Data Science for Delamination Prognosis and Online Batch Learning in Semiconductor Assembly Process

Abstract

The transformation of wafers into chips is a complex manufacturing process involving literally thousands of equipment parameters. Delamination, a leading cause of defective products, can occur between die and epoxy molding compound (EMC), epoxy and substrate, lead frame and EMC, and so on. Troubleshooting is generally on a case-by-case basis and is both time-consuming and labor-intensive. We propose a three-phase data science (DS) framework for process prognosis and prediction. The first phase is for data preprocessing. The second phase uses least absolute shrinkage and selection operator (LASSO) regression and stepwise regression to identify the key variables affecting delamination. The third phase develops a backpropagation neural network (BPNN), support vector regression (SVR), partial least squares (PLS), and gradient boosting machine (GBM) to predict the ratio of the delamination area in a die. We also investigate the imbalance between a false positive rate and a false negative rate after quality classification with BPN and GBM models to improve the tradeoff between the two types of risks. We conducted an empirical study of a semiconductor manufacturer, and the results show that the proposed framework provides an effective delamination prediction supporting the troubleshooting. In addition, for online prediction, it is necessary to determine the batch size for the timing of retraining the model, and we suggest the cost-oriented method to solve the issue.