PFAS model evaluation
PFAS model performance — and what it does not yet cover
What our coupled PFAS transport model reproduces, how we score it, what failed, and the audit we have not earned the right to publish. Every figure here is from the coupling manuscript and every limit is stated in the same breath as the result it limits.
Figures are frozen artifacts from the coupling manuscript (SSRN 7152526), read from source rather than retyped from a summary. That manuscript is a preprint: SSRN posts it for immediate access, and it has not been peer reviewed. This is a case-study performance report; the portfolio-scale audit that exists for drainage area does not exist for PFAS.
- NSE 0.91 — water-table heads, not PFAS
- 846 PFOS observations → 73 grid cells
- 1,096 of 1,096 daily steps converged
- A case series, not a population audit
A performance page that reports only good numbers gets read as marketing, and then the good numbers get discounted along with everything else. So this page is organised the other way round: each result appears next to the thing it cannot tell you, and the section that matters most is the one describing the audit we cannot yet run.
The short version. The groundwater flow field the transport runs on is fit well. The PFAS comparison itself rests on far less independent evidence than its raw observation count suggests, and we report the denominator that makes that visible. The site demonstration is uncalibrated on purpose. And what we have is two demonstrations, not a portfolio audit — which is a different and much weaker kind of claim than the one our drainage-area page makes.
FAQ
How well does the SWATGenX coupled model reproduce measured PFAS?
The comparison is made in log space, cell by cell, and at that level the model shows no 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. Most of the agreement is the prescribed background, so a single headline accuracy figure would present the background as performance. The groundwater PFOS comparison rests on 846 observations, but those aggregate to 73 model grid cells — 8 prescribed and 65 predicted — and 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.
Is NSE 0.91 the PFAS model performance?
No. NSE 0.91 (RMSE 5.6 m) is the fit of the groundwater flow field to 5,383 static water-table heads. It says the water is moving plausibly; it says nothing about whether the PFAS concentrations are right. Presenting a head-fit statistic as the headline result of a PFAS performance page would be technically true and thoroughly misleading, so it is reported here as what it is: a precondition, not the result.
Have all SWATGenX PFAS models been audited?
No, and the distinction matters. Our drainage-area audit checks every generated model against an independent national reference at portfolio scale. There is no equivalent reference dataset for PFAS to audit against, and there is one published coupled PFAS demonstration plus one uncalibrated site demonstration. That is a case series, not a population audit, and this page reports it as one.
Why is the Wurtsmith demonstration uncalibrated?
Because a calibrated fit would prove less. The source term is measured soil PFAS alone — 1,738 ng/g of measured source-soil PFOS — with no fitted flux and no parameter tuning of any kind. Nothing was adjusted to make the model match the observations. That produces a weaker fit than a calibrated model would report and a stronger claim about the physics, which is the result worth having.
What can the PFAS transport model not do?
PFOS only. Water is exchanged in both directions every day; PFAS travels one way — from a measurement-set source, through groundwater, to the streambed. That is a design choice, not a shortcut: it keeps the in-stream attribution non-circular. Transient transport runs on a steady flow field, and the source cap is a numerical bound rather than a measured ceiling. Coupled groundwater builds can also fail on solver convergence where an equivalent standalone build succeeds — we treat that as a failed delivery rather than shipping an unvalidated result.
Next steps
Last updated 2026-08-31. Figures are frozen artifacts from the coupling manuscript (SSRN 7152526), a preprint that has not been peer reviewed.