to that injected during the PETM, the CO2 would stay in the surface carbon reservoirs (atmosphere, ocean, soil, biosphere) for tens of thousands of years, long enough for the atmosphere, ocean and ice sheets to fully respond to the changed atmospheric composition. In addition, there is the potential that global warming from fossil fuel CO2 could spur release of CH4 and CO2 from methane hydrates or permafrost. Carbon release during the hyperthermals required several thousand years, but that long injection time may have been a function of the pace of the astronomical forcing, which is much slower than the pace of fossil fuel burning. The Cenozoic record also reveals the amplification of climate change that occurs with growth or decay of ice sheets, as is apparent at about 34 Myr BP when the Earth became cool enough for large-scale glaciation of Antarctica and in the most recent 3–5 Myr with the growth of Northern Hemisphere ice sheets. Global climate fluctuated in the 20 Myr following Antarctic glaciation with warmth during the Mid-Miocene Climatic Optimum (MMCO, 15 Myr BP) possibly comparable to that at 34 Myr BP, as, for example, Germany became warm enough to harbour snakes and crocodiles that require an annual temperature of about 20°C or higher and a winter temperature more than 10°C [23]. Antarctic vegetation in the MMCO implies a summer temperature of approximately 11°C warmer than today [24] and annual sea surface temperatures ranging from 0°C to 11.5°C [25]. Superimposed on the long-term trends, in addition to occasional hyperthermals, are continual high-frequency temperature oscillations, which are apparent in figure 1 after 34 Myr BP, when the Earth became cold enough for a large ice sheet to form on Antarctica, and are still more prominent during ice sheet growth in the Northern Hemisphere. These climate oscillations have dominant periodicities, ranging from about 20 to 400 kyr, that coincide with variations in the Earth's orbital elements [26], specifically the tilt of the Earth's spin axis, the eccentricity of the orbit and the time of year when the Earth is closest to the Sun. The slowly changing orbit and tilt of the spin axis affect the seasonal distribution of insolation [27], and thus the growth and decay of ice sheets, as proposed by Milankovitch [28]. Atmospheric CO2, CH4 and N2O have varied almost synchronously with global temperature during the past 800 000 years for which precise data are available from ice cores, the GHGs providing an amplifying feedback that magnifies the climate change instigated by orbit perturbations [29–31]. Ocean and atmosphere dynamical effects have been suggested as possible causes of some climate change within the Cenozoic era; for example, topographical effects of mountain building [32], closing of the Panama Seaway [33] or opening of the Drake Passage [34]. Climate modelling studies with orographic changes confirm significant effects on monsoons and on Eurasian temperature [35]. Modelling studies indicate that closing of the Panama Seaway results in a more intense Atlantic thermohaline circulation, but only small effects on Northern Hemisphere ice sheets [36]. Opening of the Drake Passage surely affected ocean circulation around Antarctica, but efforts to find a significant effect on global temperature have relied on speculation about possible effects on atmospheric CO2 [37]. Overall, there is no strong evidence that dynamical effects are a major direct contributor to Cenozoic global temperature change. We hypothesize that the global climate variations of the Cenozoic (figure 1) can be understood and analysed via slow temporal changes in Earth's energy balance, which is a function of solar irradiance, atmospheric composition (specifically long-lived GHGs) and planetary surface albedo. Using measured amounts of GHGs during the past 800 000 years of glacial–interglacial climate oscillations and surface albedo inferred from sea-level data, we show that a single empirical ‘fast-feedback’ climate sensitivity can account well for the global temperature change over that range of climate states. It is certain that over a large climate range climate sensitivity must become a strong function of the climate state, and thus we use a simplified climate model to investigate the dependence of climate sensitivity on the climate state. Finally, we use our estimated state-dependent climate sensitivity to infer Cenozoic CO2 change and compare this with proxy CO2 data, focusing on the Eocene climatic optimum, the Oligocene glaciation, the Miocene optimum and the Pliocene. The δ18O stable isotope ratio was the first palaeothermometer, proposed by Urey [38] and developed especially by Emiliani [39]. There are now several alternative proxy measures of ancient climate change, but the δ18O data (figure 1a) of Zachos et al. [4], a conglomerate of the global ocean sediment cores, is well suited for our purpose as it covers the Cenozoic era with good temporal resolution. There are large, even dominant, non-climatic causes of δ18O changes over hundreds of millions of years [40], but non-climatic change may be small in the past few hundred million years [41] and is generally neglected in Cenozoic climate studies. The principal difficulty in using the δ18O record to estimate global deep ocean temperature, in the absence of non-climatic change, is that δ18O is affected by the global ice mass as well as the deep ocean temperature. We make a simple estimate of global sea-level change for the Cenozoic era using the near-global δ18O compilation of Zachos et al. [4]. More elaborate and accurate approaches, including use of models, will surely be devised, but comparison of our result with other approaches is instructive regarding basic issues such as the vulnerability of today's ice sheets to near-term global warming and the magnitude of hysteresis effects in ice sheet growth and decay. During the Early Cenozoic, between 65.5 and 35 Myr BP, the Earth was so warm that there was little ice on the planet and the deep ocean temperature is approximated by [6] 3.1 Hansen et al. [5] made the approximation that, as the Earth became colder and continental ice sheets grew, further increase in δ18O was due, in equal parts, to deep ocean temperature change and ice mass change,3.2 Equal division of the δ18O change into temperature change and ice volume change was suggested by comparing δ18O at the endpoints of the climate change from the nearly ice-free planet at 35 Myr BP (when δ18O approx. 1.75) with the Last Glacial Maximum (LGM), which peaked approximately 20 kyr BP. The change of δ18O between these two extreme climate states (approx. 3) is twice the change of δ18O due to temperature change alone (approx. 1.5), with the temperature change based on the linear relation (??eq3.1) and estimates of Tdo∼5°C at 35 Myr BP (figure 1) and approximately −1°C at the LGM [42].This approximation can easily be made more realistic. Although ice volume and deep ocean temperature changes contributed comparable amounts to δ18O change on average over the full range from 35 Myr to 20 kyr BP, the temperature change portion of the δ18O change must decrease as the deep ocean temperature approaches the freezing point [43]. The rapid increase in δ18O in the past few million years was associated with the appearance of Northern Hemisphere ice sheets, symbolized by the dark blue bar in figure 1a. The sea-level change between the LGM and Holocene was approximately 120 m [44,45]. Thus, two-thirds of the 180 m sea-level change between the ice-free planet and the LGM occurred with formation of Northern Hemisphere ice (and probably some increased volume of Antarctic ice). Thus, rather than taking the 180 m sea-level change between the nearly ice-free planet of 34 Myr BP and the LGM as being linear over the entire range (with 90 m for δ18O<3.25 and 90 m for δ18O>3.25), it is more realistic to assign 60 m of sea-level change to δ18O 1.75–3.25 and 120 m to δ18O>3.25. The total deep ocean temperature change of 6°C for the change of δ18O from 1.75 to 4.75 is then divided two-thirds (4°C) for the δ18O range 1.75–3.25 and 2°C for the δ18O range 3.25–4.75. Algebraically, 3.3 3.4 3.5 and3.6 where SL is the sea level and its zero point is the Late Holocene level. The coefficients in equations (3.4) and (3.6) account for the fact that the mean LGM value of δ18O is approximately 4.9. The resulting deep ocean temperature is shown in figure 1b for the full Cenozoic era.Sea level from equations (3.3) and (3.4) is shown by the blue curves in figure 2, including comparison (figure 2c) with the Late Pleistocene sea-level record of Rohling et al. [47], which is based on analysis of Red Sea sediments, and comparison (figure 2b) with the sea-level chronology of de Boer et al. [46], which is based on ice sheet modelling with the δ18O data of Zachos et al. [4] as a principal input driving the ice sheet model. Comparison of our result with that of de Boer et al. [46] for the other periods of figure 2 is included in the electronic supplementary material, where we also make available our numerical data. Deep ocean temperature from equations (3.5) and (3.6) is shown for the Pliocene and Pleistocene in figure 3 and for the entire Cenozoic era in figure 1. Figure 2. (a–c) Sea level from equations (3.3) and (3.4) using δ18O data of Zachos et al. [4], compared in (b) with ice sheet model results of de Boer et al. [46] and in (c) with the sea-level analysis of Rohling et al. [47].
Figure 3. Deep ocean temperature in (a) the Pliocene and Pleistocene and (b) the last 800 000 years. High-frequency variations (black) are five-point running means of the original data [4], whereas the blue curve has a 500 kyr resolution. The deep ocean temperature for the entire Cenozoic era is in figure 1b.
Differences between our inferred sea-level chronology and that from the ice sheet model [46] are relevant to the assessment of the potential danger to humanity from future sea-level rise. Our estimated sea levels have reached +5 to 10 m above the present sea level during recent interglacial periods that were barely warmer than the Holocene, whereas the ice sheet model yields maxima at most approximately 1 m above the current sea level. We find the Pliocene sea level varying between about +20 m and −50 m, with the Early Pliocene averaging about +15 m; the ice sheet model has a less variable sea level with the Early Pliocene averaging about +8 m. A 15 m sea-level rise implies that the East Antarctic ice sheet as well as West Antarctica and Greenland ice were unstable at a global temperature no higher than those projected to occur this century [1,48]. How can we interpret these differences, and what is the merit of our simple δ18O scaling? Ice sheet models constrained by multiple observations may eventually provide our best estimate of sea-level change, but as yet models are primitive. Hansen [49,50] argues that real ice sheets are more responsive to climate change than is found in most ice sheet models. Our simple scaling approximation implicitly assumes that ice sheets are sufficiently responsive to climate change that hysteresis is not a dominant effect; in other words, ice volume on millennial time scales is a function of temperature and does not depend much on whether the Earth is in a warming or cooling phase. Thus, our simple transparent calculation may provide a useful comparison with geological data for sea-level change and with results of ice sheet models. We cannot a priori define accurately the error in our sea-level estimates, but we can compare with geological data in specific cases as a check on reasonableness. Our results (figure 2) yield two instances in the past million years when sea levels have reached heights well above the current sea level: +9.8 m in the Eemian (approx. 120 kyr BP, also known as Marine Isotope Stage 5e or MIS-5e) and +7.1 m in the Holsteinian (approx. 400 kyr BP, also known as MIS-11). Indeed, these are the two interglacial periods in the Late Pleistocene that traditional geological methods identify as probably having a sea level exceeding that in the Holocene. Geological evidence, mainly coral reefs on tectonically stable coasts, was described in the review of Overpeck et al. [51] as favouring an Eemian maximum of +4 to more than 6 m. Rohling et al. [52] cite many studies concluding that the mean sea level was 4–6 m above the current sea level during the warmest portion of the Eemian, 123–119 kyr BP; note that several of these studies suggest Eemian sea-level fluctuations up to +10 m, and provide the first continuous sea-level data supporting rapid Eemian sea-level fluctuations. Kopp et al. [53] made a statistical analysis of data from a large number of sites, concluding that there was a 95% probability that the Eemian sea level reached at least +6.6 m with a 67% probability that it exceeded 8 m. The Holsteinian sea level is more difficult to reconstruct from geological data because of its age, and there has been a long-standing controversy concerning a substantial body of geological shoreline evidence for a +20 m Late Holsteinian sea level that Hearty and co-workers have found on numerous sites [54,55] (numerous pros and cons are contained in the references provided in our present paragraph). Rohling et al. [56] note that their temporally continuous Red Sea record ‘strongly supports the MIS-11 sea level review of Bowen [57], which also places MIS-11 sea level within uncertainties at the present-day level’. This issue is important because both ice core data [29] and ocean sediment core data (see below) indicate that the Holsteinian period was only moderately warmer than the Holocene with similar Earth orbital parameters. We suggest that the resolution of this issue is consistent with our estimate of the approximately +7 m Holsteinian global sea level, and is provided by Raymo & Mitrovica [58], who pointed out the need to make a glacial isostatic adjustment (GIA) correction for post-glacial crustal subsidence at the places where Hearty and others deduced local sea-level change. The uncertainties in GIA modelling led Raymo & Mitrovica [58] to conclude that the peak Holsteinian global sea level was in the range of +6 to 13 m relative to the present. Thus, it seems to us, there is a reasonable resolution of the long-standing Holsteinian controversy, with substantial implications for humanity, as discussed in later sections. We now address differences between our sea-level estimates and those from ice sheet models. We refer to both the one-dimensional ice sheet modelling of de Boer et al. [46], which was used to calculate sea level for the entire Cenozoic era, and the three-dimensional ice sheet model of Bintanja et al. [59], which was used for simulations of the past million years. The differences most relevant to humanity occur in the interglacial periods slightly warmer than the Holocene, including the Eemian and Hosteinian, as well as the Pliocene, which may have been as warm as projected for later this century. Both the three-dimensional model of Bintanja et al. [59] and the one-dimensional model of de Boer et al. [46] yield maximum Eemian and Hosteinian sea levels of approximately 1 m relative to the Holocene. de Boer et al. [46] obtain approximately +8 m for the Early Pliocene, which compares with our approximately +15 m. These differences reveal that the modelled ice sheets are less susceptible to change in response to global temperature variation than our δ18O analysis. Yet the ice sheet models do a good job of reproducing the sea-level change for climates colder than the Holocene, as shown in figure 2 and electronic supplementary material, figure S2. One possibility is that the ice sheet models are too lethargic for climates warmer than the Holocene. Hansen & Sato [60] point out the sudden change in the responsiveness of the ice sheet model of Bintanja et al. [59] when the sea level reaches today's level (figs 3 and 4 of Hansen & Sato [60]) and they note that the empirical sea-level data provide no evidence of such a sudden change. The explanation conceivably lies in the fact that the models have many parameters and their operation includes use of ‘targets’ [46] that affect the model results, because these choices might yield different results for warmer climates than the results for colder climates. Because of the potential that model development choices might be influenced by expectations of a ‘correct’ result, it is useful to have estimates independent of the models based on alternative assumptions. Note that our approach also involves ‘targets’ based on expected behaviour, albeit simple transparent ones. Our two-legged linear approximation of the sea level (equations (3.3) and (3.4)) assumes that the sea level in the LGM was 120 m lower than today and that the sea level was 60 m higher than today 35 Myr BP. This latter assumption may need to be adjusted if glaciers and ice caps in the Eocene had a volume of tens of metres of sea level. However, Miller et al. [61] conclude that there was a sea level fall of approximately 55 m at the Eocene–Oligocene transition, consistent with our assumption that Eocene ice probably did not contain more than approximately 10 m of sea level. Real-world data for the Earth's sea-level history ultimately must provide assessment of sea-level sensitivity to climate change. A recent comprehensive review [7] reveals that there are still wide uncertainties about the Earth's sea-level history that are especially large for time scales of tens of millions of years or longer, which is long enough for substantial changes in the shape and volume of ocean basins. Gasson et al. [7] plot regional (New Jersey) sea level (their fig. 14) against the deep ocean temperature inferred from the magnesium/calcium ratio (Mg/Ca) of deep ocean foraminifera [62], finding evidence for a nonlinear sea-level response to temperature roughly consistent with the modelling of de Boer et al. [46]. Sea-level change is limited for Mg/Ca temperatures up to about 5°C above current values, whereupon a rather abrupt sea-level rise of several tens of metres occurs, presumably representing the loss of Antarctic ice. However, the uncertainty in the reconstructed sea level is tens of metres and the uncertainty in the Mg/Ca temperature is sufficient to encompass the result from our δ18O prescription, which has comparable contributions of ice volume change and deep ocean temperature change at the Late Eocene glaciation of Antarctica. Furthermore, the potential sea-level rise of most practical importance is the first 15 m above the Holocene level. It is such ‘moderate’ sea-level change for which we particularly question the projections implied by current ice sheet models. Empirical assessment depends upon real-world sea-level data in periods warmer than the Holocene. There is strong evidence, discussed above, that the sea level was several metres higher in recent warm interglacial periods, consistent with our data interpretation. The Pliocene provides data extension to still warmer climates. Our interpretation of δ18O data suggests that Early Pliocene sea-level change (due to ice volume change) reached about +15 m, and it also indicates sea-level fluctuations as large as 20–40 m. Sea-level data for Mid-Pliocene warm periods, of comparable warmth to average Early Pliocene conditions (figure 3), suggest sea heights as great as +15–25 m [63,64]. Miller et al. [61] find a Pliocene sea-level maximum of 22±10 m (95% confidence). GIA creates uncertainty in sea-level reconstructions based on shoreline geological data [65], which could be reduced via appropriately distributed field studies. Dwyer & Chandler [64] separate Pliocene ice volume and temperature in deep ocean δ18O via ostracode Mg/Ca temperatures, finding sea-level maxima and oscillations comparable to our results. Altogether, the empirical data provide strong evidence against the lethargy and strong hysteresis effects of at least some ice sheet models. The temperature of most interest to humanity is the surface air temperature. A record of past global surface temperature is required for empirical inference of global climate sensitivity. Given that climate sensitivity can depend on the initial climate state and on the magnitude and sign of the climate forcing, a continuous record of global temperature over a wide range of climate states would be especially useful. Because of the singularly rich climate story in Cenozoic deep ocean δ18O (figure 1), unrivalled in detail and self-consistency by alternative climate proxies, we use deep ocean δ18O to provide the fine structure of Cenozoic temperature change. We use surface temperature proxies from the LGM, the Pliocene and the Eocene to calibrate and check the relation between deep ocean and surface temperature change. The temperature signal in deep ocean δ18O refers to the sea surface where cold dense water formed and sank to the ocean bottom, the principal location of deep water formation being the Southern Ocean. Empirical data and climate models concur that surface temperature change is generally amplified at high latitudes, which tends to make temperature change at the site of deep water formation an overestimate of global temperature change. Empirical data and climate models also concur that surface temperature change is amplified over land areas, which tends to make temperature change at the site of deep water an underestimate of the global temperature. Hansen et al. [5] and Hansen & Sato [60] noted that these two factors were substantially offsetting, and thus they made the assumption that benthic foraminifera provide a good approximation of global mean temperature change for most of the Cenozoic era. However, this approximation breaks down in the Late Cenozoic for two reasons. First, the deep ocean and high-latitude surface ocean where deep water forms are approaching the freezing point in the Late Cenozoic. As the Earth's surface cools further, cold conditions spread to lower latitudes but polar surface water and the deep ocean cannot become much colder, and thus the benthic foraminifera record a temperature change smaller than the global average surface temperature change [43]. Second, the last 5.33 Myr of the Cenozoic, the Pliocene and Pleistocene, was the time that global cooling reached a degree such that large ice sheets could form in the Northern Hemisphere. When a climate forcing, or a slow climate feedback such as ice sheet formation, occurs in one hemisphere, the temperature change is much larger in the hemisphere with the forcing (cf. examples in Hansen et al. [66]). Thus, cooling during the last 5.33 Myr in the Southern Ocean site of deep water formation was smaller than the global average cooling. We especially want our global surface temperature reconstruction to be accurate for the Pliocene and Pleistocene because the global temperature changes that are expected by the end of this century, if humanity continues to rapidly change atmospheric composition, are of a magnitude comparable to climate change in those epochs [1,48]. Fortunately, sufficient information is available on surface temperature change in the Pliocene and Pleistocene to allow us to scale the deep ocean temperature change by appropriate factors, thus retaining the temporal variations in the δ18O while also having a realistic magnitude for the total temperature change over these epochs. Pliocene temperature is known quite well because of a long-term effort to reconstruct the climate conditions during the Mid-Pliocene warm period (3.29–2.97 Myr BP) and a coordinated effort to numerically simulate the climate by many modelling groups ([67] and papers referenced therein). The reconstructed Pliocene climate used data for the warmest conditions found in the Mid-Pliocene period, which would be similar to average conditions in the Early Pliocene (figure 3). These boundary conditions were used by eight modelling groups to simulate Pliocene climate with atmospheric general circulation models. Although atmosphere–ocean models have difficulty replicating Pliocene climate, atmospheric models forced by specified surface boundary conditions are expected to be capable of calculating global surface temperature with reasonable accuracy. The eight global models yield Pliocene global warming of 3±1°C relative to the Holocene [68]. This Pliocene warming is an amplification by a factor of 2.5 of the deep ocean temperature change. Similarly, for the reasons given above, the deep ocean temperature change of 2.25°C between the Holocene and the LGM is surely an underestimate of the surface air temperature change. Unfortunately, there is a wide range of estimates for LGM cooling, approximately 3–6°C, as discussed in §6. Thus, we take 4.5°C as our best estimate for LGM cooling, implying an amplification of surface temperature change by a factor of two relative to deep ocean temperature change for this climate interval. We obtain an absolute temperature scale using the Jones et al. [69] estimate of 14°C as the global mean surface temperature for 1961–1990, which corresponds to approximately 13.9°C for the 1951–1980 base period that we normally use [70] and approximately 14.4°C for the first decade of the twenty-first century. We attach the instrumental temperature record to the palaeo data by assuming that the first decade of the twenty-first century exceeds the Holocene mean by 0.25±0.25°C. Global temperature probably declined over the past several millennia [71], but we suggest that warming of the past century has brought global temperature to a level that now slightly exceeds the Holocene mean, judging from sea-level trends and ice sheet mass loss. Sea level is now rising 3.1 mm per year or 3.1 m per millennium [72], an order of magnitude faster than the rate during the past several thousand years, and Greenland and Antarctica are losing mass at accelerating rates [73,74]. Our assumption that global temperature passed the Holocene mean a few decades ago is consistent with the rapid change of ice sheet mass balance in the past few decades [75]. The above concatenation of instrumental and palaeo records yields a Holocene mean of 14.15°C and Holocene maximum (from five-point smoothed δ18O) of 14.3°C at 8.6 kyr BP. Given a Holocene temperature of 14.15°C and LGM cooling of 4.5°C, the Early Pliocene mean temperature 3°C warmer than the Holocene leads to the following prescription: 4.1 and4.2 This prescription yields a maximum Eemian temperature of 15.56°C, which is approximately 1.4°C warmer than the Holocene mean and approximately 1.8°C warmer than the 1880–1920 mean. Clark & Huybers [76] fit a polynomial to proxy temperatures for the Eemian, finding warming as much as +5°C at high northern latitudes but global warming of +1.7°C ‘relative to the present interglacial before industrialization’. Other analyses of Eemian data find global sea surface temperature warmer than the Late Holocene by 0.7±0.6°C [77] and all-surface warming of 2°C [78], all in reasonable accord with our prescription.Our first estimate of global temperature for the remainder of the Cenozoic assumes that ΔTs=ΔTdo prior to 5.33 Myr BP, i.e. prior to the Plio-Pleistocene, which yields a peak Ts of approximately 28°C at 50 Myr BP (figure 4). This is at the low end of the range of current multi-proxy measures of sea surface temperature for the Early Eocene Climatic Optimum (EECO) [79–81]. Climate models are marginally able to reproduce this level of Eocene warmth, but the models require extraordinarily high CO2 levels, for example 2240–4480 ppm [82] and 2500–6500 ppm [83], and the quasi-agreement between data and models requires an assumption that some of the proxy temperatures are biased towards summer values. Moreover, taking the proxy sea surface temperature data for the peak Eocene period (55–48 Myr BP) at face value yields a global temperature of 33–34°C (fig. 3 of Bijl et al. [84]), which would require an even larger CO2 amount with the same climate models. Thus, below we also consider the implications for climate sensitivity of an assumption that ΔTs=1.5×ΔTdo prior to 5.33 Myr BP, which yields Ts approximately 33°C at 50 Myr BP (see electronic supplementary material, figure S3). Figure 4. (a–c) Surface temperature estimate for the past 65.5 Myr, including an expanded time scale for (b) the Pliocene and Pleistocene and (c) the past 800 000 years. The red curve has a 500 kyr resolution. Data for this and other figures are available in the electronic supplementary material.
Climate sensitivity (S) is the equilibrium global surface temperature change (ΔTeq) in response to a specified unit forcing after the planet has come back to energy balance, 5.1 i.e. climate sensitivity is the eventual (equilibrium) global temperature change per unit forcing. Climate sensitivity depends upon climate feedbacks, the many physical processes that come into play as climate changes in response to a forcing. Positive (amplifying) feedbacks increase the climate response, whereas negative (diminishing) feedbacks reduce the response.We usually discuss climate sensitivity in terms of a global mean temperature response to a 4 W m−2 CO2 forcing. One merit of this standard forcing is that its magnitude is similar to an anticipated near-term human-made climate forcing, thus avoiding the need to continually scale the unit sensitivity to achieve an applicable magnitude. A second merit is that the efficacy of forcings varies from one forcing mechanism to another [66]; so it is useful to use the forcing mechanism of greatest interest. Finally, the 4 W m−2 CO2 forcing avoids the uncertainty in the exact magnitude of a doubled CO2 forcing [1,48] estimate of 3.7 W m−2 for doubled CO2, whereas Hansen et al. [66] obtain 4.1 W m−2, as well as problems associated with the fact that a doubled CO2 forcing varies as the CO2 amount changes (the assumption that each CO2 doubling has the same forcing is meant to approximate the effect of CO2 absorption line saturation, but actually the forcing per doubling increases as CO2 increases [66,85]). Climate feedbacks are the core of the climate problem. Climate feedbacks can be confusing, because in climate analyses what is sometimes a climate forcing is at other times a climate feedback. A CO2 decrease from, say, approximately 1000 ppm in the Early Cenozoic to 170–300 ppm in the Pleistocene, caused by shifting plate tectonics, is a climate forcing, a perturbation of the Earth's energy balance that alters the temperature. Glacial–interglacial oscillations of the CO2 amount and ice sheet size are both slow climate feedbacks, because glacial–interglacial climate oscillations largely are instigated by insolation changes as the Earth's orbit and tilt of its spin axis change, with the climate change then amplified by a nearly coincident change of the CO2 amount and the surface albedo. However, for the sake of analysis, we can also choose and compare periods that are in quasi-equilibrium, periods during which there was little change of the ice sheet size or the GHG amount. For example, we can compare conditions averaged over several millennia in the LGM with mean Holocene conditions. The Earth's average energy imbalance within each of these periods had to be a small fraction of 1 W m−2. Such a planetary energy imbalance is very small compared with the boundary condition ‘forcings’, such as changed GHG amount and changed surface albedo that maintain the glacial-to-interglacial climate change. The average fast-feedback climate sensitivity over the LGM–Holocene range of climate states can be assessed by comparing estimated global temperature change and climate forcing change between those two climate states [3,86]. The appropriate climate forcings are the changes in long-lived GHGs and surface properties on the planet. Fast feedbacks include water vapour, clouds, aerosols and sea ice changes. This fast-feedback sensitivity is relevant to estimating the climate impact of human-made climate forcings, because the size of ice sheets is not expected to change significantly in decades or even in a century and GHGs can be specified as a forcing. GHGs change in response to climate change, but it is common to include these feedbacks as part of the climate forcing by using observed GHG changes for the past and calculated GHGs for the future, with calculated amounts based on carbon cycle and atmospheric chemistry models. Climate forcings due to past changes in GHGs and surface albedo can be computed for the past 800 000 years using data from polar ice cores and ocean sediment cores. We use CO2 [87] and CH4 [88] data from Antarctic ice cores (figure 5a) to calculate an effective GHG forcing as follows: 5.2 where Fa is the adjusted forcing, i.e. the planetary energy imbalance due to the GHG change after the stratospheric temperature has time to adjust to the gas change. Fe, the effective forcing, accounts for variable efficacies of different climate forcings [66]. Formulae for Fa of each gas are given by Hansen et al. [89]. The factor 1.4 converts the adjusted forcing of CH4 to its effective forcing, Fe, which is greater than Fa mainly because of the effect of CH4 on the tropospheric ozone and the stratospheric water vapour [66]. The factor 1.12 approximates the forcing by N2O changes, which are not as well preserved in the ice cores but have a strong positive correlation with CO2 and CH4 changes [90]. The factor 1.12 is smaller than the 1.15 used by Hansen et al. [91], and is consistent with estimates of the N2O forcing in the current Goddard Institute for Space Studies (GISS) radiation code and that of the Intergovernmental Panel on Climate Change (IPCC) [1,48]. Our LGM–Holocene GHG forcing (figure 5c) is approximately 3 m−2, moderately larger than the 2.8 W m−2 estimated by IPCC [1,48] because of our larger effective CH4 forcing.Figure 5. (a) CO2 and CH4 from ice cores; (b) sea level from equation (3.4) and (c) resulting climate forcings (see text).
Climate forcing due to surface albedo change is a function mainly of the sea level, which implicitly defines ice sheet size. Albedo change due to LGM–Holocene vegetation change, much of which is inherent with ice sheet area change, and albedo change due to coastline movement are lumped together with ice sheet area change in calculating the surface albedo climate forcing. An ice sheet forcing does not depend sensitively on the ice sheet shape or on how many ice sheets the ice volume is divided among and is nearly linear in sea-level change (see electronic supplementary material, figure S4, and [5]). For the sake of simplicity, we use the linear relation in Hansen et al. [5] and electronic supplementary material, figure S4; thus, 5 W m−2 between the LGM and ice-free conditions and 3.4 W m−2 between the LGM and Holocene. This scale factor was based on simulations with an early climate model [3,92]; comparable forcings are found in other models (e.g. see discussion in [93]), but results depend on cloud representations, assumed ice albedo and other factors; so the uncertainty is difficult to quantify. We subjectively estimate an uncertainty of approximately 20%. Global temperature change obtained by multiplying the sum of the two climate forcings in figure 5c by a sensitivity of 3/4°C per W m−2 yields a remarkably good fit to ‘observations’ (figure 6), where the observed temperature is 2×ΔTdo, with 2 being the scale factor required to yield the estimated 4.5°C LGM–Holocene surface temperature change. The close match is partly a result of the fact that sea-level and temperature data are derived from the same deep ocean record, but use of other sea-level reconstructions still yields a good fit between the calculated and observed temperature [5]. However, exactly the same match as in figure 6 is achieved with a fast-feedback sensitivity of 1°C per W m−2 if the LGM cooling is 6°C or with a sensitivity of 0.5°C per W m−2 if the LGM cooling is 3°C. Figure 6. Calculated surface temperature for forcings of figure 5c with a climate sensitivity of 0.75°C per W m−2, compared with 2×ΔTdo. Zero point is the Holocene (10 kyr) mean.
Accurate data defining LGM–Holocene warming would aid empirical evaluation of fast-feedback climate sensitivity. Remarkably, the range of recent estimates of LGM–Holocene warming, from approximately 3°C [94] to approximately 6°C [95], is about the same as at the time of the CLIMAP [96] project. Given today's much improved analytic capabilities, a new project to define LGM climate conditions, analogous to the Pliocene Research, Interpretation and Synoptic Mapping (PRISM) Pliocene data reconstruction [97,98] and Pliocene Model Intercomparison Project (PlioMIP) model intercomparisons [67,68], could be beneficial. In §7b, we suggest that a study of Eemian glacial–interglacial climate change could be even more definitive. Combined LGM, Eemian and Pliocene studies would address an issue raised at a recent workshop [99]: the need to evaluate how climate sensitivity varies as a function of the initial climate state. The calculations below were initiated after the workshop as another way to address that question. Climate sensitivity must be a strong function of the climate state. Simple climate models show that, when the Earth becomes cold enough for the ice cover to approach the tropics, the amplifying albedo feedback causes rapid ice growth to the Equator: ‘snowball Earth’ conditions [100]. Real-world complexity, including ocean dynamics, can mute this sharp bifurcation to a temporarily stable state [101], but snowball events have occurred several times in the Earth's history when the younger Sun was dimmer than today [102]. The Earth escaped snowball conditions owing to limited weathering in that state, which allowed volcanic CO2 to accumulate in the atmosphere until there was enough CO2 for the high sensitivity to cause rapid deglaciation [103]. Climate sensitivity at the other extreme, as the Earth becomes hotter, is also driven mainly by an H2O feedback. As climate forcing and temperature increase, the amount of water vapour in the air increases and clouds may change. Increased water vapour makes the atmosphere more opaque in the infrared region that radiates the Earth's heat to space, causing the radiation to emerge from higher colder layers, thus reducing the energy emitted to space. This amplifying feedback has been known for centuries and was described remarkably well by Tyndall [104]. Ingersoll [105] discussed the role of water vapours in the ‘runaway greenhouse effect’ that caused the surface of Venus to eventually become so hot that carbon was ‘baked’ from the planet's crust, creating a hothouse climate with almost 100 bars of CO2 in the air and a surface temperature of about 450°C, a stable state from which there is no escape. Arrival at this terminal state required passing through a ‘moist greenhouse’ state in which surface water evaporates, water vapour becomes a major constituent of the atmosphere and H2O is dissociated in the upper atmosphere with the hydrogen slowly escaping to space [106]. That Venus had a primordial ocean, with most of the water subsequently lost to space, is confirmed by the present enrichment of deuterium over ordinary hydrogen by a factor of 100 [107], the heavier deuterium being less efficient in escaping gravity to space. The physics that must be included to investigate the moist greenhouse is principally: (i) accurate radiation incorporating the spectral variation of gaseous absorption in both the solar radiation and thermal emission spectral regions, (ii) atmospheric dynamics and convection with no specifications favouring artificial atmospheric boundaries, such as between a troposphere and stratosphere, (iii) realistic water vapour physics, including its effect on atmospheric mass and surface pressure, and (iv) cloud properties that respond realistically to climate change. Conventional global climate models are inappropriate, as they contain too much other detail in the form of parametrizations or approximations that break down as climate conditions become extreme. We use the simplified atmosphere–ocean model of Russell et al. [108], which solves the same fundamental equations (conservation of energy, momentum, mass and water substance, and the ideal gas law) as in more elaborate global models. Principal changes in the physics in the current version of the model are use of a step-mountain C-grid atmospheric vertical coordinate [109], addition of a drag in the grid-scale momentum equation in both atmosphere and ocean based on subgrid topography variations, and inclusion of realistic ocean tides based on exact positioning of the Moon and Sun. Radiation is the k-distribution method of Lacis & Oinas [110] with 25 k-values; the sensitivity of this specific radiation code is documented in detail by Hansen et al. [111]. Atmosphere and ocean dynamics are calculated on 3°×4° Arakawa C-grids. There are 24 atmospheric layers. In our present simulations, the ocean's depth is reduced to 100 m with five layers so as to achieve a rapid equilibrium response to forcings; this depth limitation reduces poleward ocean transport by more than half. Moist convection is based on a test of moist static stability as in Hansen et al. [92]. Two cloud types occur: moist convective clouds, when the atmosphere is moist statically unstable, and large-scale super-saturation, with cloud optical properties based on the amount of moisture removed to eliminate super-saturation, with scaling coefficients chosen to optimize the control run's fit with global observations [108,112]. To avoid long response times in extreme climates, today's ice sheets are assigned surface properties of the tundra, thus allowing them to have a high albedo snow cover in cold climates but darker vegetation in warm climates. The model, the present experiments and more extensive experiments will be described in a forthcoming paper [112]. The equilibrium response of the control run (1950 atmospheric composition, CO2 approx. 310 ppm) and runs with successive CO2 doublings and halvings reveals that snowball Earth instability occurs just beyond three CO2 halvings. Given that a CO2 doubling or halving is equivalent to a 2% change in solar irradiance [66] and the estimate that solar irradiance was approximately 6% lower 600 Ma at the most recent snowball Earth occurrence [113], figure 7 implies that about 300 ppm CO2 or less was sufficiently small to initiate glaciation at that time. Figure 7. (a) The calculated global mean temperature for successive doublings of CO2 (legend identifies every other case) and (b) the resulting climate sensitivity (1×CO2=310 ppm).
Climate sensitivity reaches large values at 8–32×CO2 (approx. 2500–10 000 ppm; figure 7b). High sensitivity is caused by increasing water vapour as the tropopause rises and diminishing low cloud cover, but the sensitivity decreases for still larger CO2 as cloud optical thickness and planetary albedo increase, as shown by Russell et al. [112]. The high sensitivity for CO2 less than 4×CO2 is due partly to the nature of the experiments (Greenland and Antarctic ice sheets being replaced by the tundra). High albedo snow cover on these continents largely disappears between 1×CO2 and 4×CO2, thus elevating the calculated fast-feedback sensitivity from approximately 4°C to approximately 5°C. In the real world, we would expect the Greenland and Antarctic ice sheets to be nearly eliminated and replaced by partially vegetated surfaces already at 2×CO2 (620 ppm) equilibrium climate. In other words, if the Greenland/Antarctic surface albedo change were identified as a slow feedback, rather than as a fast-feedback snow effect as it is in figure 7, the fast-feedback sensitivity at 1–4×CO2 would be approximately 4°C. Thus, the sensitivity approximately 8°C per CO2 doubling in the range of 8–32×CO2 is a very large increase over sensitivity at smaller CO2 amounts. How confident are we in the modelled fast-feedback sensitivity (figure 7b)? We suspect that the modelled water vapour feedback may be moderately exaggerated, because the water vapour amount in the control run exceeds observed amounts. In addition, the area of sea ice in the control run exceeds observations, which may increase the modelled sensitivity in the 1–4×CO2 range. On the other hand, we probably underestimate the sensitivity at very high CO2 amounts, because our model (such as most climate models) does not change the total atmospheric mass as the CO2 amount varies. Mass change due to conceivable fossil fuel loading (up to say 16×CO2) is unlikely to have much effect, but sensitivity is probably underestimated at high CO2 amounts owing to self-broadening of CO2 absorption lines. The increased atmospheric mass is also likely to alter the cloud feedback, which otherwise is a strongly diminishing feedback at very large CO2 amounts. Atmospheric mass will be important after the Earth has lost its ocean and carbon is baked into the atmosphere. These issues are being examined by Russell et al. [112]. Earth today, with approximately 1.26 times 1950 CO2, is far removed from the snowball state. Because of the increase in solar irradiance over the past 600 Myr and volcanic emissions, no feasible CO2 amount could take the Earth back to snowball conditions. Similarly, a Venus-like baked-crust CO2 hothouse is far distant because it cannot occur until the ocean escapes to space. We calculate an escape time of the order of 108–109 years even with the increased stratospheric water vapour and temperature at 16×CO2. Given the transient nature of a fossil fuel CO2 injection, the continuing forcing required to achieve a terminal Venus-like baked-crust CO2 hothouse must wait until the Sun's brightness has increased on the billion year time scale. However, the planet could become uninhabitable long before that. The practical concern for humanity is the high climate sensitivity and the eventual climate response that may be reached if all fossil fuels are burned. Estimates of the carbon content of all fossil fuel reservoirs including unconventional fossil fuels such as tar sands, tar shale and various gas reservoirs that can be tapped with developing technology [114] imply that CO2 conceivably could reach a level as high as 16 times the 1950 atmospheric amount. In that event, figure 7 suggests a global mean warming approaching 25°C, with much larger warming at high latitudes (see electronic supplementary material, figure S6). The result would be a planet on which humans could work and survive outdoors in the summer only in mountainous regions [115,116]—and there they would need to contend with the fact that a moist stratosphere would have destroyed the ozone layer [117]. GHG and surface albedo changes, which we treated as specified climate forcings in evaluating fast-feedback climate sensitivity, are actually slow climate feedbacks during orbit-instigated Pleistocene glacial–interglacial climate swings. Given that GHG and albedo feedbacks are both strong amplifying feedbacks, indeed accounting by themselves for most of the global Pleistocene climate variation, it is apparent that today's climate sensitivity on millennial time scales must be substantially larger than the fast-feedback sensitivity. Climate sensitivity including slow feedbacks is described as ‘Earth system sensitivity’ [118–120]. There are alternative choices for the feedbacks included in Earth system sensitivity. Hansen & Sato [60] suggest adding slow feedbacks one by one, creating a series of increasingly comprehensive Earth system climate sensitivities; specifically, they successively move climate-driven changes in surface albedo, non-CO2 GHGs and CO2 into the feedback category, at which point the Earth system sensitivity is relevant to an external forcing such as changing solar irradiance or human-made forcings. At each level, in this series, the sensitivity is state dependent. Our principal aim here is to use Cenozoic climate change to infer information on the all-important fast-feedback climate sensitivity, including its state dependence, via analysis of Earth system sensitivity. CO2 is clearly the dominant forcing of the long-term Cenozoic cooling, in view of the abundant evidence that CO2 reached levels of the order of 1000 ppm in the Early Cenozoic [9], as discussed in the Overview above. Thus, our approach is to examine Earth system sensitivity to CO2 change by calculating the CO2 history required to produce our reconstructed Cenozoic temperature history for alternative state-independent and state-dependent climate sensitivities. By comparing the resulting CO2 histories with CO2 proxy data, we thus assess the most realistic range for climate sensitivity. Two principal uncertainties in this analysis are (i) global temperature at the EECO approximately 50 Myr BP and (ii) CO2 amount at that time. We use EECO approximately 28°C (figure 4) as our standard case, but we repeat the analysis with EECO approximately 33°C (see electronic supplementary material, figure S3), thus allowing inference of how the conclusions change if knowledge of Eocene temperature changes. Similarly, our graphs allow the inferred climate sensitivity to be adjusted if improved knowledge of CO2 50 Myr BP indicates a value significantly different from approximately 1000 ppm. To clarify our calculations, let us first assume that fast-feedback climate sensitivity is a constant (state-independent) 3°C for doubled CO2 (0.75°C per W m−2). It is then trivial to convert our global temperature for the Cenozoic (figure 4a) to the total climate forcing throughout the Cenozoic, which is shown in the electronic supplementary material, figure S4a, as are results of subsequent steps. Next, we subtract the solar forcing, a linear increase of 1 W m−2 over the Cenozoic era due to the Sun's 0.4% irradiance increase [121], and the surface albedo forcing due to changing ice sheet size, which we take as linear at 5 W m−2 for the 180 m sea-level change from 35 Myr BP to the LGM. These top-of-the-atmosphere and surface forcings are moderate in size, compared with the total forcing over the Cenozoic, and partially offsetting, as shown in the electronic supplementary material, figure S4b. The residual forcing, which has a maximum of approximately 17 W m−2 just prior to 50 Myr BP, is the atmospheric forcing due to GHGs. Non-CO2 GHGs contribute 25% of the total GHG forcing in the period of ice core measurements. Atmospheric chemistry simulations [122] reveal continued growth of non-CO2 gases (N2O, CH4 and tropospheric O3) in warmer climates, at only a slightly lower rate (1.7–2.3 W m−2 for 4×CO2, which itself is approx. 8 W m−2). Thus, we take the CO2 forcing as 75% of the GHG forcing throughout the Cenozoic in our standard case, but we also consider the extreme case in which non-CO2 gases are fixed and thus contribute no climate forcing. A CO2 forcing is readily converted to the CO2 amount; we use the equation in table 1 of Hansen et al. [89]. The resulting Cenozoic CO2 history required to yield the global surface temperature of figure 4a is shown in figure 8a for state-independent climate sensitivity with non-CO2 GHGs providing 25% of the GHG climate forcing. The peak CO2 in this case is approximately 2000 ppm. If non-CO2 GHGs provide less than 25% of the total GHG forcing, then the inferred CO2 amount would be even greater. Results for alternative sensitivities, as in figure 8b, are calculated for a temporal resolution of 0.5 Myr to smooth out glacial–interglacial CO2 oscillations, as our interest here is in CO2 as a climate forcing. Figure 8. (a) CO2 amount required to yield a global temperature of figure 4a if fast-feedback climate sensitivity is 0.75°C per W m−2 and non-CO2 GHGs contribute 25% of the GHG forcing. (b) Same as in (a), but with temporal resolution 0.5 Myr and for three choices of fast-feedback sensitivity; the CO2 peak exceeds 5000 ppm in the case of 0.5°C sensitivity. The horizontal line is the Early–Mid-Holocene 260 ppm CO2 level.
We focus on the CO2 amount 50 Myr BP averaged over a few million years in assessing the realism of our inferred CO2 histories, because CO2 variations in the Cenozoic remain very uncertain despite the success of Beerling & Royer [9] in eliminating the most extreme outliers. Beerling & Royer [9] find a best-fit CO2 at 50 Myr BP of about 1000 ppm—see their figure 1, which also indicates that CO2 at 50 Myr BP was almost certainly in the range of 750–1500 ppm, even though it is impossible to provide a rigorous confidence interval. We conclude that the average fast-feedback climate sensitivity during the Cenozoic is larger than the canonical 3°C for 2×CO2 (0.75°C per W m−2) that has long been the central estimate for current climate. An average 4°C for 2×CO2 (1°C per W m−2) provides a good fit to the target 1000 ppm CO2, but the sensitivity must be still higher if non-CO2 GHG forcings amplify the CO2 by less than one-third, i.e. provide less than 25% of the total GHG forcing. More realistic assessment should account for the state dependence of climate sensitivity. Thus, we make the same calculations for the state-dependent climate sensitivity of the Russell climate model, i.e. we use the fast-feedback climate sensitivity of figure 7b. In addition, for the purpose of assessing how the results depend upon climate sensitivity, we consider a second case in which we reduce the Russell sensitivity of figure 7b by the factor two-thirds. The estimated 1000 ppm of CO2 at 50 Myr BP falls between the Russell sensitivity and two-thirds of the Russell sensitivity, though closer to the full Russell sensitivity. If the non-CO2 GHG forcing is less than one-third of the CO2 forcing, the result is even closer to the full Russell sensitivity. With these comparisons at 50 Myr BP in mind, we can use figure 9 to infer the likely CO2 amount at other times. The End-Eocene transition began at about 500 ppm and fell to about 400 ppm. The Mid-Miocene warmth, which peaked at about 15 Myr BP, required a CO2 increase of only a few tens of ppm with the Russell sensitivity, but closer to 100 ppm if the true sensitivity is only two-thirds of the Russell sensitivity. The higher (full Russell) sensitivity requires much less CO2 change to produce the Mid-Miocene warming for two reasons: (i) the greater temperature change for a specified forcing and (ii) the smaller CO2 change required to yield a given forcing from the lesser CO2 level of the higher sensitivity case. The average CO2 amount in the Early Pliocene is about 300 ppm for the Russell sensitivity, but could reach a few tens of ppm higher if the true sensitivity is closer to two-thirds of the Russell sensitivity. Figure 9. (a) CO2 amount required to yield the global temperature history of figure 4a if fast-feedback climate sensitivity is that calculated with the Russell model, i.e. the sensitivity shown in figure 7b, and two-thirds of that sensitivity. These results assume that non-CO2 GHGs provide 25% of the GHG climate forcing. (b) Vertical expansion for the past 35 Myr.
van de Wal et al. [123] used the same Zachos et al. [4] δ18O data to drive an inverse model calculation, including an ice sheet model to separate ice volume and temperature, thus inferring CO2 over the past 20 Myr. They find an MMCO CO2 approximately 450 ppm, which falls between the Russell and two-thirds Russell sensitivities (figure 9). The van de Wal et al. [123] model has a 30°C change in Northern Hemisphere temperature (their model is hemispheric) between the MMCO and average Pleistocene conditions driven by a CO2 decline from approximately 450 ppm to approximately 250 ppm, which is a forcing of approximately 3.5 W m−2. Thus, the implied (Northern Hemisphere) Earth system sensitivity is an implausible approximately 35°C for a 4 W m−2 CO2 forcing. The large temperature change may be required to produce substantial sea-level change in their ice sheet model, which we suggested above is unrealistically unresponsive to climate change. However, they assign most of the temperature change to slow feedbacks, thus inferring a fast-feedback sensitivity of only about 3°C per CO2 doubling. Finally, we use the largest and best documented of the hyperthermals, the PETM, to test the reasonableness of the Russell state-dependent climate sensitivity. Global warming in the PETM is reasonably well defined at 5–6°C and the plausible range for carbon mass input is approximately 4000–7000 Pg C [14]. Given that the PETM carbon injection occurred over a period of a few millennia, carbon cycle models suggest that about one-third of the carbon would be airborne as CO2 following complete injection [21]. With a conversion factor of 1 ppm CO2∼2.12 Gt C, the 4000–7000 Gt C source thus yields approximately 630–1100 ppm CO2. We can use figure 10, obtained via the same calculations as described above, to see how much CO2 is required to yield a 5°C warming. The Russell sensitivity requires approximately 800 ppm CO2 for a 5°C warming, whereas two-thirds of the Russell sensitivity requires approximately 2100 ppm CO2. Given the uncertainty in the airborne fraction of CO2 and possible non-CO2 gases, we cannot rule out the two-thirds Russell sensitivity, but the full Russell sensitivity fits plausible PETM carbon sources much better, especially if the PETM warming is actually somewhat more than 5°C (see figure 10 for quantitative implications). Figure 10. Atmospheric CO2 amount (y-axis) required to yield a given global temperature (x-axis) at the time of the PETM for (a) the Russell climate sensitivity and (b) two-thirds of the Russell sensitivity. The CO2 increment required to yield a given PETM warming above the pre-PETM temperature (25.7°C) is obtained by subtracting the CO2 amount at the desired Ts from the CO2 at Ts=25.7°C. The vertical line is for the case of 5°C PETM warming. The orange lines show the required CO2 if the CO2 increase is accompanied by a non-CO2 GHG feedback that provides 25% of the total GHG forcing.
This analysis is for Earth system sensitivity with CO2 as the forcing, as is appropriate for the PETM because any carbon injected as CH4 would be rapidly oxidized to CO2. Feedbacks in the PETM do not include large ice sheets, but non-CO2 GHGs are an unmeasured feedback. If a warming climate increases the amount of N2O and CH4 in the air, the required carbon source for a given global warming is reduced, because the amount of carbon in airborne CH4 is negligible. Any non-CO2 GHG feedback increases the CO2-forced Earth system sensitivity, potentially by a large amount (figure 10). The CO2-forced Earth sensitivity is the most relevant climate sensitivity, not only for the PETM but for human-made forcings. Although present enhanced amounts of airborne CH4 and N2O are mostly a climate forcing, i.e. their increases above the pre-industrial level are mainly a consequence of human-made sources, they also include a GHG feedback. Climate sensitivity including this GHG feedback is the most relevant sensitivity for humanity as the CO2 forcing continues to grow. If the EECO global temperature exceeded 28°C, as suggested by multi-proxy data taken at face value (see above), climate sensitivity implied by the EECO warmth and the PETM warming is close to the full Russell climate sensitivity (see electronic supplementary material, figures S7–S9). We conclude that the existing data favour a climate sensitivity of at least two-thirds of the Russell sensitivity, and probably closer to the full Russell sensitivity. That lower limit is just over 3°C for 2×CO2 for the range of climate states of immediate relevance to humanity (figure 7b). Covariation of climate, sea level and atmospheric CO2 through the Cenozoic era is a rich source of information that can advise us about the sensitivity of climate and ice sheets to forcings, including human-made forcings. Our approach is to estimate Cenozoic sea level and temperature from empirical data, with transparent assumptions and minimal modelling. Our data are available in the electronic supplementary material, allowing comparison with other data and model results. Hansen [49,50] argues that real ice sheets are more responsive to warming than in most ice sheet models, which suggests that large ice sheets are relatively stable. The model of Pollard & DeConto [124], for example, requires three to four times the pre-industrial CO2 amount to melt the Antarctic ice sheet. This stability is, in part, a result of hysteresis: as the Earth warms, the ice sheet size as a function of temperature does not return on the same path that it followed as temperature fell and the ice sheet grew. We do not question the reality of mechanisms that cause ice sheet hysteresis, but we suspect they are exaggerated in models. Thus, as an extreme alternative that can be compared with ice sheet models and real-world data, we assume that hysteresis effects are negligible in our approximation for sea level as a function of temperature. Ice sheets in question are those on Greenland and Antarctica, ice sheets that could shrink with future warming. Despite the stability of those ice sheets in the Holocene, there is evidence that sea level was much more variable during the Eemian, when we estimate the peak global temperature was only +1.0°C warmer than in the first decade of the twenty-first century. Rohling et al. [52] estimate an average rate of Eemian sea-level change of 1.4 m per century, and several studies noted above suggest that the Eemian sea level reached heights of +4–6 m or more relative to today. The MMCO provides one test of hysteresis. Our sea-level approximation (figure 2) suggests that the Antarctic ice sheet nearly disappeared at that time. John et al. [125] provide support for that interpretation, as well as evidence of numerous rises and falls of sea level by 20–30 m during the Miocene. These variations are even larger than those we find (figure 2), but the resolution of the δ18O data we use is not adequate to provide the full amplitude of variations during that period (electronic supplementary material, figure S1). The Mid-Pliocene is a more important test of ice sheet variability. We find sea-level fluctuations of at least 20–40 m, much greater than in ice sheet models (figure 2), with global temperature variations of only a few degrees. Independent analyses designed to separate ice volume and temperature change, such as Dwyer & Chandler [64], find sea-level maxima and variability comparable to our estimates. Altogether, the empirical data support a high sensitivity of the sea level to global temperature change, and they provide strong evidence against the seeming lethargy and large hysteresis effects that occur in at least some ice sheet models. Estimates of climate sensitivity cover a wide range that has existed for decades [1,48,99]. That range measures our ignorance; it does not mean that climate response from a specified state is stochastic with such inherent uncertainty. God (Nature) plays dice, but not for such large amounts. Indeed, one implication of the tight fit of calculated and measured temperature change of the past 800 000 years (figure 6) is that there is a single well-defined, but unknown, fast-feedback global climate sensitivity for that range of climate, despite large regional climate variations and ocean dynamical effects [31]. Improved empirical data can define climate sensitivity much more precisely, provided that climate-induced aerosol changes are included in the category of fast feedbacks (human-made aerosol changes are a climate forcing). Empirical assessment of fast-feedback climate sensitivity is obtained by comparing two quasi-equilibrium climate states for which boundary condition climate forcings (which may be slow feedbacks) are known. Aerosol changes between those climate states are appropriately included as a fast feedback, not only because aerosols respond rapidly to changing climate but also because there are multiple aerosol compositions, they have complex radiative properties and they affect clouds in several ways, thus making accurate knowledge of their glacial–interglacial changes inaccessible. The temporal variation of the GHG plus surface albedo climate forcing closely mimics the temporal variation of either the deep ocean temperature (figure 6) or Antarctic temperature [5,31] for the entire 800 000 years of polar ice core data. However, the temperature change must be converted to the global mean to allow inference of climate sensitivity. The required scale factor is commonly extracted from an estimated LGM–Holocene global temperature change, which, however, is very uncertain, with estimates ranging from approximately 3°C to approximately 6°C. Thus, for example, the climate sensitivity (1.7–2.6°C for 2×CO2) estimated by Schmittner et al. [94] is due largely to their assumed approximately 3°C cooling in the LGM, and in lesser part to the fact that they defined some aerosol changes (dust) to be a climate forcing. Climate sensitivity extracted from Pleistocene climate change is thus inherently partly subjective as it depends on how much weight is given to mutually inconsistent estimates of glacial-to-interglacial global temperature change. Our initial assessment is a fast-feedback sensitivity of 3±1°C for 2×CO2, corresponding to an LGM cooling of 4.5°C, similar to the 2.2–4.8°C estimate of PALAEOSENS [99]. This sensitivity is higher than estimated by Schmittner et al. [94], partly because they included natural aerosol changes as a forcing. In addition, we note that their proxies for LGM sea surface cooling exclude planktic foraminifera data, which suggest larger cooling [126], and, as noted by Schneider von Deimling et al. [95], regions that are not sampled tend to be ones where the largest cooling is expected. It should be possible to gain consensus on a narrower range for climate sensitivity via a community project for the LGM analogous to PRISM Pliocene data reconstruction [97,98] and PlioMIP model intercomparisons [67,68]. However, we suggest that an even more fruitful approach would be a focused effort to define the glacial-to-interglacial climate change of the Eemian period (MIS-5e). The Eemian avoids the possibility of significant human-made effects, which may be a factor in the Holocene. Ruddiman [127] suggests that deforestation and agricultural activities affected CO2 and CH4 in the Holocene, and Hansen et al. [91] argue that human-made aerosols were probably important. Given the level of Eemian warmth, approximately +1.8°C relative to 1880–1920, with a climate forcing similar to that for LGM–Holocene (figure 5), we conclude that this relatively clean empirical assessment yields a fast-feedback climate sensitivity in the upper part of the range suggested by the LGM–Holocene climate change, i.e. a sensitivity of 3–4°C for 2×CO2. Detailed study is especially warranted because Eemian warmth is anticipated to recur in the near term. We have shown that global temperature change over the Cenozoic era is consistent with CO2 change being the climate forcing that drove the long-term climate change. Proxy CO2 measurements are so variable and uncertain that we only rely on the conclusion that the CO2 amount was of the order of 1000 ppm during peak Early Eocene warmth. That conclusion, in conjunction with a climate model incorporating only the most fundamental processes, constrains average fast-feedback climate sensitivity to be in the upper part of the sensitivity range that is normally quoted [1,48,99], i.e. the sensitivity is greater than 3°C for 2×CO2. Strictly this Cenozoic evaluation refers to the average fast-feedback sensitivity for the range of climates from ice ages to peak Cenozoic warmth and to the situation at the time of the PETM. However, it would be difficult to achieve that high average sensitivity if the current fast-feedback sensitivity were not at least in the upper half of the range of 3±1°C for 2×CO2. This climate sensitivity evaluation has implications for the atmospheric CO2 amount throughout the Cenozoic era, which can be checked as improved proxy CO2 measurements become available. The CO2 amount was only approximately 450–500 ppm 34 Myr BP when large-scale glaciation first occurred on Antarctica. Perhaps more important, the amount of CO2 required to melt most of Antarctica in the MMCO was only approximately 450–500 ppm, conceivably only about 400 ppm. These CO2 amounts are smaller than suggested by ice sheet/climate models, providing further indication that the ice sheet models are excessively lethargic, i.e. resistant to climate change. The CO2 amount in the earliest Pliocene, averaged over astronomical cycles, was apparently only about 300 ppm, and decreased further during the Pliocene. Our climate simulations, using a simplified three-dimensional climate model to solve the fundamental equations for conservation of water, atmospheric mass, energy, momentum and the ideal gas law, but stripped to basic radiative, convective and dynamical processes, finds upturns in climate sensitivity at the same forcings as found with a more complex global climate model [66]. At forcings beyond these points the complex model ‘crashed’, as have other climate models (discussed by Lunt et al. [83]). The upturn at the 10–20 W m−2 negative forcing has a simple physical explanation: it is the snowball Earth instability. Model crashes for large positive forcings are sometimes described as a runaway greenhouse, but they probably are caused by one of the many parametrizations in complex global models going outside its range of validity, not by a runaway greenhouse effect. The runaway greenhouse effect has several meanings ranging from, at the low end, global warming sufficient to induce out-of-control amplifying feedbacks, such as ice sheet disintegration and melting of methane hydrates, to, at the high end, a Venus-like hothouse with crustal carbon baked into the atmosphere and a surface temperature of several hundred degrees, a climate state from which there is no escape. Between these extremes is the moist greenhouse, which occurs if the climate forcing is large enough to make H2O a major atmospheric constituent [106]. In principle, an extreme moist greenhouse might cause an instability with water vapour preventing radiation to space of all absorbed solar energy, resulting in very high surface temperature and evaporation of the ocean [105]. However, the availability of non-radiative means for vertical transport of energy, including small-scale convection and large-scale atmospheric motions, must be accounted for, as is done in our atmospheric general circulation model. Our simulations indicate that no plausible human-made GHG forcing can cause an instability and runaway greenhouse effect as defined by Ingersoll [105], in agreement with the theoretical analyses of Goldblatt & Watson [128]. On the other hand, conceivable levels of human-made climate forcing could yield the low-end runaway greenhouse. A forcing of 12–16 W m−2, which would require CO2 to increase by a factor of 8–16 times, if the forcing were due only to CO2 change, would raise the global mean temperature by 16–24°C with much larger polar warming. Surely that would melt all the ice on the planet, and probably thaw methane hydrates and scorch carbon from global peat deposits and tropical forests. This forcing would not produce the extreme Venus-like baked-crust greenhouse state, which cannot be reached until the ocean is lost to space. A warming of 16–24°C produces a moderately moist greenhouse, with water vapour increasing to about 1% of the atmosphere's mass, thus increasing the rate of hydrogen escape to space. However, if the forcing is by fossil fuel CO2, the weathering process would remove the excess atmospheric CO2 on a time scale of 104–105 years, well before the ocean is significantly depleted. Baked-crust hothouse conditions on the Earth require a large long-term forcing that is unlikely to occur until the sun brightens by a few tens of per cent, which will take a few billion years [121]. Burning all fossil fuels would produce a different, practically uninhabitable, planet. Let us first consider a 12 W m−2 greenhouse forcing, which we simulated with 8×CO2. If non-CO2 GHGs such as N2O and CH4 increase with global warming at the same rate as in the palaeoclimate record and atmospheric chemistry simulations [122], these other gases provide approximately 25% of the greenhouse forcing. The remaining 9 W m−2 forcing requires approximately 4.8×CO2, corresponding to fossil fuel emissions as much as approximately 10,000 Gt C for a conservative assumption of a CO2 airborne fraction averaging one-third over the 1000 years following a peak emission [21,129]. Our calculated global warming in this case is 16°C, with warming at the poles approximately 30°C. Calculated warming over land areas averages approximately 20°C. Such temperatures would eliminate grain production in almost all agricultural regions in the world [130]. Increased stratospheric water vapour would diminish the stratospheric ozone layer [131]. More ominously, global warming of that magnitude would make most of the planet uninhabitable by humans [132,133]. The human body generates about 100 W of metabolic heat that must be carried away to maintain a core body temperature near 37°C, which implies that sustained wet bulb temperatures above 35°C can result in lethal hyperthermia [132,134]. Today, the summer temperature varies widely over the Earth's surface, but wet bulb temperature is more narrowly confined by the effect of humidity, with the most common value of approximately 26–27°C and the highest approximately of 31°C. A warming of 10–12°C would put most of today's world population in regions with wet a bulb temperature above 35°C [132]. Given the 20°C warming we find with 4.8×CO2, it is clear that such a climate forcing would produce intolerable climatic conditions even if the true climate sensitivity is significantly less than the Russell sensitivity, or, if the Russell sensitivity is accurate, the CO2 amount required to produce intolerable conditions for humans is less than 4.8×CO2. Note also that increased heat stress due to warming of the past few decades is already enough to affect health and workplace productivity at low latitudes, where the impact falls most heavily on low- and middle-income countries [135]. The Earth was 10–12°C warmer than today in the Early Eocene and at the peak of the PETM (figure 4). How did mammals survive that warmth? Some mammals have higher internal temperatures than humans and there is evidence of evolution of surface-area-to-mass ratio to aid heat dissipation, for example transient dwarfing of mammals [136] and even soil fauna [137] during the PETM warming. However, human-made warming will occur in a few centuries, as opposed to several millennia in the PETM, thus providing little opportunity for evolutionary dwarfism to alleviate impacts of global warming. We conclude that the large climate change from burning all fossil fuels would threaten the biological health and survival of humanity, making policies that rely substantially on adaptation inadequate. Let us now verify that our assumed fossil fuel climate forcing of 9 W m−2 is feasible. If we assume that fossil fuel emissions increase by 3% per year, typical of the past decade and of the entire period since 1950, cumulative fossil fuel emissions will reach 10 000 Gt C in 118 years. However, with such large rapidly growing emissions the assumed 33% CO2 airborne fraction is surely too small. The airborne fraction, observed to have been 55% since 1950 [1], should increase because of well-known nonlinearity in ocean chemistry and saturation of carbon sinks, implying that the airborne fraction probably will be closer to two-thirds rather than one-third, at least for a century or more. Thus, the fossil fuel source required to yield a 9 W m−2 forcing may be closer to 5000 Gt C, rather than 10 000 Gt C. Are there sufficient fossil fuel reserves to yield 5000–10 000 Gt C? Recent updates of potential reserves [114], including unconventional fossil fuels (such as tar sands, tar shale and hydrofracking-derived shale gas) in addition to conventional oil, gas and coal, suggest that 5×CO2 (1400 ppm) is indeed feasible. For instance, using the emission factor for coal from IPCC [48], coal resources given by the Global Energy Assessment [114] amount to 7300–11 000 Gt C. Similarly, using emission factors from IPCC [48], total recoverable fossil energy reserves and resources estimated by GEA [114] are approximately 15 000 Gt C. This does not include large ‘additional occurrences’ listed in ch. 7 of GEA [114]. Thus, for a multi-centennial CO2 airborne fraction between one-third and two-thirds, as discussed above, there are more than enough available fossil fuels to cause a forcing of 9 W m−2 sustained for centuries. Most of the remaining fossil fuel carbon is in coal and unconventional oil and gas. Thus, it seems, humanity stands at a fork in the road. As conventional oil and gas are depleted, will we move to carbon-free energy and efficiency—or to unconventional fossil fuels and coal? If fossil fuels were made to pay their costs to society, costs of pollution and climate change, carbon-free alternatives might supplant fossil fuels over a period of decades. However, if governments force the public to bear the external costs and even subsidize fossil fuels, carbon emissions are likely to continue to grow, with deleterious consequences for young people and future generations. It seems implausible that humanity will not alter its energy course as consequences of burning all fossil fuels become clearer. Yet strong evidence about the dangers of human-made climate change have so far had little effect. Whether governments continue to be so foolhardy as to allow or encourage development of all fossil fuels may determine the fate of humanity. We thank James Zachos for the deep ocean oxygen isotope data; Chris Brierly, Mark Chandler, Bas de Boer, Alexey Fedorov, Chris Hatfield, Dorothy Peteet, David Rind, Robert Rohde and Cynthia Rosenzweig for helpful information; Andy Ridgwell for useful editorial suggestions and patience; Eelco Rohling for ably organizing the palaeoclimate workshop that spurred the writing of this paper; Gerry Lenfest (Lenfest Foundation), ClimateWorks, Lee Wasserman (Rockefeller Family Foundation), Stephen Toben (Flora Family Foundation) and NASA program managers Jack Kaye and David Considine for research support. FootnotesOne contribution of 11 to a Discussion Meeting Issue ‘Warm climates of the past—a lesson for the future?’. © 2013 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/3.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 6
You have to enable JavaScript in your browser's settings in order to use the eReader. Or try downloading the content offline DOWNLOAD |
The last interglacial period with temperatures similar to the present interglacial period was the
zusammenhängende Posts
Urheberrechte © © 2024 wiewird Inc.
