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

Abstract

The widespread occurrence of per- and polyfluoroalkyl substances (PFAS) in drinking-water systems has become a serious environmental and public-health concern because of their exceptional persistence, bioaccumulation potential, and resistance to conventional treatment processes. This study develops a machine learning–guided chemical engineering framework for predicting PFAS removal performance, optimizing treatment conditions, evaluating compound-specific toxicity, and estimating potential reductions in cancer and liver-disease risks. A harmonized dataset comprising 18,600 laboratory, pilot-scale, and water-quality observations was compiled for 18 commonly detected PFAS compounds. The dataset contained 42 physicochemical and operational variables, including influent concentration, molecular chain length, functional group, pH, temperature, dissolved organic carbon, ionic strength, adsorbent dosage, membrane pressure, contact time, pore size, and treatment configuration. Six treatment approaches were investigated: granular activated carbon, ion-exchange resin, nanofiltration, reverse osmosis, electrochemical oxidation, and hybrid ion-exchange–nanofiltration. After missing-value imputation, outlier removal, normalization, categorical encoding, feature engineering, and leakage-free data partitioning, linear regression, support vector regression, random forest, artificial neural network, LightGBM, and XGBoost models were evaluated. XGBoost produced the best predictive performance, achieving an R² of 0.963, a root-mean-square error of 3.18%, and a mean absolute error of 2.41% for PFAS removal efficiency. The optimized hybrid treatment achieved 97.6% total PFAS removal, outperforming activated carbon at 82.4%, ion exchange at 91.8%, nanofiltration at 94.3%, and reverse osmosis at 95.7%. Machine learning–based process optimization also reduced energy consumption by 18.9%, chemical usage by 21.6%, and estimated operating costs by 16.7% compared with non-optimized operating conditions. SHAP analysis identified PFAS chain length, influent concentration, dissolved organic carbon, adsorbent dosage, membrane pressure, and contact time as the most influential predictors. A toxicity-weighted health-risk module estimated reductions of 93.2% in cumulative non-carcinogenic hazard, 89.7% in modeled lifetime cancer risk, and 91.4% in liver-toxicity risk following optimized treatment. These health outcomes represent modeled risk estimates rather than observed clinical effects. Overall, the proposed framework integrates explainable machine learning, treatment-process optimization, and health-risk assessment to support cost-effective, transparent, and scalable PFAS management in drinking-water systems.