Machine learning modeling and statistical optimization of dye removal from contaminated water using CTAB-functionalized graphene oxide

Water pollution is a rapidly growing problem worldwide. Organic dyes are among the most commonly encountered hazardous chemicals in wastewater streams. Hence, in this study, graphene oxide (GO) was functionalized with cetyltrimethylammonium bromide (CTAB) in order to boost its performance in removing methyl orange (MO), a hazardous azo dye, from synthetic wastewater samples. Characterization techniques including XRD and FTIR analysis indicated the formation of the composite material (CTAB@GO) due to increase in peaks intensity and also a shift in the peaks position relative to the pristine GO. Zeta potential measurements revealed electrostatic attraction as one of the possible reasons for enhanced uptake of anionic MO by GO modified with CTAB. Response surface methodology (RSM) technique was used to model and optimize abatement of MO via the manipulation of key process parameters. The obtained RSM results demonstrated that the increase in the MO initial concentration and the initial pH of the adsorption medium positively affects the adsorptive removal of MO while increasing the CTAB@GO dose caused a drop in the dye removal. Besides the optimization studies, MO adsorption kinetics and isotherm were also conducted to further explore the adsorption behavior of MO onto CTAB@GO. The obtained kinetics and isotherm experimental data showed a better fit to the Avrami kinetics model and Freundlich adsorption isotherm, respectively. Furthermore, the maximum monolayer adsorption capacity computed from the Langmuir model was 1541.9 mg/g and it closely matched maximum experimental adsorption capacity. Additionally, the generated MO adsorption data were used for machine learning modelling exercise and the model generated by the support vector machine (SVM) algorithm showed a better performance in predicting MO adsorption as judged by its higher correlation coefficient (R) and lower RMSE value, 0.9601 and 108.34, respectively, relative to other machine learning models utilized in this study.