Comparative evaluation of artificial neural networks and support vector machines for predicting heavy metal adsorption dynamics in fixed beds

Abstract

This research examines the predictive capabilities of Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) as Machine Learning (ML) techniques for accurately modeling 10,668 data points derived from the breakthrough curves of 13 different heavy metals observed during dynamic adsorption in fixed-bed columns. Four specific models were designed and assessed: ANN1 (a one-hidden-layer ANN with logarithmic sigmoid hidden and tangent hyperbolic output activation functions); ANN2 (a two-hidden-layer ANN using logarithmic sigmoid in both hidden layers and tangent hyperbolic in the output); a conventional SVM; and an SVM-DA (SVM hybridized with the dragonfly optimization algorithm) utilizing an RBF-Gaussian kernel function. Achieving a remarkable correlation coefficient (R) of 0.9986, a root mean square error (RMSE) of 0.0209, and an average absolute relative deviation (AARD) of 24.608%, the ANN2's predicted breakthrough curves exhibited exceptional proximity to experimental observations, underscoring its robust performance in predicting dynamic adsorption phenomena. Furthermore, a sensitivity analysis employing the inverse artificial neural network method was conducted to quantify the influence of all input variables on the adsorption process. For enhanced practical utility and improved result traceability, a Graphical User Interface (GUI) was developed.