Yield Prediction Through the Event Sequence Analysis of the Die Attach Process

Abstract

Die attach is the process of mounting a plurality of dice to a printed circuit board (PCB) or substrate. Die attach is critical to the thermal and electrical performance of semiconductor products, significantly affecting the final yield of PCBs. In general, the die attacher records alarm events, change events, and maintenance events in a log. Alarm events occur when dice are not aligned well to the mounting positions on a PCB. Change events are recorded when product types are changed or raw materials of different suppliers are introduced. Maintenance events are recorded whenever the workers conduct corrective actions due to alarm events. We empirically observed that different sequences of events have different effects on the final yield. In this paper, we propose a data mining approach that predicts the final yield of a PCB using the event sequences recorded in the log of the die attacher. We propose a predictive association rule considering the event sequence (PARCOS) algorithm that creates a set of rules, in which each rule estimates the yield for a sequence of events. An experiment with a work-site dataset demonstrated that the PARCOS algorithm had a yield prediction accuracy that was at least 9% higher than those of the regression models that did not consider the event sequences.