Advancing Predictive Accuracy in Heavy Metal Photocatalysis through Systematic Optimization of Neural Network Architectures

Abstract

The efficiency of metal oxide photocatalysis is constrained by complex, nonlinear interactions among variables that are difficult to predict via traditional experimental methods. This study develops a robust Artificial Neural Network (ANN) model to identify the optimal training algorithm and architecture for predicting heavy metal removal efficiency. A feed-forward network was trained with 100 datasets, comparing Levenberg-Marquardt (LM), Resilient Backpropagation (Rprop), and Scaled Conjugate Gradient (SCG) algorithms. Configurations of 10, 15, and 20 hidden neurons were systematically tested to optimise the network's predictive performance. The LM algorithm with 20 hidden neurons achieved superior results, yielding a high coefficient of determination (R2) of 0.9935 and a minimal Mean Squared Error (MSE) of 0.0022893. A low average percentage error of 2.46% across diverse datasets further confirms that the model successfully reached an optimal state without overfitting. This high predictive accuracy validates the network's effectiveness in capturing the nonlinear relationships within the photocatalytic system. Optimised ANNs provide a powerful, sustainable framework for enhancing the operational efficiency of industrial wastewater treatment applications. Future research will focus on incorporating larger datasets and refining the network further to improve model predictive precision.