Comparative analysis of ANFIS, ANN, and BBD for enhanced prediction of methyl orange adsorption in water treatment

This work focused on the enhanced prediction of methyl orange removal (MO) from water by activated carbon synthesized from banana peels. Characterization was done using powder X-ray diffraction (PXRD), Fourier transform infrared spectroscopy (FTIR), thermogravimetric analysis (TGA), scanning electron microscopy (SEM), and Brunauer-Emmett-Teller (BET). Modeling and prediction of process variables, pH (5-9), time (3-60 min), and temperature (25-50 °C), was carried out using Box-Behnken design (BBD), artificial neural networks (ANN), and adaptive neuro-fuzzy inference system (ANFIS). Performance metrics of R², Adjusted R², Pearson's r, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) were used to evaluate the models. The regression coefficients from the modeling and prediction showed that BBD (R² = 0.9849), ANFIS (R² = 0.9934), and ANN (R² = 0.9921), which describes the high prediction capacity of the three models. The performance metrics showed that ANFIS had superior capacity in data modeling and prediction compared to BBD and ANN when analyzing complex non-linear relationships. The Elovich, pseudo-first-order, intraparticle diffusion, and pseudo-second-order kinetic models had high R² values. The data obtained showed that the pseudo-second-order fitted the data well; as such, chemisorption was the most dominant mechanism. In addition, the isotherm models of Freundlich, Temkin, Langmuir, and Dubinin-Radushkevich were determined. The Freundlich model shows the highest R², as such adsorption occurs on heterogeneous multilayer surfaces. This study therefore shows the efficiency of ANFIS, ANN, and BBD in the prediction of dye removal by activated carbon synthesized from banana peels.