MACHINE LEARNING–GUIDED CHEMICAL ENGINEERING SOLUTIONS FOR PFAS REMOVAL FROM DRINKING WATER: PREDICTIVE TOXICITY ASSESSMENT AND REDUCTION OF MODELED CANCER AND LIVER DISEASE RISKS

Abstract

Background: Per- and polyfluoroalkyl substances (PFAS) are persistent contaminants that resist conventional drinking-water treatment and are associated with adverse health outcomes. Coupling advanced treatment with machine learning may improve removal, lower resource demand and allow compound-specific, exposure-based risk comparison.

Aim: To develop an ML-guided chemical-engineering framework for optimizing PFAS removal from drinking water, identify influential operational and water-quality predictors, and estimate modeled reductions in toxicity-weighted cancer and liver-related risk.

Methods: Nine PFAS (PFOA, PFOS, PFHxS, PFNA, PFHxA, PFBS, PFBA, PFHpA, and PFDA) were evaluated across six treatments: granular activated carbon (GAC), ion exchange (IX), nanofiltration (NF), reverse osmosis (RO), electrochemical oxidation (EO) and hybrid systems. Five regressors (Random Forest, support vector regression, artificial neural network, LightGBM, XGBoost) predicted removal efficiency using a batch-grouped train/validation/test design with grouped cross-validation. SHAP, permutation importance and partial dependence explained the best model, which then drove multi-objective optimization. Hazard quotients, a hazard index and modeled lifetime cancer and liver-related risks were derived from estimated daily intake under a central (mean effluent) and a conservative effluent scenario.

Results (modeled): Total PFAS removal was 82% (GAC), 92% (IX), 94% (NF), 96% (RO) and 97.1% (optimized hybrid); EO (88%) is a placeholder. XGBoost performed best (test R² = 0.96, MAPE = 3.1%). Key predictors were PFAS molecular characteristics, influent concentration, dissolved organic carbon (DOC), adsorbent dosage, and membrane pressure and contact time. Optimization lowered modeled energy use by 19%, chemical use by 22% and operating cost by 17%. Under conservative effluent inputs, modeled cumulative non-carcinogenic hazard fell by 93%, lifetime cancer risk by 90% and liver-related risk by 91%; with mean modeled effluent the reductions were 97.7% for all three endpoints. The general hazard index (3.36 conservative; 1.13 central) and cancer risk (6.8 × 10⁻⁶; 1.5 × 10⁻⁶) remained above 1 and 10⁻⁶, respectively, in both scenarios.

Conclusion: Under the stated assumptions, ML-guided hybrid treatment offers high PFAS removal with lower resource demand and substantially lower modeled exposure-based risk. Independent experimental and pilot-scale validation is required before any performance claim.