Predictive models for hydrocyclone performance parameters: Evaluating the impact of training functions using artificial neural network

This study introduces a multilayer perceptron artificial neural network model to predict water split, cut size and sharpness of separation in hydrocyclone classifiers. The model is developed based on experimental data from 75, 100, 150 and 250 mm diameter hydrocyclones under different operating conditions and design parameters. The Levenberg–Marquardt training algorithm is found to be optimal based on the correlation coefficient and mean square error values. Pearson's correlation coefficient is used to assess the strength and direction between variables. The importance of input variables is estimated using the out-of-bag permuted predictor method. The ANN model employs 10, 10 and 8 hidden neurons achieving regression coefficients of R ∼ 0.92, R ∼ .99, R ∼ 0.90 and MSE values of 21.98, 38.915, 0.34 for water split, cut size and sharpness of separation, respectively. Comparative analysis indicates that the ANN model shows superior performance than empirical and semi empirical models.