The study, published in Nature Geoscience by C. Y. Fu of the University of Cambridge’s Scott Polar Research Institute, M. B. Osman of the University of Cambridge, and M. A. Aquino-López of the Centro de Investigación en Matemáticas in Mexico, tackles a problem that has plagued the latest generations of climate models. Although winter sea ice extent shows the least year-to-year variability in the observational record, the disagreement between models is widest precisely at this time of year, with the largest spread seen in March. The absolute bias in wintertime sea ice in the CMIP5 and CMIP6 model archives is substantially greater than the bias in September, meaning that the models collectively paint too much ice on the winter Arctic Ocean — and nobody knows exactly which models are to blame.

The consequences of this blind spot are not academic. Winter ice cover regulates the release of heat and moisture from the ocean to the atmosphere, and its reduction has been implicated in colder winters and heavier snowfall across Eurasia and North America, intensified Northern Hemisphere cyclones and shifting storm tracks, and extreme precipitation events in the Mediterranean. The 2025 annual maximum Arctic sea ice extent was the lowest on record, bringing the long-term winter decline into sharp focus. Yet efforts to refine projections have concentrated overwhelmingly on the summer minimum, potentially biasing model evaluation toward the season the models already handle relatively well.

Fu and colleagues’ insight was that the past can serve as an out-of-sample test of model performance. During the mid-Holocene, roughly six thousand years ago, orbitally driven changes enhanced summer insolation in the Northern Hemisphere, producing an increased annual mean radiative forcing in the Arctic comparable to the greenhouse gas forcing anticipated in the coming decades. Previous analyses of the Palaeoclimate Modeling Intercomparison Project have shown that Arctic warming in the mid-Holocene and in the future is driven by common physical processes from autumn through winter, with similar contributors to the multi-model spread across all seasons. Models that capture past sea ice dynamics, the reasoning goes, should also perform well in future projections — both periods being effectively unseen by the models during their development.

The obstacle has always been quantitative data. Palaeoclimate applications of sea ice proxies have remained largely qualitative, hampered by sparse records and large reconstruction uncertainties. To overcome this, the team compiled a new Arctic-wide database of mid-Holocene biomarker records based on the ln(PIP25) index, a fully quantitative sea ice proxy built on the relative concentrations of IP25 — an ice proxy with 25 carbon atoms produced by sea ice diatoms — and associated phytoplankton sterols such as brassicasterol and dinosterol, which serve as open-water indicators. The resulting compilation of 46 observations, spanning latitudes from roughly 54 to 83 degrees north, is to the researchers’ knowledge the largest assembly of palaeo-sea ice biomarker records ever assembled.

The technical heart of the study lies in a recently developed proxy system model called BaySIC, which translates simulated sea ice concentrations from 27 climate models of the PMIP3 and PMIP4 mid-Holocene ensembles into proxy units for direct, fully quantitative comparison with the sediment observations. Crucially, the forward modelling accounts for spatiotemporal variations in proxy seasonality: because the ln(PIP25) index carries a seasonal bias toward late winter and early spring, when sea ice extent reaches its annual maximum, the team used climatological sea ice concentrations from the months corresponding to each model’s simulated timing of local sea ice breakup rather than fixed calendar months. This matters enormously in high-latitude regions, where the timing of ice breakup can shift substantially between the modern climate and the mid-Holocene.

On top of this framework, the researchers built a Bayesian palaeoclimate model weighting scheme that adjusts the contribution of each model to the ensemble mean based on proxy evidence. The scheme maximizes the likelihood that the weighted multi-model mean of predicted proxy values reproduces the observations across all core sites, while implicitly accounting for model interdependence — related models that contribute similar information share weight, effectively penalizing redundancy without requiring explicit skill-dependence metrics. The team first derived prior weights from the models’ performance against historical observations from 1950 to 1999, then updated these with the mid-Holocene data to obtain posterior weights that maximize proxy-model agreement across both intervals.

The results are striking. In validation experiments, the weighted multi-model mean aligned with observations far more closely than the unweighted mean, and the uncertainty range was reduced by approximately 6.3 million square kilometres — an 85 percent narrowing — bringing the projection uncertainty down to the level of interannual variability. When historical and mid-Holocene constraints were combined, the weights effectively eliminated the mean positive bias in winter sea ice extent, leaving a residual of just 0.016 million square kilometres, a percentage error of 0.099 percent. The mid-Holocene period alone yielded more effective constraints than the historical record, supporting the notion that out-of-sample evaluation based on past warm periods provides information about model reliability beyond anything the short observational record can offer.

Applying the same weights to future projections of the same models under three emissions scenarios — SSP1, SSP2 and SSP5 in CMIP6, corresponding to RCP2.6, RCP4.5 and RCP8.5 in CMIP5 — the weighted mean winter sea ice extent was consistently lower than the unweighted estimate, by roughly 4 to 6 percent for the decade 2090 to 2099. For the high-emissions SSP5 scenario, the end-of-century projection was tightened by over 10 million square kilometres. The spatial pattern is equally consequential: the weighted projections show higher probabilities of sea ice loss in the Barents, Greenland and Labrador Seas, corresponding to northward ice-edge shifts of up to 5.5 degrees of latitude — regions already identified as vulnerable to Atlantic water intrusion and associated atmosphere-ocean-sea ice feedbacks.

The cascading implications reach far beyond the polar circle. Reduced winter sea ice in the Barents Sea has been linked to cold extremes in Eurasia through local atmospheric heating and weakening of the polar vortex, as well as altered cyclone tracks over the Nordic Seas and increased rainfall in the Mediterranean and East Asia. The team’s regression analyses suggest that late winter and early spring snowfall over Southeastern Greenland may be underestimated by more than 30 percent in current projections, with clear implications for Greenland Ice Sheet mass loss and future sea level rise rates, while end-of-century winter-spring precipitation in the Mediterranean may be twice what the CMIP ensemble anticipates — supporting elevated flood risks under fossil-fuelled development pathways. The palaeoclimate-weighted projections indicate that winter sea ice retreat, and the extreme mid-latitude weather it may fuel, are probably more severe than current models suggest — a finding that demands the same attention long given to the Arctic’s more famous summer decline.

Subject of Research: Cryospheric science

Article Title: Palaeoclimate constraints narrow uncertainty in future Arctic winter sea ice projections

Article References: Fu, C. Y., Osman, M. B., & Aquino-López, M. A. (2026). Palaeoclimate constraints narrow uncertainty in future Arctic winter sea ice projections. Nature Geoscience. https://doi.org/10.1038/s41561-026-02123-w

Image Credits: AI Generated

DOI: 10.1038/s41561-026-02123-w

Keywords: Arctic sea ice, palaeoclimate, climate models, mid-Holocene, Bayesian statistics, CMIP6, Nature Geoscience, winter sea ice, sea ice projections, cryosphere, constraints, narrow