Comparative study of machine learning models for removal of As(III) from potable water using spiral-wound nanofiltration and analyzing input responses

Abstract

An attention toward health disorder and numerous health illnesses is the root cause of increase in heavy metal contaminants such as Arsenic(III) in potable water. An effective remedy for this problem is a nano-filtration membrane, which is affordable and doesn’t allow heavy metal ions to permeate. However, for better outcomes input parameters should be gauged at an accurate value, which is quite challenging. This study addresses the complexity of employing artificial neural network (ANN) to model the percentage rejection of a nano-filtration membrane using deep learning toolbox in MATLAB. Three different algorithms, i.e., Levenberg-Marquardt, Bayesian regulation, and scaled conjugate gradient, have been used for training, and the best results are shown by Bayesian regulation algorithms and hence selected for this study. The number of neurons in the hidden layer is specified as 10, which provides the mean square error (4.7 * 10^-7) and coefficient of correlation (1), which signifies a well-trained model. Following an examination of trained model by verification and validation then the various input responses response was studied. The optimum percentage rejection of As(III) removal occurred when feed concentration, transmembrane pressure, and feed flow rate were between 30 to
50 mg/L, 5.71 to 7.09 bar, and 12.5 to 17 L/min, respectively, when temperature and pH are under the nominal range, i.e., 303K and 8, respectively.