Yield prediction in semiconductor manufacturing using two-step machine learning

Optimising yield is a key challenge in semiconductor manufacturing. Accurate yield predictions early in the production process enable effective interventions, improving both quality and efficiency. This study examines a two-step approach to enhance yield prediction accuracy by excluding outliers from Fault Detection and Classification (FDC) data. The methodology involves two stages: first, outcomes are classified into ‘good’ and ‘defective’ categories; then, yield predictions are based solely on the ‘good’ data. The research investigates whether this approach improves the precision of yield forecasts. Preliminary results demonstrate that this method predicts electrical yield with a coefficient of determination (R2) of 0.496 using data from only six production steps. Additionally, reducing the number of parameters further improves the R2 value to 0.538. These findings suggest that excluding outliers and focussing on high-quality data can significantly enhance prediction accuracy while requiring less data. This paper assesses the effectiveness and advantages of the proposed two-step system, aiming to refine yield predictions and optimise production processes. The approach highlights the potential for improving semiconductor manufacturing outcomes by leveraging targeted data analysis to address key challenges efficiently.