To fill the gap in multi-centennial projections of Antarctic mass changes26,27,28 and provide robust estimates of future sea-level change, we need to systematically account for as many sources of uncertainty as possible (such as future emissions, climate model, ice-sheet model structure and parametric uncertainty) while calibrating with past (observed) mass changes20. Here, we provide a robust assessment of the future evolution of the Antarctic Ice Sheet that, for the first time, jointly accounts for parametric, ice-sheet model structure, and climate model uncertainties in historically-calibrated projections through 2300 under a wide range of emissions, from low to very high. Future Antarctic climate trajectories to 2300 under the extended Shared Socio-Economic Pathways SSP1-2.6 and SSP5-8.529 are derived from four General Circulation Models (GCMs) from the sixth phase of the Coupled Model Intercomparison Project (CMIP6) and used to drive two ice-sheet models, with uncertainties in ice-climate interactions explored through a wide range of parameter perturbations. We quantify the contribution of these various sources of uncertainty to the projected range in ice-sheet changes using an analysis of variance, and assess the long-term consequences of the warming projected over the next centuries (that is, the committed ice-sheet response), by maintaining the changes in climate projected for 2300 through the end of the millennium.
The Antarctic Ice Sheet may face a wide range of future climates over the coming centuries, depending on the emission pathway and strongly modulated by the climate model used (Fig. 1). Here, we base our analysis on a subset of CMIP6 GCMs-the UK Earth System Model (UKESM1-0-LL; ref. ), the Institut Pierre-Simon Laplace global climate model (IPSL-CM6A-LR; ref. ), the Community Earth System Model (CESM2-WACCM; ref. ) and the Meteorological Research Institute Earth System Model (MRI-ESM2-0; ref. ), which provide some of the few projections up to 2300 available in CMIP6 at the time of this analysis. These GCMs span a broad range of climate sensitivities, representative of the CMIP6 ensemble. Despite significant improvements over previous generations, CMIP6 GCMs still show biases in Antarctic climate (e.g., refs. ). For example, the upper Southern Ocean is generally too warm and fresh, and the Amundsen Sea Low remains poorly captured.
During the first half of this century, projected Antarctic atmospheric and oceanic warming is similar across the two emission scenarios. Divergence between the higher-emission (SSP5-8.5) and the lower-emission (SSP1-2.6) pathway emerges after 2050, eventually leading to significant changes in Antarctic climate on multi-centennial timescales (Fig. 1). Under SSP1-2.6, the four GCMs considered project a limited Antarctic-averaged atmospheric and oceanic temperature change of +1.0 °C - +3.6 °C and +0.2 °C - +0.7 °C, respectively, by 2300 compared to the 1995-2014 period. In contrast, the SSP5-8.5 scenario leads to substantial changes after the end of this century, with Antarctic-averaged atmospheric warming ranging from +12.0 °C to +17.0 °C and ocean warming from +1.7 °C to +3.1 °C by 2300 (Fig. 1).
Across the emission pathways, the projected changes in Antarctic climate over the next three centuries show similar characteristics for each CMIP6 GCM considered. Compared to the other three CMIP6 GCMs, MRI-ESM2-0 projects relatively slow and limited changes, especially under very high emissions (with Antarctic oceanic and atmospheric warmings of +1.7 °C and 12.2 °C, respectively, by 2300). Within each emission pathway, CESM2-WACCM predicts the strongest atmospheric warming of the four GCMs, along with a steady increase in circum-Antarctic ocean temperatures, reaching up to +0.7 °C under SSP1-2.6 and +2.0 °C under SSP5-8.5 by 2300. In contrast, projections by IPSL-CM6A-LR and UKESM1-0-LL show early and abrupt ocean warming in the first half of this century under both emission pathways. Under SSP5-8.5, these two GCMs yield the strongest multi-centennial Antarctic ocean warming, reaching +2.7 °C and +3.1 °C, respectively, by 2300.
To quantify the Antarctic contribution to sea-level rise and its associated uncertainties under these various future climates, we perform a 1400-member ensemble of simulations using two state-of-the-art ice-sheet models: Kori-ULB and the Parallel Ice Sheet Model (PISM; see Methods). Future Antarctic climates are derived from the four CMIP6 GCMs described above under the emission pathways SSP1-2.6 and SSP5-8.5 (Fig. 1). Key model parameters controlling the interaction of the ice sheet with the atmosphere and the ocean (Table 2) are varied using Latin Hypercube designs, which sample efficiently across the parameter space (as used previously for ice-sheet projections by, for example, refs. ). Our simulations begin in 1950 to allow, for each ice-sheet model, a Bayesian calibration with observations from the satellite era. Each ensemble member is attributed a weight by comparing the simulated ice-sheet evolution over the past decades with a series of mass balance estimates from the latest Ice Sheet Mass Balance Inter-comparison Exercise (IMBIE; see Methods and ref. ).
Through this calibration, the ice-sheet model ensembles are brought closer to the observed ice-sheet mass changes over the past decades (Supplementary Fig. 1a, b). This improved agreement between modeled and observed mass changes is quantitatively confirmed by a reduction in the Continuous Ranked Probability Score (CRPS) following the calibration (Supplementary Fig. 1e-l; see Methods for details). In addition, conditioning on the IMBIE record is effective in reducing the uncertainty in the hindcasts, especially for the Kori-ULB ensemble (Supplementary Fig. 1a) given the wider range of ice-ocean uncertainties explored (Table 2; see Methods).
Some regional biases in Antarctic mass changes, however, persist in the calibrated ensembles. For example, PISM simulations tend to underestimate the observed West Antarctic ice-sheet thinning (Supplementary Fig. 1d, k), whereas Kori-ULB trajectories capture this trend more accurately (Supplementary Fig. 1c, g). Conversely, East Antarctic mass gain is slightly overestimated in the calibrated Kori-ULB ensemble compared to the IMBIE estimates (Supplementary Fig. 1f). These discrepancies between simulated and observed historical ice-sheet changes likely contribute to the overestimation of the ice-sheet net mass balance for both ice-sheet models, especially after 2000 (Supplementary Fig. 1a, b; see e.g., median of posterior distribution). Between 1992 and 2020, the calibrated median net mass balance amounts to -77 Gt yr (-189 Gt yr - +1 Gt yr, 5-95% percentile range) for the Kori-ULB ensemble and to -59 Gt yr (-157 Gt yr - +26 Gt yr) for the PISM ensemble, compared to the observed value of -92 ± 18 Gt yr (see Supplementary Fig. 1e-l).
These regional discrepancies may lead to compensatory effects at the continental scale. For example, the calibration tends to favor PISM ensemble members that overestimate mass loss in the Antarctic Pensinula to offset the underestimation of West Antarctic mass loss. In contrast, for Kori-ULB, ensemble members that slightly overestimate West Antarctic mass loss are favored to counteract the model's bias towards mass gain in East Antarctica. Such compensations may reduce the ability of the calibration to fully capture regional dynamics, and can lead to biases in specific regions despite overall agreement at the continental scale (see Supplementary Fig. 1e-l and Methods).
Our multi-model ensemble supports emerging evidence suggesting little scenario-dependence of the Antarctic contribution to sea-level rise by 2100 (Fig. 2a, zoom-in). In particular, in the coming decades, ice-sheet trajectories under both emission pathways substantially overlap, with projected sea-level contributions ranging between -0.03 m and +0.33 m for SSP1-2.6 and between -0.05 m and +0.37 m for SSP5-8.5 by 2100 (Table 1). On multi-decadal timescales, mass loss remains confined to West Antarctica as well as the Antarctic Peninsula and is offset to some extent by East Antarctic mass gain (Table 1, Supplementary Fig. 2).
After 2100, the projected Antarctic sea-level contribution under both emission pathways starts to diverge, potentially leading to multi-meter ice loss under very high emissions. By 2300, mass loss under SSP5-8.5 ranges from +0.73 m to +5.95 m sea-level equivalent (Fig. 2a, Table 1), primarily driven by a significant grounding-line retreat in the Amundsen Sea Embayment, the Siple Coast, and Weddell Sea regions (Fig. 3c, d). In addition, marine parts of the East Antarctic Ice Sheet, such as the Wilkes and Recovery catchments, potentially lose mass by 2300 (Fig. 3c, d). Multi-centennial ice loss from these East Antarctic marine catchments is only partially offset by accumulation in the interior of East Antarctica, with the lower bound of the 5-95% probability interval reaching less than -0.20 m sea-level equivalent by 2300 under SSP5-8.5 for both ice-sheet models (Table 1, Supplementary Fig. 2c, d).
When strongly constraining emissions according to SSP1-2.6, the projected Antarctic contribution to sea-level rise remains below +1.75 m by 2300 with 95% probability across both ice-sheet models (Fig. 2a, Table 1). Compared to SSP5-8.5, mass loss is substantially reduced due to a lower likelihood of extensive grounding-line retreat in West Antarctica. Nonetheless, the resulting sea-level rise may still require major coastal adaptation efforts. In East Antarctica, mass loss may be avoided by 2300 when following SSP1-2.6, with increased snow accumulation dominating the regional signal and leading to a negative median contribution to sea level for both ice-sheet models (Table 1, Supplementary Fig. 2c, d).
By 2100, our range of the projected Antarctic sea-level contribution is consistent with the IPCC-AR6 assessment (based on their medium confidence projections; Fig. 2a), although our upper bound under SSP5-8.5 is slightly lower. The higher IPCC-AR6 upper bound reflects the LARMIP-2 estimates, which focus on the response to sub-shelf melting under high melt sensitivities representative of the Amundsen Sea Embayment. By 2300, however, our historically calibrated ensembles of simulations yield a higher upper bound than IPCC-AR6 (+4.10 - +4.40 m versus +3.10 m in IPCC-AR6, 83rd percentiles; Fig. 2a) and than ISMIP6 under very high emissions (+5.10 - +5.95 m, 95th percentiles, versus +4.40 m, the upper end of ISMIP6 model range). The higher upper bounds presented here and in ISMIP6 likely reflect the application of a range of extended CMIP6 climate trajectories beyond 2100, which were not considered in IPCC-AR6. The sampling of both ice-sheet model structure and wide parametric uncertainties explains our ensembles' higher upper bounds compared to ISMIP6, despite being constrained by observed mass changes. Yet, even with the strong SSP5-8.5 warming trajectories projected by some of the GCMs considered here, our estimates remain well below the range projected by a single ice-sheet model that considers the Marine Ice Cliff Instability (MICI; +6.87 m - +13.55 m sea-level equivalent under RCP8.5). These projections sampled only a limited subset of MICI-related uncertainty and did not account for broader parametric or climate model uncertainties, which would likely alter these ranges.
Projected Antarctic mass loss is shaped by a shifting interplay between ice-sheet model structures, parameters governing the interactions with the ocean and the atmosphere, and diverging GCM climate trajectories. Structural differences between the ice-sheet models and parameters modulating sub-shelf melting dominate multi-decadal uncertainty. As changes in Antarctic climate intensify and diverge across emission scenarios, parametric uncertainty related to ice-atmosphere interactions and climate model uncertainty increasingly contribute to the spread in the projected Antarctic ice-sheet trajectories. Under very strong warming (SSP5-8.5), this shift in the dominant sources of uncertainty leads to a robust agreement across the ice-sheet models on a multi-meter sea-level contribution by 2300.
To quantify the relative contribution of these different sources of uncertainty in our ensembles, we perform a variance decomposition using an analysis of variance (ANOVA) framework (see Methods; Figs. 4 and 5). This approach attributes the spread in the projected Antarctic sea-level contribution to parameters governing the interaction of the ice sheet with the atmosphere and ocean, the choice of climate model, the ice-sheet model structure, and their two-way interactions (the combined effect of two sources of uncertainty acting together).
In line with previous findings, the choice of ice-sheet model is a major source of uncertainty in Antarctic sea-level projections (Fig. 4), highlighting the influence of differences in model physics (e.g., basal friction laws, calving schemes; see Supplementary Table 1), numerical methods, and initialization approaches. Over the next decades, the influence of this structural uncertainty on the projected ice-sheet trajectories increases under both emission scenarios, with its main effect accounting for up to 40% of the total variance at the continental scale (Fig. 4a, b). This fraction is smaller than in previous assessments, likely because of differences in experimental design: our setup explores a wider range of parametric uncertainty but includes only two ice-sheet models, whereas refs. sampled fewer parameters but a larger and more diverse set of models. In those earlier studies, parametric uncertainty was implicitly embedded in the structural uncertainty, whereas here both sources of uncertainty are explicitly quantified, along with their interactions. Our results suggest that the interactions between parametric and structural uncertainties significantly contribute to the total variance during the next few decades, adding to the main effect of the ice-sheet model structures. The uncertainty in the ice-sheet model structure is illustrated by the systematically higher Antarctic sea-level contribution projected by Kori-ULB compared to PISM throughout this century (Fig. 2a, zoom-in; Table 1). Under SSP1-2.6, PISM projects a zero median Antarctic sea-level contribution by 2100, with a 5-95% range of -0.03 m to +0.04 m. Under SSP5-8.5, the median increases slightly to +0.02 m (-0.05 m to +0.12 m). In contrast, Kori-ULB projects substantially higher estimates: +0.09 m (+0.01 m to +0.33 m) under SSP1-2.6 and +0.08m (-0.01 m to +0.37 m) under SSP5-8.5. This ice-sheet model divergence mainly stems from a stronger dynamic ice loss in West Antarctica in the projections by Kori-ULB than PISM (Supplementary Fig. 2a, b), as a continuation of the trends simulated by each ice-sheet model over the historical period (Supplementary Fig. 1). While the ice-sheet model structure is the dominant source of uncertainty in West Antarctica and the Antarctic Peninsula throughout this century (Fig. 4c, e, f, h), in East Antarctica, it becomes negligible by the end of this century under both emission scenarios (Fig. 4d, g). In this region, both ice-sheet models consistently project a slight mass gain by 2100 (Table 1, Supplementary Fig. 2c, d).
Beyond 2100, the evolution of the structural uncertainty diverges between the two emission scenarios. Under the SSP1-2.6 warming trajectory, it continues to increase, reaching 50% of the total variance at the continental scale by 2300 (Fig. 4a). In contrast, under SSP5-8.5, the strong atmospheric and oceanic changes projected over Antarctica progressively override the trends inherited from the historical period. Consequently, structural uncertainty decreases and eventually becomes negligible on the continental scale by 2300. At the same time, the contribution of parameters related to ice-atmosphere interactions increases, especially in East Antarctica. By 2300, these parameters become the main source of uncertainty in the projected ice-sheet response under SSP5-8.5, accounting for 31% of the total variance (Fig. 4b). This reflects the growing influence of surface melt and runoff in a warming climate. The shift in the dominant sources of uncertainty for this very strong warming scenario leads to robust agreement between the two ice-sheet models on a substantial Antarctic sea-level contribution by 2300 (Fig. 2a). Not only do we find a similar median ice loss of +2.67 m and +2.73 m sea-level equivalent in Kori-ULB and PISM, respectively, but also the 5-95% probability intervals are similar (Table 1). The increasing dominance of the climate pathway over the ice-sheet model structure is also particularly evident when explicitly including the emission pathways in the ANOVA (Supplementary Fig. 3).
Across both emission scenarios, the parameters controlling ice-ocean interactions play a key role in driving divergent mass loss trajectories, accounting for up to about 35% of the variance at the continental scale over the next three centuries (Fig. 4). This signal is largely driven by the Kori-ULB ensemble, where parametric uncertainty in ice-ocean interactions dominates the variance in the projected ice-sheet changes, except under very strong warming (SSP5-8.5 beyond 2250; Fig. 5a, b). In contrast, parameters associated with ice-atmosphere interactions consistently dominate the variance in the PISM ensemble (Fig. 5c, d). This difference in the dominant source of uncertainty between both ice-sheet models likely arises from differences in their ensemble designs: the Kori-ULB ensemble, by sampling a variety of sub-shelf melt parameterizations and associated parametric uncertainty in the effective ice-ocean heat flux, is more sensitive to ocean warming (in particular, to the early increase in circum-Antarctic ocean temperatures projected by UKESM1-0-LL and IPSL-CM6A-LR even under SSP1-2.6; Fig. 1). By comparison, the PISM ensemble relies on a single sub-shelf melt parameterization (the Potsdam Ice-Shelf Cavity mOdel, PICO; see Methods), making it less sensitive to ocean warming. This difference in the sensitivity of the two ice-sheet models ensembles to changes in the ocean is also evident in the contribution of the two-way interaction between ice-sheet model and ocean-related parameters, which quantifies the divergent model responses to ocean forcing. This interaction term specifically captures the influence of the choice of sub-shelf melt parameterization on projected ice-sheet mass changes as well as the ice-sheet model sensitivity resulting from the initialization, which were hidden within the structural uncertainty in previous assessments. These findings highlight the importance of sampling a range of sub-shelf melt parameterizations and associated parametric uncertainties in systematic ensemble designs, next to continued developments in the representation of ocean-induced melting in ice-sheet models. Similarly, the difference in the sensitivity of the two ice-sheet model ensembles to atmospheric forcing is captured by the two-way interaction between the ice-sheet model and atmosphere-related parameters, which emerges as a significant contributor to the total variance at the continental scale under very strong warming.
Beyond ice-sheet model structure and parametric uncertainties, the trajectory of future Antarctic ice loss is also strongly modulated by the choice of GCM, which explains up to 17% of the total variance in the Antarctic sea-level contribution by 2300, consistent with previous estimates. Especially in East Antarctica, the GCM divergence beyond 2100 (Fig. 1) represents an important source of uncertainty (Fig. 4d, g). This influence is clearly reflected in the distributions of the Antarctic sea-level contribution by 2300 for the different GCMs, which in some cases do not even overlap in their 17-83% probability intervals (Fig. 2b; comparing, for example, distributions under MRI-ESM2-0 and UKESM1-0-LL). For example, the limited changes in Antarctic climate projected by MRI-ESM2-0 result in consistently lower ice-sheet sea-level contributions by 2300 in both the Kori-ULB and PISM ensemble (Fig. 2b). In contrast, the strong Antarctic atmospheric warming projected by CESM2-WACCM (Fig. 1a) amplifies Antarctic mass loss, leading to a multi-meter sea-level contribution under SSP5-8.5 emissions for both ice-sheet models (Fig. 2b). The early and abrupt ocean warming characteristic of UKESM1-0-LL and IPSL-CM6A-LR (Fig. 1b) triggers high sea-level contributions compared to the other GCMs in the ocean-sensitive Kori-ULB ensemble, even under SSP1-2.6. In the PISM ensemble, however, these same GCMs lead to lower sea-level contributions than for CESM2-WACCM due to the ensemble's atmospheric rather than oceanic sensitivity (Fig. 5).
Irrespective of the wide range of future Antarctic climates as well as the uncertainties in the ice-sheet model structure and parameters sampled in the ensembles, our projections reveal certainties for the long-term evolution of the Antarctic Ice Sheet to the year 3000: a collapse of the West Antarctic Ice Sheet under very high emissions, and substantially reduced sea-level contribution under low emissions.
Under the higher-emission pathway SSP5-8.5, a collapse of the West Antarctic Ice Sheet is triggered across our entire ensemble of simulations. This collapse could already unfold by 2300, is likely (66-100% outcome probability, see ref. for definitions of likelihood terms and confidence levels followed here) by 2500 and becomes virtually certain by 3000, as a committed response to the projected SSP5-8.5 changes in Antarctic climate (Fig. 3c, d). Such ice loss results in a West Antarctic contribution to sea-level rise ranging from +2.76 m to +6.91 m by 3000 (Table 1).
Self-sustained ice loss is also triggered in East Antarctica under the strong SSP5-8.5 warming trajectory, leading to very likely long-term ice-sheet retreat (Fig. 3c, d). This includes both ocean-driven grounding-line retreat in marine catchments and potential ice loss from regions grounded above sea level, amplified by the surface melt-elevation feedback. As a result, the East Antarctic contribution to sea-level rise by 3000 shows a wide spread, ranging from -0.79 m to +18.05 m (Table 1). Our historically-calibrated projections, linking observed and projected changes to long-term dynamics, consistently indicate that grounding-line retreat in the East Antarctic Wilkes catchment is more likely than in the Aurora catchment under SSP5-8.5 (Fig. 3c, d). This suggests that the currently observed ocean-driven mass loss of Totten Glacier (also captured by Kori-ULB during the historical period; Supplementary Fig. 1) does not necessarily imply a substantial long-term retreat. In contrast, even though the Wilkes catchment currently shows limited mass changes, grounding lines in this region are very likely to retreat over the long term in response to early-millennium SSP5-8.5 warming, consistent with the distinct glaciological and topographic settings of these catchments.
Under the lower-emission pathway SSP1-2.6, a long-term collapse of the West Antarctic Ice Sheet by 3000 cannot be ruled out, but is considerably less likely (Fig. 3a, b). Uncertainty arises from the Amundsen Sea Embayment, where the likelihood of retreat ranges from about as likely as not (PISM) to virtually certain (Kori-ULB), reflecting the trends projected by the ice-sheet models through 2300 (see previous section). In East Antarctica, constraining emissions under the SSP1-2.6 pathway largely prevents substantial self-sustained ice loss at least until 3000 (Fig. 3a, c), although modest mass loss could still contribute to a sea-level contribution of up to half a meter in both Kori-ULB and PISM (Table 1).
Overall, while very high SSP5-8.5 warming could potentially lead to substantial Antarctic ice loss of up to +25.85 m sea-level equivalent by 3000, following the SSP1-2.6 scenario limits the long-term Antarctic mass loss to below +5.28 m with a 95% probability (Table 1). Note that the spread in the Antarctic sea-level contribution over multi-centennial timescales (with higher amplitudes projected by PISM compared to Kori-ULB; Table 1) is partly due to differences in the simulated magnitude of the water-expulsion effect (see Methods).