Someday, your drinking water could be completely free of toxic “forever chemicals.”
These chemicals, called PFAS (per- and polyfluoroalkyl substances), are found in common household items like makeup, nonstick cookware, dental floss, batteries, and food packaging. PFAS permeate the soil, water, food, and air, and they can remain in the environment for millennia. Once inside the human body, PFAS can persist for years, suppressing the immune system and increasing cancer risk.
A multi-university team of researchers, led by Georgia Tech, is using a cutting-edge machine learning (ML) model to design a better membrane that efficiently removes PFAS from drinking water, a significant source of human exposure.
“More than 200 million Americans in all 50 states are affected by PFAS in drinking water, with 1,400 communities having levels above health experts’ safety thresholds,” noted the study’s principal investigator Yongsheng Chen, Bonnie W. and Charles W. Moorman IV Professor in Georgia Tech’s School of Civil and Environmental Engineering. “Our research aims to provide a scalable, efficient, and sustainable solution for mitigating these toxic chemicals’ impact on human health and the environment.”
The resulting work, funded with more than $10 million in multiyear grants from the U.S. Department of Agriculture (USDA), the National Science Foundation, and the Environmental Protection Agency (EPA), was recently published in Nature Communications.
Sewage treatment limitations
Conventional water treatment processes are ineffective at removing PFAS. Too often, traditional cleansing methods, such as using chlorine to kill pathogens in water, create harmful byproducts.
“Solving one problem creates another problem,” said Chen.
He has already used ML and artificial intelligence in precision agriculture to monitor nutrient levels in plants and insists that tackling PFAS removal similarly requires new approaches. Rather than treating an entire body of water, Chen’s team first separated PFAS from the water stream. Success depended on finding the right membrane material to isolate the chemicals in the water.
UW-Madison Associate Professor Ying Li validated the ML model with molecular simulations.
Chen relied on a team of 10 PhD students and nine research scientists to perform the ML modeling. In addition to Georgia Tech, two other schools contributed people and laboratory expertise. The University of Wisconsin-Madison validated the model with molecular simulations, while Arizona State University trained it using data from scientific literature and their lab.
“Applying machine learning to membrane separation represents an exciting frontier for environmental engineering,” said Tiezheng Tong, an associate professor of environmental engineering in ASU’s School of Sustainable Engineering and the Built Environment.
This is another step in tackling PFAS pollution, a widespread problem that has recently received significant public attention due to PFAS’ toxic nature and the recent EPA ruling on PFAS in drinking water, he said.
“By integrating with molecular simulation tools, we can better understand PFAS transport across nanofiltration and reverse osmosis membranes, pushing the boundary of fundamental science relating to membrane separation,” Tong said.
ML accelerates membrane-material discoveries
Using ML modeling significantly sped up the discovery process. For instance, one PhD student in Chen’s lab used trial and error over two years to pinpoint one promising membrane. Machine learning modeling allowed the team to find eight membrane candidates 10 to 20 times faster, reducing discovery time from years to a few months.
“Our molecular dynamics simulations reveal that electrostatic interactions, size exclusion, and dehydration play critical roles in governing the transport of PFAS molecules across polyamide membranes,” said Ying Li, an associate professor of mechanical engineering at UW-Madison. “These calculations indicate that electrostatic interactions dominate PFAS rejection, with charged functional groups significantly influencing transport behavior. The simulation results provide fundamental insights that align with ML predictions, highlighting the key molecular determinants of PFAS removal efficiency.”
Addressing PFAS exposure in agriculture
By addressing PFAS contamination, this research could also benefit the agriculture industry, which depends on fertilizer sourced from water treatment plants. Wastewater biosolids are processed into fertilizer, offering farmers and ranchers a cheaper alternative to chemical fertilizers. Unfortunately, PFAS-tainted fertilizers from sewage sludge have contaminated significant amounts of land and livestock. Industry groups estimate that almost 70 million acres of U.S. farmland could be contaminated by these forever chemicals.
By funding this research, the USDA hopes that an effective membrane will help the United States reclaim this crucial resource.
“Synthesizing a very smart membrane to get rid of PFAS also allows us to recover the fertilizer from municipal wastewater treatment plants,” Chen said. “Such a membrane could enable us to get rid of things we don’t want and keep the things we need, so we can keep the water for irrigation or other applications.”
Eliminating PFAS in fertilizers also could help address the mismatch of food and water demand in urban versus rural areas since 80% of the demand resides in cities. PFAS removal could directly support urban area resource recovery and food production.
“Our goal is achieving a circular economy where materials never become waste, and nature is regenerated,” Chen said.
What’s next
The team will fine-tune the model and add more data to improve its training features. Chen will synthesize membranes in his lab to further test the model’s PFAS removal predictions.
Today, scientists have found ways to remove long chains of PFAS, but the shorter chains of these chemicals persist, explained Chen.
“If we can better understand the mechanism, we’ll be able to design a good material membrane to get rid of all PFAS. That could be game-changing.”
This work is partially supported by the NSF (Award Nos. 2112533, 2427299, 2345543, Yongsheng Chen; 2448130, Tiezheng Tong; and 2345542, Ying Li).
Yongsheng Chen acknowledges the financial support by the USDA (Award No.2018−68011-28371), NSF-USDA (Award No. 2020-67021-31526), and EPA (Award No. 840080010).
Tiezheng Tong acknowledges the support of the USDA National Institute of Food and Agriculture (Hatch Project COL00799, accession 1022591).
Ying Li acknowledges the financial support by the National Alliance for Water Innovation (NAWI), funded by the US DOE, Office of Energy Efficiency and Renewable Energy (EERE), Advanced Manufacturing Office, under Funding Opportunity announcement Number DE-FOA-0001905, through a subcontract to the University of Wisconsin-Madison.
A version of this story was originally published by Georgia Tech.