Prediction of surface roughness in duplex stainless steel face milling using artificial neural network

This study proposes the inclusion of noise variables in an experimental design to develop a predictive model of surface roughness in the face milling process of duplex stainless steel. A central composite design arrangement was conducted, incorporating controlled variables (cutting speed, feed rate, milling width, and depth of cut) and noise variables (tool flank wear, fluid flow, and protrusion length). Each experimental configuration was employed in duplex stainless steel milling, with the collection of roughness data under each condition. The collected data were used to train eight configurations of artificial neural networks, which were then applied to predict roughness. The results indicate that the 7-20-14-1 network configuration exhibited the lowest root mean square error, which is a measure of the difference between predicted and observed values of (0.063), followed by 7-64-32-1 (0.064) and 7-14-12-1 (0.068), respectively. Additionally, these configurations also demonstrated the lowest mean absolute error values, which calculate the average of the absolute differences between predicted and observed values of (0.046, 0.053, and 0.055, respectively), and the coefficient of determination, which is a statistical measure indicating the proportion of data variability explained by the statistical model of (0.914, 0.908, and 0.901, respectively). Therefore, the inclusion of noise variables alongside controllable process factors resulted in a more accurate and robust predictive model of surface roughness for duplex stainless steel face milling.