Collaborative Research: CAIG: Physics-Informed Deep Learning for Understanding Coastal Hypoxia Formation Mechanisms

Coastal “dead zones” occur when oxygen levels in the water become too low to support most marine life. Dead zones are serious environmental challenges with economic impacts to fisheries, tourism, and other industries. The largest dead zone in the United States forms each summer along the coasts of Louisiana and Texas. The zone spans an area larger than the state of Connecticut and disrupts marine food webs. Scientists know that river nutrients, ocean circulation, and seafloor processes contribute to the formation of dead zones. However, scientists still cannot predict the combination of factors that causes the zones to spread and persist. This project applies machine learning to better understand and predict coastal oxygen loss. A novel component will incorporate data on how oxygen moves in and out of the seafloor. The project will provide training to graduate and undergraduate students. Educational materials for both K-12 and graduate students will be disseminated through an artificial intelligence-powered learning platform. The researchers will release all data, software, and trained models as free, open resources to support reproducibility. The work benefits Gulf Coast communities and resource managers by enabling fast, reliable forecasts that inform fisheries management and nutrient reduction. The project develops a Multi-Architecture Physics-Informed Neural Network framework to advance mechanistic understanding of hypoxia formation on the Louisiana-Texas shelf. The approach integrates three components. First, U-Net and DeepLabv3+ semantic-segmentation networks will map spatial hypoxia patterns from gridded environmental inputs. Dynamic Graph Neural Networks will represent transport pathways that govern oxygen distribution derived from high-resolution regional circulation models and neural operator-based ocean digital twins. Finally, to ensure that predictions remain physically consistent, the researchers will use physics-informed loss functions that embed thermodynamic, advection-diffusion transport, light-attenuation, and sediment oxygen consumption constraints. A central innovation is the explicit incorporation of measured benthic-flux bounds to address the poorly understood role of benthic-pelagic coupling in bottom-water oxygen loss. This novel contribution to geometric deep learning has a graph structure derived from physical transport dynamics rather than data alone. Major activities include training and validating the framework against multi-decadal circulation-model hindcasts and independent field observations, analyzing network attention weights and feature attributions to discover the dominant physical drivers of hypoxia, and deploying interpretable, computationally efficient forecasting tools for research and operational use. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.