Iron, Not Nutrients, Drove the Ice Age Ocean’s Carbon Grab, Massive Model Ensemble Shows

Some 20,000 years ago, at the height of the last ice age, the air above Antarctica held roughly 90 parts per million less carbon dioxide than it did before the industrial revolution. Ice cores have recorded that drop with exquisite precision, yet the ocean processes responsible for locking away that carbon have resisted a definitive explanation for decades. Now a team of marine biogeochemists has run one of the most exhaustive modeling experiments ever attempted on the problem, and their verdict is striking: the trace metal iron, far more than the major nutrients nitrate and phosphate, controlled how much carbon the glacial ocean could sequester through the soft-tissue biological carbon pump.

The study, led by Chia-Te Chien of GEOMAR Helmholtz Centre for Ocean Research Kiel and National Taiwan University, together with Markus Pahlow, Christopher J. Somes, Markus Schartau, and Andreas Oschlies, is published in the journal Earth System Dynamics. Rather than running a single simulation with one best-guess set of model parameters, the team built an ensemble of 480 simulations of the Last Glacial Maximum, systematically varying both the environmental boundary conditions of the glacial world and the physiological parameters of the plankton ecosystem model embedded within an Earth system model. The design was aimed squarely at a blind spot that has haunted glacial carbon cycle modeling: most previous studies used a single parameter set calibrated against modern observations and simply assumed those values would hold in a radically different ocean.

The technical foundation of the work is the optimality-based plankton ecosystem model, or OPEM, coupled to the University of Victoria Earth system model of intermediate complexity. Unlike conventional models that fix the elemental composition of marine organic matter to the canonical Redfield ratio of 106 carbon to 16 nitrogen to 1 phosphorus, OPEM allows phytoplankton to flexibly adjust their carbon-to-nitrogen and carbon-to-phosphorus ratios in response to nutrient stress and light. This variable stoichiometry matters enormously for carbon sequestration, because if glacial plankton built organic matter with more carbon per unit of nutrient, the same nutrient supply could export more carbon into the deep ocean and pull down atmospheric carbon dioxide more efficiently.

The ensemble architecture was deliberately layered. The researchers first generated 600 parameter sets, each representing a unique combination of values for 19 parameters governing plankton physiology, detritus remineralization, and particle sinking. Each of the 600 was spun up for more than 10,000 years under pre-industrial conditions with a prescribed atmospheric carbon dioxide level of 284.3 parts per million. A likelihood-based cost function then scored each simulation against observations of phosphate, nitrate, and oxygen from the World Ocean Atlas across 17 biomes, penalizing mismatches in both mean values and spatial and temporal variability. The 20 parameter sets with the lowest costs, all within 1.3-fold of the best, were carried forward. Each was then run under 24 different combinations of glacial boundary conditions, producing the 480-simulation ensemble.

Those boundary conditions captured the major hypothesized differences between the glacial and pre-industrial ocean. On the physical side, the team applied Last Glacial Maximum orbital parameters, wind stress patterns drawn from PMIP3 models featuring a northward shift and intensification of Southern Ocean westerlies, and a halving of meridional moisture diffusivity over the Southern Ocean, which stratifies the surface and weakens deep water formation. On the biogeochemical side, the 120-meter drop in sea level exposed continental shelves, cutting benthic denitrification by roughly 40 percent and slashing sedimentary iron input by about 80 percent, while terrestrial erosion boosted the global phosphate inventory by 15 percent and glacial dust quadrupled atmospheric iron deposition to 6.1 gigamoles of iron per year.

The results were unambiguous about which lever mattered most. Iron supply, whether arriving as dust or released from sediments, exerted the most profound influence on marine biogeochemistry and atmospheric carbon dioxide. Quadrupling atmospheric iron deposition lowered carbon dioxide by about 26 parts per million on its own, while cutting sedimentary iron raised it by roughly 60 parts per million, a swing so large it dwarfed every other forcing tested. By contrast, increasing the phosphate inventory or reducing benthic denitrification, the two classic macronutrient hypotheses, produced only marginal effects. The reason lies in co-limitation: adding one major nutrient simply aggravates limitation by the others, and the model’s flexible stoichiometry partially offsets the gain by allowing plankton to incorporate relatively less carbon per unit of nutrient when nutrients become abundant.

The sedimentary iron finding carries a provocative implication. Because falling sea level stripped away shallow seafloor sources, the total dissolved iron inventory in the glacial ocean was actually about 1 percent lower than before the ice age, even as dust deposition soared. Lower temperatures also reduced productivity and therefore the organic carbon flux reaching the seafloor, further starving sedimentary iron release. Yet atmospheric carbon dioxide still fell, driven primarily by a spatial redistribution of surface dissolved inorganic carbon and altered air-sea fluxes rather than by a wholesale increase in oceanic iron fertilization. Earlier modeling studies that considered only enhanced dust deposition, without the sedimentary decline, may have substantially overestimated the glacial iron effect.

The ensemble also exposed a sobering uncertainty. Under full glacial conditions, atmospheric carbon dioxide decreased by 36 to 58 parts per million across the 20 parameter sets, a spread equal to roughly half the mean drawdown of 43.5 parts per million, even though all 20 sets reproduced pre-industrial biogeochemistry about equally well. Statistical analysis traced the variance to parameters governing phytoplankton nitrogen subsistence, temperature-dependent mortality, and zooplankton grazing, and revealed that strongly correlated parameter pairs could mask or amplify the carbon dioxide response depending on how their values co-varied. Notably, the drawdown achieved by each simulation was unrelated to its cost function score, meaning that a model fitting modern data well tells you almost nothing about how it will behave in a glacial ocean.

The full ensemble accounted for only about half of the observed 90 parts per million glacial drawdown, leaving room for processes outside the model’s scope, such as brine-induced stratification around Antarctica, air-sea carbon disequilibrium, and a larger terrestrial carbon transfer than the simulations produced. Still, the physical forcings alone, colder temperatures plus altered winds and moisture diffusivity, contributed about 35 parts per million, with the biological and stoichiometric effects adding the rest. The team estimates that flexible carbon-to-nutrient ratios contributed an additional 16 to 17 parts per million of drawdown when iron deposition alone is considered, consistent with earlier theoretical estimates.

Beyond solving an ice age puzzle, the findings speak directly to one of the most debated proposals for engineered carbon removal: ocean iron fertilization. The glacial ocean serves as a natural experiment in what happens when iron delivery to the sea surges, and the model’s behavior, including the nutrient-robbing effect in which fertilized regions deplete macronutrients downstream, mirrors the central criticism of iron fertilization as a climate intervention. Intriguingly, the variable stoichiometry feedback, which sustains productivity even as macronutrients decline, does not exist in fixed-ratio models and could partially buffer that limitation. The authors argue that the same parameter and boundary condition uncertainties that complicate glacial simulations apply equally to assessing carbon dioxide removal strategies, making ensemble approaches like this one essential for any credible accounting of what the ocean can, and cannot, be persuaded to store.

Subject of Research: Ensemble modeling of Last Glacial Maximum marine biogeochemistry and the role of the soft-tissue biological carbon pump in glacial atmospheric CO2 drawdown

Article Title: Ensemble simulation of the Last Glacial Maximum marine biogeochemistry and atmospheric pCO2 drawdown due to the soft-tissue biological carbon pump

Article References: Chien, C.-T., Pahlow, M., Somes, C. J., Schartau, M., & Oschlies, A. (2026). Ensemble simulation of the Last Glacial Maximum marine biogeochemistry and atmospheric p CO 2 drawdown due to the soft-tissue biological carbon pump. Earth System Dynamics, 17(5), 1277-1297. https://doi.org/10.5194/esd-17-1277-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1277-2026

Keywords: Last Glacial Maximum, biological carbon pump, iron fertilization, marine biogeochemistry, atmospheric CO2, paleoceanography, Earth system modeling, variable stoichiometry, ocean circulation, nutrient limitation, carbon sequestration, ensemble simulation