Parth RanaBack to home

ChemPrint

PFAS Source Attribution · Minnesota East Metro

Demonstration app. Interactive walkthrough of the ChemPrint pipeline. The classifier shown uses fingerprint matching; the full study trains a Random Forest / Gradient Boosting / MLP ensemble at 99.3% cross-validated accuracy on synthetic source profiles.
Manufacturing AFFF (firefighting foam) Landfill leachate Wastewater effluent
Water Sample Input
Concentration · ng/L (parts per trillion)
ChemPrint Analysis
Chemical Fingerprint · molar fractions
samplesource profile
Attributed Source
—
confidence —
The Basis: Four Sources, Four Fingerprints
Each source type emits a characteristic distribution of the 15 PFAS compounds. These published profiles are what every classification is measured against. The proportions, not the total concentration, carry the signal.
105 East-Metro Monitoring Sites
Manufacturing AFFF Landfill Wastewater Known disposal site
Selected site
Click any site
Each point is a monitoring location. ChemPrint assigns it the most probable source. Click to inspect.
41.9%
Manufacturing
23.8%
Landfill
22.9%
Wastewater
11.4%
AFFF
39% of sites (41 of 105) fall in high-vulnerability communities. Manufacturing contamination is most common among them (16 sites).
Model Comparison · 10-fold CV
ModelAccuracyWeighted F1
Random Forest0.9930.993
Gradient Boosting0.9870.987
Neural Network0.9770.976
ChemPrint Ensemble0.9930.993
Per-Class Performance
SourcePrec.RecallF1
AFFF1.0001.0001.000
Landfill0.9930.9870.990
Manufacturing0.9931.0000.997
Wastewater0.9870.9870.987
Confusion Matrix (%)
Rows are the true source, columns the prediction. The diagonal is correct. Every error sits between landfill and wastewater, which share short-chain PFAS, since treatment plants receive landfill leachate as influent.
Feature Importance (Gini)