.png){kind=icon}
.png){kind=icon}
Return
Global feature identification layer for mixed-type wafer bin map classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126709.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the semiconductor manufacturing process, classifying defect patterns in wafer bin maps (WBMs) is essential as they reveal the root cause of defects. Engineers can improve the process by identifying the defect type to increase the yield. Recently, as semiconductor manufacturing processes have become more complex, mixed-type defects have become increasingly prevalent in WBMs, complicating their classification. In previous studies, complex, large-scale models with numerous parameters have been employed to address this complication. Although these models exponentially increase the number of parameters, their classification performance often becomes saturated, achieving only marginal improvements. Preprocessing methodologies such as denoising and feature generation can be applied to WBMs to address this problem, but these methods can potentially reduce performance. In particular, feature generation methods need help in extracting high-quality features from mixed-type defect WBMs because of their complexity. We propose a global feature identification layer, which is a deep learning layer with limited parameters, and a preprocessing method to overcome these challenges. The proposed layer can extract feature maps from mixed-type defect WBMs, even including those with overlapping defects, by separating features for single defects in the WBMs. The proposed layer performs preprocessing and extracts feature maps by multiplying the layer elementwise with the input data. During training, the layer is also trained along with the model. The experiments using the MixedWM38 and simulated datasets demonstrate that the proposed layer can be applied to any vision model utilized for the analysis of mixed-type WBMs, including stateof-the-art models, enhancing their performance, robustness, and interpretability.
Keywords:
Wafer bin map preprocessing
Deep learning based preprocessing
Mixed-type defects
Wafer bin map classification
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available