Excellence in Research: Harnessing Data Science to Quantitatively Predict the Partitioning Behavior of Per and Polyfluoroalkyl Substances (PFAS)

This project is funded by the U.S. National Science Foundation (NSF) Historically Black Colleges and Universities-Excellence in Research (HBCU-EIR) Program. HBCU-EIR supports research at public and private HBCUs to strengthen their research capacity and promote engagement with NSF. Per- and polyfluoroalkyl substances (PFAS) are known as "forever chemicals." They are found in the drinking water of millions of Americans. They have been linked to cancer and other health problems. Engineers still lack good tools to predict how these chemicals move through soil, water, and treatment systems. Part of the problem is scale. PFAS include thousands of different chemical forms, but lab data exist for only a few of them. As a result, current models are too narrow to guide real decisions about water treatment and site cleanup. This project closes that gap in two steps. First, the team will run new lab tests to measure how a wide range of PFAS compounds stick to different kinds of materials. Second, the team will use this data to build a new machine-learning (ML) model. This model can handle small, uneven datasets. It can still make reliable predictions, even for PFAS compounds that were never tested directly. The resulting tools should help utilities, regulators, and engineers make faster, better-informed choices to protect the environment. Beyond the science, the project will train new scientists and engineers, including a PhD student, a postdoctoral researcher, and an undergraduate assistant, in integrated experimental and Artificial Intelligence (AI) methods. The project will also bring PFAS science into K-12 classrooms through a hands-on learning module built with regional STEM centers and shared with students at the NC Science Festival. This project integrates targeted experimentation with artificial intelligence (AI) to build a mechanistic, data-driven framework for predicting PFAS partitioning behavior across environmentally and industrially relevant solid-liquid interfaces. The research is organized around three objectives: (1) constructing a curated PFAS adsorption database combining systematic literature mining with an in-house isotherm campaign spanning diverse PFAS subclasses (short/long-chain, carboxylate/sulfonate, linear/branched, precursor) and sorbent families (activated carbons, ion-exchange resins, engineered sorbents), characterized via Quadrupole Time-of-Flight mass spectrometer(QTOF-MS) and combustion ion chromatography (C-IC); (2) developing a multimodal-multivariate mixture-of-experts (MoE) machine learning architecture that couples mixture density networks with hybrid response-surface-methodology/ML sub-models, dynamically weighted by molecular and surface descriptors—including Abraham solvation parameters, BET surface area, pHpzc, and pore metrics—to resolve distribution coefficients (Kd) across electrostatic and hydrophobic sorption regimes without assuming a priori mechanistic separability; and (3) translating model outputs into full-scale engineering applications by predicting adsorption bed breakthrough curves and retardation factors (R) validated against lab-scale column tests. Scientifically, the project advances PFAS fate-and-transport theory by resolving how molecular structure and sorbent surface chemistry jointly govern partitioning across mixed-mechanism regimes, while methodologically it contributes a generalizable AI framework for modeling sparse, heterogeneous environmental datasets that extends beyond PFAS to other emerging contaminants, with direct engineering application to the design and operation of full-scale environmental remediation systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.