Forecasting ‘forever chemicals’ in U.S. waterways with AI

Southeast Michigan’s Huron River abounds with picturesque natural scenes, including burbling streams, graceful trumpeter swans, towering leafy trees, and… polluted foam? More than just an eyesore, this foam—now a common sight in waterways across Michigan and much of the U.S.—often contains a group of harmful synthetic chemicals called perfluoroalkyl and polyfluoroalkyl substances, or PFAS, which have been linked to a variety of negative health effects.

Now, a team of researchers at the University of Michigan, alongside nonprofit organization Environmental Working Group (EWG), are working to develop artificial intelligence (AI) tools that can predict PFAS concentrations in water sources across the country. They have received a Catalyst Grant from the U-M Graham Sustainability Institute, which will provide $10,000 and facilitation services to support their continued developments.

“With this grant, our team is hoping to leverage the latest AI methods to create models that can accurately predict the presence of PFAS in fish or surface water,” said Elizabeth Bondi-Kelly, assistant professor of computer science and engineering at U-M and PI of the project. “Our goal is to take some of the labor and complexity out of identifying areas of high pollution, the first step toward avoiding PFAS exposure and its potential negative effects.”

PFAS are called “forever chemicals” because of their durability—they can take hundreds or even thousands of years to break down. Their presence can cause lasting damage to the environment, building up in harmful amounts in soil, water, and even as fine particulate matter in the air we breathe. Exposure to PFAS has been linked to adverse health effects in humans, including increased cholesterol, increased risk of some cancers, increased risk of pregnancy complications, and others.

Despite these issues, PFAS use shows no signs of slowing. These chemicals continue to be used in a wide variety of consumer and industrial products, from everyday cleaning and personal care items to compounds used in firefighting and military applications.

“PFAS are particularly harmful because they are both long lasting and highly prevalent—they’re pretty much everywhere,” said Bondi-Kelly. “They are used in a lot of common products, like waterproof fabrics, coatings for nonstick pans, food packaging, and more, and easily seep into surface water through industrial runoff and even just standard disposal methods.”

Now at higher concentrations than ever before, PFAS pose a significant risk to the environment and human health, with no apparent end in sight. Key to combating this persistent and growing threat is understanding where PFAS concentrations are highest, and what factors contribute most to their presence.

Over the past two decades, EWG has played a leading role in documenting the human health impact and scope of PFAS contamination, including through their innovative mapping initiatives. Through painstaking water sampling and compiling existing data from publicly available sources, the organization has built an interactive map that allows users to view PFAS concentrations at over 6,000 sites across the U.S.

While EWG’s mapping system is extensive and robust, adding to it and keeping it up to date entails monumental effort. PFAS concentrations change over time, and maintaining accurate data requires regular sampling, a time- and cost-intensive task.

This is where harnessing the power of AI could have a tremendous impact. Teaming up with EWG, Bondi-Kelly is looking to use machine learning methods to build easy-to-use and accurate models that take much of the guesswork and maintenance out of these PFAS monitoring efforts.

“The goal is to alleviate the burden of wide-scale water sampling, which requires substantial cost and effort,” said Bondi-Kelly. 

The first step toward building these powerful predictive models is leveraging existing data. Using PFAS water contamination data along with additional PFAS readings taken from fish samples, the researchers are working to feed their model as much information as possible, training it to make more accurate predictions.

Their model’s calculations will also incorporate various factors that might be relevant to PFAS pollution, such as the locations of nearby military bases or industrial sites, enabling further analysis about possible contributors to high PFAS concentrations.

“Using multiple types of data, from fish tissue samples to satellite images, we’re training our model to predict which geographic locations are more likely to have high concentrations of PFAS,” explained Bondi-Kelly.

Armed with this information, the hope is that this model will give experts, public servants, and everyday citizens alike access to up-to-date, accurate information about PFAS concentrations, enabling more targeted environmental monitoring and decision-making.

“The tools we develop will support a variety of users in making  more informed choices,” said Bondi-Kelly, “including where to deploy mitigation measures, what products to buy, and much more.”

After its initial training with existing data, the next step for Bondi-Kelly’s team is to equip their model with the ability to generalize and quantify uncertainty, predicting PFAS concentrations in locations where data is sparse or nonexistent.

Through this innovative use of AI, Bondi-Kelly’s collaboration with EWG could significantly streamline the process of understanding and combating PFAS pollution. As the team sets out to improve and validate their AI-driven models, they are paving the way for quicker, more efficient methods of identifying pollution hotspots and informing public health and policy decisions, a significant step toward protecting our environment and health from these persistent contaminants.