New MSU study shows innovative machine learning techniques more accurately monitor PFAS in Michigan drinking water

EAST LANSING, Mich. — Newly published research from Michigan State University scientists demonstrates how regionally specific machine learning-based modeling more effectively monitors levels of per- and polyfluoroalkyl (PFAS) chemicals in Michigan drinking water compared to nationwide models.

Findings were published in the journal Water Research.

The project was led by A. Pouyan Nejadhashemi, MSU Foundation Professor in the departments of Biosystems and Agricultural Engineering, and Plant, Soil and Microbial Sciences. Other participating researchers in Nejadhashemi’s laboratory were Nicolas Fernandez, who recently joined the University of Florida as a postdoctoral associate, and Christian Loveall.

PFAS, a class of thousands of chemicals, has been a topic of increasing concern over the last several years. Toxicological studies have shown links to numerous human health problems, such as cancers and disorders of the endocrine system, which controls hormone regulation, neurological development and other essential processes.

Nejadhashemi, an expert in modeling and ecohydrology — the interaction between water and ecological systems — is advancing understanding of how PFAS moves through the environment and what can be done to lessen its harmful effects.

The team utilized a branch of artificial intelligence known as machine learning that is designed to use data and algorithms to mimic the way humans learn, improving accuracy over time.

“The Michigan Department of Environment, Great Lakes, and Energy (EGLE) estimates that over 1.5 million residents have consumed PFAS-contaminated water, and more than 11,300 sites are suspected to be contaminated with PFAS,” said Nejadhashemi, a faculty member in the MSU Center for PFAS Research and with MSU AgBioResearch. “However, the silver lining lies in a newly developed PFAS predictive model that offers promise in determining regional PFAS concentrations and presence.”

Nationwide research has been conducted every five years since 2013 on behalf of the Unregulated Contaminant Monitoring Rule, which provides the Environmental Protection Agency and other organizations with scientific information on the presence of contaminants in drinking water that are not governed by the Safe Drinking Water Act.

These large-scale monitoring efforts have revealed a multitude of variables that affect the concentration of PFAS in drinking water sources. One of the primary factors is the distance of drinking water wells from the nearest fire training facility, which have historically used aqueous film-forming foams that contain PFAS to put out fires. Other variables include the concentration of possible co-contaminants and the percentage of urban land within a buffer zone around the wells.

Using data provided by EGLE, Nejadhashemi and his team tested the modeling techniques in Michigan that were employed in other areas around the country, taking into account the state’s unique geology, hydrology and soil characteristics. Researchers also sought to identify the most relevant variables for Michigan.

Four models of varying complexity were used for the study:

• The first used the smallest amount of data and only considered the total possible number of PFAS sources.
• The second was more sophisticated than the first, involving the distance to possible PFAS sources and Michigan soil characteristics.
• The third was the most data-intensive model, using hydrological and soil characteristics in combination with geospatial location and the presence of more than 200 potential co-contaminants.
• The final model introduced land use into the equation, including agricultural, natural and urban land.

Model performance was evaluated by comparing them to the previous nationwide studies using similar metrics. Researchers then worked to improve the models by factoring in new variables discovered through literature review.

A Bayesian variable reduction technique — a modeling tool that helps determine only the relevant variables — was then used to increase model precision.

“This methodology allows us to build on previous work and make it more applicable to Michigan,” Nejadhashemi said. “Our state has unique land use variables, a diverse set of industries that are using PFAS chemicals, different water use and water resources, and many other considerations that need to be made. The goal is to develop better policy to protect drinking water in Michigan, but the strategies may be considerably different from those in other states.”

The team found that proximity to confirmed PFAS sources and airports, soil drainage characteristics, distance to the nearest landfill, and the population and socioeconomic status of a given area all played critical roles in PFAS occurrence and concentration levels in drinking water wells.

“The newly developed data-driven model, along with a distributed hydrogeochemical model, stand as pivotal tools for policymakers and stakeholders,” Nejadhashemi said. “It aims to pinpoint high-risk areas, shape proactive measures, and offer enhanced monitoring guidelines. All these efforts converge on a singular mission: guaranteeing a safe water supply for everyone and safeguarding the Great Lakes from contamination.”

Funding for the project was provided by the Michigan Department of Natural Resources and the U.S. Department of Agriculture’s National Institute of Food and Agriculture.