Root Cause Detection of Excursion—An Empirical Study for Semiconductor Manufacturing

In semiconductor manufacturing, various low-yield wafers are produced inevitably. Determining the root cause of low yield is an important procedure for preventing production of low-yield wafers as well as providing feedback of process modifications. However, the procedures are complicated due to broad data to investigate and an insufficient number of faulty wafers to compare. In addition, failure analysis includes various investigating fields, such as process step information, measurement data, inspection maps, delay time, equipment error, source change, etc. Therefore the strategic examination of possible causes is required to increase the efficiency of quality control.<br/>Failure analysis of anomalies starts with defining the group of faulty wafers. Afterwards, process modifications and operating equipment assignment are reviewed generally. In some cases, causes of anomalies are not revealed by general analyzing procedures and some features of issues indicate that each process in equipment of individual wafer is needed to be screened. In this study, we propose the addition of comparing trace sensor data across the whole manufacturing processes for detecting the faulty process. Pre-processing includes labeling, process sequencing, and signal adjusting. Besides, the sensor data are enormous even for a few wafers because they pass hundreds steps during the fabrication. Therefore the comparison between faulty wafers and normal wafers requires prioritizing technique with proper scoring.<br/>As an empirical case, we analyzed an example, with a few intermittent faulty wafers having a specific fail mode, by the proposed method. The priority of sensor signal comparison was high since only a few wafers were defective and no changed process were matched with the course of the wafers group. Moreover, the measurement and inspection data of good and bad wafers were not sufficient in the early stage. Based on binary groups, we compared sensor signals with the capable computing tools combining pre-processing, comparing and prioritizing. Accordingly, the root cause process was identified by distinctive signal differences at a specific step. Consequently, the unstable point of the process was clarified with signal information and subsequently repaired, using domain knowledge.<br/>In conclusion, the addition of the sensor data comparison for failure analysis provides not only possible causes but also clues to process feedback, since sensor signals of equipment contain process information. This method is particularly advantageous when the problems are equipment related and the trace sensor data are recorded. Furthermore, the sensor signals are commonly available for every wafer or lot in consideration of recent progress of equipment sensors in the semiconductor manufacturing industry. Therefore, in the early stage of issue when only a few abnormal wafers are detected, this method is possibly used as a powerful tool to discover the root cause of excursion.