Researchers at the University of Maine are developing AI tools to determine what makes up “marine snow”—the continuous stream of tiny particles sinking from the ocean’s surface toward the deep sea. These particles contain organic matter, minerals, carbon, nutrients and pollutants, but underwater cameras can generally reveal only their visible characteristics, such as size, shape and transparency. The project aims to use AI to infer their chemical composition from those visual characteristics, potentially extracting much more information from existing oceanographic imagery.
The National Science Foundation has awarded nearly $700,000 to UMaine researchers Meg Estapa and Chaofan Chen for the three-year project, which is scheduled to run from January 2027 through December 2029. Working with University of Rhode Island researcher Melissa Omand, the team will use data from six major oceanographic field campaigns, including observations from the North Atlantic, waters off West Africa and tropical regions. The researchers hope AI will reduce the months of manual work currently required to classify marine particles and allow scientists to analyze much larger datasets.
A particularly important feature of the project is its focus on interpretable AI. Rather than developing a black-box model that simply predicts what a particle contains, Chen's team wants the system to indicate which visible characteristics influenced its conclusion. Researchers will build a database connecting underwater images with laboratory measurements, particle composition, microplastic concentrations and sampling locations. This will allow scientists to compare AI predictions with physical evidence and identify situations where the model may be misleading.
The potential scientific impact extends beyond simply identifying marine particles. Better understanding of marine snow could improve estimates of how carbon, nutrients and microplastics move from surface waters into the deep ocean and through marine food webs. The same approach could eventually be applied to other scientific fields where researchers must interpret enormous quantities of complex images. The project is therefore an example of AI moving from general-purpose automation toward AI-assisted scientific discovery, where the goal is not to replace scientists but to help them extract patterns from datasets that would be impractical to analyze manually.
