AI Learns to Decode the Chemistry Hidden in Sinking Ocean Particles

Every day, across every ocean basin on Earth, a slow and largely invisible snowfall drifts down through the water column. This is marine snow, a continuous rain of tiny particles composed of dead phytoplankton, the waste products of larger organisms, mineral grains, bacterial communities, and increasingly, fragments of plastic pollution. To the casual observer, the particles look like nothing more than drifting flecks of debris. To oceanographers, they are one of the most important conveyer belts on the planet, carrying carbon, nutrients, and contaminants from the sunlit surface waters toward the deep sea. The problem has always been that while scientists can see these particles sinking, they have struggled to determine what the particles actually contain. A research effort led from the University of Rhode Island now aims to change that, using artificial intelligence to extract chemical information from ordinary underwater photographs.

Melissa Omand, a research professor at the URI Graduate School of Oceanography, is leading the project, which is funded by a grant of nearly 700,000 dollars from the U.S. National Science Foundation. The three-year effort is scheduled to begin in January 2027 and run through December 2029. Working with collaborators Meg Estapa and Chaofan Chen at the University of Maine, Omand and her team will investigate whether machine learning models can predict the chemical composition of marine snow from characteristics that are visible in underwater images, such as particle size, shape, and transparency. If the approach succeeds, it could transform the way oceanographers interpret the enormous volumes of imagery their instruments already collect, turning pictures of drifting particles into quantitative estimates of what those particles are made of.

The stakes are considerable. Marine snow sits at the heart of the biological carbon pump, the set of processes by which the ocean sequesters carbon dioxide from the atmosphere. When phytoplankton die or are consumed, their remains aggregate into particles dense enough to sink. As Omand explains, these particles are rich in carbon and other nutrients, provide a key food source for organisms living in the deep ocean, and play an important role in marine carbon sequestration. The rate at which marine snow sinks, and the rate at which it is decomposed or consumed on the way down, determines how much carbon is locked away in the deep ocean and for how long. Better measurements of particle composition could therefore sharpen estimates of how the ocean will respond to a warming climate.

Underwater cameras have been part of the oceanographer’s toolkit for decades, and modern instruments can capture images of thousands of individual particles in a single deployment. Cameras mounted on autonomous floats and profilers can now document particle distributions across entire ocean basins. What the images cannot generally reveal, however, is chemistry. A particle that appears as a translucent blob in a photograph might be a nutrient-rich aggregate of organic matter, a mineral ballast fragment, or a piece of microplastic. Distinguishing among these possibilities has traditionally required collecting physical samples, filtering them, and analyzing them in laboratories, a slow and expensive process that limits how much of the ocean can be characterized.

The labor involved in manual image analysis is not trivial. Estapa noted that one of her graduate students spent months classifying particles and identifying what appeared in underwater images. That kind of effort, repeated across the field, represents a bottleneck at precisely the moment when imaging technology is generating more data than human analysts can process. The researchers hope that AI can automate much of this classification work, allowing scientists to collect and interpret more information while spending more of their time on scientific analysis and discovery. Crucially, the team also plans to develop models in ways that help scientists understand how the models reach their conclusions, addressing a common criticism of machine learning approaches in the sciences.

The foundation of the project is a carefully constructed database that pairs images with ground truth. Omand’s role in the first year will center on collecting marine snow images and matching physical samples across a broad range of ocean environments, spanning coastal waters off Ghana, the equatorial Atlantic, the New England shelf, and the California current system. Through collaborations with the Monterey Bay Aquarium Research Institute and the University of Ghana, the team already collected samples from two of these sites during the summer of 2026. The resulting database will link each image to information about the geochemical properties of the corresponding particles, the amount of microplastic present, and the location where the samples were gathered.

Once that database is assembled, the researchers will use data from six major oceanographic field campaigns to test whether groups of particles visible in images can accurately predict their chemical composition. The logic is that particle appearance is not random. Aggregates formed from diatom blooms look different from fecal pellets produced by zooplankton, and mineral-rich particles scatter light differently from soft organic ones. If those visual signatures correlate reliably with chemistry, a trained model could estimate carbon content, nutrient loading, or plastic contamination directly from photographs, without waiting for laboratory analysis. Validating such predictions across diverse ocean regions, from productive coastal shelves to the oligotrophic open ocean, will be the critical test of whether the method generalizes.

One of the most socially resonant applications involves plastic pollution. By identifying and quantifying plastic particles alongside naturally occurring marine snow, the researchers hope to better understand how plastic debris moves from surface waters into the deep ocean and through marine food webs. Microplastics have been found in the deepest ocean trenches and in the tissues of animals at every level of the marine food chain, yet the pathways by which they travel downward remain poorly quantified. If sinking aggregates are a major transport mechanism, as several studies suggest, then an AI system capable of flagging plastic-containing particles in routine imagery could produce basin-scale maps of plastic export that no sampling program could match.

The same tools could illuminate the natural cycling of carbon and nutrients through the ocean. As Omand puts it, gaining detailed insights into marine snow particles, their associated communities, and the links to their environment will allow scientists to better predict the movement of carbon and nutrients in the ocean and the impact that changes may have on marine life. That predictive capacity matters at a time when ocean warming, acidification, and shifting plankton communities are altering the efficiency of the biological pump in ways that global climate models struggle to capture. Particle-level observations, scaled up through automated analysis, could provide the empirical grounding those models lack.

For Omand, the project also represents a broader shift in how ocean science handles data. Research vessels, autonomous floats, and seafloor observatories now generate imagery faster than any laboratory can interpret it, and the techniques developed for marine snow could extend well beyond oceanography, to any scientific field that relies on complex images that are difficult and time-consuming to interpret. In that sense, the humble flecks drifting through the deep ocean may end up teaching researchers far from the sea a lesson about extracting knowledge from pictures. The snowfall that carries the ocean’s carbon downward may, with the help of machines that learn to read it, carry scientific understanding upward.

Subject of Research: Using artificial intelligence to determine the chemical composition of sinking marine snow particles from underwater imagery

Article Title: Reading the ocean’s snowflakes: How AI decodes marine snow

Article References: Reading the ocean’s snowflakes: How AI decodes marine snow. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: marine snow, artificial intelligence, carbon sequestration, microplastics, oceanography, biological carbon pump, underwater imaging, machine learning, National Science Foundation, University of Rhode Island, deep ocean, nutrient cycling