PFAS source attribution by spatial and chemical crosswalk

Framing (non-negotiable). The unit of analysis is the S2 Level-13 grid cell (~1.27 km²), the shared spatial key that lets PFAS measurements, EPA-regulated facilities and administrative geography be joined across graphs. Coverage is 2,949 cells across 14 states and 307 counties, dominated by Maine's state monitoring programme and the national Water Quality Portal. The level of inference is spatial co-occurrence, not causation: a facility sharing a cell with a PFAS detection is a plausible contributor, not a demonstrated one, and no hydrology, groundwater flow direction, or release record was used. Maine's PFAS sampling is moreover risk-targeted by statute — sites were chosen because contamination was suspected — which inflates any facility–detection association measured on it. Treat every ranking below as hypothesis generation for site prioritisation, not as an attribution of responsibility. Keep this caveat attached to every downstream claim.

Abbreviations. PFAS per- and polyfluoroalkyl substances · AFFF aqueous film-forming foam · FRS EPA Facility Registry Service · NAICS North American Industry Classification System · EGAD Maine DEP Environmental and Geographic Analysis Database · WQP Water Quality Portal · ICE EPA/NIEHS Integrated Chemical Environment · QSUR quantitative structure–use relationship · DTXSID DSSTox substance identifier · CAS Chemical Abstracts Service registry number · WWTF wastewater treatment facility · S2 Google S2 spherical-geometry grid · KG knowledge graph · NPL National Priorities List · PFOA/PFOS perfluorooctanoic acid / perfluorooctane­sulfonic acid · ng/L nanograms per litre (parts per trillion) · OR odds ratio.

1. Executive summary

Joining PFAS measurements (sawgraph) to EPA-regulated facilities (fiokg) on the S2 Level-13 grid produces a proximity–contamination gradient. Across the 2,102 grid cells that carry at least one PFAS detection, the median peak aqueous concentration falls from 36.8 ng/L where an EPA PFAS-relevant facility sits in the same ~1.3 km² cell, to 29.6 ng/L in an adjacent cell, 16.3 ng/L where only non-flagged regulated facilities are nearby, and 8.0 ng/L where no regulated facility falls in the window at all (Kruskal–Wallis H=53.7, p=1.3×10⁻¹¹). The same ordering holds independently for detection frequency and for the number of distinct PFAS analytes detected. A cell with a PFAS-flagged facility within the one-ring window is 2.15× more likely to return any detection (89.8% vs 80.4%, Fisher exact p=5.8×10⁻⁹).

The crosswalk resolves 598 of 2,102 detection cells (28.4%) to a named PFAS-relevant facility within ~1–3 km, drawn from 12,714 co-located EPA FRS facilities of which 435 carry EPA's PFAS-industry flag. The ranking recovers Maine's known PFAS investigation sites without being told about them: the top-ranked cells resolve to Naval Air Station Brunswick (104,265 ng/L in that cell), the Bangor Air National Guard 101st Air Refueling Wing / Bangor International Airport (53,041 ng/L), the Verso/Pixelle Androscoggin paper mill at Jay (24,800 ng/L), Tasman Leather Group and the Hartland WWTF, and the former Loring Air Force Base at Limestone — each an independently documented AFFF, paper-mill, tannery or landfill source (§8).

The chemical axis is thinner but usable: of 175 distinct analytes, 93 carry a well-formed CAS, of which 39 resolve into biobricks-ice and 32 into biobricks-toxcast (up to 1,510 assay endpoints for PFOS). ICE supplies functional-use categories for 39 of them, but **all are model-predicted (QSUR), none curated** — so use category is a weak, hypothesis-grade axis here, not evidence of what a compound was actually used for.

The honest counterweight: **68% of the fifty highest-concentration cells are not facility-attributable (tiers C/D), and the single most contaminated cell in the entire dataset — 226,000 ng/L in Kennebec County, Maine — has no PFAS-flagged facility in its cell or ring**. Our data show only the absence of a candidate facility there, not the presence of an alternative source; the biosolids/land-application pathway documented for this part of Maine (§8, F6) is the most plausible explanation in the literature, but this study carries no biosolids layer and cannot test it. Either way it marks the ceiling on what a pure co-location crosswalk can attribute.

2. Sources used

Every row above traces to at least one logged, non-exploratory SPARQL query in the reproducibility record (§11). No other federation graph contributed to any number in this report.

3. Design & rules

The analysis asks a deliberately narrow question: where does a PFAS detection sit close enough to a regulated facility, of a kind known to handle PFAS, that the facility is worth investigating as a source? Everything else — hydrology, release inventories, groundwater gradients, plume modelling — is out of scope, and that is the principal reason the output is a prioritisation list rather than an attribution.

The spatial key is the S2 Level-13 cell. Both sawgraph sample points and fiokg facilities attach to those cells through kwg:sfWithin, so a same-cell join is exact rather than a distance approximation, and the cell's ~1.27 km² footprint sets the resolution of every claim. Note that the naive owl:sameAs join suggested by the federation's join registry does not work for facilities: in fiokg, owl:sameAs is a self-link into epa-frs-data#, and only kwg:sfWithin reaches the grid. Getting this wrong silently returns zero PFAS-flagged facilities.

Two windows are used. The same-cell window is the strictest reading of co-location. The 1-ring window adds the eight S2 cells adjacent to the sample cell — roughly 1–3 km — taken from spatialkg's own spatial:connectedTo adjacency rather than computed locally, so the neighbourhood is itself part of the federated, reproducible query chain. Every cell in the study has exactly eight neighbours, which we verified rather than assumed.

Cells enter the study only if they contain a SAWGraph sample point. Of those, cells with no analyte-linked observation are excluded outright, and cells with observations but no detection are held back as a control set rather than ranked — a detection is what the ranking is trying to attribute, so a non-detect has nothing to attribute. The inventory below is rebuilt live from the extracts:

The co-location score combines five components — proximity, detection intensity, detection frequency, analyte breadth and an industry source-strength prior — renormalised over whichever components a cell actually has, so a cell measured only in a non-aqueous medium is not penalised for lacking a ng/L value. The industry prior grades each NAICS leaf code as High, Moderate or Low source strength following EPA's PFAS-industry sector list. The exact weights, saturation constants and the full NAICS prior table are specified once, in the reproducibility file — they are deliberately not restated here.

Just over a quarter of PFAS detections (598 of 2,102) sit in the same cell or an adjacent cell as an EPA PFAS-flagged facility, and half (1,061 of 2,102) sit near regulated facilities that carry no PFAS flag at all — the crosswalk narrows the field sharply but leaves most detections unattributed, which is the expected result for a screening method rather than a failure of it.

4. Confidence tiers

Tiers are ordinal in the strength of the attribution, not of the contamination: a tier-D cell can be heavily contaminated (and several are), it simply has no candidate facility to name. The tier distribution is deliberately top-heavy in C — most PFAS sampling in this federation happens near some regulated facility, because most sampling happens near people and industry.

5. Findings by axis

5.1 Proximity gradient — the primary signal

The central result is that all three independent contamination measures fall as distance from a PFAS-relevant facility increases. Peak concentration and detection frequency decline strictly at every step; analyte breadth declines across the range but plateaus within pairs (median 9, 9, 8, 8 analytes for tiers A–D), so it is monotone non-increasing rather than strictly decreasing. Peak concentration falls 36.8 → 29.6 → 16.3 → 8.0 ng/L across the four tiers; detection frequency falls 0.263 → 0.232 → 0.200 → 0.182; analyte breadth falls likewise. Pairwise, tier A exceeds tier C (Mann–Whitney one-sided p=2.0×10⁻⁵, rank-biserial 0.24) and tier D (p=1.2×10⁻⁸, 0.36), and tier B exceeds tier D (p=8.5×10⁻¹⁰, 0.31). Tier A does not significantly exceed tier B (p=0.14): at this resolution, "same cell" and "next cell over" are not distinguishable, which is a useful negative result — it says the effective attribution radius is the ~1–3 km ring, not the ~1.3 km cell.

Within the ranking itself the score and the measured peak concentration agree closely (Spearman ρ=0.817 over the 1,349 scored cells with aqueous data, p<10⁻³⁰⁰), which is a coherence check on the score rather than independent evidence — concentration is one of the score's own components.

A permutation test guards against the gradient being an artefact of the score construction: shuffling tier labels across the 1,349 scored cells with aqueous data 10,000 times gives a null median tier-A concentration of 17.0 ng/L against the observed 36.8 ng/L (p=0.0003).

The gradient is real but shallow — roughly a 4.6× median difference between the closest and the most distant tier, against within-tier spreads of three to four orders of magnitude. That is the signature of a genuine but weak spatial predictor: useful for ranking a worklist, useless for adjudicating an individual site.

5.2 Spatial distribution and hot-spots

The dataset is not national in any uniform sense. Maine contributes 1,286 of 2,949 cells and 213 of 598 facility-attributable detections, with Minnesota, Indiana and Arizona supplying most of the rest; 13 states carry at least one tier-A or tier-B cell. Maricopa County (Arizona), Cumberland County (Maine), Hennepin County (Minnesota) and Pima County (Arizona) hold the most tier-A cells.

Tier-A cells cluster tightly around the Minneapolis–St Paul, Indianapolis, Phoenix/Tucson and southern Maine industrial corridors, while the tier-D cells thin out into rural areas — a pattern that is at least partly the geography of sampling effort, not of contamination (§10, limitation 2).

5.3 Detection-frequency and compound axis

Detection is dominated by the short- and long-chain perfluoroalkyl acids: PFOS and PFOA are each detected in 57–58% of the ~20,000 observations that screen for them (11,519 and 11,581 detections respectively), with PFBA, PFPeA, PFHpA and PFHxA close behind at 50–54%. The 22.6% overall detection rate across 567,538 observations reflects the long tail of rarely-detected analytes — 175 distinct analytes are reported, but most are screened widely and found seldom.

5.4 Chemical and toxicological crosswalk

Of 175 analytes, 118 rows carry a CAS literal resolving to 93 distinct well-formed CAS numbers; 39 of those match biobricks-ice and 32 match biobricks-toxcast. Assay coverage tracks regulatory attention rather than environmental prevalence — PFOS carries 1,510 ToxCast endpoints and PFOA 1,396, while several equally-detected short-chain acids carry roughly 460–510.

Panel C shows the axis a source-attribution study most wants and least gets: the compounds that are environmentally ubiquitous are not the ones with the deepest toxicological characterisation, so the chemical crosswalk adds identity and bioactivity context but cannot by itself discriminate sources.

6. Domain analyses

Four domain analyses were planned; all four were run — industry-sector attribution, functional-use stratification, regional stratification, and the screened-negative control. A fifth, hydrologic routing of detections to upstream facilities via hydrologykg/geoconnex, was deliberately skipped: hydrologykg covers Illinois only and the federation's registry records sawgraph↔geoconnex as a verified non-join (reference-IRI vs materialised-node mismatch), so no reproducible flow-path query was available.

6.1 Industry-sector attribution

Ranked by the number of PFAS sample cells they touch (the metric plotted in Figure 5; facility counts differ and are given second), sewage treatment is the most frequent PFAS-flagged same-cell neighbour (47 cells / 48 facilities), followed by metal coating and electroplating (24 / 27) and waste treatment and disposal (22 / 34). Widening to the 1-ring changes the ordering: metal coating and electroplating leads (122 cells / 149 facilities), then sewage treatment (121 / 91), plastics and rubber products (94 / 94), and airport operations (89 / 76). Airports rank far higher in the ring than in the cell, which is exactly what an AFFF fire-training source looks like when the release point and the monitoring well are a kilometre or two apart.

Industry coverage is the weakest link in this axis: only 2,452 of the 12,714 co-located facilities (19.7%) carry any fio:ofIndustry link at all, so roughly four-fifths of the co-located inventory is industrially unclassified and the sector counts are lower bounds.

6.2 Functional-use stratification

ICE supplies predicted functional-use categories for 39 CAS across 5 categories only: emulsion stabilizer (39 analytes, 25.5% detection frequency), flame retardant (30, 21.9%), surfactant (1, 16.4%), foamer (2, 7.7%) and antimicrobial (1, 0%). The apparent "emulsion stabilizer > flame retardant > surfactant" ordering in Figure 4B is almost entirely an artefact of which analytes fall in which category — the categories overlap heavily and 32 of the 39 CAS carry the emulsion-stabilizer label (the other seven carry only a flame-retardant, surfactant, foamer or antimicrobial label). Every one of these assignments is model-predicted (QSUR); none of the PFAS in this set carries a curated OECD functional use. The axis is therefore reported for completeness and explicitly not used as evidence of source type.

6.3 Regional stratification

Detection rate varies about three-fold between states (11% Vermont to 31% South Carolina) but the small-n states are unstable; the interpretable contrast is Maine and Massachusetts at 24–27% against Minnesota at 14% and Arizona at 15%, which most plausibly reflects differences in what each programme sampled — Maine targeted suspected sludge and AFFF sites, the WQP states sampled ambient water — rather than a real regional difference in contamination.

6.4 Screened-negative control

The control set is the sharpest available test of the co-location hypothesis. Of 435 cells with PFAS screening and zero detections, 68 still have a PFAS-flagged facility within the window (8 in the same cell, 60 in the ring only) — the false positives. These include a resin manufacturer in York County (Maine), a paper mill in Crow Wing County (Minnesota) and a military installation in Penobscot County (Maine) where screening returned nothing. Their existence sets a practical ceiling: proximity to a flagged facility raises the odds of a detection 2.15-fold but is far from determinative.

7. Discussion

Read together, the axes describe a method that works about as well as its inputs permit. The spatial crosswalk is the load-bearing element: it is exact (a shared grid key, not a distance heuristic), it is reproducible entirely inside the federation, and it produces a statistically clear gradient in three independent contamination measures. The chemical crosswalk adds identity, bioactivity breadth and a use-category axis, but the use categories are predicted rather than curated and the assay coverage tracks regulatory history rather than environmental behaviour, so chemistry contextualises the ranking without sharpening it.

The most useful practical output is the tier structure rather than the score. Tier A and B together name 598 cells with a specific candidate facility, and the top of that list is dominated by exactly the sectors the regulatory literature identifies — AFFF at military airfields and airports, paper mills, tanneries, landfills, sewage treatment. That the ranking independently rediscovers NAS Brunswick, Bangor ANG, the Jay paper mill, Tasman Leather/Hartland and Loring AFB — without any prior site list — is the strongest evidence that the crosswalk is picking up signal rather than sampling density.

The most important finding for anyone intending to use this is the failure mode. Facility proximity is not where the extreme values live: 68% of the fifty highest-concentration cells fall in tiers C or D, and the two most contaminated cells in the dataset (226,000 and 167,000 ng/L, both in Kennebec County, Maine) have no PFAS-flagged facility in cell or ring. Tier A is only modestly enriched in that extreme tail (1.45× over its base rate) — and so, strikingly, is tier D (1.38×). The tail is bimodal: facility-proximal contamination and a second population of severe, facility-distant contamination. The literature identifies that second population readily (§8, F6): PFAS-bearing biosolids and septage spread on farmland tens of kilometres from the mill or treatment plant that produced them — the pathway Maine has been investigating in the Kennebec/Fairfield area. We cannot confirm that mechanism here, because no biosolids or septage land-application layer exists in these graphs; what our data establish is only that the most severe contamination has no regulated facility to point at. A co-location model is structurally blind to it, because the proximate source is a field, and fields are not regulated facilities.

Three testable predictions follow. First, adding a biosolids/septage land-application layer — licensed spreading sites, which Maine DEP holds — should reclaim a large share of the tier-C/D extreme tail and is the single highest-value extension. Second, because tier A and tier B are statistically indistinguishable, widening the window to a 2-ring (~3–5 km) should add candidate facilities without degrading the gradient, consistent with the ~4–5 km critical distance reported in the European surface-water literature. Third, the gradient should flatten measurably on the WQP layer relative to the EGAD layer, because WQP sampling is not risk-targeted; a stratified re-run is the cleanest available internal control on the ascertainment bias described in §10.

8. Comparison with prior work

The comparison used WebSearch and direct retrieval of primary regulatory documents (Maine DEP, EPA, ATSDR, ITRC, National Guard/Air Force records) rather than the PubMed/Paperclip connectors, which were not reachable in this session; the sources are listed in §12. Each finding was checked against the retrieved sources' own text; none against a paywalled full text, and figures that could not be independently confirmed are marked Unresolved rather than asserted. The per-finding detail behind each row is in PFAS_literature_comparison.md.

F6 is the finding that matters most, and our own data are consistent with it: the 68% of extreme-value cells that no facility explains (§7) is exactly the gap a diffuse, non-facility pathway would leave. That is corroboration of the limitation, not confirmation of the mechanism — identifying biosolids as the actual source would require a land-application layer this study does not have. F4 is the second: because Maine chose where to sample partly on suspicion of nearby sources, the association measured in §5.1 is an upper bound on what a randomly sited monitoring network would show, and the effect size should not be transported to other states.

9. Full ranked results

The complete ranked table — 2,102 scored cells with all five score components, tier, geography, facility counts, industry attribution and named facilities — is in PFAS_results.xlsx (sheet Ranked Results, tier-coloured with autofilter), alongside the screened-negative control, per-analyte chemistry, industry and NAICS detail, regional tables, statistical tests and a Methods & Rules sheet. The machine-readable extracts and every intermediate are in data/.

Tip: click a column header to sort, type in the search box to filter, and use the drop-downs to restrict to a confidence tier, a state, or an attribution window. The sources (n) column counts the federation graphs backing each row — sawgraph supplies the measurement, spatialkg the grid cell and its administrative geography, and fiokg the co-located facility and its industry.

The interactive map below plots the 150 highest-ranked cells on OpenStreetMap tiles; each marker is clickable and carries that cell's rank, score, tier, county, detection counts, peak concentration, facility counts and the names of its nearest PFAS-flagged facilities.

The named facilities at the top of the ranking are overwhelmingly AFFF sites (Air National Guard, Naval Air Station Brunswick, Bangor International Airport), pulp and paper mills, wastewater treatment facilities, landfills and a tannery — the method converges on the sectors regulators already prioritise, without having been given a sector list.

A representative slice of the top of the ranking:

The ranking's top is a list of named, independently documented PFAS sites, which is the intended behaviour — but note that the highest score (89.7) and the highest concentration in this slice (104,265 ng/L) are different cells — and that the dataset-wide maximum, 226,000 ng/L, belongs to a tier-C cell that does not appear here at all, and the score's median across all scored cells is only 22.7. The score orders a worklist; it does not measure contamination severity.

10. Summary of findings & limitations

Findings. Joining 567,538 PFAS observations to 12,714 EPA-regulated facilities on a shared ~1.3 km² spatial grid resolves 598 of 2,102 detection cells (28.4%) to a candidate PFAS-relevant facility within roughly 1–3 km. Peak aqueous concentration and detection frequency decline strictly, and analyte breadth non-strictly, with distance from such a facility (Kruskal–Wallis H=53.7, p=1.3×10⁻¹¹ for concentration; H=79.6 for detection frequency; H=17.85 for analyte breadth), and the presence of a flagged facility in the window raises the odds of any detection 2.15-fold. The sectors that dominate the attributable set — military AFFF sites and airports, pulp and paper mills, sewage treatment, landfills, tanning, metal plating, petroleum terminals — match the regulatory and peer-reviewed literature, and the highest-ranked cells resolve to independently documented PFAS investigation sites (NAS Brunswick, Bangor ANG, the Jay paper mill, Tasman Leather/Hartland, Loring AFB) that were never supplied to the method.

The chemical crosswalk resolves 39 and 32 of 93 CAS into biobricks-ice and biobricks-toxcast respectively, adding standardised identity and up to 1,510 assay endpoints per compound, but supplies only model-predicted use categories and so cannot discriminate sources on chemistry alone. Against that, 68% of the fifty most contaminated cells have no facility to attribute at all. The literature's leading candidate for that population is biosolids land application (§8, F6), which these graphs do not represent — a structural blind spot rather than a tuning problem.

Limitations.

1. Co-location is not causation. No hydrology, groundwater gradient, release record or temporal

ordering entered the analysis. A facility sharing a cell with a detection may be downgradient of it, may postdate it, or may be irrelevant.

1. Ascertainment bias is severe and directional. Maine's PFAS sampling is risk-targeted by statute

(§8, F4) — sites were chosen because a source was suspected. The measured facility–detection association is therefore an upper bound, and the Maine-dominated geography (1,286 of 2,949 cells) means the national picture is not a national sample.

1. The extreme tail is not facility-attributable. 68% of the top-50 concentration

cells are tiers C/D; the biosolids/septage pathway that most plausibly explains them is absent from the federation graphs used here.

1. Industry coverage is thin. Only 2,452 of 12,714 co-located facilities (19.7%) carry a

fio:ofIndustry link, and 76,167 of fiokg's PFAS-flagged facilities have none at all — every sector count in §6.1 is a lower bound.

1. Functional use is predicted, not curated. All 39 use assignments come from QSUR

models; none of these PFAS carries a curated OECD assignment (§8, F5).

1. The chemical crosswalk is sparse. 39/93 CAS reach ICE and

32/93 reach ToxCast; aggregate parameters (sum-of-6 PFAS, PFOA+PFOS) carry no CAS and drop out of every chemical axis despite being among the most-detected quantities.

1. Concentration comparability. The maxNgL axis is restricted to ng/L results so that values are

comparable; cells whose detections were only in soil, sediment or tissue therefore lack that score component (it is renormalised away, not zeroed). Twenty-six cells returned a coso:non-detect sentinel IRI where a numeric maximum was expected and were coerced to missing.

1. Facility itemisation is incomplete. The per-facility extract covers 12,430 of the

12,714 facilities the aggregate count establishes (97.8%); the headline counts use the exact aggregates, the industry breakdown uses the itemised subset.

1. 412 cells excluded. These carry a sample point but no analyte-linked observation in

sawgraph, so they can be neither ranked nor used as controls.

1. Snapshot, not a time series. All graphs are pinned releases (§2); PFAS monitoring and the FRS

facility registry both change continuously, and no temporal alignment between a facility's operating period and a sample's date was attempted.

1. The tier-A/tier-B distinction is not statistically supported (p=0.14). The ~1.3 km cell is

finer than the data can resolve; treat A and B as one "facility-proximal" class.

1. Grid-cell artefacts. S2 Level-13 cells vary in true area with latitude, boundary cells straddle

counties (the first county alphabetically is taken as primary), and a facility just outside the 1-ring is treated identically to one a hundred kilometres away.

11. Reproducibility

Everything needed to replicate this analysis — the originating prompt, the replicator specification (join keys, window definitions, score weights, the NAICS prior), every supporting SPARQL query verbatim with its row count, the verified quantities, the pinned KG versions and the timing — is in PFAS_reproducibility.md, with the analysis scripts in scripts/ and the intermediate extracts in data/.

12. References

Retrieved by WebSearch and direct fetch of primary regulatory documents; the full annotated set with per-finding mapping is in PFAS_literature_comparison.md. None was verified against paywalled full text.

1. Hu X.C. et al. (2016). Detection of poly- and perfluoroalkyl substances (PFASs) in US drinking water linked to industrial sites, military fire training areas, and wastewater treatment plants. Environmental Science & Technology Letters. doi:10.1021/acs.estlett.6b00260
2. Tokranov A.K. et al. / USGS (2024). Predictions of groundwater PFAS occurrence at drinking water supply depths in the United States. Science. doi:10.1126/science.ado6638
3. Breitmeyer S.E. et al. (2023). Predicting PFAS occurrence in private wells using machine learning. Science of the Total Environment. doi:10.1016/j.scitotenv.2023.167839
4. Chen Q. et al. (2023). Spatial distribution and attenuation of PFAS in soil and groundwater around a fluorochemical industrial park. Journal of Hazardous Materials. doi:10.1016/j.jhazmat.2023.131372
5. Sunderland-style EU surface-water ML study (2025). Critical distance thresholds for point-source PFAS influence in European surface waters. Environment International. doi:10.1016/j.envint.2025.109312
6. Garrett J. et al. (2025). The Landscape of PFAS Contamination in the United States: Sources and Spatial Patterns. Environmental Science & Technology. doi:10.1021/acs.est.4c14474
7. Anderson R.H. et al. (2016). Occurrence of select PFAAs at US Air Force AFFF-impacted sites. Chemosphere. doi:10.1016/j.chemosphere.2016.01.014
8. California Bayesian spatial PFAS model (2024). Facility-buffer predictors of PFAS in California drinking-water sources. Environmental Research. doi:10.1016/j.envres.2024.118762
9. US EPA (2021–2024). Multi-Industry PFAS Study — Preliminary and Final Reports. link
10. National Academies of Sciences, Engineering, and Medicine (2022). Guidance on PFAS Exposure, Testing, and Clinical Follow-Up. doi:10.17226/26156
11. US EPA (2024). Effluent Guidelines Program Plan 15 — PFAS priority categories (landfills, leather tanning). link
12. Maine DEP (2024–2025). Naval Air Station Brunswick PFAS response and monitoring results. link
13. US EPA (n.d.). Brunswick Naval Air Station Superfund site profile (NPL). link
14. US Navy / MRRA (2024). Hangar 4 AFFF release — incident and sampling reports, Brunswick Landing. link
15. Maine ANG / 101st Air Refueling Wing (2018–2024). PFAS site investigation, Bangor ANGB. link
16. Maine CDC / DEP (2019–2024). Bangor-area private well PFAS sampling results. link
17. Maine DEP (2021–2024). PFAS in Maine — programme background, sludge land-application history and commissioner statements. link
18. Maine Superior Court (2024). State of Maine v. paper-mill defendants — PFAS biosolids litigation filings. link
19. Maine DEP (2022). PFAS in wastewater treatment facility effluent and landfill leachate — statewide screening. link
20. US EPA ECHO / NPDES (n.d.). Tasman Leather Group, Hartland ME — permit record. link
21. Town of Hartland / Maine DEP (2022). Hartland WWTF and landfill leachate PFAS results. link
22. US Air Force Civil Engineer Center (2018). Former Loring AFB PFAS site inspection report. link
23. Maine DEP (2024). Statewide PFAS residential well testing results table. link
24. Maine DEP (n.d.). Environmental and Geographic Analysis Database (EGAD). link
25. National Water Quality Monitoring Council (n.d.). Water Quality Portal. link
26. Maine DEP (2022). PFAS sludge and septage land-application site prioritisation — tiering methodology. link
27. State of Maine (2021). Public Law 2021 c.478 — An Act To Investigate PFAS Contamination of Land and Groundwater. link
28. Phillips K.A. et al. (2017). Suspect screening analysis of chemicals in consumer products / QSUR models for functional use. Environmental Science & Technology. doi:10.1021/acs.est.7b04781
29. Isaacs K.K. et al. (2016). Chemical Product and Function Database (CPDat). Journal of Exposure Science & Environmental Epidemiology. doi:10.1038/jes.2015.72
30. US EPA (n.d.). CompTox Chemicals Dashboard — Functional Use and Predicted Functional Use. link
31. NIEHS/NICEATM (n.d.). Integrated Chemical Environment (ICE) — Functional Use Explorer. link
32. US EPA (n.d.). CPDat / Factotum documentation — curated vs predicted functional use. link
33. ITRC (2023). PFAS Technical and Regulatory Guidance Document — fate and transport, plume length. link
34. Chemours Fayetteville Works air-deposition studies (2019–2023). NC DEQ consent-order sampling. link
35. Sunderland E.M. et al. (2019). A review of the pathways of human exposure to PFAS and present understanding of health effects. Journal of Exposure Science & Environmental Epidemiology. doi:10.1038/s41370-018-0094-1
36. NH DES (2018–2022). Saint-Gobain Performance Plastics, Merrimack NH — air-deposition PFAS investigation. link
37. Schaider L.A. et al. (2016). Septic systems as sources of organic wastewater compounds including PFAS. Science of the Total Environment. doi:10.1016/j.scitotenv.2016.04.104
38. Wisconsin DNR (2022–2024). PFAS in private wells and septic-influenced groundwater. link
39. Houtz E.F. & Sedlak D.L. (2012). Oxidative conversion as a means of detecting precursors to PFAAs in urban runoff (TOP assay). Environmental Science & Technology. doi:10.1021/es302274g
40. Benskin J.P. et al. (2010). Perfluorinated acid isomer profiling in water and quantitative assessment of manufacturing source. Environmental Science & Technology. doi:10.1021/es102582x
41. Maine DEP (2021–2023). Fairfield-area PFAS investigation — residential well results. link
42. Maine Department of Agriculture, Conservation and Forestry (2022–2024). PFAS in agricultural soils from biosolids land application. link