Estimating PFAS Exposure Risk Under Climate Stress Using Machine Learning - Poster #27R-G

Sep 17, 2026

Uploaded by Skyler

Poster #27R-G evaluates how hydrologic and climate conditions shape PFAS (“forever chemicals”) concentrations and where potential exposure hotspots may emerge across CONUS. The study integrates EPA UCMR occurrence and concentration data, demographic and military site information, land cover, and remote-sensing climate/hydrologic variables. K-means clustering (3 clusters) defines spatial PFAS risk groupings, then machine-learning models (Random Forest and XGBoost) quantify predictive skill and feature importance against observed PFAS levels.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Introduction

Data & Methods

Vulnerability Index and Remote Sensing

Clustering and Machine Learning Models

Results

Estimating PFAS Exposure Risk Under Climate Stress Using Machine Learning Kathleen I. Shank1 , Himmat Basnet, Eunsang Cho, Sangchul S. Hwang Ingram School of Engineering | College of Science and [Engineering | Texas State University |](Engineering | Texas State University |1 nat98@txstate.edu)[1](Engineering | Texas State University |1 nat98@txstate.edu)[ nat98@txstate.edu](Engineering | Texas State University |1 nat98@txstate.edu) Poster \#27R-G INTRODUCTION • Per-and polyfluoroalkyl substances (PFAS) are persistent “forever chemicals” that resist environmental degradation and accumulate in natural systems • Due to their strong carbon-fluorine bonds, PFAS have been detected throughout the hydrologic cycle, raising concerns about widespread exposure (Tokranov et al. , 2024) • PFAS exposure risks are elevated in communities near known point sources (e.g. , industrial facilities, military installations, airports, wastewater treatments plants, agricultural areas with irrigation or biosolids application, wildfire prone regions) • Objective: Evaluate the influence of hydrologic and climate variables on PFAS concentrations and identify potential hotspots across CONUS DATA & METHODS • Data Integration: PFAS occurrence and concentration data from Environmental Protection Agency’s 3rd and 5th Unregulated Contaminant Monitoring Rule (UCMR), covering years 2013-15 and 2023-25, respectively. Concentrations of multiple compounds were averaged to generate a single composite PFAS concentration, allowing comparisons across locations • U.S. Census tract-level demographic data, population, and military site locations within a ZIP code • Vulnerability Index: Equally weighted %Population below poverty line,%Black,%Hispanic or Other,%Population without health insurance,%Renter occupied households, Normalized median household income • Remote Sensing Products: Annual dominate land cover type of each ZIP code from National Landcover Database (NLCD) • Climate & Hydrologic Variables: Annual averaged precipitation and temperature from Parameter-elevation Regressions on Independent Slopes Model (PRISM), Annual maximum and minimum drought indices from the Standardized Precipitation Evapotranspiration Index (SPEI) Table 1: Categorical variables used in K-Means clustering, RF, and XG Boost models Category Variables Tem poral year Climate/Hydro logic prec ip_ ann , tem p_ ann , Variables spei_ m in , spei_ max Location lat , lon Land Cover lc_ code Socio-Demographic population Potential PFAS military_ site_ count Sources Identifier (excluded zip_ code from training) Figure 1: Mean PFAS concentrations across ZIP codes, values categorized using EPA-informed screening thresholds • Geospatial & Statistical Analysis: Spatial joining and aggregation of 5878 ZIP codes to census tract scale • K-Means Clustering: Identify spatial PFAS risk groupings utilizing 3 optimal clusters, determined by Elbow and Silhouette methods • Machine Learning Models: Random Forest (RF) with number of trees = 200 to 300, Extreme Gradient Boosting (XGBoost), models ran 80/20 sample split for train/test Training Set: 80% Sam ples Testing Set: 20% Sam ples Figure 2: UC MR sampling locations across CON US divided by training and testing sets used for RF and XG Boost models • K-Means Clustering: Moderate PFAS risk (Cluster 1) are ZIP codes with average PFAS levels, climate stress, and mid-level population; High PFAS risk (Cluster 2) have elevated PFAS levels, higher population, and more military sites; Low PFAS risk (Cluster 3) have low PFAS levels, lower population, and less industrialized areas • Clustering pattern generally follows national climate patterns as high risk dominates the south with hotspots scattered amongst the northeast corridor, and moderate risk is distributed amongst the north with hotspots focused across the Midwest • The top 5% of vulnerable populations reside in the western half of CONUS, overlapping

FAQs

What is the main objective of the study?

Evaluate how hydrologic and climate variables influence PFAS concentrations and identify potential PFAS hotspots across CONUS.

How are PFAS concentrations constructed for analysis?

Concentrations of multiple PFAS compounds are averaged to produce a single composite PFAS concentration for comparison across locations.

What clustering approach is used to define PFAS risk groups?

K-means clustering with 3 optimal clusters, selected via Elbow and Silhouette methods, groups ZIP codes into low, moderate, and high PFAS risk patterns.

Sep 17, 2026

Uploaded by Skyler