Fuzzy Few-Shot Learning for Predictive Maintenance and Cost-Benefit Analysis in Semiconductor Manufacturing: A Case Study on Photomask Haze

Abstract

Predictive Maintenance (PdM) plays a key role in production management by extending the lifespan of components and reducing maintenance costs. In the manufacture of semiconductors, the scarcity of data related to haze defects makes it difficult to draw correlations between environmental factors and haze formation. The uncertainty associated with a reliance on indirect evidence (i.e., cumulative time in the environment) has seldom been explored in the literature. Therefore, we developed a PdM framework based on fuzzy few-shot learning to deal with photomask haze in the semiconductor industry. The robustness of the model was evaluated in a three-month pilot study conducted in a wafer foundry. This paper also provides cost-benefit analysis of model implementation. The results demonstrate that the proposed haze photomask detection model outperforms existing models in terms of hit rates and false alarm rates. It also proved effective in lowering labor costs and power consumption, as evidenced by the fact that the number of haze candidates was well below the daily inspection cap, which allowed the decommissioning of one photomask inspection device, reducing the photomask inspection volume by 21%, which is equivalent to an annual labor cost reduction of USD 18,780. These results should help to promote the adoption of predictive maintenance applications, even in situations with small sample sizes.

Keywords

• Fuzzy few-shot learning
• photomask haze
• predictive maintenance
• imbalanced data
• cost-benefit analysis