We use a published experiment protocol in which global mean surface air temperatures (GSAT) are increased through emission of CO2 into the model atmosphere at the rate required by a given ESM (based on each model's transient climate response to cumulative emissions of CO2) to achieve a fixed rate of 2 °C per century. At specified global warming levels (GWLs), CO2 emissions are set to zero, with a view to stabilising global warming. Fifty years into each zero-emission run (at different GWLs) CO2 emissions become negative and the climate is cooled back to preindustrial temperatures. This idealised framework allows Earth system changes to be expressed as a response to a simple global forcing. The protocol can be readily used across different ESMs, with a common global warming rate despite model-dependence in climate sensitivity to CO2 and zero and negative emission runs branched off at the same GWLs across models. This experiment protocol forms the TIPMIP ESM Tier 1 experiment protocol . It is important to note that changes in GSAT are realised via idealised CO2 emissions and sequestration alone. Neither aerosol forcing, trace gas emissions, or any other greenhouse gas or anthropogenic boundary condition change form part of this idealised framework. In this paper we deviate from the formal TIPMIP naming convention when referring to our experiments on grounds of readability; a mapping between the names used here and the formal TIPMIP ESM designations can be found in Table .

The control simulation represents an indefinite preindustrial climate, with no anthropogenic CO2 emission (experiment “PI”). Specifications for boundary conditions come from the CMIP6 piControl experiment . The initial conditions for this simulation are derived from the UKESM1.1 concentration-driven preindustrial simulation which has spun up long enough for global mean surface air temperatures to be stable, and for fluxes into and out of the model carbon sources and sinks to be in balance, so global mean atmospheric CO2 does not drift when allowed to evolve prognostically (Fig. b, c).

In UKESM we use a constant emission of 8 Gt C yr⁻¹ CO2 to realise a GSAT increase of ∼2 °C per century from the PI baseline (experiment “Up8”). This warming rate of is roughly equivalent to the rate of warming seen in both the CMIP6 1pctCO2 experiment and observations of the global warming rate over the past half-century . The emissions have the same spatial pattern as used in the final year of the CMIP6 CO2 Historical experiment , scaled to give a constant emission rate of 8 Gt C yr⁻¹.

As the Up8 simulation proceeds, when GSAT anomalies with respect to PI reach certain thresholds, our simulations attempt to stabilise at that GWL by cutting the CO2 emission rate to zero again (experiments “ZE-x”, for x=[1.5,2,3,4,5] °C). Net zero CO2 emissions are not guaranteed to produce zero temperature trends in an Earth system simulation of this sort, as ongoing removal of excess CO2 from the atmosphere by the land and ocean carbon sinks provides a cooling tendency whilst opposing factors such as the adjustment of the ocean heat budget tend to lead to surface air warming . Initial testing with UKESM suggested that in this idealised framework these two tendencies largely cancel out, such that zero CO2 emission produces a reasonably stable GSAT (at least for GWLs at or below 2 °C, Fig. c).

To represent possible overshoot pathways of GSAT in our protocol and investigate the reversibility of simulated changes, other simulations allow the climate to adjust to zero emissions at a GWL for some time, before reducing GSAT again by removing CO2 from the atmosphere. We do this by reversing the sign of CO2 emissions in our system, simulating sequestration of CO2 from the lower atmosphere, using the same spatial pattern for emissions used in the ramp-up. The simulations described in this paper come from a larger UKESM ensemble which follows and extends the TIPMIP ESM protocol , sampling a wider range of GWLs, zero-emission periods and warming and cooling rates. In the interests of brevity, in this overview paper our results are drawn from one member of each warming, zero-emission and cooling phase for each GWL. In the members described, cooling simulations are branched after 50 years of zero CO2 emissions, using a negative emission rate of -4 Gt C yr⁻¹ (experiments “Dn4-x”, for each GWL x), half of the rate used during the ramp-up. The zero-emission simulations are carried on for a total of 500 years. We chose to describe these members for practical reasons, since their results were available first from our wider body of simulations. The evolution of these members can be taken as generally representative of the major features of all related variants at each GWL in the full UKESM ensemble. The full ensemble behaviour will be extensively analysed in subsequent papers. Although we intended for the cooling runs to all continue until negative GSAT anomalies were achieved, the length of simulations required to realise this from the higher GWLs and decommissioning of the computing facility used for this study means that we were not able to meet this criterion for all Dn4-x simulations.

We use UKESM1.2, a version formed by combining UKESM1.1 as described in and , with the two-way coupled Greenland and Antarctic ice sheet components described for variants of UKESM1 in and . UKESM1.2 uses the N96 resolution (1.875° longitude 1.25° latitude) of the Unified Model atmosphere component version 3.1 with 85 vertical levels extending to the mesosphere. For the ocean, we use the G07 version of NEMO on an eORCA1 (equivalent 1° resolution) ocean grid with 75 vertical levels, including explicit modelling of the ocean and melt under floating Antarctic ice shelves. The land surface and terrestrial carbon cycle is modelled by JULES using the same horizontal grid as the atmosphere model and 13 plant functional types. Our configuration of JULES includes an online downscaling of ice sheet surface exchange based on sub-gridscale elevation classes and a multilayer snow model that allows for percolation and refreezing of liquid water within the snowpack. The distribution of vegetation is interactively calculated by the TRIFFID model . We use a complex, interactive atmospheric chemistry and aerosol scheme provided by UKCA and ocean biogeochemistry from MEDUSA which uses five marine functional types. The Greenland and Antarctic ice sheets are simulated by BISICLES , a vertically integrated model of ice-sheet dynamics which uses an L1L2 solver with an adaptive mesh that can refine to a horizontal resolution of 1 km to accurately model processes such as grounding line movement. We do not use the thermodynamics component of BISICLES in these simulations so the ice sheets maintain constant internal temperature and effective viscosity fields throughout. We refer readers to the listed references for a complete description of each of these components and how they are coupled together.

This ensemble uses a baseline simulation with constant preindustrial forcing of indefinite duration, similar to that defined by the CMIP6 protocol, from which the Up8 simulation branches. For an ESM with interactive ice sheets, this concept raises two issues. Firstly, there is insufficient evidence of what the preindustrial states of the Greenland and Antarctic ice sheets (GrIS, AIS) actually were, importantly including a lack of information on the position of the grounding lines of key outlet glaciers. The sensitivity of the ice sheet response to changes in boundary conditions is known to depend on the details of the initial state . Moreover, the concept of the CMIP6 preindustrial control simulation assumes that the key metrics of the Earth system in which we are interested are in some quasi-stable equilibrium with the preindustrial forcing. In reality, the GrIS and AIS in the preindustrial would have been transiently evolving away from their Last Glacial Maximum states at some rate. Within our imperfect modelling systems, it is impossible to realise a preindustrial simulation with indefinitely stable ice sheet states that are simultaneously realistic and in balance with the biases inherent to a given ESM.

Developing best-practice approaches to the problem of ice sheet initialisation in global ESM simulations is an area of active research that we do not attempt to address here. In the simulations described in this paper, as in previous UKESM simulations with interactive ice sheets, we use the initialisations of the GrIS and AIS developed to represent the states of the ice sheets at the beginning of the 21st century . This was felt to be the simplest and cleanest approach at the time of initialising our ensemble. Preliminary attempts to grow our simulated ice sheets from their modern states to expanded states that might be more appropriate for the preindustrial and use these to initialise 20th century simulations have so far resulted in biases in ice thickness and velocity that are worse than those that arise from simply starting from the modern states. During the baseline PI simulation we do not allow the ice sheet geometry to evolve away from those modern states, although we do interactively compute the surface mass balance (SMB) and basal mass balance (BMB) at the ice sheet surfaces that result from their presence in the preindustrial climate and pass the resultant freshwater fluxes to the ocean. When the Up8 simulation is branched from PI the ice sheet geometries are then allowed to start evolving, and this evolution then continues through all subsequent ZE-x and Dn4-x branches.

One of the key issues with using a modern initialisation of the ice sheets when starting our simulations from preindustrial conditions is that the modern BISICLES AIS initialisation includes grounding lines for key ice streams in the West AIS that are already retreating (e.g. Pine Island Glacier and Thwaites Glacier). This retreat is not halted by applying UKESM preindustrial boundary conditions to the ice sheet, so that our realisation of this sector of the ice sheet is effectively already past a tipping point with respect to the UKESM climatic forcing. As a result, every AIS simulation in our ensemble has a historically recent rate of ice loss built into it from the beginning, a factor that we take into consideration in our analysis. To enable this analysis, we conduct an “ice control” simulation that allows the ice sheet geometry to evolve under continued preindustrial boundary conditions to isolate the unforced behaviour of our AIS component (experiment “ZE-0”).

Although the focus of this paper is on polar climate and ice sheets, for context we first present the basic global surface temperature responses of UKESM under the TIPMIP ESM protocol.

The accumulating burden of emitted CO2 in UKESM illustrates the positive emission (warming), zero-emission (stabilisation) and negative emission (cooling) phases of the forcing protocol we are using (Fig. a). The concentration of CO2 in the atmosphere increases in Up8 and then slowly drops in the ZE-x simulations when emissions cease as the excess CO2 is taken up by the land and ocean, with the rate of uptake being slightly higher for experiments that have stabilised at higher GWLs (i.e. with greater emitted CO2, Fig. b). Sequestration of CO2 adds strongly to the effect of the natural carbon sinks and CO2 concentrations drop rapidly in the Dn4-x simulations.

Global mean air surface temperature broadly follows the atmospheric CO2 burden (Fig. c). Despite the slow fall in atmospheric CO2 in the ZE-x simulations, surface temperature trends are near zero for GWLs at or below 2 °C, and slightly positive for higher GWLs. This comes about from a balance between the surface warming trend associated with heat sequestered into the deep ocean reemerging at the surface as the climate comes into equilibrium, and cooling from the slow decrease in atmospheric CO2 in ZE-x. Surface temperatures generally cool in line with the rate of CO2 removal in the Dn4-x simulations as radiative forcing from excess CO2 in the atmosphere is rapidly reduced. At GWLs less than 4 °C there is a consistent relationship between GSAT and the accumulated amount of CO2 in the system, regardless of whether CO2 is being emitted or sequestered (Fig. d). In the cooling Dn4-4 and Dn4-5 GSAT is slightly higher for any given amount of accumulated CO2 than it was for that burden as Up8 warmed, reflecting warming in the ocean that cannot be as rapidly lost when atmospheric CO2 is reduced.

Surface air temperature (SAT) change at high latitudes is usually greater than the global mean due to several polar amplification mechanisms, especially in the Northern Hemisphere . In our simulations Arctic SAT is a factor of two to three times higher than the global average (Fig. a, c). The amplification factor is highest for zero-emission runs at lower GWL simulations, likely because at higher GWLs there is less Arctic sea-ice and high latitude snow available to drive a positive ice albedo feedback and insulate the atmosphere from the warmer ocean. For the highest GWLs the amplification factor declines with time because GSAT slowly increases in these runs whilst Arctic SAT itself remains constant. The warming trend seen in GSAT in ZE-x simulations at higher GWLs is largely driven by warming at high southern latitudes (Fig. b, d), which have a lower amplification factor than the Northern Hemisphere SAT. This warming, which is especially marked at higher GWLs, is due to the reemergence in the Southern Ocean upwelling zones of heat that has been taken up by the ocean during the warming phase of the protocol and sequestered below the thermocline. The reemergence of this heat at the surface additionally triggers positive feedbacks in Antarctic sea-ice melt and reduces local cloud fraction which both act to amplify the surface heating.

There is a large multi-decadal warming in Southern Hemisphere high latitude SAT of around 2 °C in the Dn4-3 simulation around when temperatures have returned to PI levels (Fig. b). This appears to be due to an episode of rapid sea-ice loss and surface warming in the Ross and Amundsen Seas, likely caused by a release of ocean heat that has built up below the surface during the warm phases of this simulation. Although the PI simulation shows significant unforced temporal variability of similar magnitude, the event in Dn4-3 is clearly different, featuring a much more widespread loss of sea-ice occurring at a global atmospheric CO₂ concentration lower than that of the PI.

The polar amplification of GSAT anomalies is not zonally uniform with latitude, and the factor by which warming is increased is significantly reduced over the GrIS and AIS themselves. This is common to previous studies, which show that it is related to a number of factors, including the height and geographical position of the ice sheets affecting heat transport, surface radiative feedbacks and the constraint of the melting temperature of the ice surface . These factors especially limit temperature change over Antarctica, resulting in initial polar amplification factors less than one for high southern latitudes (Fig. d).

During Up8, Arctic sea-ice area declines approximately linearly with global temperature change (Fig. ), as reported in previous modelling studies (e.g. ). As GSAT increases, Arctic SAT increases disproportionately (as discussed in Sect. ), this exacerbates sea-ice melt. Given that in our simulations GSAT is approximately linearly correlated with CO2 emissions, our results are also consistent with the observed linear relationship between cumulative anthropogenic CO2 emissions and Arctic sea-ice decline .

The ZE-x simulations generally show a rapid stabilisation of Arctic sea-ice once emissions are halted (Fig. a, b). This is in contrast to previous studies in which sea-ice area continues to decline during stabilisation runs following the emission phase . In ZE-5, the highest GWL in our ensemble, the sea-ice appears to recover by approximately 2.5 million km² before sequestration starts, suggesting that there are sea-ice restorative feedbacks in response to large-scale melt or global temperature increase. This reversibility of sea-ice decline is consistent with which finds that Arctic sea-ice recovers from dramatic perturbations imposed in the form of a summer sea-ice free state.

The Dn4-x simulations highlight the reversibility of Arctic sea-ice with global mean temperature (Fig. b) and cumulative emissions. Ramp-down simulated sea-ice area tends to follow a reversed trajectory to that of the ramp-up simulations for all GWLs.

There is a clear forced signal of Greenland mass loss in the zero-emission runs at all GWLs in the ensemble, compared to the background drift of this ice sheet under constant preindustrial forcing in ZE-0 (Fig. ). As CO2 concentrations increase, so does the rate of mass loss, with ZE-x mass change curves appearing to split from the Up8 curve as local tangents at the time that emissions are zeroed (Fig. a). Stabilising GSAT by zeroing CO2 emissions does not stop mass loss from the GrIS; rather, it simply holds the rate of mass loss fixed at that level. At each GWL there is an approximately linear rate of ice loss that is set by the value of the rate when emissions cease. At a GWL of 1.5 °C this amounts to a 0.2 mm yr⁻¹ sea level equivalent volume (SLE) loss, whilst at 5 °C the GrIS loses 2 mm yr⁻¹ SLE. Like some other variables (e.g. ocean heat content, see Sect. ) there is a considerable time lag between the start of CO2 removal and a change in the sign of the mass trend for the GrIS. In Dn4-1.5, the forced GrIS contribution to global mean sea level (GMSL) only stops after approximately 100 years of negative emissions, whilst for Dn4-3 it takes approximately 200 years of negative emissions, with even longer stabilisation timescales for the GMSL contribution at higher GWLs. In our ensemble, however, there is no clear evidence of abrupt acceleration of mass loss from the GrIS as a whole, or that any state is reached from which cooling of the climate does not initiate the reversal of GrIS change. It should be noted that the centennial timescales of our simulations are shorter than those of previous studies in which irreversible loss of the GrIS was assessed (e.g. ).

The evolution of the mass of the GrIS over the next hundred years or so is expected to be largely controlled by the surface mass balance (SMB, i.e. the difference between surface accumulation and ablation) it experiences rather than dynamical change . There is a clear correlation across our ensemble between the rate of GrIS mass loss and the SMB, supporting this expectation. SMB becomes increasingly negative as GSAT increases and settles to a constant rate when GSAT stabilises, with an apparently similar form to the relationship derived by . Increases in air temperature are associated with greater precipitation which can offset some of the loss of mass from increased melt, but melting increases more quickly with temperature than precipitation and at higher GWLs a significant proportion of the precipitation over the ice sheet falls as rain which instantly runs off in our model. SMB itself responds rapidly to reductions in GSAT in the Dn4-x simulations with the trajectory of GrIS SMB simply reversing, even when reducing from the highest GWLs (Fig. b).

GSAT needs to be brought back close to PI values before SMB becomes net positive again, and the ice sheet has the possibility of regaining mass, dependent on its rate of dynamic discharge which does not vary much on these timescales in UKESM. At PI GSAT, individual components of SMB after the warming and cooling are not significantly different from those in the unperturbed PI simulation. When the climate is cooled further the rate of regrowth is slow; the gradient of the relationship between SMB and GSAT is small for our negative GSAT anomalies. In this regime, further cooling to a GWL of -1 °C reduces runoff from the ice sheet slightly whilst the snowfall rate is constant.

The GrIS dynamically responds to this change in surface forcing in UKESM with a characteristic pattern. Ablation at the low-lying, warm edges of the ice sheet thins and slows the ice, which both act to reduce the discharge of solid ice at much of the boundary, especially in the west and south. In UKESM1.2, GrIS marine-terminating glaciers do not directly feel the evolving ocean boundary conditions because the necessary fjord regions cannot be resolved by the ocean model, so we do not expect to see the acceleration of these glaciers or the associated increase in discharge in those regions in our simulations. This is a limitation of the model because, in reality, ocean-forced changes in ice discharge are responsible for a significant amount of the current observed ice loss , although in the future ocean-forced mass loss is expected to be small in comparison to SMB .

The Up8 simulation shows an approximately linear decline of Antarctic sea-ice with increasing GSAT (Fig. ), similar to that of the Arctic. In our zero-emission runs, however, with the exception of the ZE-1.5, Antarctic sea-ice continues to decline despite the stabilisation of global mean temperatures. This is in contrast to the Arctic, a regional disparity that is also reported by . In the Antarctic, we find that the rate of sea-ice decline per degree of global mean surface temperature warming is increased during zero-emission runs when compared to the Up8 simulation, as reported by . This indicates that lags in the model Earth system, for example in local surface air temperature, continue to cause sea-ice melt in the stabilisation simulations beyond that caused by the increase of GSAT.

The Dn4-x simulations highlight the reversibility of Arctic sea-ice change with respect to GSAT and cumulative emissions (Fig. ), but the Antarctic response differs, displaying a significant lag in the relationship (Fig. ). Antarctic sea-ice continues to decrease in the ZE-x runs, and when plotted against GSAT the sea-ice in Dn4-x simulations does not recover along the same curve as Up8 since it starts to recover from a lower ice extent (especially above the 3 °C GWL), likely due to the accumulation of sequestered heat in the Southern Ocean. suggested that Antarctic sea-ice change is reversible on longer timescales, with sea-ice lagging surface temperature change by several centuries. This appears to be consistent with our simulations, in which sea-ice only begins to recover in the Dn4-4 and Dn4-5 simulations after 150–200 years.

One of the major components of the climate system at high southern latitudes is the belt of mid-latitude westerlies that encircles the Antarctic continent, referred to as the “westerly jet”. It interacts strongly with not only the Southern Ocean, but also parts of the Antarctic continent as its poleward flank extends over important coastal zones . The broad behaviour of the Southern Hemisphere westerly jet can be captured using two indices based on lower-tropospheric zonal mean westerly wind: jet latitude index (JLI) and jet speed index (JSI) as described in .

There is a general poleward shift of the westerly jet during Up8, which is maintained during the 50-year zero emission periods, followed by a return towards lower latitudes during CO2 removal phases (Fig. a). The simulations exhibit broad reversibility, which is apparent across the range of GWLs shown.

The jet speed strengthens as global temperatures rise during Up8 (Fig. b), but this strengthening reduces as GSAT increases beyond 4 °C. During the periods of zero emissions these enhanced speeds are not in general maintained. Instead, the jet gradually weakens, particularly at the higher GWLs. During global cooling the jet weakens further and at high GWLs does not recover along the same curve, particularly during Dn4-5. In this case the jet returns to a weaker state than it was initially, which is consistent with the weakened storm track found after CO2 removal by .

This behaviour of jet diagnostics likely results from competing processes, including low-latitude upper-tropospheric warming, which tends to strengthen the jet and shift it poleward, and high-latitude lower-tropospheric warming, which tends to weaken the jet and induce an equatorward shift , as well as interactions with changes in sea-ice area which is shown to also exhibit variations over the ZE-x runs (Fig. ). As suggested by , a weakening jet during stabilisation at a GWL is consistent with continued SH sea-ice decreases in the ZE-x runs. The more prominent lack of simple reversibility in JSI is consistent with the closer link between JSI and sea-ice extent found by .

In all the simulations presented here, the AIS loses mass, providing an increasing flux of freshwater into the Southern Ocean (Fig. a). The magnitude of that flux is dependent on the climate forcing, particularly at higher GWLs. However, the majority of this mass loss comes from the basal melting of floating ice shelves, and this part of the mass loss does not directly contribute to changes in GMSL. To assess the impact on GMSL, we need to examine the change in mass above flotation (Fig. b).

Across the range of GWLs, and on the centennial timescales explored in these simulations, the overall change in the ice mass above flotation of the AIS is a result of the interplay between the ice dynamical response arising from the initial condition (as discussed in Sect. ); changes in the SMB, including snowfall and surface melt; and changes in ice discharge and grounding line position in response to changes in the basal mass balance (BMB) of the floating ice shelves, melted from beneath by the ocean. Over the first 550 years of these simulations, the ZE-0 baseline simulation contributes ∼280 mm to GMSL, despite the absence of external climate forcing. This is the result of the unsteady initial condition that leads to mass loss, predominantly from Thwaites Glacier, combined with the ice sheet adjusting to the preindustrial climate state in UKESM when it is coupled. All other simulations presented here contain the imprint of this background ice dynamic behaviour in this sector of the ice sheet for the duration of the simulation. This background mass loss does not cause significant drift in any of the climate features analysed elsewhere in this paper. The ZE-0 simulation has the greatest loss of AIS mass above flotation in our ensemble, and therefore the largest contribution to GMSL (∼280 mm after 550 years). This counter-intuitive result occurs for two reasons: (1) the ice initialisation response, and (2) increased snowfall in a warmer climate. That is, all scenarios lose mass in response to the initial conditions, predominantly from Thwaites Glacier, but the greenhouse gas forcing partially offsets this mass loss by causing increased accumulation of snowfall (positive SMB anomaly) over the grounded regions of the ice sheet. In one simulation (ZE-3), the net GMSL contribution from Antarctica is briefly negative (-2 mm after 550 years), as the cumulative increase in SMB temporarily offsets the dynamic ice loss.

The SMB of the AIS is likely to change in a future, warmer climate, with the primary response on centennial timescales being increased levels of precipitation (in the form of snowfall) from the accelerated hydrological cycle (e.g. ). However, quantitative model estimates of the present-day SMB, and projections of future SMB, vary significantly . UKESM has previously been shown to project a large increase in Antarctic snowfall compared to other models, which becomes a dominant factor in its simulation of forced change in Antarctic grounded mass balance over the 21st century . As with Greenland, our simulations show that SMB changes on the grounded portion of the AIS closely track changes in the local air temperature, and are reversible on similar timescales: grounded SMB stabilises as soon as local temperatures stabilise, and start to fall once negative emissions cause local air temperatures to decrease. However, due to the continued warming of the Antarctic in the ZE-3, ZE-4 and ZE-5 simulations (Fig. b), the grounded SMB continues to increase (and the floating SMB to decrease) even under zero emissions. This can be seen by the larger spread in both floating and grounded SMB for ZE-3 and ZE-4 in Figs. c and d. It is the local SAT that influences the SMB response.

It is worth noting that the SMB response to increasing atmospheric temperature can be split into two types of behaviour : a tendency to decrease in low-lying coastal areas (e.g. ice shelves) and to increase on higher ground further inland. Integrated over the grounded ice sheet, SMB simply increases with local SAT due to the dominance of the precipitation component, as the seasonal runoff remains small (Fig. c). Quantitatively, the grounded SMB in the ZE-5 simulation increases by ∼40 % compared to the ZE-0 experiment, and the reduction in grounded SMB in the ramp-down experiments (e.g. in Dn4-4) demonstrates the reversibility of this change as surface temperatures are reduced.

On the floating ice shelves, however (Fig. d), surface melting begins to play an important role as GWLs, and local air temperatures, increase. At first, this increased melt cancels the impact of the increased precipitation, and then in the ZE-3 to ZE-5 simulations, melt starts to reduce the net SMB over the ice shelves. At the highest GWLs explored, ZE-4 and ZE-5, the SMB integrated over the ice shelves becomes negative, and they experience widespread ablation from both the air above and the ocean below. Where significant surface melting occurs, there is the possibility of hydrofracturing and mechanical breakup of the shelf, depending on the stress regime in the ice . This effect is not included in our ice sheet model.

The other notable features of the mass loss curves in Fig. a are the abrupt changes in the gradients due to changes in the ice dynamics. After ∼475 years in this set of simulations, there is an acceleration in the loss of mass above flotation that is synchronous across simulations, i.e. independent of the climate forcing (Fig. b). This corresponds to Thwaites Glacier entering a period of accelerated thinning and grounding line retreat. The fact that this acceleration also occurs in the ZE-0 experiment suggests that this is an inherent feature of our ice sheet initialisation that takes many centuries to unfold.

Inflection points in the mass loss curves in Fig. a occur in the Up8, ZE-4 and ZE-5 experiments between 250 and 300 years into the simulations. Here, the drivers are regime shifts in the ocean conditions underneath the Ross and Filchner-Ronne ice shelves, as simulated previously by the UKESM for future climate scenarios . Warm water intrudes into the previously cold ice shelf cavities, causing a substantial increase in basal melt rates, reduced ice-shelf buttressing, accelerated upstream ice flow, and grounding-line retreat. The Up8, ZE-4 and ZE-5 experiments switch from a period of gain in net mass above flotation, to becoming net contributors to GMSL again following the transition (Fig. b).

Over the past few decades, the global ocean has stored about 89 % of the heat added to the Earth system which is closely related to the thermosteric component of GMSL. The unique capabilities of the ice sheet coupling in UKESM mean that, for the first time, these simulations allow a systematic investigation of thermosteric and ice sheet contributions to GMSL at different GWLs under zero CO2 emission and subsequent cooling.

Figure  shows the time-series of ocean heat content (OHC) evolution for the ensemble, relative to the content at the start of the PI simulation. The PI has a small (+0.12 W m⁻²) positive top of atmosphere budget which leads to a small accumulation of heat in the ocean under this baseline forcing. For the purpose of our analysis, however, it is the OHC changes arising from the time-dependent forcing that are of interest.

The ocean warms most rapidly in the Up8 simulation but, in contrast to the surface air temperatures, warming does not stop in any of the ZE simulations; rather it persists at a substantial rate and thermosteric sea level rise continues. The warming behaviour in the surface layer is qualitatively similar to that of the full ocean.

In the Dn4-x simulations, global ocean warming begins to stabilise with a significant time lag after sequestering CO2. This OHC reaction time scale increases with depth and GWL. In the deepest layer, it takes about 40 years or so for the forcing signal to reach that layer at the start of Up8. Even after 100 years into the Dn4-x simulations, layers below 2000 m continue to warm. In the top layer the reaction time scale is shorter and the ocean cooling begins more quickly after the start of CO2 sequestration, for example after only 10 years in Dn4-1.5.

Although timeseries of ice sheet mass loss in the simulations are presented in the individual sections above, it is also instructive to view all these components together in a framework that emphasises their relationship to global temperatures (Fig. ). In this presentation it is clear that sea level rise continues from all of these components under zero CO2 emissions. For the thermosteric component the rate of sea level rise starts to drop, but on the multi-centennial timescales of these simulations this component remains positive throughout (Fig. a, d). Sequestering CO2 and reducing surface air temperature does eventually produce a negative rate of thermosteric sea level rise (i.e. a contribution that acts to reduce GMSL) but this does not even begin to happen until GSAT has reduced significantly. The multi-century timescale of our simulations does not see the thermosteric sea level contribution return to original preindustrial values, even from the lowest GWL in our framework.

The contribution from Greenland is even harder to slow (Fig. b, e). Net zero CO2 emission stabilises the rate of loss of the GrIS and this rate does not decline with time as the thermosteric contribution does. Removal of CO2 from the atmosphere and cooling reduces the rate of sea level rise from GrIS, but only very small negative rates are ever achieved, even once the climate has cooled significantly below the PI baseline. Mass loss from GrIS and its contribution to sea level rise are therefore essentially irreversible in these simulations.

This presentation also highlights the fact that the GMSL contribution from Antarctica should not be viewed as a simple function of global warming level (Fig. c, f). More so than Greenland on the timescales of our simulations, the AIS contribution arises from a complex interaction of the evolving boundary SMB and BMB with the internal dynamics of the ice sheet, some of which are inherent to our ice sheet configuration and initialisation and almost unaffected by the wider Earth system. Indeed, the modern trends in mass loss from the Amundsen sector of the AIS that come with our ice sheet initialisation may be masking a lot of the dynamical response to this climate forcing that might be expected to arise in reality. These are crucial aspects to any decadal or centennial projection of ice sheet behaviour and remain hugely uncertain. Developing techniques for the initialisation of ice sheet models in coupled climate models is an active area of research.

It is, however, worth noting that one of most impactful aspects of sea level rise is the rate at which it happens, which sets the timescale on which societies must adapt to it. Overall, warming stabilisation and cooling do act to reduce the rate of sea level rise in our simulations, so even if the world is committed to an irreversible magnitude of sea level rise we can still act to control the rate at which this is realised, with potentially important benefits for adaption.