Correcting model drift shifts Antarctic sea-level uncertainty

A study explains why removing systematic drift makes climate assumptions account for more of the remaining uncertainty—and what that correction leaves out.

A study explains why removing systematic drift makes climate assumptions account for more of the remaining uncertainty—and what that correction leaves out.
Removing systematic drift from Antarctic ice-sheet simulations changes which assumptions dominate their uncertainty. In a study published in Nature Geoscience on September 30, 2026, researchers report that ice-model features account for an average 64% of the variance—the spread among simulated outcomes—in raw simulations, compared with 23% after correction over 2015–2100. Climate assumptions then account for nearly half of the smaller remaining variance.
These are estimates of Antarctica’s contribution to sea-level change, not projections of total sea-level rise. Antarctica accounted for about 10% of global mean sea-level rise during 2006–2018, but its enormous ice reservoir makes it a major source of uncertainty about future rise. Understanding where that uncertainty comes from matters when researchers compare projections used in coastal planning.
Analysis
Raw simulations use the ice-sheet models’ outputs directly. They retain nonlinear ice dynamics: the ice sheet’s response need not change in simple proportion to the warming applied. But many raw simulations fail to reproduce observed mass-loss trends or show unrealistic model drift, a systematic change that complicates interpretation of the response to climate change.
Drift correction subtracts a model’s control run from its raw simulation to isolate the response to climate change and remove systematic biases. The tradeoff is that subtraction cannot fully preserve nonlinear ice dynamics. It also removes background Antarctic mass loss. Following the IPCC Sixth Assessment Report approach, the researchers restore that background contribution with a linear sea-level trend of 0.33 ± 0.16 millimetres per year. The uncertainty is one standard deviation, and the trend is assumed to be independent of other model features.
The researchers attribute the reduction in variance to replacing control-run variability, which accounts for 45% of raw-simulation variance, with that historical trend. The control-run variability comes from ice-model features, not climate forcing or ice-shelf collapse. Uncertainty from climate forcing and ice-shelf collapse remains unchanged between the two simulation types, so it becomes a larger share of the reduced total.
| Source of uncertainty | Raw simulations | Drift-corrected simulations |
|---|---|---|
| Ice-model features | 64% | 23% |
| Atmosphere–ocean climate-model choice | 9% | 31% |
| Global mean surface temperature | 6% | 17% |
Global mean temperature describes worldwide warming, while atmosphere–ocean general circulation models determine regional conditions around Antarctica, including snowfall and ocean warming. Those regional differences affect how much ice accumulates or is lost. The substituted historical trend also brings its own uncertainty, which peaks at 30–33% of variance during 2030–2040.
The researchers trained a machine-learning emulator on simulations from ISMIP6-2300, an ice-sheet model-intercomparison ensemble. An emulator approximates the underlying models’ outputs, allowing researchers to explore many combinations of climate and ice-model assumptions without rerunning the full simulations. This emulator links those assumptions to both global-average and location-specific Antarctic sea-level contributions.
The team assessed performance with tenfold cross-validation: dividing the data into ten parts and repeatedly testing predictions on a held-out part. The researchers report robust performance across prediction targets. The emulator is still an imperfect approximation, and the analysis includes uncertainty from its predictions. That uncertainty dominates before 2030 in the raw simulations because the overall projection uncertainty is then small.
To estimate each assumption’s contribution, the researchers used a Bayesian hierarchical framework, a probability-based method for separating linked sources of uncertainty. Its starting assumptions, called priors, treat every included ice-model option as equally plausible. Warming relative to 2015 is assigned equal probability across the range from no further warming to the maximum climate-model output at each time step. These priors generate the large ensemble used to calculate the variance shares; the shares therefore depend on both the ISMIP6 options and the chosen priors.
The same total Antarctic ice loss can produce different sea-level changes in different cities, depending on where the ice disappears. Changes in ice mass affect gravity, Earth’s rotation and deformation, creating regional patterns known as sea-level fingerprints. The researchers used pattern matching to connect the emulator’s global contribution to local changes while accounting for different spatial patterns of mass loss.
| City | Share of variance |
|---|---|
| London | Less than 1% |
| Singapore | Less than 1% |
| New York | 5% |
| Shanghai | 3% |
| Buenos Aires | 17% |
| Melbourne | 21% |
In Buenos Aires and Melbourne, uncertainty about where ice is lost rivals ice-dynamic features as a major contributor. These percentages describe uncertainty in the local pattern relative to the global contribution—the normalized fingerprint—not the additional uncertainty about the total amount of Antarctic mass loss.
The researchers restrict the emulator to 2100 because the ensemble is too small to capture the strong nonlinear behaviour expected from the twenty-second century onward. None of the included models incorporates Marine Ice Cliff Instability, a mechanism involving ice-cliff collapse. ISMIP6 does include a simplified scheme for ice-shelf collapse, but that is not the same process.
The uncertainty shares thus describe the choices represented in this ensemble, not every possible Antarctic response. The researchers say separating the contributions of individual assumptions can inform future experiment design, helping identify which modelling choices to investigate next.

The commission’s October vote could advance new protected areas and tribal harvest rules, but approval and shared management remain unsettled.
4 min

An Eos comparison traces hillside failures into river channels and the bay after heavy rain in North Korea. Clouds and delayed access to imagery limit the initial assessment.
2 min

NHS England’s review found consent procedures did not follow its digital principles. Its guidance calls for decisions for each patient, rather than blanket use of monitoring technology.
4 min