Researchers at the University of Manchester have developed a new physics-based approach to artificial intelligence. This makes it possible for the first time to accurately predict on a global scale how dissolved organic carbon, an important but previously unquantifiable component of the Earth’s carbon cycle, moves between seawater and marine sediments. This effort was led by Dr. Peyman Babakani It was carried out with cooperation from the Department of Civil Engineering Management. Dr. Majid Sedighireveals how relatively simple AI algorithms can successfully emulate complex mechanical environment models that are typically too computationally intensive to run on a planetary scale.
Solving mechanistic models of natural environments is notoriously time-consuming and often unstable under diverse real-world conditions. To overcome this, the research team trained an AI “emulator” to reproduce the behavior of existing mechanistic models that describe carbon cycling in ocean sediments. These emulators can be applied globally after training to predict the behavior of dissolved organic carbon at resolutions and scales not achievable with the original numerical models alone.
The study found that 11% of particulate organic carbon that reaches the seafloor is returned to seawater as dissolved organic carbon, and 24% is adsorbed by seawater. mineral. Remarkably, about half of all solid-phase organic carbon in the upper meters of marine sediments appears to be derived from dissolved carbon adsorbed on minerals. These findings are the first global quantification of dissolved organic carbon cycling in sediments and highlight its importance in the Earth’s long-term carbon budget.
In developing the modeling framework, the researchers compared deep learning architectures, random forest models, and simpler feedforward artificial neural networks. Unexpectedly, the simplest algorithm produced the most accurate predictions. The team confirmed these results by validating the emulator’s output against a low-resolution global map (where the mechanistic model can be solved numerically) and against variable algebraic solutions using known analytical equations. We also found that predictive accuracy consistently decreases as the complexity of neural network structures increases, providing rare empirical support for the principle of parsimony, also known as Occam’s Razor, in AI model development.
These insights have important implications for climate science. Quantifying carbon budgets across the sediment-water interface is essential to understanding global climate change, but has historically been hampered by computational limitations. By providing a fast, scalable, and accurate way to represent sediment carbon processes, the new AI-based framework can be integrated into global circulation models and used to explore potential ocean-based climate change mitigation strategies. This study opens new avenues for simulating and testing how ocean carbon reservoirs will respond to environmental changes in the coming decades.
