Theme: 07. Per- and Polyfluorinated Alkyl Substances (PFAS)

Topic Area: 7n. PFAS Source and Forensic Considerations

Formatted Title:

Seeing the Forest for the (Decision) Trees: Machine-Learning Enhances PFAS Analytics

Background/Objectives:

Complex mixtures of per- and polyfluoroalkyl substances (PFAS) can enter the environment through a wide variety of sources. Once in the environment, different fate-and-transport characteristics alter the composition and spatial distribution of original impacts, further increasing the number of chemical signatures observed at a given site. Ubiquitous environmental presence and diminutive regulatory limits place a burden on site owners to establish site specific knowledge in at least three areas: 1) relative responsibility for PFAS impacts, 2) magnitude of recontamination potential (i.e., ambient levels), and 3) data-driven boundaries for remediation decision units. The intrinsic high dimensionality of PFAS environmental chemistry makes source tracking difficult or infeasible with traditional site investigation methods. Increasingly accessible machine-learning tools have popularized big-data approaches that “learn” the patterns of various source types (like a specific aqueous film-forming foam). However, complex PFAS fate-and-transport presents a formidable obstacle for site-specific application of supervised machine-learning algorithms. The dearth of quality, comprehensive source profile data, the extremely high number of potential sources, and the potential for other onsite sources (i.e., multiple generations of AFFF, additional waste streams) likewise creates significant challenges for differentiation of onsite and offsite sources from precursor transformations and ambient levels.

Approach/Activities:

A new PFAS fingerprinting tool developed by WSP combines robust multivariate algorithms to characterize impacts on contaminated sites through extraction of ambient levels and fate-and-transport patterns. This process initially discriminates ambient from elevated concentration ranges, then differentiates compositional signatures and correlates them with environmental processes (i.e., precursor transformation) to distill probable sources, and finally delineates unique spatial zones. Finite-mixture models (FMM) rely on probability theory to estimate concentration-based groups in complex datasets, and a version of FMM capable of handling censored data produces a range of ambient levels. Nonnegative matrix factorization (NMF) performs blind-source separation of different signatures and can estimate both their chemical profiles and highlight spatial hot spots. Machine learning feature importance compares NMF signatures with precursor (and other) data. Finally, a novel spatially informed clustering algorithm partitions the site into zones of similar impacts that may be used to target further site investigation and remediation. The framework presented here has implications for general site investigations, where an established background location is not present or is suspect, where fingerprints are mixed or an offsite source is conceptualized, and when data-driven site partitioning is required. This tool was applied to PFAS sites to assess the variability of potential site-related sources, fate-and-transport patterns, and ambient levels.

Results/Lessons Learned:

Matrix factorization identified multiple potential onsite PFAS source areas and distinct downgradient fingerprints. Due to the spatial patterns and chemical compositions of these fingerprints, it is likely that they represent precursor transformations or differential transport effects rather than offsite sources. NMF and FMM both identified a low-concentration signature that dominated offsite, which was used to estimate ambient levels. Feature importance calculations indicated differential links between supplemental analyses and NMF results, clarifying potential processes that may play a role in the composition of specific signatures. Spatially informed clustering partitioned sites into multiple data-driven focus areas for further investigation and remedial design. As with other analyses, rigorous preprocessing and quality checking were important safeguards against artificial bias. Distillation of major themes from these complex results was key to communicating findings to technical and nontechnical stakeholders. Together, these results indicate a significant potential for the differentiation of sources, fate-and-transport patterns, and ambient levels to inform effective site management, reduce or apportion liability, and improve engineering considerations early in the remediation project life cycle.

Stefano Marconetto

Affiliation

Track

Refine Results

Seeing the Forest for the (Decision) Trees: Machine-Learning Enhances PFAS Analytics