Experimental and interpretable machine learning analysis of sulfamethoxazole adsorption on mineral-modified activated carbon

Sulfamethoxazole (SMX), a widely detected sulfonamide antibiotic in aquatic and terrestrial environments, poses long-term ecological risks due to its persistence and potential to induce antibiotic resistance. Adsorption onto carbonaceous materials has been widely applied for controlling the migration and environmental fate of antibiotics in soil-water systems. In natural environments, carbonaceous materials frequently interact with mineral phases, forming mineral-carbon associations that can significantly influence adsorption behavior. However, the combined effects of material properties and environmental conditions on antibiotic adsorption remain difficult to quantify using conventional mechanistic approaches. In this study, eight types of mineral-modified activated carbons were prepared by loading four typical minerals (montmorillonite, kaolinite, quartz sand, and hematite) at 5% and 10% mass fractions. The structural parameters of the synthesized samples ranged within 0.14 < V micro < 0.23 cm 3 /g, 0.44 < V meso < 0.61 cm 3 /g, and 734.4 < S BET < 966.3 m 2 /g. An interpretable machine learning (IML) workflow integrating batch adsorption experiments with Bayesian optimization-assisted CatBoost (BO-CatBoost) modeling was developed. The model incorporated six key input variables: material structural parameters (S BET , V meso , V micro ) and operating conditions (adsorption time, initial pH, stirring speed). Evaluated through a rigorous 7-fold group cross-validation combined with independent external validation, the optimized BO-CatBoost model achieved robust predictive accuracy, yielding an R 2 of 0.8153 on the test set and 0.5799 on the external validation set. SHAP analysis revealed that operational conditions exerted a more dominant influence on SMX adsorption efficiency than intrinsic pore-structure descriptors. Partial dependence analysis further identified nonlinear interactions and synergistic regimes among key variables, elucidating the coupled effects of mass-transfer limitations and material properties. Overall, this study provides mechanistic insights into the adsorption behavior of antibiotics on mineral-carbon composite systems and demonstrates the potential of IML for guiding the rational design and optimization of adsorbents for emerging contaminant control in complex environmental media.