How much sea-level rise could Antarctica commit this century?
A Nature Geoscience study uses a machine-learning emulator across 340 ice-sheet simulations to separate physical and modelling uncertainty. It finds committed mass loss is very likely, while the upper projections remain conditional on emissions and model assumptions.
The short version
- What happened
- The emulator was trained on 29,240 annual samples from 340 simulations, each covering 86 time steps and described by 21 forcing and ice-model features.
- Who it affects
- Coastal residents and infrastructure planners need ranges and high-end stress tests because the Antarctic contribution remains deeply uncertain even when mass loss is very likely.
- What remains unproven
- The machine-learning emulator approximates an existing ice-sheet model ensemble and cannot recover physical processes absent from that ensemble.
Keep in mind
The machine-learning emulator approximates an existing ice-sheet model ensemble and cannot recover physical processes absent from that ensemble.
The main analysis stops at 2100 because the available ensemble was too limited to emulate stronger twenty-second-century nonlinearities reliably.
Marine ice-cliff instability was not included, and the prior treats available modelling options as equally plausible where evidence cannot distinguish them.
Source analysis · 1 linked source. Primary analysed source date: 30 September 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
The study asks why credible ice models disagree
Antarctic sea-level projections diverge not only because emissions differ, but because modelling teams make different choices about ocean forcing, ice flow, sliding at the bed, ice-shelf melting, resolution and model initialisation. A Nature Geoscience study published on 30 September uses a machine-learning emulator to separate those contributions and sample combinations more quickly than full ice-sheet models permit. The objective is not to replace physical modelling with a black box; it is to approximate an existing international simulation ensemble so that its assumptions and uncertainty can be interrogated systematically.
The authors used the ISMIP6-2300 intercomparison archive but restricted their principal analysis to 2015–2100. Their database contains 29,240 annual samples drawn from 340 simulations, each with 86 time steps. Every sample has 21 features: 18 describing ice dynamics plus the atmosphere–ocean model, emissions scenario and projection year. The emulator was tested with tenfold cross-validation that withheld whole simulations, a stricter design than randomly splitting annual rows from the same model run.[1]
Mass loss is very likely; the amount remains conditional
After combining the emulator with a Bayesian framework and satellite observations, the study estimates a probability of at least 0.92 that Antarctica is committed to net mass loss during this century, including under aggressive emissions reductions. It also estimates at least 0.89 probability that higher emissions produce greater Antarctic loss by 2100. Those probabilities express the model framework and assumptions; they are not direct measurements of a future event and should not be rounded into certainty.
Under a very high-emissions pathway, the paper reports an Antarctic contribution of up to 25.4 centimetres of global mean sea-level rise at the 95th percentile by 2100, with a median of 15.7 centimetres. The 95th percentile is a high-end point in a conditional distribution, not the central forecast and not total global sea-level rise. Oceans expand as they warm and other glaciers and ice sheets also contribute, while regional sea level differs because gravity, land movement and ocean circulation redistribute water.[1]
Model choices can dominate the uncertainty people see
In raw simulations, ice-model features accounted for an average 64% of twenty-first-century Antarctic projection variance, with sliding-law choices alone contributing 21%. After drift correction, the balance changed: climate forcing components contributed nearly half of variance, including 31% from the selected atmosphere–ocean model and 17% from global temperature change. This contrast shows why a single uncertainty range can conceal different physical and numerical explanations.
The team evaluated random forests, gradient boosting, LightGBM, XGBoost, CatBoost and a multilayer perceptron. Complete simulation time series were kept together during cross-validation so a model could not see neighbouring years from the same run in training and testing. The final emulator enabled large Monte Carlo samples and fast sensitivity analysis. Even so, it learns from the available ensemble; if that ensemble omits a process or clusters around similar modelling conventions, rapid sampling cannot create missing physics.[1]
What coastal communities can use—and what would change our assessment
For coastal planning, the result reinforces a risk-management approach rather than a single-number forecast. Authorities can test infrastructure and evacuation plans across median and high-end sea-level pathways, update them as observations improve and avoid interpreting emissions cuts as an immediate end to committed ice loss. Rapid mitigation still matters because the study finds higher emissions increase loss and because long-lived coastal assets are sensitive to cumulative change beyond 2100.
Our assessment would strengthen as new satellite observations narrow calibration, intercomparison projects sample more independent model structures and held-out real-world periods test predictive performance. It would change materially if evidence supports or rejects rapid marine ice-cliff instability, a debated process not included in this ensemble. It would weaken if the emulator fails on new model configurations or if uniform priors over modelling choices prove unrealistic. The paper clarifies uncertainty inside a major ensemble; it does not eliminate deep uncertainty or provide a site-specific flood forecast.[1]