Ancient Ocean Clues Reveal a Blind Spot in Climate Models: Melting Ice Sheets and the Labrador Sea

Deep beneath the stormy surface of the North Atlantic, one of the planet’s most important climate machinery quietly operates each winter. In a handful of regions — the Labrador Sea, the Irminger Sea and the Greenland Sea — surface waters cool, become denser, and plunge downward in violent convective chimneys that help drive the Atlantic Meridional Overturning Circulation, the great conveyor belt of currents that regulates climate on both sides of the ocean. The depth to which this winter mixing penetrates, known as the mixed layer depth, is a critical diagnostic of how well climate models capture these processes. Now a new study published in the journal Climate of the Past has put fifteen state-of-the-art climate models through an unusual test: simulating the mid-Holocene, a warm period roughly 6,000 years ago, and comparing their results against reconstructions of past ocean conditions extracted from the seafloor. The verdict is sobering, and it points to a missing ingredient that could matter enormously for our future.

The research, led by Xiner Wu of the Université du Québec à Montréal together with Anne de Vernal and Paul G. Myers, forms part of the Paleoclimate Modelling Intercomparison Project, or PMIP4, a coordinated international effort that uses past climates as out-of-sample tests for the same numerical models that project future warming. The mid-Holocene is an ideal target for such a test. At 6,000 years before present, continental geometry was essentially identical to today’s and greenhouse gas concentrations were close to preindustrial levels, but subtle differences in Earth’s orbit redistributed sunlight across latitudes and seasons, producing warmer Northern Hemisphere summers and cooler winters. If a model cannot faithfully reproduce how the ocean responded to that modest, well-documented forcing, confidence in its response to the far larger forcing of the coming century is necessarily shaken.

To establish what the ocean actually did during that interval, the team turned to an ingenious biological archive: dinoflagellate cysts. These microscopic, extraordinarily resistant structures are produced by single-celled plankton in the surface ocean, and the species composition of the assemblages preserved in layered marine sediments reflects the environmental conditions under which the organisms lived. Crucially, recent work by the same group showed that dinocyst assemblages respond to wintertime deepening of the mixed layer associated with deep convection, allowing quantitative reconstruction of past mixed layer depths. Drawing on twenty-four Holocene sediment records from the modern deep convection regions of the subpolar North Atlantic, the researchers applied a modern analogue technique with a demonstrated error of about 41 metres for winter mixed layer depth — a precision that makes meaningful model comparison possible for the first time.

The comparison was conducted using anomalies rather than absolute values, subtracting the preindustrial baseline from the mid-Holocene simulation for each model and from each proxy record. This sidesteps systematic biases that plague both models and reconstructions. The team recomputed mixed layer depth from modelled temperature and salinity fields using a standardized density threshold, since different modelling groups define the mixed layer in incompatible ways, and they evaluated agreement with two complementary statistics: root mean square error, which measures the magnitude of disagreement, and Cohen’s kappa, which tests whether a model captures the direction of change — deeper or shallower — better than random chance would allow.

The results revealed a strikingly wide spread across the fifteen PMIP4 models. Some, such as ACCESS-ESM1.5, produced very large winter mixed layer anomalies, while others, like INM-CM4-8, barely responded at all. A recurring pattern emerged in several models: deepening in the Labrador Sea alongside a northwestward shift of the deep-mixing zone in the Nordic Seas, a response the authors link to reduced winter sea ice. Higher Northern Hemisphere summer insolation melts more Arctic sea ice, reducing its winter export into the subpolar North Atlantic; where ice retreats, the ocean loses heat to the atmosphere more readily, buoyancy fluxes intensify, and convection deepens — a self-reinforcing feedback in which rising warm, salty subsurface water further suppresses local ice formation.

Yet when the models were confronted with the proxy data, the picture fractured along a geographic line that carries real weight for future projections. In the Nordic Seas, the ensemble mean aligned broadly with the reconstructions, which indicate increased winter mixing at western and southeastern sites and shallower mixing in the east. In the Labrador Sea, however, the disagreement was stark and systematic. All four proxy sites there recorded shallower winter mixed layers during the mid-Holocene, while nearly every model — from both the current PMIP4 generation and the older PMIP3 ensemble — simulated the opposite, a deeper mixed layer. Only one model, MRI-ESM2.0, achieved statistically significant agreement with the direction of reconstructed change, and even the ensemble mean failed the test, demonstrating that averaging models together does not guarantee better performance.

The explanation the authors propose is meltwater — a forcing that the mid-Holocene experiment deliberately omits. The protocol prescribes modern ice sheets in equilibrium, but the real Labrador Sea of 6,000 years ago was anything but equilibrium. Residual ice caps lingered in eastern North America until roughly 6,000 to 5,500 years ago, and the Greenland Ice Sheet did not reach its minimum extent until about 4,000 years ago. Marine records show persistently low sea surface salinity in the Labrador Sea through the Early and Middle Holocene, and earlier proxy work suggests a buoyant, stratified surface layer that suppressed deep-water formation until well into the Holocene. Freshwater discharged from melting ice sheets would have capped the ocean with a light, low-density lid, preventing the dense overturning that models, deprived of this input, freely simulate. Notably, the Nordic Seas appear largely insulated from this effect, consistent with modern studies showing that most meltwater from the East Greenland Current ends up in the Labrador Sea and Baffin Bay rather than crossing eastward.

Transient simulations of the last deglaciation, which do include time-varying meltwater fluxes, lend qualified support to this hypothesis: they overestimate Labrador Sea mixed layer depth less severely than the equilibrium experiments. But the deglaciation models disagree wildly with each other and with the records, and none reproduces the reconstructed late-Holocene trend toward modern-like convection in the Labrador Sea, suggesting that the ocean’s memory in these long simulations may be unrealistically long. The team also found that model details matter enormously — different tuning of the same model changed the mixed layer response more than different ice sheet reconstructions did, and swapping the atmospheric component of a coupled model could reverse the sign of simulated change, underscoring that air-sea-ice interactions, not ocean physics alone, govern open-ocean convection.

Why should this matter beyond the paleoclimate community? Because the Labrador Sea is one of the few places on Earth where the atmosphere can reach down and set the deep ocean in motion, and because the same mechanism — ice sheet melt freshening the surface — is now operating in reverse time. As Greenland’s ice sheet accelerates its melt under global warming, the buoyant lid that suppressed convection 6,000 years ago could reassemble, with consequences for the overturning circulation, European climate and ocean carbon uptake. The study’s central message is that deep convection in the Labrador Sea may be uniquely and disproportionately sensitive to meltwater, and that current climate models, which handle this forcing imperfectly even in purpose-built paleoclimate experiments, may be underestimating that vulnerability. The mid-Holocene, it turns out, was not just an ancient curiosity but a rehearsal — and the models, on this crucial measure, are still learning their lines.

Subject of Research: Evaluation of mixed layer depth in PMIP4 mid-Holocene climate simulations against proxy reconstructions in North Atlantic deep convection regions

Article Title: Mixed layer depth in the PMIP4 midHolocene simulations: comparison to proxy data in North Atlantic deep convection regions

Article References: Wu, X., de Vernal, A., & Myers, P. G. (2026). Mixed layer depth in the PMIP4 midHolocene simulations: comparison to proxy data in North Atlantic deep convection regions. Climate of the Past, 22(9), 1609-1629. https://doi.org/10.5194/cp-22-1609-2026

Image Credits: AI Generated

DOI: 10.5194/cp-22-1609-2026

Keywords: mixed layer depth, PMIP4, mid-Holocene, Labrador Sea, Nordic Seas, deep convection, AMOC, dinocyst proxies, meltwater forcing, sea ice, climate models, paleoclimate