Physics-Based AI Quantifies the Ocean Floor's Hidden Carbon Sink for the First Time

How much carbon does the ocean floor actually trap, and how much leaks back into the water? A new physics-based artificial intelligence model has provided the first global answer: 11% of the particulate organic carbon that reaches the seafloor dissolves back into the ocean, while 24% is permanently locked away by binding to minerals. The exact exchange of dissolved organic carbon between seawater and marine sediments has long remained a missing variable in global climate models, largely because the underlying mechanistic equations were too computationally expensive to run at a planetary scale.[1][2]

The breakthrough, published in the journal The Innovation on September 8, 2026, relies on training AI emulators to replicate the behavior of existing, highly complex mechanistic models. Researchers at the University of Manchester, led by Dr. Peyman Babakhani, recognized that solving traditional numerical models of natural environments is notoriously time-consuming and often unstable under diverse real-world conditions. By shifting the computational load to an AI framework, the team was able to predict dissolved organic carbon behavior at a global resolution that was previously impossible.[1][2]

The data reveals the sheer scale of the seafloor's carbon sink. According to the study, approximately 50% of all solid-phase organic carbon in the upper meter of marine sediments originates from dissolved carbon that has been sorbed onto minerals. This mineral sorption acts as a massive, long-term storage mechanism, preventing the carbon from re-entering the water column and eventually the atmosphere.[1][2][3]

The research also highlights the overlooked role of the deep ocean. The findings demonstrate that the contribution of the abyssal plain to global dissolved organic carbon efflux and preservation is equivalent to half the contribution of the continental margins. Historically, carbon cycle studies have focused heavily on coastal shelves, but the AI model confirms that the vast, deep-ocean floor plays a mathematically significant role in the Earth's long-term carbon budget.[2][3]

Unexpectedly, the project also yielded a finding about artificial intelligence itself. In developing the modeling framework, the Manchester researchers compared complex deep learning architectures, random forest models, and simpler feedforward artificial neural networks. The simplest algorithms consistently produced the most accurate predictions, extending the mathematical principle of parsimony into the realm of physics-based AI.[1][2]

The primary limitation of the new framework lies in its reliance on emulation. Because the AI is trained to reproduce the outputs of an existing mechanistic model rather than direct physical sampling across the entire global seafloor, its accuracy is inherently bounded by the assumptions of the original numerical model. The researchers acknowledge that while the AI solves the computational bottleneck, physical validation through deep-sea sediment core sampling remains necessary to confirm the exact percentages.[2][3]

Despite this constraint, the tool offers immediate utility for climate science. Quantifying carbon budgets across the sediment-water interface allows researchers to integrate these sediment processes directly into global circulation models. 'With this approach, we can finally explore global-scale carbon cycling processes that were previously impossible to quantify,' Dr. Babakhani noted in the university's release.[1][2]

The next phase of the research will apply the emulators to future climate scenarios. By providing a fast, scalable, and accurate way to represent sediment carbon processes, the framework can be used to test potential ocean-based climate change mitigation strategies in silico, simulating how marine carbon reservoirs will respond to warming oceans and changing acidity over the coming decades.[1][2]