Wafer-Level Root Cause Detection with Association Rule Mining Based on Recalculated Metrics and Interaction Effects

Abstract

In the semiconductor manufacturing process, yield drops and wafer quality degradation can occur, making it essential to identify the root causes. Key factors that impact wafer quality include processing steps, equipment, and recipes. A single factor may be responsible, while multiple factors may interact synergistically at other times, complicating accurate analysis. Traditionally, most semiconductor defect detection has relied on engineers’ experience and domain expertise, so it is challenging to analyze defect-causing equipment and interaction effects, often leading to failures in root-cause analysis. We propose a pattern mining-based root cause detection methodology that identifies interaction effects between the process steps using defective wafers as input data. By leveraging the traditional frequent item detection algorithm from Apriori-based association analysis, we can identify suspect paths that are likely to cause defects based on the frequency of defective wafers passing through. Due to performance issues with the naive version that processed both bad and good wafers, frequent item sets were extracted from bad wafers, and the association rule metrics were recalculated through a good wafer count search. This approach not only detects the paths causing defects by analyzing the commonalities and differences between defective and high-yield wafers but also quantifies the interaction effects between root cause factors. The experimental results demonstrate that the proposed root cause search methodology identifies defect root causes and verifies interaction effects based on 200 semiconductor defect wafers. Additionally, the improvements over the naive version, which applied association rule mining, are shown by a 95.27-fold improvement in runtime and a 28.97-fold improvement in memory usage.