Analysis methodology for semiconductor yield by data mining

The conventional semiconductor yield analysis is a hypothesis verification process, which heavily depends on engineers' knowledge. Data mining methodology, on the other hand, is a hypothesis discovery process that is free from this constraint. This paper proposes a data mining method for semiconductor yield analysis, which consists of the following two phases: discovering hypothetical failure causes by regression tree analysis and verifying the hypotheses by visualizing the measured data based on engineers' knowledge. It is shown, through experiment under the real environment, that the proposed method detects hypothetical failure causes, which were considered practically impossible to detect, and that yield improvement is achieved by taking preventive actions based on the detected failure causes.