An interdisciplinary team led by North Carolina State University has secured more than $1 million from the National Science Foundation to build AI tools that forecast coastal dead zones faster and more accurately than existing methods.
The three-year project, a collaboration between NC State and Louisiana State University, brings together oceanographers, computer scientists and coastal experts. Paul Liu, a professor in NC State's Department of Marine, Earth and Atmospheric Sciences and co-director of the university's AI Hub for Science, serves as lead principal investigator. The combined NSF investment totals $1,049,875, with $800,013 going to NC State and $249,862 to LSU.
Why predicting dead zones is hard
Each summer, the largest coastal dead zone in the United States forms along the Louisiana-Texas continental shelf. Nutrients from the Mississippi and Atchafalaya rivers trigger biological growth, while ocean currents, water-column stratification and seafloor processes determine where oxygen gets depleted. Scientists understand each factor individually, but predicting how they interact to control the size and persistence of hypoxic zones has remained elusive.
Hypoxia - dangerously low dissolved oxygen - can kill marine life and disrupt fisheries. Gulf Coast communities, fisheries managers and environmental planners all need better forecasts to make decisions about fishing closures and nutrient-reduction strategies.
How the AI system works
Liu's team is developing what they call a Multi-Architecture Physics-Informed Neural Network framework. The system combines image-segmentation models, dynamic graph neural networks, high-resolution ocean-circulation simulations and ocean digital twins. Crucially, physical constraints - including water movement, oxygen transport, light attenuation and sediment oxygen consumption - will be embedded directly into the AI framework.
"Physical constraints governing water movement, oxygen transport, light attenuation and sediment oxygen consumption will be embedded directly into the AI framework," the team reported.
A central part of the innovation is incorporating measurements of oxygen exchange between the seabed and the water above. Sediment deposition, resuspension and consumption can significantly affect bottom-water oxygen levels, particularly during storms. By modeling these processes more accurately, the researchers hope to identify the mechanisms that cause hypoxic waters to form, grow and persist.
The project also aims to produce forecasting tools that run faster than traditional numerical models while staying physically consistent and scientifically interpretable. This speed could directly support fisheries management, nutrient-reduction planning and decision-making in Gulf Coast coastal communities.
Training the next generation
The project provides interdisciplinary training for graduate and undergraduate students in oceanography, coastal science and AI. Educational materials will be distributed through an AI-powered learning platform. The team plans to release project data, software and trained models as open resources for the wider scientific community.
For researchers in the AI for Science & Research field, this project demonstrates how physics-informed deep learning can tackle environmental problems that traditional models struggle with.
Why this matters for science and research professionals
This NSF-funded project represents a concrete test case for embedding physical constraints directly into deep learning systems rather than treating neural networks as black boxes. For scientists and engineers working on environmental modeling, the project's success or failure in the Gulf of Mexico will provide a clear benchmark for whether physics-informed AI can deliver faster, more interpretable forecasts without sacrificing accuracy. That proof of concept could prove critical for other high-stakes domains - from flood prediction to climate modeling - where the balance between speed and physical consistency is necessary.
Your membership also unlocks:
