A generalised approach on kerf geometry prediction during CO2 laser cut of PMMA thin plates using neural networks

Abstract This study presents an application of feedforward and backpropagation neural network (FFBP-NN) for predicting the kerf characteristics, i.e. the kerf width in three different distances from the surface (upper, middle and down) and kerf angle during laser cutting of 4 mm PMMA (polymethyl methacrylate) thin plates. Stand-off distance (SoD: 7, 8 and 9 mm), cutting speed (CS: 8, 13 and 18 mm/sec) and laser power (LP: 82.5, 90 and 97.5 W) are the studied parameters for low power CO2 laser cutting. A three-parameter three-level full factorial array has been used, and twenty-seven (33) cuts are performed. Subsequently, the upper, middle and down kerf widths (Wu, Wm and Wd) and the kerf angle (KA) were measured and analysed through ANOM (analysis of means), ANOVA (analysis of variances) and interaction plots. The statistical analysis highlighted that linear modelling is insufficient for the precise prediction of kerf characteristics. An FFBP-NN was developed, trained, validated and generalised for the accurate prediction of the kerf geometry. The FFBP-NN achieved an R-all value of 0.98, in contrast to the ANOVA linear models, which achieved Rsq values of about 0.86. According to the ANOM plots, the parameter values which optimize the KA resulting in positive values close to zero degrees were the 7 mm SoD, 8 mm/s CS and 97.5 W LP.