A Neural-Network Approach to Better Diagnosis of Defect Pattern in Wafer Bin Map

Wafer bin map (WBM) represents specific defect patterns that provide information for diagnosing root causes of low yield in semiconductor manufacturing. In practice, most semiconductor engineers use subjective and time-consuming eyeball analysis to assess defect patterns. Given shrinking feature sizes, various types of WBMs with different defect patterns occur; therefore, relying on human vision to judge defect patterns become more complicated, inconsistent, and unreliable. To bridge the gap, a system is proposed to facilitating WBM patterns extraction and assisting engineer to recognize defect patterns efficiently. We propose an individual classification model which is trained by Deep Belief Network (DBN) to diagnose the patterns on wafer bin maps. By setting up six single classifiers with different thresholds, the individual classifiers were combined for both single and mixed-type patterns recognition. The numerical results showed that the single classifiers outperform MLP method both on single and mix-type patterns. Simultaneously, the results have shown the validity and practical viability of the proposed combined classifiers.