EART

H,A

TMO

SPH

ERIC

,A

ND

PLA

NET

ARY

SCIE

NCE

S

Glacial ocean circulation and stratification explainedby reduced atmospheric temperatureMalte F. Jansena,1

aDepartment of the Geophysical Sciences, The University of Chicago, Chicago, IL 60637

Edited by Mark H. Thiemens, University of California, San Diego, La Jolla, CA, and approved November 7, 2016 (received for review June 27, 2016)

Earth’s climate has undergone dramatic shifts between glacial andinterglacial time periods, with high-latitude temperature changeson the order of 5–10 ◦C. These climatic shifts have been asso-ciated with major rearrangements in the deep ocean circulationand stratification, which have likely played an important role inthe observed atmospheric carbon dioxide swings by affecting thepartitioning of carbon between the atmosphere and the ocean.The mechanisms by which the deep ocean circulation changed,however, are still unclear and represent a major challenge to ourunderstanding of glacial climates. This study shows that vari-ous inferred changes in the deep ocean circulation and stratifica-tion between glacial and interglacial climates can be interpretedas a direct consequence of atmospheric temperature differences.Colder atmospheric temperatures lead to increased sea ice coverand formation rate around Antarctica. The associated enhancedbrine rejection leads to a strongly increased deep ocean strati-fication, consistent with high abyssal salinities inferred for thelast glacial maximum. The increased stratification goes togetherwith a weakening and shoaling of the interhemispheric over-turning circulation, again consistent with proxy evidence for thelast glacial. The shallower interhemispheric overturning circula-tion makes room for slowly moving water of Antarctic origin,which explains the observed middepth radiocarbon age maximumand may play an important role in ocean carbon storage.

LGM | AMOC | stratification | cooling | sea-ice

The deep ocean today is ventilated mainly by two watermasses. North Atlantic deep water (NADW) is formed in

the North Atlantic before flowing southward at a depth of about2–3 km and eventually rising back up to the surface in theSouthern Ocean. Antarctic bottom water (AABW) is formedaround Antarctica and spreads northward into the abyssal basins.The AABW that makes it into the Atlantic then slowly upwellsinto the lower NADW before returning southward and resur-facing again around Antarctica (1, 2). The glacial equivalentof NADW was likely confined to shallower depths, leavingmore of the deep Atlantic filled with water masses originat-ing primarily from around Antarctica (3–5). Moreover, the twowater masses appear more distinct, with less mixing betweenthem (6).

Multiple studies have pointed toward the potential impor-tance of sea ice and surface buoyancy fluxes around Antarc-tica in controlling changes in the deep ocean stratification andcirculation between the present and Last Glacial Maximum(LGM) (7–11). Specifically, we recently showed that enhancedbuoyancy loss around Antarctica is expected to lead to anincrease in the abyssal stratification, an upward shift of NADW,and a clearer separation between NADW and southward-flowingAABW (11)—all in agreement with inferences made for differ-ences in circulation and stratification between the present andLGM (3, 6, 12, 13). This gives rise to the hypothesis that strongcooling around Antarctica led to an increased net freezing rate.The resulting net salt flux into the ocean amounts to an increasedbuoyancy loss, which then led to the observed changes in deepocean stratification and circulation. This hypothesis is tested inthis study.

We test the connection between atmospheric temperatureand ocean circulation and stratification changes, using idealizednumerical simulations, which allow us to isolate the proposedmechanism. We use a coupled ocean–sea-ice model, with atmo-spheric temperature, winds, and evaporation–precipitation pre-scribed as boundary conditions (Materials and Methods). Themodel uses an idealized continental configuration resembling theAtlantic and Southern Oceans, where the most elemental circu-lation changes have been inferred (3, 5, 6, 12).

ResultsWe first focus on the model’s ability to reproduce key featuresof the modern (interglacial) ocean state and circulation. Fig. 1Ashows the sea surface temperature (SST) over the simu-lated domain, with boundary conditions resembling present-dayatmospheric forcing. The chosen forcing gives SSTs in broadagreement with observations, varying from about 28◦C in thetropics to the freezing point (∼ −2 ◦C) around Antarctica, wheresea ice forms (Fig. 1A, white contours). The North Atlantic issomewhat warmer and ice free. Fig. 2 shows zonally averageddepth–latitude sections of potential temperature (Fig. 2A) andsalinity (Fig. 2C), which reproduce the basic features observedin the present-day Atlantic. In particular, we see a tongue ofsalty water at middepth, penetrating southward into the South-ern Ocean and leading to a reversal of the salinity stratifica-tion in the abyss (2). The associated overturning streamfunc-tion (Fig. 3A) reveals that this salty tongue is associated withNADW that flows southward as part of the Atlantic meridionaloverturning circulation (AMOC). Below the clockwise AMOCcell lies an anticlockwise abyssal cell, representing the pathwayof AABW. The peak overturning transports of +17.5 Sverdrup

Significance

To understand climatic swings between glacial and interglacialclimates we need to explain the observed fluctuations inatmospheric carbon dioxide (CO2), which in turn are mostlikely driven by changes in the deep ocean circulation. Thisstudy presents a model for differences in the deep oceancirculation between glacial and interglacial climates consis-tent with both our physical understanding and various proxyobservations. The results suggest that observed changes inocean circulation and stratification are caused directly byatmospheric cooling or warming, which has important impli-cations for our interpretation of glacial–interglacial transi-tions. In particular, the direct link between atmospheric tem-perature and ocean circulation changes supports the notion ofa positive feedback loop between atmospheric temperature,ocean circulation, and atmospheric CO2.

Author contributions: M.F.J. designed research, performed research, analyzed data, andwrote the paper.

The author declares no conflict of interest.

This article is a PNAS Direct Submission.1Email: mfj@uchicago.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1610438113/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1610438113 PNAS | January 3, 2017 | vol. 114 | no. 1 | 45–50

Fig. 1. (A and B) Sea surface temperature (colors) and sea-ice concentration(white contours), for simulations with boundary conditions representingpresent-day conditions (A) and LGM conditions with reduced atmospherictemperature (B). Contour interval for sea-ice concentration is 20%.

(SV) and −5.2 SV are roughly consistent with the observed rateof NADW formation and the transport of AABW into the NorthAtlantic (1, 2).

We now analyze the response of the model’s equilibrium solu-tion to a reduction in atmospheric temperature. Consistent with

Fig. 2. (A–D) Zonal mean temperature (A and B) and salinity (C and D) for simulations with boundary conditions representing present-day forcing (A andC) and simulations with reduced atmospheric temperature resembling LGM conditions (B and D). Contour intervals are 1◦C in A and B and 0.05 g/kg in Cand D. Note that the colorbar ranges have been cropped to focus on the deep ocean.

proxy data for the LGM, the prescribed atmospheric cooling ispolar amplified, ranging from 2◦C in the tropics to 6◦C aroundAntarctica (14, 15). All other boundary conditions are held fixed.The reduction in atmospheric temperature leads to a reductionin ocean surface temperature and an expansion of sea ice aroundAntarctica, as well as the appearance of sea ice in the NorthAtlantic (Fig. 1B). Moreover, the model suggests a cooling andsalinification of AABW, leading to a strong salt stratification inthe deep ocean, which replaces the reversed salinity stratifica-tion observed with present-day forcing (Fig. 2B). The strong saltstratification is consistent with pore-fluid data for the LGM (12).(To compare the absolute salinity values here to those inferredfor the LGM, one needs to account for the increased bulk oceansalinity resulting from the eustatic drop in sea level, which is notincluded in the simulations and would add around 1 g/kg glob-ally.) The AMOC cell weakens and shoals (Fig. 3B), again consis-tent with inferences for the LGM (3–5). The abyssal cell slightlystrengthens and becomes somewhat more confined to the abyss,leading to an increased separation between the two overturningcells, which may have played an important role in the increasedocean carbon storage (6).

In the real ocean the abyssal overturning cell is distributedover multiple basins and currently overlaps in depth and densitywith the AMOC cell, leading to a continuous exchange of waterbetween the two cells in the Southern Ocean (1, 2). Whereasthis interbasin overlap between the present-day overturning cellscannot be modeled in a single-basin model, the simulated con-traction of both the AMOC and abyssal cells is likely to berobust (11). In a multibasin configuration, this contraction ofboth cells is expected to remove or reduce the overlap betweenthe two cells, thus generating a more isolated abyssal water mass.The rearrangement of the overturning circulation would likely

46 | www.pnas.org/cgi/doi/10.1073/pnas.1610438113 Jansen

EART

H,A

TMO

SPH

ERIC

,A

ND

PLA

NET

ARY

SCIE

NCE

S

Fig. 3. (A and B) Meridional overturning streamfunction (colors) and potential density referenced to 2 km depth (black lines), for simulations with boundaryconditions representing present-day forcing (A) and LGM conditions with reduced atmospheric temperature (B). The overturning streamfunction is computedfrom the sum of the Eulerian zonal mean velocity and the parameterized eddy-induced bolus velocity. [Note that the isopycnal overturning transport includesan additional component associated with standing meanders (11), which is not included in Fig. 3. The contribution of standing meanders largely cancels theapparent diapycnal zonal-mean transport in the channel region.]

resemble that sketched in Ferrari et al. (9), although the mecha-nism proposed here is somewhat different.

The changes in the deep ocean circulation and stratificationhere result from an increased buoyancy loss rate around Antarc-tica, which in turn results primarily from enhanced brine rejec-tion associated with sea-ice formation and export. Sea-ice exportis proportional to the product of the ice load (here defined as thetime- and zonal-mean mass of sea-ice and snow per unit area)and equatorward transport velocity. Both increase as the atmo-spheric temperature is reduced, with the dominant role playedby differences in the ice load, which (near its maximum) goesup from about 300 kg/m2 in the “present” simulation to about800 kg/m2 in the “LGM” simulation. The ice export velocity alsoincreases, as sea ice extends farther northward where the west-erly winds are stronger. As a result of the larger ice load andexport velocity, the peak ice export rate from around Antarcticaincreases from about 3 ×107 kg/s to about 14 ×107 kg/s.

To compute the effective net buoyancy loss around Antarc-tica it is important to consider the nonlinearity of the equationof state and in particular the pressure dependence of the ther-mal expansion coefficient (16). If surface buoyancy fluxes arecomputed using the surface haline and thermal expansion coef-ficients, virtually no buoyancy loss around Antarctica is found inthe present-day–like simulation. This lack of buoyancy loss wouldappear to be at odds with the presence of an abyssal cell and thetransformation of upwelling circumpolar deep water (CDW) toAABW. The apparent contradiction can be resolved by notingthat the density increase associated with the transformation ofCDW to AABW is dominated by a cooling and counteracted bya freshening. Whereas cooling has a small effect on surface den-sities at cold temperatures, the temperature effect is amplifiedas a water parcel sinks into the deep ocean. We can estimatethe effect of heat and salt fluxes on a parcel at depth by com-puting buoyancy fluxes based on changes in potential densitiesreferenced to 2 km depth (consistent with the potential densitiesshown in Fig. 3), which yields an integrated surface buoyancyloss rate around Antarctica for the present-day–like simulationof about 4.4 ×103 m4· s−3 (Materials and Methods). In the LGMsimulation the integrated surface buoyancy loss rate increases to2.1× 104 m4· s−3.

The larger buoyancy loss rate around Antarctica in the LGMsimulation gives rise to the observed changes in deep ocean cir-culation and stratification. Because surface buoyancy loss around

Antarctica has to be balanced by vertical diffusion in the basin,the deep ocean stratification is expected to depend approxi-mately linearly on the buoyancy loss rate (11). This relationshipexplains the increase in the deep ocean stratification between thepresent and LGM simulations (although a quantitative compar-ison of changes in buoyancy loss and stratification is somewhatcomplicated by the nonlinearity in the equation of state). Theincreased stratification in the LGM then leads to an upward shiftof NADW, consistent with the results of Jansen and Nadeau (11).

Sensitivity ExperimentsTo test the robustness of the results to additional modificationsin the boundary conditions, we consider a number of sensitivityexperiments, varying the wind stress and the vertical turbulentdiffusivity, as well as the spatial structure of atmospheric temper-ature change. The results of these simulations are summarized inTable 1 and are briefly discussed in the following.

In a seminal paper, Toggweiler et al. (17) proposed an equa-torward shift in the latitude of the Southern Hemisphere sur-face westerlies as a potential mechanism for differences in theocean circulation between the present and LGM. Observationalevidence, however, allows for an equatorward shift of at mostabout 3◦ (18, 19), which in turn is here found to have negligibleimpact on the solution (experiment “LGM windN” in Table 1).Proxy observations and climate models instead indicate a slightpoleward shift and strengthening of the surface westerlies overthe Southern Ocean (18, 19). A simulation incorporating a 3◦

southward shift and 20% strengthening of the Southern Hemi-sphere westerlies shows a moderate increase in ice export andassociated buoyancy loss around Antarctica (experiment “LGMwindS” in Table 1). The increased buoyancy loss rate amplifiesthe differences between the present and LGM simulations dis-cussed above.

The loss of shallow shelf seas during the LGM has likely led toincreased tidal energy dissipation in the deep ocean, which in turnmay have caused enhanced vertical mixing (20, 21). On the otherhand, the increased deep ocean stratification during the LGMmay have suppressed vertical mixing, as more turbulent kineticenergy dissipation would be required to mix the more stratifiedwater column (22). In our simulations, which assumed unchangedvertical diffusivities, the implied energy input to mixing below theupper thermocline (300 m) almost doubles between the presentand LGM simulations (Table 1, last column).

Jansen PNAS | January 3, 2017 | vol. 114 | no. 1 | 47

Table 1. Summary of results from sensitivity experiments

Experiment∫

BdA N2deep ΨNADW ΨAABW zNADW zΨ=0 zAABWreturn emix

Present 0.4× 104 0.6× 10−4 17.5 5.2 1,450 2,050 2,370 0.53× 10−6

LGM 2.1× 104 2.3× 10−4 9.6 6.2 1,070 1,590 2,480 1.02× 10−6

LGM windS 3.4× 104 3.8× 10−4 9.5 7.2 1,000 1,500 2,530 1.42× 10−6

LGM windN 1.7× 104 2.0× 10−4 9.6 6.1 1,040 1,550 2,430 0.95× 10−6

LGM κ F02D 50% 2.3× 104 3.9× 10−4 6.4 4.8 810 1,250 2,490 0.73× 10−6

LGM κ + 50% 2.2× 104 1.8× 10−4 12.2 8.6 1,210 1,710 2,250 1.35× 10−6

LGM dTSH 1.6× 104 2.1× 10−4 10.9 6.2 1,120 1,660 2,520 0.98× 10−6

LGM dTconst 1.8× 104 2.2× 10−4 8.7 6.3 1,080 1,620 2,500 0.96× 10−6

Present seas 0.3× 104 0.4× 10−4 15.2 5.0 1,550 2,220 2,530 0.46× 10−6

LGM seas 1.6× 104 1.8× 10−4 9.9 6.1 1,190 1,750 2,410 0.85× 10−6

Warm 0 0.1× 10−4 19.9 — 2270 — — 0.35× 10−6

Each row indicates a different numerical simulation: Present denotes the simulation with boundary conditions resembling present-day conditions. LGMdenotes the simulation with reduced atmospheric temperature, resembling LGM conditions. LGM windS denotes a simulation with atmospheric tempera-tures as in LGM, but also including a 3◦ southward shift and 20% strengthening of the Southern Hemisphere (SH) westerlies. In LGM windN SH westerliesinstead are shifted 3◦ northward (Fig. S3B). LGM κ-50% and LGM κ + 50% denote simulations with boundary conditions analogous to LGM but withdiapycnal diffusivities reduced or enhanced by 50%, respectively (Fig. S5). LGM dTSH denotes a simulation where atmospheric temperatures have beenreduced only in the Southern Hemisphere, whereas LGM dTconst denotes a simulation where temperatures have been reduced homogeneously over thewhole domain by 5◦C relative to Present. Present seas and LGM seas denote simulations analogous to Present and LGM, but with seasonally varying airtemperature forcing (Fig. S1A). Warm denotes a global warming simulation, with similar but opposite signed changes in the atmospheric temperature as inthe LGM simulation (Fig. S3A). Columns show various diagnostics:

∫BdA (in m4s−3) is the integrated buoyancy loss rate around Antarctica (Materials and

Methods). N2deep is the mean stratification (referenced to 2 km depth) averaged over the deep ocean basin below 1,500 m and north of 40◦S (in s−2). ΨNADW

indicates the maximum AMOC overturning transport and ΨAABW is the maximum overturning transport in the abyssal cell (both in Sv). zNADW denotes themedian depth of NADW, computed as the depth (in meters) at which the basin-averaged overturning streamfunction is reduced to 50% of its maximumvalue. zΨ=0 is the depth at which the basin-averaged overturning streamfunction changes sign; and zAABWreturn is the depth where the streamfunctionreaches 50% of its minimum value, thus denoting the median depth of the southward return flow of AABW. emix is the globally averaged energy input perunit area by vertical mixing below 300 m (in units m3·s−3).

To test the sensitivity of our results to the rate of verticalmixing, we consider two additional LGM simulations in whichthe vertical diffusivity is reduced or increased by 50% (“LGM κ−50%” and “LGM κ + 50%”, respectively). Due to the strongerstratification, the implied energy input to mixing below 300 mis still slightly larger than for present-day conditions in LGMκ −50%, and the energy input is almost tripled in LGM κ +50%. Reduced vertical mixing weakens and shoals the AMOCcell during the LGM and increases the abyssal stratification—thus amplifying the effects of atmospheric cooling. Enhancedvertical mixing instead strengthens and deepens the AMOC celland reduces the abyssal stratification—thus counteracting theeffects of atmospheric cooling. However, even the simulationwith enhanced mixing maintains a stronger stratification andweaker and shallower AMOC cell compared with the simulationrepresenting present-day conditions, indicating that the effect ofatmospheric cooling remains dominant.

Significant uncertainty also exists in the spatial pattern ofatmospheric temperature change between the present and LGM(14, 15). We consider two sensitivity experiments, which repre-sent extreme cases: one where atmospheric cooling is restrictedto the Southern Hemisphere (experiment “LGM dTSH” inTable 1) and one where a globally constant cooling is applied(experiment “LGM dTconst”). Both experiments show roughlysimilar results as the LGM reference case with symmetric polaramplified cooling, as long as the reduction in atmospheric tem-perature at high southern latitudes remains about the same. Thisresult confirms our interpretation that circulation and stratifica-tion changes are controlled primarily by differences in the sur-face boundary conditions around Antarctica.

The simulations here are highly idealized and, among otherthings, do not include a seasonal cycle, which may affect eventhe mean sea-ice growth and export rate around Antarctica. Toanalyze the effect of seasonality, the present and LGM simula-tions were repeated using seasonally varying air temperatures,but leaving the annual mean temperatures unchanged (experi-ments “Present seas” and “LGM seas” in Table 1 and Fig. S1).Even though the idealized representations of surface boundary

conditions, mixed layer dynamics, and sea-ice thermodynamicsmake this model arguably less suited to represent the seasonalcycle, the simulations do exhibit strong seasonality in sea-icecover around Antarctica (Fig. S1B). Whereas the addition of aseasonal cycle causes some minor changes in the deep oceancirculation and stratification in both the present and LGM sim-ulations, the main results regarding differences in the stratifi-cation and circulation between the present and LGM remainunchanged by the inclusion of a seasonal cycle (Fig. S1C).

To address the potential implications of our results for long-term future climate change, we finally examine a “global warm-ing” simulation (experiment “Warm” in Table 1), which showsessentially reversed results from the cooling experiments: Above-freezing temperatures around Antarctica lead to ice-free condi-tions and net buoyancy gain. As a result AABW formation is shutdown, leaving the entire deep ocean filled with nearly unstrati-fied water of North Atlantic origin (Fig. S2). The effect of thesepotential rearrangements on ocean carbon storage represents animportant topic for future research.

Discussion and ConclusionsThe results of this study suggest that atmospheric coolingalone—via a modification of the buoyancy loss rate aroundAntarctica—leads to a strong increase in the deep ocean strat-ification, a shoaling of NADW, and an increased separationbetween NADW and the underlying abyssal overturning cell.The simulated response of the deep ocean circulation and strat-ification to atmospheric temperature change is consistent withdifferences between the present and LGM inferred from paleo-proxy observations (3, 6, 12, 13). A series of sensitivity experi-ments suggests that the dominant control over circulation andstratification changes is exerted by the atmospheric temperaturearound Antarctica. Temperature changes in other regions, aswell as differences in the atmospheric wind stress, instead playonly a relatively minor role.

The results here are broadly consistent with some complexcoupled climate model simulations (7, 8), as well as with globalocean-only simulations under LGM boundary conditions (10).

48 | www.pnas.org/cgi/doi/10.1073/pnas.1610438113 Jansen

EART

H,A

TMO

SPH

ERIC

,A

ND

PLA

NET

ARY

SCIE

NCE

S

Coupled LGM climate simulations using the Community Cli-mate System Model (CCSM)1.4 and CCSM3 show a stronglyincreased abyssal stratification and shallower NADW (7, 23, 24).Consistent with the results discussed here, these changes havebeen attributed to enhanced sea-ice export around Antarctica(7) and ultimately reduced CO2 concentrations (8). However,whereas Shin et al. (7) argue that enhanced sea-ice export iscaused “ultimately by increased westerlies,” (ref. 7, p. 1) we heresuggest that cooling alone is sufficient and likely to be the domi-nant driver. Changes in freshwater fluxes around Antarctica havealso been argued to explain the high deep ocean stratification in adecoupled Community Earth System Model (CESM)1.1.2 oceansimulation under LGM boundary conditions (10). Many othercoupled climate models, however, show different and widelydiverging changes in the deep ocean circulation and stratifica-tion between preindustrial and LGM simulations (25, 26). Pos-sible reasons for this disagreement include differences betweentransient and fully equilibrated solutions (27, 28), insufficientincrease in Antarctic sea-ice cover and formation (29), and/orcompensating effects due to other differences in the boundaryconditions, such as changes in the ice sheets (26, 30). A moredetailed analysis of comprehensive LGM climate simulations isneeded to better understand differences between them, but isbeyond the scope of this study.

The dominant role of atmospheric temperature change as thedriver of differences in the ocean circulation and stratificationhas important implications for our understanding of glacial–interglacial climate swings. Although the exact mechanisms arestill debated, it is likely that changes in the ocean circulation andstratification have played a key role in modulating ocean carbonstorage and thus atmospheric CO2 concentrations (31, 32). Ifocean circulation changes are themselves directly driven by tem-perature swings, as argued here, they are likely to play a cru-cial role in a positive feedback loop between global temperature,ocean circulation, and atmospheric CO2. A better comprehen-sion of this feedback loop will be central to our understandingof past and future climatic changes.

Materials and MethodsNumerical Model Configuration. The numerical simulations use the Mas-sachusetts Institute of Technology (MIT) general circulation model (MIT-gcm) (33), in a hydrostatic Boussinesq configuration. The idealized domainextends from 70◦S to 65◦N, covers 72◦ in longitude, and is 4 km deep. Thehorizontal resolution is 1◦×1◦ and the vertical resolution ranges from 20 mnear the surface to 200 m in the deep ocean, with a total of 29 levels. Theocean is bounded by a 1◦ strip of land on all sides, which is interrupted onlybetween 69◦S and 48◦S, where, above 3 km depth, zonally periodic bound-ary conditions give rise to a reentrant channel, representing the SouthernOcean.

Atmospheric temperatures, net freshwater flux (i.e., precipitation–evaporation), and surface winds are prescribed in the form of idealized ana-lytical functions shown in Fig. S3. Except in the sensitivity simulations withseasonal cycle in air temperatures, all atmospheric forcing fields are con-stant in time. Momentum and sensible heat exchange between the atmo-sphere and the ocean are described using standard bulk formulas (34, 35).In addition to the sensible heat flux, an idealized radiative restoring is pre-scribed as Frad = σ(T4

s − T4a ), where Ts is the ocean/ice surface tempera-

ture, Ta is the prescribed atmospheric temperature, and σ = 5.67 × 10−8

W·m−2·K−4 is the Stefan–Boltzmann constant. In reality the ocean expe-riences a net radiative heating, which is balanced primarily by latent heatloss. Both are ignored here, which in effect simplifies the thermal boundaryconditions to a restoring to the prescribed atmospheric temperature—albeitwith a restoring rate that is modified by wind speed and boundary layer sta-bility. This idealization was chosen to avoid additional assumptions aboutchanges in downward short- and long-wave radiation and boundary layer-specific humidity. Moreover, use of a bulk formula for moisture and latentheat transfer would require artificial flux adjustments to close the globalsalt budget.

To test the sensitivity of the results to the details of the thermal bound-ary conditions, additional simulations were performed with a simple linearrestoring of the ice/ocean surface temperature (36) (Fig. S4). The surface

heat flux is computed as F = 25 W/(m2K)(Ts − Ta). This amounts to asomewhat faster effective restoring than what is on average implied bythe original boundary conditions. To still obtain a realistic present-day–likesimulation the meridional air temperature gradient was slightly reduced,decreasing the air temperature at the equator by 1◦C while increasing thetemperature at the highest latitudes by 2.5 ◦C (which keeps the global meanair temperature approximately unchanged). The air temperature differencebetween the present and LGM simulations, however, is identical to that inthe simulations discussed in this article. The results suggest that the mainconclusions are not sensitive to the details of the thermal boundary condi-tions (Fig. S4).

Sea-ice dynamics and thermodynamics are described using the MITgcmsea-ice package, which is based on the viscous-plastic rheology introducedby Hibler (37) and modified by Zhang and Hibler (38) and Losch et al. (39).Sea-ice thermodynamics are based on Zhang et al. (40) and use a zero-heat–capacity approximation, but account for the different heat conduc-tivities of ice and snow cover. Mesoscale eddy fluxes are parameterizedusing the Gent and McWilliams (GM) (41) and Redi (42) parameterizations,with a variable eddy diffusivity formulated following Visbeck et al. (43),with α= 0.01 and mixing length l = 160 km. The eddy diffusivity is cappedbetween Kmin = 200 m2·s−1 and Kmax = 2,000 m2·s−1, and the GM param-eterization is tapered in the presence of steep isopycnal slopes, followingGerdes et al. (44). Diapycnal mixing is represented by a vertical diffusiv-ity, which is strongly enhanced in the abyss, but reduces to a backgroundvalue of 2×10−5 m2·s−1 in the thermocline (45, 46). The diffusivity profileis shown in Fig. S5. Convection is parameterized using a diffusive adjust-ment with a convective diffusivity κconv = 10 m2·s−1 whenever the stratifi-cation is statically unstable. Dissipation of momentum in the bottom bound-ary layer is represented by a linear bottom drag, with a piston velocity of1 mm·s−1. No-slip horizontal boundary conditions are applied at the walls,and grid-scale noise is suppressed using Laplacian and biharmonic viscosi-ties, with viscosity coefficients ν2 = 2×104 m2·s−1 and ν4 = 2×1013 m4·s−1,respectively.

All simulations are integrated to equilibrium for at least 4,000 y, and aver-ages are taken over the last 500 y of each simulation. Whereas the simula-tion with present-day forcing reaches a steady equilibrium solution, the coldLGM simulation exhibits some decadal to centennial variability in the AMOC.The exact nature of this variability, which appears to be associated withsea ice in the North Atlantic, is beyond the scope of this study, but may beaddressed in follow-up work. The magnitude of the variability is small com-pared with the differences obtained between the present and LGM simula-tions and is averaged out in Figs. 1–3. Nevertheless, the lack of a steady-statesolution puts into question the use of an accelerated tracer time stepping(47), which has been used to speed up the equilibration. To test the impactof the accelerated tracer time step, the LGM simulation was initialized fromthe statistical equilibrium solution obtained with tracer acceleration andintegrated for another 1,000 y without tracer acceleration. Whereas prop-erties of the internal variability appear to be affected by the tracer accel-eration, the mean state and circulation discussed in this paper are virtuallyunaffected. The sensitivity to tracer acceleration was also tested explicitly inthe present seas simulation, which includes a seasonal cycle in the thermalforcing. Again, no significant effect on the deep ocean circulation and strat-ification (or on the seasonality of sea ice) was found. Due to the negligibleeffect of tracer acceleration on the mean state and circulation, all sensitiv-ity experiments discussed in this study use tracer acceleration for the entireintegration.

Computation of Buoyancy Loss Rates. Buoyancy loss from the ocean is com-puted based on the net heat and freshwater fluxes as

B = gαQρ−10 C−1

p + gβFSρ−10 , [1]

whereQ is the net heat flux and F the net freshwater flux out of the ocean(which include the effects of sea-ice formation and melt), α is the thermalexpansion coefficient, β is the haline contraction coefficient, g is the grav-itational acceleration, ρ0 = 1,035 kg·m−3 is a reference density, S is thesalinity, and Cp is the heat capacity of seawater. Because we are interestedin the effect of surface fluxes on the density of a parcel at depth, the thermalexpansion and haline contraction coefficients are computed for an ambientpressure of 2,000 dbar.B is generally positive (denoting ocean buoyancy loss) within a strip

around Antarctica, but negative farther northward in the circumpolar chan-nel. The “total buoyancy loss rate around Antarctica” is computed byintegrating B over the entire region of buoyancy loss around Antarctica,defined to include every gridpoint south of 55◦S where B> 0 (because the

Jansen PNAS | January 3, 2017 | vol. 114 | no. 1 | 49

areas of buoyancy loss and gain are well separated, the results are not sen-sitive to the exact choice of this latitude).

ACKNOWLEDGMENTS. I thank Alice Marzocchi, K. Thomas, and two anony-mous reviewers for their very valuable comments. The MITgcm config-

uration files and model output data are available from the authorupon request. Funding for this work was provided by the National Sci-ence Foundation under Award 1536454, and computational resourceswere provided by the Research Computing Center at the University ofChicago.

1. Lumpkin R, Speer K (2007) Global ocean meridional overturning. J Phys Oceanogr37:2550–2562.

2. Talley LD (2013) Closure of the global overturning circulation through the Indian,Pacific and Southern Oceans: Schematics and transports. Oceanography 26:80–97.

3. Curry WB, Oppo DW (2005) Glacial water mass geometry and the distribution of δ13Cof

∑CO2 in the western Atlantic Ocean. Paleoceanography 20(1):PA1017.

4. Lynch-Stieglitz J, et al. (2007) Atlantic meridional overturning circulation during thelast glacial maximum. Science 316:66–69.

5. Burke A, et al. (2015) The glacial mid-depth radiocarbon bulge and its implicationsfor the overturning circulation. Paleoceanography 30(7):1021–1039.

6. Lund DC, Adkins JF, Ferrari R (2011) Abyssal Atlantic circulation during the LastGlacial Maximum: Constraining the ratio between transport and vertical mixing. Pale-oceanography 26:PA1213.

7. Shin SI, Liu Z, Otto-Bliesner BL, Kutzbach JE, Vavrus SJ (2003) Southern Ocean sea-ice control of the glacial North Atlantic thermohaline circulation. Geophys Res Lett30(2):1096.

8. Liu Z, Shin SI, Webb RS, Lewis W, Otto-Bliesner BL (2005) Atmospheric CO2 forcing onglacial thermohaline circulation and climate. Geophys Res Lett 32(2).

9. Ferrari R, et al. (2014) Antarctic sea ice control on ocean circulation in present andglacial climates. Proc Natl Acad Sci USA 111(24):8753–8758.

10. Sun S, Eisenman I, Stewart AL (2016) The influence of southern ocean surface buoy-ancy forcing on glacial-interglacial changes in the global deep ocean stratification.Geophys Res Lett 43(15):8124–8132.

11. Jansen M, Nadeau LP (2016) The effect of Southern Ocean surface buoyancy loss onthe deep ocean circulation and stratification. J Phys Oceanogr 46(11):3455–3470.

12. Adkins JF, McIntyre K, Schrag DP (2002) The salinity, temperature and δ18O contentof the glacial deep ocean. Science 298:1769–1773.

13. Roberts J, et al. (2016) Evolution of South Atlantic density and chemical stratificationacross the last deglaciation. Proc Natl Acad Sci USA 113(3):514–519.

14. Schmittner A, et al. (2011) Climate sensitivity estimated from temperature reconstruc-tions of the Last Glacial Maximum. Science 334(6061):1385–1388.

15. Annan JD, Hargreaves JC (2013) A new global reconstruction of temperature changesat the Last Glacial Maximum. Clim Past 9(1):367–376.

16. Killworth PD (1977) Mixing of the Weddell Sea continental slope. Deep Sea Res24(5):427–448.

17. Toggweiler J, Russell JL, Carson S (2006) Midlatitude westerlies, atmospheric CO2, andclimate change during the ice ages. Paleoceanography 21(2):PA2005.

18. Kohfeld KE, et al. (2013) Southern Hemisphere westerly wind changes during the LastGlacial Maximum: Paleo-data synthesis. Quat Sci Rev 68:76–95.

19. Sime LC, et al. (2013) Southern Hemisphere westerly wind changes during the LastGlacial Maximum: Model-data comparison. Quat Sci Rev 64:104–120.

20. Egbert GD, Ray RD, Bills BG (2004) Numerical modeling of the global semidiurnal tidein the present day and in the last glacial maximum. J Geophys Res Oceans 109:C03003.

21. Schmittner A, Green JAM, Wilmes SB (2015) Glacial ocean overturning inten-sified by tidal mixing in a global circulation model. Geophys Res Lett 42(10):4014–4022.

22. Wunsch C, Ferrari R (2004) Vertical mixing, energy, and the general circulation of theoceans. Annu Rev Fluid Mech 36:281–314.

23. Otto-Bliesner BL, et al. (2006) Last glacial maximum and Holocene climate in CCSM3.J Clim 19(11):2526–2544.

24. Brandefelt J, Otto-Bliesner BL (2009) Equilibration and variability in a Last GlacialMaximum climate simulation with CCSM3. Geophys Res Lett 36:L19712.

25. Weber SL, et al. (2007) The modern and glacial overturning circulation in the Atlanticocean in PMIP coupled model simulations. Clim Past 3(1):51–64.

26. Muglia J, Schmittner A (2015) Glacial Atlantic overturning increased by wind stress inclimate models. Geophys Res Lett 42(22):9862–9868.

27. Stouffer R, Manabe S (2003) Equilibrium response of thermohaline circulation to largechanges in atmospheric CO2 concentration. Clim Dynam 20(7-8):759–773.

28. Zhang X, Lohmann G, Knorr G, Xu X (2013) Different ocean states and transient char-acteristics in Last Glacial Maximum simulations and implications for deglaciation. ClimPast 9:2319–2333.

29. Roche DM, Crosta X, Renssen H (2012) Evaluating Southern Ocean response for theLast Glacial Maximum and pre-industrial climates: PMIP-2 models and data evidence.Quat Sci Rev 56:99–106.

30. Klockmann M, Mikolajewicz U, Marotzke J (2016) The effect of greenhouse gas con-centrations and ice sheets on the glacial AMOC in a coupled climate model. Clim Past,1829–1846.

31. Brovkin V, Ganopolski A, Archer D, Rahmstorf S (2007) Lowering of glacial atmo-spheric CO2 in response to changes in oceanic circulation and marine biogeochem-istry. Paleoceanography 22:PA4202.

32. Watson AJ, Vallis GK, Nikurashin M (2015) Southern Ocean buoyancy forcing of oceanventilation and glacial atmospheric CO2. Nat Geosci 8(11):861–864.

33. Marshall J, Hill C, Perelman L, Adcroft A (1997) Hydrostatic, quasi-hydrostatic, andnonhydrostatic ocean modeling. J Geophys Res 102:5753–5766.

34. Large WG, Pond S (1981) Open ocean momentum flux measurements in moderate tostrong winds. J Phys Oceanogr 11(3):324–336.

35. Large WG, Pond S (1982) Sensible and latent heat flux measurements over the ocean.J Phys Oceanogr 12(5):464–482.

36. Haney RL (1971) Surface thermal boundary condition for ocean circulation models.J Phys Oceanogr 1:241–248.

37. Hibler IIIWD (1979) A dynamic thermodynamic sea ice model. J Phys Oceanogr 9(4):815–846.

38. Zhang J, Hibler WD (1997) On an efficient numerical method for modeling sea icedynamics. J Geophys Res Oceans 102(C4):8691–8702.

39. Losch M, Menemenlis D, Campin JM, Heimbach P, Hill C (2010) On the formulation ofsea-ice models. Part 1: Effects of different solver implementations and parameteriza-tions. Ocean Model 33(1):129–144.

40. Zhang J, Hibler IIIWD, Steele M, Rothrock DA (1998) Arctic ice-ocean modeling withand without climate restoring. J Phys Oceanogr 28(2):191–217.

41. Gent PR, McWilliams JC (1990) Isopycnal mixing in ocean circulation models. J PhysOceanogr 20:150–155.

42. Redi MH (1982) Oceanic isopycnal mixing by coordinate rotation. J Phys Oceanogr12:1154–1158.

43. Visbeck M, Marshall J, Haine T, Spall M (1997) Specification of eddy transfer coeffi-cients in coarse-resolution ocean circulation models. J Phys Oceanogr 27:381–402.

44. Gerdes R, Koberle C, Willebrand J (1991) The influence of numerical advectionschemes on the results of ocean general circulation models. Clim Dynam 5(4):211–226.

45. Nikurashin M, Ferrari R (2013) Overturning circulation driven by breaking internalwaves in the deep ocean. Geophys Res Lett 40:3133–3137.

46. Waterhouse AF, et al. (2014) Global patterns of diapycnal mixing from measurementsof the turbulent dissipation rate. J Phys Oceanogr 44(7):1854–1872.

47. Bryan K (1984) Accelerating the convergence to equilibrium of ocean-climate models.J Phys Oceanogr 14(4):666–673.

50 | www.pnas.org/cgi/doi/10.1073/pnas.1610438113 Jansen