Experimental and ALO-optimized machine learning interpretable models for pharmaceutical adsorption onto raw...

Area of Science:

• Environmental Science
• Materials Science
• Computational Chemistry

Background:

• Pharmaceuticals in aquatic systems are persistent environmental pollutants.
• Conventional wastewater treatment struggles with complete removal of these bioactive compounds.

Purpose of the Study:

• To investigate the adsorptive removal of ketotifen fumarate (KF), doxycycline hyclate (DXC), and nystatin (Nyst) using raw bentonite (RB).
• To evaluate adsorption kinetics, equilibrium, and thermodynamics using batch experiments and machine learning.
• To develop an interpretable ML model predicting adsorption capacity and linking molecular structure to removal efficiency.

Main Methods:

• Batch adsorption experiments were conducted with RB for KF, DXC, and Nyst.
• Adsorption data were analyzed using isotherm, kinetic, and thermodynamic models.
• Four Ant Lion Optimizer (ALO)-optimized machine learning models (ANN, SVR, RF, XGBoost) were trained to predict adsorption.
• SHAP (SHapley Additive exPlanations) analysis was used for model interpretability.

Main Results:

• RB demonstrated high adsorption capacities: 178.86 mg/g (KF), 222.91 mg/g (DXC), and 190.25 mg/g (Nyst).
• Freundlich isotherm best described equilibrium, indicating multilayer adsorption on a heterogeneous surface.
• XGBoost model achieved high accuracy (R²=0.972), with SHAP analysis identifying adsorbent dosage, initial concentration, pH, and molecular size (nC) as key factors.
• Adsorption mechanisms involved electrostatic attraction, hydrogen bonding, cation exchange, and van der Waals forces.

Conclusions:

• Raw bentonite is a cost-effective and scalable adsorbent for pharmaceutical removal.
• The interpretable ML framework successfully linked pharmaceutical molecular structure to adsorption behavior.
• This approach offers a promising strategy for understanding and mitigating pharmaceutical contamination in aquatic environments.