The study, published in The Innovation, was led by Dr Peyman Babakhani from the Department of Civil Engineering and Management and carried out in collaboration with Dr Majid Sedighi. Their starting point was an existing mechanistic model, a set of coupled differential equations describing how organic carbon reacts, dissolves, diffuses, and binds to mineral surfaces within sediment columns. Models of this kind are the gold standard for environmental process research because they are grounded in physical and biogeochemical reality. Yet they are also notoriously slow to solve and often numerically unstable when pushed across the enormous diversity of real-world conditions found in the ocean, from shallow coastal shelves to the deepest abyssal plains.

The Manchester team’s solution was to train AI emulators, machine learning systems designed to reproduce the input-output behaviour of the full mechanistic model at a tiny fraction of the computational cost. Instead of replacing physics with data-driven guesswork, the approach keeps the mechanistic model at its core: the emulator learns directly from the model’s own solutions, effectively compressing years of expensive simulation into a fast surrogate that can be evaluated millions of times. Once trained, the emulators could be applied across the entire global ocean, predicting dissolved organic carbon behaviour at resolutions and spatial scales that the original numerical model alone could never achieve.

The results quantify, for the first time, the global magnitude of this previously unmeasured carbon pathway. According to the study, 11 percent of the particulate organic carbon arriving at the seafloor is returned to the water column as dissolved organic carbon, while 24 percent becomes sorbed onto mineral surfaces within the sediment. Perhaps most striking of all, the researchers estimate that roughly half of all solid-phase organic carbon in the upper metre of marine sediments originates not from intact particles settling from above, but from dissolved carbon that was once in seawater or porewater and later bound itself to mineral grains. In other words, the dissolved pool is not a footnote to the sediment carbon budget; it is one of its principal architects.

These numbers carry real weight for Earth’s long-term carbon accounting. Marine sediments hold one of the planet’s largest reservoirs of organic carbon, and the fraction that escapes remineralisation determines how much carbon is buried for millions of years versus how much cycles back into the ocean and atmosphere. By demonstrating that mineral sorption of dissolved organic carbon supplies about half of the solid-phase carbon stock in the uppermost metre of sediment, the study reshapes how scientists should think about the seafloor’s role as a carbon sink. A pathway that was too small, too slow, or too complex to resolve in global models turns out to be a first-order control on the planet’s carbon budget.

What makes the technical achievement even more remarkable is the choice of algorithm. In developing the modelling framework, the researchers systematically compared deep learning architectures, random forest models, and simpler feedforward artificial neural networks. The expectation in much of modern machine learning is that bigger, deeper, more heavily parameterised models perform better. Here, the opposite occurred: the simplest algorithms produced the most accurate predictions. The team found that increasing the complexity of the neural network structures consistently reduced prediction accuracy, a result that offers rare empirical support for the Principle of Parsimony, better known as Occam’s Razor, within AI model development. When the underlying process is governed by well-defined physics, a lean model that respects that structure can outperform far more elaborate machinery.

Accuracy was not taken on faith. The researchers validated their emulator outputs in two independent ways. First, they compared predictions against low-resolution global maps in regions where the original mechanistic model remained numerically solvable, confirming that the surrogate faithfully reproduced the full model’s behaviour wherever a ground truth existed. Second, they checked emulator outputs against algebraic solutions for variables with known analytic expressions, providing an exact mathematical benchmark rather than a merely approximate one. This dual validation strategy is what allows the team to describe their global predictions as accurate rather than merely plausible, and it reflects a broader lesson for the use of machine learning in the geosciences: an emulator is only as trustworthy as the evidence that it has genuinely learned the physics it was meant to imitate.

The implications extend well beyond sediment geochemistry. Quantifying carbon budgets across the sediment-water interface has long been recognised as essential for understanding global climate dynamics, but progress has been hindered by exactly the computational limitations this study overcomes. General circulation models, the workhorses of climate projection, cannot currently afford to embed full mechanistic sediment models at every grid cell of the ocean. A fast, scalable, and accurate surrogate changes that calculus. The new AI-based framework can be integrated into global circulation models, allowing sediment carbon processes to be represented explicitly rather than crudely parameterised, and opening the door to systematic exploration of how the ocean’s carbon reservoirs may respond to warming, acidification, and altered productivity in the coming decades.

The framework also creates new opportunities for evaluating ocean-based climate change mitigation strategies before any real-world intervention is attempted. Dr Babakhani noted that the modelling framework can play a substantial role in testing potential ocean-based climate change mitigation scenarios in silico, adding that with this approach the team can finally explore global-scale carbon cycling processes that were previously impossible to quantify. The ability to run thousands of counterfactual simulations, each asking how sediment carbon storage might shift under a different management or emissions scenario, is precisely the kind of capability that policy-relevant climate science has lacked for the sediment realm.

There is a broader lesson here for how artificial intelligence enters the environmental sciences. Rather than chasing ever-larger architectures trained on ever-larger datasets, the Manchester study suggests that the most powerful role for AI in many geophysical problems is as a disciplined student of mechanistic models, learning their behaviour, compressing their cost, and extending their reach to scales where the underlying equations could never be solved directly. In doing so, physics-based AI has turned an intractable computational problem into an answerable scientific question, and in the process revealed that the quiet chemistry of the seafloor plays a far larger role in Earth’s carbon cycle than anyone had been able to measure before.

Subject of Research: Physics-based artificial intelligence modelling of dissolved organic carbon cycling between seawater and marine sediments

Article Title: Physics based AI unlocks first global predictions of carbon cycling in ocean sediments

Article References: Physics based AI unlocks first global predictions of carbon cycling in ocean sediments. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: dissolved organic carbon, marine sediments, carbon cycle, physics-based AI, machine learning emulators, neural networks, Occam's Razor, carbon sequestration, mineral sorption, climate modelling, ocean biogeochemistry, University of Manchester