Palaeoclimate constraints narrow uncertainty in future Arctic winter sea ice projections and point to lower ice extent

In March 2025, Arctic winter sea ice reached the lowest annual maximum extent in the satellite-era record, and the press release accompanying a new study notes that the Arctic matched that record again in 2026. Yet winter is the season in which climate models disagree most about the future. A study in Nature Geoscience now uses sediment evidence from about 6,000 years ago to test those models and narrow the disagreement.

What the researchers did

Most attention on Arctic ice loss has gone to summer. Winter ice, however, regulates how much heat and moisture leave the ocean for the atmosphere, and the authors note that models show their largest biases and uncertainties in that season. Historical observations alone are a short record strongly shaped by internal variability, which limits how well models can be tested against them.

The team, from the University of Cambridge and Mexico's CIMAT, instead turned to the mid-Holocene (5–7 thousand years ago). Then, orbital changes enhanced Northern Hemisphere summer insolation and produced an Arctic radiative forcing the authors describe as comparable to anticipated future greenhouse gas forcings. They compiled what they describe as the largest compilation of palaeo-sea ice biomarker records so far: 46 observations, made up of 27 ln(PBIP25) and 19 ln(PDIP25) constraints, spanning roughly 54–83° N. The index is based on IP25, a biomarker produced by algae associated with sea ice, relative to sterols from open-water phytoplankton.

They compared these records with mid-Holocene simulations from 27 PMIP3 and PMIP4 climate models. A proxy system model called BaySIC translated each model's simulated sea ice concentration into proxy units, so models and sediment data could be compared quantitatively while accounting for proxy uncertainty and for how the proxy's seasonal bias shifts with the timing of ice breakup.

A Bayesian weighting scheme then assigned each model a weight that maximised the likelihood that the weighted multi-model mean matched the proxy observations at all sites. The scheme implicitly handles model interdependence: related models tend to share weight. The weights were first derived from historical performance (1950–1999) and then updated with the mid-Holocene data. They were then applied to historical simulations and to future projections under three scenarios (SSP1, SSP2 and SSP5, or RCP2.6, 4.5 and 8.5).

What they found

The unweighted ensemble simulated too much winter ice. In the authors' words, collectively the models predicted more winter sea ice than the ancient evidence supported.

Weighting changed that. The weighted multi-model mean for March sea ice extent aligned more closely with observations than the unweighted mean in all tests, and the uncertainty range shrank by about 6.3 million km², or 85%, relative to the 95% confidence interval of the inter-model spread. Sea ice extent was never used within the weighting itself, so this improvement is an emergent result. The mid-Holocene data gave stronger constraints than the historical data alone. Combining both periods effectively eliminated the mean positive bias in winter extent, leaving a residual of 0.016 million km² (a 0.099% error).

“When we apply the model weights to historical simulations, we completely eliminated the positive bias.”

— Crystal Fu, Scott Polar Research Institute, University of Cambridge

For future projections, the weighted mean was consistently lower than the unweighted mean across all scenarios, by roughly 4–6% for 2090–2099. The persistent bias removed was nearly 1 million km². Under SSP5, the end-of-century projection range was tightened by over 10 million km². Observed interannual variations in the historical record largely stayed within the narrowed bounds, which the authors take to suggest that the weighting does not produce overconfident predictions.

Spatially, the weighted ensemble showed higher probabilities of ice loss in the Barents, Greenland and Labrador Seas, corresponding to northward ice-edge shifts of up to 5.5° of latitude.

The authors also used relationships between sea ice extent and weather in the model ensemble to explore knock-on effects. Under SSP5 (2070–2099), late winter–early spring snowfall over southeastern Greenland may be underestimated by more than 30%, and end-of-century winter–spring precipitation in the Mediterranean may be twice the amount anticipated by the unweighted CMIP ensemble. The paper stresses that more research is needed to verify these relationships.

“The future projections similarly end up being lower than previous estimates, by about a million square kilometers.”

— Crystal Fu, Scott Polar Research Institute, University of Cambridge

Why it matters

Winter sea ice influences atmospheric circulation, and the paper cites links between its loss and colder winters and increased snowfall across Eurasia and North America, storm-track shifts, and extreme Mediterranean precipitation. If winter ice is declining faster than unweighted models suggest, these connections may matter more than assumed. The authors say the results point to an even higher probability of cold and wet winters in the mid-latitudes, while noting that those links need further verification.

More snowfall over Greenland could affect ice sheet stability, and more Mediterranean rainfall could raise flood risk in southern Europe. For Arctic ecosystems and the communities that depend on them, the authors argue that winter retreat is probably more severe than current models suggest.

Methodologically, the study offers a quantitative way to use palaeoclimate evidence, which the authors say has remained largely qualitative, to test models out of sample.

“It doesn't matter if the warming is natural variation or caused by human activity: if it warms, sea ice is going to melt.”

— Crystal Fu, Scott Polar Research Institute, University of Cambridge

What this study can't tell us

• The weights are not skill scores. Because the ensemble was biased towards too much ice, models with less ice at the core sites were upweighted to compensate. The authors caution against reading higher weight as better skill.
• The method works best when reality lies within the ensemble's spread. The authors say this appears to be the case for winter ice and CMIP models, but if the truth fell outside, no weighting could converge.
• The projections assume that the sea ice sensitivity seen in the past, and the ensemble diversity, persist into the near future.
• The proxy network is sparse and uneven, with informative sites concentrated near the winter ice edge, such as the Labrador and Greenland Seas. The proxy has a seasonal bias towards late winter and early spring, so the results apply to March extent, not to other seasons.
• Proxy calibration is highly uncertain. One site (MD99-2274) had an outlying value; it was retained, with an exclusion sensitivity test reported.
• The mid-Holocene warming was orbitally driven, not caused by greenhouse gases. The researchers argue that sea ice responds by the same physical principles, but this is an assumption underlying the approach.
• The Greenland snowfall and Mediterranean precipitation results come from regression relationships in model output, not from direct simulation, and the paper itself says they need verification.