ANN based RUL assessment for copper-aluminum wirebonds subjected to harsh environments

Abstract

Copper (Cu) wire bonding is new alternative to traditional Gold (Au) wirebonds. Since Cu is not as inert as Au, material selection in the package plays key role in reliability of packages. Researchers have reported individual effect of the variables such as pH value, ionic contamination, and filler content of EMC etc. on reliability of Cu wirebonds. However, since all these parameters have combined effect on reliability, understanding of joint effect of all parameters on reliability of Cu wirebond is necessary for smooth transition to Cu wirebond system. In this paper, predictive model for life prediction of copper wirebond system based on neural network is presented. A set of parts, molded with eight different EMC's were subjected to high temperature environment (temperature range of 150°C-225°C). Resistance, IMC change and shear strength change were monitored during this study. Resistance spectroscopy was used for accurate resistance measurement. Dage 2400PC was used to calculate change in shear strength. Parts were cross-sectioned and polished along Cu-Al interface using SEM and EDX system after the failure. Relation between resistance changes with change in shear strength was established. 20% change in resistance was considered as failure threshold. All parts were tested till failure. Evolution of resistance was considered as leading indicator of failure. Variable selection for the model was done using principle component analysis. Scree plot was used to identify and retain influential variables in the model and to ignore non-significant variables. The shortlisted variables along with resistance evolution and time-to-failure data were used to build predictive model. Neural network regression model was trained with input feature vectors. Supervised learning was used during training. Feedforward multilayer network was trained using Bayesian regularization in conjuncture with Levenberg Marquardt algorithm. Self-validation and cross validations were performed multiple times to avoid overfitting of the data. Prediction model will be able to predict remaining useful life when environmental conditions, properties of EMC and current state of leading indicator are known. This model will provide, educated estimation of remaining useful life (RUL) for Cu wirebonded molded packages, at desired operating condition.