Improving Wirebonding Failure Classification Across Bond Recipe via Machine Learning

Wirebonding is an essential process in the Integrated Circuits (IC) packages assembly. Failures can occur during the bonding process, and they are mainly detected via electrical currents. In this paper, we study one of the common failures, Non-Stick-On-Pad (NSOP). The three main failure modes of NSOP are Capillary-Marked-On-Die (CMOD, caused by capillary tip striking the pad directly), equipment failure, and foreign material contamination. Equipment failure can be detected by anomaly detection while foreign material can be visually inspected. Hence, in the context of this paper, we focus on CMOD classification. We explore three regularised machine learning models, namely logistic regression, decision tree, and XGBoost by using the process signals extracted from the wirebonding machine. Our best model can achieve accurate classification of CMOD and also explain the contributing process signals, which enables the engineers to better identify the root causes and perform necessary corrective actions, thereby reducing yield loss. In addition, we apply transfer learning techniques to transform the distinctive process signals from different recipes into a similar feature space for training a general model.