Main

Ice loss from the grounded regions of the Antarctic ice sheet is impacted by changes in the thickness and extent of the adjoining floating ice shelves, due to their capacity to buttress ice flow inland1,2. Mass loss from the Antarctic ice sheet has accelerated in recent decades3, and some of the recently observed ice loss has been related to a reduction in ice shelf buttressing caused by ocean-induced ice shelf thinning4. Current ice loss is primarily concentrated in the Amundsen Sea Embayment region of West Antarctica3,5. Uniquely, ice shelf cavities in this region are in partial contact with relatively warm deep water (for example, ref. 6), whereas many other ice shelves, in particular the large Filchner–Ronne and Ross, are in contact with comparatively colder water masses. Several previous ocean modelling studies have now shown the potential for near-future (on the order of decades) thermal regime shifts whereby intrusions of warm deep water replace the currently cold shelf waters within the Filchner–Ronne and Ross ice shelf cavities7,8,9,10,11. Resulting temperature changes would increase sub-shelf melt rates from a few metres to tens of metres per year, thereby causing ice shelf thinning, an associated loss of ice shelf buttressing and an increase in ice flow across the grounding lines. Over time this could lead to the inland migration of grounding lines and a possible onset of unstable and irreversible ice loss. However, whereas potentially profound, the implications of an immediate regime shift in ocean conditions on the dynamics and future mass loss from the Antarctic ice sheet remains unquantified.

In this Article, our goal is to quantify how a cold to warm thermal regime shift within ice shelf cavities could impact future mass balance and stability of the Antarctic ice sheet. We use a recently published ensemble of ocean-circulation model-derived melt rates (Fig. 1) from two independent modelling groups that used two different models (FESOM and NEMO; ref. 12). Both groups found a transition to warm water inflow into many ice shelf cavities around Antarctica on decadal timescales. However the precise timing of the onset of warm water intrusions varied depending on the forcing. Nonetheless, the resulting warm states are similar; all find temperatures in ice shelf cavities to increase by 2 to 4 °C and order-of-magnitude increases in sub-shelf melt rates, particularly at the large, currently cold Filchner–Ronne and Ross ice shelves.

Fig. 1: Regime shifts in ocean model melt rates.
figure 1

a–f, Perturbed sub-shelf melt rates in m yr−1 from each of the respective ocean model scenarios shown in the plot titles. Note that FESOMσ and FESOMz are configurations of the same ocean circulation model using either σ or z coordinates, respectively. Reference ocean model melt rates with respect to observations are shown in Extended Data Fig. 3. Blue, orange and red polygons denote the catchment boundaries for the Filchner–Ronne, Ross and Amundsen Sea Embayment regions, respectively. g, The total integrated ice shelf melt (gigatonnes per year) in each of our three main catchments of interest. The error bars represent the spread of melt rates based on the three melt rate parameterizations (Methods), where the centre of the bar is the middle value and the upper and lower limits are the maximum and minimum values, respectively, of the integrated melt rates calculated using the three different melt parameterizations. Note the quadratic parameterization overestimates melt in the Ross ice shelf under the NEMO4 forcing, and the error bar extends beyond the y-axis scale). (Extended Data Fig. 2).

Full size image

Here we immediately apply the warm ocean-induced melt rates—representative of plausible conditions for the late twenty-second and twenty-third centuries—to a present-day configuration of the Antarctic ice sheet in the ice sheet model Úa13 (Extended Data Fig. 1). Then we run an ensemble of forward-in-time ice sheet model experiments (Methods) to explore the timescales of the ice sheet response, immediately after this regime shift in sub-shelf melt rates has occurred. To account for uncertainties in melt rates as the geometry evolves, we use three conceptually different parameterizations for the melt rates through time (Methods and Extended Data Fig. 2). One of these is a new approach to estimating the evolution of sub-shelf melt rates that adopts a modern analogue technique previously used for palaeoclimate reconstructions (for example, ref. 14). Our simulations reveal the Filchner–Ronne and Ross catchments—that are currently in approximate mass balance5—could become large contributors to future ice loss from Antarctica within 100s of years. Under these high-melt-rate conditions, grounding lines in several catchments in West Antarctica could also enter phases of irreversible retreat, whereby reverting to current melt rate conditions is not sufficient for ice loss to be recovered.

Ocean melt rate forcing

Here we provide a brief summary of the published ocean model melt rates12 we use as forcings for our ice sheet model simulations. These comprise of six perturbed states (Fig. 1), all of which are 10-year averages from the end of 100-year idealized perturbation experiments under different climate forcing scenarios. Further details on the conditions and forcing scenarios under which these changes in thermal state occur can be found in refs. 8,11,15 and are described briefly in our Methods.

In this Article we focus only on changes in West Antarctica, particularly on the impact of regime shifts in the currently cold Filchner–Ronne and Ross ice shelf cavities. In the three high-end scenarios12 (NEMO4, \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\), \({{\rm{FESOM}}}_{z}^{{\rm{high}}}\); Fig. 1g) integrated melt rates at the Filchner–Ronne ice shelf increased to 2,000−2,700 Gt yr−1 from 71−135 Gt yr−1 in the reference configurations (Extended Data Fig. 3). In all cases melt rates are highest along the path of warm deep water inflow beneath the Filchner ice shelf (Fig. 1a–c). In the \({{\rm{FESOM}}}_{z}^{{\rm{mid}}}\) scenario melt rates increased to 1,200 Gt yr−1 and to approximately 340 Gt yr−1 in both low scenarios (Fig. 1g). At the Ross ice shelf, increases in melt rates varied between the two high-end ocean states; increasing from 58−148 Gt yr−1 in the reference simulations to 2,500 Gt yr−1 in NEMO4 or 1,300 Gt yr−1 in \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\) (Fig. 1g). In both cases high melt rates are concentrated along the ice front and western portion of the shelf (Fig. 1a,b).

Elsewhere, in the Amundsen Sea Embayment region of West Antarctica, the ice shelf cavities are currently in a warm state. The reference ocean model states all underestimate present-day conditions (Extended Data Fig. 3) due to the limited resolution and small size of the ice shelves, and accurate simulations require high-resolution regional model configurations (for example, ref. 16). Nonetheless the NEMO4 reference state is able to capture intrusions of warm circumpolar deep water onto the continental shelf11. In the perturbed NEMO4 simulation the Amundsen Sea warms by a further + 2 °C (ref. 11), and integrated melt rates doubled to 550 Gt yr−1 with respect to observations (note this is a tripling with respect to the NEMO4 reference run; Extended Data Fig. 3). In this region we therefore only focus on the NEMO4 ocean model states.

Impact of regime shift on ice loss

Our forward-in-time perturbation experiments show that an instantaneous switch from cold to warm conditions in the cavities of the Filchner–Ronne and Ross ice shelves causes them to become positive contributors to global sea-level rise (Fig. 2). This contrasts with the currently limited ice loss from these catchments, captured by our control simulations using reference ocean model melt rates (Methods and Extended Data Fig. 3). These combined contributions to sea-level rise from the Filchner–Ronne and Ross catchments after 100 years could outweigh that of the Amundsen Sea Embayment region under present-day conditions (Fig. 2).

Fig. 2: Volume changes during 100-year melt perturbation simulations for the Filchner–Ronne, Ross and Amundsen Sea Embayment catchments.
figure 2

a, Shaded red shows the total loss of ice above flotation in meters water equivalent (m w.e.) after 100 years in the NEMO4 perturbed simulation using the modern analogue technique melt (Mmat) parameterization. Coloured lines (legend in b) show the final 100-year grounding line positions for all perturbed simulations. Control simulations are in black. Line styles (shown in the legend in c) denote the melt rate parameterization (see Methods). Gl, glacier; IS, ice stream. b–d, Sea-level equivalent (SLE) volume loss (negative value indicates sea-level rise) for the Filchner–Ronne, Ross and Amundsen Sea Embayment catchments using the same line colour and styles outlined above.

Full size image

In the Filchner–Ronne region, under the high-end scenarios, ice volume loss is 64–97 mm sea-level equivalent (SLE) after 100 years (ranges provided are always those of the whole ensemble and not confidence intervals). Alongside this, in all three high-end scenarios and using all three melt parameterizations, the grounding lines in the Robin subglacial basin reach a similar location approximately 160 km inland from present day (Fig. 2a). By comparison, under control conditions and in the low-end scenarios, volume loss is limited to −2.7 to 13 mm SLE. Grounding line retreat occurs predominantly across the Filchner and eastern Ronne ice shelves (Fig. 2a), coincident with the highest melt rates downstream (Fig. 1). Directly upstream of the current grounding lines, the ice is lightly grounded (<100 m above flotation) on the underlying topography (Supplementary Fig. 12). Hence, a small migration from this position in the control and low-end scenarios does not coincide with an increase in the rate of volume loss (Fig. 2a). Migration of the grounding lines further inland in the mid–high scenarios, particularly in the Möller and Institute ice streams, is simultaneous with an increase in volume loss from these regions (Fig. 2a).

Grounding line retreat, particularly at the Möller ice stream, varies through time; retreat is limited during the first two decades after which retreat becomes more rapid up until 100 years (Fig. 3a). This is coincident with an increase in the rate of volume loss after approximately 25 years (Fig. 2b). There are two phases of accelerated retreat (grey panels Fig. 3b); an initial smaller phase between 6 and 34 km along the profile and a second larger phase immediately inland (approximately 34–150 km along the profile). Crucially, rapid retreat occurs in all simulations and is insensitive to both the choice of ocean model scenario and the melt parameterization (Fig. 3b). However, the onset of rapid retreat depends on the magnitude of the perturbation; in the low-end scenarios, the grounding line does not reach 34 km inland within 100 years. In scenarios where the grounding line reaches a topographic high point at 150–160 km along the profile, the rate of retreat decreases thereafter (Fig. 3b).

Fig. 3: Grounding line retreat and irreversibility.
figure 3

a, Grounding line positions through time at ten-year intervals during the 100-year NEMO4 perturbed simulation using the Mmat parameterization (Methods). The background is bed topography from BedMachine v229 in meters above sea level (m a.s.l.) shown on x and y coordinates in polar stereographic (ps) projection. White lines show profiles along key ice streams, and white dots are at 20-km intervals inland of the initial grounding lines. Coloured circles show the location of reversibility experiments, which correspond to the coloured circles shown in panels b–g. b–d, The grounding line positions extracted along these profiles during the forward-in-time experiments, for the Möller ice stream (b), Thwaites Glacier (c) and Bindschadler ice stream (d) profiles. Coloured lines represent the different forward perturbation experiments (legend in c). Grey shading shows phases of accelerated retreat. e–g, The grounding line positions during the reversibility experiments (the timing of which is denoted by the coloured circles), where purple and orange lines are when the melt rates were reverted to control conditions for the NEMO4 and \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\) experiments, respectively. Red lines are where the control melt rates were reduced by 50%. The line style in b–g denotes the choice of melt rate parameterization (legend in e). The shaded orange shows periods of continued retreat during the reversibility experiments. In g, the grey shaded region is a period of re-advance. Note the difference in y-axis scales for Thwaites Glacier (c,f).

Full size image

Consistent with present-day observations, the Ross catchment gains mass in our control and low-end perturbed simulations (≈4.42 ± 1.71 mm SLE after 100 years; Fig. 2c). In the two high-end scenarios (NEMO4 and \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\)) and for all melt rate parameterizations, our simulations show the Ross catchment to switch from mass gain to mass loss. However, the magnitude of mass loss varies from 7 to 39 mm SLE after 100 years, depending on the melt rate distribution (Fig. 2c). Many of the final 100-year grounding line positions are similar between the two scenarios (for example, at MacAyeal and Bindschadler ice streams), but there are some regions, for example, the Kamb ice stream, where the grounding line positions show clear dependence on the melt rate parameterization (Fig. 2a). Across an extensive distance (90–150 km) inland of many of the current grounding lines, the ice is lightly grounded on the underlying topography (<100 m above flotation; Supplementary Fig. 12), and retreat over these regions does not coincide with a substantial increase in volume loss (Fig. 2c). The exception is Bindschadler ice stream, where the inland ice is >200 m above flotation (Supplementary Fig. 12), and retreat is concurrent with an increase in ice loss (Fig. 2a). Here after migrating 5 km inland, grounding line retreat becomes more rapid until reaching approximately 36 km inland (Fig. 3d). Retreat then slows for the following 2–3 decades (Fig. 3d). In most simulations, a second phase of accelerated retreat occurs between 55 and 85 km inland (Fig. 3d).

As the ocean model outputs do not project a regime shift in the Amundsen Sea Embayment, most of our simulations continue to lose ice at approximately the same rate as present day, amounting to 32 ± 45 mm SLE volume after 100 years (Fig. 2d). This large uncertainty is due to the NEMO4 control simulations, where, depending on the melt rate parameterization, some show up to 120 mm SLE volume loss (Fig. 2d) and 110 km of grounding line retreat at Thwaites Glacier (Fig. 3c). In the perturbed NEMO4 simulations, a twofold increase in melt rates is sufficient to accelerate ice volume loss, reaching up to 160 mm SLE after 100 years, and the Thwaites grounding line retreats a further 20 km than in the control simulations. In addition, the grounding lines of Pope, Smith and Kohler glaciers also migrate up to 120 km inland (Fig. 2a). At Thwaites Glacier, in all simulations, once the grounding lines reach approximately 35 km inland of the present-day position (between 7 and 14 years), retreat (Fig. 3c) and ice volume loss (Fig. 2d) both accelerate. In the two perturbed simulations—that impose melting as the geometry changes—the final 100-year grounding line positions are in similar locations, 120–130 km inland of present day. This suggests that retreat, at least eventually, becomes insensitive to the exact prescription of the melt rates.

Irreversible grounding lines

In the previous section, we pinpointed several regions in West Antarctica where grounding line retreat accelerates in all forward simulations. For the high-end perturbation experiments (NEMO4 and \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\)) in both cold-cavity basins, we test whether these changes are irreversible by removing the perturbation in ice shelf melt rates just before phases of accelerated retreat. From this we can assess if phases of accelerated retreat are committed to occur irrespective of the melt rates. Melt rates are reverted to control conditions and simulations continued up to 1,000 years.

For the Filchner–Ronne region, we perform reversibility experiments at time stamps (earliest 6 years, latest 43 years) when the grounding lines reach 6 and 34 km along the Möller ice stream profile (blue and red circles in Fig. 3a,b). In all experiments, irrespective of the scenario or melt parameterization, retreat continues (example shown in Extended Data Fig. 4a,b), and in no simulations do the grounding lines appear to stabilize anywhere between 6 and 110 km inland of the present-day position (Fig. 3e). Whereas retreat briefly slows once grounding lines reach 34 km inland (when the reverse experiments started at 6 km), they nevertheless all continue to retreat (Fig. 3e) across an overdeepened section of the bed (Extended Data Fig. 4b). In two additional reversibility experiments, we reduce the control melt rates by 50%, and find that this remains insufficient to halt the retreat of the grounding line further inland (red lines, Fig. 3e). In all reversibility experiments, retreat slows once the grounding lines reach approximately 110 km, and they are able to persist at this retreated location (120–140 km) until the end of the 1,000-year simulations (Fig. 3e).

At Bindschadler ice stream, we remove the perturbation in ice shelf melt rates once the grounding lines migrate 5 km inland (earliest 26 years, latest 50 years) and revert to control conditions. In all such reversal experiments, the grounding lines continue to retreat, but after 100–180 years settle at approximately 45 km inland of the present-day location and persist there for decades to centuries (Fig. 3g). After 400–600 years in simulations using control melt rates from NEMO4, all grounding lines then re-advance close to their initial present-day locations (Fig. 3g). In contrast, in the \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\) simulations (which overestimate observed integrated ice shelf melt rates; Extended Data Fig. 3), after 160–430 years (depending on the melt parameterization), the grounding lines enter a second phase of accelerated retreat and continue to migrate, reaching approximately 85–91 km inland from present day. After reaching this location, they do not migrate any further within 1,000 years (Fig. 3g). Given the different responses depending on the melt scenario, there may be two separate regions of irreversible retreat. It appears that the grounding lines can remain constant at two locations, both the present-day position (consistent with ref. 17) and at approximately 90 km inland but after retreating to 45 km will diverge to either one of these positions. The existence of several regions of irreversible retreat within one glacier system was previously shown for Pine Island Glacier18. Establishing this for Bindschadler ice stream will require an in-depth regional study, but nonetheless, our results clearly show a regime shift in ice shelf melt rates will cause the grounding line retreat.

At Thwaites Glacier, phases of accelerated retreat occur in both our control and perturbed experiments. We explore whether grounding line positions under current conditions are irreversible by reducing the control melt rates by 50% after 100 years to simulate a cooling of the cavity (pink, orange, brown circles in Fig. 3). In simulations where melt rates were fixed through time, retreat is limited to 6 km within 100 years, and in our reversibility experiments, there is no further retreat inland (solid red line Fig. 3f). This confirms previous work that at its present-day location, the Thwaites Glacier grounding line is not yet retreating irreversibly17,19. Crucially, however, in simulations where grounding lines migrate 60–100 km inland during the forward control experiments (Fig. 3c), this 50% reduction in control melt rates does not prevent further retreat. Instead, the grounding lines migrate up to 460 km inland of the 100-year control run positions and show no sign of halting within 1,000 years (Fig. 3f and Extended Data Fig. 4c). Whereas we cannot pinpoint the exact location where retreat becomes irreversible, our experiments indicate that once the grounding lines reach approximately 75 km inland of present day, retreat is likely to continue across the deep marine sections of the catchment (Extended Data Fig. 4). Moreover, the likelihood of grounding lines undergoing irreversible retreat once they migrate further inland under present-day melt rates aligns with earlier studies suggesting Thwaites Glacier could collapse under current climatic conditions19,20,21,22,23.

Implications for extent and timing of ice sheet collapse

Previous studies have suggested a regime shift in the thermal state of ice shelf cavities could occur within the twenty-first and twenty-second centuries7,8,10. Whereas our forcing scenarios here do not focus on the exact timing of a regime shift itself, nor follow specific shared socio-economic pathways (which themselves are associated with large uncertainties), we are able to explore the timescales associated with grounding line retreat after such a regime shift in ocean-induced melt rates has happened. By applying these melt rates instantaneously, our results show that if high-end melt rates typical of a sudden regime shift—were to persist for a further 100 years, we could see substantial contributions to global mean sea-level rise from the Filchner–Ronne and Ross ice shelf catchments. Additionally, we find that the onset of accelerated and irreversible retreat occurs in all our simulations and not only for certain combinations of ocean model scenario and melt rate parameterization, provided that a shift from cold to warm conditions has occurred. In particular, once retreat at the Möller ice stream is set in motion, without a strong reduction in sub-shelf melt rates (>50% lower than present day over 100s of years), this retreat will continue up to 110 km inland of present day.

Despite our model results showing widespread and probably irreversible retreat of the grounding lines in the Filchner–Ronne and Ross regions, this does not lead to a runaway collapse of the West Antarctic ice sheet via these catchments. Both regions start to exhibit a slowdown in retreat rates within 100 years, provided there is no further perturbation in sub-shelf melt. Extended simulations show the grounding lines can remain close to these retreated locations over a further 100 years (Supplementary Fig. 20) Unlike Thwaites Glacier, where a runaway collapse of the West Antarctic ice sheet19,20,22 may be facilitated by a deep marine basin inland of the current grounding line, the topography further inland of the Filchner–Ronne and Ross ice shelves may limit the potential for runaway retreat.

Here we explore the ice sheet response to an immediate step change in melt rates to identify the potential for irreversible ice loss. In reality, as shown by the ocean models, such a transition to a warm state may take several decades. Should melt rates increase more gradually, the response of the ice sheet will be more gradual as well, but our statements about irreversible grounding line locations will not be affected. Furthermore, whereas our results are robust across a suite of melt rate parameterizations, coupled ice–ocean models—that capture all melt-geometry feedbacks24—will help to further narrow down the timeframes of ice sheet response to a regime shift. Other model choices such as mesh resolution and melting close to the grounding line affects some details of our modelled transient retreat rates (Supplementary Section 6.1), but irreversible retreat of the grounding lines over centennial timescales is a robust numerical aspect of our simulations. Future ice loss could be compensated for by increases in snow accumulation25,26,27 expected in a warmer climate (not accounted for in our simulations). In a sensitivity test we find that increased precipitation could indeed extend timescales of grounding line retreat but does not affect our statements about irreversibility once grounding lines migrate far enough inland (Supplementary Figs. 18 and 19). Timescales of irreversible ice loss may also be sensitive to other modelling choices not addressed here, for example, parameterizations of basal sliding and iceberg calving and uncertainties in the values of unknown parameters in our ice sheet model. However, recent work has shown that using models initialized using data assimilation, as used here, the impact of the choice of basal sliding law on ice loss projections may be tightly bounded on decadal to centennial timescales28. Whereas we do not include calving processes, we note that calving is often limited to the frontal regions of ice shelves, which, in general, do not provide substantial buttressing1. Ocean-induced melt across the entire ice shelf area (which we do include) is likely to have a greater impact on upstream ice flow due to thinning of highly buttressed regions across the ice shelf1,2.

This study has shown that a switch from cold to warm ocean conditions has important implications for the large Filchner–Ronne and Ross regions of the Antarctic ice sheet, leading to ice loss larger than present-day loss from the Amundsen Sea Embayment region. We have also identified crucial regions inland of these large ice shelves over which retreat becomes irreversible and ice loss continues even under restored cold conditions. Both the Filchner–Ronne and Ross regions are not currently contributing to global sea-level rise, and we find no indication that this is likely to change in the near future under current climate conditions. However, a switch to a warm ocean state will not only result in substantial ice loss from those regions but act as a trigger for the irreversible retreat of some of their grounding lines.

Methods

Ocean model states

The published ensemble of sub-shelf melt rates from the state-of-the-art ocean-circulation models used here is provided on Zenodo12 and shown in Fig. 1. Note that the NEMO4 simulation presented here is the updated state presented in ref. 11. Reference simulations for these ocean states are also provided, and integrated ice shelf melt rates are broadly in agreement with one another and observations (Extended Data Fig. 3), especially in the basins in which a regime shift exists between the reference and perturbed melt rates (Filchner–Ronne, Ross). In all reference ocean model simulations, the difference between modelled and observed integrated melt rates in the Amundsen Sea is greater. This is probably due to coarse mesh resolution and the exclusion of subglacial run-off11,30,31. Each of these perturbed ocean model simulations used an anomaly approach to perturb the model configurations and found a regime shift in ocean conditions, characterized by at least an order-of-magnitude increase in sub-shelf melt rates. The NEMO4 melt rates are the last 10 years (2089–2099) of a 100-year ‘high-end’ perturbation simulation using anomalies from the IPSL CMIP6 SSP5–8.5 model between 2260 and 2300. These melt rates are representative of late twenty-third century conditions in the low-probability scenario SSP5–8.5 (ref. 11 provides details). \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\) melt rates (Fig. 1b) are the last 10 years (2190–2199) of a HadCM3 21C-A1B scenario simulation with respect to reference simulation using atmospheric forcing from the HadCM3 twentieth century climate model simulation (ref. 10 provides details). In both scenarios, warm circumpolar deep water flows beneath many ice shelves around Antarctica (Fig. 1a,b). The \({{\rm{FESOM}}}_{\sigma }^{{\rm{low}}}\) scenario is the last ten years of a continuation of the reference run (2190–2199) also forced with HadCM3 A1B, from ref. 15. All FESOMz perturbed states (high, mid, low) are with respect to a single reference run using ERA Interim forcing (Extended Data Fig. 3). The \({{\rm{FESOM}}}_{z}^{{\rm{high}}}\) simulation perturbed the atmospheric forcing using HadCM3 21C-A1B from 2050 onwards (ref. 8 provides details). Unlike the previous two high-end ocean states (NEMO4 and \({{\rm{FESOM}}}_{\sigma }^{{\rm{high}}}\)), substantial increases in melting occur only in the Filchner–Ronne ice shelf, and there is no regime shift in the Ross ice shelf cavity. Additional perturbations were made to the ERA-Interim forcing (used in the reference simulation) to create the mid and low scenarios for the FESOMz configuration8.

Parameterized sub-shelf melt rates

Standalone ice sheet models rely on sub-shelf melt parameterizations, and evaluating the best approach remains the focus of ongoing work32,33. During our experiments, we capture uncertainties in the prescription of the ocean model-derived sub-shelf melt rates as the geometry evolves through time by using three approaches (Supplementary Section 5 provides details):

  1. (1)

    Moc: ocean model melt rates are applied in our ice sheet model at the beginning of the simulation and remain fixed (do not evolve) through time. No melt is applied to cells that become afloat during the simulation. Whereas this has the advantage of retaining the spatial distribution of melt from the ocean models, it does not account for any melt-geometry feedbacks.

  2. (2)

    Mmat: melt rates are calculated using a modern analogue technique (for example, ref. 14), where at an ice sheet grid point, the melt rate is the average of melt rates at k selected grid points in the ocean model training dataset (Moc). Selection of the best grid points is based on the strength of matches to geometric characteristics. Melt rates are recalculated as the ice shelf geometry evolves through time.

  3. (3)

    Mquad: melt rates are parameterized using a standard quadratic dependency32 on thermal forcing and evolve through time.

At the beginning of our simulations, all parameterizations closely replicate the integrated ocean model melt rates, especially at the Filchner–Ronne ice shelf (Extended Data Fig. 2). At the Ross ice shelf, the quadratic parameterization overestimates melt rates (Extended Data Fig. 2), consistent with previous studies33. In all cases the Mmat approach captures both the integrated and spatial pattern of sub-shelf melt rates (Extended Data Fig. 2).

Ice sheet model initialization

Here we initialize the shallow-shelf ice sheet model Úa13 using observed ice geometry from BedMachine v229,34 and observed ice velocities35. First, we perform a model inversion in which we estimate unknown parameters in the regularized Coulomb sliding law (Supplementary equation 3) and in Glen’s flow law (Supplementary Section 2). We use a finite element mesh with element sizes ranging from 2.1 km close to the grounding line and areas of high strain rates to a maximum of 160 km inland. After the inversion, we relax the model by making an iterative modification to the surface mass balance term as where we subtract observed rates of thickness change36 from the RACMO37 surface mass balance field (Supplementary Section 3). We test the sensitivity of our results to this mass balance modification and find our statements regarding the onset of irreversibily are not affected by modelling choices related to the initialization procedure (Supplementary Section 6.2). During forward experiments, we use adaptive mesh refinement to retain high resolution around the grounding line as it migrates. The calving front position remains fixed throughout all simulations. We also do not account for any solid earth feedbacks in our simulations. Using the four reference ocean model melt rates (Extended Data Fig. 3) and three different melt rate parameterizations, we run 12 forward-in-time control simulations. Under control conditions, the spatial rates of ice volume loss replicate observations (Extended Data Fig. 1) and the integrated loss 0.91 ± 0.64 mm yr−1 is also in good agreement with observed rates of volume loss of 0.7 ± 0.07 mm yr−1 between 2009 and 2017 (Supplementary Table 1)5.

Experimental design

In our perturbation experiments, we immediately apply the regime-shifted melt rates to our initialized ice sheet model configuration and run simulations forward in time for 100 years. From this we can assess the ice sheet response to a sudden and dramatic regime shift and determine whether such a large amplitude perturbation has the potential to initiate instabilities in the ice sheet. During all experiments, the surface mass balance as remains fixed through time. Over the floating ice shelves, the sub-shelf melt rates ab are prescribed as

$${a}_\mathrm{b}={M}_{{\rm{OC}}}^{{\rm{ref}}}+{M}_{{\rm{OC}}}^{{\rm{anom}}}$$
(1)

where \({M}_{{\rm{OC}}}^{{\rm{ref}}}\) and \({M}_{{\rm{OC}}}^{{\rm{anom}}}\) are the reference sub-shelf melt rates and sub-shelf melt rate anomalies (with respect to those reference melt rates), respectively, from the ocean models. During the control experiments \({M}_{{\rm{OC}}}^{{\rm{anom}}}=0\). In total we run 40 forward-in-time simulations (18 perturbed and 12 control) based on six perturbed melt rate scenarios (Fig. 1), four reference model configurations (Extended Data Fig. 3) and three melt rate parameterizations (Extended Data Fig. 2).

During these forward-perturbed simulations we identify the timing of phases of accelerated retreat. To test whether changes in the ice sheet are irreversible, we perform reversibility experiments. For each region (Filchner–Ronne, Ross and Amundsen Sea Embayment), we identify the time at which retreat appears to become irreversible and perform a series of reversibility experiments in which we either remove the perturbation in ice shelf melt rates and revert to control conditions or further reduce the control melt rates by 50%. The timing and detail of these experiments for each region is shown by the coloured circles and text in Fig. 2.