Optimal laser-cutting parameters for QFN packages by utilizing artificial neural networks and genetic algorithm

Abstract

In this paper, a multiple regression analysis (MRA) and an artificial neural network (ANN) were employed to build a predicting model for cutting Quad Flat Non-lead (QFN) packages by using a Diode Pumped Solid State Laser (DPSSL) System. The predicting model includes three input variables of the current, the frequency and the cutting speed, and six cutting qualities of depths of the cutting line, widths of heat affected zone (HAZ) and cutting line for epoxy and for copper-compounded epoxy. After the training process from 27 sets of training data including input data and its output qualities, the average training error is 0.822% by using a back-propagation (BP) neural network with Levenberg–Marquardt (LM) algorithm, which leads to the best results. The testing accuracy is then verified with extra 14 sets of experimental data and the average predicting error is 1.512%. The results show that the ANN model has the predicting ability to estimate the laser-cutting qualities of QFN packages. Finally, a genetic algorithm (GA) is applied to find the optimal cutting parameters that lead to least HAZ width and fast cutting speed with complete cutting. The optimal combination found is the current of 29 A, the frequency of 2.7 kHz and the cutting speed of 3.49 mm/s. The GA is helpful to determine the ideal laser-cutting parameters in order to meet the desired cutting qualities and to avoid unnecessary adjustments in the subsequent cutting process.