Ultrapure water (UPW) is indispensable for semiconductor manufacturing, where even trace oxidants can compromise metal stability and device reliability. In the polishing stage, vacuum-ultraviolet (VUV) and ultraviolet (UV) irradiation at wavelength...
Ultrapure water (UPW) is indispensable for semiconductor manufacturing, where even trace oxidants can compromise metal stability and device reliability. In the polishing stage, vacuum-ultraviolet (VUV) and ultraviolet (UV) irradiation at wavelengths of 185 and 254 nm effectively degrades low-molecular-weight organic compounds (LMWOCs) but simultaneously promotes hydrogen peroxide (H₂O₂) formation through radical pathways. In this study, we systematically investigated how pH, hydraulic retention time (HRT), and LMWOC identity govern H₂O₂ accumulation under oxic conditions (DO = 8.5 mg/L). Experiments were conducted using four representative LMWOCs—methanol, isopropyl alcohol, acetonitrile, and nitromethane—over a concentration range of 1–10 μM, pH 4–10, and HRTs of 30–60 s. Lower pH, longer HRT, and alcohol-based organics consistently enhanced H₂O₂ formation, whereas acetonitrile and nitromethane resulted in comparatively lower levels, reflecting their distinct radical reactivity.
To generalize these relationships, we developed a machine learning framework that integrates operational variables (pH, HRT, concentration, molecular weight) with molecular descriptors (hydroxyl radical reaction rate constant, polar surface area, redox potential). We organized these inputs into eight feature sets by systematically combining the molecular descriptors with the operational variables. Among four algorithms evaluated—Gradient Boosting (GB), Random Forest, Gaussian Process Regression, and Support Vector Regression—GB exhibited the highest predictive performance. All four algorithms were evaluated across all feature sets under an identical training/validation procedure with Bayesian hyperparameter optimization. Feature sets incorporating all descriptors produced the strongest agreement with the experimental data, and the GB model trained on this full-descriptor feature set achieved the best generalization performance and was selected as the final predictive model. Global sensitivity analysis using the Morris method identified pH as the most influential variable, followed by HRT and k•OH, highlighting the joint importance of acidity control, residence time, and radical reaction propensity. Overall, this study combines mechanistic understanding and data-driven modeling to provide a practical predictive tool for minimizing H₂O₂ formation and improving process reliability in UPW production.