The ocean remains one of Earth’s largest carbon sinks. The difficult question is not whether it absorbs a substantial share of human carbon emissions, but whether scientists can reliably determine how that uptake changes from year to year, from decade to decade and from one region to another.
A new study offers a more detailed view of carbon stored inside the ocean. But it also illustrates why climate measurements require more than an impressive number, a sophisticated algorithm or a map covering the entire planet.
The study, published September 2, 2026, as a discussion-stage preprint in EGUsphere, uses a machine-learning system called ResNet-DIC to reconstruct dissolved inorganic carbon throughout the ocean from the 1990s through 2019. The authors estimate that the ocean contained approximately 37,508 petagrams of dissolved inorganic carbon in 2019, with an average increase of 2.64 petagrams per year across the study period.
They also report pronounced variation in the rate of carbon accumulation, particularly during the early 2000s and in intermediate waters between 300 and 1,000 meters deep. The Atlantic, Indian and Southern oceans contributed strongly to the reconstructed pattern.
That result is potentially important. Carbon entering the ocean does not remain only at the surface. Through circulation, chemistry and biological processes, it can be transported into the interior, where it may remain for decades or longer. A full-depth reconstruction could therefore reveal changes that surface measurements alone cannot show.
What the study establishes
The preprint demonstrates that machine learning can be used to estimate the distribution of dissolved inorganic carbon across regions and depths that have not been measured continuously. The model was trained using observations from the GLODAPv2.2023 database, with atmospheric carbon dioxide, temperature, salinity and location among its predictors.
The approach is useful because the ocean is severely undersampled. Ships and autonomous instruments collect important measurements, but observations are unevenly distributed across basins, depths and years. A reconstruction can make those observations more informative by estimating what lies between them.
It cannot, however, turn an estimate into a direct measurement.
Dissolved inorganic carbon is a broad quantity. It includes carbon associated with human emissions, but it is also affected by natural circulation, biological activity, water-mass formation, remineralization, alkalinity and air-sea exchange. A rise in interior-ocean dissolved inorganic carbon is evidence of changing carbon storage. It is not automatically a precise measurement of the annual amount of human carbon dioxide absorbed from the atmosphere.
That distinction matters because the preprint sometimes places its reconstructed accumulation rates alongside estimates of the anthropogenic ocean carbon sink. The comparison is informative, but the quantities are not interchangeable. Total carbon inventory changes can include redistribution within the ocean as well as newly absorbed atmospheric carbon.
The ocean sink is established. Its stability is not.
The broad scientific conclusion is secure: the ocean absorbs a substantial portion of human-produced carbon dioxide. The UNESCO Intergovernmental Oceanographic Commission’s 2026 assessment places the share at roughly one-quarter of anthropogenic emissions, while emphasizing that estimates differ among models and observational approaches.
Those differences are not evidence that the carbon sink is imaginary. They are evidence that the system is more difficult to measure than a single global percentage suggests.
Surface-ocean carbon observations, ship-based measurements, repeated hydrographic surveys, atmospheric carbon budgets and ocean-biogeochemistry models each observe a different part of the system. They may converge on the existence of substantial uptake while disagreeing about its precise magnitude, timing and regional distribution.
Earlier research has raised the possibility that the ocean’s efficiency in absorbing additional carbon dioxide may change as seawater chemistry and circulation change. A 2023 analysis summarized by NOAA’s Atlantic Oceanographic and Meteorological Laboratory found that the ocean’s sensitivity to rising atmospheric carbon dioxide may have declined between two study periods. The authors also identified uncertainty about natural variability and the role of large-scale circulation.
That is different from showing that the global carbon sink has collapsed. It is also different from showing that the ocean will continue absorbing carbon at a fixed rate.
The reviewers’ objection is methodological, not dismissive
The preprint remains under discussion, and its referee reports identify several questions that bear directly on the strength of its conclusions.
One reviewer questioned whether a roughly 30-year reconstruction can establish clear decadal variability, particularly because the beginning and end of the period may be affected by edge effects. The reviewer also asked whether using globally averaged atmospheric carbon dioxide and relatively limited biogeochemical predictors adequately captures the processes controlling dissolved inorganic carbon.
A second reviewer raised a more fundamental concern: whether the reconstruction is detecting changes in the carbon sink or primarily reproducing changes in water masses, circulation and other physical conditions. Temperature and salinity contain valuable information about the ocean, but they may not capture all of the biological and chemical processes that affect interior carbon.
The reviewer also called for stronger tests using entire cruises or time periods withheld from the model, rather than relying mainly on randomly withheld profiles. That distinction matters. If nearby measurements from the same expedition appear in both the training and testing data, the model may perform well without demonstrating that it can reconstruct genuinely unobserved regions and years.
These criticisms do not show that the reconstruction is wrong. They identify what would need to be demonstrated before its strongest interpretation could be accepted: that the apparent peaks and slowdowns remain present when observation density, predictor choice, validation design and the distinction between total carbon storage and anthropogenic uptake are tested more rigorously.
Why sparse data can create convincing patterns
Machine-learning systems are particularly useful when observations are incomplete. They are also capable of producing smooth, coherent patterns from incomplete information. That is both their strength and their danger.
If the distribution of observations changes over time, the apparent history of the system may partly reflect the changing record rather than the changing ocean. The problem is not unique to machine learning. Climate scientists have long had to reconstruct global conditions from measurements that were geographically uneven, historically discontinuous and collected using changing instruments.
The solution is not to abandon reconstruction. It is to test it against independent observations, alternative model structures, withheld time periods and multiple physical indicators. A reconstruction becomes more credible when different methods converge on the same pattern for independent reasons.
The preprint’s authors provide validation experiments and an uncertainty estimate, but the referee comments show why those tests remain under examination. In particular, the paper’s reported late-period accumulation rate is substantially higher than established estimates of anthropogenic ocean uptake. That does not necessarily make it impossible, but it raises the evidentiary standard for claiming that the result represents a real change in the global sink.
The institutional problem beneath the scientific one
The larger issue is not whether one algorithm succeeds or fails. It is whether the international observing system is capable of detecting a sustained change in the ocean carbon cycle before that change becomes obvious in atmospheric carbon dioxide.
The answer depends on long-running ship surveys, autonomous platforms, shared data standards, laboratory calibration and sustained support for researchers in regions that remain poorly observed. The UNESCO assessment identifies missing observations, incomplete biological understanding and weaknesses in the representation of coastal, polar and land-ocean processes as continuing sources of uncertainty.
Those are institutional limitations as much as technical ones. A model can only reconstruct what the available data and assumptions allow it to reconstruct. Improving the answer therefore requires both better computation and better measurements.
This matters for climate policy. Ocean-carbon estimates enter global carbon budgets, climate projections and evaluations of proposed marine carbon-dioxide-removal techniques. If the ocean absorbs less carbon than expected, more of each unit of human emissions remains in the atmosphere. If the sink appears temporarily stronger, that should not be treated as permission to emit more. Natural sinks vary, and their future behavior is not under political control.
The responsible conclusion is narrower than either an alarmist headline or a reassuring one. We know that the ocean is absorbing a large share of human carbon emissions. We know that carbon storage varies across regions and time. We do not yet know whether the recent changes reconstructed in this preprint represent a sustained weakening or strengthening of the global ocean sink, because the observations remain sparse and the methods used to fill the gaps are still being tested.
Uncertainty here is not ignorance. It is a measurement problem with consequences. The scientific task is to determine which changes belong to the ocean and which belong to the reconstruction.
