Explainable Machine Learning and Mechanistic Interpretation of Congo Red Removal by CTABr-Modified Bio-Apatite-Based Adsorbent

Waste-derived adsorbents represent a promising strategy for sustainable wastewater treatment; however, adsorption processes are frequently governed by complex nonlinear interactions that remain difficult to interpret using conventional approaches. The present study developed an integrated experimental and explainable machine learning framework to investigate Congo red (CR) removal using cetyltrimethylammonium bromide (CTABr)-modified bio-apatite-based adsorbent. The adsorbent was characterized using SEM–EDX, FTIR, XRD, and thermal analyses, confirming a heterogeneous Ca–P-rich structure associated with hydroxyapatite and CTABr-derived surface functionalities. A dataset comprising batch adsorption experiments was generated by varying solution pH, contact time, initial CR concentration, CTABr concentration, adsorbent dose, and temperature. The experimental results demonstrated that CR removal efficiency ranged from approximately 22.9% to 100%, with the experimental conditions giving the highest capacity (pH ≈ 2, contact time ≈ 120 min, initial CR concentration ≈ 10 mg L−1, CTABr concentration ≈ 0.5 meq g−1, adsorbent dose ≈ 0.2 g L−1, and temperature ≈ 25 °C). Elastic Net, Random Forest (RF), and Gradient Boosting (GB) models were developed and evaluated using nested repeated five-fold cross-validation. The nonlinear ensemble models outperformed the regularized linear approach, indicating that is associated with complex nonlinear relationships among operational variables. RF provided the best predictive performance (R2 = 0.675 ± 0.167, RMSE = 14.63 ± 5.81 percentage points) and was subsequently interpreted using SHapley Additive exPlanations (SHAP). Explainable analysis identified initial CR concentration as the dominant predictor within the fitted model and the investigated experimental domain, followed by contact time, adsorbent dose, pH, CTABr concentration, and temperature. SHAP dependence analysis revealed nonlinear behaviors, including concentration-dependent suppression, optimum-like dose effects, and CTABr saturation responses. Integration of characterization and modeling results suggested a multimodal adsorption mechanism involving electrostatic attraction between CTABr-derived quaternary-ammonium groups and CR sulfonate groups, complemented by interactions with the hydroxyapatite matrix, hydrogen bonding, and hydrophobic interactions. The proposed framework demonstrates how explainable machine learning can bridge predictive modeling and adsorption process interpretation while supporting the valorization of waste-derived materials for water treatment.