Adaptive burn-in time decision system based on pattern recognition for intelligent reliability control

Abstract

Most semiconductor companies usually screen out early field failures of semiconductor device by conducting burn-in test for all manufactured devices. Burn-in is a production process whereby all manufactured devices operated under accelerated stresses for constant periods of time and accordingly crucial in productivity, on-time delivery and quality of semiconductor device. Many researches on the determination of optimal burn-in time and schedule of its operations have been conducted. Most of them, however, had some limitations to apply to real world because of their complexity and practical difficulties. Our study aims at providing an easy, efficient and more practical alternative approach. We present a multi-agent-based system, called adaptive burn-in time decision system (ABITDS) that predict the reliability level of a newly manufactured semiconductor lot on the basis of fail patterns of previously manufactured lots and then make an adaptive decision of burn-in time according to the predicted reliability level. The ABITDS uses SOM (Self-Organizing Map) neural network to firstly extract the patterns of defective chips within the wafer, wafer bin map patterns, of previously manufactured lots with good, normal and bad reliability level and predict the reliability levels of newly manufactured lots by measuring the similarity degrees between their wafer bin map patterns and the extracted ones. We implemented a web-based ABITDS prototype and validated the effectiveness of our approach through its application to real data of a semiconductor company.