A new study concludes that Antarctica’s net ice loss this century is effectively locked in, even under the Paris Agreement’s ambitious 1.5°C warming limit, and that extra snowfall from a warmer atmosphere will not offset it enough to reverse the trend. This will add to global sea- level rise that threatens low- lying coastal areas.

The article, by Yucheng Lin (City University of Hong Kong) and Robert Kopp (Rutgers University), was published in The Conversation on September 30, 2026. It summarizes their research in Nature Geoscience.

Key context

• Antarctica holds enough ice to raise global sea level by about 58 meters (190 feet) if it all melted.
• Roughly 1 billion people live in coastal areas, with about 100 million within 1 meter of current sea level and highly exposed to flooding.
• Scientists have long disagreed on whether Antarctica will gain or lose mass this century, because warmer air holds more moisture (favoring more snowfall) while ocean warming melts ice shelves, speeds glacier flow, and drives ice loss.

How the study addressed uncertainty

Ice- sheet models involve a chain of assumptions (emissions → ocean/air warming around Antarctica → ice- shelf melting und or collapse → glacier sliding over bedrock). Different assumptions cascade into large differences in outcomes, and full simulations are computationally expensive.

The researchers trained a physics- informed machine- learning emulator on the large archive of Antarctic simulations from the Ice Sheet Model Intercomparison Project (ISMIP6). This allowed them to rapidly explore hundreds of thousands to millions of combinations of assumptions. They then filtered the results against satellite gravity data (GRACE) that has measured Antarctic mass changes since 2002. Only scenarios consistent with observed ice- sheet behavior were retained.

Main findings

• The viable scenarios consistently show net ice loss from Antarctica this century, even if warming is limited to 1.5°C.
• Extra snowfall does not compensate enough for dynamic losses driven by ocean warming and ice-shelf processes.
• Greenhouse- gas emissions still matter: every avoided ton reduces the amount of Antarctic ice loss this century, with benefits that persist for centuries.
• Under very high- emissions scenarios, cascading effects (rapid ocean warming → thinner and weaker ice shelves → faster glacier flow) could contribute up to ~25 cm (10 inches) of sea- level rise from Antarctica alone by 2100, enough to permanently flood the homes of more than 10 million people.
• The three most important physical uncertainties for future projections are: how ice shelves respond to a warming ocean, how ice slides over bedrock, and how quickly the ocean and atmosphere around Antarctica warm.

Limitations and implications

The authors note that some processes (e.g., ice fracturing and calving and the evolution of subglacial rivers and lakes) are not fully represented in current models, so the emulator cannot capture them. These could alter the timing or magnitude of ice loss and the window for coastal adaptation.

The work provides a faster way to test assumption combinations and offers a clearer basis for the adaptation decisions already facing coastal communities worldwide. Sea- level rise is already driven by Greenland melt, mountain glaciers, and thermal expansion of seawater; Antarctica’s locked- in contribution adds to that total.

In short, the study reinforces that limiting emissions remains valuable for slowing the rate and ultimate amount of Antarctic contribution to sea- level rise, even though some ice loss this century appears unavoidable based on current observations and modeled physics.

ISMIP6 (Ice Sheet Model Intercomparison Project for CMIP6) is a coordinated multi- model effort that produced projections of 21st- century (and some longer- term) sea- level contributions from the Greenland and Antarctic ice sheets. It systematically samples uncertainties arising from climate forcing, ice- sheet model structure and physics, parameters, and ice- ocean interactions.

The Antarctic Ice Sheet (AIS) is the largest single source of uncertainty in future sea- level rise projections. Under high- emissions scenarios, ISMIP6 ensembles span roughly −5 to +43 cm sea- level equivalent (SLE) by 2100 (or −7.8 to +30 cm in core multi- model results relative to a control). Greenland contributions are more consistently positive and better constrained.

Main categories of uncertainty

ISMIP6 was designed to quantify uncertainty from:

• Emissions scenarios (e.g., RCP2.6 vs. RCP8.5; later linked to CMIP6/SSPs).
• Climate models (AOGCMs/ESMs from CMIP5/CMIP6 that supply atmospheric and oceanic forcing).
• Ice- sheet models (structural differences across ~13+ international groups and models).
• Parameters and parameterizations, especially for ice- ocean interactions (basal melt under ice shelves).
• Ice- climate interactions (how models translate ocean thermal forcing into melt rates, grounding- line migration, etc.).

Quantified contributions to uncertainty (Antarctica)

Analyses of the ISMIP6 ensemble (focusing on additional dynamic mass loss driven by changing oceanic conditions) break down total uncertainty at 2100 approximately as follows:

Source of uncertainty	Share of total (AIS scale)	Notes / glacier-scale variation
Choice of ice- sheet model	~52%	Dominant overall. Ranges 14% (some ice streams) to 56% (Pine Island Glacier). Reflects differences in physics, resolution, sliding laws, calving, initial ice-shelf extent, and treatment of melt near the grounding line.
Ice- climate interaction	~22%	Includes melt parameterization choice/calibration and simulated ice- shelf geometries. Higher for sensitive glaciers (e.g., ~36- 39% for Institute Ice Stream and Thwaites).
Choice of climate model	~13% (rising over time)	Grows through the century. Highly variable by glacier (4% for Thwaites to 53% for Whillans Ice Stream).

Other important contributors include:

• Ocean- induced basal melt rates and their calibration (often based on conditions outside ice-shelf cavities).
• Initial ice- shelf extent (varies by a factor of ~2.5 across models).
• Model spatial resolution and numerical treatment near the grounding line.
• Basal sliding laws (e.g., Weertman, Coulomb, or hybrid) and related parameters.
• Calving and ice- shelf collapse representations (ISMIP6 used idealized schemes; many models lack full hydrofracture or marine ice- cliff instability processes).

Key vulnerable glaciers identified include Thwaites and Pine Island (West Antarctica) plus Totten and Moscow University (East Antarctica), which show high sensitivity to ice-shelf basal melt.

Differences between Greenland and Antarctica

Greenland: Mass loss is more strongly driven by surface mass balance (atmospheric forcing). Model and climate uncertainties are substantial, but the sign is consistently positive. Ocean forcing plays a secondary role in the ISMIP6 design. One analysis attributes spreads of ~40 mm (ice-sheet model), ~36 mm (climate model), and ~19 mm (ocean forcing) under RCP8.5.

Antarctica: Competing effects of increased snowfall (atmospheric warming) versus ocean- driven dynamic loss creates sign uncertainty and weaker scenario dependence in many ensembles. Ocean forcing and ice-dynamics response dominate the spread.

Known limitations and gaps

Incomplete process representation: Many simulations omit or simplify ice fracturing and calving, subglacial hydrology/lakes/rivers, tidal water intrusion beyond the grounding line, and full marine ice- cliff instability (MICI). These can alter rates of ice discharge.

Underestimation of uncertainty: Some critiques note that the ensemble may under- sample parametric uncertainty or be biased low relative to recent observations of mass loss.

Computational constraints: Full exploration of all combinations is expensive, which is why later work (including machine- learning emulators trained on ISMIP6 archives) has been used to expand sampling, as in the recent study on locked- in Antarctic loss.

Forcing biases: Climate- model biases in polar ocean and atmosphere fields, and how melt rates are calibrated from them, remain major issues.

Longer projections (to 2300) show even larger spreads (e.g., −0.6 to +4.4 m under high emissions in some ensembles), with melt sensitivity and dynamic response factors explaining much of the variance.

How ISMIP6 addressed (and continues to inform) these issues

The protocol used “standard” experiments with prescribed melt parameterizations (sampling low/mid/high parameter values) and “open” experiments allowing groups to use their preferred schemes. Forcing came from a curated subset of CMIP models selected for polar skill. Results fed into IPCC assessments and subsequent studies that refine or extend the ensemble (e.g., more comprehensive sampling of melt sensitivity or climate- ice interactions).

In short, ISMIP6 showed that ice- sheet model structural differences and ice- ocean coupling are the leading sources of uncertainty for Antarctica, while climate forcing and surface processes matter more for Greenland. Reducing these uncertainties requires better observations of sub-shelf melt and grounding-line processes, improved basal sliding and calving physics, higher- resolution models, and tighter constraints from satellite mass- change records.

The machine learning model referenced in the The Conversation article (and the underlying Nature Geoscience study by Yucheng Lin, Robert Kopp, and colleagues) is a physics- informed machine learning emulator trained on the large archive of Antarctic ice- sheet simulations from ISMIP6.

Purpose and design

Full ice- sheet models are computationally expensive, one simulation can take days on a supercomputer,so exploring every combination of physical assumptions (emissions scenarios, climate model forcings, basal sliding laws, ice-shelf melt parameterizations, etc.) is impractical. The emulator acts as a fast surrogate:

• It learns the mapping from each simulation’s physical assumptions/inputs to its outcomes (primarily sea- level contribution or ice- mass change).
• Once trained, it evaluates in a fraction of a second while retaining accuracy comparable to the original numerical models.
• This enables running millions of combinations and filtering hundreds of thousands of possible futures against satellite gravity observations (GRACE/GRACE-FO mass-change data since 2002).

Only scenarios consistent with the observed ice- sheet behavior are retained. This “transient calibration” or bias- correction step is central to the study’s finding that net Antarctic ice loss this century is locked in even under low- warming pathways.

Key capabilities highlighted in the work

• Systematic quantification of cascading uncertainties: It traces how each physical assumption propagates into spatiotemporally variable sea-level rise uncertainties.
• Bias correction: By calibrating to the observed long- term Antarctic mass- loss rate (approximately 0.44 mm/yr sea- level equivalent, climate- variability corrected, 2002- 2021), the approach reduces projection uncertainty (by ~30- 42% in the 5th- 95th percentile range in related analyses) and increases median sea- level contributions relative to uncorrected ensembles.
• Identification of dominant uncertainty drivers: AOGCM (climate model) selection, basal sliding laws, and ice- shelf basal melt parameterizations emerge as primary sources.
• Scenario dependence and tail risks: Higher emissions produce greater mass loss (with high probability). Under very high- emissions pathways, the upper tail reaches ~25 cm of Antarctic contribution by 2100 in some bias- corrected results.

The preprint describing this work is titled “Bias-corrected Antarctic sea-level projections reveal greater impact of historical warming” (Lin et al.). It emphasizes that historical warming has already committed the Antarctic Ice Sheet to continued net mass loss through 2100 with high probability (P ≥ 0.92 even under aggressive mitigation such as SSP1-1.9), unlike some prior uncorrected ensembles that allowed for possible net gain.

Broader context of ML emulators for ice sheets

Similar techniques have been applied elsewhere in the ISMIP6 ecosystem and related research:

• Neural networks, LSTMs, Gaussian processes, and graph convolutional networks have been used to emulate ISMIP6 ensembles, regional sea- level fingerprints (GRD effects), surface mass balance, or specific glaciers.
• “Physics- informed” approaches incorporate physical constraints (mass conservation, flow laws, etc.) rather than treating the problem as a pure black-box regression, improving physical consistency and data efficiency.

In this study, the emulator’s speed and ability to perform large- scale sampling and observational constraint are what allowed the authors to conclude that extra snowfall will not reverse the locked- in ice loss and that emissions reductions still matter for limiting the rate and magnitude of Antarctic contributions to sea- level rise.

Committed Antarctic Ice Sheet mass loss by the end of the twenty-first century

The Antarctic Ice Sheet (AIS) is the largest source of uncertainty in sea- level rise projections. Uncertainties cascade through emissions scenarios, atmosphere– ocean general circulation models (AOGCMs), ice-sheet dynamics, and sea- level physics. Traditional intercomparison ensembles can be biased toward common modelling choices regardless of observational consistency.

This study uses a machine- learning emulation framework trained on ISMIP6 simulations to quantify how each individual physical assumption propagates into projection uncertainty. It also applies Bayesian calibration against satellite observations to reduce bias.

Key findings:

• It is very likely (≥0.92 probability) that the AIS is committed to net mass loss by 2100, even under aggressive emissions- reduction scenarios.
• Higher emissions drive greater Antarctic mass loss by 2100 (≥0.89 probability), elevating near- term coastal risks.
• Under very high- emissions scenarios, cascading mechanisms consistent with satellite observations could produce up to 25.4 cm of sea- level rise by 2100 (95th percentile; median = 15.7 cm).
• Managing these risks requires rapid emissions reductions plus better constraints on AOGCM selection, basal sliding laws, and ice- shelf melt parameterizations (the primary uncertainty drivers identified).

Core approach

The authors built a physics- informed machine-learning emulator on the ISMIP6 archive. This allowed them to:

• Rapidly explore millions of combinations of modelling assumptions.
• Systematically attribute uncertainty to specific choices (e.g., sliding laws ~21%, ice-shelf melt parameterizations ~20% of variance in raw simulations).
• Perform Bayesian transient calibration against GRACE/GRACE-FO satellite mass-change observations (2002 onward).
• Produce bias-corrected, probabilistic projections that better reflect historical performance.

They examined both raw and drift- corrected simulations and mapped how uncertainties cascade from global mean surface temperature → AOGCM choice → ice-sheet model features → regional sea-level fingerprints.

This is the peer- reviewed scientific paper underlying the The Conversation article by Yucheng Lin and Robert Kopp that started this discussion.

Title: Committed Antarctic Ice Sheet mass loss by the end of the twenty- first century

Journal information: Nature Geoscience

DOI: 10.1038/s41561-026-02102-1

Published: 30 September 2026

Provided: The Conversation

Authors: Yucheng Lin,
Xuebin Zhang,
Nicholas R. Golledge,
Robert E. Kopp,
John A. Church,
Yi Jin,
Chen Zhao &
Chris R. Stokes 

Abstract

The Antarctic Ice Sheet is the largest source of uncertainty in sea-level rise projections, with uncertainties propagating through emissions scenarios, atmosphere–ocean general circulation models, ice-sheet dynamics and sea-level physics. In ice-sheet model intercomparison exercises, these uncertainties—typically attributed to intermodel differences—stem from modelling choices that may bias ensembles towards commonly adopted approaches regardless of observational consistency. Here we quantify how each individual physical assumption cascades into projection uncertainty using a machine-learning emulation framework, which also enables Bayesian calibration against satellite observations to reduce projection bias. Our results suggest it is very likely (≥0.92 probability) that the Antarctic Ice Sheet is committed to twenty-first-century mass loss, even under aggressive emissions-reduction scenarios. Higher emissions drive greater Antarctic mass loss by 2100 (≥0.89 probability), directly elevating near-term coastal risks. Under very high-emissions scenarios, we identify cascading mechanisms that could produce up to 25.4 cm of sea-level rise by 2100 (95th percentile; median = 15.7 cm) while remaining consistent with satellite observations. Effective management of these risks to densely populated coastal communities requires rapid emissions reductions and improved constraints on climate model selection, sliding laws and ice-shelf melt parameterizations.