Public research tool accompanying Same Unit, Different Quantity: How the Definition of a Field Distribution Coefficient Creates Apparent PFAS Sorption in Soils.
Select a CAS-resolved PFAS, enter soil pH and organic carbon (%), and obtain the frozen laboratory model's log10 Kd and Kd (L/kg). Single predictions, local CSV batches, result exports, and one-variable response curves are supported.
The artifact is the original source-only structural-core RBF SVR: C=3,
epsilon=0.1, with its original median imputation, standardization and one-hot
encoder. It is never refit, reselected or recalibrated here. SHA-256:
A274A76CD6243011F540BF2A175789FF07CE6C0731EE17BB511DED61842A59C3.
This estimates laboratory batch Kd. Field apparent Kd is a different endpoint.
The same locked model had MAE 0.416 log10 units on the controlled D1 cohort and
1.463 log10 units on Lyon's field endpoint, with mean signed error -1.433.
The latter is a substantial underprediction. These aggregate numbers trace to
validation_summary.csv; D0 performance is descriptive, after source model
selection. No predictive confidence interval or calibrated field estimate is
claimed. Curves describe the model, not a causal experiment.
Inputs outside marginal training ranges or identities absent from training are flagged. Being within these ranges does not prove joint applicability. The catalog omits two identities without a resolved structure, and preserves the original representation (including acid/salt form); it makes no name-similarity or protonation-state substitutions.
Inference uses the visitor browser CPU, with no Python server, paid hardware or external API. Files and inputs stay in the browser. The exported RBF computation is regression-tested against the original scikit-learn artifact.
No raw observations, source workbooks, AFFF publications, manuscript files or credentials are included. Model parameters necessarily encode fitted training information. Public source files and model weights are available in this Space.