Modeling of adsorption capacity, reaction kinetics and thermodynamic studies on Ni(II) removal with GO@Fe3O4@pluronic-F68 nanocomposite via machine learning methods

This study presents a comprehensive machine learning (ML) modelling framework for predicting the adsorption behaviour of Ni2 + ions onto the GO@Fe3O4@Pluronic-F68 nanocomposite. Fourteen ML regression algorithms spanning linear, support vector, tree-based, and ensemble methods were trained and compared using experimental batch adsorption data across contact time, pH, adsorbent dose, initial concentration, and temperature. Domain-informed features — a logarithmic time transformation (log time) capturing nonlinear kinetics and a concentration-to-dose ratio (C0/dose) encoding the pollutant load–capacity balance — were introduced. Model performance was evaluated with an 80/20 train–test split and 5-fold cross-validation on R2, RMSE, MAE, and MAPE metrics. The SVR-RBF model achieved the highest hold-out test accuracy for both residual Ni2+ concentration (R2 = 0.9711) and adsorption capacity (R2 = 0.9694); under nested cross-validation, however, Gradient Boosting (nested-CV R2 = 0.968 on qe) matched or slightly exceeded SVR-RBF (0.940), leaving the two as jointly top-tier candidates whose ranking depends on the evaluation protocol. Feature importance identified C0/dose and log time as the most influential predictors — mechanistically consistent with Freundlich isotherm dominance and pseudo-second-order kinetics. Residual analysis confirmed no systematic bias, and 3D response-surface modelling identified optimal conditions at C0 ≥ 300 mg L−1 and t ≥ 60 min. Within the sampled laboratory design space, the integrated experimental-ML framework serves as a laboratory-scale decision-support tool for nanocomposite-based Ni2+ adsorption; its extension to industrial wastewater treatment requires independent-laboratory validation, real-matrix experiments, adsorbent-regeneration cycling, and pilot-scale trials.