This study uses a modified version of the coupled climate model GFDL CM2.1 (Delworth et al., 2006), using boundary conditions for the late Eocene. The ocean component is updated to the modular ocean model (MOM) version 5.1.0, while the other components of the model are the same as in CM2.1, namely Atmosphere Model 2, Land Model 2 and the Sea Ice Simulator. In the ocean and sea ice components, the horizontal grid is modified to have a resolution of 1∘ latitude × 1.5∘ longitude. We use a tripolar grid as shown in Fig. 7 of Murray (1996), with a regular latitude–longitude mesh south of 65∘ N, and a transition to a bipolar Arctic grid north of 65∘ N. The poles then lie over North America and Siberia to avoid convergence of meridians in the Arctic Ocean. This enables the model to simulate Arctic Ocean velocities without damping in the vicinity of the North Pole. Unlike in CM2.1, there is no refinement of the latitudinal grid spacing in the tropics. The ocean retains the original 50 vertical levels with the same grid spacing as in CM2.1. The atmospheric horizontal grid resolution is 3∘ × 3.75∘ , with 24 vertical levels. The atmosphere grid is identical to that used in CM2Mc (Galbraith et al., 2010). This choice of grid resolutions gives good load balancing between the atmosphere and ocean components, while enabling better resolution of coastlines and straits than most existing Eocene models (∼ 3∘ ocean resolution).

The topography (both land and ocean) uses the late Eocene (38 Ma) reconstruction of Baatsen et al. (2016). This topography is distinct from previous reconstructions (e.g. Markwick, 2007) in that it uses a palaeomagnetic reference frame to position the continents (Torsvik et al., 2012; van Hinsbergen et al., 2015), rather than a hotspot reference frame (Seton et al., 2012). Continental elevation is based on ETOPO modern-day topography, which is then relocated to its 38 Ma position by plate tectonic motion using GPlates. For the deep ocean, an age–depth relationship is applied (Müller et al., 2008) and adjusted to the palaeomagnetic reference frame. Manual adjustments are then applied to areas where elevation changes are well constrained by geological evidence. Specific regions of adjustment include Antarctica, the Himalayas, the Amazon, Turgai Strait and the Tethys Sea. Where palaeo-elevation data is missing or unknown, the topography defaults to modern-day elevation. One region of uncertainty is in the gateway between the Arctic and Atlantic Ocean, where other reconstructions have a more constricted throughflow (e.g. Fig. 41 of Markwick, 2007). In particular, deepening of the Greenland–Scotland Ridge is hypothesised as a trigger of North Atlantic deep water formation at the EOT, via its impact on freshwater transport from the Arctic into the Atlantic (Abelson and Erez, 2017; Stärz et al., 2017).

Manual adjustments were made to the topography to ensure that all straits are at least two grid cells wide, so that they have corresponding velocity grid points, and no ocean grid cell is isolated. The resulting ocean bathymetry is shown in Fig. 1a. In the atmosphere, the topography is interpolated onto the horizontal grid and then smoothed using a 3-point mean filter to ensure a smoother interaction with the wind field. This filtering was mainly needed on the Antarctic continent, due to convergence of meridians on the topography grid, which caused numerical noise in the wind field during initial testing. This filter will likely be removed outside of the polar regions in future versions of the model. The resulting topography is shown in Fig. 1b. Vegetation types are founded on a dataset based on Sewall et al. (2000), with modifications where data are available from Thorn and DeConto (2006); Utescher and Mosbrugger (2007); and Gomes Rodrigues et al. (2012). The Sewall et al. (2000) dataset was originally configured for the CESM plant functional types, and we have adapted these to the corresponding vegetation types in CM2.1. The river run-off is determined by a relocation map, where each land point is assigned a corresponding coastal location for returning run-off to the ocean. This relocation map was determined by a downslope relocation algorithm from the model topography. Aerosol forcing was taken from the Eocene (55 Ma) reconstruction of Herold et al. (2014), and adapted to our model's input types.

We use a simplified version of the tidal mixing scheme of Simmons et al. (2004). The CM2.1 code distribution provides a heterogeneous distribution of seafloor roughness amplitude based on high resolution maps of the modern seafloor. We set this to be uniform for simplicity, since estimating the late Eocene seafloor roughness is not straightforward. We thus calculate an average roughness amplitude of 210.512  (the h parameter from Eq. (1) of Simmons et al. 2004), and set a background diffusivity of 1.0 × 10-5 m2 s-1. In parallel we also simulate the model using the standard Bryan–Lewis diffusivity in CM2.1, commonly used in deep-time palaeoclimate studies (e.g. Lunt et al., 2012). We prefer a bottom roughness mixing scheme to Bryan–Lewis, since the former is more physically realistic. The Bryan–Lewis scheme uses the following values: Keqz=10-4[0.65+1.15πtan⁡-1(4.5×10-3z-2500)]Kpolez=10-4[0.75+0.95πtan⁡-1(4.5×10-3z-2500)], where Keq and Kpole are the equatorial and polar diffusivities respectively, in m2 s-1. The depth z is in metres, which has a transition level at 2500 m as shown above. The low latitudes use Keq and the high latitudes use Kpole, with a transition at 35∘ latitude.

The initial conditions of the ocean model are configured to a modified version of the DeepMIP experimental design (Lunt et al., 2017). The temperature and salinity are set to the following profiles: T[∘C]=(5000-z)/5000*20cos(φ)+10;ifz≤5000m.=10;ifz>5000m.S[psu]=35.0, where φ is latitude, z is depth of the ocean (metres below surface – positive downwards).

The model is run at three different levels of CO2: 400, 800 and 1600 ppm, where 800 ppm is deemed as the control simulation. The atmosphere and land surface components are initialised from a previous test simulation.

We used the following procedure to spin up the model simulations:

The model is run in coupled mode for 50 years.

In order to assess the robustness of this acceleration method, we also ran the control simulation from the same initial conditions for 1000 years in coupled mode only. This coupled-only simulation yielded an ocean state that was very similar in key metrics (i.e. temperature, salinity, age tracer) to the iteratively coupled ocean run at year 2000 (see Fig. 2). In other words, the iterative coupling procedure achieved a climate state that was similar to that of an ordinary coupled run, with roughly twice the number of model years needed to reach the same state. This approximately cancels out the computational speed-up achieved by the decoupling procedure. We therefore abandoned the iterative coupling procedure after six cycles and completed the spin-up in coupled mode.

The spin-up evolution of ocean temperature across the three levels of CO2 is shown in Fig. 2. The simulations all have surface climates in quasi-equilibrium, although the deep ocean is gradually cooling in all cases. All simulations also have a temperature trend of less than 0.1 ∘C per century at 4000 m, although the warmer climates are trending more slowly than the colder ones. This is because the initial conditions are designed to be warmer rather than colder than equilibrium, so that convection can readily occur due to surface cooling. The step changes in SST over the first 3300 years are due to the coupling/decoupling procedure described above. We also examined the evolution of the meridional overturning circulation during spin-up (Fig. 2e). This shows that the magnitude of Northern Hemisphere overturning remains steady over the last ∼ 2000 years, while the Southern Hemisphere overturning is gradually reducing in the 1600 and 400 ppm cases, and is steady in the 800 ppm case. Southern Ocean sinking is discussed further in Sects. 3 and 4. The trend in Arctic salinity is also shown in Fig. 2f. This indicates a rapid adjustment towards a fresh Arctic surface in all three cases, followed by minimal change over the last 3000 years.

One way to estimate the final equilibrium in a transient climate simulation is to derive a “Gregory plot” (Gregory et al., 2004), which compares the top-of-atmosphere (TOA) net radiative forcing with the temperature anomaly. This is shown for each CO2 level in Fig. 3. The data points are derived from 10-year averages during the iterative coupling phase of the spin-up, and then from 100-year averages during the fully coupled phase. There is a trend towards zero net radiative forcing during much of the spin-up, but the curves begin to flatten before they reach zero. This makes extrapolation towards an equilibrium value difficult, since there is no definitive slope from which to project. In addition, we note the CM2.1 model has been documented to have a radiative imbalance under pre-industrial conditions on the order of 0.5 W m-2 (Delworth et al., 2006). This is due to a number of factors, including a systematic loss of energy (∼ 0.3 W m-2) in the atmosphere model, and a heat sink that occurs when water is returned to the ocean from precipitation or run-off (∼ 0.14 W m-2). Therefore, we do not expect the model to reach a perfect radiative forcing balance at equilibrium.

In this section, we describe the control climate (800 ppm), key features of the circulation and how they compare with proxy evidence. Fig. 4a and b show the sea surface temperature (SST) and surface air temperature (SAT) respectively. Equatorial SSTs reach as high as 38 ∘C in the western Pacific warm pool, with tropical temperatures around 30–35 ∘C elsewhere. The warm pool extends west into the Indian Ocean due to an open Indonesian Seaway. The meridional temperature gradient is a little lower than in the modern climate, with the high latitude SSTs of the North Pacific and Southern Ocean being around 15–20 ∘C, even along the coast of Antarctica.

The sea surface salinity (Fig. 4c) indicates very fresh conditions in the Arctic Ocean, around 20 psu. The fresh Arctic surface waters primarily flow out into the North Atlantic, where the surface salinities around Greenland and above ∼ 45∘ N are generally below 30 psu. Under these conditions deep sinking does not occur in the North Atlantic. Conversely, the high latitude North Pacific surface water is mostly around 35 psu, apart from a limited region in the north-east where there is strong run-off entering the ocean from the North American mountain ranges. Under modern conditions, orography preferentially freshens the surface water in the North Pacific over the North Atlantic (Sinha et al., 2012; Wills and Schneider, 2015; Maffre et al., 2018). This is primarily due to its influence on precipitation and drainage pathways, especially in the North American mountain ranges. Notably, the low surface salinities prevent deep water formation in the present-day North Pacific (Warren, 1983). However, our control simulation shows a higher river run-off (0.28 Sv) going into the Arctic than the North Atlantic (0.14 Sv) and North Pacific (0.15 Sv) (here the cut-off latitude is chosen to be 45∘ N). The same comparison with a simulation of CM2.1 in its modern configuration produces a river run-off of 0.10 Sv into the Arctic, only slightly higher than the North Atlantic value of 0.07 Sv and the North Pacific value of 0.08 Sv. This high Arctic run-off, combined with enhanced precipitation minus evaporation (P-E) forcing of the warm climate creates the very fresh Arctic surface waters. Transport of this fresh Arctic water mass into the North Atlantic causes a state reminiscent of the present-day North Pacific with low surface salinity, a strong halocline and no deep water formation.

Zonal wind stress (Fig. 4e) in the Southern Ocean has a more zonal structure than in the present day, with a weaker standing wave meander and a lower magnitude overall. The wind stress peak sits around 50∘ S, similar to present day, and the wind stress curl drives subtropical gyre circulations in the South Pacific and Indian oceans analogous to the present day, as shown in the barotropic streamfunction of Fig. 4f. The Southern Ocean gateways – Drake Passage and the Tasman Seaway – are open, but they are narrow and shallow and thus allow only a weak circumpolar flow around Antarctica of 15 Sv.

The mean state of the model has an east–west temperature contrast across the equatorial Pacific in excess of 6 ∘C, slightly larger than in the present day. Empirical orthogonal function (EOF) analysis of global SST indicates the strongest variability in the central equatorial Pacific (Fig. 5a), while precipitation varies most strongly in a dipole spanning the western equatorial Pacific and Indian oceans (Fig. 5b). Using the first EOF of SST, we define an “El Niño Index” region between 170∘ E and 140∘ W, and 5∘ S to 5∘ N, as being representative of highest Pacific equatorial SST variability. This index is analogous to the present-day Niño-3.4 index, designed to bracket the region of highest El Niño–Southern Oscillation (ENSO) variability. This index indicates the occurrence of multi-year El Niño and La Niña events (Fig. 5c), defined by a threshold of +0.5 ∘C for an El Niño and -0.5 ∘C for a La Niña, as used in present day. The magnitude of temperature deviations is generally weaker than present-day Niño-3.4 variability. The east–west SST gradient in the equatorial Pacific is closely anti-correlated with the El Niño index, with a reduction in east–west gradient during El Niño and vice versa for La Niña (Fig. 5d). However, the magnitude of these variations is also weaker than present day. The areal extent of the western Pacific warm pool in our model is larger than the present day. We suggest that this partly due to fewer land barriers in the western Pacific and a wider basin. We argue that this creates a larger thermal inertia in the western Pacific, implying a reduction in variance in the east to west thermal gradient. This result agrees broadly with von der Heydt and Dijkstra (2006), who also found a stronger western Pacific warm pool in an Oligocene model simulation.

The “Weddell Sea” equivalent in the EOT model, where deep water forms in the model, has annual mean sea surface temperature below 10 ∘C and develops a small area of seasonal sea ice close to the coast. The Arctic Ocean temperatures are typically around 4–6 ∘C, and it is mostly ice free all year round, apart from some embayments around the Siberian coast. The average maximum monthly sea ice thickness for each hemisphere is plotted in Fig. 6, showing March thickness for the Northern Hemisphere and September thickness for the Southern Hemisphere. We note that there is evidence of ice-rafted debris in the Arctic from the middle Eocene onwards (Stickley et al., 2009), indicating the likely presence of seasonal sea ice at this time. Our 400 ppm simulation has Arctic-wide sea ice in summer (Fig. 6b), in agreement with this proxy evidence. However, the ice-rafted debris emanates from the central Arctic (Stickley et al., 2009), where there is still some sea ice in the 800 ppm case, making it difficult to distinguish which case is more realistic.

The meridional overturning circulation (Fig. 7) shows a structure of sinking in the North Pacific and Southern Ocean. There is no deep overturning cell in the North Atlantic, due to the surface waters being far too fresh to allow for local deep water formation. The structure of the Pacific overturning is analogous to modern-day North Atlantic deep water (NADW) formation: the northern component water is warmer and saltier than the southern component water and is heavier at the surface. This difference in deep water properties is caused by colder winter temperatures in the Southern Ocean sinking regions than the North Pacific, since deep water forms exclusively in winter. The colder winter temperatures are due to the asymmetry of land masses between the poles. One factor is that the Southern Ocean extends further poleward than the North Pacific (in both Pacific and Atlantic sectors). Furthermore, the Antarctic continental interior becomes colder in winter than the Arctic Ocean, and therefore the seasonal cycle in high latitude (> 60∘ ) surface air temperature is stronger in the Southern Hemisphere. The Northern Hemisphere also has very cold winters in continental interiors, but its polar winters are moderated by the presence of an ocean and its seasonal cycle is milder.

As in modern climate, thermobaric effects control the layering of the abyssal ocean (Nycander et al., 2015). The Southern Hemisphere sourced deep waters are colder and therefore more compressible than the warmer and slightly more saline ones sourced from the Northern Hemisphere. Thus bottom waters are predominantly formed in the Southern Ocean. Sinking occurs in both the Pacific and Atlantic sectors of the Southern Ocean, as indicated by the age tracer at 2000 m and winter mixed layer depths (Fig. 8). The shallow gateways of Drake Passage and Tasman Seaway inhibit the exchange of deep water between the basins, preventing Pacific- and Atlantic-sourced bottom waters from directly reaching the opposite ocean basin. The poleward heat transport is similar to that of the modern climate (not shown), with the partition between ocean and atmosphere heat transport being dominated by the ocean in the tropics only, and by the atmosphere in the mid to high latitudes. As in the modern ocean, poleward heat transport is skewed towards the Northern Hemisphere, with a peak value of 2.2 PW (northward) and a minimum value of -0.9 PW in the Southern Hemisphere. We suggest that this asymmetry is due to the analogous asymmetry of deep water formation; i.e. the northern overturning cell traverses a larger temperature difference and hence enhances northward ocean heat transport relative to weaker temperature gradient across the southern overturning cell. In the tropics, we do not find any warm salt-driven deep water formation, in line with previous studies (Table 1).

We also compare the results of our control run with the widely used Bryan–Lewis mixing scheme, as used in the default case of CM2.1 (Delworth et al., 2006). The Bryan–Lewis simulation was spun up using the same procedure as the control run, as described in Sect. 2.3. Prior to this study, we expected that changes to the bathymetry would alter the deep ocean mixing distribution due to intensified mixing over ridges and seamounts. Furthermore, we anticipated that changes to surface radiative forcing would alter the thermocline structure. For these reasons we expected that a bottom-enhanced vertical mixing scheme would be more adaptable to late Eocene climate conditions and perform better than the fixed vertical mixing structure of the Bryan–Lewis scheme. However, we find that the meridional overturning circulation and stratification are largely insensitive to this change in vertical mixing. The magnitude of the overturning, distribution of sinking regions and age tracer were all very similar between the mixing schemes (not shown). This was a surprising result, given theoretical predictions of change due to bottom-enhanced mixing (de Boer and Hogg, 2014). We do find that the abyssal ocean is slightly cooler in the bottom-enhanced mixing scheme (0.5 ∘C), suggesting a weaker diffusive heat penetration into the abyss, in line with theoretical predictions. We also find abyssal salinity differences on the order of 0.1 psu. Given the very long spin-up time of the simulations, the lack of divergence between the schemes is surprising. In the Arctic Ocean, the differences in deep ocean temperature and salinity are greater. This is because the deep waters in the Arctic are very stably stratified by the halocline and can only communicate with the surface diffusively.

In order to examine the effective diffusivity Keff of each mixing scheme, we computed the stratification-weighted diffusivity (de Lavergne et al., 2016) between the two schemes as follows: Keff=∬KvN2dxdy∬N2dxdy where Kv is the total vertical diffusivity, and N2 is the buoyancy frequency. This was computed offline using monthly output of Kv and N2 and averaged across the last 50 years of simulation (Fig. 9). This calculation shows that the effective mixing across the thermocline is indeed similar between the two schemes. Mixing is slightly stronger in the Bryan–Lewis scheme between 2500 and 3000 m, while mixing near the seafloor is indeed stronger in the bottom-enhanced scheme. There is also larger effective mixing in the Atlantic than the Indo-Pacific in both schemes, and also a larger difference between the bottom-enhanced mixing and the Bryan–Lewis scheme in the Atlantic.

We further note that the Simmons et al. (2004) mixing scheme used in our control run assumes that the energy input to the mixing within each grid column is proportional to the simulated bottom stratification. An increase in bottom stratification (e.g. as a result of increased surface density gradients) is, therefore, immediately paralleled by an increase in mixing energy. This can have the effect of maintaining approximately constant diffusivities and circulation rates despite larger density contrasts. In this way, the Simmons et al. (2004) scheme is similar to the Bryan–Lewis scheme in that it fixes diffusivity rather than mixing energy. A better approach may be to fix the energy consumed by vertical mixing, since energy constraints provide a physically consistent framework to relate the drivers of mixing to the abyssal circulation (de Lavergne et al., 2016).

A comparison of proxy data with the model SSTs for the 400, 800 and 1600 ppm experiments is shown in Fig. 10. The data are taken from the SST proxy compilation from 38 to 34 Ma shown in Table A2 of Baatsen et al. (2018a). The compilation combines UK'37 , TEX86, Mg / Ca, δ18O and Δ47 data (Kamp et al., 1990; Pearson et al., 2001, 2007; Tripati and Zachos, 2002; Kobashi et al., 2004; Bijl et al., 2009; Liu et al., 2009; Okafor et al., 2009; Wade et al., 2012; Douglas et al., 2014; Petersen and Schrag, 2015; Hines et al., 2017; Evans et al., 2018). The palaeolocations of these proxies are consistent with the model boundary conditions (Baatsen et al., 2016). Figure 10 shows the proxy data points compared with the nearest model ocean point, and the model zonal mean temperature. The 1600 and 800 ppm cases are too warm compared with the proxies in the tropics. In the control case (800 ppm) the model is roughly 3 ∘C warmer at the equator. However, the TEX86 data of Pearson et al. (2007) from Tanzania (∼ 32 ∘C) agree better with the control case. In the mid to high latitudes, the control case is colder than the proxy estimates. The 1600 ppm case agrees better with the mid to high latitude proxy estimates, however its warm bias at the equator is very large. In the 400 ppm simulation, the tropical temperatures agree well with the proxies, while in the midlatitudes, it is too cold.

The model's inability to capture the low meridional temperature gradient in the proxies is a common problem in simulating warm climates of the Eocene (Huber and Caballero, 2011). However, several aspects of the proxy data could help to improve this situation. First, alternative calibrations of the TEX86 data from Tanzania may yield temperatures as high as 35 ∘C in the late Eocene (Bijl et al., 2009), which would reduce the low gradient problem. We also note that there are no available proxy estimates from 38 to 34 Ma in the western Pacific warm pool, where the ocean's warmest temperatures are found. Middle Eocene (42 to 38 Ma) SST estimates from Java are as high as 35 to 36 ∘C (Evans et al., 2018). Second, high latitude TEX86 data may be affected by summer bias due to seasonal growth of their underlying species (Kim et al., 2010), which may have the effect of reducing the meridional temperature gradient seen in the proxies.

The global mean SST in the control run (800 ppm) is 27.8 ∘C; halving CO2 to 400 ppm produces 3.5 ∘C of cooling, and doubling CO2 to 1600 ppm yields 4.2 ∘C of warming. The global mean SAT in the control is 25.6 ∘C; halving CO2 to 400 ppm produces 4.0 ∘C of cooling, and doubling produces 4.8 ∘C of warming. The higher sensitivity of temperature change in the atmosphere is due to both stronger polar amplification in the atmosphere than the ocean, and a stronger temperature change over land than over the ocean. In its modern configuration, CM2.1 has an equilibrium sensitivity of 3.4 ∘C (Winton et al., 2010), although we note that this sensitivity may be different when using the present lower resolution atmospheric-model component. The lower sensitivity in the modern case may be due to many factors, for example higher albedo due to the presence of ice sheets, and the higher percentage of land versus ocean. The net incoming shortwave radiation from our late Eocene model has an extra 12 W m-2 of insolation compared with the modern CM2.1, indicating a significantly lower albedo. Furthermore, our simulations indicate some state dependence, with higher sensitivity at higher CO2.

The response of SST and SAT to CO2 changes is shown in Fig. 11. In the 1600 ppm case (Fig. 11a, c), the tropics warm by between 3 and 4 ∘C, with a strong amplification in both of the polar regions. SAT over the Arctic Ocean and over Antarctica are 7–9 ∘C warmer, with the magnitude of polar amplification being similar in both hemispheres. This approximate doubling of warming in the polar regions aligns with expectations from idealised climate modelling of the polar amplification response (Alexeev et al., 2005). The polar amplification is a combination of the local radiative forcing and the enhanced moist energy transport from the tropics. The Arctic Ocean warming is especially pronounced in the 1600 ppm case, despite the absence of ice–albedo feedbacks. Cloud radiation feedbacks may provide additional polar amplification at high levels of CO2 (Abbot et al., 2009). Warming over land shows a mixed response; warming over North America and Siberia is enhanced compared with oceanic regions at the same latitude, as is the case in southern Africa. By contrast, other regions such as Australia, equatorial Africa and parts of South America show little change to the land–ocean contrast. This is linked directly with the land–ocean monsoon response in low to mid latitudes. Regions of enhanced land warming are closely associated with enhanced drying, while regions of reduced land warming generally also become wetter, as shown in the P-E forcing (Fig. 12). In the higher latitudes, the monsoonal forcing is weaker, and thus the P-E changes are not as closely linked with the temperature response.

In the 400 ppm case (Fig. 11b, d), the magnitude of tropical cooling and polar amplification are approximately as strong as in the warming case. The Antarctic SAT response is, however, somewhat stronger than the warming case due to the triggering of snow albedo feedbacks. Under 400 ppm, a far greater proportion of the Antarctic continent is covered in snow in winter. The model does not sustain freezing temperatures over Antarctica in summer. However, this should not be directly interpreted as evidence that an ice sheet cannot be triggered under this climate; the model lacks the necessary feedbacks to simulate the accumulation of long term snow and ice (e.g. Gasson et al., 2014). One area of strong cooling occurs in the Weddell Sea, where SSTs drop by ∼ 8 ∘C, with a commensurate change in SAT. There is also a cessation of deep water formation in this region, and the formation of seasonal sea ice in the 400 ppm case (see Fig. 6). The Pacific sector of the Southern Ocean cools to a similar magnitude as the warming case. The SST cooling in the Arctic is less pronounced than in the warming case, simply because the sea surface freezes for much of the year and therefore the surface change is limited to ∼ 4 ∘C. The SAT response in the Arctic is similarly strong, showing that the atmosphere polar amplification can be much stronger than the Arctic SST change.

In response to a doubling of CO2 to 1600 ppm, surface salinity decreases markedly over the mid and high northern latitudes as shown in Fig. 12a. Arctic Ocean salinities decrease by around 1 psu, due to the enhancement of the high latitude precipitation and river run-off into the Arctic. The river run-off accumulates from North America, Europe and Siberia, so the net effect of the increased P-E (Fig. 12c) on the salinity is particularly concentrated in the Arctic. This Arctic freshening also has a marked impact on North Atlantic salinity. Conversely when CO2 is halved, Arctic salinities are enhanced by 1–2 psu, due to the weakening of precipitation and run-off into the Arctic.

Salinity changes in other regions also reflect changes in the P-E balance to a lesser extent. The north-east Pacific freshens in response to warming and becomes saltier in response to cooling, due to the same forcing mechanisms. The interiors of the gyre circulations in the Pacific show signatures of the P-E forcing, but the salinity differences are not as clearly correlated here, due to advection and lateral mixing altering the signal. One area of notable increase in salinity under the warming scenario is the southern tropical Pacific, where salinity is greatly enhanced. This corresponds to an area of net drying in the P-E fields, but other regions do not show nearly the same salinity response to the same magnitude of forcing. Clearly advection feedbacks are at play, as also evidenced by the net freshening of the South Atlantic in the 1600 ppm case. This occurs despite a net decrease in P-E over the subtropical South Atlantic gyre. One possible explanation is the enhanced African monsoon, which leads to large river run-off increases from West Africa into the Atlantic. Furthermore, sinking in the Atlantic sector of the Southern Ocean decreases markedly in the 1600 ppm case, which is evident in both the mixed layer depths and age tracers at intermediate depth (not shown). This reduction in deep water formation in turn reduces the salt advection into the Weddell Sea, leading to freshening at the surface. In the 400 ppm case, the tropical South Atlantic becomes saltier, which is likely due to a decrease in river run-off from West Africa. However, the Weddell Sea becomes fresher in this case, due to a reduction of sinking in the South Atlantic and its associated salt advection.

Greenhouse warming enhances the hydrological cycle, due to the increased capacity of the atmosphere to hold water vapour. This effect is known as the Clausius–Clayperon relation. In the zonal mean, this effect has been shown to create a pattern of change where wet regions get wetter and dry regions get drier. A simple scaling relation can be derived that relates the change (denoted by δ) in precipitation minus evaporation (P-E) to its original distribution (Held and Soden, 2006): δ(P-E)=αδT(P-E), where α is a constant and δT is the change in temperature. This scaling relation has been shown to hold true over ocean regions in modern observations (Durack et al., 2012), but not over land where the dynamics of rainfall distribution are more complex (Greve et al., 2014). In line with the above scaling argument, the oceanic patterns of wet vs dry regions (Fig. 4d) are enhanced in most ocean regions when CO2 is doubled. The equatorial response of P-E has some exceptions to the scaling relation, as the western Pacific warm pool and equatorial cold tongue do not respond according to the same paradigm. However, we note that the equatorial Pacific is subject to a strong zonal asymmetry due to El Niño variability, and these patterns do not follow the scaling relation as cleanly as the latitudinal relationship even in modern climate models (Held and Soden, 2006).

The basin integrated forcing of P-E and river run-off upon the surface ocean is summarised in Fig. 13. This illustrates the combined effects of P-E forcing and river run-off in creating much fresher Arctic conditions compared to present day. In comparing the control run (800 ppm) with the 1600 and 400 ppm cases, the warmer simulations give rise to increased freshwater forcing in the North Pacific to a similar degree as in the Arctic. However, the Arctic Ocean has a much smaller surface area and restricted outflow; therefore the impacts on surface salinity of this forcing are strongest. The connection of Arctic surface waters to the Atlantic gives rise to large salinity changes in the North Atlantic, whereas the North Pacific salinity is less impacted by this change. Furthermore, the freshwater forcing is concentrated in the north-east of the basin, away from the deep water formation region in the north-west. Thus, the North Pacific is able to maintain deep water formation despite the enhanced freshwater forcing.

Overall, these changes in salinity are not enough to alter the preferred regions of northern sinking. The palaeogeography employed here robustly generates sinking in the North Pacific and in the Southern Ocean. In all cases, the northern cell is warmer and saltier, and the southern cell is colder and fresher and thus is heavier in the abyss, as in the present day. The cooling-induced weakening of the hydrological cycle does bring the North Atlantic closer to a regime of deep water formation, but these changes of 1–2 psu are not enough to trigger sinking. If we hypothesise that NADW did form at the EOT (Borrelli et al., 2014), then other mechanisms are needed to first raise the control state salinities in the North Atlantic.

Meridional overturning circulation is weaker in the 1600 ppm case, by around 10 Sv in the northern cell, and up to 15 Sv in the southern cell (Fig. 14a). Although this is a substantial reduction, the overall structure remains similar to that seen in the control run (Fig. 7), with sinking in the North Pacific and Southern Ocean, albeit with a lower magnitude. The reduction in magnitude may be partly due to the 400 and 800 ppm runs being further from equilibrium, and therefore the deep ocean is expelling more heat. Otherwise, the high CO2 case has a lower meridional temperature gradient, which may alter the forcing of the inter-hemispheric MOC (Wolfe and Cessi, 2014). In the 400 ppm case, the overall magnitude of the MOC is similar in magnitude to the control, however there is a northward shift of the southern cell, which manifests as a positive change in the high southern latitudes in Fig. 14b. As the climate cools, a stable halocline develops in the Weddell Sea which ceases to form deep/bottom water. In winter, however, the halocline supports the growth of seasonal sea ice in the region.

In the control 800 ppm simulation, the surface waters in the Ross Sea are slightly warmer and more saline than those in the Weddell Sea, but they have densities close enough to allow both regions to induce deep convection in winter (Fig. 15). With lowered temperatures in 400 ppm simulation, the nonlinear dependence on the temperature of sea water density will, for the same temperature decrease, cause a larger density increase of the warmer Ross Sea surface waters. Presumably, this feature combined with the salt-advection feedback enhances the surface density difference between the two regions to the point where deep convection ceases in the Weddell Sea. As this transition occurs a stable halocline develops, supporting the growth of seasonal sea ice in the Weddell Sea (Goosse and Zunz, 2014). However, the net sea ice formation is too weak in the 400 ppm simulation to produce any brine enriched deep water, which is an important feature for the present-day dense-water formation around Antarctica. In the 1600 ppm case, freshening occurs due to enhancement of the hydrological cycle, leading to cessation of Weddell Sea convection. Vigorous sinking occurs in the South Pacific in all cases, analogous to Ross Sea convection in the present day. This area is more predisposed to sinking due to a relative lack of river run-off ending up in that region, while the strong subpolar gyre circulation allows lower latitude water to maintain a salt feedback into the region. North Pacific surface waters are warmer and saltier than in the Southern Ocean sinking regions, and thus establish a deep water mass of comparable density to that of the Southern Ocean. However, the Southern Ocean water mass, having lower temperature, becomes the densest water mass in the abyss due to thermobaric effects.