.png){kind=icon}
.png){kind=icon}
Return
Adaptive Open-World Learning for Analyzing Wafer Map Variations
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.131994.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the semiconductor manufacturing industry, the wafer map is essential for visualizing wafer failure patterns and diagnosing root causes of wafer failures. However, existing researches usually rely on datasets with the binary representation, which cannot capture failure details from different items in the wafer testing process. Besides, researchers use the closed-set classification models to learn information in wafer maps, which are inadequate for the dynamic production environment. In this paper, we introduce the data preprocessing method based on the item-wise layering designed for Chip Probe Map (CP Map) and Delta Fuse Count Map (DFC Map) from the industrial production environment, and then conduct a comprehensive semantic analysis for two types of wafer maps. Meanwhile, we design an open-world learning framework for two types of wafer maps, consisting of two components: (1) M3Net for known-class image classification, which explores unique properties of CP Map and DFC Map from three perspectives: Multi-task learning, Multi-scale feature fusion and Multi-label classification loss function, and (2) Parallel Support Vector Machine (PSVM) for novel-class image detection. The experimental results on the industrial dataset demonstrate that M3Net outperforms existing methods in classifying known failure patterns, while the PSVM module effectively detects potential novel failure types for expert review. This work provides a scalable and adaptive solution for analyzing wafer map variations from the industrial production environment.
Keywords:
Wafer map analysis
Open-world learning
Semiconductor manufacturing
Failure pattern detection
Industrial data preprocessing
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W