Atmosphere–ocean climate models, including ours, commonly include a fixed freshwater flux from the Greenland and Antarctic ice sheets to the ocean. This flux is chosen to balance snow accumulation in the model's control run, with the rationale that approximate balance is expected between net accumulation and mass loss including icebergs and ice shelf melting. Global warming creates a mass imbalance that we want to investigate. Ice sheet models can calculate the imbalance, but it is unclear how reliably ice sheet models simulate ice sheet disintegration. We forgo ice sheet modeling, instead adding a growing freshwater amount to polar oceans with alternative growth rates and initial freshwater amount estimated from available data.

Change of freshwater flux into the ocean in a warming world with shrinking ice sheets consists of two terms: term 1 being net ice melt and term 2 being change in P-E (precipitation minus evaporation) over the relevant ocean. Term 1 includes land based ice mass loss, which can be detected by satellite gravity measurements, loss of ice shelves, and net sea ice mass change. Term 2 is calculated in a climate model forced by changing atmospheric composition, but it is not included in our pure freshwater experiments that have no global warming.

IPCC (Vaughan et al., 2013) estimated land ice loss in Antarctica that increased from 30 Gt year-1 in 1992–2001 to 147 Gt year-1 in 2002–2011 and in Greenland from 34 to 215 Gt year-1, with uncertainties discussed by Vaughan et al. (2013). Gravity satellite data suggest Greenland ice sheet mass loss ∼ 300–400 Gt year-1 in the past few years (Barletta et al., 2013). A newer analysis of gravity data for 2003–2013 (Velicogna et al., 2014), discussed in more detail in Sect. 5.1, finds a Greenland mass loss 280 ± 58 Gt year-1 and Antarctic mass loss 67 ± 44 Gt year-1.

One estimate of net ice loss from Antarctica, including ice shelves, is obtained by surveying and adding the mass flux from all ice shelves and comparing this freshwater mass loss with the freshwater mass gain from the continental surface mass budget. Rignot et al. (2013) and Depoorter et al. (2013) independently assessed the freshwater mass fluxes from Antarctic ice shelves. Their respective estimates for the basal melt are 1500 ± 237 and 1454 ± 174 Gt year-1. Their respective estimates for calving are 1265 ± 139 and 1321 ± 144 Gt year-1.

This estimated freshwater loss via the ice shelves (∼ 2800 Gt year-1) is larger than freshwater gain by Antarctica. Vaughan et al. (1999) estimated net surface mass balance of the continent as +1811 and +2288 Gt year-1 including precipitation on ice shelves. Vaughan et al. (2013) estimates the net Antarctic surface mass balance as +1983 ± 122 Gt year-1 excluding ice shelves. Thus comparison of continental freshwater input with ice shelf output suggests a net export of freshwater to the Southern Ocean of several hundred Gt year-1 in recent years. However, substantial uncertainty exists in the difference between these two large numbers.

An independent evaluation has recently been achieved by Rye et al. (2014) using satellite measured changes of sea level around Antarctica in the period 1992–2011. Sea level along the Antarctic coast rose 2 mm year-1 faster than the regional mean sea level rise in the Southern Ocean south of 50∘ S, an effect that they conclude is almost entirely a steric adjustment caused by accelerating freshwater discharge from Antarctica. They conclude that an excess freshwater input of 430  ± 230 Gt year-1, above the rate needed to maintain a steady ocean salinity, is required. Rye et al. (2014) note that these values constitute a lower bound for the actual excess discharge above a “steady salinity” rate, because numerous in situ data, discussed below, indicate that freshening began earlier than 1992.

Term 2, change in P-E over the Southern Ocean relative to its preindustrial amount, is large in our climate simulations. In our ensemble of runs (using observed GHGs for 1850–2003 and scenario A1B thereafter) the increase in P-E in the decade 2011–2020, relative to the control run, was in the range 3500 to 4000 Gt year-1, as mean precipitation over the Southern Ocean increased ∼ 35 mm year-1 and evaporation decreased ∼ 3 mm year-1.

Increasing ice melt and increasing P-E are climate feedbacks, their growth in recent decades driven by global warming. Our pure freshwater simulations indicate that their sum, at least 4000 Gt year-1, is sufficient to affect ocean circulation, sea ice cover, and surface temperature, which can spur other climate feedbacks. We investigate these feedbacks via climate simulations using improved estimates of freshwater flux from ice melt. P-E is computed by the model.

We take freshwater injection to be 720 Gt year-1 from Antarctica and 360 Gt year-1 in the North Atlantic in 2011, with injection rates at earlier and later times defined by assumption of a 10-year doubling time. Resulting mean freshwater injection around Antarctica in 1992–2011 is ∼ 400 Gt year-1, similar to the estimate of Rye et al. (2014). A recent estimate of 310 ± 74 km3 volume loss of floating Antarctic ice shelves in 2003–2012 (Paolo et al., 2015) is not inconsistent, as the radar altimeter data employed for ice shelves do not include contributions from the ice sheet or fast ice tongues at the ice shelf grounding line. Greenland ice sheet mass loss provides most of the assumed 360 Gt year-1 freshwater, and this would be supplemented by shrinking ice shelves (Rignot and Steffen, 2008) and small ice caps in the North Atlantic and west of Greenland (Ohmura, 2009) that are losing mass (Abdalati et al., 2004; Bahr et al., 2009).

We add freshwater around Antarctica at coastal grid boxes (Fig. S13) guided by the data of Rignot et al. (2013) and Depoorter et al. (2013). Injection in the Western Hemisphere, especially from the Weddell Sea to the Ross Sea, is more than twice that in the other hemisphere (Fig. 14). Specified freshwater flux around Greenland is similar on the east and west coasts, and small along the north coast (Fig. S13).

Actual GHG forcing is less than scenario A1B, because CH4 and minor gas growth declined after IPCC scenarios were defined (Fig. 5; Hansen et al., 2013c, update at http://www.columbia.edu/~mhs119/GHGs/). As a simple improvement we decreased the A1B CH4 scenario during 2003–2013 so that subsequent CH4 is reduced 100 ppb, decreasing radiative forcing ∼ 0.05 W m-2.

Stratospheric aerosol forcing to 2014 uses the data set of Sato et al. (1993) as updated at http://www.columbia.edu/~mhs119/StratAer/. Future years have constant aerosol optical depth 0.0052 yielding effective forcing -0.12 W m-2, implemented by using fixed 1997 aerosol data. Tropospheric aerosol growth is assumed to slow smoothly, leveling out at -2 W m-2 in 2100. Future solar forcing is assumed to have an 11-year cycle with amplitude 0.25 W m-2. Net forcing exceeds 5 W m-2 by the end of the 21st century, about 3 times the current forcing (Fig. S16).

Global temperature has a maximum at +1.2 ∘C in the 2040s for the modified forcings (Fig. 15). Ice melt cooling is advanced as global ice melt reaches 1 m of sea level in 2060, 1/3 from Greenland and 2/3 from Antarctica. Global temperature rise resumes in the 2060s after cessation of freshwater injection.

Global temperature becomes an unreliable diagnostic of planetary condition as the ice melt rate increases. Global energy imbalance (Fig. 15b) is a more meaningful measure of planetary status as well as an estimate of the climate forcing change required to stabilize climate. Our calculated present energy imbalance of ∼ 0.8 W m-2 (Fig. 15b) is larger than the observed 0.58 ± 0.15 W m-2 during 2005–2010 (Hansen et al., 2011). The discrepancy is likely accounted for by excessive ocean heat uptake at low latitudes in our model, a problem related to the model's slow surface response time (Fig. 4) that may be caused by excessive small-scale ocean mixing.

Large scale regional cooling occurs in the North Atlantic and Southern oceans by mid-century (Fig. 16) for 10-year doubling of freshwater injection. A 20-year doubling places similar cooling near the end of this century, 40 years earlier than in our prior simulations (Fig. 7), as the factor of 4 increase in current freshwater from Antarctica is a 40-year advance.

Cumulative North Atlantic freshwater forcing in sverdrup years (Sv years) is 0.2 Sv years in 2014, 2.4 Sv years in 2050, and 3.4 Sv years (its maximum) prior to 2060 (Fig. S14). The critical issue is whether human-spurred ice sheet mass loss can be approximated as an exponential process during the next few decades. Such nonlinear behavior depends upon amplifying feedbacks, which, indeed, our climate simulations reveal in the Southern Ocean.

Amplifying feedbacks in the Southern Ocean and atmosphere contribute to dramatic climate change in our simulations (Fig. 16). We first summarize the feedbacks to identify processes that must be simulated well to draw valid conclusions. While recognizing the complexity of the global ocean circulation (Lozier, 2012; Lumpkin and Speer, 2007; Marshall and Speer, 2012; Munk and Wunsch, 1998; Orsi et al., 1999; Sheen et al., 2014; Talley, 2013; Wunsch and Ferrari, 2004), we use a simple two-dimensional representation to discuss the feedbacks.

Climate change includes slowdown of AABW formation, indeed shutdown by mid-century if freshwater injection increases with a doubling time as short as 10 years (Fig. 17). Implications of AABW shutdown are so great that we must ask whether the mechanisms are simulated with sufficient realism in our climate model, which has coarse resolution and relevant deficiencies that we have noted. After discussing the feedbacks here, we examine how well the processes are included in our model (Sect. 3.8.5). Paleoclimate data (Sect. 4) provide much insight about these processes, and modern observations (Sect. 5) suggest that these feedbacks are already underway.

Large-scale climate processes affecting ice sheets are sketched in Fig. 18. The role of the ocean circulation in the global energy and carbon cycles is captured to a useful extent by the two-dimensional (zonal-mean) overturning circulation featuring deep water (NADW) and bottom water (AABW) formation in the polar regions. Marshall and Speer (2012) discuss the circulation based in part on tracer data and analyses by Lumpkin and Speer (2007). Talley (2013) extends the discussion with diagrams clarifying the role of the Pacific and Indian oceans.

Wunsch (2002) emphasizes that the ocean circulation is driven primarily by atmospheric winds and secondarily by tidal stirring. Strong circumpolar westerly winds provide energy drawing deep water toward the surface in the Southern Ocean. Ocean circulation also depends on processes maintaining the ocean's vertical density stratification. Winter cooling of the North Atlantic surface produces water dense enough to sink (Fig. 18), forming North Atlantic Deep Water (NADW). However, because North Atlantic water is relatively fresh, compared to the average ocean, NADW does not sink all the way to the global ocean bottom. Bottom water is formed instead in the winter around the Antarctic coast, where very salty cold water (AABW) can sink to the ocean floor. This ocean circulation (Fig. 18) is altered by natural and human-made forcings, including freshwater from ice sheets, engendering powerful feedback processes.

A key Southern Ocean feedback is meltwater stratification effect, which reduces ventilation of ocean heat to the atmosphere and space. Our “pure freshwater” experiments show that the low-density lid causes deep-ocean warming, especially at depths of ice shelf grounding lines that provide most of the restraining force limiting ice sheet discharge (Fig. 14 of Jenkins and Doake, 1991). West Antarctica and Wilkes Basin in East Antarctica have potential to cause rapid sea level rise, because much of their ice sits on retrograde beds (beds sloping inland), a situation that can lead to unstable grounding line retreat and ice sheet disintegration (Mercer, 1978).

Another feedback occurs via the effect of surface and atmospheric cooling on precipitation and evaporation over the Southern Ocean. CMIP5 climate simulations, which do not include increasing freshwater injection in the Southern Ocean, find snowfall increases on Antarctica in the 21st century, thus providing a negative term to sea level change. Frieler et al. (2015) note that 35 climate models are consistent in showing that warming climate yields increasing snow accumulation in accord with paleo-data for warmer climates, but the paleo-data refer to slowly changing climate in quasi-equilibrium with ocean boundary conditions. In our experiments with growing freshwater injection, the increasing sea ice cover and cooling of the Southern Ocean surface and atmosphere cause the increased precipitation to occur over the Southern Ocean, rather than over Antarctica. This feedback not only reduces any increase in snowfall over Antarctica but also provides a large freshening term to the surface of the Southern Ocean, thus magnifying the direct freshening effect from increasing ice sheet melt.

North Atlantic meltwater stratification effects are also important, but different. Meltwater from Greenland can slow or shutdown NADW formation, cooling the North Atlantic, with global impacts even in the Southern Ocean, as we will discuss later. One important difference is that the North Atlantic can take centuries to recover from NADW shutdown, while the Southern Ocean recovers within 1–2 decades after freshwater injection stops (Sect. 3.7).

Realistic representation of these feedbacks places requirements on both the atmosphere and ocean components of our climate model. We discuss first the atmosphere, then the ocean.

There are two main requirements on the atmospheric model. First, it must simulate P-E well, because of its importance for ocean circulation and the amplifying feedback in the Southern Ocean. Second, it must simulate winds well, because these drive the ocean.

Simulated P-E (Fig. S15b) agrees well with meteorological reanalysis (Fig. 3.4b of Rhein et al., 2013). Resulting sea surface salinity (SSS) patterns in the model (Fig. S15a) agree well with global ocean surface salinity patterns (Antonov et al., 2010, and Fig. 3.4a of Rhein et al., 2013). SSS trends in our simulation (Fig. S15c), with the Pacific on average becoming fresher while most of the Atlantic and the subtropics in the Southern Hemisphere become saltier, are consistent with observed salinity trends (Durack and Wijffels, 2010). Recent freshening of the Southern Ocean in our simulation is somewhat less than in observed data (Fig. 3.4c, d of Rhein et al., 2013), implying that the amplifying feedback may be underestimated in our simulation. A likely reason for that is discussed below in conjunction with observed sea ice change.

Obtaining accurate winds requires the model to simulate well atmospheric pressure patterns and their change in response to climate forcings. A test is provided by observed changes of the Southern Annular Mode (SAM), with a decrease in surface pressure near Antarctica and a small increase at midlatitudes (Marshall, 2003) that Thompson et al. (2011) relate to stratospheric ozone loss and increasing GHGs. Our climate forcing (Fig. S16) includes ozone change (Fig. 2 of Hansen et al., 2007a) with stratospheric ozone depletion in 1979–1997 and constant ozone thereafter. Our model produces a trend toward the high index polarity of SAM (Fig. S17) similar to observations, although perhaps a slightly smaller change than observed (compare Fig. S17 with Fig. 3 of Marshall, 2003). SAM continues to increase in our model after ozone stabilizes (Fig. S17), suggesting that GHGs may provide a larger portion of the SAM response in our model than in the model study of Thompson et al. (2011). It would not be surprising if the stratospheric dynamical response to ozone change were weak in our model, given the coarse resolution and simplified representation of atmospheric drag and dynamical effects in the stratosphere (Hansen et al., 2007a), but that is not a major concern for our present purposes.

The ocean model must be able to simulate realistically the ocean's overturning circulation and its response to forcings including freshwater additions. Heuze et al. (2013, 2015) point out that simulated deep convection in the Southern Ocean is unrealistic in most models, with AABW formation occurring in the open ocean where it rarely occurs in nature. Our present ocean model contains significant improvements (see Sect. 3.2) compared to the GISS E2-R model that Heuze et al. include in their comparisons. Thus we show (Fig. 19) the maximum mixed-layer depth in winter (February in the Northern Hemisphere and August in the Southern Hemisphere) using the same criterion as Heuze et al. to define the mixed-layer depth, i.e., the layers with a density difference from the ocean surface layer less than 0.03 kg m-3.

Southern Ocean mixing in the model reaches a depth of ∼ 500 m in a wide belt near 60∘ S stretching west from the southern tip of South America, with similar depths south of Australia. These open-ocean mixed-layer depths compare favorably with observations shown in Fig. 2a of Heuze et al. (2015), based on data of de Boyer Montegut et al. (2004). There is no open-ocean deep convection in our model.

Deep convection occurs only along the coast of Antarctica (Fig. 19). Coastal grid boxes on the continental shelf are a realistic location for AABW formation. Orsi et al. (1999) suggest that most AABW is formed on shelves around the Weddell–Enderby Basin (60 %) and shelves of the Adélie–Wilkes Coast and Ross Sea (40 %). Our model produces mixing down to the shelf in those locations (Fig. 19b), and also on the Amery Ice Shelf near the location where Ohshima et al. (2013) identified AABW production, which they term Cape Darnley Bottom Water.

With our coarse 4∘ stair step to the ocean bottom, AABW cannot readily slide down the slope to the ocean floor. Thus dense shelf water mixes into the open-ocean grid boxes, making our modeled Southern Ocean less stratified than the real world (cf. temporal drift of Southern Ocean salinity in Fig. S18), because the denser water must move several degrees of latitude horizontally before it can move deeper. Nevertheless, our Southern Ocean is sufficiently stratified to avoid the unrealistic open-ocean convection that infects many models (Heuze et al., 2013, 2015).

Orsi et al. (1999) estimate the AABW formation rate in several ways, obtaining values in the range 8–12 Sv, larger than our modeled 5–6 Sv (Fig. 17). However, as in most models (Heuze et al., 2015), our SMOC diagnostic (Fig. 17) is the mean (Eulerian) circulation, i.e., excluding eddy-induced transport. Rerun of a 20-year segment of our control run to save eddy-induced changes reveals an increase in SMOC at 72∘ S by 1–2 Sv, with negligible change at middle and low latitudes, making our simulated transport close to the range estimated by Orsi et al. (1999).

We conclude that the model may simulate Southern Ocean feedbacks that magnify the effect of freshwater injected into the Southern Ocean: the P-E feedback that wrings global-warming-enhanced water vapor from the air before it reaches Antarctica and the AABW slowdown that traps deep-ocean heat, leaving that heat at levels where it accelerates ice shelf melting. Indeed, we will argue that both of these feedbacks are probably underestimated in our current model.

The model seems less capable in Northern Hemisphere polar regions. Deep convection today is believed to occur mainly in the Greenland–Iceland–Norwegian (GIN) seas and at the southern end of Baffin Bay (Fig. 2b of Heuze et al., 2015). In our model, perhaps because of excessive sea ice in those regions, open-ocean deep convection occurs to the southeast of the southern tip of Greenland and at less deep grid boxes between that location and the United Kingdom (Fig. 19). Mixing reaching the ocean floor on the Siberian coast in our model (Fig. 19) may be realistic, as coastal polynya are observed on the Siberian continental shelf (D. Bauch et al., 2012). However, the winter mixed layer on the Alaska south coast is unrealistically deep (Fig. 19). These model limitations must be kept in mind in interpreting simulated Northern Hemisphere climate change.

Ice melt in the North Atlantic increases simulated sea level pressure in that region in all seasons (Fig. 20). In summer the Bermuda high-pressure system (Fig. S2) increases in strength and moves northward. Circulation around the high pressure creates stronger prevailing northeasterly winds at latitudes of Bermuda and the Bahamas. A1B climate forcing alone (Fig. S21, top row) has only a small impact on the winds, but cold meltwater in the North Atlantic causes a strengthening and poleward shift of the high pressure.

The high pressure in the model is located further east than needed to produce the fastest possible winds at the Bahamas. Our coarse-resolution (4∘ × 5∘) model may be partly responsible for the displacement. However, the location of high pressure also depends on meltwater placement, which we spread uniformly over all longitudes in the North Atlantic between 65∘ W and 15∘ E, and on the specific location of ocean currents and surface temperature during the Eemian.

North Atlantic cooling from AMOC shutdown creates faster winds in our simulations, with a seasonal-mean increment as much as 10–20 %. Such a percentage translates into an increase in storm power dissipation by a factor ∼ 1.4–2, because dissipation is proportional to the cube of wind speed (Emanuel, 1987, 2005). Our simulated changes refer to mean winds over large grid boxes, not individual storms, for which the change in the most extreme cases might be larger.

Increased North Atlantic high pressure strengthens prevailing northeasterly winds blowing onto the Bahamas in the direction of Eemian wave-formed deposits (Sect. 4.1.2). Consistent increase in these winds would contribute to creation of long-wavelength, deep-ocean waves that scour the ocean floor as they reach the shallow near-shore region. However, extreme events may require the combined effect of increased prevailing winds and tropical storms guided by the strengthened blocking high pressure and nurtured by the unusually warm late-Eemian tropical sea surface temperatures (Cortijo et al., 1999), which would favor more powerful tropical storms (Emanuel, 1987). This enhanced meridional temperature gradient – warmer tropics and cooler high latitudes – was enhanced by low obliquity of Earth's spin axis in the late Eemian.

If GHGs continue to increase rapidly and ice melt grows, our simulations yield shutdown or major slowdown of the AMOC in the 21st century, implying an increase in severe weather. This is shown by zonal-mean temperature and eddy kinetic energy changes in simulations of Sects. 3.3–3.6 with and without ice melt (Fig. 21). Without ice melt, surface warming is largest in the Arctic (Fig. 21, left), resulting in a decrease in lower tropospheric eddy energy. However, the surface cooling from ice melt increases surface and lower tropospheric temperature gradients, and in stark contrast to the case without ice melt, there is a large increase in midlatitude eddy energy throughout the midlatitude troposphere. The increase in zonal-mean midlatitude baroclinicity (Fig. 21) is in agreement with the localized, North Atlantic-centered increases in baroclinicity found in the higher-resolution simulations of Jackson et al. (2015) and Brayshaw et al. (2009).

Increased baroclinicity produced by a stronger temperature gradient provides energy for more severe weather events. Many of the most significant and devastating storms in eastern North America and western Europe, popularly known as superstorms, have been winter cyclonic storms, though sometimes occurring in late fall or early spring, that generate near-hurricane-force winds and often large amounts of snowfall (Chapter 11, Hansen, 2009). Continued warming of low-latitude oceans in coming decades will provide a larger water vapor repository that can strengthen such storms. If this tropical warming is combined with a cooler North Atlantic Ocean from AMOC slowdown and an increase in midlatitude eddy energy (Fig. 21), we can anticipate more severe baroclinic storms. Increased high pressure due to cooler high-latitude ocean (Fig. 20) can make blocking situations more extreme, with a steeper pressure gradient between the storm's low-pressure center and the blocking high, thus driving stronger North Atlantic storms.

Freshwater injection into the North Atlantic and Southern oceans increases sea level pressure at middle latitudes and decreases it at polar latitudes (Figs. 20, S22), but the impact is different in the North Atlantic than in the Southern Ocean. In the Southern Ocean the increased meridional temperature gradient increases the strength of westerlies in all seasons at all longitudes. In the North Atlantic Ocean the increase in sea level pressure in winter slows the westerlies (Fig. 20). Thus instead of a strong zonal wind that keeps cold polar air locked in the Arctic, there is a tendency for a less zonal flow and thus more cold air outbreaks to middle latitudes.

Eemian sea level is of special interest because Eemian climate was little warmer than today. Masson-Delmotte et al. (2013) conclude, based on multiple data and model sources, that peak Eemian temperature probably was only a few tenths of a degree warmer than today. Yet Eemian sea level reached heights several meters above today's level (Land et al., 1967; Chen et al., 1991; Neumann and Hearty, 1996; Hearty et al., 2007; Kopp et al., 2009; Dutton and Lambeck, 2012; O'Leary et al., 2013; Dutton et al., 2015).

Change of sea level within the Eemian period is particularly relevant to concerns about ice sheet stability and the potential for rapid sea level rise. Hearty et al. (2007) used data from 15 sites around the world to construct an Eemian sea level curve that had sea level rising in the early Eemian to +2–3 m (“+” indicates above today's sea level), possibly falling in the mid-Eemian to near today's sea level, rapidly rising in the late Eemian to +6–9 m, and then plummeting as Earth moved from the Eemian into the 100 000-year glacial period preceding the Holocene. Evidence from a variety of sources supports this interpretation, as discussed in the Supplement.

The most comprehensive analyses of sea level and paleoclimate storms are obtained by combining information from different geologic sources, each with strengths and weaknesses. Coral reefs, for example, allow absolute U/Th dating with age uncertainty as small as 1–2 ky, but inferred sea levels are highly uncertain because coral grows below sea level at variable depths as great as several meters. Carbonate platforms such as Bermuda and the Bahamas, in contrast, have few coral reefs for absolute dating, but the ability of carbonate sediments to cement rapidly preserves rock evidence of short-lived events such as rapid sea level rise and storms.

The important conclusion, that sea level rose rapidly in the late Eemian by several meters, to +6–9 m, is supported by records preserved in both the limestone platforms and coral reefs. Figure 6 of Hearty and Kindler (1995), for example, based on Bermuda and Bahamas geological data from marine and eolian limestone, reveals the rapid late-Eemian sea level rise and fall. Based on the limited size of the notches cut in Bahamian shore during the rapid late-Eemian level rise and crest, Neumann and Hearty (1996) inferred that this period was at most a few hundred years. Independently, Blanchon et al. (2009) used coral reef “back-stepping” on the Yucatán Peninsula, i.e., movement of coral reef building shoreward as sea level rises, to conclude that sea level in the late Eemian jumped 2–3 m within an “ecological” period, i.e., within several decades.

Despite general consistency among these studies, considerable uncertainty remains about absolute Eemian sea level elevation and exact timing of end-Eemian events. Uncertainties include effects of local tectonics and glacio-isostatic adjustment (GIA) of Earth's crust. Models of GIA of Earth's crust to ice sheet loading and unloading are increasingly used to improve assessments. O'Leary et al. (2013) use over 100 corals from reefs at 28 sites along the 1400 km west coast of Australia, incorporating minor GIA corrections, to conclude that sea level in most of the Eemian was relatively stable at +3–4 m, followed by a rapid late-Eemian sea level rise to about +9 m. U-series dating of the corals has peak sea level at 118.1 ± 1.4 ky b2k.

A more complete discussion of data on Eemian sea level is provided in the Supplement.

Late-Eemian sea level rise may seem a paradox, because orbital forcing then favored growth of Northern Hemisphere ice sheets. We will find evidence, however, that the sea level rise and increased storminess are consistent, and likely related to events in the Southern Ocean.

Geologic data indicate that the rapid end-Eemian sea level oscillation was accompanied by increased temperature gradients and storminess in the North Atlantic region. We summarize several interconnected lines of evidence for end-Eemian storminess, based on geological studies in Bermuda and the Bahamas referenced below. It is important to consider all the physical evidence of storminess rather than exclusively the transport mechanism of the boulders; indeed, it is essential to integrate data from obviously wave-produced runup and chevron deposits that exist within a few kilometers on North Eleuthera, Bahamas, as well as across the Bahamas Platform.

The Bahama Banks are flat, low-lying carbonate platforms that are exposed as massive islands during glacials and largely inundated during interglacial high stands. From a tectonic perspective, the platforms are relatively stable, as indicated by near-horizontal +2–3 m elevation of Eemian reef crests across the archipelago (Hearty and Neumann, 2001). During MIS 5e sea level high stands, an enormous volume of aragonitic oolitic grains blanketed the shallow, high-energy banks. Sea level shifts and storms formed shoals, ridges, and dunes. Oolitic sediments indurated rapidly (∼ 101 to ∼ 102 years) once stabilized, preserving detailed and delicate lithic evidence of these brief, high-energy events. This shifting sedimentary substrate across the banks was inimical to coral growth, which partially explains the rarity of reefs during late MIS 5e.

The preserved regional stratigraphic, sedimentary and geomorphic features attest to a turbulent end-Eemian transition in the North Atlantic. As outlined below, a coastal gradient of sedimentological features corresponds with coastal morphology, distance from the coast, and increasing elevation, reflecting the attenuating force and inland “reach” of large waves, riding on high late-Eemian sea levels. On rocky, steep coasts, giant limestone boulders were detached and catapulted onto and over the coastal ridge by ocean waves. On higher, Atlantic-facing built-up dune ridges, waves ran up to over 40 m elevation, leaving meter-thick sequences of fenestral beds, pebble lenses, and scour structures. Across kilometers of low-lying tidal inlets and flats, “nested” chevron clusters were formed as stacked, multi-meter thick, tabular fenestral beds.

The complexity of geomorphology and stratigraphy of these features are temporal measures of sustained sea level and storm events, encompassing perhaps hundreds of years. These features exclude a single wave cluster from a local point-source tsunami. Here we present data showing the connections among the megaboulders, runup deposits, and chevron ridges.

In North Eleuthera enormous boulders were plucked from seaward middle Pleistocene outcrops and washed onto a younger Pleistocene landscape (Hearty and Neumann, 2001). The average 1000 t megaclasts provide a metric of powerful waves at the end of MIS 5e. Evidence of transport by waves includes that (1) they are composed of recrystallized oolitic–peloidal limestone of MIS 9 or 11 age (300–400 ky; Kindler and Hearty, 1996) and hammer-ringing hardness; (2) they rest on oolitic sediments typical of early to mid-MIS 5e that are soft and punky under hammer blows; (3) Cerion land snail fossils beneath boulder #4 (Hearty, 1997) correlate with the last interglacial period (Garrett and Gould, 1984; Hearty and Kaufman, 2009); (4) calibrated amino acid racemization (AAR) ratios (Hearty, 1997; Hearty et al., 1998; Hearty and Kaufman, 2000, 2009) confirm the last interglacial age of the deposits as well as the stratigraphic reversal; (5) dips of bedding planes in boulders between 50 and 75∘ (Fig. 22) far exceed natural angles; and (6) some of the largest boulders are located on MIS 5e deposits at the crest of the island's ridge, proving that they are not karstic relicts of an ancient landscape (Mylroie, 2008).

The ability of storm waves to transport large boulders is demonstrated. Storms in the North Atlantic tossed boulders as large as 80 t to a height +11 m on the shore on Ireland's Aran Islands (Cox et al., 2012), this specific storm on 5 January 1991 being driven by a low-pressure system that recorded a minimum 946 mb, producing wind gusts to 80 kn and sustained winds of 40 kn for 5 h (Cox et al., 2012). Typhoon Haiyan (8 November 2013) in the Philippines produced longshore transport of a 180 t block and lifted boulders of up to ∼ 24 t to elevations as high as 10 m (May et al., 2015). May et al. (2015) conclude that these observed facts “demand a careful re-evaluation of storm-related transport where it, based on the boulder's sheer size, has previously been ascribed to tsunamis”.

The situation of the North Eleuthera megaboulders is special in two ways. First, all the large boulders are located at the apex of a horseshoe-shaped bay that would funnel energy of storm waves coming from the northeast, the direction of prevailing winds. Second, the boulders are above a vertical cliff at right angles to the incoming waves, a situation that allows constructive interference of reflected and incoming waves (Cox et al., 2012). The ability of waves hitting that cliff to produce large near-vertical splash is shown by a photograph in the Supplement taken on 31 October 1991, when a storm in the North Atlantic produced large waves impacting Eleuthera.

It is generally accepted that the boulders were wave-transported in the late Eemian. The boulders were deposited near complex chevron ridges and widespread runup deposits, which must be considered in analyzing wave-generating mechanisms. Lower elevation areas such as tidal inlets would have been flooded and scoured by the same waves, forming chevron ridges, and such large waves would also wash up onto higher, older ridges.

Across several hundred kilometers of the Bahama Islands, older built-up dune ridges are mantled with wave runup deposits that reach heights over +40 m (Fig. 23). They are generally 1–5 m thick, fenestrae-filled, and seaward-sloping tabular beds (Wanless and Dravis, 1989; Chen et al., 1991; Neumann and Hearty, 1996; Tormey and Donovan, 2015). These stratigraphically youngest Eemian deposits mantle older MIS 5e dune deposits on the shore-parallel ridges, and are the upland correlative to wave-generated boulders and chevron formations.

If these are deposits of powerful storms driven by an unusually warm tropical ocean and strong temperature gradients in the North Atlantic, as opposed to a tsunami, should there not be evidence of comparable end-Eemian storms in Bermuda? Indeed, along several kilometers of the north coast of Bermuda (Land et al., 1967; Vacher and Rowe, 1997; Hearty et al., 1998) there are seaward sloping planar beds rising to about +20 m. Although interpretations of these beds vary, they are filled with beach fenestrae and stratigraphically of latest MIS 5e carbonate sediments equivalent to runup in the Bahamas. These planar beds contrast with older MIS 5e sedimentary (dune) structures that underlie them (Hearty et al., 1998). Massive subtidal cross beds comprise the seaward facies of the elevated beach beds, pointing to an exceptional energy anomaly on the normally tranquil, shallow, broad and protected north shore platform of Bermuda.

In the Bahama Islands, extensive oolitic sand ridges with a distinctive landward-pointing V shape are common, standing ∼ 5–15 m high across several kilometers on broad, low-lying platforms or ramps throughout the Atlantic-facing, deep-water margins of the Bahamas (Hearty et al., 1998). Hearty et al. (1998) examined 35 areas with chevron ridges across the Bahamas, which all point generally in a southwest direction (S65∘ W) with no apparent relation to the variable aspect of the coastline, nor to a point source event as would be expected for a tsunami generated by flank margin collapse.

These chevron formations are the lowland correlative to the wave-generated rocky coast boulder deposits. The chevron ridges often occur in nested groups of several ridges (e.g., North Eleuthera and Great Exuma; Hearty et al., 1998) and show multiple complex sets and subsets of fenestrae-filled beds, indicating the passage of a sustained interval of time late in the interglacial. Their definitive and complex characteristics preclude formation during a single tsunami event.

The character of fenestral beds in both the Eemian chevron ridges and runup deposits change with increasing elevation and distance from shore, as do the abundance and geometry of fenestral pores (Tormey and Donovan, 2015): (1) at low elevations and in proximal locations, the chevron ridges are dominated by multiple truncated, thick, tabular, fenestrae-rich beds (Fig. 24a, b); (2) at moderate elevations and further inland, fenestrae are concentrated in discrete packages within eolianites, often associated with scour (Fig. 23) and rip-up clasts; and (3) in the highest and most distal eolian ridges, only rare, thin, discontinuous fenestrae beds can be found (Fig. 24c of Tormey and Donovan, 2015). This spatial transition is improbable if torrential rain was falling across the area during a storm as asserted by Bain and Kindler (1994); rather this is exactly the pattern expected as waves attenuate with greater distance and elevation inland.

Presence of a few eolian structures (Engel et al., 2015) does not imply that the chevron ridges are parabolic dunes; it suggests the deposits were sub-aerially exposed and wind blew during periods of relative quiescence (as commonly observed on today's beaches after a storm). Unlike parabolic dunes, the chevron ridges are dominated by thick, low-angle (< 10∘) seaward-dipping, aggradational oolitic bedding (Hearty et al., 1998; Tormey, 1999; Fig. 24a–c). Foreset beds, diagnostic of migrating parabolic dunes, are rare or absent from many chevron ridges, supporting formation primarily by waves (Hearty et al., 2002). Furthering the distinction, fenestral porosity in low-angle bedding is prevalent throughout chevron ridges, occurring in repeated cycles of centimeter-thick beds that onlap the underlying strata, and often comprise meter-thick fenestrae-rich packages that can be followed in outcrop for tens of meters (Fig. 24a, b).

Alternative interpretations of the geologic data have been made (Bain and Kindler, 1994; Kindler and Strasser, 2000, 2002; Engel et al., 2015); specifically, (1) the boulders were thrown by a tsunami caused by flank margin collapse in North Eleuthera, (2) beach fenestrae in runup and chevron beds were caused by heavy rainfall, and (3) the chevron beach ridges are parabolic dunes. These views are challenged by Hearty et al. (2002) and again here for the following reasons. (1) extensive research in the Bahamas has revealed no geologic evidence of a point-source tsunami radiating from North Eleuthera. A slow speed margin failure is possible, without a tsunami, and indeed such a flank margin collapse could have been initiated by massive storm waves impacting an over-steepened margin. (2) If heavy rainfall was a significant process in the formation of fenestrae in dunes, they should commonly occur in all dunes of all ages, which is not the case. (3) Carbonate dunes, particularly oolitic ones, generally do not migrate unless exposed to extremely arid climates, which contradicts point 2, and chevrons lack the most diagnostic feature of migration – foreset bedding.

It is too random and chronologically coincidental to argue that the trilogy of evidence – boulders, runup deposits, and chevron ridges – was caused by unconnected processes. If large, long-period waves lifted 1000 t boulders onto and over the coastal ridge, as is generally agreed, the same waves must have also impacted large areas of the eastern Bahamas, for which there is abundant documentation. A radiating pattern of landforms outward from a North Eleuthera point source, as from a tsunami generated from a local bank margin collapse, is not observed in the area or broader region. Absence of evidence for tsunamis on the United States East Coast refutes the possibility of a large remote tsunami source.

Our interpretation of these features is the most parsimonious, and we have argued that it is most consistent with the data. A common, synchronous, and non-random set of super-storm-related processes best explains boulder transport by waves, emplacement of runup deposits on older built-up ridges, and the formation of complex chevron deposits over time across lower areas of the Bahamas. Indeed, given the geologic evidence of high seas and storminess from Bermuda and the Bahamas, Hearty and Neumann (2001) suggested “steeper pressure, temperature, and moisture gradients adjacent to warm tropical waters could presumably spawn larger and more frequent cyclonic storms in the North Atlantic than those seen today”.

We now seek evidence about end-Eemian climate change to help clarify how North Atlantic storms could have dispersed such strong long-period, well-organized waves to the southwest.

We present here evidence for rapid end-Eemian cooling in the North Atlantic at a time with the tropics warmer than today. The cooling marked initial descent from interglacial conditions toward global ice age conditions, occurring at ∼ 118 ky b2k in ocean cores with uncertainty ∼ 2 ky. It is identified by Chapman and Shackleton (1999) as cold event C26 in Greenland ice cores.

This section discusses ocean core data, but we first note the relation with ice core data and dating uncertainties. Ice cores have become of great value for climate studies, partly because the relative timing of events in ice cores at different locations can be determined very accurately via marker events such as volcanic eruptions and CH4 fluctuations, even though the absolute dating error in ice cores is comparable to the dating uncertainty in ocean cores.

C26 is the cold phase of Dansgaard–Oeschger (D-O) climate oscillation D-O 26 in the NGRIP (North Greenland Ice Core Project) ice core (NGRIP, 2004). C26 begins with a sharp cooling at 119.14 ky b2k on the GICC05modelext timescale (Rasmussen et al., 2014). The GICC05 timescale is based on annual layer counting in Greenland ice cores for the last 60 ky and an ice-flow-model extension for earlier times. An alternative timescale is provided by Antarctic ice core chronology AICC2012 (Bazin et al., 2013; Veres et al., 2013), on which Greenland ice core records are synchronized via global markers, mainly oscillations of atmospheric CH4 amount, which is globally well mixed. C26 on Greenland is at 116.72 ky b2k on the AICC2012 timescale. Figure S19 shows the difference between GICC05 and AICC2012 timescales versus time.

This age uncertainty for C26 is consistent with the ice core 2σ error estimate of 3.2 ky at Eemian time (Bazin et al., 2013). Despite this absolute age uncertainty, we can use Greenland data synchronized to the AICC2012 timescale to determine the relative timing of Greenland and Antarctic climate changes (Sect. 4.2.1) to an accuracy of a few decades (Bazin et al., 2013).

Sediment cores from multiple locations provide information not only on ocean temperature and circulation (Fig. 25) but also on ice sheet via information inferred from ice-rafted debris. Comparison of data from different sites is affected by inaccuracy in absolute dating and use of different age models. Dating of sediments is usually based on tuning to the timescale of Earth orbital variations (Martinson et al., 1987) or “wiggle matching” to another record (Sirocko et al., 2005), which limits accuracy to several thousand years. Temporal resolution is limited by bioturbation of sediments; thus resolution varies with core location and climate (Keigwin and Jones, 1994). For example, high deposition rates during ice ages at the Bermuda Rise yield a resolution of a few decades, but low sedimentation rates during the Eemian yield a resolution of a few centuries (Lehman et al., 2002). Lateral transport of sedimentary material prior to deposition complicates data interpretation and can introduce uncertainty, as argued specifically regarding data from the Bermuda Rise (Ohkouchi et al., 2002; Engelbrecht and Sachs, 2005).

Adkins et al. (1997) analyzed a sediment core (MD95-2036; 34∘ N, 58∘ W) from the Bermuda Rise using an age model based on Martinson et al. (1987) orbital tuning with the MIS stage 5–6 transition set at 131 ky b2k and the stage 5d–5e transition at 114 ky b2k. They found that oxygen isotope δ18O of planktonic (near-surface dwelling) foraminifera and benthic (deep-ocean) foraminifera both attain full interglacial values at ∼ 128 ky b2k and remain nearly constant for ∼ 10 ky (their Fig. 2). Adkins et al. (1997) infer that “late within isotope stage 5e (∼ 118 ky b2k), there is a rapid shift in oceanic conditions in the western North Atlantic…”. They find in the sediments at that point an abrupt increase in clays indicative of enhanced land-based glacier melt and an increase in high nutrient “southern source waters”. The latter change implies a shutdown or diminution of NADW formation that allows Antarctic Bottom Water (AABW) to push into the deep North Atlantic Ocean (Duplessy et al., 1988; Govin et al., 2009). Adkins et al. (1997) continue: “The rapid deep and surface hydrographic changes found in this core mark the end of the peak interglacial and the beginning of climate deterioration towards the semi-glacial stage 5d. Before and immediately after this event, signaling the impeding end of stage 5e, deep-water chemistry is similar to modern NADW.” This last sentence refers to a temporary rebound to near-interglacial conditions. In Sect. 4.2.4 we use accurately synchronized Greenland and Antarctic ice cores, which also reveal this temporary end-Eemian climate rebound, to interpret the glacial inception and its relation to ice melt and late-Eemian sea level rise.

Ice-rafted debris (IRD) found in ocean cores provides a useful climate diagnostic tool (Heinrich, 1988; Hemming, 2004). Massive ice rafting (“Heinrich”) events are often associated with decreased NADW production and shutdown or slowdown of the Atlantic meridional overturning circulation (AMOC) (Broecker, 2002; Barreiro et al., 2008; Srokosz et al., 2012). However, ice rafting occurs on a continuum of scales, and significant IRD is found in the cold phase of all the 24 D-O climate oscillations first identified in Greenland ice cores (Dansgaard et al., 1993). D-O events exhibit rapid warming on Greenland of at least several degrees within a few decades or less, followed by cooling over a longer period. Chapman and Shackleton (1999) found IRD events in the NEAP18K core for all D-O events (C19–C24) within the core interval that they studied, and they also labeled two additional events (C25 and C26). C26 did not produce identifiable IRD at the NEAP18K site, but it was added to the series because of its strong surface cooling.

Lehman et al. (2002) quantify the C26 cooling event using the same Bermuda Rise core (MD95-2036) and age model as Adkins et al. (1997). Based on the alkenone paleo-temperature technique (Sachs and Lehman, 1999), Lehman et al. (2002) find a sharp SST decrease of ∼ 3 ∘C (their Fig. 1) at ∼ 118 ky BP, coinciding with the end-Eemian shoulder of the benthic δ18O plateau that defines stage 5e in the deep ocean. The SST partially recovered after several centuries, but C26 marked the start of a long slide into the depths of stage 5d cold, as ice sheets grew and sea level fell ∼ 50 m in 10 ky (Lambeck and Chappell, 2001; Rohling et al., 2009). Lehman et al. (2002) wiggle-match the MD95-2036 and NEAP18K cores, finding a simple adjustment to the age model of Chapman and Shackleton (1999) that maximizes correlation of the benthic δ18O records with the Adkins et al. (1997) δ18O record. Specifically, they adjust the NEAP timescale by +4 ky before the MIS 5b δ18O minimum and by +2 ky after it, which places C26 cooling at 118 ky b2k in both records. They give preference to the Adkins et al. (1997) age scale because it employs a 230Th-based timescale between 100 and 130 ky b2k.

We do not assert that the end-Eemian C-26 cooling was necessarily at 118 ky b2k, but we suggest that the strong rapid cooling observed in several sediment cores in this region of the subtropical and midlatitude North Atlantic Drift at about this time were all probably the same event. Such a large cooling lasting for centuries would not likely be confined to a small region. The dating models in several other studies place the date of the end-Eemian shoulder of the deep-ocean δ18O and an accompanying surface cooling event in the range 116–118 ky b2k.

Kandiano et al. (2004) and Bauch and Kandiano (2007) analyze core M23414 (53∘ N, 17∘ W), west of Ireland, finding a major SST end-Eemian cooling that they identify as C26 and place at 117 ky b2k. The 1 ky change in the timing of this event compared with Lehman et al. (2002) is due to a minor change in the age model; specifically, Bauch and Kandiano say that “the original age model of MD95-2036 (Lehman et al., 2002) has been adjusted to our core M23414 by alignment of the 4 per mil level in the benthic δ18O records (at 130 ka in M23414) and the prominent C24 event in both cores”. Bauch and Erlenkeuser (2008) and H. Bauch et al. (2012) examine ocean cores along the North Atlantic Current including its continuation into the Nordic Seas. They find that, in the Greenland–Iceland–Norwegian (GIN) seas, unlike middle latitudes, the Eemian was warmest near the end of the interglacial period. The age model employed by Bauch and Erlenkeuser (2008) has the Eemian about 2 ky younger than the Adkins et al. (1997) age model, Bauch and Erlenkeuser (2008) having the benthic δ18O plateau at ∼ 116–124 ky BP (their Fig. 6). Rapid cooling they illustrate there at ∼ 116.6 ky BP for core M23071 on the Vøring Plateau (67∘ N, 3∘ E) likely corresponds to the C26 end-Eemian cooling event.

Identification of end-Eemian cooling in ocean cores is hampered by the fact that Eemian North Atlantic climate was more variable than in the Holocene (Fronval and Jansen, 1996). There were at least three cooling events within the Eemian, each with minor increases in IRD, which are labeled C27, C27a and C27b by Oppo et al. (2006); see their Fig. 2 for core site ODP-980 in the eastern North Atlantic (55∘ N, 15∘ W) near Ireland. High- (sub-centennial) resolution cores in the Eirik drift region (MD03-2664, 57∘ N, 49∘ W) near the southern tip of Greenland reveal an event with rapid cooling accompanied by reduction in NADW production (Irvali et al., 2012; Galaasen et al., 2014), which they place at ∼ 117 ky b2k. However, their age scale has the benthic δ18O shoulder at ∼ 115 ky b2k (Fig. S1 of Galaasen et al., 2014), so that event may have been C27b, with C26 being stronger cooling that occurred thereafter.

Glacial–interglacial climate cycles are related to insolation change, as shown persuasively by Hays et al. (1976). Each “termination” (Broecker, 1984) of glacial conditions in the past several hundred thousand years coincided with a large positive warm-season insolation anomaly at the latitude of North American and Eurasian ice sheets (Raymo, 1997; Paillard, 2001). The explanation is that positive summer insolation anomalies (negative in winter) favor increased summer melting and reduced winter snowfall, thus shrinking ice sheets.

Termination timing is predicted better by high Northern Hemisphere late spring (April–May–June) insolation than by summer anomalies. For example, Raymo (1997) places midpoints of Termination I and II (preceding the Holocene and Eemian) at 13.5 and 128–131 ky b2k. Late spring insolation maxima are at 13.2 and 129.5 ky b2k (Fig. 26a). The AICC2012 ice core chronology (Bazin et al., 2013) places Termination II at 128.5 ky b2k, with 2σ uncertainty of 3.2 ky. Late spring irradiance maximizes warm-season ice melt by producing the earliest feasible warm-season ice sheet darkening via snow melt and snow recrystallization (Hansen et al., 2007b).

Summer insolation anomalies are also shown in Fig. 26, because interglacial periods can be expected to continue as long as summer insolation is large enough to prevent ice sheet genesis. Summer insolation anomalies at 60∘ N became negative at ∼ 118 ky b2k (Fig. 26a). Dating of insolation anomalies has high absolute accuracy, unlike ocean and ice core dating, as orbital anomalies are based on well-known planetary orbital mechanics (Berger, 1978).

Late Eemian sea level rise might appear to be a paradox, because glacial–interglacial sea level change is mainly a result of the growth and decay of Northern Hemisphere ice sheets. Northern warm-season insolation anomalies were declining rapidly in the latter part of the Eemian (Fig. 26a), so Northern Hemisphere ice should have been just beginning to grow. We suggest that the explanation for a late-Eemian sea level maximum is a late-Eemian collapse of Antarctic ice facilitated by the positive warm-season insolation anomaly on Antarctica and the Southern Ocean during the late Eemian (Fig. 26b) and possibly aided by an AMOC shutdown, which would increase warming of the Southern Ocean.

Persuasive evidence for this interpretation is provided by detailed paleoclimate data discussed in the next section, and is supported by modeling of relevant climate mechanisms. We will show that these mechanisms in turn help to explain ongoing climate change today, with implications for continuing climate change this century.

There is ample evidence that reduced atmospheric CO2 in glacial times, at least in substantial part, results from increased stratification of the Southern Ocean that reduces ventilation of the deep ocean (Toggweiler, 1999; Anderson et al., 2009; Skinner et al., 2010; Tschumi et al., 2011; Burke and Robinson, 2012; Schmitt et al., 2012; Marcott et al., 2014). Today the average “age” of deep water, i.e., the time since it left the ocean surface, is ∼ 1000 years (DeVries and Primeau, 2011), but it was more than twice that old during the Last Glacial Maximum (Skinner et al., 2010). The Southern Ocean dominates exchange between the deep ocean and atmosphere because ∼ 80 % of deep water resurfaces in the Southern Ocean (Lumpkin and Speer, 2007), as westerly circumpolar winds and surface flow draw up deep water (Talley, 2013).

Mechanisms causing more rapid deep-ocean ventilation during interglacials include warmer Antarctic climate that increases heat flux into the ocean and buoyancy mixing that supports upwelling (Watson and Garabato, 2006), poleward shift of the westerlies (Toggweiler et al., 2006), and reduced sea ice (Keeling and Stephens, 2001). Fischer et al. (2010) question whether the latitudinal shift of westerlies is an important contributor; however, the basic point is the empirical fact that a warmer interglacial Southern Ocean produces faster ventilation of the deep ocean via a combination of mechanisms.

Although a complete quantitative understanding is lacking for mechanisms to produce the large glacial–interglacial swings of atmospheric CO2, we can safely assume that deep-ocean ventilation acts to oppose the sequestration of carbon in the ocean by the “pumps” that move carbon from the surface to ocean depths, and changes in the ventilation rate have a significant effect on atmospheric CO2. Ridgwell and Arndt (2015) describe several conceptual pumps: (1) the organic carbon pump, in which the sinking material also controls nutrient cycling by the ocean; (2) the carbonate pump, with biological precipitation of mainly calcium carbonate, a fraction of which escapes dissolution to form a new geological carbon reservoir; (3) the simple solubility pump, as CO2 is more soluble in the cold polar waters where deep water forms; and (4) a microbial carbon pump that seems capable of altering deep-ocean dissolved organic carbon.

No doubt the terrestrial biosphere also contributes to glacial–interglacial atmospheric CO2 change (Archer et al., 2000; Sigman and Boyle, 2000; Kohler et al., 2005; Menviel et al., 2012; Fischer et al., 2015). Also, the efficacy of the ocean pumps depends on terrestrial conditions, e.g., dust-borne iron fertilization of the biological pump (Martin and Fitzwater, 1988) contributes to millennial and full glacial CO2 drawdown (Martinez-Garcia et al., 2014). Moreover, the Southern Ocean is not the only conduit to the deep ocean; for example, AMOC changes are associated with at least two rapid CO2 increases of about 10 ppm, as revealed by a high-resolution West Antarctic ice core (Marcott et al., 2014). Nevertheless, it is reasonable to hypothesize that sequestration of CO2 in the glacial ocean is the largest cause of glacial–interglacial CO2 change, and it is known that ocean ventilation occurs mainly via the Southern Ocean.

Southern Ocean ventilation, as the dominant cause of atmospheric CO2 change, helps explain temperature-CO2 leads and lags. Temperature and CO2 rises are almost congruent at ice age terminations (Masson-Delmotte et al., 2010; Pedro et al., 2012; Parrenin et al., 2013). Southern Ocean temperature is expected to lead, spurring deep-ocean ventilation and atmospheric CO2 increase, with global temperature following. Termination I is dated best and Shakun et al. (2012) have reconstructed global temperature for that period, finding evidence for this expected order of events.

Correlation of δ18O and CO2 over the past 140 ky (Fig. 27c) is 84.4 %, with CO2 lagging by 760 years. For the period 100–20 ky b2k, which excludes the two terminations, the correlation is 77.5 %, with CO2 lagging by 1040 years. Briefer lag for the longer period and longer lag during glacial inception are consistent with the rapid deep-ocean ventilation that occurs at terminations.

CO2 is the principal determinant of Earth's climate state, the radiative “control knob” that sets global mean temperature (Lacis et al., 2010, 2013). Degree of control is shown by comparison of CO2 amount with Antarctic temperature for the past 800 000 years (Fig. 28a). Control should be even tighter for global temperature than for Antarctic temperature, because of regional anomalies such as Antarctic temperature overshoot at terminations (Masson-Delmotte et al., 2006, 2010), but global data are not available.

The tight fit of CO2 and Antarctic temperature (Fig. 28a) implies an equilibrium Antarctic sensitivity of 20 ∘C for 2 × CO2 (4 W m-2) forcing (200 → 300 ppm forcing is ∼ 2.3 W m-2; Table 1 of Hansen et al., 2000) and thus 10 ∘C global climate sensitivity (Antarctic temperature change is around twice that of global change) with CO2 taken as the ultimate control knob, i.e., if snow/ice area and other GHGs are taken to be slaves to CO2-driven climate change. This implies a conventional climate sensitivity of 4 ∘C for 2 × CO2, as GHG and albedo forcings are similar for glacial-to-interglacial climate change and non-CO2 GHGs account for ∼ 20 % of the GHG forcing. The inferred sensitivity is reduced to 2.5–3 ∘C for 2 × CO2 if, as some studies suggest, global mean glacial–interglacial temperature change is only about one-third of the Antarctic temperature change (Palaeosens, 2012; Hansen et al., 2013b).

The CO2 dial must be turned to ∼ 260 ppm to achieve a Holocene-level interglacial. CO2 ∼ 250 ppm was sufficient for quasi-interglacials in the period 800–450 ky b2k, with sea level 10–25 m lower than in the Holocene (Fig. S18 of Hansen et al., 2008). Interglacials with CO2 ∼ 280 ppm, i.e., the Eemian and Holsteinian (∼ 400 ky b2k), were warmer than the Holocene and had sea level at least several meters higher than today.

CO2 and albedo change are closely congruent over the last 800 000 years (Fig. S18 of Hansen et al., 2008). GHG and albedo forcings, which are both amplifying feedbacks that boost each other, are each of amplitude ∼ 3 W m-2. So why do we say that CO2 is the control knob?

First, CO2, in addition to being a slow climate feedback, changes independently of climate. Natural CO2 change includes an increase to ∼ 1000 ppm about 50 million years ago (Zachos et al., 2001) as a result of plate tectonics, specifically volcanic emissions associated with movement of the Indian plate across the Tethys Ocean and collision with Asia (Kent and Muttoni, 2008). Humankind, mainly by burning fossil fuels, also moves the CO2 control knob.

Second, CO2 is more recalcitrant than snow and ice, i.e., its response time is longer. CO2 inserted into the climate system, by humans or plate tectonics, remains in the climate system of the order of 100 000 years before full removal by weathering (Archer, 2005). Even CO2 exchange between the atmosphere (where it affects climate) and ocean has a lag of the order of a millennium (Fig. 27). In contrast, correlations of paleo-temperatures and sea level show that lag of sea level change behind temperature is of the order of a century, not a millennium (Grant et al., 2012).

We suggest that limitations on the speed of ice volume (and thus sea level) changes in the paleo-record are more a consequence of the pace of orbital changes and CO2 changes, as opposed to being a result of lethargic ice physics. “Fast” changes of CO2 have been identified, e.g., an increase of ∼ 10 ppm in about a century at ∼ 39.6 ky b2k (Ahn et al., 2012) and three increases of 10–15 ppm each within 1–2 centuries during the deglaciation following the last ice age (Marcott et al., 2014), but the magnitude of these CO2 increases is not sufficient to provide a good empirical test of ice sheet sensitivity to the CO2 forcing.

Dominance of SMOC, the Southern Ocean meridional overturning circulation, in affecting the CO2 control knob and thus glacial–interglacial change is contrary to the idea that the AMOC is a prime driver that flips global climate between quasi-stable glacial and interglacial states, yet AMOC retains a significant role. AMOC can affect CO2 via the volume and residence time of NADW, but its largest effect is probably via its impact on the Southern Ocean. When AMOC is not shut down it cools the Southern Hemisphere, transferring heat from the Southern to the Northern Hemisphere at a rate of ∼ 1 PW, which is ∼ 4 W m-2 averaged over a hemisphere (Crowley, 1992). However, the Southern Ocean slowly warms when AMOC shuts (or slows) down; the response time is of the order of 1000 years because of the Southern Ocean's large thermal inertia (Stocker and Johnson, 2003). These mechanisms largely account for the nature of the “bipolar seesaw” (Broecker, 1998; Stocker, 1998; Stenni et al., 2011; Landais et al., 2015), including the lag between AMOC slowdown and Antarctic warming.

The magnitude and rapidity of Greenland climate change during D-O events would deter prediction of human-made climate effects if D-O events remained a mystery. Instead, however, enough is now understood about D-O events that they provide insight related to the vulnerability of ice shelves and ice sheets, including the role of subsurface ocean warming.

Broecker (2000) inferred from the rapidity of D-O warmings that a reduction of sea ice cover was probably involved. Li et al. (2005, 2010) modeling showed that removal of Nordic Seas ice cover is needed to yield the magnitude of observed Greenland warming. The spatial gradient of D-O warming, with smaller warming in northwest Greenland, agrees with that picture (Guillevic et al., 2013; Buizert et al., 2014). Such sea ice change is consistent with changes in deuterium excess in Greenland ice cores at D-O transitions, which indicate shifts of Greenland moisture source regions (Masson-Delmotte et al., 2005; Jouzel et al., 2007).

Fluckiger et al. (2006), Álvarez-Solas et al. (2010, 2011, 2013) and Marcott et al. (2011) noted modern and paleo-data that point to ocean–ice shelf interaction as key to the ice discharge of accompanying Heinrich events, and they used a range of models to support this interpretation and overturn earlier suggestions of a central role for ice sheets via binge–purge oscillations (MacAyeal, 1993) or outburst flooding from subglacial reservoirs (Alley et al., 2006). Shaffer et al. (2004) and Petersen et al. (2013) conclude that subsurface ocean warming in the North Atlantic takes place during the stadial (cold) phase of all D-O events, and eventually this subsurface warming leads to ice shelf collapse or retreat, ice rafting, sea level rise, and sea ice changes. Rasmussen et al. (2003) examined ocean cores from the southeast Labrador Sea and found that for all 11 Heinrich events “the icy surface water was overlying a relatively warm, poorly ventilated and nutrient rich intermediate water mass to a water depth of at least 1251 m”. Collapse of a Greenland ice shelf fronting the Jakoshavn ice stream during the Younger Dryas cold event has been documented (Rinterknecht et al., 2014), apparently due to subsurface warming beneath the ice shelf leading to rapid discharge of icebergs.

Some D-O details are uncertain, e.g., the relation between changing sea ice cover and changing location of deep water formation (Rahmstorf, 1994) and whether an ice shelf between Greenland and Iceland contributed to the sea ice variability (Petersen et al., 2013). However, ocean–ice interactions emerge as key mechanisms, spurred by subsurface ocean warming, as ocean stratification slows but does not stop northward heat transport by AMOC.

We consider a specific D-O event for the sake of discussing mechanisms. D-O 22 cold phase, labeled C22 in ocean cores and coinciding with Heinrich H8 (Fig. 27), occurred as Northern Hemisphere insolation was rising (Fig. 27a). The North Atlantic surface was cooled by rapid ice discharge; sea level rose more than 40 m, a rate exceeding 1.6 m per century (Cutler et al., 2003). Ice discharge kept the North Atlantic highly stratified, slowing AMOC. Antarctic warming from a slowed AMOC increases almost linearly with the length of the D-O cold phase (Fig. 3 of EPICA Community Members, 2006; Fig. 6 of Capron et al., 2010) because of the Southern Ocean's large heat capacity (Stocker and Johnson, 2003). Antarctic warming, aided by the 2500-year D-O 22 event, spurred SMOC enough to raise atmospheric CO2 40 ppm (Fig. 27c).

As the Antarctic warmed, ocean heat transport to the North Atlantic would have increased, with most heat carried at depths below the surface layer. When the North Atlantic became warm enough at depth, stratification of cold fresh surface water eventually could not be maintained. The warming breakthrough may have included change in NADW formation location (Rahmstorf, 1994) or just large movement of the polar front. Surface warming east of Greenland removed most sea ice and Greenland warmed ∼  10 ∘C (Capron et al., 2010). As the warm phase of D-O 21 began, AMOC was pumping heat from the Antarctic into the Nordic Seas and Earth must have been slightly out of energy balance, cooling to space, so both Antarctica and Greenland slowly cooled. Once the North Atlantic had cooled enough, sea ice formed east of Greenland again, ice sheets and ice shelves grew, sea level fell, and the polar front moved southward.

Sea level rise associated with D-O events covers a wide range. Sea level increases as large as ∼ 40 m were associated with large insolation forcings at 107 and 86 ky b2k (Fig. 27). However, rapid sea level change occurred even when forcing was weak. Roche et al. (2004) conclude from analyses of δ18O that H4, at a time of little insolation forcing (∼ 40 ky b2k, Fig. 27), produced 1.9 ± 1.1 m sea level rise over 250  ± 150 years. Sea level rise as great as 10–15 m occurred in conjunction with some other D-O events during 65–30 ky b2k (Lambeck and Chappell, 2001; Yokoyama et al., 2001; Chappell, 2002).

Questions about possible D-O periodicity and external forcing were raised by a seeming 1470-year periodicity (Schulz, 2002). However, improved dating indicates that such periodicity is an artifact of ice core chronologies and not statistically significant (Ditlevsen et al., 2007), and inspection of Fig. 27b reveals a broad range of timescales. Instead, the data imply a climate system that responds sensitively to even weak forcings and stochastic variability, both of which can spur amplifying feedbacks with a range of characteristic response times.

Two conclusions are especially germane. First, subsurface ocean warming is an effective mechanism for destabilizing ice shelves and thus the ice sheets buttressed by the ice shelves. Second, large rapid sea level rise can occur as a result of melting ice shelves.

However, ice shelves probably were more extensive during glacial times. So are today's ice sheets much more stable? The need to understand ice sheet vulnerability focuses attention on end-Eemian events, when ice sheets were comparable in size to today's ice sheets.

Termination II, ushering in the Eemian, was spurred by a late spring 60∘ N insolation anomaly peaking at +45 W m-2 at 129.5 ky b2k (Fig. 27a), the largest anomaly in at least the past 425 ky (Fig. 3 of Hansen et al., 2007b). CO2 and albedo forcings were mutually reinforcing. CO2 began to rise before Antarctic δ18O, as deglaciation and warming began in the Northern Hemisphere. Most of the total CO2 rise was presumably from deep-ocean ventilation in the Southern Ocean, aided by meltwater that slowed the AMOC and thus helped to warm the Southern Ocean.

The northern late spring insolation anomaly fell rapidly, becoming negative at 123.8 ky b2k (Fig. 27a), by which time summer insolation also began to fall (Fig. 26). Northern Hemisphere ice sheets must have increased intermittently while Southern Hemisphere ice was still declining, consistent with minor, growing ice rafting events C27, C27a, C27b and C26 and a sea level minimum during 125–121 ky b2k (Sect. 4.1.1). High Eemian climate variability in the Antarctic (Pol et al., 2014) was likely a result of the see-saw relation with North Atlantic events. CO2 (Fig. 27c) remained at ∼ 270 ppm for almost 15 ky as the positive insolation anomaly on the Southern Ocean (Fig. 27a) kept the deep ocean ventilated. Sea level in the Red Sea analysis (Grant et al., 2012) shown in Fig. 27f seems to be in decline through the Eemian, but that must be a combination of dating and sea level error, as numerous sea level analyses cited in Sect. 4.1.1, our Supplement, and others (e.g., Chen et al., 1991; Stirling et al., 1998; Cutler et al., 2003) indicate high sea level throughout the Eemian and allow a possible late-Eemian maximum. Chen et al. (1991), using a U-series dating with 2σ uncertainty ±1.5 ky, found that the Eemian sea level high stand began between 132 and 129 ky b2k, lasted for 12 ky, and was followed by rapid sea level fall.

We assume that C26, the sharp cooling at 116.72 ky b2k in the NGRIP ice on the AICC2012 timescale, marks the end of fully interglacial Eemian conditions, described as 5e sensu stricto by Bauch and Erlenkeuser (2008). δ18O in Antarctica was approaching a relative minimum (-46.7 ‰ at EDML; see Fig. S20 for details) and CO2 was slowly declining at 263 ppm. In the next 300 years δ18O increased to -45.2 and CO2 increased by 13 ppm with lag ∼ 1500 years, which we interpret as see-saw warming of the Southern Ocean in response to the C26-induced AMOC slowdown and resulting increased SMOC ventilation of CO2.

Freshwater causing the C26 AMOC shutdown could not have been Greenland surface melt. Greenland was already 2000 years into a long cooling trend and the northern warm-season insolation anomaly was in the deepest minimum of the last 150 ky (Fig. 26a). Instead, C26 was one event in a series, preceded by C27b and followed by C25, each a result of subsurface North Atlantic warming that melted ice shelves, causing ice sheets to discharge ice. Chapman and Shackleton (1999) did not find IRD from C26 in the mid-Atlantic, but Carlson et al. (2008) found a sharp increase in sediments near the southern tip of Greenland that they identified with C26.

We suggest that the Southern Hemisphere was the source for brief late-Eemian sea level rise. The positive warm-season insolation anomaly on the Southern Ocean and AMOC slowdown due to C26 added to Southern Ocean heat, causing ice shelf melt, ice sheet discharge, and sea level rise. Rapid Antarctica ice loss would cool the Southern Ocean and increase sea ice cover, which may have left telltale evidence in ice cores. Indeed, Masson-Delmotte et al. (2011) suggest that abrupt changes of δ18O in the EDML and TALDICE ice cores (those most proximal to the coast) indicate a change in moisture origin, likely due to increased sea ice. Further analysis of Antarctic data for the late Eemian might help pinpoint the melting and help assess vulnerability of Antarctic ice sheets to ocean warming, but this likely will require higher-resolution models with more realistic sea ice distribution and seasonal change than our present model produces.

Terrestrial records in Northern Europe reveal rapid end-Eemian cooling. Sirocko et al. (2005) find cooling of 3 ∘C in summer and 5–10 ∘C in winter in southern Germany, with annual layering in a dry Eifel maar lake revealing a 468-year period of aridity, dust storms, bushfires, and a decline of thermophilous trees. Similar cooling is found at other German sites and La Grande Pile in France (Kuhl and Litt, 2003). Authors in both cases interpret the changes as due to a southward shift of the polar front in the North Atlantic corresponding to C26. Cooling of this magnitude in northern Europe and increased aridity are found by Brayshaw et al. (2009) and Jackson et al. (2015) in simulations with high-resolution climate models forced by AMOC shutdown.

While reiterating dating uncertainties, we note that the cool period with reduced NADW formation identified in recent high-resolution ocean core studies for Eirik Drift site MD03-2664 (Fig. 25) near Greenland (Irvali et al., 2012; Galaasen et al., 2014) at ∼ 117 ky b2k has length similar to the 468-year cold stormy period found in a German lake core (Sirocko et al., 2005).

The Eirik core data show a brief return to near-Eemian conditions and then a slow decline, similar to the oscillation in the NGRIP ice core at 116.72 ky b2k on the AICC2012 timescale.

The principal site of NADW formation may have moved from the GIN seas to just south of Greenland at end of the Eemian. Southward shift of NADW formation and the polar front is consistent with the sudden, large end-Eemian cooling in the North Atlantic and northern Europe, while cooling in southern European was delayed by a few millennia (Brauer et al., 2007). Thus end-Eemian midlatitude climate was characterized by an increased meridional temperature gradient, an important ingredient for strengthening storms.