Modeling of Fixed-Bed Column System of Hg(II) Ions on Ostrich Bone Ash/nZVI Composite by Artificial Neural Network

Determination of removal percentage (RP) of pollutants in a fixed-bed column system is time-consuming, difficult, and subject to errors. To overcome this problem, an artificial neural network (ANN) with different learning algorithms, activation functions, input variables, neurons in the hidden layers, and number of hidden layers was employed. For this purpose, the RP of Hg(II) ions by ostrich bone ash–nanoscale zero-valent iron composite (OBA/nZVI), as a novel adsorbent, was measured in a fixed-bed column experiment. Four effective variables, including inflow rate (F), initial pollutant concentration (C), bed height (Z), and pH were taken as input data and the RP of the composite was taken as output. Four ANN models, including different combinations of effective variables, were constructed to reveal the sensitivity analysis of the models. Normalized root mean square error (NRMSE), mean residual error (MRE), and linear regression ($R^2$ ) were used as criteria for comparison of estimated data by the models and the experimental data. Results indicated that the ANN4 model, comprising a trainlm learning algorithm and a log sigmoid activation function with all four input data, accomplished the best prediction of RP ($R^2=0.996$ , $\mathrm{NRMSE}=0.028$ , $\mathrm{MRE}=0.008$ ). The sensitivity analysis indicated that the predicted RP is more sensitive to pH, followed by F, Z, and C. This study demonstrated that the ANN model can be a more accurate and faster alternative to the laborious and time-consuming laboratory measurements for RP of Hg(II) ions in a fixed-bed column system.