Suspect Recommendation For Memory Devices

During manufacturing of memory devices, there may be various defects which may be introduced at multiple stages of manufacturing such as fabrication, production and testing. These defects such as assembly line defects, external defects, equipment or process defects, may results in the failure of that memory device and may be identified using automated or manual mass production tests. For large-scale manufacturing, manual inspection of these defects is time-inefficient and error-prone. This process of identifying probable causes of failure in embedded memory devices is known as suspect recommendation. In this paper, a machine learning approach to identify defective device lots from non-defective device lots is proposed. The proposed algorithm consolidates the predictions from three different feature importance methods such that critical features affecting the manufacturing memory devices can be identified with confidence. Moreover, class-specific feature importance scores are derived from the three feature importance methods so that critical features affecting defective and non-defective device lots can be detected and a novel way of combining them is proposed. Furthermore, we design a domain specific performance criterion, namely Feature Match Index (FMI) to compare the subject matter expert’s (SME) view with the critical features identified by the proposed solution and achieve an 88.89% match. The proposed solution has been deployed successfully in the large-scale production environment of memory devices.