Water-treatment engineers trying to remove PFAS have been stuck in a deeply annoying loop: one membrane study says “great rejection,” another says “meh,” and a third shows the same chemical slipping through like it knows the building code.
That is the frustration behind Tomsovic, Gu, Doudrick, and Straub’s new paper, “Mechanistic Insights into PFAS Rejection in Nanofiltration and Reverse Osmosis from Data-Driven Analysis” in Environmental Science & Technology (DOI: 10.1021/acs.est.6c02287, PMID: 42059125). The authors pulled together 2,353 literature data points on PFAS rejection by nanofiltration and reverse osmosis membranes, then used machine learning to ask a basic but slippery question: what actually controls whether these “forever chemicals” get blocked?
Spoiler: the answer is not “just buy a better membrane and call it lunch.”
The Membrane Is a Bouncer, but the Guest List Is Weird
PFAS are a sprawling family of fluorinated chemicals famous for resisting heat, oil, water, and apparently our collective desire for a tidy cleanup problem. They show up in water because they have been used in firefighting foams, coatings, textiles, packaging, and industrial processes. Their carbon-fluorine bonds are stubborn little overachievers.
Membrane treatment sounds simple: push contaminated water through a very selective barrier. Clean-ish water goes through. PFAS stay behind. Reverse osmosis is the stricter club bouncer; nanofiltration is still picky, but a little more willing to let small things pass. Wikipedia describes nanofiltration membranes as sitting between ultrafiltration and reverse osmosis, with pore sizes roughly in the 1-10 nanometer range, which is both tiny and somehow still large enough for chemistry to cause drama (nanofiltration background).
And drama matters because the EPA’s 2024 PFAS drinking water rule set enforceable limits for several PFAS and identified treatment technologies including granular activated carbon, ion exchange, reverse osmosis, and nanofiltration (EPA PFAS rule). Translation: this is no longer “interesting lab problem” territory. Utilities need treatments that work outside the spotless fantasy realm of beaker water.
The Machine Learning Part: Less Robot Oracle, More Organized Suspicion
The authors analyzed 13 descriptors covering PFAS properties, membrane properties, operating conditions, and water chemistry. Machine learning here is not a magical crystal ball. It is more like a very patient spreadsheet detective with better posture.
The strongest main effects came from membrane water permeance and PFAS molecular volume. In plain English: size matters. Bigger PFAS molecules are easier to reject, and the membrane’s water-moving behavior says a lot about how permissive the barrier is. That supports steric exclusion, the satisfyingly physical idea that some molecules are simply too bulky to squeeze through.
That sounds obvious until you remember PFAS are not one chemical. They are a chemical extended universe, and not every installment is equally blockable.
Earlier machine-learning work on PFOS nanofiltration found strong predictive performance from random forest, gradient boosting, and AdaBoost models, with pressure, PFOS concentration, and membrane type ranking highly among variables (DOI: 10.1016/j.seppur.2022.120775). The new study broadens the lens: not just PFOS, not just one narrow setup, but a meta-analysis across many PFAS, membranes, and experimental conditions.
The Water Itself Is Meddling
Here is the eyebrow-raising part: background ions and dissolved organic matter did not behave like polite supporting characters.
At low concentrations, ions and organic matter could increase PFAS rejection, possibly by forming complexes that make PFAS act larger. Great. Chemistry puts a winter coat on the contaminant, and the membrane says, “Too bulky, try another door.”
But at higher concentrations, rejection could drop. Why? The authors point to mechanisms like charge shielding and concentration polarization. Charge shielding means ions can dampen electrostatic repulsion, letting PFAS approach or pass more easily. Concentration polarization means rejected material piles up near the membrane surface, creating a local mess right where you least want a local mess.
So yes, “more stuff in the water” can help until it suddenly hurts. Sure, 95% rejection sounds great until the other 5% is the part your compliance officer is staring at with a calculator and no sense of humor.
Recent experimental work backs up the idea that water chemistry is not background scenery. A 2024 study found that surfactants, ion valency, and temperature changed PFOA and PFBA rejection in commercial RO/NF systems, with high-valence ions sometimes improving rejection and elevated temperature reducing PFBA rejection for nanofiltration (DOI: 10.1016/j.jwpe.2024.106039). A 2023 closed-circuit membrane study also examined high-recovery PFAS rejection, which matters because real systems often run at recoveries where concentrate effects stop being cute (DOI: 10.1016/j.seppur.2023.124867).
Why This Paper Is Useful, With the Usual Caveats Wearing Steel-Toed Boots
The best part of this study is not that it declares one master rule. It does the opposite. It gives researchers a framework for why prior studies disagreed without implying everyone before them forgot how membranes work.
That matters for membrane design. If molecular volume and permeance dominate, engineers can focus on tuning membrane structure for size exclusion. If ions and organic matter flip behavior depending on concentration, then testing only in clean lab water is like crash-testing a car by gently rolling it into a pillow.
The limitations are real. Literature meta-analysis inherits the quirks of the literature: inconsistent reporting, uneven coverage of PFAS types, different membranes, different water matrices, and probably a few experimental details hiding in supplementary tables like they owe someone money. Machine learning can reveal patterns, but it cannot invent missing experiments. The authors explicitly identify data gaps, especially around underexplored PFAS and realistic feedwaters.
A 2024 critical review of nanoengineered membranes makes a similar point: PFAS removal often depends on membrane structure, electrostatics, and realistic water chemistry, not just a headline rejection percentage (DOI: 10.1016/j.jwpe.2024.105471). And a 2026 review of nanofiltration for PFAS removal argues that steric exclusion leads the mechanism list, with electrostatic repulsion close behind, while calling for more realistic water-matrix studies (DOI: 10.1016/j.jiec.2026.01.046).
The Takeaway: The Filter Works, but the Water Has Opinions
This paper’s value is its refusal to treat PFAS rejection like a single-number beauty contest. It says: look at the molecule, look at the membrane, look at the soup the molecule is swimming in, and maybe stop pretending pure-water tests are destiny.
If the results hold up and expand, utilities and researchers could use this kind of data-driven framework to choose membranes more intelligently, design better experiments, and avoid expensive “wait, really?” moments during scale-up. For PFAS treatment, that is not flashy. It is better: useful.
References
- Tomsovic Y, Gu S, Doudrick K, Straub AP. “Mechanistic Insights into PFAS Rejection in Nanofiltration and Reverse Osmosis from Data-Driven Analysis.” Environmental Science & Technology, 2026. DOI: 10.1021/acs.est.6c02287, PMID: 42059125.
- “Machine learning-based modeling and analysis of PFOS removal from contaminated water by nanofiltration process.” Separation and Purification Technology, 2022. DOI: 10.1016/j.seppur.2022.120775.
- “Effects of surfactants, ion valency and solution temperature on PFAS rejection in commercial reverse osmosis (RO) and nanofiltration (NF) processes.” Journal of Water Process Engineering, 2024. DOI: 10.1016/j.jwpe.2024.106039.
- “Rejection of Perfluoroalkyl Acids by Nanofiltration and Reverse Osmosis in a High-Recovery Closed-Circuit Membrane Filtration System.” Separation and Purification Technology, 2023. DOI: 10.1016/j.seppur.2023.124867.
- U.S. EPA. “Per- and Polyfluoroalkyl Substances (PFAS).” EPA PFAS drinking water information.
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.