Decoding pharmaceutical removability in full-scale wastewater treatment plants via machine learning model integrating...

Assessing the pharmaceutical removal potential in wastewater treatment plants (WWTPs) is critical for risk management, but remains technically challenging. In this study, a machine learning (ML)-based framework was developed to predict pharmaceutical removability using over 4000 spatiotemporal measurements of 80 pharmaceuticals across 17 full-scale WWTPs. Modified quantum chemistry (QC) descriptors, which couple electronic properties with treatment processes, were integrated with conventional molecular descriptors and WWTP parameters for modeling. The developed classification model achieved satisfactory predictive performance with an accuracy of 0.81 and demonstrated robust generalizability in external validation. Model interpretation revealed that site-specific electronic features significantly improved predictive performance, with the modified electrophilicity index emerging as the dominant driving factor of pharmaceutical removability in WWTPs. Pharmaceuticals with high molecular stability and low electro-surface activity tended to exhibit low removability, particularly under conditions of high site-specific electrophilicity and lower temperatures. Scenario-dependent predictions from the model provide valuable guidance for optimizing treatment strategies. Specifically, pharmaceuticals such as carbamazepine, diazepam, and citalopram exhibited higher predicted amenability to chemical treatment, whereas mifepristone, erythromycin, and sulfacetamide were more likely to be effectively removed under biological processes. Extended sludge and hydraulic retention time, along with optimized influent loading, were expected to further enhance the removal of persistent pharmaceuticals. This study offers predictive insights into pharmaceutical removal potential through the developed ML model and provides scientific support for pharmaceutical management in WWTPs.