Optimization of Azare low-grade barite beneficiation: comparative study of response surface methodology and artificial neural network approach

This study examined the efficacy of response surface methodology (RSM) and artificial neural network (ANN) optimization approaches on barite composition optimization from low-grade Azare barite beneficiation. The Box-Behnken Design (BBD) and Central Composite Design (CCD) approaches were used as RSM methods. The best predictive optimization tool was determined via a comparative study between these methods and ANN. Barite mass (60–100 g), reaction time (15–45 min) and particle size (150–450 μm) at three levels were considered as the process parameters. The ANN architecture is a 3-16-1 feed-forward type. Sigmoid transfer function was adopted and mean square error (MSE) technique was used for network training. Experimental data were divided into training, validation and testing. Batch experimental result revealed maximum barite composition of 98.07% and 95.43% at barite mass, reaction time and particle size of 100 g, 30 min and 150 μm; and 80 g, 30 min and 300 μm for BBD and CCD respectively. The predicted and experimental barite compositions of 98.71% and 96.98%; and 94.59% and 91.05% were recorded at optimum predicted point for BBD and CCD respectively. The analysis of variance revealed high significance of developed model and process parameters. The correlation of determination recorded by ANN for training, validation and testing were 0.9905, 0.9419 and 0.9997; and 0.9851, 0.9381 and 0.9911 for BBD and CCD. The best validation performance was 48.5437 and 5.1777 at epoch 5 and 1 for BBD and CCD respectively. In conclusion, the overall mean squared error of 14.972, 43.560 and 0.255; R2 value of 0.942, 0.9272 and 0.9711; and absolute average deviation of 3.610, 4.217 and 0.370 recorded for BBD, CCD and ANN respectively proved ANN to be the best.