Mechanistic insights into sulfamethoxazole removal in activated sludge systems through machine learning and microcosm...

Antibiotic residues in wastewater pose serious ecological and public health risks and may even induce the more severe threat of antimicrobial resistance (AMR). The activated sludge (AS) processes, the backbone of global wastewater treatment, play a central role in mitigating these risks. However, antibiotic removal in AS systems remains highly variable and mechanistically unclear. Here, sulfamethoxazole (SMX) was selected as a representative antibiotic to develop a comprehensive machine learning (ML)-based framework for identifying key factors and elucidating process-level mechanisms governing SMX biodegradation in AS processes using an integrated dataset from lab-, pilot-, and full-scale systems. The random forest regression (RFR) model outperformed other models and achieved the highest predictive performance (testing R² = 0.83) after full-factor grid optimization, 5-fold cross-validation and feature engineering, with strong generalization confirmed by independent (prediction error < 20 %). Feature attribution, dependency, and causal analyses revealed that process control parameters (43 %) dominated SMX removal, followed by influent (30 %) and process performance parameters (27 %). Key factors including acclimation period (AP), pH, influent SMX, and ammonia nitrogen (NH₄⁺-N) were identified, with threshold values (15 days, 7.3-8.0, 0.2 mg/L, 43 mg/L) revealed for the first time using the RFR model. Model-identified threshold-like effects of influent SMX concentration and AP governed the initiation of biodegradation, while influent NH₄⁺-N regulated SMX removal via co-metabolism. Causal inference further indicated that NH₄⁺-N removal rate was a stronger performance indicator of AS processes for SMX biodegradation than COD removal rate. The 60-day microcosm validation confirmed that the RFR-guided acclimation achieved near 100 % SMX removal (0.2 mg/L-10 mg/L) and supported the biological plausibility of the model-inferred threshold-like behavior, providing mechanistically informed insights beyond model. This framework integrates data-driven prediction with mechanistic understanding, providing a generalizable approach and offering actionable guidance for optimizing AS processes to mitigate antibiotic pollution.