A team led by Abigail Swann of the University of Washington, together with Charles Koven of Lawrence Berkeley National Laboratory and colleagues across fifteen institutions, analyzed simulations from ten Earth system models run under the new flat10MIP experimental protocol, an idealized framework adopted for the Coupled Model Intercomparison Project Phase 7. Unlike earlier experiments that prescribed atmospheric carbon dioxide concentrations, flat10MIP specifies emissions directly, letting each model calculate how much carbon dioxide actually accumulates in the air as its land and ocean components absorb or release carbon. That design allows scientists to probe two crucial climate metrics, the transient climate response to cumulative emissions and the zero emissions commitment, while simultaneously tracking where carbon ends up on land.

The experimental architecture is elegantly simple. In the core flat10 experiment, carbon dioxide is emitted at a constant 10 petagrams of carbon per year, close to present-day rates, for at least a century, accumulating 1000 petagrams of cumulative emissions. From that point, two branches diverge. In the zero-emissions branch, emissions drop instantly to zero and the simulation runs for another 200 years. In the carbon-dioxide-removal branch, emissions decline steadily, turn negative, and eventually reach minus 10 petagrams per year, so that by year 300 the cumulative emissions over the whole experiment return to zero. Ten models, including ACCESS-ESM1-5, CESM2, NorESM2-LM, GFDL-ESM4, UKESM1.2, MPI-ESM1-2-LR, MIROC-ES2L, CNRM-ESM2-2, NASA-GISS-E2.1-G-CC2, and HadCM3LC-Bris, completed all three scenarios.

During the positive emissions phase, the models agreed on a broad pattern: land gained carbon, on average about 246 petagrams by year 100, and the majority of that gain, roughly 60 percent, landed in vegetation rather than soil. The mechanism is familiar. Elevated atmospheric carbon dioxide enhances photosynthesis, a process represented in every model, and the carbon flows first into leaves, wood, and roots before slowly leaking into litter and soil pools. The gains were largest in the tropics and mid-latitudes, and in most models carbon accumulated on land nearly linearly with cumulative emissions, at least through the first 1000 petagrams. A few models, notably ACCESS-ESM1-5, showed the accumulation rate tapering off earlier, as tropical gross primary productivity began to saturate under heat and moisture stress.

Then the picture darkened. When emissions were cut to zero, atmospheric carbon dioxide continued to fall as land and ocean sinks kept drawing it down, but the land’s internal accounting shifted dramatically. In seven of the ten models, tropical latitudes lost carbon during the net-zero phase, and in every single model, tropical vegetation carbon declined. Meanwhile, mid- and high-latitude regions kept gaining carbon, mostly in soils. The explanation is a cruel asymmetry. As carbon dioxide concentrations fall, the fertilization effect that boosted tropical growth partially reverses, yet temperatures remain high, accelerating respiration and stressing plants in the hottest regions. The tropics, which gained the most during the emissions phase, become the biggest losers once emissions stop.

The negative emissions branch told a similar story at a different baseline. By the point of cumulative zero emissions, when all the carbon ever emitted had hypothetically been removed, vegetation carbon had declined in most models and tropical carbon had fallen in all of them, relative to the pre-industrial starting state. Mid-latitude soils and high-latitude vegetation and soils, by contrast, often ended up above their initial stocks. The similarity between the net-zero and cumulative-zero responses suggests that the same processes, faster turnover of vegetation under heat stress and the reversal of carbon dioxide fertilization, operate in both cases, just from different starting points.

Perhaps the most striking result concerns timing. Peak atmospheric carbon dioxide arrived about 15 years before net-zero emissions, and global temperature peaked roughly 6 years after net-zero. Vegetation carbon peaked a remarkably consistent 10 years after net-zero, with a spread of only about 7 years across the ensemble, suggesting that the processes governing vegetation growth and decay are represented similarly across models. Soil carbon was another matter entirely. Its peak arrived on average 33 years after net-zero, but ranged from 9 years before to a staggering 126 years after, with GFDL-ESM4’s late peak reflecting its unusual soil model in which turnover slows as productivity rises. Because the zero emissions commitment depends on century-scale carbon sink behavior, this soil carbon disagreement directly undermines confidence in projecting how much warming will persist after emissions stop.

The study also exposed uncomfortable gaps in what the models actually contain. Pre-industrial land carbon stocks varied enormously, from 978 petagrams in GFDM-ESM4 to 3119 petagrams in NorESM2-LM, against an IPCC benchmark of about 3350 petagrams for the present day, and many models appear to hold too little soil carbon, particularly in high latitudes where permafrost processes are missing from most land components. Only three models resolve soil carbon by depth, and just two of those represent permafrost; both showed stagnation or losses of high-latitude soil carbon under zero emissions, consistent with warming-driven permafrost thaw. Even more puzzling, none of the structural differences the team examined, whether nutrient limitation, dynamic vegetation, fire, or soil complexity, correlated with model behavior. Models with and without nitrogen cycling, with one soil pool or eighty, spread across the full range of outcomes, implying that parametric uncertainty and unrepresented processes swamp the visible architecture.

That last point carries a warning for carbon dioxide removal efforts. The models in this study lack many mechanisms that would make tropical losses worse: heat damage to enzymes and reproduction, hydraulic failure and mortality, pests and pathogens, and some fire-related disturbances. The simulated tropical declines are therefore likely underestimates. And because most afforestation and reforestation projects are planned for low latitudes, precisely the regions where the models show carbon being gained and then lost, the study raises the prospect that tree-planting schemes could see their stored carbon evaporate under the sustained heat of a zero-emissions world. The idealized flat10MIP scenarios also impose removal as a boundary condition rather than modeling the physical side effects of specific removal methods, which could add further complications.

The takeaway is sobering but precise. Tropical carbon is the most likely to be gained during rising emissions and the most likely to be lost under zero, declining, and negative emissions, making it a fragile foundation for net-zero accounting. Mid- and high-latitude soils, meanwhile, behave more reliably but on timescales so uncertain and so model-dependent that they dominate the spread in the zero emissions commitment. As nations design removal portfolios and net-zero pledges, the message from this ten-model ensemble is clear: stopping emissions does not stop the carbon cycle from rearranging itself, and the places we most depend on to store carbon may be the places least able to keep it.

Subject of Research: Land carbon sink responses to positive, zero, and negative CO2 emissions across Earth system models

Article Title: Land carbon response to positive, zero, and negative CO2 emissions across Earth system models

Article References: Swann, A. L. S., Koven, C. D., Proistosecu, C., Fisher, R. A., Sanderson, B. M., Brovkin, V., Hajima, T., Jones, C. D., Kiang, N. Y., Lawrence, D. M., Liddicoat, S., Liddy, H., Romanou, A., Séférian, R., Sentman, L. T., Steinert, N. J., Tjiputra, J., & Ziehn, T. (2026). Land carbon response to positive, zero, and negative CO 2 emissions across Earth system models. Earth System Dynamics, 17(5), 1237-1275. https://doi.org/10.5194/esd-17-1237-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1237-2026

Keywords: land carbon sink, Earth system models, net-zero emissions, negative emissions, tropical forests, soil carbon, CO2 fertilization, flat10MIP, zero emissions commitment, carbon dioxide removal, permafrost, climate modeling