Evidentiary methods
PFAS model methods, limits and reproducibility
Written for a reviewer looking for overstatement. What the model does not do comes first, every figure carries the unit that makes it checkable, and each artifact is named with its actual access status rather than implied to be downloadable.
Methods and fit statistics: SSRN preprint 7152526, which SSRN labels as not peer reviewed. Data: DOI 10.5281/zenodo.21536787 (open) and DOI 10.5281/zenodo.21096358 (files embargoed until 31 December 2026).
- PFOS only — not all PFAS
- NSE 0.91 is water-table heads, not PFAS
- 846 observations → 73 grid cells
- Source term measured, not fitted
A methods page that reports only what went well is read as marketing and discounted whole, including the parts that were true. So this one is ordered the way a reviewer reads rather than the way that flatters us: the limits first, then what was measured and at what unit, then what was deliberately left untuned, then what you can actually download.
Two figures on this page are easy to carry away wrong, so they are written to resist it. NSE 0.91 describes water levels and not PFAS. And 846 observations aggregate to 73 independent model cells — the second number is the one that bounds what the comparison can support.
FAQ
What does this PFAS transport model not do?
It simulates PFOS only; compound-specific transport with per-analyte isotherms is future work. Water is exchanged in both directions at every daily step, but PFAS travels one way — it enters groundwater as a prescribed, measurement-anchored source and moves to the streambed. That direction is a design choice rather than a simplification: it keeps in-stream attribution non-circular. Transient transport runs on a steady flow field, which omits non-stationary pumping and recharge and is a stated limitation for plume timing. The source cap is a numerical bound, not a measured ceiling.
What was validated, and against what?
The groundwater flow field the transport runs on was fit to 5,383 static water-table observations, giving NSE 0.91 and RMSE 5.6 m. That is a head statistic — it describes water levels, not PFAS. The PFOS comparison itself draws on 846 observations, and those aggregate to 73 model grid cells, 8 prescribed and 65 predicted; 73 is the honest denominator for how much independent evidence the comparison carries. The simulated groundwater PFOS spans the observed range above the 10 ng/L background, largely because most predicted cells sit at that background, and it shows no cell-level skill: a constant at the observed mean scores lower error than the model. The groundwater side is constrained by a head calibration (NSE 0.91 over 5,383 observed heads) and supported independently only by the multi-analyte fingerprint attribution. A streambed porewater check at 43 sites is the most direct available test of the modelled pathway. The coupled run converged on 1,096 of 1,096 daily steps with heads agreeing within 0.16 m.
Why is there no single accuracy percentage?
Because there is no skill for a percentage to report. The manuscript compares the model with the observations in log space, cell by cell, and reports a permutation test showing no cell-level skill: re-pairing the simulated values at random does as well in 10% of permutations, and a constant at the observed mean scores lower error. It reports the agreement with the observed range as largely the prescribed background, so a headline accuracy figure would present the background as performance, and this page does not offer one.
What was the model tuned to reproduce?
Very little, and that is the argument rather than an apology. The source term is set from measurement, not fitted. There is no calibrated groundwater concentration multiplier. The only quantity fitted to in-stream observations is the fraction of discharged load reaching the stream. The site demonstration is uncalibrated outright: its source term comes from measured soil concentrations with no fitted flux. A model that was not tuned to its validation target is harder to attack than one with a better fit and more free parameters.
What can a third party actually obtain?
The methods and fit statistics are in a preprint on SSRN, abstract 7152526, which SSRN itself labels as not peer reviewed. The site well-level dataset is an open Zenodo deposit, DOI 10.5281/zenodo.21536787, holding 35 wells and 4,275 observations. The national soil inventory has a citable DOI, 10.5281/zenodo.21096358, but its files are embargoed until 31 December 2026 — the description is public, the data is not, and citing it as downloadable would be wrong. The groundwater lithology inventory is in open discussion as preprint essd-2026-527. What is not archived is a specific run: reproducing one needs the model bundle for that run, which the paper does not contain.
Next steps
Last updated 2026-08-31. Identifiers on this page were resolved rather than assumed; see the note under Table 2.